Testing Macroeconometric Models

Testing Macroeconometric Models

Ray C. Fair

Harvard University Press
Cambridge, Massachusetts London, England 1994

Copyright © 1994 by the President and Fellows of Harvard College All rights reserved Printed in the United States of America This book is printed on acid-free paper, and its binding materials have been chosen for strength and durability. Library of Congress Cataloging-in-Publication Data Fair, Ray C. Testing macroeconometric models / Ray C. Fair. p. cm. Includes bibliographical references and index. ISBN 0-674-87503-6 (acid-free paper) 1. United States—Economic conditions—1945– —Econometric models. 2. Econometric models. I. Title HC106.5.F32 1994 330.973—dc20 94-13914 CIP

To my son, John, who wasn’t born in time to make the dedication of the previous book

. . 2. . . .8 Further Aggregation . . . . . . . . . . . . . . 2. . . . . . . . .1 Background . . . . . . . . . . . . . 2 Theory 2.2. . . . . . . . . .1. . . .5 Looking Ahead . . . . .1. . . . . . 2. . 2.3 Firm Behavior . . . . . .5 Links in the Model . . . . . . . . . . . .1 Transition from Theory to Empirical Specifications . .1 Background .5 The Complete Model . . Variables. . . . . . . . . . . . . . . . . . . . . .3 The Real Business Cycle Approach 1. . . . . . . . . . . . . . . . . . 3 The Data. . . . . . . . . . . . . . .2 Household Behavior .Contents Preface 1 Introduction 1. . . . . . . . . . . 2. . . . . . . . .4 The New Keynesian Economics . . . . . . . . . . . . 1. . . . 2. . . .1 One Country . . . .2. .2. . . . . . . . . . . . . .1. . . . . .7 The Use of Reaction Functions 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . . .1. . . . . . . . . . . . .2 Two Countries . . . . . .2. .2 The US Model . . . . . . .3 Equations . . . 1. . . . . . . . . . . . .4 Bank and Government Behavior 2. . . . . xxi 1 1 4 11 14 15 17 17 17 19 20 21 22 24 24 25 25 29 30 30 37 38 39 39 40 . . . . .2. .4 Closing the Model . . .1. .6 Properties of the Model . 2. . . . . 2. . .2. . . . . .2. 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Notation . . . . . . . 2. . . . . . . . and Equations 3. . . . .1 Background . .2 The Cowles Commission Approach 1. . . . 3. . 2. . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . 3. . . . . . . . . . . . 4. . . . 5.2. . . . . . . . . . . .3. .2 Equation 1. . . .2. . . . . . . . . 3. . . . . . . . .4 Equation 3: I : Fixed Investment . . . . . 3. . . . . . . . 6. . . . 6. . . . . . . . . . . . 5. . . . . . . . . .10 Additional Comments . . . . . . 4. . . . . . . .3 4 Estimating and Testing Single Equations 4. . 5. .5 Chi-Square Tests . .4 The Main Firm Sector Equations . .4 The Identities . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . 3. . . . . . . . . The ROW Model .1 Introduction . .3 Variable Construction . . .3 Estimation of Equations with Rational Expectations 4. . . . . . . . . . . . . . . .1 The Tables (Tables B. 5. .3. 4. . . . . . . . .1 The Tables (Tables A. . M: Merchandise Imports . . . . . . .7) . .5 The Stochastic Equations . .2. . . .2. . .6 The Linking Equations .1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 . . . . . .5 The Stochastic Equations . . . . . .3 Equation 2: C: Consumption . . . . . . . 5. . 5. . . . . .3. . . .2 Household Expenditure and Labor Supply Equations 5. .4 Two Stage Least Absolute Deviations . . . . . . . . . . . . . . .7 The Import Equation .3 Variable Construction . . .8) . .viii 3. .9 Interest Payments Equations . . . . . . . . . . . . . . . . The Stochastic Equations of the ROW Model 6. . . . .1–B. . 3. . .4 The Identities . 40 40 43 53 54 55 55 56 56 59 60 61 63 63 65 65 70 71 75 78 83 83 88 102 106 121 125 128 130 133 137 141 141 145 145 149 3. . . . . . .1–A. . . . . . . . . . . . . . . . . . . 5. . . . 3. . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . .6 Stability Tests . .1 Introduction . . . . . CONTENTS . . . . . . . 4. . . .2 Two Stage Least Squares . . . 5 The Stochastic Equations of the US Model 5. . . .2 The Raw Data . . . . . . . . . . . . . . . . . . . . . . . .3. . . 5. . . . . . . . . . . . . . . . . . . . .8 Government Sector Equations . . . . . . . . . .3 Money Demand Equations . . . . 4. . . . . . . . . . . . . . . . .6 Financial Sector Equations . .2 The Raw Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . .5 Other Firm Sector Equations . . . . 3. . . . . . . . . . 3. . . . . . . 6. . . .2. . . . . 3. . . . . . . . . . . .7 Tests of Age Distribution Effects . . . . . . . . . . . . . .

. . . .7 Comparing Predictive Accuracy . . . .4 Median Unbiased Estimates . 8. . . . . . 8. . .5 6. . . . . . . . . . . . . . . . 192 7. . . . . . 6. . . . . . . . . . . . . . . ix 156 156 157 159 163 164 169 170 171 175 . 201 7. .14 6. 8. . .8 Comparing Information in Forecasts . . Equation 8: RB: Long Term Interest Rate . . . . . Equation 14: L1: Labor Force—Men. . . . . . . . . .10 6. Equation 9 E: Exchange Rate . . . . . . . . . 183 7. . . . .16 The Trade Share Equations . . . .6 VAR and AC Models for Comparison Purposes . . . . . . . . . . . . . . . Equation 13: J : Employment . . . . . . . . .15 Equation 4: Y : Production . . . . . . . .4 Asymptotic Distribution Accuracy . . . . . . . . . . . . . . . .12 6. . . . . . 8. . . . . . . . . Equation 6: M1: Money . . . . . Equation 7: RS: Short Term Interest Rate . . . . . . . . . . . . . . . . . . 176 . . . . . . . . . . . . .1 Notation . . . .6 6. . . . . . . . . 184 7.8 Estimating Event Probabilities . . . . . . . . . 6. . . . . . . .7 6. . . .CONTENTS 6. . . . . . . . . .11 6. . . . . . .3 Stochastic Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . .3 MU Estimates of the US Model . . . . . . .9 Summary of the Test Results . . . . . . . . 186 7. . . . . . . . . . . . . Equation 11 P X: Export Price Index . . . . . .2 3SLS and FIML . . . . . . . . . Equation 5: P Y : Price Deflator . . . . . . . . . . . . . . . . 8. . . . . . . . . . . . . . . . . . . 219 219 220 221 223 225 229 237 241 243 8 . . . . . 189 7. . . . . . . . . 8. . . . . . . . . . . . . . .2 US+ Model . . . . . . . 195 7. . . . . . . . Equation 15: Labor Force—Women . . . . . . .17 Additional Comments . . . . 177 . . . . . . 184 7. . . . . . . . . . . . . . .10 Full Information and Solution of Rational Expectations Models 203 Estimating and Testing the US Model 8. . Equation 10 F : Forward Rate . . . . . 191 7. . . . . 8. Equation 12: W : Wage Rate . . . .6 Predictive Accuracy . . . . . . . . . . . . . . . . . . L2: .1 Introduction .5 Examining the Accuracy of Asymptotic Distributions . . . . . . . . . 180 7 Estimating and Testing Complete Models 183 7. . . . . . . . . .9 Estimating Event Probabilities . .7 Comparing Information in Forecasts 8. . . . . . . . . .13 6. . . . . . . . . . . . . . . .8 6. . . . . . . . .9 6. . . . . . . . . . . . . . . . .5 A Comparison of the Estimates . . . . .

. 10. . . . . . 9. 11. . . . . . . . . . . . .3. . . .3 Sensitivity of Fiscal Policy Effects to Assumptions about Monetary Policy . . . . . . . 11. .3. . . . . .3. . . . .2 A Monetary-Policy Experiment: RS Decrease . . . . . .7 The Deficit Response to Spending and Tax Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Multipliers from a Price Shock: P I M Increase . . . . . 11. . . . . . . . . . . .2 A General Discussion of the US Model’s Properties .1 Fiscal Policy Variables .2. . . . . . . . . . . 11. . 10. . . .7 Is Monetary Policy Becoming Less Effective? . . . . . 10. . 11. .3. . . . .4 Sources of Economic Fluctuations in the US Model . . . . .2 Stochastic Simulation . . . . . . . . . . . . 9. CONTENTS 247 247 247 251 252 256 261 261 262 262 263 265 268 270 273 275 275 276 281 281 292 293 293 297 298 298 303 311 315 320 328 . . . . . . . . 10 Analyzing Properties of Models 10. . . . 11. . .8 What if the Fed had Behaved Differently in 1978 and 1990? .6 Sensitivity of Multipliers to the Rational Expectations Assumption in the US Model .3 Computing Multipliers and Their Standard Errors . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . .1 Introduction . . . .4 Within Sample RMSEs . .1 Introduction . . . . . . . . . . . . . . . . .3. . .x 9 Testing the MC Model 9. . . . . . . . . . . . . 11. . . . . . . . . . . . .4 Sensitivity of Fiscal Policy Effects to the Specification of the Import Equation . . . . . . . .1 Deterministic Simulation . . 11. . . . . . . .3 The ARMC Model . . . .3 Sources of Economic Fluctuations .3. 11. . . . .2 The Size and Solution of the MC model 9. . . . . 11.1 Introduction .2. . . . . . . . . . 11. . .5 Optimal Choice of Monetary-Policy Instruments in the US Model . . . . . . . .4 Optimal Monetary Instruments . . 10. .3. . . . . . . . . . . . . .5 Outside Sample RMSEs . . . . 10. . . . . . 11. . . . . . . . . . .5 Optimal Control . . . . . . . . . . . . . . . . . 10. . . . . . . . . 11 Analyzing Properties of the US Model 11. . . . . . . . .5 Sensitivity of the Multipliers to Alternative Coefficient Estimates . . .6 Counterfactual Multiplier Experiments . . . . . . . . . . . . . . . . . 11. . . . . . 9.2 Computing Multipliers and Their Standard Errors 10. . . . . . . . . . . . . . . . . .

12. . . .4 Common Results Across Countries . . . . . . . . . .3 Computing Multipliers . . . . . . . . . . . .2 A General Discussion of the MC Model’s Properties . . . . 13 Conclusion Appendix A Tables for the US Model Appendix B Tables for the ROW Model Bibliography Index xi 333 333 334 338 348 349 353 393 407 419 . . . . . . . .CONTENTS 12 Analyzing Properties of the MC Model 12. . . . . . 12. . . . . . . . . . . . . . . . 12.1 Introduction . . . . . . .

.

. . . . . .9 5. . . . . . . . . . . . . . .5 5. . . . . . . . . . . . Estimated Percentages of Excess Labor and Capital . . . . . . . . . . . . . . . 33 36 50 77 91 93 95 96 98 99 100 101 104 108 112 114 117 118 119 121 105 122 123 124 125 126 . . . . . . .7 5. . . . . . . . . . . . . . . . . . . . . . . Equation 7 .2 3. . . . . . . . . . .14 5. . . . . . . Critical Values for the AP Statistic for λ = 2. . . .13 5. . . . . . . . . . .6 5. . . . . . . . . . . Equation 9 . . . . . . . . . . .10 5. . . . . . . .1 5. . . . . . . . . .2 5. . .12 5. . . .List of Tables 2. . . . . . . . . . . . . . . . . . . Equation 8 . . . . . . . . . . . . . . Equation 4 . . . .3 5. . . . . .23 Simulation Results for the Two Country Model . . . . . . Equation 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .75 . . . . . . . . . . . . . . . . . . . . . . . . . .21 5. . . . . . . . . . . . . . . . . . . . . . . . . . Equation 1 . . . . . . . . . . . . . . . . . . . . . . . . . Equation 10 Equation 11 Equation 12 Equation 13 Equation 14 Equation 15 Equation 16 Equation 17 Equation 18 Equation 20 Equation 21 Equation 22 Equation 23 . . . . . . . . . . . . . . . . . . . .1 5. . . . . . . . . . . . . . . . . . . . . . . . Equation 6 . . . . . . . . . . . . . . . Equation 5 . . . . . . . . . . .8 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 5. .22 5. . Equation 3 . . . .17 5. . . .4 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15 5. . . . .1 2. . .18 5. . . . . . . . .1 4.11 5. . . . Summary of the Experiments . . . . . . . . . . . . . .20 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .1a 6. . . . . . . . . . . . . . . . . .7a 6. . . . . . . . . . . . . . .13b 6. . . . . . . . Test Results for Equation 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .26 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equation 11 . . . . . . . . . . Test Results for Equation 13 Equation 14 . . . Equation 6 . . . . . . . . . . . . . . . .30 6. . . . . . .xiv 5. . . . . Equation 4 . . . . . . . . . .14b 6. . . . . . . . . . . . . . . Test Results for Equation 1 . . . . . . . . . . . . . . . . .13a 6. . . . . . . . . . Equation 5 . . . . . . . . . . . . . . . .15b 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25 5. . . . . . . . . . . . . . . . . . . . . . . . . . . Test Results for Equation 7 . . .4b 6. . . . . . . . . 212 Model 3: Six Equations . . . . . . . . . . . . . . . . . Test Results for Equation 11 Equation 12 . . . . . . . .14a 6. . .2b 6. . . . . . . . . . . . Equation 2 . . . . . Test Results for Equation 4 . . . . . . . . .11a 6. . . . . . . . . . . . . . . Equation 10 . . . . Test Results for Equation 6 . . . . . . .3a 6. . . . . . . . . . . . . . . . . . . . . . . . . . .5a 6. . . . . . . . . Test Results for Equation 14 Equation 15 . . . . . . . . . . .27 5. . . . Test Results for Equation 15 FIML Results for Three Models . . . . . . Test Results for Equation 8 . . . . . . .28 5. . Test Results for Equation 5 . . . . . . . .12a 6. . . . . . . . . . . . . . . Test Results for Equation 9 . . . . .9b 6. . . . . . . . . . . . . . . . . .8b 6. . . . . . . . . . . . . . . . . . . . . Equation 7 . . .15a 6. . . . . . .1b 6. . . . . . . . . Equation 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equation 8 . . . . . . . . . . . . . LIST OF TABLES . .4a 6. . . . . . . .2a 6. . . .24 5. . . . . . . .6a 6. . . . . .12b 6. . . . . . . . . .11b 6. . . . . . . . . . . . . . . . . . . . .3b 6. . . . .6b 6. . . . . . . . . . . . . . . . . . . . Test Results for Equation 2 . . . . Test Results for Equation 12 Equation 13 . . . . . . . . . . . . . . . . . . . . . . . Equation 3 . . . . . . . . . . . . . . . . .10a 6. . . . .7b 6. . . . . . . . . . . . . . . .8a 6. . . . . .1 7.9a 6. . . . . . . . . . . . .5b 6. . .2 Equation 24 Equation 25 Equation 26 Equation 27 Equation 28 Equation 30 . . 127 128 106 129 131 133 146 147 148 149 150 151 152 153 154 155 158 159 160 161 162 163 166 167 170 172 173 174 175 176 177 178 179 180 181 Equation 1 . . . . 213 . . . . . . . . . . . . . . . . . . .

. . . . . . . . .12 AC Versus VAR5/2 and VAR4 . . .1 xv Stochastic Simulation Results for Model 3 . . . . . . . . . . . 222 224 226 228 230 236 238 239 242 Ratios of Within Sample RMSEs 1972–1990: MC/ARMC 254 US RMSEs: US Alone Versus US in MC Model . . . . . 11. . . . . . . . . . . . .3 11. . . .3 8. . . . . . . . Estimates of Probability Accuracy . . . . . .6 Estimated Multipliers and Their Standard Errors for an Increase in P I M of 10 percent . . .4 Estimated Multipliers for Three Experiments and Two Versions of the Import Equation . . . . . .9 9. . . . . . . . . and FIML Estimates RMSEs for Five Sets of Coefficient Estimates for 1954:1– 1993:2 for the US Model . . . .1 8. .2 8. . . . . . . . . . . . . . Comparison of 2SLS. . . . . . . . 11.10 Estimated Precision of the Stochastic Simulation Estimates for Real GDP . . . 257 Ratios of Outside Sample RMSEs 1987–1990: MC/ARMC 258 Estimated Multipliers and Their Standard Errors for Eight Fiscal Policy Experiments . .3 Estimated Multipliers and Their Standard Errors for an Increase in COG under Three Monetary Policy Assumptions 11. . . . . . . . . . . . . 282 291 294 296 297 299 301 304 306 311 . . . . . 2SLAD. . . . . . . . .3 8. . . . . . . . .2 Estimated Multipliers and Their Standard Errors for a Decrease in RS of One Percentage Point . . . . . .LIST OF TABLES 7. . . . 11. . . . 11.7 Estimated Effects on the Federal Government Deficit of Six Fiscal Policy Experiments . . . . . . . . . .8 8. . . . . . . . . . RMSEs of Outside Sample Forecasts for Four Models for 1976:3–1993:2 . . . . .6 8. . . Asymptotic Distribution Accuracy . .7 8. . . . . . . . . . . . . . . . 11. . . . US Model Versus Three Others: Estimates of Equation 7. . . . . . . . . . . . . . . . . . . . . Estimated Standard Errors of Forecasts for Four Models .2 9. . .9 Variance Decomposition for the Nonfarm Price Deflator . . . .5 8. 217 Estimated Bias of 2SLS Lagged Dependent Variable Coefficient Estimates . . . 11. 11. . . 3SLS. . . . . . .1 9. . . . . . 11. . . . . . . . .4 8. . . . . .8 Variance Decomposition for Real GDP .5 Estimated Multipliers for a COG Increase for Alternative Sets of Coefficient Estimates .

. . . . . . .5 A. 344 The Six Sectors of the US Model .7 A. .18 Estimated Economy if the Fed had Lowered Interest Rates in 1990:3–1993:2 . . . . . 11. . . . . The Countries and Variables in the MC Model . 11.1 12. . . . . . . . . . . .3 A. Government Spending Increase . . . . . . . . The Raw Data Variables for the US Model . . . . . . . . .S. . . . . . . . . .xvi LIST OF TABLES 11. 11. . 343 Multipliers for Own Government Spending Increases . . 11. . . . . .13 Percentage Difference Between the Variance Under the Optimal Policy and the Variance Under the Interest Rate Policy 11. . . .17 Estimated Economy if the Fed had not Raised Interest Rates in 1978:3–1983:4 . . First Stage Regressors for the US Model for 2SLS and 3SLS Solution of the US Model Under Alternative Monetary Policy Assumptions . . . . 12. . . . .2 B. . . The Equations of the US Model .3 B. . . . . .1 A. . . . The Variables in the US Model in Alphabetical Order . . . . . The Equations for a Given Country . . . .2 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Government Spending Increase . .2 A. . . . . . . . 339 Multipliers for a U. . . . . . . . . . . . . . The Variables for a Given Country in Alphabetical Order . . 11. . . . . . . .8 A. . .9 B. Construction of the Variables for the US Model . Cross-Reference Chart for the US Model . . . . . .3 A.15 Comparison of the Actual and Alternative Economies . . . . Links Between the National Income and Product Accounts and the Flow of Funds Accounts . . . . . . . . .6 A. . . . .4 A. .12 Percentage Difference Between the Variance Under the Money Supply Policy and the Variance Under the Interest Rate Policy: No Shocks to the Money Equations . . . .1 B. .4 312 313 314 316 322 326 329 330 Multipliers for a U. . . . . . . . . . . . . .11 Percentage Difference Between the Variance Under the Money Supply Policy and the Variance Under the Interest Rate Policy . . . . Equations that Pertain to the Trade and Price Links Among Countries . . . .S. . . . . . . . .16 Estimated Multipliers in the Actual and Alternative Economies for a Decrease in RS of One Percentage Point . 355 356 362 370 378 380 386 389 390 395 398 400 402 . . . . . . 11. . . . . . . . . . . . . .14 Estimated Multipliers for the US Model with Rational Expectations . . . . . . . . . . .

. . . . .6 B. . . . . . . . . . . . . . . 403 The Procedure Used to Create Quarterly Data from Annual Data . . . . . . . . . . . . . . . . . . . . . . . . 405 . . . . . . .5 B. . . . . . . . . . . . .LIST OF TABLES B.7 xvii Links Between the US and ROW Models . 404 Construction of the Balance of Payments Data: Data for S and T T . .

.

245 . . . . 244 Estimated Probabilities for Event A for VAR5/2 . . .List of Figures 8. . .1 8. . . .2 Estimated Probabilities for Event A for US+ . . .

.

and it discusses and applies various econometric techniques to it.S. Taylor. Some of the articles are joint. Dominguez. Kathryn M. The coauthors are Donald W. with particular emphasis on testing. This book is in large part an application of the Cowles Commission approach to current macro data. including my U. It presents the current version of my multicountry econometric model. Some of the work is new for this book and has not been published elsewhere. Robert J. I have indicated in a footnote at the beginning of each relevant section the article upon which the material in the section is based. Andrews.Preface This book brings together my macroeconometric research of roughly the last decade. Shiller. Chapter 1 gives a general view of where I think my work fits into the literature. It is a sequel to my previous book. Fair March 1994 . It is a rallying cry for the Cowles Commission approach. an approach I feel too many academic researchers abandoned in the 1970s. Parke. model. William R. and John B. K. which brought together my macroeconometric research through the early 1980s. Fair (1984). All the empirical work using the model has been updated for this book. Ray C.

.

Testing Macroeconometric Models .

.

The first was the commercialization of macroeconometric models. From Tinbergen’s (1939) model building in the late 1930s through the 1960s.2 Structural econometric models were specified. and at its peak it contained nearly 400 equations. Basic research gave way to the day to day needs of keeping the models up to date. building on the earlier work of Klein (1950) and Klein and Goldberger (1955). 1969)]. and the field of macroeconomics is no exception. One of the major macroeconometric efforts of the 1960s. and it was laid to rest around 1972. Although much was learned from this exercise. The dominant methodology of this period was what I will call the “Cowles Commission” approach. Fromm. which has the same title as this book 2 See Arrow (1991) and Malinvaud (1991) for interesting historical discussions of econometric research at the Cowles Commission (later Cowles Foundation) and its antecedents. 3 The commercialization of models has been less of a problem in the United Kingdom than 1 The 1 .1 Background1 Interest in research topics in different fields fluctuates over time.3 The second event was Lucas’s (1976) discussion in this section and some of the discussion in the rest of this chapter is taken from Fair (1993d). Two important events in the 1970s contributed to the decline in popularity of the Cowles Commission approach. Klein. the model never achieved the success that was initially expected. was the Brookings model [Duesenberry. and then analyzed and tested in various ways. This changed the focus of research on the models. estimated. This model was a joint effort of many individuals. there was considerable interest in the construction of structural macroeconomic models. and Kuh (1965.” and of meeting the special needs of clients. of subjectively adjusting the forecasts to make them “reasonable.1 Introduction 1.

T. and it discusses various econometric techniques. in the mid 1960s. model. and I have continued to try to make progress using this approach. The program was initiated by Robert Hall. I have chosen techniques that I think are important for macroeconometric work. which in turn generated a counter response in the form of new Keynesian economics More will be said about these latter two areas later in this chapter. The book is a sequel to Fair (1984). I have never lost interest in structural models. which brought together my macroeconometric research through the early 1980s. I played a minor role in this development.I. Most of the techniques that are discussed are new since the earlier book was written.2 1 INTRODUCTION critique. a computer program that provided an easy way of using various econometric techniques.I. models and their associated databases are made available to academic researchers through the Bureau. and all of this work is discussed. Various U.T. This book brings together my macroeconometric research of roughly the last decade.T. Many hours were spent by many students in the basement of the Sloan building at M. which argued that the models are not likely to be useful for policy purposes. The theory behind the econometric model has changed very little from that described in the earlier book. including my U. Advances in computer hardware have considerably lessened the computain the United States. . In the choice of econometric techniques to discuss. and it soon attracted many others to help in its development. I have been idiosyncratic in the present book. I continue to believe that the Cowles Commission approach is the best way of trying to learn how the macroeconomy works. Perhaps because of fond memories of my time in the basement of Sloan. On the other hand.K. as I was in the earlier book. working on various macroeconometric equations using an IBM 1620 computer (punch cards and all).S. all the empirical work is new (because there is nearly a decade’s worth of new data). In 1983 the Macroeconomic Modelling Bureau of the Economic and Social Research Council was established at the University of Warwick under the direction of Kenneth F. but these by no means exhaust all relevant techniques. and so the theory is only briefly reviewed in the present book. The Lucas critique led to a line of research that culminated in real business cycle (RBC) theories. This was a period in which there was still interest in the Brookings model project and in which intensive work was being carried out on the MPS (M. This was also the beginning of the development of TSP (Time Series Processor).–Penn–SSRC) model. Wallis. It presents the current version of my multicountry econometric model.I. My interest in structural macroeconomic model building began as a graduate student at M.

The United States part of the MC model will be called the “US” model. The non United States part of the MC model will be called the “ROW” (rest of world) model. There is considerable stress in this book on testing (hence the title of the book). There are also estimated trade share equations for 44 countries plus an “all other” category. and this has greatly expanded the ways in which models can be tested and analyzed. In particular. and it does not include the trade share equations. and this discussion is not repeated in the present book. are much more straightforward to use in FP than they are in programs like TSP. System wide techniques.1 BACKGROUND 3 tional burden of working with large scale models. The FP program is discussed in Fair and Parke (1993). . labelled “AO. deterministic and stochastic simulation. This program is joint work with William R. It also partly serves as an outline of the book. The complete multicountry econometric model will be called the “MC” model. 3SLS estimation. and this is reflected in the current book. and testing is clearly an essential ingredient in this process. the availability of fast. Many of the techniques discussed in this book require the use of stochastic simulation.1. optimal control techniques. All the techniques discussed in the earlier book and in the present book are programmed into the Fair-Parke (FP) program. This model consists of estimated structural equations for 33 countries. It consists of estimated equations for the United States only. Much of my work in macroeconomics has been concerned with testing. The rest of this chapter is a discussion and defense of the Cowles Commission approach and a criticism of the alternative approaches of real business cycle theorists and new Keynesian economists. inexpensive computers has made stochastic simulation routine. Whereas TSP was designed with single equation estimation in mind. The FP program expands on TSP in an important way. and techniques for rational expectations models. such as FIML estimation. FP was designed to treat all equations of a model at the same time. Some of the more advanced techniques are applied only to the US model. My primary aim in macroeconomics is to develop a model that is a good approximation of how the macroeconomy works.” The trade share matrix is thus 45 × 45. Parke. both the testing of single equations and the testing of overall models.

2 The Cowles Commission Approach4 Specification Some of the early macroeconometric models were linear. . it will be useful to consider a simpler example.1) where yt is an n–dimensional vector of endogenous variables. . I am using the phrase to mean the actual approach used by structural macro model builders. . and 6 form an example of the use of theory in the specification of an econometric model. Heckman (1992). xt . of material in this section and in Sections 1. and the specification of the stochastic equations is discussed in Chapters 5 and 6. xt is a vector of predetermined variables (including lagged endogenous variables). .4 1 INTRODUCTION 1. Chapters 2. If in fact some variables are not stationary. and 3) the probability structure for uit . of most macroeconometric model building work is that the variables are trend stationary. The model will be written as fi (yt . where the normalization used is to set the coefficient of this variable equal to minus one. 5. either explicit or implicit. αi ) = uit . . testing it. 4 Part . and uit is the error term for equation i for observation t. and then stopping. 5 In modern times one has to make sufficient stationarity assumptions about the variables to make time series econometricians happy. but this soon gave way to the specification of nonlinear models. n). I will thus use “Cowles Commission approach” in a general way. αi is a vector of unknown coefficients. Before moving to the theory in Chapter 2. uses the phrase to mean the procedure of forming a hypothesis (from some theory). 2) the functional form of each equation. For equations that are identities. (t = 1. Specification consists of choosing 1) the variables that appear in each equation with nonzero coefficients.5. T ) (1. The theory is discussed in Chapter 2. . this may make the asymptotic distributions that are used for hypothesis testing inaccurate. uit is identically zero for all t. The assumption. Perhaps a better phrase would have been “traditional model building approach. but it should be kept in mind that there are narrower definitions in use. It should be noted that I am using the phrase “Cowles Commission approach” in a much broader way than it is sometimes used. and this is done in Section 7. where there is much back and forth movement between specification and empirical results. for example. In most cases there is an obvious left hand side variable for the equation. Consequently.5 is taken from Fair (1992).3–1. Heckman argues (correctly in my view) that this is a very rigid way of doing empirical work.5 Economic theory is used to guide the choice of variables. however. the accuracy of the asymptotic distributions that are used in macroeconometric work can be examined. This is the variable considered to be “explained” by the equation. (i = 1. only the nonlinear case will be considered here. Fortunately. .” but this is awkward. It will be seen that the asymptotic distributions appear fairly accurate. .

. β) 1 (1. it is not generally ∗ possible to find analytic expressions for C1 and L∗ . 1 this problem is not analytically tractable. . W is the wage rate. CT . . Maximize E0 U (C1 .3) where C is consumption. and 3) the parameters of the utility function. A is the terminal value of assets.5 are not in general known. · · · E0 PT . P is the price level. and t = 1. . are replaced by their expected values. H is the total number of hours in the period. A. · · · . . The aim of the empirical work is to try to estimate equations that are approximations of equations 1. one can write the expressions for C1 and L∗ as 1 ∗ C1 = g1 (A0 .4 and 1. . E0 PT .6 The functional forms of equations 1. 6 If . however. . E0 P1 . . E0 W1 . P . and rt . the random variables. E0 is the expectations operator conditional on information available through time 0. . First. E0 WT . E0 P1 .5 simply state that the optimal values for the first period are a function of 1) the initial and terminal values of assets. t = 1. 2) the expected future values of the wage rate.4 and 1.2 THE COWLES COMMISSION APPROACH 5 Consider the following maximization problem for a representative household. E0 rT . E0 Wt . E0 Pt. . 1 The approach that I am calling the Cowles Commission approach can be thought of as specifying and estimating approximations of the decision equations.5. . . E0 WT . β) (1. . . A is the level of assets. and E0 rt . and the interest rate.2) (1. E0 rT . E0 W1 . L is leisure.4) L∗ = g2 (A0 . · · · .4 and 1. t = 1. and E0 r1 should be replaced by their actual values in equations 1. E0 P1. · · · . A. T . S is saving. . T .4 and 1.5) where β is the vector of parameters of the utility function. it is possible in principle to solve for the optimal ∗ values of C and L for period 1. . LT ) subject to St = Wt (H − Lt ) + rt At−1 − Pt Ct At = At−1 + St ¯ AT = A (1. This approach in the context of the present example is the following. T . . Experimentation consists in trying different information for period 1 is available at the time the decisions are made.1. E0 r1 . L1 . Given this replace∗ ment. . · · · . Wt . denoted C1 and L∗ . Equations 1. In other words. r is the one period interest rate. . In general. Given A0 and the conditional distributions of the future values of W .5. then E0 W1 . Pt . · · · . the price level. . and r. E0 r1 .

E0 WT might be assumed to lie on a low order polynomial or to be geometrically declining.4 and 1. Since C0 and L0 do not appear in 1.5 1 are long run “desired” values. Let Et−1 y2t+1 denote the expected value of y2t+1 . . the parameters of the expected future values might be restricted in order to lessen the number of free parameters to be estimated. 0<λ<1 (1. the entire model is estimated at the same time by full information maximum likelihood.) If expectations are assumed to be rational. and they can be interpreted as approximation errors. as specified in equation 1. . . This procedure is ad hoc in the sense that the adjustment equation is not explicitly derived from utility maximization. One can. For example. This setup can handle the assumption of rational expectations in the following sense. The usual adjustment equation for consumption would be ∗ C1 − C0 = λ(C1 − C0 ). . and y might be the wage rate. which would justify the use of lagged dependent variables in the empirical approximating equations for 1. Et−1 y2t+1 is replaced by y2t+1 .4 and 1.4 and 1.6) which adds C0 to the estimated equation.1. . (L2 − L1 )2 . however. Because of the large number of expected values in equations 1. In the limited information case. In the full information case. (L1 − L0 )2 . in the utility function 1. It is often the case when equations like 1.4 and 1.5 are estimated that lagged dependent variables are used as explanatory variables. (C2 −C1 )2 . and the equation is estimated by Hansen’s (1982) generalized method of moments (GMM) procedure. there is only a partial adjustment of actual to desired values.6 1 INTRODUCTION functional forms and in trying different assumptions about how expectations are formed. where the expectation is based on information through period t − 1. (This equation might be an equation explaining consumption.5. the expectational assumptions usually restrict the number of free parameters to be estimated. . how can one justify the use of lagged dependent ∗ variables? A common procedure is to assume that C1 in 1. the parameters for E0 W1 . . .4 and L∗ in 1.5.4 and 1.5.1.5. this equation and equations like it can be estimated by either a limited information or a full information technique. This would add the variables C0 and L0 to the right hand side of equations 1. Again.2. It is then assumed that because of adjustment costs. assume that there are utility costs to large changes in consumption and leisure and thus put terms like (C1 −C0 )2 . The error terms are usually assumed to be additive. where the restriction is imposed that the expectations of future values of variables are equal to the model’s predictions of the future values. and assume that Et−1 y2t+1 appears as an explanatory variable in equation i in 1.

1. As noted above.10 that computational advances have made even the estimation of models with rational expectations by FIML computationally feasible. These distributions can then be compared to the asymptotic distributions that are typically used for hypothesis testing. to obtain median unbiased (MU) estimates of the coefficients of macroeconometric models.4. 2SLS is discussed in Section 4.2. which is strictly valid only in the linear quadratic setup. it is now possible using stochastic simulation and reestimation to compute “exact” distributions of estimators that are used for macroeconometric models. and FIML is discussed in Section 7.5 and applied to the 2SLS estimates of the US model in Section 8. . These techniques are discussed in Fair (1984). estimation techniques are available that handle the assumption of rational expectations. It is also possible. simultaneous. The procedure for computing exact distributions is explained in Section 7. The specification also uses the certainty equivalence procedure. the asymptotic distributions may not be good approximations. Techniques that do not take account of the correlation of the error terms across equations (limited information techniques) include two stage least squares (2SLS) and two stage least absolute deviations (2SLAD). More will be said about this below. although they are only briefly discussed here. Only approximations of the decision equations are being estimated.10. as discussed in Section 7.4. and these estimates are also computed in this book. even under the assumption of rational expectations. nonlinear. If some variables are not stationary. Techniques that do account for this correlation (full information techniques) include full information maximum likelihood (FIML) and three stage least squares (3SLS). Finally. including their modifications to handle the case in which the error terms follow autoregressive processes.3. and has error terms that may be correlated across equations and with their lagged values. The specification is thus subject to the Lucas (1976) critique. They are used in the current book. It will be seen in Section 7. A number of techniques have been developed for the estimation of such models.2. Estimation A typical macroeconometric model is dynamic. and 3SLS and FIML in Section 7. Hansen’s method is discussed in Section 4.2 THE COWLES COMMISSION APPROACH 7 The specification that has just been outlined does not allow the estimation of “deep structural parameters. 2SLAD in Section 4.” such as the parameters of utility functions.4.

” The expected values of the error terms are usually assumed to be zero. There is always a danger in this business of “data mining. Other criteria include mean absolute error and Theil’s inequality coefficient.8 Testing 1 INTRODUCTION Testing has always played a major role in applied econometrics. The danger with this type of searching is that one finds a model that fits well within the estimation period that is in fact a poor approximation of the economy.” which means specifying and estimating different versions of a model until a good fit has been achieved (say in terms of the RMSE criterion). and 4) lagged values of the endogenous variables. If a model is poorly specified. Before a complete model is tested. if the properties of the estimated residuals are as expected. Complete models can also be tested. A “stochastic” simulation is one in which 1) the error terms are drawn from an estimated distribution. and if the solution is “dynamic. it should not predict well outside the period for . Equations are discarded or modified if they do not seem to approximate the process that generated the data very well. 3) values of the error terms. if its coefficient estimates are significant and of the expected sign. Given 1) a set of coefficient estimates. 2) values of the exogenous variables. To guard against this. A standard procedure for evaluating how well a model fits the data is to solve the model by performing a dynamic. deterministic simulation and then to compare the predicted values of the endogenous variables with the actual values using the root mean squared error (RMSE) criterion. predictions are many times taken to be outside of the estimation period.” the actual values of the lagged endogenous variables are used for each period solved. When an equation is estimated. the simulation is said to be “deterministic. this is evidence in favor of model A over model B.” the values of the lagged endogenous variables are taken to be the predicted values of the endogenous variables from the previous periods.5–4. If two models are being compared and model A has lower RMSEs for most of the variables than model B. If the solution (simulation) is “static. and 3) the predicted value of each endogenous variable is taken to be the average of the solution values. but here things are more complicated. If one set of values of the error terms is used. one examines how well it fits the data. 2) the model is solved for each set of draws. Sections 4.7 discuss the methods used in this book to test the individual equations. and so in most cases the error terms are set to zero for a deterministic solution. a model can be solved for the endogenous variables. it must be solved. and Chapters 5 and 6 present the results of the tests. and so on.

that accounts for these RMSE difficulties. and because of variation in the initial conditions. for example. which uses stochastic simulation. Even if this type of searching is not formally done. it may be that information beyond the estimation period has been implicitly used in specifying a model. and 3) choosing the version that best fits the period beyond the point. one would also have to wait for more observations to see how accurate the model is. because of variation in the exogenous variables. If a particular model’s estimated variances are in general smaller than estimated variances from other models. The method can handle a variety of assumptions about exogenous variable uncertainty. If this type of searching is done. A more serious problem with the RMSE criterion as a means of comparing models is that models may be based on different sets of exogenous variables. If. The method accounts for the four main sources of uncertainty of a forecast from a model: uncertainty due to 1) the error terms. then one has to wait for more observations to provide a good test of the model. . or the estimated variances from one structural model can be compared to those from an autoregressive or vector autoregressive model. One polar assumption is that there is no uncertainty attached to 7 This is assuming that one does not search by 1) estimating a model up to a certain point. A by-product of the method is an estimate of the degree of misspecification of a model for each endogenous variable.2 THE COWLES COMMISSION APPROACH 9 which it was estimated. and 4) the possible misspecification of the model. and so it can be compared across models.1.7 One problem with the RMSE criterion (even if the predictions are outside of the estimation period) is that it does not take account of the fact that forecast error variances vary across time. this is evidence in favor of the particular model. 2) solving the model for a period beyond this point. and the method can estimate the quantitative importance of the misspecification. one model takes investment as exogenous and a second does not. The forecast error variance for each variable and each period that is estimated by the method accounts for all four sources of uncertainty. even though it may fit well within the period. Forecast error variances vary across time because of nonlinearities in the model. This might then lead to a better fitting model beyond the estimation period than is warranted. Although RMSEs are in some loose sense estimates of the averages of the variances across time. 3) the exogenous variables. Any model is likely to be somewhat misspecified. The estimated variances from different structural models can be compared. In this case. I have developed a method. This type of searching may lead to a model that predicts well outside the estimation period even though it is in fact a poor approximation. the first model has an unfair advantage when computing RMSEs. 2) the coefficient estimates. no rigorous statistical interpretation can be placed on them: they are not estimates of any parameters of a model.

for example. An in between case is to estimate the variance of an exogenous variable forecast error from actual forecasting errors made by a forecasting service—say the errors made by a commercial forecasting service in forecasting defense spending. It is briefly reviewed in Section 7.7. there are a variety of ways in which it can be analyzed. Again. It is sometimes felt that analyzing the properties of a model is a way of testing it. A model to the Cowles Commission was a null hypothesis to be tested. and solving optimal control problems. Hendry and Richard (1982).10 1 INTRODUCTION the exogenous variables. Analysis Once a model has been estimated. stochastic simulation is used for many of these methods. developed in Fair (1993c). but this does not necessarily mean that it is a . examining the sources of economic fluctuations. Another test. A model may be specified and constrained in ways that lead it to have “reasonable” properties from the point of view of the model builder. such as a recession or severe inflation. examining the optimal choice of monetary-policy instruments. developed in Fair and Shiller (1990). of some policy variables. for example.9 and applied in Section 8. One can. Methods for analyzing the properties of models are discussed in Chapter 10. estimate autoregressive equations for each exogenous variable and add these equations to the model. Another method of comparing complete models is to regress the actual value of an endogenous variable on a constant and forecasts of the variable from two or more models. is discussed in Section 7. It is related to the literature on encompassing tests—see.7 of the current book and then used in Section 8. This method was developed in Fair (1980). Tests of the sort just described seem clearly in the spirit of the Cowles Commission approach. This test uses stochastic simulation to estimate event probabilities from macroeconometric models.8. The other polar assumption is that the exogenous variables are in some sense as uncertain as the endogenous variables.8 and applied in Section 8. which could then be tested.6 to compare the US model to other models. This method. is discussed in Section 7. The methods include computing multipliers and their standard errors. Davidson and MacKinnon (1981). where the estimated probabilities are then compared to the actual outcomes. but one must be very careful here. for example. and it is also discussed in Fair (1984). This would produce a model with no exogenous variables. This might be true. and Chong and Hendry (1986). It examines how well a model predicts various economic events.

Chapters 11 and 12 examine the properties of the models.2. estimation. and this is what is done in Chapters 5 and 6. but again this is not a substitute for testing the model. Because specification. Finally. estimation. it is not likely to be a good approximation even if it has reasonable properties. In practice there is considerable movement back and forth from analysis to specification. it will be useful to begin with the utility maximization model in Section 1. This is the analysis part of the Cowles Commission approach. and testing are so closely linked. The theory that has been used to guide the empirical specifications is discussed in Chapter 2. it will be useful to consider the real business cycle approach and the approach of new Keynesian economists from the perspective of the Cowles Commission approach. Before proceeding to a discussion of the theory. a model has what seem to be bizarre properties. Unless a model tests well.3 THE REAL BUSINESS CYCLE APPROACH 11 good approximation of the economy. and testing. on the other hand. the model may be changed and then analyzed again. If. this may mean that the model is not a good approximation even if it has done well in the tests. and tested in these two chapters—Chapter 5 for the US model and Chapter 6 for the ROW model. There is no further specification in these two chapters. The RBC approach to this model would be to specify a particular functional form for the .3 The Real Business Cycle Approach As noted in Section 1.1. it is generally useful to discuss these together. This may indicate that the tests that were performed have low power. It also briefly discusses the transition from theory to empirical specifications. and this is the subject matter for the rest of this chapter. 1. Chapters 5 and 6 combine elements of specification. the RBC approach is a culmination of a line of research that was motivated by the Lucas critique. The complete models are then tested in Chapters 8 and 9—the US model in Chapter 8 and the entire MC model in Chapter 9. The individual stochastic equations are specified. Use of the Cowles Commission Approach for the MC Model To review. If a model’s properties do not seem reasonable. In discussing this approach. This procedure usually results in a model with “reasonable” properties. Chapter 3 presents the data to which the specifications are to be applied. estimated.1. the use of the Cowles Commission approach for the MC model is as follows.

A recent example is Christiano and Eichenbaum (1990). Take the simple RMSE procedure. Although there is some parameter estimation in the RBC literature. The properties of the computed paths of the decision variables are then compared to the properties of the actual paths of the variables. If the parameters are estimated. and Mankiw. deep structural parameters are being estimated (or calibrated). The tone of these lectures is that there is an exciting sense of progress in macroeconomics and that there is hope that in the end there will be essentially no distinction between microeconomics and macroeconomics.g. which are an elegant argument for dynamic economic theory. This procedure would compute a prediction error for a given variable for . where the parameters of their model are estimated using Hansen’s (1982) GMM procedure. Rotemberg. Is the RBC approach a good way of testing models? At first glance it might seem so. since computed paths are being compared to actual paths. and Summers (1985). for example. and autocovariances). there is usually some searching over parameters to find that set that gives good results in matching the computed paths to the actual paths in terms of the particular criterion used. It is hard to overestimate the appeal this has to many people. The RBC approach meets the Lucas critique in the sense that. similar variances. covariances. If the computed paths have similar properties to the actual paths (e. most of the studies calibrate rather than estimate.2 are estimated from the first order conditions—see. Anyone who doubts this appeal should read Lucas’ 1985 Jahnsson lectures [Lucas (1987)]. one would solve the utility maximization problem for the optimal consumption and leisure paths. they are estimated from the first order conditions. Hansen and Singleton (1982). Altug (1989) estimates the parameters of her model using a likelihood procedure.12 1 INTRODUCTION utility function in equation 1. given the various assumptions. Finn. and Pagan (1991) contain interesting discussions of the estimation of RBC models. this is judged to be a positive sign for the model. In this sense the calibrated parameters are also estimated.2. Chow (1991) and Canova. in the spirit of the seminal article by Kydland and Prescott (1982). There is also a slightly earlier literature in which the parameters of a utility function like the one in equation 1. But the paths are being compared in a very limited way from the way that the Cowles Commission approach would compare them. The parameters of this function would then be either estimated or simply chosen (“calibrated”) to be in line with parameters estimated in the literature.. There will simply be economic theory applied to different problems. Hall (1978). by whatever means. If the parameters are chosen by calibration. Once the coefficients are chosen. In the example in Section 1.2. the overall model is solved.

3 THE REAL BUSINESS CYCLE APPROACH 13 each period and then calculate the RMSE from these prediction errors. but it will eventually reach a dead end unless it comes around to developing models that can compete with other models in explaining the economy observation by observation. The RBC literature has focused so much on solving one problem that it may have exacerbated the effects of a number of others. The literature may take a long time to play itself out. is the RBC literature estimating deep structural parameters if a representative agent utility function is postulated and used that is independent of demographic changes over time? (A way of examining the possible problem of coefficients in macroeconomic equations changing as the age distribution changes is discussed in Section 4. How would.7 and applied in Chapter 5. This RMSE might then be compared to the RMSE from another structural model or from an autoregressive or vector autoregressive model. The disturbing feature of the RBC literature is there seems to be no interest in computing RMSEs and the like. it has in my view dropped out of the race for the model that best approximates the economy. Even the best econometric model is only an approximation of how the economy works. for example. Having the computed path mimic the actual path for a few selected moments is a far cry from beating even a first order autoregressive equation (let alone a structural model) in terms of fitting the observations well according to the RMSE criterion. but they proceed anyway. In what sense. but the key question for empirical work is the quantitative importance of this critique. People generally seem to realize that the RBC models do not fit well in this sense. say. and these problems may swamp the problem of coefficients changing when policy rules change. the coefficients in aggregate equations are likely to change.) . Why waste one’s time in working with models whose coefficients change over time as policy rules and other things change? The logic of the Lucas critique is certainly correct. the currently best fitting RBC model compare to a simple first order autoregressive equation for real GDP in terms of the RMSE criterion? Probably very poorly. Put another way. and this is a source of error in the estimated equations. This problem may be quantitatively much more important than the problem raised by Lucas. If this literature proceeds anyway. One of the main reasons people proceed anyway is undoubtedly the Lucas critique and the general excitement about deep structural parameters. I have never seen this type of comparison done for a RBC model. As the age and income distributions of the population change. the representative agent model that is used so much in macroeconomics has serious problems of its own.1. Another potential source of coefficient change is the use of aggregate data.

Given my views of how the economy works. 1. If the policy variable has not changed or changed very little. but something seems missing. and so methods like those mentioned in Section 1. Assume also that the policy variable changes frequently. whose study uses quite good data. Assume that for an equation or set of equations the parameters change considerably when a given policy variable changes. In a completely different literature—the estimation of production smoothing equations—Krane and Braun (1989). I do not think this is my main source of uneasiness. but the misspecification will not have been given a chance to be picked up in the data. many of the results of the new Keynesian literature seem reasonable. The main problem is that this literature is not really empirical in the Cowles Commission sense. many of the conclusions do not seem robust to small changes in the models. It’s like coming out of a play that many of your friends liked and feeling that you did not really like it. and it’s hard to remember each one. report that their attempts to estimate first order conditions were unsuccessful. however. It may simply not be sensible to use aggregate data to estimate utility function parameters and the like. one encouraging feature regarding the Lucas critique is that it can be tested. the articles in the two volumes of New Keynesian Economics. One problem is that it is hard to get a big picture. Finally. But otherwise. In this case the model is obviously misspecified. and Summers (1985) for the utility parameters are not supportive of the approach. of the 34 papers in these two volumes. Consider. for example. By my count. The results in Mankiw. but not knowing quite why. This literature has moved macroeconomics away from its econometric base. In addition. Rotemberg. the results have not always been very good even when judged by themselves. edited by Mankiw and Romer (1991). only eight have anything . models that suffer in an important way from the Lucas critique ought to be weeded out by various tests.2 should be able to pick up this misspecification if the policy variable has changed frequently. There are many small stories.4 The New Keynesian Economics I come away from reading new Keynesian articles feeling uneasy.14 1 INTRODUCTION When deep structural parameters have been estimated from the first order conditions. then the model will be misspecified. Upon further reflection.

5 LOOKING AHEAD 15 to do with data. I should hasten to add that I do not mean by the above criticisms that there is no interesting empirical work going on in macroeconomics. This is probably an excessively generous interpretation. it is not likely to have much long run impact. one does not see. it is surely not the essence of new Keynesian economics. 1. “Efficiency Wages and the Interindustry Wage Structure”) is more labor than macro. presents and briefly discusses data in one figure. As is also true in the RBC literature of RBC models. Nevertheless. The RBC literature is only interested in testmight argue nine. the literature on production smoothing. but unless the literature does move in a more econometric and larger model direction.8 Of these eight. has produced some important results and insights.” which I did not count in the eight. Even if one wanted to be generous and put some of this empirical work in the new Keynesian literature. it seems clear that there is very little in the new Keynesian literature similar to the structural modeling outlined in Section 1. given the focus of this literature on small theoretical models. Okun’s article “Inflation: Its Mechanics and Welfare Costs. both having left the original path laid out by the Cowles Commission and its predecessors.2. But here one does not see it because no econometric models of real GDP are constructed! So this literature is in danger of dropping out of the race not because it is necessarily uninterested in serious tests but because it is uninterested in constructing econometric models. which is largely empirical. but they are not really empirical macroeconomics. For example. “The Rigidity of Prices”) is more industrial organization than macro and one (Krueger and Summers. 9 By Olivier Blanchard. 8 One . predictions of real GDP from some new Keynesian model compared to predictions of real GDP from an autoregressive equation using a criterion like the RMSE criterion. one (Carlton. It has been pointed out to me9 that the Mankiw and Romer volumes may be biased against empirical papers because of space constraints imposed by the publisher. These two studies provide some interesting insights that might be of help to macroeconomists. It is simply that literature of this type is not generally classified as new Keynesian. say. One might argue that new Keynesian economics is just getting started and that the big picture (model) will eventually emerge to rival existing models of the economy.5 Looking Ahead So I see the RBC and new Keynesian literatures passing each other like two runners in the night.1.

The RBC literature should entertain the possibility of testing models based on estimating deep structural parameters against models based on estimating approximations of decision equations. the tests should be more than just observing whether a computed path mimics the actual path in a few ways. what follows is an application of the Cowles Commission approach. and there are procedures for weeding out inferior models.16 1 INTRODUCTION ing in a very limited way. If one is worried about coefficients in structural equations changing. and analyzed. estimate. and the new Keynesian literature is not econometric enough to even talk about serious testing. Even the quantitative importance of the Lucas critique can be tested. A structural macroeconomic model is specified. At any rate. But I argue there is hope. estimated. Models can be tested. The new Keynesian literature should entertain the possibility of putting its various ideas together to specify. tested. Also. it seems unlikely that getting rid of the structural detail in large scale models is going to get one closer to deep structural parameters. . I have always thought it ironic that one of the consequences of the Lucas critique was to narrow the number of endogenous variables in a model from many (say a hundred or more) to generally no more than three or four. and test structural macroeconometric models. both literatures ought to consider bigger models. Finally.

By requiring that the decisions of agents be “choice theoretic” (i.. and Clower (1965). based on the maximizing postulates of microeconomics) and by relaxing the assumption that markets are always in equilibrium. The problem with these early disequilibrium studies is that they did not provide an explanation of why prices and wages may not always clear markets.e. and the existence of excess supply in the goods market is a justification for the existence of unemployment. The treatment of prices and wages as exogenous or in an ad hoc manner is particularly restrictive in a disequilibrium context because disequilibrium questions are inherently 17 . 2) allowing for the possibility of disequilibrium in some markets.1 One Country Background The theory that has guided the specification of the US model was first presented in Fair (1974) and then in Chapter 3 in Fair (1984). This was followed by the work of Barro and Grossman (1971). and 3) accounting for all balance-sheet and flow of funds constraints. Prices and wages were either taken to be exogenous or determined in an ad hoc manner. This was also true of the related literature on fixed price equilibria (see Grandmont (1977) for a survey of this literature). Chapter 13. The implications of the first two ideas were first worked on by Patinkin (1956). this work provided a more solid theoretical basis for the existence of the Keynesian consumption function and for the existence of unemployment. The existence of excess supply in the labor market is a justification for including income as an explanatory variable in the consumption function.1 2. This work stresses three ideas: 1) basing macroeconomics on solid microeconomic foundations.1.2 Theory 2.

employment. 335). The maximization problems require that firms form expectations of various variables before the problems are solved. this does not result in an immediate gain of everyone else’s customers. In these studies prices and wages are part of the decision variables of firms. which in turn were influenced by Stigler’s classic article (1961) on imperfect information and search. (1970) by adding investment. hence the relevance of Stigler’s article. Christ (1968) was one of the first to emphasize the government budget constraint. the model accounts for all balance-sheet and flow of funds constraints among the sectors. provides an explanation of disequilibrium. and output. . prices and wages are decision variables of firms. and these errors can persist for more than one period. (1970). this does not result in an immediate loss of all its customers. A similar statement holds for wages. along with investment.18 2 THEORY concerned with whether prices always get set in such a way as to clear markets. This explanation draws heavily on the studies in Phelps et al. The model is also more general in its treatment of household behavior and the behavior of financial markets. Also. Chapter 3. a tendency for high price firms to lose customers over time and for low price firms to gain customers. disequilibrium can occur because of expectation errors. Models of this type are in Phelps and Winter (1970) for prices and in Phelps (1970) and Mortensen (1970) for wages. In other words. In the Phelps and Winter model. as noted above. Errors can occur if firms do not know the exact processes that generate the variables for which they must form expectations. The theoretical work in Fair (1974) and Fair (1984). disequilibrium occurs if the average price set by firms differs from the expected average price (1970. If a firm’s market share is a function of its price relative to the prices of other firms. and expectation errors may lead to the setting of prices and wages that do not clear markets. and if a firm lowers its price below prices of other firms. This means that the government budget constraint is automatically accounted for. and it expands on the studies in Phelps et al. Firms choose these variables in a multiperiod profit maximization context. This model thus expands on the fixed price disequilibrium studies by adding prices and wages as decision variables. employment. This feature is likely to be true if customers and workers have imperfect information about prices and wages. If a firm raises its price above prices of other firms. then a firm’s optimal price strategy is a function of this relationship. Disequilibrium can occur in models of this type. and output as decision variables. for example. however. There is. Chapter 3. In the model in Fair (1974) and Fair (1984). p.

Chapter 3. then an increase in the after tax wage rate relative to the price level leads a household to work and consume more. which is that it may not be able to work as many hours as it would like (i. which it does in the simulation runs. There is also a possible labor constraint on a household. This is in contrast to the stochastic simulation of econometric models. Simulating a theoretical model is simply a way of learning about its properties for the particular set of parameters used. and 5) the level of transfer payments. If the labor constraint is binding after the household has solved its maximization problem ignoring the constraint. 1 It . It works as much as the labor constraint allows. The initial and terminal values of wealth also affect these decisions. There are thus three things a household can do with its time: work.2 Household Behavior The household maximization problem is similar to the example presented in Section 1.1 ONE COUNTRY 19 The properties of the theoretical model were analyzed in Fair (1974) and Fair (1984). A household’s “unconstrained” was assumed for the simulation exercises that households do not expect the labor constraint to be binding in future periods. 2) the price level. An increase in transfer payments or in the initial value of wealth leads it to work less and consume more. can be used in the testing of models.. other things being equal. 2. the household reoptimizes subject to the constraint. The treatment of money holdings provides a choice theoretic explanation of the interest sensitivity of the demand for money. using simulation techniques.1. The following is a brief outline of the model and the simulation results. as much as the solution of its maximization problem implies).e. The variables that influence how much a household consumes and works include current and expected future values of 1) the wage rate. and a relationship has been added between the level of money holdings and time spent taking care of these holdings. 3) the interest rate. which. 4) the tax rate. has a positive effect on the household’s saving rate at the beginning of the horizon and a negative effect near the end.2. Regarding the use of simulation techniques in this context.1 This leads the household to consume less than it otherwise would. An increase in the interest rate. Taxes and transfers have been added to the problem. take care of money holdings. note that the simulation of theoretical models is not a test of the models in any way. The possible labor constraint can also affect household behavior.2. and engage in leisure. as discussed in Chapter 7. If the substitution effect dominates the income effect.

The main expected costs from doing this.3 Firm Behavior As noted above. 2. then the unconstrained and constrained decisions are the same. Because of these adjustment costs. Regarding money demand.” A firm expects that it will gain customers by lowering its price relative to the expected prices of other firms. The demand for money is also a function of the level of transactions. Similarly. some of the time the number of machine hours available for use may be greater than the number of machine hours actually used. A household solves its multiperiod optimization problem based on expectations of future prices and wages (as well as of nominal interest rates). time spent taking care of money holdings responds negatively to the wage rate and positively to the interest rate. a household spends more time keeping money balances low when the wage rate is low or the interest rate is high. and a household’s “constrained” decision is defined to be the decision it actually makes taking into account the possible labor constraint. investment. The worker-machine ratio is assumed to be fixed for each type of machine. wages. Note that real interest rate effects on household behavior are accounted for in the model. firms solve profit maximization problems in which prices. and so future inflation effects are captured. . In other words. The technology of a firm is assumed to be of a “putty-clay” type. There are assumed to be costs involved in changing the size of the work force and the size of the capital stock. it may be optimal for a firm to operate some of the time below capacity and “off” its production function. where at any one time there are a number of different types of machines that can be purchased. in the number of workers that must be used per machine per unit of time. The difference between hours paid for by a firm and hours worked is “excess labor. employment. and output are decision variables. and in the amount of output that can be produced per machine per unit of time. are the adjustment costs.20 2 THEORY decision is defined to be the decision it would make if the labor constraint were not binding.” and the difference between the number of machines on hand and the number of machines required to produce the output is “excess capital. This means that some of the time the number of worker hours paid for may be greater than the number of hours that the workers are effectively working. If the labor constraint is not binding. aside from the lower price it is charging per good.1. The machines differ in price.

and so both labor and loan constraints were considered.2. In other words. A firm expects that it will gain (lose) workers if it raises (lowers) its wage rate relative to the expected wage rates of other firms. On the plus side. The firm also expects that other firms will follow it if it raises (lowers) its wage rate. the possible existence of loan constraints was dropped. a firm. a firm expects that it will lose customers by raising its price relative to the expected prices of other firms. . may choose in the current period not to lower its employment and capital stock to the minimum levels required.1 ONE COUNTRY 21 The firm also expects that other firms over time will follow it if it lowers its price. and so it does not expect to lose an ever increasing share of the market with its higher price. This considerably simplifies the model. I did not find in the empirical work much evidence of the effects of loan constraints on the economy. aside from the lost customers. The household and firm maximization problems are easier to specify. it may be optimal for the firm to hold either excess labor or excess capital or both during certain periods. and excess capital on hand has a negative effect on current investment decisions. Similar to household behavior. real interest rate effects on firm behavior are accounted for in the model through the multiperiod optimization problem of a firm. and it responds to an increase in demand by raising its price and expanding. are again the adjustment costs. Some of the main properties of the model of firm behavior that result from the simulation runs are the following. The main costs from doing this. 4) A firm responds to a decrease in demand by lowering its price and contracting.1. Similar reasoning holds for wages. Because of the adjustment costs. and so in presenting the model in Fair (1984). 1) A change in the expected prices (wages) of other firms leads the given firm to changes its own price (wage) in the same direction. 3) An increase (decrease) in the interest rate leads to a substitution away from (toward) less labor intensive machines and a decrease (increase) in investment expenditures. Conversely.4 Bank and Government Behavior The model in Fair (1974) allowed for the possible existence of credit rationing. and so it does not expect to be able to capture an ever increasing share of the market with its lower price. the firm expects that other firms over time will follow it if it raises its price. if it chooses to lower output. and so it does not expect to capture (lose) an ever increasing share of the market with its wage rate increase (decrease). 2. 2) Excess labor on hand has a negative effect on current employment decisions. Chapter 3.

1. The percent of bank reserves borrowed from the monetary authority is a function of the spread between the bank loan rate and the discount rate. and so the government budget constraint is met. This simpler model of bank behavior is outlined here. The addition of an interest rate reaction function to the theoretical model grew out of empirical work I had done in estimating a reaction function of the Federal Reserve [Fair (1978)]. It sets the reserve requirement ratio and the discount rate. . The monetary authority earns interest on its loans to the banks. They must hold a certain portion of these deposits in the form of bank reserves. and the government interact. Some of the main properties of the complete model as gleaned from the simulation results are the following. as a policy variable of the monetary authority. It chooses its spending for goods and labor. on the other hand. banks. an estimated reaction function is part of the US model. an interest rate reaction function was postulated for the monetary authority. i. and it pays interest on its debt. In this version the amount of government securities outstanding is endogenous—the amount is whatever is needed to have the interest rate value from the reaction function be met. Banks loan money to households. As will be seen in Chapter 4. where the interest rate implied by the reaction function was attained through open market operations. 2 The phrase “government securities” is used for convenience here and in what follows even though there is no distinction in the model between government securities and other types of securities. It can also engage in open market operations by buying and selling government securities. firms. and the government. In Fair (1984). All the flows of funds among the sectors are accounted for. firms. 2.e. Banks receive money from households and firms in the form of demand deposits. The constraint states that any nonzero level of saving of the government must result in the change in non borrowed reserves or the amount of government securities outstanding.. firms.5 The Complete Model The complete model consists of the maximization problems of households and firms and the specification of how households. The fiscal authority in the government sets the tax rates and collects taxes from households.2 In Fair (1974) the amount of government securities outstanding was taken as exogenous. and they are assumed never to hold excess reserves.22 2 THEORY and it no longer necessary to specify a maximization problem for banks. and banks. Chapter 3.

Firms are not assumed to have this much knowledge. and so they can make expectation errors. which leads them to consume even less. In order for a firm to form correct expectations. increasing the income tax rate lowers labor supply (assuming the substitution effect dominates).1 ONE COUNTRY 23 1. A multiplier reaction can thus get started from an initial labor constraint on households. they consume less than they otherwise would. Given the many fac- . less labor intensive machines. 4. which increases investment expenditures. which in turn is a function of such variables as the after tax real wage and the level of transfer payments. For example. and firms respond in the next period by lowering output and their demand for labor. A fall in the interest rate results in a capital gain on stocks. When households reoptimize subject to the labor constraint. The unemployment rate is a positive function of the supply of labor. This lowers firms’ sales. 2. This further constrains the households. and so on. the labor constraint is binding on households. The lower interest rate may also lead firms to switch to more expensive.2. the interest rate is lowered by the reaction function of the monetary authority. Note that this explanation of disequilibrium does not rely on price and wage rigidities. whereas decreasing the level of transfer payments raises it. If the quantity of labor demanded from firms and the government (banks do not demand labor) is less than the quantity of labor that households want to supply from their unconstrained maximization problems. As unemployment increases. Once the economy begins to contract. it would have to know the maximization problems of all the other firms and of the households. this is another reason for the existence of disequilibrium. Both the lower interest rate and the higher wealth have a positive effect on consumption. It would also have to know the exact way that transactions take place once the decisions have been solved for. although if there are such rigidities. which comes about from expectation errors. 3. the interest rate is one of the key variables that prevents it from contracting indefinitely. Unemployment in the model can be thought of as the difference between the unconstrained and constrained supply of labor from the households. Disequilibrium of the kind just described can occur because firms do not necessarily set the correct prices and wages. The effects of a policy change on the unemployment rate thus depend in part on the labor supply response to the policy change.

and exchange rate links among the countries. In other words. Branson (1974).24 2 THEORY tors that affect labor supply. for example. Dornbusch (1976). Frenkel and Rodriguez (1975). There is. A straightforward way of doing this is to link one single country model to another. Kouri (1976). and they are also accounted for when two single country . This completes the outline of the single country theoretical model. price. on the other hand. p. The two country model that is developed is sufficient to capture the main links among the countries that exist in the MC model. It accounts for the trade. [Other studies in the 1970s in which the stock-flow distinction was important included Allen (1973).] The monetary approach to the balance of payments stressed the stock market determination of the exchange rate. [See. This model was in part a response to the considerable discussion in the literature that had taken place in the 1970s as to whether the exchange rate is determined in a stock market or in a flow market. and in no rigorous sense can it be said to be the variable that clears a particular market. 686)]. there is no natural distinction between stock market and flow market determination of the exchange rate. and Girton and Henderson (1976).] The reason there is no stock-flow distinction in the model is the accounting for all flow of funds and balance-sheet constraints. there is no need for a stock-flow distinction in the model.2. in other words. the model needs to be opened to the outside world. These constraints are accounted for in the single country model. and this will now be done. 2. and the survey by Myhrman (1976). The exchange rate is merely one endogenous variable out of many. stock and flow effects are completely integrated. interest rate. Simulation results for the model can be found in Fair (1984). In the model in Fair (1979). no stable Okun’s law and no stable Phillips curve in the model.1 Two Countries Background The theoretical two country model that has guided the specification of the MC model was first presented in Fair (1979). Frenkel and Johnson (1976).2 2. Black (1973). which was contrasted with “the popular notion that the exchange rate is determined in the flow market so as to assure a balanced balance of payments” [Frenkel and Rodriguez (1975. To complete the theory. Chapter 3. there is no stable relationship in the model between the unemployment rate and real output and between the unemployment rate and the rate of inflation.

financial (b). and increase in e is a depreciation of country 1’s currency. however. x). The private non financial sector includes both households and firms. lower case letters denote variables for country 2.2. The equations for country 1 are as follows. (The derivative indicates the expected effect of the particular variable on the left hand side variable. and collects taxes (T . y). and an asterisk (*) on a variable denotes the other country’s holdings or purchase of the variable. b). Only the private sector of the given country holds the money of the country. This should help in understanding the properties of the theoretical model before it is used to guide the specification of the MC model. then the value of B or b for that sector is negative. Section 3. The price of X is P and of x is p. e is the price of country 1’s currency in terms of country 2’s currency. It will be called the “private sector. 2. for example. Again.2.3 Equations There are 17 equations per country and one redundant equation.e. The interest rate on B is R and on b is r. 2. so that. Each country has its own money (M. The bonds are one period securities. If a sector is a debtor with respect to a bond (i.2 Notation In what follows capital letters denote variables for country 1. and government (g).2.” Members of the financial sector will be called “banks. In this section a version of the two country model is presented that will be analyzed by simulation techniques. the two country theoretical model was not analyzed by simulation techniques in Fair (1984).” Each country specializes in the production of one good (X. the simulation of the theoretical model is not meant to be a test of the model in any sense. The government of each country holds a positive amount of the international reserve (Q.. m) and its own bond (B. a supplier of the bond). The government of a country does not hold the bond of the other country and does not buy the good of the other country. t) as a proportion of income (Y. which is denominated in the units of country 1’s currency.) The demands for the .2. Contrary to the case for the single country theoretical model.2 TWO COUNTRIES 25 models are put together to form a two country model. The main features of the model in Fair (1979) that are relevant for the construction of the MC model were discussed in Fair (1984). q). There are three sectors per country: private non financial (h). fij is the derivative of fi with respect to argument j .

e · p: P = f3 (Y.26 two goods by the private sector of country 1 are Xh = f1 (P . e · p). e · p. Mb = Mh (2. so that BR = RR · Mb (2. Y − T ). f72 < 0 (2. and xh is the purchase of country 2’s good by the private sector of country 1. f14 > 0 (2.7) Since the private sector is assumed to be the only sector holding money.8) where Mb is the money held in banks. where EP+1 is the expected value of P for the next period based on current period information. f13 < 0. f32 > 0 (2.6) Borrowing by the banks from the monetary authority (BO) is assumed to be a function of R and of the discount rate RD: BO = f7 (R.2) R is the real interest rate. Taxes paid to the government are T = TX ·Y (2. f31 > 0. f22 < 0.9) . f62 > 0 (2. ∗ xh = f2 (P . The demand for real balances is assumed to be a function of the interest rate and income: Mh = f6 (R. Y − T ). f23 < 0. Equation 2. e · p. The domestic price level is assumed to be a function of demand pressure as measured by Y and of the level of import prices. f71 > 0.1) f21 > 0. f12 > 0. P f61 < 0. 2 THEORY f11 < 0. RD). Y ). The equations state that the demands are a function of the two prices.3) There is assumed to be no inventory investment. Xh is the purchase of country 1’s good by the private ∗ sector of country 1. and after tax income.5) where T X is the tax rate. R − (EP+1 − P ). so that production is equal to sales: ∗ (2. the real interest rate.4) Y = Xh + Xg + Xh ∗ where Xg is the purchase of country 1’s good by its government and Xh is the purchase of country 1’s good by country 2. R .8 simply says that all money is held in banks. R . f24 > 0 (2. Banks are assumed to hold no excess reserves.

. Let Ee+1 be the expected exchange rate for the next period based on information available in the current period.e. where r is the interest rate on the bond of country 2. the expected (one period) return on the bond of country 2. Er). on the other hand.2 TWO COUNTRIES 27 where BR is the level of bank reserves and RR is the reserve requirement rate. This will be seen in Chapter 6 in the estimation of the forward rate equations. f10. denoted Er. There is an exact relationship between the expected future exchange rate. minus import costs. plus interest received (or minus interest paid) on the holdings of country 1’s bond. 154–155. The demand for country 2’s bond is assumed to be a function of R and Er : ∗ bh = f10 (R. and plus interest received on the holdings of country 2’s 3 It was incorrectly stated in Fair (1984). If there were uncovered interest parity between. The next three equations determine the financial saving of each sector: ∗ ∗ ∗ S h = P · Xg + P · X h − e · p · x h − T + R · B h + e · r · b h (2.S.1 < 0. the bonds of countries A and B. plus export revenue. If capital mobility is such as to lead to uncovered interest parity almost holding (i. is Ee+1 e (1 + r) − 1. Covered interest parity.3 then equation 2. It would not make sense in this case to estimate three equations. the MC model consists of estimated interest rate and exchange rate equations (reaction functions) for a number of countries (all exchange rates are relative to the U. which is not assumed here.2. pp.2 > 0 (2. The correct assumption is that uncovered interest parity does not hold.11 states that the saving of the private sector is equal to revenue from the sale of goods to the government. it would not be possible to estimate interest rate equations for countries A and B plus an exchange rate equation. the two interest rates. Then from country 1’s perspective. and there is effectively only one interest rate in the model.. Equation 2. f10. minus taxes paid.10 and the equivalent equation for country 2 drop out.10) ∗ bh is the amount of country 2’s bond held by country 1. dollar).12) (2. and so given a value of the expected future exchange rate. that the version of the model that is used to guide the specification of the MC model is based on the assumption of perfect substitution of the two bonds. say. As will be seen in Chapter 6.11) (2. only two of the other three values are left to be determined. does roughly hold in the data used here. If uncovered interest parity holds exactly.13) Sb = R · Bb − RD · BO Sg = T − P · Xg + R · Bg + RD · BO Equation 2. and the spot exchange rate if uncovered interest parity holds. R ∗ almost equal to Er).10 and the equivalent equation for country 2 are important in the model. then large changes in bh will result from small changes in the difference between R and Er.

28 2 THEORY bond.16 states that any nonzero value of saving of the government must result in the change in bond holdings.13 determines the government’s surplus or deficit.12 states that the saving of banks is equal to interest revenue on bond holdings (assuming Bb is positive) minus interest payments on borrowings from the monetary authority. There are thus 34 independent equations in the model. which says that someone’s asset is someone else’s liability with respect to the bond of country 1: ∗ 0 = B h + Bb + Bg + Bh (2. Equation 2. The next three equations are the budget constraints facing each sector: 0 = Sh − 0 = Sb − 0 = Sg − Mh − Bh − e · ∗ bh (2. Equation 2. Q is replaced by q/e. nonborrowed reserves (which are a liability to the government). It states that the saving of the government is equal to tax revenue. This equation is implied by equations 2. and so it is redundant.17) These same 17 equations are assumed to hold for country 2.16) Bb + Bg + Mb − (BR − BO) Q (BR − BO) − Equation 2. There is also a constraint across all sectors.11–2.11. minus expenditures on goods. or nonborrowed reserves.) The last equation of the model is 0= Q+ q which says that the change in reserves across countries is zero. . and plus interest received on loans to banks.14 states that any nonzero value of saving of the private sector must result in the change in its money or bond holdings. or international reserve holdings.14) (2. with lower case and upper case letters reversed except for Q and with 1/e replacing e. Equation 2.15) (2. minus interest costs (assuming Bg is negative). money deposits (which are a liability to banks). Remember that the private sector (h) is a combination of households and firms. and so transactions between households and firms net out of equation 2.17 and the equivalent equations for country 2. Equation 2. (Remember that Q and q are in the units of country 1’s currency. If the private sector is a net debtor with respect to the bond of country 1.15 states that any nonzero value of saving of the financial sector must result in the change in bond holdings. then Bh is negative and R · Bh measures interest payments.

Xh . Xh .17 yields the first derived equation: 0=S+ ∗ Bh − e · ∗ bh − Q (i) This equation simply says that any nonzero value of saving of country 1 must result in the change in at least one of the following three: country 2’s holdings of country 1’s bond. government purchases of goods. The model can then be closed by taking Bg . these same 18 variables for country 2. minus interest paid to country 2. Ee+1 . Y . e. minus import costs. Summing equations 2. Sh . In order to close the model one needs to make an assumption about how the three expectations are determined and to take three other variables as exogenous. R .16 and using 2. Bg . (Remember there are 34 independent equations in the model. and plus interest received from country 2.17: ∗ ∗ ∗ ∗ S = P · Xh − e · p · xh − R · Bh + e · r · bh (ii) This equation says that the saving of country 1 is equal to export revenue. ∗ Sg .2.13 and using 2.11–2. . RR. there ∗ are 40 variables in the model: Bb . The second derived equation is obtained by summing equations 2. These are the three main tools of the monetary authorities. First.) Assume also that the two price expectations are static. BR. RD. Bh . Not counting these variables. P . the reserve requirement rate. Mb . T . and Ep+1 . EP+1 . Taking these three tools of the monetary authorities as exogenous thus closes the model. let S denote the financial saving of country 1. the tax rate. Remember that R does not necessarily equal r since uncovered interest parity is not necessarily assumed to hold. Q. which is the sum of the saving of the three sectors: S = Sh + Sb + Sg S is the balance of payments on current account of country 1. T X. and country 1’s holding of the international reserve.Sb .) Assume for now that exchange rate expectations are static in the sense that Ee+1 = e always.4 Closing the Model The exogenous government policy variables are: Xg . country 1’s holding of country 2’s bond. and the same variables for country 2. (This implies that Er = r and ER = R. EP+1 = P and Ep+1 = p. Bh .2 TWO COUNTRIES 29 It will be useful in what follows to consider two equations that can be derived from the others. the discount rate. and Q as exogenous. BO.14–2. 2.2. Mh . bg .

) The interest rate and exchange rate equations are interpreted as reaction functions. The exchange rate e is just one of the many endogenous variables. r. they are likely to take market forces into account in setting their target values of R. . along with the other endogenous variables.2. where the explanatory variables in the equations are assumed to be variables that affect the monetary authorities’ decisions. and these will be discussed next in the context of the overall properties of the model. In other words. r. (Thus. Country 1 buys country 2’s good ∗ ∗ (xh ). bg . r. The interest rate and exchange rate links are less straightforward. and e in the theoretical model in order to guide the choice of explanatory variables in the estimated reaction functions in the MC model.30 2 THEORY Instead of taking the three tools to be exogenous. and e.3 and the equivalent equation for country 2. and it is determined. and Q must be taken to be endogenous. Country 2’s price affects country 1’s price. by the overall solution of the model. and vice versa. If reaction functions for these three variables are used (or the three variables are taken to be exogenous). r. bg . and e much different from what the market would otherwise achieve. the exchange rate and interest rate equations in the MC model are based on the assumption that the monetary authorities manipulate R. The solution values of Bg . and e met. The price links come through equation 2. Bg . Note that in closing the model no mention was made of stock versus flow effects. r. and country 2 buys country 1’s good (Xh ). and e. one needs to know the market forces that affect R. then Bg . one can assume that the monetary authorities use the tools to manipulate R. they are likely to be sensitive to and influenced by market forces.2. 2. it will take large changes in the three tools to achieve values of R. and e.5 Links in the Model The trade links in the model are standard. and Q are endogenous in the MC model. r. Therefore.6 Properties of the Model As will be discussed in the next section. however. 2. Since the monetary authorities are likely to want to avoid large changes in the tools. from above. and Q are whatever is needed to have the target values of R. bg . The key question in this work is what variables affect the monetary authorities’ decisions. If capital mobility is high in the sense that uncovered interest parity almost holds.

and e in the theoretical model. the experiments are based on the assumption that Ee+1 = e.4 The model was solved using the Gauss-Seidel technique. Taking Q to be exogenous means that the monetary authorities are not manipulating e.1 · log e · p + . Xg = xg = 20. r.75 · log(Y − T ) (2.0 · R + .07. all prices = 1. for a discussion of the Gauss-Seidel technique.1 · log Y log Mh = a6 − 1.10 and the equivalent equation for country 2 that ∗ ∗ bh and Bh are simply a function of R and r.2.5 · log Y P BO = a7 + 50 · R − 50 · RD (2.2. In all but the last experiment. a simulation version has been analyzed. e = 1. 2. This means from equation 2. 2. and all other variables.6) (2. Y = y = 100. Xh = xh = 20.3) (2. This is a way of examining the market forces on e without intervention. equal to zero. The ai coefficients were chosen so that when the model was solved using the base values of all the variables. including lagged values when appropriate. 5 See Fair (1984).0 · log P − 1.7) (2. RR = rr = .0 · log e · p − 1. and e in the model.3. all interest rates = . 2.10 and the equivalent equations for country 2. solving the model. Bg and bg are always endogenous for the experiments because all the experiments ∗ ∗ base values were Xh = xh = 60. The following experiments were chosen with the aim of learning about the market forces affecting R.0 · R + . and 2.75 · log(Y − T ) (2.5 The properties of the model can be examined by changing one or more exogenous variables.25 · log P + . e is endogenous and Q is exogenous.1) ∗ log xh = a2 + 1. Section 7.0 · R + .10) ∗ bh = a10 − 100 · R + 100 · Er The same functional forms and coefficients were used for country 2. 4 The .2. Particular functional forms and coefficients have been chosen for equations 2.6. Unless otherwise noted.2. the solution values were the base values.2) log P = a3 + . T X = tx = . T = t = 20. Mh = Mb = mh = mb = 100.25 · log e · p − 1. and comparing the solution values to the base values.7. The experiments are also based on the assumptions that EP+1 = P and Ep+1 = p. The solution value of e for each experiment is the value that would pertain if the monetary authorities did not intervene at all in the foreign exchange market in response to whatever change was made for the experiment. r. BR = br = 20. The five equations for country 1 are: log Xh = a1 − .1.2 TWO COUNTRIES 31 In order to examine the market forces on R.2. 2.

1 may be reversed with different coefficients. Only signs are presented in the table because the magnitudes mean very little given that the coefficients and base values are not empirically based.1 and discussed below. S is increased by increasing country 1’s exports and decreasing its imports—equation ii—which is accomplished by a depreciation.1 for e for experiment 1 thus means that country 1’s currency depreciated. but the results from this work should not be taken as evidence that all the signs in the table hold for all possible coefficient values.32 2 THEORY either have R and r exogenous or Mb and mb exogenous. and vice versa. When R and r are exogenous.3 it can be seen that this must result in an increase in S. are not necessarily robust to alternative choices of the coefficients. The + in Table 2. In two cases it is necessary to know which interest rate (R or r) changed the most. (Both interest rates are exogenous in this experiment. however. and these cases are noted in Table 2. 6 Remember . it is always assumed that the monetary authorities either keep interest rates or money supplies unchanged in response to whatever change was made for the experiment. which resulted in a depreciation of country 1’s currency.6 The fall in R relative to r led to an ∗ increase in the demand for the bond of country 2 by country 1 (bh increased) ∗ and a decrease in the demand for the bond of country 1 by country 2 (Bh decreased). The simulation work is meant to help in understanding the theoretical model. The results of all the experiments are reported in Table 2.2.1. since Q is exogenous and unchanged. At least some of the signs in Table 2.) This change resulted in a depreciation of country 1’s currency. Mb and mb are endogenous. and the demand for money increased because of the lower interest rate and the higher level of inthat a rise in e is a depreciation of country 1’s currency. Another way of looking at this is that the fall in R relative to r led to a decreased demand for country 1’s currency because of the capital outflow. All shocks in the experiments are for country 1. Experiment 1: R decreased. r unchanged For this experiment the interest rate for country 1 was lowered (from its base value) and the interest rate for country 2 was assumed to remain unchanged. country 1’s balance of payments. From equation i in Section 2. In other words. The simulation experiments are simply meant to be used to help in understanding the qualitative effects on various variables. and the following discussion of the experiments relies on this table. Output for country 1 (Y ) increased because of the lower interest rate and the depreciation. Even the qualitative results.

1.2 shocked by . the higher is the degree of capital mobility.10. Although not shown in Table 2. the larger is the depreciation of the exchange rate and the larger is the increase in Bg for the same drop in R. experiments with alternative coefficients ∗ ∗ in the equations explaining bh and Bh —equation 2.3(+) + 0 0 0 0 0 0 0 0 0 0 + 0 + 0 0 0 0 0 + +b + − + − + + − − + + + 0 0 0 0 − + 5 Eq2.10.2 TWO COUNTRIES Table 2.001. r and e exogenous a R decreased more than did r. Mb raised by 1.2. Mb and mb exogenous 6.10 and the equivalent equation for country 2—showed that the more sensitive are the demands for the foreign bonds to the interest rate differential.001.10.3(+) Eq2.3 shocked by .3 shocked by . Equation 2. R and r exogenous 4. come. Equation 2. r exogenous 2. b R increased more than did r. mb exogenous 3. Equation 2.0. Size of changes: 1. Mb and mb exogenous 5. The monetary authority of country 1 bought bonds to achieve the reduction in R (Bg increased).2(+) + + − − + − + − + + − + − 0 0 0 0 + − 6 R(−) 0 − 0 − + + − + + + + + + + + − + + − 33 Q is exogenous exept for experiment 6. R lowered by .1 Simulation Results for the Two Country Model 1 R(−) e R r S s ∗ bh ∗ Bh ∗ xh ∗ Xh Y y P p Mh mh Q q Bg bg + − 0 + − + − − + + − + − + − 0 0 + − 2 Mb (+) + −a − + − + − − + + − + − + 0 0 0 + − Experiment 3 4 Eq2. the larger is the . R lowered by . In other words.

which means that the exchange rate is expected to appreciate in the next period relative to the value in the current period (i. the expected return on country 2’s bond (Er) falls.. The increase in Mb led to a decrease in R. The results of this experiment are similar to those of experiment 1. which leads to a smaller increase in bh and a smaller decrease ∗ in Bh . There is thus less downward pressure on country 1’s currency and thus a smaller depreciation. which leads to a depreciation of country 1’s exchange rate. i. and the money supply of country 1 was increased. The expected return on country 2’s bond rises. R and r unchanged For this experiment the price equation for country 1 was shocked positively. which decreases the demand for country 1’s bond and increases the demand for country 2’s bond. which means that the exchange rate is expected to depreciate further in the next period. both absolutely and relative to r. where Ee+1 = e. which led to a depreciation of country 1’s currency. then the depreciation in the current period is less. Experiment 3: Positive price shock. The positive price shock resulted in a depreciation of country 1’s currency. Experiment 2: Mb increased. The interest rate of country 1 falls relative to that of country 2. there is more of a depreciation in the current period. This is because if Ee+1 is less than e. which leads to greater downward pressure on country 1’s exchange rate. mb unchanged For this experiment the monetary authorities are assumed to target the money supplies (Mb and mb are exogenous). The differential between R and Er thus falls less as a result of the ∗ decrease in R. The monetary authorities were assumed to respond to this by keeping interest rates unchanged. Given the coefficients and base values that are used for the . Country 1’s output increased as a result of the depreciation and the fall in R.e. The monetary authority of country 1 bought bonds to increase the money supply (Bg increased).34 2 THEORY size of open market operations that is needed to achieve a given target value of the interest rate. reverse at least some of the depreciation in the current period). Note that the effect of a change in the money supply on the exchange rate works through the change in relative interest rates. If expectations are formed in such a way that Ee+1 turns out to be greater than e.e. If instead expectations are formed in such a way that Ee+1 turns out to be less than e. Remember that the above experiment is for the case in which exchange rate expectations are static.

and there was no change in any real magnitudes. One of the key differences between the results for this experiment and the results for experiment 1 is that the . The reason for the exchange rate depreciation is the following. Mb and mb unchanged This experiment is the same as experiment 3 except that the money supplies rather than the interest rates are kept unchanged. The negative effects of the price shock offset the positive effects of the interest rate changes. led to an increase in R. r unchanged.2. a positive price shock leads to a decrease in the demand for exports and an increase in the demand for imports. which means from equation i that S cannot change. worsened. Experiment 6: R decreased. bg . which puts downward pressure on S. but merely an increase in demand for the imported good. The increased demand for imports led to a depreciation of country 1’s currency. Experiment 5: Positive import demand shock. Other things being equal. then bh and Bh do not change. It may at first glance seem odd that a positive import shock would lead to an increase in Y . since there was an increased demand for country 2’s currency. Experiment 4: Positive price shock. which ∗ ∗ increased Bh and decreased bh . the exchange rate depreciated by the same percent that P increased. Therefore. R also increased relative to r. and Q so as to lower R and keep r and e unchanged. In this case the monetary authorities choose Bg . country 1’s currency depreciated.2 TWO COUNTRIES 35 simulation model. The latter results in an increase in Y because of the stimulus from the depreciation. but remember that the shock does not correspond to any shock to the demand for the domestic good. The balance of payments. a depreciation must take place to decrease export demand and increase import demand enough to offset the effects of the price shock. Even though R increased relative to r. The experiment is not a substitution away from the domestic good to the imported good. S. e unchanged This experiment is the same as experiment 1 except that e rather than Q is exogenous. ∗ ∗ however. Mb and mb unchanged For this experiment the import demand equation of country 1 was shocked positively. interest rates are unchanged. If. which with an unchanged money supply. The depreciation led to an increase in Y and P . The positive price shock with the money supplies unchanged led to an increase in R.

The net effect is that S decreases (and thus Q decreases). money supply unchanged 5.2 presents a summary of these effects in the model (experiment 6 is not included in the table because both R and e are exogenous in it). money supply unchanged 2 THEORY Effect on: Domestic Interest Rate Exchange Rate — Lowered — Depreciation Depreciation Depreciation Raised Depreciation Raised Depreciation balance of payments. In the present experiment there is still an increase in the demand for country 2’s bond and a decrease in the demand for country 1’s bond—because R falls relative to r—but S does not necessarily have to increase because Q can change. Of main concern here are the effects of the various changes on the domestic interest rate and the exchange rate. This then worsens the balance of payments. The decrease in R is an expansionary action in country 1. Table 2. Positive price shock. Money supply raised 3. There is no offsetting effect from a depreciation of the currency to reverse this movement. Positive import shock.36 Table 2. This completes the discussion of the experiments. decreases rather than increases.2 Summary of the Experiments Experiment 1. . In experiment 1 S had to increase because of the increase in the demand for country 2’s bond by country 1 and the decrease in the demand for country 1’s bond by country 2. They should give one a fairly good idea of the properties of the model. and among other things it increases the country’s demand for imports. S. The reason for the decrease in S is fairly simple. Interest rate lowered 2. Positive price shock. The increase in S is accomplished by a depreciation. In experiment 1 S must increase because Q is exogenous—equation i. interest rates unchanged 4.

there are essentially two options open. bg . 2. reaction functions for interest rates and exchange rates have been estimated in the MC model. are always exogenous.2.10 in order to determine interest rates and exchange rates if one is willing to give up determining Bg . and Q are taken to be exogenous or equations estimated for them. No attempt has been made in the construction of the MC model to try to estimate money supply reaction functions. If one believes that monetary authorities intervene at least somewhat. r. and Q as exogenous. The present work is based on the implicit assumption that interest rate reaction functions provide a better approximation of the way monetary authorities behave than do money supply reaction functions. There is also a practical reason for taking the present approach. bg . and e equations are added to the model and the model is solved for Bg . One of the main problems is that data on bilateral holdings of securities either do not exist or are not very good. and the equivalent equations for country 2 were estimated. and this is the assumption made for the MC model. R. bg . If Bg . it will help to consider an alternative approach that in principle could have been followed. the R.10.2. equations like 2. To put this approach in perspective. In doing this. For many applications one can get by without knowing the amounts 7 It is in the spirit of the present approach to estimate money supply reaction functions rather than interest rate reaction functions. Their values would be whatever is needed to clear the two bond markets and the market for foreign exchange. bg .2 TWO COUNTRIES 37 2. and Q on the left hand side. i. which determine the bilateral demands for securities. however. 2. and e (and the other endogenous variables) by taking Bg . If equations 2. one would be making the rather extreme assumption that the monetary authorities’ choices of Bg . and Q are never influenced by the state of the economy. If the first option is followed.7 The Use of Reaction Functions As noted in the previous section. One is to estimate equations with Bg . and Q equations are added to the model and the model is solved for R. one can avoid estimating equations like 2. then the Bg . and e would thus be determined without having to estimate any direct equations for them. If the second option is followed. and Q. It is more natural to think of them having target values of interest rates (or money supplies7 ) and exchange rates. In practice it is very difficult to estimate such equations.2.6. and the other is to estimate equations with R. . 2. 2. one could solve the model for R.1. If instead equations for interest rates and exchange rates are estimated. must be estimated. 2. bg . and Q. bg . r. and e on the left hand side.3. r. The first option is awkward because one does not typically think of the monetary authorities having target values of the instruments themselves. bg . In either case Bg is endogenous.10.e. r.7. and e. r.

where an initial value for A for each country was first chosen. One can simply keep in mind that the values of these variables are whatever is needed to have the interest rate and exchange rate values be met.38 2 THEORY of government bonds outstanding and government reserve holdings. A = − Bh + e · bh + Q.2. 2. . a net asset variable. In ∗ ∗ terms of the variables in the theoretical model. The cost of doing this is that capital gains and losses on bonds from exchange rate changes are not accounted for. and ∗ ∗ so data on variables like Bh and bh are not available. Given the current data. there is little that can be done about this limitation. denoted A in the MC model. Equation i thus becomes 0=S− A (i) Data on S are available for each country. This aggregation is very convenient because it allows A to be easily constructed.8 Further Aggregation Data on bilateral security holdings were not collected for the MC model. was constructed for each country. and A was constructed as A−1 + S. Instead.

considerable “theorizing” involved in this transition process.1 The first step in the transition.1 Transition from Theory to Empirical Specifications The transition from theory to empirical work in macroeconomics is not always straightforward. is to choose the data and variables. The countries are listed 1 This transition is discussed in detail in Fair (1984). 39 . There are 30 stochastic equations for the United States and up to 15 each for each of the other countries. so compromises have to be made in moving from theory to empirical specifications. The discussion in the present chapter relies heavily on the tables in Appendices A and B. and Equations 3. There usually is. There are 101 identities for the United States and up to 19 each for each of the others. is to choose the explanatory variables in the stochastic equations and the functional forms of the equations.3 The Data.2. The third step. The second step. which are to be determined by stochastic (estimated) equations. Variables. which is where most of the theory is used. All the equations in the two models are listed in this chapter. is to choose which variables are to be treated as exogenous. in particular about unobserved variables like expectations and about dynamics. The quality of the data are never as good as one might like. This is the task of Chapters 5 and 6. extra assumptions usually have to be made.1. Also. and which are to be determined by identities. the overall MC model consists of estimated structural equations for 33 countries. Section 2. also taken in this chapter. There are 44 countries in the trade share matrix plus an all other category called “all other” (AO). As noted in Section 1. All the data and variables in the US and ROW models are presented in this chapter. The trade share matrix is thus 45×45. in other words. which is taken in this chapter.

6 show how the variables were constructed from the raw data.1–A. and equations for the US model are discussed in this section.2 and A. and the data for the other countries begin in 1960:1. foreign (r). and Table A. but not the coefficient estimates. It will be useful to begin with Tables A.9 shows which variables appear in which equations.1 shows how the six sectors in the US model are related to the sectors in the FFA.2 The US Model The data.1 The Tables (Tables A. The functional forms of the stochastic equations are given.2 lists all the variables in the US model in alphabetical order. The notation on the right side of this table (H1.9) Table A. The relevant tables are Tables A. AND EQUATIONS in Table B.4–A. Table A.9 in Appendix A.4 in the description of the FFA data.40 3 THE DATA. Table A.4–A.3 lists all the stochastic equations and identities. As will be discussed.FA.2.S. 3. firm (f).2 and A. Table A.1. In order to account for the flow of funds among these sectors and for their balance-sheet constraints. 3.1 presents the six sectors in the US model: household (h). Table A.6 before turning to Tables A.S. Tables A. federal government (g).7 lists the first stage regressors that were used for the 2SLS and 3SLS estimates. etc. Flow of Funds Accounts (FFA) and the U. VARIABLES. The data for the United States begin in 1952:1.30 in Chapter 5. some of the country models are annual rather than quarterly. and Table A.1–A. 3. Tables A. Finally.1–5. The coefficient estimates are presented in Tables 5. and these will be briefly outlined first.8 shows how the model is solved under various assumptions about monetary policy.2 The Raw Data The NIPA Data Table A.3 are the main reference tables for the US model.) is used in Table A. The variables from the NIPA are presented first. the U. financial (b).3. National Income and Product Accounts (NIPA) must be linked.2.4 lists all the raw data variables. variables. Of the remaining tables. in the order in which they appear in the Survey of Current . and state and local government (s). Many of the 101 identities in the US model are concerned with this linkage.

The alternatives to the “fixed 1987 weights” are “chain type annual weights” and “benchmark year weights. a discrepancy variable.” There are a number of problems with using the fixed 1987 weights over a period as long as that used in this study (1952:1–1993:2). Second. R19–R22). (ST AT P is constructed using equation 83 in Table A. the regular data from 1988 on were used (based on fixed 1987 weights). slowly decreasing in absolute value during the period.4 shows how this was done. The Other Data The variables from the FFA are presented next in Table A. August 1993. 2 See . At the time of this writing the alternative weights are not available before 1959 and after 1987. In early 1992 the NIPA data were revised. The old data for this period that were in units of 1982 dollars were multiplied up to be in units of 1987 dollars. One of the alternative set of weights—the chain type annual weights—was thus used in the construction of the data for the model. of course. and the old price indices that were 100 in 1982 were multiplied up to be 100 in 1987. which is the difference between real GDP and the sum of its real components. with the benchmark year changed from 1982 to 1987. ST AT P is taken to be exogenous in the model.3. since they are closer to the period. In the absence of alternative weights for this period. zero before 1959 and after 1987. the 1982 weights seemed a better choice than the 1987 weights. and the nominal variables were deflated by these indices (see variables R11–R16.2 and so the alternative weights have considerable appeal for present purposes.) ST AT P is. was created.1 (variables R84–R93). Table A. Consequently. Third. The procedure that was followed to create the real variables from the NIPA data is a follows. At the same time the Bureau of Economic Analysis began publishing quantity and price indices based on other than fixed weights. ordered by their code numbers.4. the pre-revised data (based on 1982 weights) were used between 1952 and 1958. First.2 THE US MODEL 41 Business. Some of these variables are NIPA variables that are not pubYoung (1992) and Triplett (1992) for a good discussion of these problems and of the proposed alternative weights. The chain type price indices were taken from NIPA Table 7. denoted ST AT P . the chain type weights were used for the data between 1959 and 1987. The use of the chain type price indices in this way means that between 1959 and 1987 real GDP is not the sum of its real components. Between 1959 and 1987 it is a fairly smoothly trending variable.3.

The quarterly social insurance variables R195–R200 were constructed from the annual variables R78–R83 and the quarterly variables R38. denoted FRB in the table. The other interest variables in the model are net payments. Otherwise.” and “other employer. Finally. It is implicitly assumed in the construction of the quarterly data that the quarterly pattern of the level of interest receipts of the rest of the world is the same as the quarterly pattern of the level of net interest payments of the firm and federal government sectors. state and local governments do not contribute to the federal government and vice versa. Quarterly data on net interest receipts of the rest of the world do not exist.” “government employer. Only annual data are available on the breakdown of social insurance contributions between the federal and the state and local governments with respect to the categories “personal.4 in alphabetical order along with their numbers.4. and R71. not payments. denoted EE in the table.” It is thus necessary to construct the quarterly variables using the annual data. These are explained beginning in the next paragraph. This allows one to find a raw data variable quickly. is constructed from the annual variable R96 and the quarterly and annual data on the I N T F and I N T G variables. R45 and R65. It is implicitly assumed in this construction that as employers. R60. Note that I N T ROW is the level of net receipts. Interest rate variables are presented next in the table. and these are indicated as such in the table. all the raw data variables are presented at the end of Table A. which is the level of net interest receipts of the rest of the world. one has to search through the entire table looking for the particular variable. VARIABLES. The tax variables R57 and R62 were adjusted to account for the tax sur- . followed by employment and population variables. Some adjustments that were made to the raw data are presented next in Table A. AND EQUATIONS lished in the Survey but that are needed to link the two accounts. I N T ROW . The constructed tax variables R201 and R202 pertain to the breakdown of corporate profit taxes of the financial sector between federal and state and local. The quarterly variable R203. The source for the interest rate data is the Federal Reserve Bulletin. Data on this breakdown do not exist. The main source for the employment and population data is Employment and Earnings.42 3 THE DATA. Some of these data are unpublished data from the Bureau of Labor Statistics (BLS). It is implicitly assumed in this construction that the breakdown is the same as it is for the total corporate sector. The adjustments that were made to the raw data are as follows. All the raw data variables are numbered with an “R” in front of the number to distinguish them from the variables in the model.

5 may look easy. If the variable is constructed by an identity. For example. Although the checks in Table A. Table A. In the model equation 49 determines T F G. Table 1. The financial stock variables in the model that are constructed from flow . labor force. where the equation number is the identity in Table A. All the receipts from sector i to sector j must be determined for all i and j (i and j run from 1 through 6). pp.. the variables in Table A. 414–415. and these must be accounted for. The variables in this table are raw data variables. considerable work is involved in having them met..6 explains the construction of the variables in the model (i. T F G being constructed directly from raw data variables.” appears. Also. the notation “Def.2) from the raw data variables (i.2 THE US MODEL 43 charge of 1968:3–1970:3 and the tax rebate of 1975:2. If any of the checks are not met. The multiplication factors in Table A. the variables in Table A. The multiplication factors are designed to make the old data consistent with the new data. is both a variable in the model and a raw data variable.e. the exogenous variables D2G is constructed by equation 49. P OP 2. For example. p.4). equation 85 constructs LM.. The equations in the table provide the main checks on the collection of the data. P OP 1. one or more errors have been made in the collection process. CD.5 presents the balance-sheet constraints that the data satisfy. some of the identities construct exogenous variables. and CE were made a few times.4 or are constructed by identities. Eq. For example.3.e. If a variable in the model is the same as a raw data variable. Equation 85 instead determines E. and employment variables. The tax surcharge numbers were taken from Okun (1971). The tax rebate was 7. see Fair (1984). consumption expenditures on durable goods. 3. 171. With a few exceptions. Official adjustments to the data on P OP . the variables in the model are either constructed in terms of the raw data variables in Table A. For further discussion. the same notation is used for both.4 pertain to the population.2.3 Variable Construction Table A.2 that constructs the variable. The tax surcharge and the tax rebate were taken out of personal income taxes (T P G) and put into personal transfer payments (T RGH ).8 billion dollars at a quarterly rate. CL2. whereas stochastic equation 8 determines LM in the model. E being constructed directly from raw data variables. CL1. In a few cases the identity that constructs an endogenous variable is not the equation that determines it in the model.

The equation is also estimated under the assumption of a first order autoregressive error term. is used in equation 15. H F S: Peak to Peak Interpolation of H F H F S is a peak to peak interpolation of H F . VARIABLES. The sample period 1952:1–1993:2 was divided into three subperiods. AND EQUATIONS identities need a base quarter and a base quarter starting value. 1971:1–1971:4. H O: Overtime Hours Data are not available for H O for the first 16 quarters of the sample period (1952:1–1955:4). The deviation of H F from H F S. which explains overtime hours. equation 80. There are a few variables in Table A. the ratio of state and local personal income taxes to taxable income. and H F F lagged once on the right hand side. and 1972:1–1993:2. H F F . the saving variables in Table A. The values for 1952:1 and 1952:2 were taken to be the 1952:3 predicted value. MB. and P U S. P I EF . The peaks are listed in Table A. 1952:1– 1970:4.5 (SH . The variables for which there are both raw data and an identity available are GDP .6 (aside from the balance-sheet checks in Table A. hours per job. P U G. There are also a few internal checks on the data in Table A.6.5). T AU S: Progressivity Tax Parameter—s T AU S is the progressivity tax parameter in the personal income tax equation for state and local governments (equation 48). and so on) must match the saving variables of the same name in Table A. The base quarter values are indicated in Table A. The base quarter was taken to be 1971:4. which is variable H F F in the model. These were judged from a plot of T H S/Y T . which the variables must satisfy. Two assumptions were then . to be periods of no large tax law changes. In addition. SF .6 whose construction needs some explanation. There is also one redundant equation in the model. The equation that explains H O in the model has log H O on the left hand side and a constant. and the stock values for this quarter were taken from the FFA stock values.6.44 3 THE DATA. It was obtained as follows.6. The missing data for H O were constructed by estimating the log H O equation for the 1956:1–1993:2 period and using the predicted values from this regression for the (outside sample) 1952:3–1955:4 period as the actual data.

D1S is defined to be T H S/Y T −(T AU S ·Y T )/P OP (see Table A. this procedure is crude. where D1 and T AU S are constants. This treatment allows a state and local marginal tax rate to be defined in equation 91: D1SM = D1S + (2 · T AU S · Y T )/P OP . 1982:3– 1983:2. Again.00111. This procedure is. The second is that changes in the tax laws affect D1 but not T AU S. 1985:2–1985:2. The 27 subperiods are: 1952:1–1953:4. and T H G is explained by equation 47 as [D1G + (T AU G · Y T )/P OP ]Y T . 1977:1–1977:1. Given T AU G. 1991:1–1991:4. 1968:1–1970:4. (The estimation period was 1952:1–1993:2 excluding 1987:2. 1971:1– 1971:4. DU M3(Y T /P OP ). DU M2(Y T /P OP ). These two assumptions led to the estimation of an equation with T H S/P OP on the left hand side and a constant. 1975:2–1976:4. crude. and 1992:1–1993:2. of course. 1987:3–1987:4. D1G is defined to be T H G/Y T − (T AU G · Y T )/P OP (see Table A. with a t-statistic of 1. and (Y T /P OP )2 on the right hand side.2 THE US MODEL 45 made about the relationship between T H S and Y T .3.) The estimate of the coefficient of DU Mi(Y T /P OP ) is an estimate of D1 for subperiod i. The estimate of T AU G was .6). where DU Mi is a dummy variable that takes on a value of one in subperiod i and zero otherwise. 1985:3–1987:1. The estimate of the coefficient of (Y T /P OP )2 is the estimate of T AU S. This . 1977:2–1978:2. In the model D1G is taken to be exogenous. and T H S is explained by equation 48 as [D1S + (T AU S · Y T )/P OP ]Y T .84. 1983:3–1984:4. 1981:4–1982:2. A similar estimation procedure was followed for T AU G as was followed above for T AU S. DU M1(Y T /P OP ). 1988:1–1988:4. The observation for 1987:2 was excluded because it corresponded to a large outlier. 1987:2–1987:2.34. 1978:3–1981:3.00270. The estimate of T AU S was . T AU G: Progressivity Tax Parameter—g T AU G is the progressivity tax parameter in the personal income tax equation for the federal government (equation 47). 1989:1–1989:4. but it provides a rough estimate of the progressivity of the federal personal income tax system. 1985:1–1985:1. 1990:1– 1990:4. Given T AU S. where 27 subperiods where chosen. 1965:1–1965:4. In the model D1S is taken to be exogenous. 1973:1–1973:4. 1964:1–1964:4.6). 1966:1–1967:4. but at least it provides a rough estimate of the progressivity of the state and local personal income tax system. The first is that within a subperiod T H S/P OP equals [D1 + T AU S(Y T /P OP )](Y T /P OP ) plus a random error term. 1954:1–1963:4. 1974:1–1975:1. 1972:1–1972:4. with a t-statistic of 11.

End of year estimates of the stock of durable goods are available from 1925 through 1990 from the Survey of Current Business. Given the value of KD at. KH : Stock of Housing KH is an estimate of the stock of housing of the household sector. which is the value published in the Survey for 1952. say.46 3 THE DATA. fourth quarter) values of KD could be matched to the values published in the Survey. The value of DELD that was chosen as achieving a good match is . First. Table 4. which are available from the NIPA.. the end of 1952 and given a value of DELD. A quarterly series for KD was then constructed using this value and a base quarter value of 313. since this series matched the published stock data. the value for DELH was chosen using total residential investment as the investment series. January 1992. quarterly observations for KD can be constructed once a base quarter value and a value for the depreciation rate DELD are chosen. which in the model is I H H + I H K + I H B. p.) The residential stock data that is published in the Survey is for total residential investment. AND EQUATIONS treatment allows a federal marginal tax rate to be defined in equation 90: D1GM = D1G + (2 · T AU G · Y T )/P OP . whereas equation 59 pertains only to the residential investment of the household sector. The procedure that was used for dealing with this difference is as follows. KD: Stock of Durable Goods KD is an estimate of the stock of durable goods.049511. The value of DELH that was chosen as achieving a good match of the created stock data to the published stock data is . a quarterly series for KD can be constructed using the above equation and the quarterly series for CD. 137. Second. VARIABLES. . (The housing stock data are also in Table 4 in the January 1992 issue of the Survey.006716. It is defined by equation 59: KH = (1 − DELH )KH−1 + I H H (59) A similar procedure was followed for estimating DELH as was followed for estimating DELD.7 in 1952:4.e. This was done for different values of DELD to see how close the constructed end of year (i. It is defined by equation 58: KD = (1 − DELD)KD−1 + CD (58) Given quarterly observations for CD.

014574) for the 1952:1–1970:4 period.71 is the average of I KF /(I KF + I KH + I KB + I KG) over the sample period. one (. The nonresidential stock data that is published in the Survey is for total fixed nonresidential investment. A similar procedure was followed here as was followed for residential investment above. once the values for DELK were chosen. one (. and one (. A base quarter value of 887.13. The . the data on V in Table 5. the nonfarm stock). which is .0 in 1988:4.98 is the average of I H H /(I H H + I H K + I H B) over the sample period. whereas equation 92 pertains only to the fixed nonresidential investment of the firm sector. A base quarter value of 1270. KH was constructed using I H H (not total residential investment). the stock data are in Table 4 in the January 1992 issue of the Survey..3. inventory investment (I V F ) is equal to the change in the stock.023068) for the 1981:1–1993:2 period.571 in 1952:4 was used.71 times the value published in the Survey for 1952. First. Second. KK was constructed using I KF (not total fixed nonresidential investment). since this series matched the published stock data. KK: Stock of Capital KK is an estimate of the stock of capital of the firm sector.018428) for the 1971:1–1980:4 period. It is determined by equation 92: KK = (1 − DELK)KK−1 + I KF (92) A similar procedure was followed for estimating DELK as was followed for estimating DELD and DELH . The value .e.) It turned out in this case that three values of DELK were needed to achieve a good match. V : Stock of Inventories V is the stock of inventories of the firm sector (i. which is .2 THE US MODEL 47 once DELH was chosen. (Again.98 times the value published in the Survey for 1952. which is equation 117: I V F = V − V−1 (117) Both data on V and I V F are published in the Survey. The value . which in the model is I KF + I KH + I KB + I KG.276 in 1952:4 was used. For present purposes V was constructed from the formula V = V−1 + I V F using the I V F series and base quarter value of 870. By definition. the values for DELK were chosen using total fixed nonresidential investment as the investment series.

J F . The variables Y . In order to test for this in the empirical work. Excess Labor and Excess Capital In the theoretical model the amounts of excess labor and excess capital on hand affect the decisions of firms. This in turn requires an estimate of the technology of the firm sector. denoted J H MI N in the model. but not on the number of hours actually worked per worker (H F a ).13 in the July 1992 issue of the Survey. µ(KK · H K a )] (3. The measurement of the capital stock KK is discussed above. Y /(J F · H F ). To be precise with this technology one has to keep track of the purchase date of each machine and its technological coefficients. Equation 92 for KK and the production function 3. the estimate of the number of worker hours required to produce the output of the quarter.3 Output per paid for worker hour. This kind of detail is not possible with aggregate data. The values of λ other than those at the peaks were assumed to lie on straight lines between the peaks. AND EQUATIONS base quarter value was taken from Table 5. H K a is the number of hours each unit of KK is utilized.48 3 THE DATA. the others are not. H F a is the number of hours worked per worker. KK is the capital stock discussed above. which implies that the values of λ in equation 3. In the model.1 are not consistent with the putty-clay technology of the theoretical model. λ is estimation of excess labor in the following way was first done in Fair (1969) using three digit industry data. and one must resort to simpler specifications. The production function of the firm sector for empirical purposes is postulated to be Y = min[λ(J F · H F a ). one needs to estimate the amounts of excess labor and capital on hand in each period. is simply Y /λ. For example.1. was plotted for the 1952:1– 1993:2 period.1) where Y is production. The peaks of this series were assumed to correspond to cases in which the number of hours worked equals the number of hours paid for. and KK are observed. Given the production function 3. data on the number of hours paid for per worker exist (H F in the model). excess labor was measured as follows. Given an estimate of λ for a particular quarter and given equation 3. VARIABLES. 3 The .1. and λ and µ are coefficients that may change over time due to technical progress. This gives an estimate of λ for each quarter. J F is the number of workers employed.1 are observed at the peaks.

The largest value for excess capital was 10. where H K a is the average number of hours that each machine is utilized. finally. from the production function 3. and the equation determining J H MI N is equation 94 in Table A. for some of the interpolations the values after the last peak were taken to be the value at the last peak. In ¯ the model. Interpolation between peaks can then produce a ¯ .6 in the description of LAM.5 before the first peak were assumed to lie on the backward extension of the line connecting the first and second peaks. A negative value occurs when the actual value of output per paid for worker hour or output per capital is above the interpolation line.9 percent in 1960:4. This is. Contrary to the case for LAM. The peak to peak interpolation lines were not always drawn so that every point between the peaks lay on or below the line.3.1. If it is assumed that ¯ at each peak of this series H K a is equal to the same constant. then one ¯ observes at the peaks µ · H .5 percent in 1982:4.6 in the description of MU H . The actual capital stock (KK) can be compared to KKMI N to measure the amount of excess capital on hand. H is assumed to be the maximum number ¯ complete series on µ · H ¯ of hours per quarter that each unit of KK can be utilized.3. Similarly. The estimated percentages of excess labor and capital by quarter are presented in Table 3. then Y /(µ · H ) is the minimum amount of capital required to produce Y (denoted KKMI N ). 5 A few values in Table 3.1. and for capital each figure is 100 times (KK/KKMI N − 1. and thus one must be content with plotting Y /KK. The peaks that were used for the interpolations are listed in Table A. for some of the peak to peak interpolations in this study the values before the first peak were taken to be the value at the first peak.0).0]. say H . µ · H is denoted MU H . This is denoted “flat end” in Table A.6. If.4 For the measurement of excess capital there are no data on hours paid for or worked per unit of KK. This is denoted “flat beginning” in Table A. The largest value for excess labor during the entire 1952:1–1993:2 period was 4.1 are negative. The actual number of workers hours paid for (J F ·H F ) can be compared to J H MI N to measure the amount of excess labor on hand. a plot of µ · H K a . and the equation determining KKMI N is equation 93 in Table A.3. Also.2 THE US MODEL 49 denoted LAM. The table shows that in the most recent recession both excess labor and capital peaked at 3. For labor each figure in the table is 100 times [(J F · H F )/J H MI N − 1. 4 The values of LAM . The peaks that were used for the interpolations are listed in Table A.6 percent in 1991:1.6. the values of LAM after the last peak were assumed to lie on the forward extension of the line connecting the second to last and last peak.

1 1.1 1.0 2.2 3.0 1.2 -1.6 .2 .4 Quar.1 9.2 1. ExL ExK 1962:3 1962:4 1963:1 1963:2 1963:3 1963:4 1964:1 1964:2 1964:3 1964:4 1965:1 1965:2 1965:3 1965:4 1966:1 1966:2 1966:3 1966:4 1967:1 1967:2 1967:3 1967:4 1968:1 1968:2 1968:3 1968:4 1969:1 1969:2 1969:3 1969:4 1970:1 1970:2 1970:3 1970:4 1971:1 1971:2 1971:3 1971:4 1972:1 1972:2 1972:3 1972:4 .1 -.1 -.1 10.0 1. ExL ExK Quar.2 1.4 10.7 .2 .0 . usually maintenance work.5 .9 1.0 1.4 3.9 .1 .1 2.8 3.3 .0 2. AND EQUATIONS Table 3.2 4.0 .1 .0 .9 1.8 1.3 2.8 1.5 .2 3.1 -.8 4.8 5.1 .9 2. The estimates of excess labor in Table 3.2 1.3 1.8 1.2 -.0 1.0 5.7 1. Some of these hours were used for other worthwhile work.7 3.0 -1.0 .3 2.2 .0 . the typical plant paid for about 8 percent more blue collar hours than were needed for regular production work.2 4.7 1.9 2.6 .2 .6 2.0 3.3 2. 1983:3 1983:4 1984:1 1984:2 1984:3 1984:4 1985:1 1985:2 1985:3 1985:4 1986:1 1986:2 1986:3 1986:4 1987:1 1987:2 1987:3 1987:4 1988:1 1988:2 1988:3 1988:4 1989:1 1989:2 1989:3 1989:4 1990:1 1990:2 1990:3 1990:4 1991:1 1991:2 1991:3 1991:4 1992:1 1992:2 1992:3 1992:4 1993:1 1993:2 ExL 1.4 2.5 1.8 1.2 .5 1.2 .1 -.2 1.2 3.1 .0 1.0 1973:1 1973:2 1973:3 1973:4 1974:1 1974:2 1974:3 1974:4 1975:1 1975:2 1975:3 1975:4 1976:1 1976:2 1976:3 1976:4 1977:1 1977:2 1977:3 1977:4 1978:1 1978:2 1978:3 1978:4 1979:1 1979:2 1979:3 1979:4 1980:1 1980:2 1980:3 1980:4 1981:1 1981:2 1981:3 1981:4 1982:1 1982:2 1982:3 1982:4 1983:1 1983:2 -.6 3.4 1.9 5.6 .2 .8 2.4 .9 6.5 7.3 2.7 2.0 1.4 3.3 2.8 2.7 .8 Comparisons to the Fay-Medoff Estimates6 It is of interest to compare the estimates of excess labor in Table 3.5 1.5 2.4 1.2 6.0 .9 .2 .8 .6 1.9 1.4 1.1 .0 .5 1.5 1.3 .6 6.1 with the survey results of Fay and Medoff (1985).3 2.2 2.9 1.4 5.2 1.0 2.5 3.4 . and after taking account of this.5 .7 6.0 .0 2.2 2.5 3.2 2.1 .0 1.2 -1.6 1.0 1.0 1.2 2.3 3.1 -1.1 .7 3.S.2 4.3 .9 2.4 9.8 4.1 2.2 .6 3.3 2.8 1.2 ExK 4.2 1.5 2.2 .0 .0 1.1 3.0 1.2 .5 3.0 -.1 probably correspond more to the concept behind the 8 percent number of Fay and Medoff than to the concept 6 The following discussion is updated from Fair (1985).1 1.1 2.3 1.4 2.9 4.0 1.3 . VARIABLES.1 .4 .3 2.0 1.1 2.7 4.2 .4 2.9 3.3 4.2 1.0 .7 .0 .0 4.7 .8 2. They found that during its most recent trough quarter.3 .4 2.5 2.4 Quar.3 .0 3.1 1.0 .6 .9 2.9 3.9 6.6 3. manufacturing plants to examine the magnitude of labor hoarding during economic contractions.8 1.2 .9 3.5 1.0 1.5 .7 -.0 2.8 8.2 4.7 ExK 1.3 1.7 5.0 -.2 3.3 5.0 1.1 .1 .6 1.9 2.0 -.6 3.3 1.1 -.7 1.3 .5 .8 1. Fay and Medoff surveyed 168 U.0 .2 5.3 .7 .1 1.9 4.6 1.3 2.6 .0 2.6 1.4 .5 .0 . 4 percent of the blue collar hours were estimated to be hoarded for the typical plant.7 .1 .4 .5 -.3 -.0 3.6 2.9 7.7 1.4 2.1 1.0 .2 2.4 .9 2.4 2.4 .9 3.6 -.0 1.3 3.9 4.3 2.2 1.1 .1 2.5 3.50 3 THE DATA.3 5.4 1.2 1.5 -.3 .3 8.6 3.3 7.6 1.1 1.6 .1 1.5 .7 .0 -.8 -.1 3.7 -.4 6.8 .4 1.0 6.3 .1 .0 2.1 1.6 3.5 6.8 -.3 .6 6.0 1.4 .0 .0 .7 1.1 . 1952:1 1952:2 1952:3 1952:4 1953:1 1953:2 1953:3 1953:4 1954:1 1954:2 1954:3 1954:4 1955:1 1955:2 1955:3 1955:4 1956:1 1956:2 1956:3 1956:4 1957:1 1957:2 1957:3 1957:4 1958:1 1958:2 1958:3 1958:4 1959:1 1959:2 1959:3 1959:4 1960:1 1960:2 1960:3 1960:4 1961:1 1961:2 1961:3 1961:4 1962:1 1962:2 ExL 1.1 2.3 2.4 5.2 .0 1.4 2.9 5.5 .5 1.4 .6 .0 2. .1 2.6 1.1 Estimated Percentages of Excess Labor and Capital Quar.3 .2 9.0 -.6 .8 1.7 .0 2.1 2.9 2.0 .5 1.6 2.4 3.5 9.5 3.

3 percent values in Table 3. For the last responses the trough might be in 1982. If the same percentage error has been made at each peak. and the aggregate estimates in Table 3. the error will merely be absorbed in the estimates of the constant terms in the equations.5 percent in 1982:1. maintenance work is shifted from high to low output periods. The first survey upon which the Fay-Medoff results are based was done in August 1981. The plant managers were asked to answer the questionnaire for the plant’s most recent trough. whereas for the earlier ones the trough is likely to be in 1980. One would thus expect the Fay-Medoff estimate to be somewhat higher than the aggregate estimates. First.7 The Fay-Medoff estimate of 8 percent is thus compared to the 3.5 percent in 1980:4 and 5. like the concept behind the peak to peak interpolation work.1 shows that the percentage of excess labor reached 3. and the second (larger) survey was done in April 1982. A follow up occurred in October 1982. and 8 percent versus a number around 3 to 3. since not all troughs are likely to occur in the same quarter across plants.2 THE US MODEL 51 behind the 4 percent number.4 percent in 1980:2 and 3. The error is also not serious for the Fay-Medoff comparisons as long as the Fay-Medoff concept behind the 8 percent number is used.1. the one in mid 1980 and the one in early 1982. . the trough in output for a given plant is on average likely to be deeper than the trough in aggregate output. then J H MI N is a misleading estimate of worker hour requirements. Also. The Fay-Medoff results appear to provide strong support for the excess labor hypothesis. The use of more recent data has thus lowered the excess labor estimates by a little over a percentage point.1 are for the total private sector. There are two possible troughs that are relevant for the Fay-Medoff study. J H MI N is too low because it has been based on the incorrect assumption that the peak productivity values could be sustained over the entire business cycle. These two sets of results seem consistent in that there are at least two reasons for expecting the Fay-Medoff estimate to be somewhat higher. not just manufacturing. In a long run sense. which is likely to be approximately the case. This error is not a serious one from the point of estimating the labor demand equations in Chapter 5. Table 3.3. does not account for maintenance that is shifted from high to low output periods. At a micro level Fay and Medoff found labor hoarding and 7 The estimates in Fair (1985) using earlier data were 4. Second.3 percent in 1982:1. the Fay-Medoff estimate of 4 percent hoarded labor cited above was 5 percent in an earlier version of the paper cited in Fair (1985).5 percent seems consistent with this. This concept. for example. the manufacturing sector may on average face deeper troughs than do other sectors. If.4 and 3.

1 − J J P /J J ). The terms that are subtracted from J J P · P OP in equation 98 are. VARIABLES. Z is a labor constraint variable in the sense that it is zero or close to zero when the worker hours-population ratio is at or near its peak and gets progressively larger in absolute value as the ratio moves below its peak. where Z is called the “labor constraint variable.6 in the description of J J P . and the peaks that were used for this interpolation are presented in Table A. The exact use of Z is explained in Chapter 5. This is one of the few examples in macroeconomics where a hypothesis has been so strongly confirmed using detailed micro data. Since Y S in equation 98 is LAM times the potential number of workers in the firm sector.52 3 THE DATA. the number of federal military worker hours. however. In the empirical work one needs some way of measuring this constraint. J J was first plotted for the 1952:1–1993:2 period. AND EQUATIONS of a magnitude that seems in line with aggregate estimates. the predicted value of J J may be greater than J J P . Y S is defined by equation 98: Y S = LAM(J J P · P OP − J G · H G − J M · H M − J S · H S) (98) J J P is the peak or potential ratio of worker hours to population (as constructed from the peak to peak interpolation of J J ). in order. Y S: Potential Output of the Firm Sector A measure of the potential output of the firm sector. Labor Market Tightness: The Z Variable An important feature of the theoretical model is the possibility that households may at times be constrained in how much they can work. and so J J P · P OP is the potential number of worker hours. In the solution of the model. The variable J J in the model is the ratio of the total number of worker hours paid for in the economy to the total population 16 and over (equation 95). LAM is the coefficient λ in the production function 3. and the number of state and local government worker hours. the number of federal civilian worker hours. A variable Z was then defined as min(0. The entire number in parentheses is thus the potential number of worker hours in the firm sector. Y S. is used in the price equation (equation 10).1. in which case Z is taken to be zero. The interpolation series is denoted J J P . it can be interpreted as the .” In the data Z is always nonpositive because J J P is constructed from the peak to peak interpolations and is always greater than or equal to J J . and a peak to peak interpolation was made. The approach taken here is the following.

2 THE US MODEL 53 potential output of the firm sector unless the capital input is insufficient to produce Y S. of course. 43. 33. The variable D2S is then interpreted as a tax rate and is taken to be exogenous. Data exist for P I H and P D.2. and D2S was constructed as T F S/P I EF . 44-54.3 are of two types. They are used in the interest payments equations. 3. and 89–131. and P SI 5 is a ratio. and 88. 55. and D2S is the “tax rate. but no attempt has been made to explain these changes. 56. Consider equation 38: P I H = P SI 5 · P D (38) where P I H is the price deflator for housing investment. These identities are equations 31.3. 57.” Data exist for T F S and P I EF . A similar procedure was followed to handle relative price changes. varies over time as tax laws and other things that affect the relationship between T F S and P I EF change.4 The Identities The identities in Table A. BG is an estimate of the value of long term bonds issued by the federal government sector in the current period. 35-42. These variables are determined by equations 55 and 56 respectively. P D is the price deflator for total domestic sales. Consider. 58–87. The construction of BF and BG is somewhat involved. equation 50: T F S = D2S · P I EF (50) where T F S is the amount of corporate profit taxes paid from firms (sector f) to the state and local government sector (sector s). This rate. and this discussion is presented in Chapter 5 in the context of the discussion of equations 19 and 29. 34. where the rate or ratio has been constructed to have the identity hold. This construction of Y S is thus based on the assumption that there is always sufficient capital on hand to produce Y S. . P I EF is the level of corporate profits of the firm sector. 19 and 29. This general procedure was followed for the other identities involving tax rates. The Bond Variables BF and BG BF is an estimate of the value of long term bonds issued by the firm sector in the current period. One type simply defines one variable in terms of others. These identities are equations 32. The other type defines one variable as a rate or ratio times another variable or set of variables. for example. Similarly.

MB is the net value of demand deposits of the financial sector.54 3 THE DATA. VARIABLES. The sum of the saving variables across the six sectors is zero. reconciles the two data sets. Chapter 5 discusses the specification. (Prices relative to wages are not exogenous. etc. is taken to be exogenous. AND EQUATIONS and P SI 5 was constructed as P I H /P D. 3. Equation 66 states that any nonzero value of saving of the household sector must result in a change in AH or MH . P SI 5.2. Of the .) is determined by an identity. and 79 for the state and local government sector. SH is from the NIPA.3 that the saving of each sector (SH . This procedure was followed for the other identities involving prices and wages. estimation. for example. and the other variables are from the FFA. (MB is negative. and testing of these equations. 77 for the federal government sector. There are equations like 66 for each of the other five sectors: equation 70 for the firm sector. is the budget constraint of the federal government sector. Note also from Table A. and so the minus sign makes G1 positive.) It is beyond the scope of the model to explain relative prices and wages. DI SH . Equation 77. An identity like equation 66 0 = SH − AH − MH + CG − DI SH (66) is concerned with this linkage. This treatment means that relative prices and relative wages are exogenous in the model. and G1 is a “reserve requirement ratio. The discrepancy variable. which is from the FFA. and the foregoing treatment is a simple way of handling these changes.5 The Stochastic Equations A brief listing of the stochastic equations is presented in Table A.” Data on BR and MB exist. SF . Many of the identities of the first type are concerned with linking the FFA data to the NIPA data. since the financial sector is a net debtor with respect to demand deposits.3. Another identity of the second type is equation 57: BR = −G1 · MB (57) where BR is the level of bank reserves. and G1 was constructed as −BR/MB. which is the reason that equation 80 is redundant.) G1 is taken to be exogenous. It varies over time as actual reserve requirements and other features that affect the relationship between BR and MB change. The left hand side and right hand side variables are listed for each equation. however. 75 for the foreign sector. which varies over time as the relationship between P I H and P D changes. 73 for the financial sector.

1–B.7) Table B.7 explains the construction of the balance of payments data—data for variables S and T T . and the others will be referred to as “annual” countries. 3. the next to the foreign sector. When quarterly data were needed but only annual data were available.3. and annual data were collected for the others. and these will be outlined first. the next to the state and local government sector.3 The ROW Model The data. Countries 34 through 45 are countries for which only trade share equations are estimated.4 lists the equations that pertain to the trade and price links among the countries.7 in Appendix B.3. did not exist. Table B. where the variables are listed in alphabetical order. the next twelve to the firm sector. Finally. A detailed description of the variables per country is presented in Table B. and it explains the construction of the trade share variables—the αij variables.3 lists the stochastic equations and the identities.1–B. which are countries 2 through 33.2. Table B.1 The Tables (Tables B. Data permitting. The way in which each variable was constructed is explained in brackets in Table B.5 lists the links between the US and ROW models. variables. the first nine pertain to the household sector. In some cases. and the final two to the federal government sector. Countries 2 through 14 will be referred to as “quarterly” countries. but not the coefficient estimates. It also explains how the quarterly and annual data were linked for the trade share calculations.2. The functional forms of the stochastic equations are given. each of the 32 countries has the same set of variables. Table B. All of the data with potential seasonal fluctuations have been seasonally adjusted.1 lists the countries in the model and provides a brief listing of the variables per country. 3. such as a population variable. and equations for the ROW model are discussed in this section.3 THE ROW MODEL 55 thirty equations. . The 32 countries for which structural equations are estimated are Canada (CA) through Thailand (TH). Table B. Quarterly data were collected for countries 2 through 14. the next five to the financial sector. quarterly observations were interpolated from annual data using the procedure described in Table B. Remember that the ROW model includes structural models for 32 countries. quarterly data for a particular variable.6. The relevant tables for the model are Tables B. The coefficient estimates for all the countries are presented in Chapter 6.

e. Data from the OECD are quarterly national accounts data. Data from the IMF are international financial statistics (IFS) data and direction of trade (DOT) data. As noted above. AND EQUATIONS It will be useful to begin with a discussion of the construction of some of the variables in Table B. It is important to note that A measures only the net asset position of the country vis-à-vis the rest of the world.7. The two key variables involved in this process are S.3. and A: Balance of Payments Variables One important feature of the data collection is the linking of the balance of payments data to the other export and import data. When “IFS” precedes a number or letter in the table. is not included.3. It was constructed by summing past values of S from a base period value of zero. and T T . Quarterly balance of payments data do not generally begin as early as the other data. 3. this refers to the IFS variable number or letter.2.2 The Raw Data The data sets for the countries other than the United States (i. T T .56 3 THE DATA. such as the domestically owned housing stock and plant and equipment stock. the balance of payments on current account. directly from the raw data). and the procedure in Table B. the value of net transfers..e.2. The summation begins in the first quarter for which data on S exist.3 Variable Construction S. 3. The variable A is the net stock of foreign security and reserve holdings. These are the “raw” data. and some were constructed from other (already constructed) variables. . Domestic wealth..7 allows quarterly data on S to be constructed as far back as the beginning of the quarterly data for merchandise imports and exports (M$ and X$). Some variables were constructed directly from IFS and OECD data (i. The sources of the data are the IMF and OECD. The construction of these variables and the linking of the two types of data are explained in Table B. the countries in the ROW model) begin in 1960. This means that the A series is off by a constant amount each period (the difference between the true value of A in the base period and zero). and annual labor force data. the way in which each variable was constructed is explained in brackets in Table B. quarterly labor force data. In the estimation work the functional forms were chosen in such a way that this error was always absorbed in the estimate of the constant term. annual national accounts data. VARIABLES.

equation 92 for the US model. and 3) data on I .3 THE ROW MODEL K: The Capital Stock 57 If depreciation is proportional to the capital stock K.015 (1.5 percent) per quarter for fixed nonresidential and residential capital combined. It provides at least a rough estimate of the capital stock of each country.) Given this equation and data for V 1.015 (. then K = (1−δ)K−1 +I . This rate is about . 2) a base value for K. and K was constructed for each period using this base value and the depreciation rate of . the preliminary base value was changed. but it is about the best that can be done given the available data. V : Stock of Inventories Data on inventory investment. are available for each country. the U. If the ratios in the two periods were similar. where δ is the depreciation rate and I is gross investment. Although. for example. This means that the constructed data for . (For countries with annual data. a base value of K was constructed. denoted V 1 in the ROW model.2. denoted V .06 for the annual countries). and so the appropriate depreciation rate is for the sum of the two.) Given 1) a value for δ. V can be constructed once a base period and base period value are chosen. By definition V = V−1 + V 1.) This rate is the value that was used for δ in the construction of K for each of the other countries. (See. The capital-output ratio (K/Y ) was then computed for the first and last periods. First. and the process was repeated. Therefore. the base value was used. (See.3. as discussed in Section 3. but not data on the stock of inventories. some way must be found for constructing K for the other countries that does not require direct data on K.) Second. the value used for δ was . only data on investment exist for most other countries. In other words. (The data on I for the other countries includes both fixed nonresidential and residential investment. for example. This procedure for constructing data on K is obviously crude. equation 117 for the US model.06. Otherwise. data were used to compute an implicit depreciation rate. and the base period value was taken to be zero.3. This was done as follows. data on both the capital stock and investment exist for the United States. The process was stopped when the ratios in the first and last periods were similar. the base value was chosen so as to make the capital-output ratio have no long run trend. The base period was chosen for each country to be the quarter or year prior to the beginning of the data on V 1. A preliminary base value was chosen. K can be constructed from this formula.S.

A similar procedure was followed to measure excess capital. VARIABLES. J . µ(K · H K a )] (3. and K are observed (or. J is the number of workers employed. If. contrary to the case for the United States.2) where Y is production. Given the production function 3. Y /K was plotted over the sample period. and peaks of this series were chosen. H J a is used in place of H F a . Data on Y . and λ · H J was assumed to lie on the lines. and λ and µ are coefficients that may change over time due to technical progress. Also. H J is assumed to be the maximum number of hours that each worker can work. Note also that Y refers here to the total output of the country (real GDP). If it is assumed that at each peak H K a is equal to . and peaks of this series were chosen. J is used in place of J F because there is no disaggregation in the ROW model between the firm sector and other sectors. data on the number of hours paid for per worker (denoted H F in the US model) are not available. finally. Y/J was plotted over the sample period. H K a is the number of hours each unit of K is utilized.1) for the United States. Straight lines were drawn between the peaks (peak to peak interpolation). say H J . This is from 3.2) is changed slightly from that in (3. K is the capital stock discussed above.2 a plot of λ · H J a .2.2 a plot of µ · H K a . Excess Labor and Excess Capital As was the case for the United States. If it is assumed that at each peak H J a is equal to the same constant. K is used in place of KK because there is no disaggregation in the ROW model between types of capital. not just the output of the firm sector.3. H J a is the number of hours worked per worker.58 3 THE DATA. The notation in equation (3. Similarly. then Y/(λ · H J ) is the minimum number of workers required to produce Y . excess labor was measured as follows for each country. Finally. and the equation determining J MI N is equation I-13 in Table B. which is the production equation 4. The actual number of workers on hand (J ) can be compared to J MI N to measure the amount of excess labor on hand. the others are not. then one observes at the peaks λ · H J . in the case of K. constructed). This error is absorbed in the estimate of the constant term in the equation in which V appears as an explanatory variable. the short run production function for each country is assumed to be one of fixed proportions: Y = min[λ(J · H J a ). which is denoted J MI N in the ROW model. This is from 3. AND EQUATIONS V are off by a constant amount throughout the sample period ( the difference between the true value in the base period and zero). λ · H J is denoted LAM.

It is exogenous in the model. which is equation I-16 in Table B.3. The only difference is that here output refers to the total output of the country rather than just the output of the firm sector. and J J P (the peak to peak interpolation series) was constructed. H K is assumed to be the maximum number of hours that each machine can be utilized.3.4 The Identities The identities for each country are listed in Table B.3.3 except that the level of imports (I M) has to be subtracted in I-3 because C.e.. and µ · H K was assumed to lie on the lines. . Straight lines were drawn between the peaks. ZZ is determined by equation I-18 in Table B.4 in Section 2. Y S: Potential Output A measure of potential output (Y S) was constructed for each country in the same manner as was done for the United States. finally. There are up to 19 identities per country. Equation I-2 is a similar equation for exports.3 THE ROW MODEL 59 the same constant. Equation I-1 links the non NIPA data on imports (i. Given Y S.. The income identity I-3 is the empirical version of equation 2. The actual capital stock (K) can be compared to KMI N to measure the amount of excess capital on hand. data on M and MS) to the NIPA data (i. If. then Y/(µ · H K) is the minimum amount of capital required to produce Y . equation I-4 defines inventory investment as the difference between production and sales.e. which is equation I-17 in Table B. and equation I-5 defines the stock of inventories as the previous stock plus inventory investment.2. and the equation determining KMI N is equation I-11 in Table B. which is denoted ZZ in the ROW model. The equation for Y S is Y S = LAM · J J P · P OP . µ · H K is denoted MU H .3.3. For each country a peak to peak interpolation of J J (= J /P OP ) was made. then one observes at the peaks µ · H K. Equation I-3 is the income identity. I . Labor Market Tightness: The Z variable A labor market tightness variable (the Z variable) was constructed for each country in the same manner as was done for the United States. data on I M).3.3 about the Z variable. 3. which is denoted KMI N in the ROW model. See the discussion in Section 3.3. and G include imports. say H K.2. a gap variable can be constructed as (Y S − Y )/Y S. The variable I MDS in the equation picks up the discrepancy between the two data sets. Z is then equal to the minimum of 0 and 1 − J J P /J J .

merchandise imports from the countries in the trade share matrix in 85 $. the import price index computed from the trade share calculations. Equation I-10 defines the capital stock.5 The Stochastic Equations The stochastic equations for a given country are listed in Table B. P SI 2. U R. J is total employment. which is generally the case in the data. Section 2. Equation I-7 defines A. the average exchange rate for the period. then E is approximately equal to (EE + EE−1 )/2. These have all been discussed above.3. (Remember that A is constructed by summing past values of S. It is exogenous in the model. to M85$A. Equation I-8 links M.3.2. As will be discussed in Chapter 6. These two equations were discussed above. estimation. the net stock of foreign security and reserve holdings. 3. the balance of payments on current account. including the armed forces. U R is equal to the number of people unemployed divided by the civilian labor force. and potential output. the end of period exchange rate. The variable M85$B is the difference between total merchandise imports (in 85$) and merchandise imports (in 85$) from the countries in the trade share matrix. AND EQUATIONS Equation I-6 defines S.60 3 THE DATA. There are up to 15 estimated equations per country. the labor constraint variable. equation I-19 links P M.2. which is equation I-9. If the exchange rate changes fairly smoothly within the period. many of these equations . total merchandise imports in 85 lc.8. to P MP . It will be useful to relate some of the equations in the table to those in the theoretical model in Chapter 2. and AF is the level of the armed forces. A variable P SI 1 was defined to make the equation E = P SI 1[(EE+EE−1 )/2] exact.3. to EE. One would expect P SI 1 to be approximately one and not to fluctuate much over time. the import price index obtained from the IFS data. as equal to last period’s value plus S. and equation I-11 defines the minimum capital stock needed to produce the output. Chapter 6 discusses the specification. Equations I-13 through I-18 pertain to the measurement of excess labor.3. The variable that links the two. and testing of these equations. Equation I-9 links E. is taken to be exogenous.2. L1 is the labor force of men. Finally. and L2 is the labor force of women. Equation I-12 defines the civilian unemployment rate. the saving of the country.) This is the empirical version of equation i in Section 2. VARIABLES. This is the empirical version of equation ii in Section 2.

3.3 THE ROW MODEL

61

are similar to the corresponding equations in the US model. Equation 1 in Table B.3 explains the demand for imports. It is matched to equation 2.2 of the theoretical model. Equation 2 explains consumption. It is matched to equation 2.1 except that consumption for equation 2 includes consumption of imported goods. In the theoretical model Xh is only the value of domestically produced goods consumed. Equation 3 explains fixed investment, and equation 4 explains production with sales as an explanatory variable, which is in effect an inventory investment equation. Neither of these equations was included in the theoretical model. The price equation 5 is matched to equation 2.3. Equation 6 explains the demand for money, and it is matched to equation 2.6. Equation 7 is an interest rate reaction function, explaining the short term interest rate RS. RS is equivalent to R in the theoretical model. (Interest rate reaction functions are discussed in Section 2.2.7.) Equation 8 is a term structure of interest rates equation, explaining the long term interest rate RB. The theoretical model does not contain a long term rate. Equation 9 is an exchange rate reaction function, explaining the exchange rate E. E is equivalent to e in the theoretical model. (Exchange rate reaction functions are also discussed in Section 2.2.7.) Equation 10 is an estimated arbitrage condition and explains the forward exchange rate. In the theoretical model this equation would be F = e 1+R , where F is the forward rate. 1+r Equation 11 explains the price of exports. In the theoretical model the price of exports is simply the price of domestic output, but this is not true in practice and an additional equation has to be introduced, which is equation 11. Equation 12 explains the wage rate; equation 13 explains the demand for employment; and equations 14 and 15 explain the labor force participation rates of men and women, respectively. These equations are not part of the theoretical model because it has no labor sector.

3.3.6

The Linking Equations

The equations that pertain to the trade and price links among countries are presented in Table B.4. (All imports and exports in what follows are merchandise imports and exports only.) The equations L-1 define the export price index for each country in U.S. dollars, P X$i . i runs from 1 through 44, and so there are 44 such equations. P X$i depends on the country’s exchange rate and on its export price index in local currency. The equations L-2 are the trade link equations. The level of exports of country i in 85 $, X85$i , is the sum of the amount that each of the other 44

62

3 THE DATA, VARIABLES, AND EQUATIONS

countries imports from country i. For example, the amount that country j imports from country i is αij M85$Aj , where αij is the fraction of country i’s exports imported by j and M85$Aj is the total imports of country j from the countries in the trade share matrix. There are 33 of these trade link equations. The αij values are determined from the trade share equations. These equations are discussed in Section 6.16, and the use of these equations in the solution of the model is discussed in Section 9.2. The equations L-3 link export prices to import prices, and there are 33 such equations. The price of imports of country i, P MPi , is a weighted average of the export prices of other countries (except for country 45, the “all other” category, where no data on export prices were collected). The weight for country j in calculating the price index for country i is the share of country j ’s exports imported by i. The equations L-4 define a world price index for each country, which is a weighted average of the 33 countries’ export prices except the prices of Saudi Arabia and Venezuela, the oil exporting countries. (As discussed in Section 6.12, the aim is to have the world price index not include oil prices.) The world price index differs slightly by country because the own country’s price is not included in the calculations. The weight for each country is its share of total exports of the relevant countries. Table B.5 explains how the US and ROW models are linked. When the two models are combined (into the MC model), the price of imports P I M in the US model is endogenous and the level of exports EX is endogenous.

4

Estimating and Testing Single Equations
4.1 Notation

This chapter discusses the estimation and testing of single equations. The notation that will be used is the same as that used in Section 1.2. The model is written as fi (yt , xt, αi ) = uit , (i = 1, . . . , n), (t = 1, . . . , T ) (4.1)

where yt is an n–dimensional vector of endogenous variables, xt is a vector of predetermined variables (including lagged endogenous variables), αi is a vector of unknown coefficients, and uit is the error term for equation i for observation t. It will be assumed that the first m equations are stochastic, with the remaining uit (i = m + 1, . . . , n) identically zero for all t. The following notation is also used. ui denotes the T –dimensional vector (ui1 , . . . , uiT ) . Gi denotes the ki × T matrix whose tth column is ∂fi (yt , xt , αi )/∂αi , where ki is the dimension of αi . α denotes the vector of all the unknown coefficients in the model: α = (α1 , . . . , αm ). The dimension of α is k, where k = m ki . Finally, Zi denotes a T × Ki matrix of i=1 predetermined variables that are to be used as first stage regressors for the two stage least squares technique. It will sometimes be useful to consider the case in which the equation to be estimated is linear in coefficients. In this case equation i in 4.1 will be written as yit = Xit αi + uit , (i = 1, . . . , n), (t = 1, . . . , T ) (4.2) 63

64

4 ESTIMATING AND TESTING SINGLE EQUATIONS

where yit is the left hand side variable and Xit is a ki –dimensional vector of explanatory variables in the equation. Xit includes both endogenous and predetermined variables. Both yit and the variables in Xit can be nonlinear functions of other variables, and thus 4.2 is more general than the standard linear model. All that is required is that the equation be linear in αi . Note from the definition of Gi above that for equation 4.2 Gi = Xi , where Xi is the ki × T matrix whose tth column is Xit . Each equation in 4.1 is assumed to have been transformed to eliminate any autoregressive properties of its error term. If the error term in the untransformed version, say wit in equation i, follows a rth order autoregressive process, wit = ρ1i wit−1 + . . . + ρri wit−r + uit , where uit is iid, then equation i is assumed to have been transformed into one with uit on the right hand side. The autoregressive coefficients ρ1i , . . . , ρri are incorporated into the αi coefficient vector, and the additional lagged values that are involved in the transformation are incorporated into the xt vector. This transformation makes the equation nonlinear in coefficients if it were not otherwise, but this adds no further complications to the model because it is already allowed to be nonlinear. It does result in the “loss” of the first r observations, but this has no effect on the asymptotic properties of the estimators. uit in 4.1 can thus be assumed to be iid even though the original error term may follow an autoregressive process. Many nonlinear optimization problems in macroeconometrics can be solved by general purpose algorithms like the Davidon-Fletcher-Powell (DFP) algorithm. This algorithm is discussed in Fair (1984), Section 2.5, and this discussion will not be repeated here. Problems for which the algorithm seems to work well and those for which it does not are noted below. Unless otherwise stated, the goodness of fit measures have not been adjusted for degrees of freedom. For the general model considered here (nonlinear, simultaneous, dynamic) only asymptotic results are available, and so if any adjustments were made, they would have to be based on analogies to simpler models. In many cases there are no obvious analogies, and so no adjustments were made. Fortunately, in most cases the number of observations is fairly large relative to numbers that might be used in the subtraction, and so the results are not likely to be sensitive to the current treatment.

4.2 TWO STAGE LEAST SQUARES

65

4.2

Two Stage Least Squares1

Probably the most widely used estimation technique for single equations that produces consistent estimates is two stage least squares (2SLS).2 The 2SLS estimate of αi (denoted αi ) is obtained by minimizing ˆ Si = ui Zi (Zi Zi )−1 Zi ui = ui Di ui (4.3)

with respect to αi . Zi can differ from equation to equation. An estimate of ˆ the covariance matrix of αi (denoted V2ii ) is ˆ ˆ ˆ ˆ ˆ V2ii = σii (Gi Di Gi )−1 (4.4)

ˆ where Gi is Gi evaluated at αi , σii = T −1 T u2 , and uit = fi (yt , xt , αi ). ˆ ˆ ˆ ˆ t=1 ˆ it The 2SLS estimate of the k × k covariance matrix of all the coefficient ˆ estimates in the model (denoted V2 ) is
    ˆ V2 =   

ˆ ˆ V211 . . . V21m . . . . . . ˆ ˆ V2m1 . . . V2mm

      

(4.5)

where

ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ V2ij = σij (Gi Di Gi )−1 (Gi Di Dj Gj )(Gj Dj Gj )−1
T ˆ ˆ t=1 uit uj t .

(4.6)

and σij = T −1 ˆ

4.3

Estimation of Equations with Rational Expectations3

With only slight modifications, the 2SLS estimator can be used to estimate equations that contain expectational variables in which the expectations are formed rationally. As discussed later in this chapter, this estimation technique
Fair (1984), Section 6.3.2, for a more detailed discussion of the two stage least squares estimator, especially for the case in which the equation is linear in coefficients and has an autoregressive error. 2 Ordinary least squares is used a lot in practice in the estimation of commercial models even when the estimator does not produce consistent estimates. This lack of care in the estimation of such models is undoubtedly one of the reasons there has been so little academic interest in them. 3 The material in this section is taken from Fair (1993b), Section 3 and Appendix A.
1 See

66

4 ESTIMATING AND TESTING SINGLE EQUATIONS

can be used to test the rational expectations hypothesis against other alternatives. The modifications of the 2SLS estimator that are needed to handle the rational expectations case are discussed in this section. It will be useful to begin with an example. Assume that the equation to be estimated is yit = X1it α1i + Et−1 X2it+j α2i + uit , (t = 1, . . . , T ) (4.7)

where X1it is a vector of explanatory variables and Et−1 X2it+j is the expectation of X2it+j based on information through period t − 1. j is some fixed positive integer. This example assumes that there is only one expectational variable and only one value of j , but this is only for illustration. The more general case will be considered shortly. A traditional assumption about expectations is that the expected future values of a variable are a function of its current and past values. One might postulate, for example, that Et−1 X2it+j depends on X2it and X2it−1 , where it assumed that X2it (as well as X2it−1 ) is known at the time the expectation is made. The equation could then be estimated with X2it and X2it−1 replacing Et−1 X2it+j in 4.7. Note that this treatment, which is common to many macroeconometric models, is not inconsistent with the view that agents are “forward looking.” Expected future values do affect current behavior. It’s just that the expectations are formed in fairly simply ways—say by looking only at the current and lagged values of the variable itself. Assume instead that Et−1 X2it+j is rational and assume that there is an observed vector of variables (observed by the econometrician), denoted here as Zit , that is used in part by agents in forming their (rational) expectations. The following method does not require for consistent estimates that Zit include all the variables used by agents in forming their expectations. Let the expectation error for Et−1 X2it+j be
t−1 it+j

= X2it+j − Et−1 X2it+j

(t = 1, . . . , T )

(4.8)

where X2it+j is the actual value of the variable. Substituting 4.8 into 4.7 yields yit = X1it α1i + X2it+j α2i + uit −t−1 it+j α2i = Xit αi + vit (t = 1, . . . , T ) (4.9) where Xit = (X1it X2it+j ), αi = (α1i α2i ) , and vit = uit −t−1 it+j α2i . Consider now the 2SLS estimation of 4.9, where the vector of first stage regressors is the vector Zit used by agents in forming their expectations. A

4.3 EQUATIONS WITH RATIONAL EXPECTATIONS

67

necessary condition for consistency is that Zit and vit be uncorrelated. This will be true if both uit and t−1 it+j are uncorrelated with Zit . The assumption that Zit and uit are uncorrelated is the usual 2SLS assumption. The assumption that Zit and t−1 it+j are uncorrelated is the rational expectations assumption. If expectations are formed rationally and if the variables in Zit are used (perhaps along with others) in forming the expectation of X2it+j , then Zit and t−1 it+j are uncorrelated. Given this assumption (and the other standard assumptions that are necessary for consistency), the 2SLS estimator of αi in equation 4.9 is consistent. The 2SLS estimator does not, however, account for the fact that vit in 4.9 is a moving average error of order j − 1, and so it loses some efficiency for values of j greater than 1. The modification of the 2SLS estimator to account for the moving average process of vit is Hansen’s (1982) generalized method of moments (GMM) estimator, which will now be described. Write 4.9 in matrix notation as yi = Xi αi + vi (4.10)

where Xi is T × ki , αi is ki × 1, and yi and vi are T × 1. Also, let Zi denote, as above, the T × Ki matrix of first stage regressors. The assumption in 4.9 that there is only one expectational variable and only one value of j can now be relaxed. The matrix Xi can include more than one expectational variable and more than one value of j per variable. In other words, there can be more than one led value in this matrix. The 2SLS estimate of αi in 4.10 is αi = [Xi Zi (Zi Zi )−1 Zi Xi ]−1 Xi Zi (Zi Zi )−1 Zi yi ˆ (4.11)

This use of the 2SLS estimator for models with rational expectations is due to McCallum (1976). As just noted, this use of the 2SLS estimator does not account for the moving average process of vit , and so it loses efficiency if there is at least one value of j greater than 1. Also, the standard formula for the covariance matrix of αi is not correct when at least one value of j is greater than 1. If, ˆ for example, j is 3 in 4.9, an unanticipated shock in period t + 1 will affect t−1 it+3 , t−2 it+2 , and t−3 it+1 , and so vit will be a second order moving average. Hansen’s GMM estimator accounts for this moving average process. The GMM estimate in the present case (denoted αi ) is ˜ αi = (Xi Zi Mi−1 Zi Xi )−1 Xi Zi Mi−1 Zi yi ˜ (4.12)

68

4 ESTIMATING AND TESTING SINGLE EQUATIONS

where Mi is some consistent estimate of lim T −1 E(Zi vi vi Zi ). The estimated covariance matrix of αi is ˜ T (Xi Zi Mi−1 Zi Xi )−1 (4.13)

There are different versions of αi depending on how Mi is computed. To ˜ compute Mi , one first needs an estimate of the residual vector vi . The residuals can be estimated using the 2SLS estimate αi : ˆ vi = yi − Xi αi ˆ ˆ (4.14)

A general way of computing Mi is as follows. Let fit = vit ⊗ Zit , ˆ where vit is the tth element of vi . Let Rip = (T − p)−1 T fit fit−p , ˆ ˆ t=p p = 0, 1, . . . , P , where P is the order of the moving average. Mi is then (Ri0 + Ri1 + Ri1 + . . . + RiP + RiP ). In many cases computing Mi in this way does not result in a positive definite matrix, and so αi cannot be computed. ˜ I have never had much success in obtaining a positive definite matrix for Mi computed in this way. There are, however, other ways of computing Mi . One way, which is discussed in Hansen (1982) and Cumby, Huizinga, and Obstfeld (1983) but is not pursued here, is to compute Mi based on an estimate of the spectral density matrix of Zit vit evaluated at frequency zero. An alternative way, which is pursued here, is to compute Mi under the following assumption: E(vit vis | Zit , Zit−1 , . . .) = E(vis vis ) , t ≥s (4.15)

which says that the contemporaneous and serial correlations in vi do not depend on Zi . This assumption is implied by the assumption that E(vit vis ) = 0, t ≥ s, if normality is also assumed. Under this assumption Mi can be computed as follows. Let aip = (T − p)−1 T vit vit−p and Bip = t=p ˆ ˆ T −1 (T − p) t=p Zit Zit−p , p = 0, 1, . . . , P . Mi is then (ai0 Bi0 + ai1 Bi1 + ai1 Bi1 + . . . + aiP BiP + aiP BiP ). In practice, this way of computing Mi usually results in a positive definite matrix. The Case of an Autoregressive Structural Error Since many macroeconometric equations have autoregressive error terms, it is useful to consider how the above estimator is modified to cover this case. Return for the moment to the example in 4.7 and assume that the error term uit in the equation follows a first order autoregressive process: uit = ρ1i uit−1 + ηit (4.16)

4.3 EQUATIONS WITH RATIONAL EXPECTATIONS

69

Lagging equation 4.7 one period, multiplying through by ρ1i , and subtracting the resulting expression from 4.7 yields yit = ρ1i yit−1 + X1it α1i − X1it−1 α1i ρ1i + Et−1 X2it+j α2i − Et−2 X2it+j −1 α2i ρ1i + ηit (4.17)

Note that this transformation yields a new viewpoint date, t − 2. Let the expectation error for Et−2 X2it+j −1 be
t−2 it+j −1

= X2it+j −1 − Et−2 X2it+j −1

(4.18)

Substituting 4.8 and 4.18 into 4.17 yields yit = ρ1i yit−1 + X1it α1i − X1it−1 α1i ρ1i + X2it+j α2i − X2it+j −1 α2i ρ1i +ηit −t−1
it+j α2i

+t−2

it+j −1 α2i ρ1i

= ρ1i yit−1 + Xit αi − Xit−1 αi ρ1i + vit

(4.19)

where Xit and αi are defined after 4.9 and now vit = ηit −t−1 it+j α2i +t−2 it+j −1 α2i ρ1i . Equation 4.19 is nonlinear in coefficients because of the introduction of ρ1i . Again, Xit can in general include more than one expectational variable and more than one value of j per variable. Given a set of first stage regressors, equation 4.19 can be estimated by 2SLS. The estimates are obtained by minimizing Si = vi Zi (Zi Zi )−1 Zi vi = vi Di vi (4.20)

4.20 is just 4.3 rewritten for the error term in 4.19. A necessary condition for consistency is that Zit and vit be uncorrelated, which means that Zit must be uncorrelated with ηit , t−1 it+j , and t−2 it+j −1 . In order to insure that Zit and t−2 it+j −1 are uncorrelated, Zit must not include any variables that are not known as of the beginning of period t − 1. This is an important additional restriction in the autoregressive case.4 In the general nonlinear case 4.20 (or 4.3) can be minimized using a general purpose optimization algorithm. In the particular case considered here, however, a simple iterative procedure can be used, where one iterates between
is a possibly confusing statement in Cumby, Huizinga, and Obstfeld (1983), p. 341, regarding the movement of the instrument set backward in time. The instrument set must be moved backward in time as the order of the autoregressive process increases. It need not be moved backward as the order of the moving average process increases due to an increase in j .
4 There

70

4 ESTIMATING AND TESTING SINGLE EQUATIONS

estimates of αi and ρ1i . Minimizing vi Di vi with respect to αi and ρ1i results in the following first order conditions: αi = [(Xi −Xi−1 ρ1i ) Di (Xi −Xi−1 ρ1i )]−1 (Xi −Xi−1 ρ1i ) Di (yi −yi−1 ρ1i ) ˆ ˆ ˆ ˆ ˆ (4.21) (yi−1 − Xi−1 αi ) Di (yi − Xi αi ) ˆ ˆ ρ1i = ˆ (4.22) (yi−1 − Xi−1 αi ) Di (yi−1 − Xi−1 αi ) ˆ ˆ where the −1 subscript denotes the vector or matrix of observations lagged one period. Equations 4.21 and 4.22 can easily be solved iteratively. Given the estimates αi and ρ1i that solve 4.21 and 4.22, one can compute the 2SLS ˆ ˆ estimate of vi , which is vi = yi − yi−1 ρ1i − Xi αi + Xi−1 αi ρ1i ˆ ˆ ˆ ˆ ˆ (4.23)

Regarding Hansen’s estimator, given vi , one can compute Mi in one of the ˆ number of possible ways. These calculations simply involve vi and Zi . Given ˆ Mi , Hansen’s estimates of αi and ρ1i are obtained by minimizing5 SSi = vi Zi Mi−1 Zi vi = vi Ci vi (4.24)

Minimizing 4.24 with respect to αi and ρ1i results in the first order conditions 4.21 and 4.22 with Ci replacing Di . The estimated covariance matrix is T (Gi Ci Gi )−1 (4.25)

where G = (Xi − Xi−1 ρ1i yi−1 − Xi−1 αi ). ˆ ˆ To summarize, Hansen’s method in the case of a first order autoregressive structural error consists of: 1) choosing Zit so that it does not include any variables not known as of the beginning of period t − 1, 2) solving 4.21 and 4.22, 3) computing vi from 4.23, 4) computing Mi in one of the number of ˆ possible ways using vi and Zi , and 5) solving 4.21 and 4.22 with Ci replacing ˆ Di .

4.4

Two Stage Least Absolute Deviations6

Another single equation estimator that is of interest to consider is two stage least absolute deviations (2SLAD). This estimator is used for comparison purposes in Chapter 8. The following is a brief review of it.
5 The estimator that is based on the minimization of 4.24 is also the 2S2SLS estimator of Cumby, Huizinga, and Obstfeld (1983). 6 See Fair (1984), Sections 6.3.6 and 6.5.4, for a more detailed discussion of the two stage least absolute deviations estimator.

4.5 CHI-SQUARE TESTS

71

It is assumed for the 2SLAD estimator that the model in 4.1 can be written: yit = hi (yt , xt , αi ) + uit , (i = 1, . . . , n), (t = 1, . . . , T ) (4.26)

where in the ith equation yit appears only on the left hand side. ˆ Let yi = Di yi and hi = Di hi , where, as above, Di = Zi (Zi Zi )−1 Zi , ˆ where Zi is a matrix of first stage regressors. There are two ways of looking at the 2SLAD estimator. One is that it minimizes
T t=1

ˆ yit − hit ˆ

(4.27)

and the other is that it minimizes
T t=1

ˆ yit − hit

(4.28)

Amemiya (1982) has proposed minimizing
T t=1

ˆ qyit + (1 − q)yit − hit ˆ

(4.29)

where q is chosen ahead of time by the investigator. The estimator that is based on minimizing 4.29 will be called the 2SLAD estimator. For the computational results in Chapter 8, q = .5 has been used. The 2SLAD estimator weights large outliers less than does 2SLS, and so it is less sensitive to these outliers. It is a robust estimator in the sense that its properties are less sensitive to deviations of the distributions of the error terms from normality than are the properties of 2SLS.

4.5

Chi-Square Tests

Many single equation tests are simply of the form of adding a variable or a set of variables to an equation and testing whether the addition is statistically significant. Let Si∗∗ denote the value of the minimand before the addition, let Si∗ denote the value after the addition, and let σii denote the estimated ˆ variance of the error term after the addition. Under fairly general conditions, as discussed in Andrews and Fair (1988), (Si∗∗ − Si∗ )/σii is distributed as χ 2 ˆ with k degrees of freedom, where k is the number of variables added. For the 2SLS estimator the minimand is defined in equation 4.3, i.e., Si = ui Di ui .

72

4 ESTIMATING AND TESTING SINGLE EQUATIONS

For Hansen’s estimator the minimand is defined in equation 4.24, i.e., SSi = vi Ci vi . In this case (SSi∗∗ − SSi∗ )/T is distributed as χ 2 , where T is the number of observations. When performing this test the Mi matrix that is used in the construction of Ci must be the same for both estimates. For the results in Chapter 5, Mi was always estimated using the residuals for the unrestricted case (i.e., using the residuals from the equation with the additions). The following is a list of tests of single equations that can be made by adding various things to the equations and performing χ 2 tests. Dynamic Specification Many macroeconomic equations include the lagged dependent variable and other lagged endogenous variables among the explanatory variables. A test of the dynamic specification of a particular equation is to add further lagged values to the equation and see if they are significant. For equation 4.2, for example, one could add the lagged value of yi if the lagged value is not already included in Xi and the lagged values of the variables in Xi . (If the lagged value of yi is in Xi , then the value of yi lagged twice would be added for the test.) Hendry, Pagan, and Sargan (1984) show that adding these lagged values is quite general in that it encompasses many different types of dynamic specifications. Therefore, adding the lagged values and testing for their significance is a test against a fairly general dynamic specification. Time Trend Long before units roots and cointegration became popular, model builders worried about picking up spurious correlation from common trending variables. One check on whether the correlation might be spurious is to add a time trend to the equation. If adding a time trend to the equation substantially changes some of the coefficient estimates, this is cause for concern. A simple test is to add the time trend to the equation and test if this addition is significant. Serial Correlation of the Error Term As noted in Section 4.1, if the error term in an equation follows an autoregressive process, the equation can be transformed and the coefficients of the autoregressive process can be estimated along with the structural coefficients. Even if, say, a first order process has been assumed and the first order coefficient estimated, it is still of interest to see if there is serial correlation of

4.5 CHI-SQUARE TESTS

73

the (transformed) error term left. This can be done by assuming a more general process for the error term and testing its significance. For the results in Chapters 5 and 6 a fourth order process was used. If the addition of a fourth order process over, say, a first order process results in a significant increase in explanatory power, this is evidence that the serial correlation properties of the error term have not been properly accounted for. Other Explanatory Variables Variables can obviously be added to an equation and their statistical significance tested. This is done for the equations in the next two chapters. If a variable or set of variables that one does not expect from the theory to belong in the equation is significant, this is evidence against the theory. Variables can also be added that others have found to be important explanatory variables in similar contexts. For example, Friedman and Kuttner (1992), (1993) have found the spread between the six month commercial paper rate and the six month Treasury bill rate is significant in explaining real GNP in a vector autoregressive system. If the spread is significant in explaining real GNP, then it should be in explaining some of the components of real GNP. It is thus of interest to add the spread to equations explaining consumption and investment to see if it has independent explanatory power. This is done in the next chapter for some of the equations in the US model.7 Leads8 Adding values led one or more periods and using Hansen’s method for the estimation is a way of testing the hypothesis that expectations are rational. Consider the example in equation 4.7 above, and consider testing the RE hypothesis against the simpler alternative that Et−1 X2it+j is only a function of X2it and X2it−1 , where both of these variables are assumed to be known at the time the expectation is made. Under the simpler alternative, X2it and X2it−1 are added as explanatory variables to 4.7. Under the RE alternative, X2it+j is added as an explanatory variable, and the equation is estimated using Hansen’s method. A test of the RE hypothesis is thus to add X2it+j to the equation with X2it and X2it−1 included and test the hypothesis that
7 The six month commercial paper rate and the six month Treasury bill rate are not variables in the US model, and they are not presented in Appendix A. The data are available from the Federal Reserve. 8 The material in this subsection is taken from Fair (1993b), Section 3.

74

4 ESTIMATING AND TESTING SINGLE EQUATIONS

the coefficient of X2it+j is zero. The Zit vector used for Hansen’s method should include the predetermined variables in X1it in 4.7—including X2it and X2it−1 —plus other variables assumed to be in the agents’ information sets.9 The test is really whether these other variables matter. If agents do not use more information than that contained in the predetermined variables in X1it in forming their expectation of X2ti+j , then the use of the variables in Zit as first stage regressors for X2it+j adds nothing not already contained in the predetermined variables in X1it . The test of the RE hypothesis is thus to add variable values led one or more periods to an equation with only current and lagged values and estimate the resulting equation using Hansen’s method. If the led values are not significant, this is evidence against the RE hypothesis. It means essentially that the extra variables in Zit do not contribute significantly to the explanatory power of the equation. An implicit assumption behind this test is that Zit contains variables other than the predetermined variables in X1it . If, say, the optimal predictor of X2it+j were solely a function of X2it and X2it−1 , then the above test would not be appropriate. In this case the traditional approach is consistent with the RE hypothesis, and there is nothing to test. The assumption that Zit contains many variables is consistent with the specification of most macroeconometric models, where the implicit reduced form equations for the endogenous variables contain a large number of variables. This assumption is in effect maintained throughout this book. The tests in Chapters 5 and 6 have nothing to say about cases in which there is a very small number of variables in Zit . As an example of the test, consider the wage variable W in the consumption equation 1.4 in Chapter 1. Assume that Wt is known, where t is period 1. The wage variables in equation 1.4 are then Wt , Et−1 Wt+1 , Et−1 Wt+2 , etc. If
9 Remember that X 2it is assumed to be known at the time the expectations are made, which is the reason for treating it as predetermined. In practice, a variable like X2it is sometimes taken to be endogenous, in which case it is not part of the Zit vector. When a variable like X2it is taken to be endogenous, an interesting question is whether one can test the hypothesis that agents know it at the time they make their expectations as opposed to having only a rational expectation of it. It is not possible to test this if the Zit vector used for the 2SLS method is the same vector used by the agents in forming their expectations. It would, however, be possible to test this hypothesis if there were some contemporaneous exogenous variables in the model that agents forming rational expectations do not know at the time they make their forecasts. These variables are appropriate first stage regressors for 2SLS (since they are exogenous), but they are not used by agents. In practice, however, this is likely to be a small difference upon which to base a test, and no attempt is made here to do so. The focus here is on values dated t + 1 and beyond.

4.6 STABILITY TESTS

75

agents use only current and lagged values of W in forming expectations of future values of W , then candidates for explanatory variables are Wt , Wt−1 , Wt−2 , etc. Under the RE hypothesis, on the other hand, agents use Zit in forming their expectations for periods t + 1 and beyond, and candidates for explanatory variables are Wt+1 , Wt+2 , etc., with Hansen’s method used for the estimation. The test is to test for the joint significance of the led values. The test proposed here is quite different from Hendry’s (1988) test of expectational mechanisms. Hendry’s test requires one to postulate the expectation generation process, which is then examined for its constancy across time. If the structural equation that contains the expectations is constant but the expectations equations are not, this refutes the expectations equations. As noted above, for the test proposed here Zit need not contain all the variables used by agents in forming their expectations, and so the test does not require a complete specification of the expectations generation process. The two main requirements are only that Zit be correlated with X2it+j but not with t−1 it+j .

4.6

Stability Tests

One of the most important issues to examine about an equation is whether its coefficients change over time, i.e., whether the structure is stable over time. A common test of structural stability is to pick a date at which the structure is hypothesized to have changed and then test the hypothesis that a change occurred at this date. In the standard linear regression model this is an F test, usually called the Chow test. More general settings are considered in Andrews and Fair (1988). One test in the more general setting is simply the χ 2 test discussed in the previous section, where Si∗∗ is the value of the minimand under the assumption of no structural change and Si∗ is the value of the minimand under the assumption that the change occurred at the specified date. Assume, for example, that the estimation period is from 1 through T and that the specified date of the structural change is T ∗ . Assume also that the equation is estimated by 2SLS. Computing the χ 2 value in this case requires estimating the equation for three (1) periods: 1 through T ∗ , T ∗ + 1 through T , and 1 through T . Let Si be the (2) value of the minimand in 4.3 for the first estimation period, and let Si be (1) (2) the value for the second estimation period. Then Si∗ = Si + Si . Si∗∗ is the value of the minimand in 4.3 that is obtained when the equation is estimated over the full estimation period. When estimating over the full period, the Zi matrix used for the full period must be the union of the matrices used

76

4 ESTIMATING AND TESTING SINGLE EQUATIONS

for the two subperiods in order to make Si∗∗ comparable to Si∗ . This means that for each first stage regressor Qit , two variables must be used in Zi for the full estimation period, one that is equal to Qit for the first subperiod and zero otherwise and one that is equal to Qit for the second subperiod and zero otherwise. If this is done, then the χ 2 value is (Si∗∗ − Si∗ )/σii , where σii is ˆ ˆ equal to the sum of the sums of squared residuals from the first and second estimation periods divided by T − 2ki , where ki is the number of estimated coefficients in the equation. Recently, Andrews and Ploberger (1994) have proposed a class of tests that does not require that the date of the structural change be chosen a priori. Let the overall sample period be 1 through T . The hypothesis tested is that a structural change occurred between observations T1 and T2 , where T1 is an observation close to 1 and T2 is an observation close to T . If the time of the change (if there is one) is completely unknown, Andrews and Ploberger suggest taking T1 very close to 1 and T2 very close to T . This puts little restriction on the time of the change. If, on the other hand, the time of the change is known to lie in a narrower interval, the narrower interval should be used to maximize power. One of the main advantages of the Andrews-Ploberger tests is that they have nontrivial power asymptotically and have been designed to have certain optimality properties. The particular Andrews-Ploberger test used here is easy to compute. The test is carried out as follows: 1. Compute the χ 2 value for the hypothesis that the change occurred at observation T1 . This requires estimating the equation three times— once each for the estimation periods 1 through T1 − 1, T1 through T , and 1 through T . Denote this value as χ 2(1) .10 2. Repeat step 1 for the hypothesis that the change occurred at observation T1 + 1. Denote this χ 2 value as χ 2(2) . Keep doing this through the hypothesis that the change occurred at observation T2 . This results in N = T2 − T1 + 1 χ 2 values being computed—χ 2(1) , . . . , χ 2(N ) . 3. The Andrews-Ploberger test statistic (denoted AP ) is AP = log[(e 2 χ
1 2(1) 1 2(N )

+ . . . + e2χ

)/N]

(4.30)

In words, the AP statistic is a weighted average of the χ 2 values, where there is one χ 2 value for each possible split in the sample period between observations T1 and T2 .
10 This

χ 2 value is computed in the regular way as discussed above.

1 Critical Values for the AP Statistic for λ = 2.43 14.14 12.50 No. This choice yields a value of λ of 2. .47 15.23 11.of coefs.5)/T .73 13.20 12.59 7.4.00 7.00 4.62 12.5)/T and π2 = (T2 − . These values are interpolated from Table I in Andrews and Ploberger (1994).69 5. Andrews and Ploberger’s Table I shows that the critical values are not very sensitive to different values of λ.1 are for just one particular value of λ (namely.55 10.80 6.33 11.65 8.75. Since the AP test is used in the next two chapters.20 Asymptotic critical values for AP are presented in Tables I and II in Andrews and Ploberger (1994).72 9. where π1 = (T1 − . 2. If the AP value is significant.01 3. it will be useful to give a few critical values.75 No.1 presents the 5 percent and 1 percent asymptotic critical values for this value of λ and various values of the number of estimated coefficients in the equations. The critical values depend on the number of coefficients in the equation and on a parameter λ.62 7. For the work in the next chapter the basic sample period is 1954:1–1993:2.75).31 1% 3.37 1% 10. 1 2 3 4 5 6 7 5% 2.36 4.95 5. Table 4.22 9. 8 9 10 11 12 13 14 5% 8. This is likely to give χ one a general idea of where the structural change occurred even though the AP test does not reveal this in any rigorous way.03 11.07 4.6 STABILITY TESTS 77 Table 4. Although the values in Table 4. and for the stability tests T1 was taken to be 1970:1 and T2 was taken to be 1979:4. The above critical values are thus approximately correct for different choices of T1 and T2 than the one made here.01 9.of coefs. where in the present context λ = [π2 (1 − π1 )]/[π1 (1 − π2 )] . it may be of interest to examine the individual 2 values to see where the maximum value occurred.

. but it is less restrictive than a typical macroeconomic equation. . As noted in Chapter 1. society has been the baby boom of the late 1940s and the 1950s and the subsequent falling off of the birth rate in the 1960s. consumption of individual h in period t). and so on through DJht . Let equation i for individual h be: yhit = Xhit αi + β0i + β1i D1ht + . . .31. The procedure is as follows.2. Since 1980 the percentage of prime aged workers has risen sharply as the baby boomers have begun to pass the age of 25. and uhit is the error term. The number of births in the United States rose from 2. Nt ). Because most macroeconomic variables are not disaggregated by age groups. which also assumes that the constant term is the same across individuals.g. .S. Consider equation i in 4. . . The Lucas critique focuses on policy changes. .7 Tests of Age Distribution Effects11 A striking feature of post war U. . Given Xhit . For example.1 million in 1974. Nt is the total number of people in the population in period t. + βJ i DJht + uhit (h = 1. Let D1ht be 1 if individual h is in age group 1 in period t and 0 otherwise.5.78 4 ESTIMATING AND TESTING SINGLE EQUATIONS 4. yhit is allowed to vary across age groups in equation 4.9. one cannot test for age sensitive αi ’s.31 is restrictive because it assumes that αi is the same across all individuals. an equation that is linear in coefficients. suppose that one of the variables in Xhit 11 The material in this section is taken from Fair and Dominguez (1991). This section discusses a procedure for examining the effects of the changes in the U. whereas by 1977 it had fallen to 49. . This birth pattern implies large changes in the percentage of prime age (25–54) people in the working age (16+) population.S. Xhit is a vector of explanatory variables excluding the constant. Divide the population into J age groups. Equation 4.31) where yhit is the value of variable i in period t for individual h (e. population age distribution on macroeconomic equations. In 1952 this percentage was 57.5 million in 1945 to 4. but other possible changes are changes in the age distribution of the population. T ) (4. The constant term in the equation is β0i + βj i for an individual in age group j in period t. (t = 1. . let D2ht be 1 if individual h is in age group 2 in period t and 0 otherwise. αi is a vector of coefficients. an important issue in macroeconomics is whether the coefficients of macroeconomic equations change over time as other things change.2 million in 1961 and then fell back to 3. .

Otherwise. and reinterpret yit . for which data are not generally available. .S. + βJ i pJ t + uit . . Nt . .+βJ i NJ t +uit . . . (All sums are for h = 1.4. yit varies as the age distribution varies. Equation 4. If.32 is divided through by Nt . let yit be the sum of yhit .) Given this notation. One is thus restricted to assuming that age group differences are reflected in different constant terms in equation 4. (t = 1. on the other hand. given Xit .32 in per capita terms can then be written: yit = Xit αi + β0i + β1i p1t + . (t = 1. . . βJ i coefficients in equation 4. If the coefficient of Yht is the same across individuals. . then the term entering the equation is γ11i D1ht Yht + . T ) (4. Let pj t = Nj t /Nt . and uit as being the original values divided by Nt . In the estimation work J below. + γ1J i DJht Yht .32) If equation 4. . . Bureau of the Census. The sum of a variable like D1ht Yht across individuals is the total income of individuals in age group 1. and it can be summed in the manner discussed in the next paragraph.33) A test of whether age distribution matters is simply a test of whether the β1i . the coefficient differs across age groups. Since the sum of pj t across j is one and there is a constant in the equation. . T ) (4. the age group coefficients are restricted to sum to zero: j =1 βj i = 0. summing equation 4. .12 If the coefficients are zero.7 TESTS OF AGE DISTRIBUTION EFFECTS 79 is Yht . . . The data from 12 Stoker (1986) characterizes this test (that all proportion coefficients are zero) as a test of microeconomic linearity or homogeneity (that all marginal reactions of individual agents are identical). the income of individual h in period t. This means that if the distributional variables do not matter. then γ1i Yht enters the equation. a restriction on the βj i coefficients must be imposed for estimation. . then adding them to the equation will not affect the constant term. it is converted into an equation in per capita terms. Series P-25. let Xit be the vector whose elements are the sums of the corresponding elements in Xhit . say γ1i .33 are significantly different from zero. He shows that individual differences or more general behavioral nonlinearities will coincide with the presence of distributional effects in macroeconomic equations. the variables in Xit . Let Nj t be the total number of people in age group j in period t.31 yields: yit = Xit αi +β0i Nt +β1i N1t +. . . . Current Population Reports. . . . .31. one is back to a standard macroeconomic equation. The Age Distribution Data The age distribution data that are used in the next chapter are from the U. and let uit be the sum of uhit . .

are updated yearly using data provided by the National Center for Health Statistics. A second degree polynomial in which the coefficients sum to zero is determined by two coefficients. (j = 1. The “total” population. . . and so on through age group 86. . age group 2 consists of individuals between 1 and 2 years of age. The variables AGE1t and AGE2t are as follows. Constraints on the Age Coefficients Since there are 55 βj i coefficients to estimate. . and 70+. . . . the age variables 55 55 enter equation i as j =1 βj i pj t . 17. where the 55 variables sum to one for a given t. First. . 69. 55 pj t variables (j = 1. . γ1 .34) where γ0 . and so there are two unconstrained coefficients to estimate per equation. and the Immigration and Naturalization Service. Another constraint. in each of 86 age groups. which consists of individuals 85 years old and over. no i subscripts are used for the γ coefficients. The published data are annual (July 1 of each year). Age group 1 consists of individuals less than 1 year old. In terms of the above notation. is that the coefficients lie on a second degree polynomial.35) ease of notation.80 4 ESTIMATING AND TESTING SINGLE EQUATIONS the census surveys. The polynomial constraint is βj i = γ0 + γ1 j + γ2 j 2 . The second degree polynomial constraint allows enough flexibility to see if the prime age groups behave differently from the young and old groups while keeping the number of unconstrained coefficients small. which was used in Fair and Dominguez (1991). The data are estimates of the total population of the United States. .13 The zero sum constraint on the βj i ’s implies that γ0 = − γ1 13 For 1 55 1 55 2 j − γ2 j 55 j =1 55 j =1 (4. and γ2 are coefficients to be determined. . Nt . One constraint is that the coefficients sum to zero. 55) have been constructed. . The two variables that are associated with two unconstrained coefficients will be denoted AGE1t and AGE2t . Because the equations estimated below are quarterly. 55) (4. the Department of Defense. is taken to be the population 16+. some constraints must be imposed on them if there is any hope of obtaining sensible estimates. including armed forces overseas. quarterly population data have been constructed by linearly interpolating between the yearly points. Fifty five age groups are considered: ages 16. . which are taken every ten years. where j =1 βj i = 0.

and γ3 the coefficient of P 66+. AG2 = P 5665 − P 1625. γ2 the coefficient of P 5665. The coefficient of AG1 in an equation is γ1 − γ0 . 26–55. .36) and AGE2t = j 2 pj t − j =1 1 55 2 55 ( j )( pj t ) 55 j =1 j =1 (4.7 TESTS OF AGE DISTRIBUTION EFFECTS The way in which the age variables enter the estimated equation is then γ1 AGE1t + γ2 AGE2t where AGE1t = j =1 55 81 jpj t − 55 55 1 55 ( j )( pj t ) 55 j =1 j =1 (4. γ1 the coefficient of P 2655. The disadvantage of the present approach over the quadratic approach is that the coefficients are not allowed to change within the four age ranges. The summation constraint can be imposed by entering three variables in the estimated equation: AG1 = P 2655−P 1625. . AG2. Let γ0 denote the coefficient of P 1625 in the estimated equation. one can calculate the four γ coefficients. 55) rather than coefficients of the lagged values of some variable. where γ0 +γ1 +γ2 +γ3 = 0. the coefficient of AG2 is γ2 − γ0 .4. the 55 βj i coefficients can be computed. and similarly for P 2655. and AG3 = (P 66+) − P 1625. where the coefficients that are constrained are the coefficients of the pj t variables (j = 1. I have found that the quadratic constraint sometimes leads to extreme values of βj i for the very young and very old ages. Imposing the constraints in the manner just described has an advantage over imposing the quadratic constraint of allowing more flexibility in the sense that three unconstrained coefficients are estimated instead of two. Also. One test of whether age distribution matters is thus to add AGE1t and AGE2t to the equation and test if the two variables are jointly significant. . The population 16+ was divided into four groups (16– 25. 56–65. From the estimated coefficients for AG1. and 66+) and it was assumed that the coefficients are the same within each group. this leaves three unconstrained coefficients to estimate. and AG3 are variables in the US model. This technique is simply Almon’s (1965) polynomial distributed lag technique. . Let P 1625 denote the percent of the 16+ population aged 16–25. Given the constraint that the coefficients sum to zero. AG1. and P 66+. and AG3 and the summation constraint.37) Given the estimates of γ1 and γ2 . For the work in the next chapter a different set of constraints was imposed on the βj i coefficients. . AG2. P 5665. and the coefficient of AG3 is γ3 − γ0 .

.

the joint significance of the age distribution variables is examined in the household expenditure and money demand equations. one table per equation except for equations 19 and 29. and the variables are defined in Table A.5 The Stochastic Equations of the US Model 5. The construction of the variables is discussed in Chapter 3. The empirical results for the equations are presented in Tables 5. (There are no tables for equations 19 and 29. The basic tests are 1) adding lagged values.2. and tested in this chapter. 5) adding additional variables. and 6) testing for structural stability.30 in this chapter. 3) adding a time trend.) Each table gives the left hand side variable.1 through 5. This chapter thus uses the theory in Chapter 2 plus additional theory in the specification of the stochastic equations. As discussed 83 . Also. As noted at the beginning of Chapter 3.1 Introduction The stochastic equations of the US model are specified. extra “theorizing” is involved in going from theory like that in Chapter 2 to empirical specifications. The stochastic equations are listed in Table A. 4) adding values led one or more quarters. 2) estimating the equation under the assumption of a fourth order autoregressive process for the error term. and the results of the tests described in Chapter 4. estimated. There are 30 stochastic equations in the US model.3 in Appendix A. the right hand side variables that were chosen for the “final” specification. Remember that “adding lagged values” means adding lagged values of all the variables in the equation (including the left hand side variable if the lagged dependent variable is not an explanatory variable).

If. the equation with the additional lagged values does not always have what one would consider sensible dynamic properties.5. for a discussion of this. If this is done. when an equation fails a test.2 For the most part no attempt is made in what follows to separate the two effects. Section 2. Obviously less confidence should be placed on equations that fail a number of the tests than on those that do not. It will be seen that only a few of the equations pass all the tests. These probabilities are labeled “p-value” in the tables. They are generally highly significant. In many cases. this is a test against a quite general dynamic specification. one possibility is to divide the sample period into two parts and estimate two separate equations.2. Similarly. the tests of the significance 1 Using a value of . It is well known that they can be accounting for either partial adjustment effects or expectational effects and that it is difficult to identify the two effects separately. In the following discussion of the results. for example. the resulting coefficient estimates are not always sensible in terms of what one would expect from theory. . . but this is an important area for future work. the best choice seems to be to stay with the original equation even though it failed the test. A small p-value is evidence against the null hypothesis and thus evidence against the specification of the equation. My experience is that it is hard to find macroeconomic equations that do.01 will be taken as a rejection of the null hypothesis and thus as a rejection of the specification of the equation.05 gives the benefit of the doubt to the equations. a p-value of less than . If an equation does not pass a test. which will lessen over time as sample sizes increase. say. however. but. the hypothesis of structural stability is rejected.84 5 US STOCHASTIC EQUATIONS in Section 4. it is not always clear what should be done. as indicated above. Also presented for each test is the probability that the χ 2 value would be whatever it is if the null hypothesis that the additional variables do not belong in the equation is true. In other words. the equations need all the help they can get. For the autoregressive test. The χ 2 value is presented for each test along with its degrees of freedom. when the additional lagged values are significant. the notation “RHO=4” will be used to denote the fact that a fourth order autoregressive process was used. even after accounting for any autoregressive properties of the error terms. 2 See Fair (1984). the change in the equation that the test results suggest may not produce what seem to be sensible results. My feeling (being optimistic) is that much of this difficulty is due to small sample problems.01 instead of. as discussed in Chapter 4. although.2.1 It will be seen that lagged dependent variables are used as explanatory variables in many of the equations.

denoted p4 . and e e so p4 = 100[(P D/P D−4 ) − 1] and p8 = 100[(P D/P D−8 ). that the inflation expectations variables are not good approximations of actual expectations and that if better expectations variables were used they would be significant. e and p e variables were never significant when As will be seen. some measure of expected future inflation must be used in constructing real interest rate variables. Consider RMA. The same reasoning holds for p4 . In other words. the first possible quarter for the break. was taken to be 1970:1 and T2 . For the AP stability tests. Two measures were used in the following work: the actual e rate of inflation in the past four quarters. If real rates instead of nominal rates matter and if p8 e proxy for actual inflation expectations. . The significance of nominal versus real interest rates was tested as follows. it was generally the case that the χ 2 values monotonically rose to the largest value and monotonically fell after that. the last possible quarter for the break. for a total of 158 observations.1 INTRODUCTION 85 of the led values are tests of the rational expectations hypothesis. The test of real versus nominal rates e was first to estimate the equation with RMA included and then to add p8 and e is a good reestimate. It will also be seen below that the real interest rate does affect nonresidential fixed investment. denoted p8 . Assume that p8 is an adequate proxy for inflation expectations of the household sector. the p4 8 added to the household expenditure equations.5 − 1]. and if nominal interest rates affect household behavior. although the estimated effect is small and may be the result of data mining. then p8 should be significant and have a coefficient estimate that equals (aside from sampling error) the negative of e the coefficient estimate of RMA. was taken to be 1979:4. It could be. of course. A ∗ after the AP value in the tables denotes that the value is significant at the one percent level. the price deflator for domestic sales. The “break” quarter that is presented in the tables for the AP test is the quarter at which the break in the sample period corresponds to the largest χ 2 value. then RMA alone should enter.5. The basic estimation period was 1954:1–1993:2. This is an open question and an area for future research. whereas the nominal interest rate variables were. Although not shown in the tables. T1 . and so the data support the use of nominal over real interest rates. If real e interest rates affect household behavior. the after tax mortgage rate. In testing for the significance of nominal versus real interest rates. which is used in the model as e the long term interest rate facing the household sector. then RMA − p8 should enter the household expenditure equations. The price deflator used for this purpose is P D. a ∗ means that the hypothesis of no break is rejected at the one percent level: the equation fails the stability test. and the actual rate of e inflation (at an annual rate) in the past eight quarters.

and seven for leads +8). .3) F2t = (j 2 − 81)X2it+j j =1 (5. the led values enter the equation as γ1 F1t + γ2 F2t (5.2) where F1t = 8 8 (j − 9)X2it+j j =1 (5. . To see what was done for the third set. the estimation period was taken to be shorter by one. First. . Given this constraint. For the second set the equation is estimated under the assumption of a moving average error of order three. or eight observations. For the second set the values led one through four times were added. .3 that results from this estimation was then used in the estimation of the equation without the led values by Hansen’s method for the same (shorter) estimation period.1) where βj = γ0 + γ1 j + γ2 j 2 . as discussed at the beginning of Section 4. and so on. with the coefficients for each variable constrained to lie on a second degree polynomial with an end point constraint of zero. Then for the third set the term added is 8 βj X2it+j j =1 (5. The Mi matrix discussed in Section 4. (When values led once are added the sample period has to be shorter by one to allow for the led values. For the third set the values led one through eight times were added.5. and for the third set the equation is estimated under the assumption of a moving average error of order seven. four. The χ 2 value is then (SSi∗∗ − SSi∗ )/T . For the first set the values of the relevant variables led once were added. when the values led four times are added the sample period has to be shorter by four. assume that one of the variables for which the led values are used is X2i . It may be helpful to review the exact procedure that was followed for the leads test. The end point constraint of zero implies that γ0 = −9γ1 − 81γ2 . β9 = 0. 8. .86 Tests of the Leads 5 US STOCHASTIC EQUATIONS Three sets of led values were tried per equation. j = 1.) The equation with the led values added was then estimated using Hansen’s GMM estimator under the appropriate assumption about the moving average process of the error term (zero for leads +1.4) There are thus two unconstrained coefficients to estimate for the third set. three for leads +4.

This version is called Version 1. This version is called Version 2. and so these equations are the same for both Versions 1 and 2.5. Autoregressive Errors Each equation was first estimated under the assumption of a first order autoregressive error term. Section 11. and “RHO3” to the third order coefficient. For the most part the current version of the model is quite similar to the previous version. In the notation in the tables “RHO1” refers to the first order coefficient.1 INTRODUCTION 87 The results of adding the led values to the stochastic equations are used in Chapter 11.7 in Appendix A. and in one case (equation 11) a third order process was used. Previous Version of the US Model The previous version of the US model is presented in Fair (1984). price. First Stage Regressors The first stage regressors (FSRs) that were used for each equation are listed in Table A. In one case (equation 4) a second order process was used in the final specification. “RHO2” to the second order coefficient. The particular variables for which led values were used are mentioned in this chapter in the discussion of each equation. The choice of FSRs for large scale models is discussed in Fair (1984). The first consists of the “base” equations in Tables 5. 215–216. with F1t and F2t added).30. Three of the main changes are 1) the use of disposable income in the household expenditure equations instead of the wage. and so this previous material does not have to be read. The second version consists of the equations with the third set of led values added (i.e. to examine the economic significance of the rational expectations assumption in the US model. and the assumption was retained if the estimate of the autoregressive coefficient was significant.6.3 The question asked in Section 11.. . and labor constraint variables 3 This work is an updated version of the material in Section 5 in Fair (1993b). nonlabor income. and this discussion will not be repeated here.6 is: How much difference to the properties of the US model does the addition of the led values make? Two versions of the model are examined. The present discussion of the model is self contained.1–5. For some equations no led values were tried because none seemed appropriate. pp. Chapter 4. which have no led values in them.

The means that. which occurred in two of the four equations. are in per capita but not log form.2 Household Expenditure and Labor Supply Equations The two main decision variables of a household in the theoretical model are consumption and labor supply. The CD and I H H equations. The labor constraint also affects the decisions if it is binding. Consider first the four expenditure equations. In the econometric model the expenditures of the household sector are disaggregated into four types: consumption of services (CS). the interest rate in the CD equation was multiplied by CDA. and 3) the different treatment of the interest payment variables of the firm and federal government sectors (equations 19 and 29). 5. and if the interest rates were entered additively in these equations. the effect of. The variable is expected to have a positive sign. and different functional forms have been used in a few cases. These eight variables are determined by eight estimated equations. it was dropped. In addition. a long term rate. the tax rate.88 5 US STOCHASTIC EQUATIONS separately. The CS and CN equations are in log form per capita. Since this does not seem sensible. the price level. and the level of transfer payments. The household after tax interest rate variables in the model are RSA. and RMA was used in the others. and investment in housing (I H H ). The household wealth variable in the model is AA. the labor force of all others 16+ (L3). a one percentage point change in the interest rate has the same percent change over time in each of the two equations. on the other hand. and the number of people holding more than one job. consumption of durable goods (CD). and if it did not. the interest rate. a few more coefficient constraints have been imposed in the current version. which is a variable constructed from peak to peak interpolations of . The determinants of these variables include the initial value of wealth and the current and expected future values of the wage rate. and the lagged value of this variable was tried in each of the equations. a one percentage point change in the interest rate would have a smaller and smaller percent effect over time on per capita durable consumption and on per capita housing investment as both increase in size over time. and the interest rates were entered additively in these equations. say. consumption of nondurable goods (CN ). the labor force of women 25–54 (L2). RSA was used in the CS equation. a short term rate. Four labor supply variables are used: the labor force of men 25–54 (L1). 2) the use of the age distribution variables. called “moonlighters” (LM). and RMA. say.

and they are taken to be exogenous. The lagged dependent variable and the constant term were included in each of the four expenditure equations. The price deflators for the four expenditure categories are P CS. having both W A and Z in the equation is similar to having a labor income variable in the equation in loose labor markets. Both CDA and I H H A are merely scale variables. there are at least two basic approaches that can be taken in specifying the expenditure equations. correspond to the “Keynesian” case. and labor constraint variables separately to the equations. These interest rate variables are nominal rates. Under the first approach one might add W A/P H . as J J falls relative to J J P ). the after tax nonlabor income variable is Y NL. Loose labor markets. The age distribution variables were tried in the four expenditure equations. the interest rate in the I H H equation was multiplied by I H H A. The after tax nominal wage rate variable is W A. and Z to the equation. Z is zero or nearly zero in tight labor markets (i. and the labor constraint variable. Y NL/P H .2 HOUSEHOLD EQUATIONS 89 CD/P OP . As discussed above. discussed in Chapter 3. price.. The second. Regarding the wage. where Z is large in absolute value. more traditional. P CS/P H . the price deflator for total household expenditures is P H . the insignificant results occurring in the I H H equation. is Z. the e e inflation expectations variables p4 and p8 were added in the testing of the equations to test for real interest rate effects. Y D/P H .e. which is a variable constructed from peak to peak interpolations of I H H /P OP . and they were jointly significant at the five percent level in three of the four. In this case the labor constraint is not binding and Z has no effect or only a small effect in the equation. This approach in effect assumes that labor markets are always loose and that the responses to changes . Since Z is highly correlated with hours paid for in loose labor markets. when J J is equal to or nearly equal to J J P . Z falls and begins to have an effect in the equation. P CD. price. The justification for including Z is the following. and income variables.e. The first is to add the wage. nonlabor income. where J J is the actual ratio of worker hours paid for to the total population and J J P is the potential ratio). They were retained in the three equations in which they were significant. Similarly. As labor markets get looser (i. By construction.. This is the “classical” case. on the other hand. P CN. and the results of these tests are reported below. approach is to replace the above four variables with real disposable personal income. Consider the CS equation. These variables in the model are as follows.5. and PIH.

The dominance of Y D/P H does not necessarily mean that the classical case never holds in practice. An interesting question for future work is whether the classical case can be captured in some other way. The results of doing this in the four expenditure equations generally supported the use of Y D/P H over the other variables.0570) and a long run coefficient of roughly one [.4 The short term interest rate is significant. 4 Since the equation is in log form.90 5 US STOCHASTIC EQUATIONS in labor and nonlabor income are the same. each equation was estimated under the assumption of a first order autoregressive error term. The equation is in real. per capita terms and is in log form. The series for CS is quite smooth. The age variables are jointly significant at the five percent level (but not at the one percent level) according to the χ 2 value. and variables were dropped from the equation if they had coefficient estimates of the wrong expected sign in both the unlagged and lagged cases. Explanatory variables lagged once as well as unlagged were generally tried.9425)]. . Equation 1. The disposable income variable has a small short run coefficient (.1. and most of the explanatory power in equation 1 comes from the lagged dependent variable.991 = . Remember that the coefficient of AG1 is the coefficient for people 26–55 minus the coefficient for those 16–25. Also. This is a change from the version of the model in Fair (1984). where the first approach was used.0570/(1 − . as noted above. One can test whether the data support Y D/P H over the other variables by including all the variables in the equation and examining their significance. Some searching was done in arriving at the “final” equations presented below. What it does suggest is that trying to capture the classical case through the use of Z does not work. It will be seen below that the Z variable does work in the labor supply equations. and the assumption was retained if the estimate of the autoregressive coefficient was significant. these coefficients are elasticities. consumer expenditures: services The results of estimating equation 1 are presented in Table 5. All this searching was done using the 2SLS technique. and so the equations reported below use Y D/P H . CS. where it is picking up “discouraged worker” effects when labor markets are loose.

30 17.1 Equation 1 LHS Variable is log(CS/P OP ) Equation RHS Variable cnst AG1 AG2 AG3 log(CS/P OP )−1 log[Y D/(P OP · P H )] RSA Est.00412 . .99 1.8611 .0000 . p-value = . 2.01 χ 2 (AGE) = 10.5. . This suggests that 5 Remember that for the lags test all the variables in the equation lagged once are added to the equation (except for the age variables).) Equation 1 passes the lags test5 and the RHO=4 test.29 22. which is consistent with the life cycle idea that people in their prime working years spend less relative to their incomes than do others. This means that for equation 1 three variables are added: log(CS/P OP )−2 .47 (df = 3. log[Y D/(P OP · P H )]−1 .0870 -.58 1.1.0000 . log(W A/P H ).999 2. (Remember that an equation will be said to have passed a test if the p-value is greater than . a log(AA/P OP ) . On the other hand. The age coefficient estimates for the CS equation suggest that.2 HOUSEHOLD EQUATIONS Table 5.30 2.2242 . the coefficient of AG2 is the coefficient for people 56–65 minus the coefficient for those 16–25. and the coefficient of AG3 is the coefficient for people 66+ minus the coefficient for those 16–25.0362 .86 0.9425 . and RSA−1 .17 -2. log(P CS/P H ).01.88 -3. log[Y N L/(P OP · P H )]. the equation dramatically fails the T test: the time trend is highly significant when it is added to equation 1.14 29.53 1.82 3 4 1 1 4 2 1 1 5 4 91 p-value .53 25.1299 .60 28. Z.2347 .25 6.0570 -.9362 SE R2 DW . These results thus suggest that the dynamic specification of the equation is fairly accurate. Consider now the test results in Table 5.0692 .93 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 e p8 Othera Spread χ 2 Tests χ 2 df 8.2293 .63 0.49∗ T1 1970:1 T2 1979:4 λ 2.0150) Stability Test: AP 14. −1 Similarly.0009 t-stat.0004 .0106 . other things being equal. people 26–55 spend less than others (the coefficient estimate for AG1 is negative and the other two age coefficient estimates are positive).0000 .75 Break 1971:4 Estimation period is 1954:1–1993:2.92 3.

which compares with the one percent critical value in Chapter 4 (for 7 coefficients) of 9. Both the level and change of the lagged dependent variable are significant in the equation. are the variables that one might use in place of disposable income. the coefficient estimates for the variables are all of the wrong expected sign. and so the version of the equation with these variables added is not sensible. consumer expenditures: nondurables Equation 2 is also in real. which. the age coefficients show that people 26–55 spend less than others. The largest χ 2 value occurred for 1971:4.6 As can be seen. Again. For this test the estimation period began in 1960:2 rather than 1954:1 because data on the spread were only available from 1959:1 on. The AP value is 14. as discussed above.9362. the equation fails the stability test. and so the dynamic specification is more complicated than that of equation 1. None of the other specifications that were tried eliminated this problem. The asset. with a p-value of . For the "spread" test in Table 5.6 to examine the sensitivity of the model’s properties to the use of the led values. and log terms. Equation 2.2. per capita. which is near the beginning of the test period of 1970:1–1979:4. CN . and it is an interesting area for future research. Finally.49. disposable income. the spread values are not close to being significant. are significant. which is evidence against the use of real versus nominal interest rates. The largest χ 2 value was for 8 leads. The results are presented in Table 5.50. an estimation period had to be changed for a test. However. The inflation expectations variables are not significant. This is the equation that is used for Version 2 in Section 11. along with the age variables and the lagged dependent variable. and interest rate variables are significant in this equation. This is thus evidence in favor of the rational expectations assumption.92 5 US STOCHASTIC EQUATIONS the trending nature of the CS series has not been adequately accounted for in the equation. The additional variables (“Other”). Disposable income was the variable for which led values were tried—in the form log[Y D/(P OP · P H )]—and the test results show that leads +4 and leads +8 are highly significant. although not shown in the table. There appears to be too much collinearity among these variables to be able to get sensible estimates. the basic equation was always reestimated for this period in calculating the χ 2 value for the test. 6 Whenever . other things being equal.1 and in the other relevant tables that follow. the current and first three lagged values of the spread between the commercial paper rate and the Treasury bill rate were added to the equation.

51 8. The equation fails the stability test.28∗ T1 1970:1 T2 1979:4 λ 2.15 0.1229 -. log(P CN/P H ).58 -1.78 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 e p8 Othera Spread χ 2 Tests χ 2 df 9.2383 -. per capital terms. nonlabor income.997 1. One of the explanatory variables is the lagged stock of durable goods.02 4 4 1 1 4 2 1 1 4 4 93 p-value .0511 .53 2.68 (df = 3. representing the wage.0019 . price.2 Equation 2 LHS Variable is log(CN/P OP ) Equation RHS Variable cnst AG1 AG2 AG3 log(CN/P OP )−1 log(CN/P OP )−1 log(AA/P OP )−1 log[Y D/(P OP · P H )] RMA SE R2 DW Est. Z.43 17. Let KD ∗∗ denote the stock of durable goods that would be desired if there were no adjustment costs of any kind.0509 .7983 . which compares to the one percent critical value (for 9 coefficients) of 11.93 10.28 11.99 14. and leads +1 and +8 are significant.0000) Stability Test: AP 14.2 HOUSEHOLD EQUATIONS Table 5.4791 1.54 1.0400 χ 2 (AGE) = 44. consumer expenditures: durables Equation 3 is in real. but it fails the RHO=4 test. log[Y N L/(P OP · P H )].6203 .41 -4. Likewise.0031 .17 4.40 0.20. and labor constraint variables are not significant at the one percent level.16 12.00557 . the spread values are not significant at the one percent level. The maximum χ 2 value occurs for 1973:2.0416 .0025 .14 4. and the justification for including this variable is as follows.0016 . -.3364 .75 Break 1973:2 Estimation period is 1954:1–1993:2. The equation passes the lags test and the T test.1374 . -1.04 9.28. The AP value is 14. CD. The additional variables.07 9.17 -3.4067 -. a log(W A/P H ).2841 . If durable consumption is proportional to the stock of durables.8516 . p-value = . The variable for which led values were tried is again disposable income. then the determinants of consumption . The inflation expectations variables are not significant. Equation 3.5.87 t-stat.0154 .

5746. and so γ above is . the depreciation rate.94 5 US STOCHASTIC EQUATIONS can be assumed to be the determinants of KD ∗∗ : KD ∗∗ = f (. The first is an adjustment of the durable stock: KD ∗ − KD−1 = λ(KD ∗∗ − KD−1 ) (5. Otherwise.) (5.5) where the arguments of f are the determinants of consumption.5–5. This implies an adjustment of the durable stock to its desired value of 7. Two types of partial adjustment are then postulated.9) (5. As discussed in Chapter 3. . If equations 5. This is. Given these two values and given the coefficient of the lagged stock variable in Table 5. a minor problem.8 yields: CD = (1 − γ )CD−1 + γ (DELD − λ)KD−1 + γ λf (. then the present per capita treatment of equation 5. however. is equal to . where P OP is population. and equation 5.8) The specification of the two types of adjustment is a way of adding to the durable expenditure equation both the lagged dependent variable and the lagged stock of durables. DELD. All variables must be divided by the same population variable for the definition to hold. The second type of adjustment is an adjustment of gross investment to its desired value: CD − CD−1 = γ (CD ∗ − CD−1 ) Combining equations 5.074.5–5.7 does not hold if the lagged stock is divided by P OP−1 . The same holds for equation 4. .7) where DELD is the depreciation rate. the implied value of λ is .6) where KD ∗ is the stock of durable goods that would be desired if there were no costs of adjusting gross investment. the explanatory variables are the same as they are in the other expenditure equations.0106.7 is merely the same equation for the desired values. . where the current values are divided by P OP and the lagged values are divided by P OP−1 . Given KD ∗ .4 percent per quarter. . .3 of −. By definition CD = KD − (1 − DELD)KD−1 .) (5. desired gross investment in durable goods is CD ∗ = KD ∗ − (1 − DELD)KD−1 (5.4254. and it has been ignored here.049511.7 The disposable income and interest rate variables are significant in Table 5. The only problem with this is that the definition used to justify 5.3.8 are defined in per capita terms. The coefficient of the lagged dependent variable is .9 follows. 7 Note in equation 3 that CD is divided by P OP and CD −1 and KD−1 are divided by P OP−1 .

For example.0063 t-stat. The other variables. The pattern here is thus different than the pattern for service and nondurable consumption.1709 -. .97 20.7377 .25 0. The equation fails the stability test by a wide margin.28 14.01105 . other things being equal. and labor constraint variables.0034 . nonlabor income.0105 .34 18. which are the asset.0850 . I H H .49 15.65 9. KH−1 .2 HOUSEHOLD EQUATIONS Table 5.05 -1.0701 .96 3.4769 . the coefficient of the lagged housing stock variable. W A/P H .5903 .43 3 4 1 1 4 2 1 1 5 4 95 p-value . which adds both the lagged dependent variable and the lagged stock of housing to the housing investment equation.75 Break 1977:1 Estimation period is 1954:1–1993:2. a (AA/P OP ) . The inflation expectations variables are not significant.74 22. but it fails the RHO=4 and T tests.0004 .40 3. equation 3 passes the lags test.5746 -.0002 .0000) Stability Test: AP 28.0650 . price. than the others.993 2.5. are significant at the five percent level but not quite at the one percent level.24 2.93 -3. The variable for which led values were tried is disposable income.57∗ T1 1970:1 T2 1979:4 λ 2. The age coefficients show people 26–55 spending more. -3.3 Equation 3 LHS Variable is CD/P OP Equation RHS Variable cnst AG1 AG2 AG3 (CD/P OP )−1 (KD/P OP )−1 Y D/(P OP · P H ) RMA · CDA SE R2 DW Est. P CD/P H .67 21.0000 . -. p-value = . residential investment—h The same partial adjustment model is used for housing investment than was used above for durable expenditures.0106 . Regarding the tests.2590 -1. The spread values are highly significant.28 -2. −1 The age variables are jointly highly significant.35 21.0000 .14 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 · CDA e p8 · CDA Othera Spread χ 2 Tests χ 2 df 2.12 (df = 3. Equation 4. and the led values are significant.78 6. Z.00 χ 2 (AGE) = 35.0000 . wage. Y N L/(P OP · P H ).

−1 −1 −1 −1 is γ (DELH − λ).8797 .96 5 US STOCHASTIC EQUATIONS Table 5.9869 . and so γ is .00 3.02 11. This is not necessarily what one would expect.8151) Stability Test: AP 3. income.0267 .8493 . The variable .81 7. Given these two values and given the coefficient of the lagged housing stock variable of −.39 0. and it may suggest that the estimated speed for the durable goods stock is too low.85 1. DELH .17 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4−1 · I H H A e p8−1 · I H H A Othera Spread χ 2 Tests χ2 1. 3. the implied value of λ is .4.75 Break 1974:1 Estimation period is 1954:1–1993:2.55 4.0185 . p-value = .180.94 (df = 3.4 Equation 4 LHS Variable is I H H /P OP Equation RHS Variable cnst (I H H /P OP )−1 (KH /P OP )−1 (AA/P OP )−1 Y D/(P OP · P H ) RMA−1 · I H H A RH O1 RH O2 Est. Equation 4 passes the lags.19 0.47 T1 1970:1 T2 1979:4 λ 2. The χ 2 test for the age variables shows that the age variables are not jointly significant.00855 .01 9.8063 .9101 . (P I H /P H ) .06 -4.92 4.5322 -.5349 . and interest rate variables are significant in Table 5.53 7. as are the lagged dependent variable and the lagged housing stock variable. where DELH is the depreciation rate of the housing stock.0026 . The asset.5322. the depreciation rate for the housing stock.006716. [Y N L/(P OP · P H )] .3519 t-stat. is .1001 . The coefficient of the lagged dependent variable is . 1.1124 -.957 1.59 -5.15 2.0809 .61 0. The estimated adjustment speed of the housing stock to its desired value is thus greater than the estimated adjustment speed of the durable goods stock. Z .99 χ 2 (AGE) = 0. This is the reason they were not included in the final specification of the equation.2267 .0603 . and T tests.4678.8334 SE R2 DW .46 df 4 2 1 1 4 2 1 1 4 4 p-value .78 2. RHO=4.0809. The equation is estimated under the assumption of a second order autoregressive error term.6394 .97 0. a (W A/P H ) . As discussed in Chapter 3.

but it is also not significant. and the led values are not significant. as expected.5 has log P H added as an explanatory variable. There is a slight negative trend in the labor force participation of men 25–54 that does not seem to be explained by other variables. Equation 5. this is a rejection of the hypothesis that the 8 In Section II in Fair and Dominguez (1991) the age distribution data discussed above were used to examine some of Easterlin’s (1987) ideas regarding the effects of cohort size on wage rates. the only expenditure equation of the four to do so. Another test reported in Table 5. The coefficient estimate for the time trend is negative and significant. on the after tax wage rate. among other things. and so the time trend was included in the equation. and wealth. The next four equations of the household sector are the labor supply equations. This is an area of future research. In addition. and the lagged dependent variable. . but fails the RHO=4 test. which will now be discussed. a time trend. implying that the income effect dominates the substitution effect for men 25–54. labor force—men 25–548 One would expect from the theory of household behavior for labor supply to depend. Equation 5 explains the labor force participation rate of men 25–54. Equation 4 thus passes all the χ 2 tests.2 HOUSEHOLD EQUATIONS 97 for which led values were tried is disposable income. if the labor constraint is at times binding on households. although the estimate is not significant. L1. however. It is in log form and includes as explanatory variables the real wage (W A/P H ). the price level. The coefficient estimate of the labor constraint variable is positive. This is a test of the use of the real wage in the equation. the labor constraint variable (Z). and the led values are not significant. This was done in the context of specifying equations for L1 and L2. Equation 5 passes the lags test.5. The coefficient estimate for the real wage is negative. The variable for which led values were tried is the real wage [log(W A/P H )]. again the only expenditure equation to do so. one would expect a labor constraint variable like Z to affect labor supply through the discouraged worker effect. and in the present work the age distribution data have not been used in the specification of equations 5 and 6. and the spread values are not close to being significant. the “other” variables are not significant. I now have. The inflation expectations variables are not significant. It also passes the stability test. (for reasons that will not be pursued here) some reservations about the appropriateness of the specifications that were used. I am indebted to Diane Macunovich for helpful discussions in this area. If log P H is significant.

Regarding the tests. and the led values are not significant. the RHO=4 test.63 -1. which is implied by the use of the real wage.35 39. log P H is significant at the five percent level but not the one percent level.7763 -.5667 .98 5 US STOCHASTIC EQUATIONS Table 5. with the maximum χ 2 value occurring for 1970:3.9872). and the lagged dependent variable. Equation 5 fails the stability test. -. Equation 6. The final χ 2 test in Table 5. labor force—women 25–54 Equation 6 explains the labor force participation rate of women 25–54. log[Y N L/(P OP · P H )] .0386 .86 1.59 10. -3. The equation thus . This is contrary to the case for men 25–54.36 5.3401 .22 T1 1970:1 T2 1979:4 λ 2.50 -3.5 Equation 5 LHS Variable is log(L1/P OP 1) Equation RHS Variable cnst log(L1/P OP 1)−1 log(W A/P H ) Z T Est. the equation passes the lags test.0139 -. These variables are not significant at the five percent level.93 3 4 1 4 2 1 2 p-value . a log(AA/P OP ) .11 1.984 2.0001 .30 1.73 Test Lags RH O = 4 Leads +1 Leads +4 Leads +8 log P H Othera χ 2 Tests χ 2 df 3.34∗ Estimation period is 1954:1–1993:2.0000 . L2. The variable for which led values were tried is the real wage (log(W A/P H )).11 15. −1 −1 coefficient of log W A is equal to the negative of the coefficient of log P H . It is also in log form and includes as explanatory variables the real wage. the labor constraint variable. where the income effect dominates. The coefficient estimate for the real wage is positive.75 Break 1970:3 t-stat.2067 . and the T test.0117 .0060 . As can be seen. The coefficient estimate for the labor constraint is positive but not significant.5 has asset and nonlabor income variables added to the equation.00196 .0515 SE R2 DW Stability Test: AP 17.0036 . The coefficient estimate for the lagged dependent variable is quite high (. implying that the substitution effect dominates the income effect for women 25–54.14 6.

828—when log P H is added.0177 .51 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 log P H Othera χ 2 Tests χ 2 df 9. Therefore.9872 .0087.21 2.77 0.999 2.01 11. when log P H is added to the equation. asset and nonlabor income variables were added to the equation.00615 . For a final χ 2 test.15 3 4 1 1 4 2 1 2 p-value .97 7. the real wage constraint was imposed on the equation. but the coefficient estimates were positive.74 2.0008 . −1 −1 does well on these tests.0023 SE R2 DW Stability Test: AP 11. This is.0610 and the coefficient for log P H is −. However.00.3652 . .20∗ . These variables are significant. which compares to the critical one percent value for 4 coefficients of 7. but (although not shown in the table) the coefficient estimates were of the wrong expected sign.75 Break 1973:4 Estimation period is 1954:1–1993:2.29 5.4816 . it is highly significant.5.2 HOUSEHOLD EQUATIONS 99 Table 5.0403 t-stat. The equation fails the stability test. a log(AA/P OP ) . One expects the level of assets and nonlabor income to have a negative effect on participation.43 1.26 192.1251 . even though it is strongly rejected by the data. of course. thus rejecting the real wage constraint. Although not shown in the table.20.0022 .27 12. and it certainly does not seem sensible that in the long run participation rises simply from an overall rise in prices and wages. log[Y N L/(P OP · P H )] .0849 . 1. when log P H is added to the equation. (The coefficient estimate for the lagged dependent variable is noticeably smaller—. the coefficient for log W A is .) It thus appears that it is primarily the nominal wage that is affecting participation.49 2.15 T1 1970:1 T2 1979:4 λ 2.0257 . . contrary to what one expects from most theories.6 Equation 6 LHS Variable is log(L2/P OP 2) Equation RHS Variable cnst log(L2/P OP 2)−1 log(W A/P H ) Z Est. The AP value is 11.2170 .

9180 . −1 Equation 7.0477 . as expected.63 Test Lags RH O = 4 Leads +1 Leads +4 Leads +8 log P H Othera χ 2 Tests χ 2 df 3. number of moonlighters Equation 8 determines the number of moonlighters. the time trend. it is insignificant.0481 Estimation period is 1954:1–1993:2.88 T1 1970:1 T2 1979:4 λ 2.48 14.74 -3.0158 -.0663 -. The equation passes the stability test.80 Est.58 9. labor force—all others 16+ Equation 7 explains the labor force participation rate of all others 16+.8896 . Equation 8.0075 .75 Break 1979:2 t-stat.29 0.11 3. the labor constraint variable. an asset variable.00533 . The variable for which led values were tried is the real wage. The asset variable has a negative coefficient estimate and the labor market tightness variable has a positive one.1153 . and the lagged dependent variable. a log[Y N L/(P OP · P H )] .73 3. L3. and the values led 4 and 8 are significant.5110 .91 4 4 1 4 2 1 1 p-value .100 5 US STOCHASTIC EQUATIONS Table 5.94 2. LM.7 Equation 7 LHS Variable is log(L3/P OP 3) Equation RHS Variable cnst log(L3/P OP 3)−1 log(W A/P H ) Z log(AA/P OP )−1 T SE R2 DW Stability Test: AP 3.0162 . It is also in log form and includes as explanatory variables the real wage. 0.0002 .78 0. It is in log form and includes as explanatory variables the real wage. The “other” variable that is added is the lagged value of nonlabor income.66 24. and so. When log P H is added to the equation. and it is not significant at the one percent level. . the labor constraint variable.58 2. .981 1. like L1. L3 appears to have a negative trend that is not explained by other variables.0057 . Equation 7 passes the lags test and the RHO=4 test.36 -1. The coefficient estimate for the time trend is negative and significant. The coefficient estimate for the real wage is positive.7428 . implying that the substitution effect dominates the income effect for all others 16+.

87 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 log P H Othera χ 2 Tests χ 2 df 4. The coefficient estimate for the real wage is positive.97 0.7920 . a log(AA/P OP ) .05647 .32 .8634 . the equation passes the stability test.8919 . and the T test. and they are not significant at the one percent level. -.92 25.54 0. The variable for which led values were tried is the real wage.10. The “other” variables that were added are the lagged value of wealth and the lagged value of nonlabor income.6380 . and the led values are not significant. -3. This completes the discussion of the household expenditure and labor supply equations.00 2.8 Equation 8 LHS Variable is log(LM/P OP ) Equation RHS Variable cnst log(LM/P OP )−1 log(W A/P H ) Z Est.98 T1 1970:1 T2 1979:4 λ 2. . Equation 8 does brilliantly in the tests.02 0. The coefficient estimate for the labor constraint variable is positive and significant.858 1. and so the real wage constraint is supported.4584 . it is not significant. Finally. the RHO=4 test. When log P H is added to the equation.0403 SE R2 DW Stability Test: AP 3.04 0.47 3.8164 .2 HOUSEHOLD EQUATIONS 101 Table 5.5.2618 . It passes the lags test.42 3 4 1 1 4 2 1 2 p-value .75 Break 1973:2 Estimation period is 1954:1–1993:2.1128 . the fewer are the number of people holding two jobs. suggesting that the substitution effect dominates for moonlighters. although the variable is not significant. log[Y N L/(P OP · P H )] .05 7.62 3.0396 t-stat.0185 1.07 6.2185 . The larger is the labor constraint. −1 −1 and the lagged dependent variable. A summary of some of the general results across the equations is in Section 5.

02 to . the interest rate.10 to .3 Money Demand Equations9 In the theoretical model a household’s demand for money depends on the level of transactions. If. the age distribution variables have been added to the household money demand equation. The key question about the dynamics is whether the adjustment of actual to desired values is in nominal or real terms.11. Assume that the equation determining desired money balances is in log form and write log(Mt∗ /Pt ) = α + β log yt + γ rt (5.09. The use of the age distribution variables in equation 9 is taken from Fair and Dominguez (1991).10.02 rather than . and a demand for currency equation. and in the empirical work the demand for money has simply been taken to be a function of the interest rate and a transactions variable.91 to −3. 9 The . the interest rate rises from . One does not necessarily expect a one percentage point rise in the interest rate to have four times the effect on the log of desired money holdings when the change is from a base of . this may pick up a wage rate effect. a change of . High wage rate households spend less time taking care of money holdings than do low wage rate households and thus on average hold more money. one for the firm sector. the interest rate rises from . In practice the results of estimating money demand equations do not seem to be very sensitive to whether the level or the log of the interest rate is used. let yt denote a measure of real transactions. the material in this section on the test of real versus nominal adjustment is taken from Fair (1987). If the adjustment of actual to desired money holdings is in real terms. for example. For the work in this book the level of the interest rate has been used. and it may not be sensible to take the log of the interest rate. Before presenting these equations it will be useful to discuss how the dynamics were handled. the log of the rate rises from −2.51. If. a change of only . and let rt denote a short term interest rate. and.30 to −2. Let Mt∗ /Pt denote the desired level of real money balances. The model contains three demand for money equations: one for the household sector.102 5 US STOCHASTIC EQUATIONS 5. However.40. on the other hand.21. and the household’s wage rate. Interest rates can at times be quite low.03.10) Note that the log form has not been used for the interest rate. as discussed below. With aggregate data it is not possible to estimate this wage rate effect on the demand for money. the log of the rate rises from −3.

which is the nominal adjustment specification.11 yields log(Mt /Pt ) = λα + λβ log yt + λγ rt + (1 − λ) log(Mt−1 /Pt−1 ) + Combining 5. If the real adjustment specification is correct. the adjustment equation is log Mt − log Mt−1 = λ(log Mt∗ − log Mt−1 ) + µt Combining 5. Disposable income is used as the transactions variable. The nominal adjustment hypothesis is also supported in Chapter 6. 17. but I have found that this is rarely the case.14. Three equations were estimated for the United States (versions of equations 9. The nominal adjustment specification has thus been used in the model. the nominal adjustment hypothesis dominates in 13. It is in per capita terms and is in log form.12) t (5. Mt−1 is divided by Pt−1 .12 yields log(Mt /Pt ) = λα + λβ log yt + λγ rt + (1 − λ) log(Mt−1 /Pt ) + µt (5. whereas in 5. This test may.10 and 5. Of the 19 countries for which the demand for money equation (equation 6) is estimated. and 26) and one for each of the other 26 countries.13 and 5. (5.10 and 5. Of the 29 estimated equations.5. the real adjustment dominated for 3. and the after tax three month Treasury bill rate is used 10 The nominal adjustment hypothesis is also supported in Fair (1987). demand deposits and currency—h Equation 9 is the demand for money equation of the household sector. the nominal adjustment dominated for 25. of course.13. and vice versa if the nominal adjustment specification is correct. This test was performed on the three demand for money equations. In 5. which is the real adjustment specification. log(Mt−1 /Pt−1 ) should be significant and log(Mt−1 /Pt ) should not.3 MONEY DEMAND EQUATIONS adjustment equation is 103 log(Mt /Pt ) − log(Mt−1 /Pt−1 ) = λ[log(Mt∗ /Pt ) − log(Mt−1 /Pt−1 )] + t (5.13) . and there was 1 tie. and in each case the nominal adjustment specification won. be inconclusive in that both terms may be significant or insignificant. A test of the two hypotheses is simply to put both lagged money variables in the equation and see which one dominates.11) If the adjustment is in nominal terms.14) Equations 5.14 differ in the lagged money term. where demand for money equations were estimated for 27 countries. MH .10 Equation 9. Mt−1 is divided by Pt .

8164 .17 -3.65 -1. 11 A .0796.6533 -.0002 .11 The test results show that the lagged dependent variable that pertains to the real adjustment specification—log[MH /(P OP · P H )]−1 —is insignificant.0 − . The sign pattern was as expected. -1. p-value = . this supports the nominal adjustment hypothesis. As discussed above.23 Test MH log( P OP ·P H )−1 Lags RH O = 4 T χ 2 Tests χ2 0.72 -3. The age variables are not. Equation 9 passes the lags and T tests. Table 3. other things being equal.66 0.104 5 US STOCHASTIC EQUATIONS Table 5.67∗ T1 1970:1 T2 1979:4 λ 2.9 Equation 9 LHS Variable is log[MH /(P OP · P H )] Equation RHS Variable cnst AG1 AG2 AG3 MH log P OP −1 H −1 ·P log[Y D/(P OP · P H )] RSA RH O1 SE R2 DW Est.22 22. but the χ 2 value of 4. and the long run elasticity is .87 (df = 3. The coefficients for the age variables show that people 26–55 hold more money. than do others.92 was less than the five percent critical value.59 1.75 Break 1975:3 Estimation period is 1954:1–1993:2.1159 . The age distribution variables are included in the equation to pick up possible differences in the demand for money by age.06 df 1 3 3 1 p-value . For the stability test the largest χ 2 value occurred for 1975:3. similar result was obtained in Fair and Dominguez (1991).767 = . jointly significant at the five percent level. which is as expected if people 26– 55 have on average higher wage rates than the others. The short run income elasticity of the demand for money in Table 5.88 1.7462 .78 -0. but it fails the RHO=4 and stability tests.2929 .2677 .5728 -.8962 .94 t-stat.8141 χ 2 (AGE) = 3.0796/(1. -. The equation is estimated under the assumption of a first order autoregressive error term. however.0796 -.8962).05 5.0035 -. as the interest rate.9 is .2763) Stability Test: AP 16.91 19.02318 .902 1. and so not much confidence should be placed on this result.

0000 .3 MONEY DEMAND EQUATIONS 105 Table 5.0073 .79 3. The interest rate variable used is RSA.17.0552 -. and the interest rate variable is the after tax three month Treasury bill rate.75 Break 1974:4 t-stat.26.956 2. . and the stability test. The results are presented in Table 5.5. It is in per capita terms and is in log form. and the equation is estimated under the assumption of a first order autoregressive error term.41 df 1 3 4 1 p-value . The equation passes the lags test. The test results show that the lagged dependent variable that pertains to the real adjustment specification—log[CU R/(P OP · P F )]−1 —is insignificant at the one percent level.0648 Estimation period is 1954:1–1993:2. The transactions variable is the level of nonfarm firm sales.57 7. not the personal tax rates used for RSA in equation 9. the test results show that the lagged dependent variable that pertains to the real adjustment specification [log(MF /P F )−1 ] is insignificant.0518 . Again. X − F A. . The equation passes all the tests.4511 . the T test.00 -2.03776 . Equation 17. The tax rates used in this equation are the corporate tax rates.1784 .42 Test log(MF /P F )−1 Lags RH O = 4 T χ 2 Tests χ2 0. 2. All the variables in the equation are significant. The results for this equation are presented in Table 5. The equation is in log form. MF . Equation 26.01 30.77 Est. All the variables in the equation are significant.74 25. currency held outside banks Equation 26 is the demand for currency equation.9038 .21 T1 1970:1 T2 1979:4 λ 2.17 Equation 17 LHS Variable is log(MF /P F ) Equation RHS Variable cnst log(MF−1 /P F ) log(X − FA) RS(1 − D2G − D2S) SE R2 DW Stability Test: AP 1.89 3. D2G and D2S. The transactions variable that is used is the level of nonfarm firm sales. CU R. It fails the RHO=4 test. demand deposits and currency—f Equation 17 is the demand for money equation of the firm sector.

10 -4. 6) the average number of overtime hours paid per job.75 Break 1977:3 Estimation period is 1954:1–1993:2.1702 . and 5) the firm’s expectations of other firms’ future price and wage decisions. 2) production.0499 -. H O.106 5 US STOCHASTIC EQUATIONS Table 5. I KF .91 CU R−1 P OP−1 ·P F Est. and 7) the wage rate of the firm sector. H F .4 The Main Firm Sector Equations In the maximization problem of a firm in the theoretical model there are five main decision variables: the firm’s price. excess labor.32 Test CU R log( P OP ·P F )−1 χ 2 Tests χ2 6.25 5. -7. W F . P F . and wage rate.9572 .95 -2. 3) the current and expected future demand schedules for the firm’s output. 5) the average number of hours paid per job. and these are the main stochastic equations of the firm sector. demand for employment. 4) the current and expected future supply schedules of labor facing the firm.0009 -. J F .67 86.4122 Lags RH O = 4 T T2 1979:4 λ 2.67 df 1 3 3 1 p-value . 5.0529 . production. 3) investment in nonresidential plant and equipment.26 Equation 26 LHS Variable is log[CU R/(P OP · P F )] Equation RHS Variable cnst log log[(X − F A)/P OP ] RSA RH O1 SE R2 DW Stability Test: AP 5.21 7. 4) the number of jobs in the firm sector.0124 . Each of these variables is determined by a stochastic equation. and inventories. In the econometric model seven variables were chosen to represent the five decisions: 1) the price level of the firm sector. and a number of approximations have to . The variables that affect this solution include 1) the initial stocks of excess capital.02 6. -.3262 .02 T1 1970:1 t-stat.989 2. investment.06 0. Moving from the theoretical model of firm behavior to the econometric specifications is not straightforward. Y .1085 .00966 . 2) the current and expected future values of the interest rate. These five decision variables are determined jointly in that they are the result of solving one maximization problem.

the optimal wage path is chosen.5. In this way of looking at the problem. the natural decision variables of a firm would seem to be the levels of prices and wages. This path implies a certain expected sales path. For example. not the price change. and the wage rate and import price variables are meant to pick up cost effects. the demand pressure variable approaches minus infinity.04. An important feature of the price equation is that the price level is explained by the equation. the price of imports. the market share equations in the theoretical model have a firm’s market share as a function of the ratio of the firm’s price to the average price of other firms. Given the optimal production path. Equation 10. The sequence is from the price decision. and a demand pressure variable.) This functional form implies that as actual output approaches four percent more than potential.10. The equation is in log form. This functional form effectively prevents actual output from ever exceeding potential output by more than four percent. P F . the firm first chooses its optimal price path. the optimal paths of investment and employment are chosen. Given the theory outlined in Chapter 2. This treatment is contrary to the standard Phillips-curve treatment. The equation is estimated under the assumption of a first order autoregressive error term. given the optimal employment path. These are price levels. and to the wage rate decision. One of the key approximations is that the econometric specifications in effect assume that the five decisions of a firm are made sequentially rather than jointly. and the results for this equation are in Table 5. to the investment and employment decisions. to the production decision.4 THE MAIN FIRM SECTOR EQUATIONS 107 be made. where the price (or wage) change is explained by the equation. The demand pressure variable is lagged one quarter in equation 10 because this gave slightly better results than did the use of the variable unlagged. and the objective of the firm is to choose the price level path (along with the paths of the other decision variables) that maximizes the 12 The material on the level versus change specification in this section is taken from Fair (1993a).04]. the wage rate inclusive of the employer social security tax rate. The lagged price level is meant to pick up expectational effects. Finally. log[(Y S−1 − Y−1 )/Y S−1 + . from which the optimal production path is chosen. which implies that the price level approaches plus infinity. price deflator for X − F A12 Equation 10 is the key price equation in the model. is the log of the percentage gap between potential and actual output plus . . The price level is a function of the lagged price level. (Remember that Y S is the potential value of Y . The demand pressure variable.

Fortunately. and log P I M −1 −1 −1 added to the equation with log P F on the left hand side and a constant. .8150 . b log P F .1395 . a log[W F (1 + D5G)] −1 and log P I M−1 added to the equation.30 df 2 4 3 1 1 4 2 1 1 3 p-value . log[W F (1 + D5G)].2253 .0361 -. If β1 in 5.15 and 5.04 0.0885 .68 6. The price equation in level form is p = β0 + β1 p−1 + β2 w + β3 D and the equation in change form is p = η 0 + η1 w + η 2 D (5.83 Test Level Lags RH O = 4 T Leads +1 Leads +4 Leads +8 U R−1 ( Y S−Y )−1 YS Change formb forma χ 2 Tests χ2 3. 155.94 8.46 6.16) (5.1418 t-stat. it is possible to test whether the price level or price change specification is better.78 1.0294 .1719 .09 10.0000 SE R2 DW Stability Test: AP 12. In 5. multiperiod objective function.12 -3.108 5 US STOCHASTIC EQUATIONS Table 5. and log[((Y S − Y )/Y S)−1 + . log[W F (1 + D5G)] .00400 . log P I M. and let D denote the level of some demand pressure variable.67 92.04] on the right hand side (with the RHO1 assumption). A firm decides what its price level should be relative to the price levels of other firms.8499 .39 2.16.15) The key difference between 5.15 is less than one. a permanent change in D results in a permanent change in the change . on the other hand.04] YS RH O1 Est.999 1.98 17.16 is that D and not D is in 5.06 6.0137 .67 0. a permanent change in D results in a permanent change in the level of P but not in the change in P .97 12.0051 .9194 .96 T1 1970:1 T2 1979:4 λ 2.1142 . let w denote the log of the wage rate.75 Break 1972:2 Estimation period is 1954:1–1993:2.0098 .10 Equation 10 LHS Variable is log P F Equation RHS Variable log P F−1 log[W F (1 + D5G)] cnst log P I M log[( Y S−Y )−1 + .28∗ .61 7.0000 . Let p denote the log of the price level.16.

. This equation is p = δ0 + δ1 p−1 + δ2 w + δ3 w−1 + δ4 D (5. as will be seen. then neither specification is supported by the data. Also. The change form is thus strongly rejected. When the lagged values of the price level. these two variables are not significant. the lagged values of the wage rate and price of imports—log[W F (1 + D5G)]−1 and log P I M−1 —are added.16 are δ1 = 1 and δ2 = −δ3 . is not supported.15 into 5. the result that the level specification is supported over the change specification has important implications for the long run properties of the economy.16 or vice versa. on the other hand. The next test shows that the simple percentage gap variable 13 The level versus change specification was also tested in Fair (1993a) for 40 disaggregate price equations. As can be seen. The test results next show that the unemployment rate lagged once is significant when added to the equation. If neither set is accepted. but they can each be nested in a more general equation. with a χ 2 value of 92.17 implied by the level specification in 5. These restrictions can be tested. and the T test.10 show that all the explanatory variables are significant. The constant term η0 in 5. The test results are as follows. the RHO=4 test. The restrictions in 5. to test the level specification. If both sets of restrictions are accepted.17) The restriction in 5. the wage rate. It is not possible to nest 5. one specification will be selected over the other. as can be seen in the last χ 2 test in Table 5. The change specification. but overall slightly favored the level specification. and the led values are not significant.13 Equation 10 passes the lags test. Although not shown in the table. As discussed later in this chapter. and the price of imports are added to the change form of the price equation. the addition of the unemployment rate makes the demand pressure variable insignificant. and so the level specification is supported over the more general specification. The variable for which led values were tried is the wage rate. The results were somewhat mixed. The disaggregate results thus provide some support for the current aggregate estimates. The results in Table 5. Otherwise.10.5.15 is δ3 = 0.16 accounts for any trend in the level of P not captured by the other variables. then the test has not discriminated between the two specifications.4 THE MAIN FIRM SECTOR EQUATIONS 109 in P .30.15 above except that the price of imports is also an explanatory variable. they are highly significant. This is a variable like w in that it is assumed to pick up cost effects. the results in Chapter 6 of estimating price equations for different countries strongly support the level specification over the change specification. Equation 10 in the model is like equation 5. First.17 implied by the change specification in 5.

the addition of this variable also makes the demand pressure variable insignificant. [(Y S − Y )/Y S]−1 .. it would produce according to the following equation (assuming that sales for the current period are known): Y = X + βX − V−1 (5. then knows what its sales for the current period will be. and costs of having the stock of inventories deviate from some proportion of sales. In the theoretical model production is smoothed relative to sales. The equation is based on the assumption that the firm sector first sets it price. V − V−1 = Y − X. Although not shown in the table. near the beginning of the first oil price shock. which include costs of changing employment. and it suggests that any policy experiments with the model that push the economy to very high levels of activity should be interpreted with considerable caution. the price equation fails the stability test. best fit) are obtained when no nonlinearity is introduced. of course. Y . This problem is. and from this latter information decides on what its production for the current period will be. The way that this problem has been handled in the model is simply to use the nonlinear functional form described above even though simpler forms do somewhat better. and so they are of no help for this problem. This situation is unsatisfactory from a theoretical perspective in that one expects that there is some degree of demand pressure beyond which prices rise faster for a further increase in demand pressure than they did before this degree was reached. Equation 11. costs of changing the capital stock. If a firm were only interested in minimizing inventory costs. . is significant. an important area for future work. The large price increases in the 1970s were primarily cost driven. The largest χ 2 value occurs for 1972:2. The functional form chosen for the demand pressure variable is thus not supported by the data. The reason for this is various costs of adjustment. production—f The specification of the production equation is where the assumption that a firm’s decisions are made sequentially begins to be used. Finally. X is the level of sales.110 5 US STOCHASTIC EQUATIONS lagged once. The problem is that the U. The behavior of prices at very high activity levels has probably not been very accurately estimated.18) where Y is the level of production. and β is the inventory-sales ratio that minimizes inventory costs. economy has not experienced enough periods in which demand pressure was very high to allow one to estimate adequately how prices behave in these extreme periods. Since by definition.g. The best results (e. V−1 is the stock of inventories at the beginning of the period.S.

2926.18 would ensure that V = βX.5. Note also that the spread values are not significant.21 states that actual production partially adjusts to desired production each period. Equation 5.18 because its optimal production plan just resulted.14 The hypothesis that firms have rational expectations regarding future values of sales is rejected.22) Equation 11 in Table 5.22.21) where ∗ denotes a desired value. The implied value of λ is . which means that actual production adjusts 70. The variable for which led values were used is the level of sales. In the theoretical model there was no need to postulate explicitly how a firm’s production plan deviated from 5. along with the other optimal paths.7074 = 1.4727 = . For the empirical work. it is generally not optimal for a firm to produce according to 5. . which means that desired production is equal to sales plus 47.0 − .27 percent of the desired change in inventories. however.20 states that the desired level of production is equal to sales plus some fraction of the difference between the desired stock of inventories and the stock on hand at the end of the previous period. The equation is estimated under the assumption of a third order autoregressive process of the error term. The implied value of α is . from the direct solution of its maximization problem.7074.18.11 is the estimated version of equation 5. Because of the other adjustment costs.20) (5. it is necessary to make further assumptions. which means that the desired stock of inventories is estimated to equal 76.4 THE MAIN FIRM SECTOR EQUATIONS 111 producing according to 5.29 percent of the (quarterly) level of sales. Combining the three equations yields Y = (1 − λ)Y−1 + λ(1 + αβ)X − λαV−1 (5.7629. The estimates of equation 11 are consistent with the view that firms smooth production relative to sales. Equation 5. The implied value of β is .74 percent of the way to desired production in the current quarter. The estimated production equation is based on the following three assumptions: V ∗ = βX (5. Equation 5.3344/. The view that production is smoothed relative to 14 Collinearity problems prevented Leads +4 from being calculated for equation 11.19) Y ∗ = X + α(V ∗ − V−1 ) Y − Y−1 = λ(Y − Y−1 ) ∗ (5. X.19 states that the desired stock of inventories is proportional to current sales. and it is interesting that the led values are not significant. Equation 11 passes all of the tests except the stability test.

5392 .3344 .6189 .3906 .25 Test Lags RH O = 4 T Leads +1 Leads +8 Spread p-value .2585 3. In the theoretical model there was no need to postulate explicitly how investment deviates from this amount. because of costs of changing the capital stock. In the theoretical model.0323 .23) . nonresidential fixed investment—f Equation 12 explains nonresidential fixed investment of the firm sector.41 3.2926 .75 Break 1979:3 χ 2 Tests χ 2 df 1.9625 -.69 3.11 Equation 11 LHS Variable is Y Equation RHS Variable Est.98 T1 1970:1 T2 1979:4 λ 2.84 -8.05348 . It is based on the assumption that the production decision has already been made.25 0. I KF .0418 . 1.2442 Estimation period is 1954:1–1993:2. sales has been challenged by Blinder (1981) and others.999 1. and this work has in turn been challenged in Fair (1989) as being based on faulty data.49 5.4753 .45 2 1 1 1 2 4 t-stat. The estimated investment equation is based on the following three equations: (KK − KK−1 )∗ = α0 (KK−1 − KKMI N−1 ) + α1 Y + α2 Y−1 + α3 Y−2 + α4 Y−3 + α5 Y−4 (5. The results in Fair (1989) using physical units data for specific industries suggests that production is smoothed relative to sales. The results using the physical units data thus provide some support for the current aggregate estimates. cnst Y−1 X V−1 RH O1 RH O2 RH O3 SE R2 DW Stability Test: AP 13. investment each period would merely be the amount needed to have enough capital to produce the output of the period. but for the empirical work this must be done.79∗ 28. If there were no such costs. it may sometimes be optimal for a firm to hold excess capital.2788 .50 19.58 1.06 4.26 4.85 6.24 0. Equation 12.8033 .112 5 US STOCHASTIC EQUATIONS Table 5.

By definition. and multiplying both variables by I KF A is a way of accounting for this.1 that p4 equals 100[(P D/P D−4 ) − 1].4 THE MAIN FIRM SECTOR EQUATIONS I KF ∗ = (KK − KK−1 )∗ + DELK · KK−1 I KF = λ(I KF ∗ − I KF−1 ) 113 (5. Equation 5. RB(1 − D2G − D2S). Equation 5.25 yields I KF = λα0 (KK−1 − KKMI N−1 ) + λα1 Y + λα2 Y−1 +λα3 Y−2 + λα4 Y−3 + λα5 Y−4 − λ(I KF−1 − DELK · KK−1 ) (5.24 is merely this same equation for the desired values.24 relates desired gross investment to desired net investment. and the real bond rate lagged three quarters. T XCR. RB−3 .25 is a partial adjustment equation relating the desired change in gross investment to the actual change. DELK · KK−1 is the depreciation of the capital stock during period t − 1.25) where again ∗ denotes a desired value.26 with two additions. which is a variable constructed by peak to peak interpolations of I KF . It is meant to approximate cost of adjustment effects. and KKMI N is the minimum amount of capital needed to produce the output of the period. then desired net investment is zero. and 5. . one would expect a given change in T XCR or RB to have an effect on I KF that increases over time.23 states that desired net investment is a function of the amount of excess capital on hand and of five change in output terms. Equation 5.12 are cost of capital variables: an investment tax credit dummy variable. and I KF ∗ is desired gross investment. How can the use of the cost of capital variables be justified? In the theoretical model the cost of capital affects investment by affecting the kinds of machines that are purchased. I KF A is exogenous. I KF is gross investment of the firm sector.23–5. If output has not changed for four periods and if there is no excess capital. I KF = KK − KK−1 + DELK · KK−1 . more capital intensive 15 RB is equal to the after tax bond rate. one 4 e of the measures of inflation expectations.12 is the estimated version of 5. it is merely a scale variable.5. (KK − KK−1 )∗ is desired net investment. Combining 5. The two additional variables in Table 5.24) (5.26) Equation 12 in Table 5. minus p e . Since I KF has a trend and T XCR and RB do not.15 Both of these variables are multiplied by I KF A. Remember from Section 5. If the cost of capital falls. The change in output terms are meant to be proxies for expected future output changes. KK is the capital stock.

52 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 cnst Spread χ 2 Tests χ 2 df 19.0616 .24 0.0396 .76 Est.83 3.6559 . -0. which says that 3.75 Break 1978:2 t-stat. when the cost of capital falls.26) may be warranted.0016 1.1 in Chapter 3. data are not available by types of machines.23863 .00 0.0018 .0396.30 0. The key approximation is the postulation of the production function 3. The implied value of α0 is −. KKMI N may underestimate the desired amount of capital.99 2.20 6. and these are accounted for through the peak to peak interpolations discussed in Chapter 3.99 T1 1970:1 T2 1979:4 λ 2.96 percent to its desired value each quarter.436 1.31 2.14 1.1203 . The estimate of λ in equation 12 is .0013 -.114 5 US STOCHASTIC EQUATIONS Table 5. For the empirical work. For example. -.28 percent of the amount of excess capital on hand is desired to be eliminated each quarter. then adding cost of capital variables to equation 5.0396.0308 . Technical change and changes in the cost of capital relative to the cost of labor affect over time the λ and µ coefficients in the equation.0346 .0328 = −. ¯ the peak to peak interpolation of µ · H . which says that gross investment adjusts 3. as seems quite likely. and at least part of this error may be picked up by adding cost of capital variables to the equation.0515 .59 6. and approximations have to be made.83 1. machines are purchased and investment expenditures increase.23 (and thus equation 5. KKMI N in equation 93 in the model is determined using MU H .0428 .19 5 4 1 1 4 2 1 4 p-value .51 -2.0660 .45 -2.80 3.12 Equation 12 LHS Variable is I KF Equation RHS Variable (KK − KKMI N )−1 I KF−1 − DELK · KK−1 Y Y−1 Y−2 Y−3 Y−4 T XCR · I KFA RB−3 · I KFA SE R2 DW Stability Test: AP 2. If.8792 Estimation period is 1954:1–1993:2. the effects of cost of capital changes on firms’ decisions are not completely captured through the peak to peak interpolations.19 7.1592 . This production function is one of fixed proportions in the short run.04 2.0013 -.1341 .0013/. The change in .7039 .

They are also based on the assumption that the production decision is made first. The tax credit variable is thus not significant. although its coefficient estimate is of the expected sign. and the results bear this out.83.27) ∗ ∗ J F−1 = J H MI N−1 /H F−1 ∗ ¯ H F−1 = H eδt (5. Even here the lag of three quarters for the bond rate seems a little long. Note also in Table 5. and the bond rate is significant. The results may thus be spurious and merely the result of data mining. This is evidence against the hypothesis that firms have rational expectations with respect to future values of output. including the stability test. The estimated employment equation is based on the following three equations: ∗ log J F = α0 log(J F−1 /J F−1 ) + α1 log Y (5. number of jobs—f The employment equation 13 and the hours equation 14 are similar in spirit to the investment equation 12. Were it not for the costs of changing employment. According to equation 5.26 there should be no constant term in the equation. In the theoretical model there was no need to postulate explicitly how employment deviates from this amount. Because of adjustment costs.5.45.29) where J H MI N is the number of worker hours required to produce the output of the period. I have found it very difficult over the years to obtain significant cost of capital effects in equation 12. but it passes all the others. but shorter lags gave poorer results. and it is interesting to see that the future output changes are not significant. but they are retained because it is embarrassing not to have cost of capital effects in the investment equation. the optimal level of employment would merely be the amount needed to produce the output of the period. The tax credit variable has a tstatistic of 0. The spread values are also not significant.12 that the constant term is not significant. which has a t-statistic of 1. it is sometimes optimal in the theoretical model for firms to hold excess labor. The χ 2 test for the addition of the constant term is not significant.28) (5. and the real bond rate lagged three times has a t-statistic of −2. Equation 12 fails the lags test. H F ∗ is the average number of hours per job that the firm would like to be worked if there were no adjustment costs. Equation 13. and the current results are probably the best that I have ever done. and J F ∗ is the number of . but this must be done for the empirical work.4 THE MAIN FIRM SECTOR EQUATIONS 115 output terms have t-statistics greater than or equal to two except for the change lagged twice. The variable used for the led values was the change in output. J F .52.

When testing the equation for structural stability. Combining 5. where the χ 2 value is quite large. Equation 5. The second addition is accounting for what seemed to be a structural break in the mid 1970s.28 defines the desired number of jobs. DD772.” Equation 5.13 is the estimated version of equation 5.29 yields ¯ log J F = α0 log H + α0 log(J F−1 /J H MI N−1 ) + α0 δt + α1 log Y (5.27 states that the change in employment is a function of the amount of excess labor on hand and the change in output (all changes are in logs). the results for equation 13 suggested that the break could be modeled fairly simply. This means that in the latter period 18. Experimenting with various specifications of this equation reveals that it is very fragile with respect to adding lagged values in that adding these values changes the values of the other coefficient estimates substantially and in ways that do not seem sensible. log J F−1 .116 5 US STOCHASTIC EQUATIONS workers the firm would like to employ if there were no adjustment costs. but the others did.0867 for the period before 1977:2 and −. Equation 13 does not pass the lags test. which is simply the required number of worker hours divided by the desired number of hours worked per job. The equation also fails the RHO=4 test.0867 −.30 with two additions.27–5.1843 = −. the change in employment is zero.30) Equation 13 in Table 5. up substantially from the earlier period. This was modeled by creating a dummy variable.43 percent of the amount of excess labor on hand is eliminated each quarter. The variable . the coefficient of the excess labor variable. This was added to pick up dynamic effects that did not seem to be captured by the original specification. with the largest χ 2 value occurring in 1977:2. In particular. The ∗ term log(J F−1 /J F−1 ) in 5. If there is no change in output and if there is no excess labor on hand. there was evidence of a structural break in the middle of the sample period. The first addition is the use of the lagged dependent variable. is −.29 postulates that ¯ the desired number of hours worked is a smoothly trending variable. Equation 5. Contrary to the case for most equations that fail the stability test. the coefficient of the change in output did not appear to change.0976 after that. where H and δ are constants. that is one from 1977:2 on and zero otherwise and adding to the equation all the explanatory variables in the equation (except the change in output) multiplied by DD772 as additional explanatory variables.27 will be referred to as the “amount of excess labor on hand.13 show that the estimate of α0 . The results in Table 5.

31 yields ¯ log H F = (α0 − λ) log H + λ log H F−1 + α0 log(J F−1 /J H MI N−1 ) + (α0 − λ)δt + α1 log Y (5.5418 -. average number of hours paid per job—f The hours equation is based on equations 5.4 THE MAIN FIRM SECTOR EQUATIONS 117 Table 5.29 and the following equation: log Y (5.0427 Estimation period is 1954:1–1993:2.69 -3.31 5 4 1 4 2 p-value . H F .0071 .0348 .07 t-stat.48 -1.56 -2.32) ∗ ∗ log H F = λ log(H F−1 /H F−1 ) + α0 log(J F−1 /J F−1 ) + α1 .18 6.00329 .29.0976 .5775 -. H F fluctuates around a slowly trending level of hours. The other two terms are the amount of excess labor on hand and the current change in output.2634 .28 and 5.03 14. Combining 5. Both of these terms affect the employment decision.755 2.5. unlike J F .46 20.0001 .0867 -.66 -1.07 4. -3. and the values led four quarters (but not one and eight) are significant at the one percent level. for which led values were tried is the change in output variable.0000 . -.75 7. The reason for the inclusion of this term in the hours equation but not in the employment equation is that.0005 .34 Test Lags RH O = 4 Leads +1 Leads +4 Leads +8 χ 2 Tests χ 2 df 38.4233 -.3037 .31 is the (logarithmic) difference between the actual number of hours paid for in the previous period and the desired number.13 Equation 13 LHS Variable is log J F Equation RHS Variable cnst DD772 F log( J HJMI N )−1 F DD772 · log( J HJMI N )−1 log J F−1 DD772 · log J F−1 T DD772 · T log Y SE R2 DW Est.65 -1.28 9.93 3.28. This restriction is captured by the first term in 5.31. Equation 14. and they should also affect the hours decision since the two are closely related. and 5.0001 -.31) The first term on the right hand side of 5. 5.

27 2.32 with the addition of the terms multiplied by DD772 to pick up the structural break. other things being equal. and the led values are not significant.97 t-stat. Equation 15.0001 . and to increase roughly one for one for high values of H F .02 7. Equation 14 passes the lags and RHO=4 tests.1837. The excess labor variable is significant in the equation.466 1. One would expect H O to be close to zero for low values of total hours.14 is the estimated version of 5. which means that. actual hours are adjusted toward desired hours by 18.85 Test Lags RH O = 4 Leads +1 Leads +4 Leads +8 χ 2 Tests χ 2 df 3.7219 .3345 .5394 .14 Equation 14 LHS Variable is log H F Equation RHS Variable cnst DD772 log H F−1 F log( J HJMI N )−1 F DD772 · log( J HJMI N )−1 T DD772 · T log Y RH O1 SE R2 DW Est. The variable for which led values were tried is the change in output variable.0681 .07 3.11 3.2402 .118 5 US STOCHASTIC EQUATIONS Table 5.1694 -.0521 -.03 4.37 percent per quarter. Equation 14 in Table 5.34) (5. 4.32 4 3 1 4 2 p-value . H O. The equation is estimated under the assumption of a first order autoregressive error term. The estimated value of λ is −. H O.1837 -.22 -5.33) .30 -2.3034 -.00257 .1323 .7236 . This is contrary to the case for equation 13.19 -3.10 -5.0002 .40 2. as are the time trend and the change in output. H F .64 2. average number of overtime hours paid per job—f Equation 15 explains overtime hours.1899 Estimation period is 1954:1–1993:2. . where the values led four quarters are significant. An approximation to this relationship is H O = eα1 +α2 H F which in log form is log H O = α1 + α2 H F (5.50 2.

To account for the irregular trend. both H F F and H F F−1 were included in the equation. and then H F − H F S.0476 Estimation period is 1956:1–1993:2. The equation thus seems to be a reasonable approximation to the way that H O is determined.90 Break 1975:3 χ 2 Tests χ 2 df 4.8834 . It is in log form. cnst HFF H F F−1 RH O1 SE R2 DW Stability Test: AP 1.930 1.) H F F is defined by equation 100 in Table A.9159 .5. time. and T tests.0825 .99 4.91 25. It also passes the stability test.40 3. 78. In the final specification. W F was simply taken to be a function of a constant. (The peak quarters used for the interpolation are presented in Table A. W F . Labor market tightness variables like the unemployment rate were . a variable H F S was constructed from peak to peak interpolations of H F . H F has a negative trend over the sample period.0201 .3.34.4 THE MAIN FIRM SECTOR EQUATIONS 119 Table 5. which is denoted H F F in the model. and the first four lagged values of the price level and the wage rate.92 2 3 1 t-stat. The equation is estimated under the assumption of a first order autoregressive error term.76 T1 1970:1 T2 1979:4 λ 2. which appeared to capture the dynamics better.6. The equation passes the lags.15.2611 . average hourly earnings excluding overtime—f Equation 16 is the wage rate equation. H F was detrended before being used in 5.15 Equation 15 LHS Variable is log H O Equation RHS Variable Est.08 4. was included in equation 15. It is the deviation of H F from its peak to peak interpolations. The coefficient estimates are significant in equation 15.00 3. although the trend appears somewhat irregular.04761 . although the estimate of the autoregressive coefficient of the error term is quite high.0122 . Equation 16.83 Test Lags RH O = 4 T p-value . the current value of the price level.31 8. First.34 to equation 15 in Table 5. RHO=4. Two modifications were made in going from equation 5. Second.

since all the equations are estimated together.36) (5. (5. these constraints were imposed by first estimating the price equation to get estimates of β1 and β2 and then using these estimates to impose the constraint on γ3 in the wage equation. and this is the reason for the use .36 that impose this restriction are: γ3 = [β1 /(1 − β2 )](1 − γ2 ) − γ1 γ5 = −γ4 γ7 = −γ6 γ9 = −γ8 When using 2SLS or 2SLAD. . The time trend is added to account for trend changes in the wage rate relative to the price level. Let p = log P F and w = log W F .16.37) which says that the real wage is a function of its own lagged values plus other terms. since it along with some of the lags identifies the price equation. equation 10. Equation 16 is estimated under the assumption of a first order autoregressive error. The real wage in 5. . .120 5 US STOCHASTIC EQUATIONS not significant in the equation. Its inclusion is important. Constraints were imposed on the coefficients in the wage equation to ensure that the determination of the real wage implied by equations 10 and 16 is sensible.35) The implied real wage equation from these two equations should not have w − p as a function of either w or p separately.35 and 5. . The wage rate lagged four times is significant. The constraints on the coefficients in equations 5. The relevant parts of the price and wage equations regarding the constraints are p = β1 p−1 + β2 w + . . w = γ1 w−1 + γ2 p + γ3 p−1 + γ4 w−2 + γ5 p−2 + γ6 w−3 + γ7 p−3 + γ8 w−4 + γ9 p−4 + . since one does not expect the real wage to grow simply because the level of w and p are growing. The desired form of the real wage equation is thus w − p = δ1 (w−1 − p−1 ) + δ2 (w−2 − p−2 ) + δ3 (w−3 − p−3 ) + δ4 (w−4 − p−4 ) + . No sequential procedure is needed to impose the constraints when using 3SLS and FIML. .37 is not a function of the level of w or p separately. The results for equation 16 (using 2SLS) are presented in Table 5. (5.

5. no demand pressure variables were found to be significant in the wage equation.5 Other Firm Sector Equations Equation 18.33 – – – – Test Real Wage Restr. As noted above.1757 .0005 .2444 .2843 -. where DF ∗ . which is picking up a trend in the real wage.05 2.0619 .18 -1. then one can write DF ∗ = π . DF .90 5. If in the long run firms desire to pay out all of their after tax profits in dividends.b Lags RH O = 4 U R−1 df 4 1 3 1 p-value . 5. the equation passes the stability test. The equation also passes the lags and RHO=4 tests. see the discussion in the text.1506 .1757 -. Finally.49 -0.1625 . b Equation estimated with no restrictions on the coefficients.0038 . The time trend is highly significant.00628 .54 1.1180 . and it is not significant. The final χ 2 test in the table has the unemployment rate lagged once added as an explanatory variable. of four lags even though lag 2 is not significant.3269 .6637 . dividends paid—f Let π denote after tax profits.91 T1 1970:1 T2 1979:4 λ 2.9863 Estimation period is 1954:1–1993:2 a Coefficient constrained.92 2.35 7.0038 -.5 OTHER FIRM SECTOR EQUATIONS 121 Table 5.96 2. The χ 2 test results show that the real wage restrictions discussed above are not rejected by the data.34 0.04 1.1506 -.0142 .999 1.13 .5. log W F−1 log P F log W F−2 log W F−3 log W F−4 cnst T RH O1 log P F−1 a log P F−2 a log P F−3 a log P F−4 a SE R2 DW Stability Test: AP 6.75 Break 1972:3 χ 2 Tests χ2 6.16 Equation 16 LHS Variable is log W F Equation RHS Variable Est.00 t-stat.

25 1 2 4 1 1 p-value . The test results also show that the constant term is not significant.39.65 .02616 . 9.29 31. which implies a fairly slow adjustment of actual to desired dividends.2636 SE R2 DW Stability Test: AP 1. The level of after tax profits in the notation of the model is P I EF − T F G − T F S. . then the restriction does not hold.39) (5. the coefficient of log(P I EF − T F G − T F S) is restricted to be the negative of the coefficient of log DF−1 in equation 18. .18 is the estimated version of equation 5. If it is assumed that actual dividends are partially adjusted to desired dividends each period as DF /DF−1 = (DF ∗ /DF−1 )λ then the equation to be estimated is log(DF /DF−1 ) = λ log(π/DF−1 ) (5.48 T1 1970:1 T2 1979:4 λ 2.38) Equation 18 in Table 5.75 Break 1976:1 Estimation period is 1954:1–1993:2.080 1..0058 . The estimate of λ is . a test that γ = 1). The first test in Table 5.30 Test Restriction Lags RH O = 4 T cnst χ 2 Tests χ 2 df 0. and the test is passed.e. Because of the assumption that DF ∗ = π . is the long run desired value of dividends for profit level π.18 Equation 18 LHS Variable is log DF Equation RHS Variable log P I EF −T F G−T F S DF −1 Est.122 5 US STOCHASTIC EQUATIONS Table 5. The above specification does not call for a constant term.0251. If instead DF ∗ = π γ .0251 t-stat.18 is a test of the restriction (i.0000 .5175 . where γ is not equal to one. and this is supported by the data.14 10. The equation fails the lags and RHO=4 tests.7065 .42 1. and it passes the T and stability tests.66 0.

The restriction is also imposed that the sum of the lag coefficients is one. in which CCF is noticeably higher than would be predicted by the equation with only log[(P I K · I KF )/CCF−1 ] in it.2757 .5 OTHER FIRM SECTOR EQUATIONS 123 Table 5.33 2 3 1 p-value .0006 . Equation 20 in Table 5. The use of the lagged dependent variable in the equation is meant to approximate the dependence of capital consumption allowances on past values of investment. The equation is estimated under the assumption of a first order autoregressive error term.5664 Estimation period is 1954:1–1993:2. Equation 21.37 Test Lags RH O = 4 T χ 2 Tests χ 2 df 4.5.75 -3.04 14. including the stability test. Equation 20. and two dummy variables. -1. where P is the price of the good and V is the stock of inventories of the good.1316 .711 2. It passes all but RHO=4 test.2649 -. which means that capital consumption allowances are assumed to be taken on all investment in the long run. This specification implies that the lag structure is geometrically declining.20 Equation 20 LHS Variable is I V A Equation RHS Variable cnst (P X − P X−1 )V−1 RH O1 SE R2 DW Stability Test: AP 3. As an approximation.7687 1.21 is an attempt to approximate this for the firm sector. where P X is the price deflator for the sales of the firm sector. I V A is regressed on a constant and (P X − P X−1 )V−1 . -1. I V A. capital consumption—f In practice capital consumption allowances of a firm depend on tax laws and on current and past values of its investment.75 Break 1974:4 t-stat. the equation seems fairly good.05 17.80039 .03 T1 1970:1 T2 1979:4 λ 2.49 0. P I K · I KF is the current value of investment. D811824 and . There are two periods. inventory valuation adjustment In theory I V A = −(P − P−1 )V−1 . Equation 21 in Table 5. CCF .20 is meant to approximate this.59 Est. 1981:1–1982:4 and 1983:1–1983:4.

0346 .0525 .21 says that capital consumption allowances are taken on 5. a crude procedure. . which the coefficient estimates for the two dummy variables suggest. however.0568(1 − . Tax law changes have effects on CCF that are not captured in the equation. The stability test was not performed because of the use of the dummy variables.89 6. then 5.02 4.271 1. then 5. then one might expect CCF to be lower at some later point (since capital consumption allowances can be taken on only 100 percent of investment in the long run).04 1 2 4 1 1 p-value .21 is a test of the restriction that the sum of the lag coefficients is one.43 3. No attempt. 16.05 percent [. and so on.124 5 US STOCHASTIC EQUATIONS Table 5.0339 .68 percent of new investment in the current quarter. This is done by adding log CCF−1 to the equation. The results of the last χ 2 test in the table show that the constant term is not significant in the equation.21 Equation 21 LHS Variable is log CCF Equation RHS Variable log[(P I K · I KF )/CCF−1 ] D811824 D831834 Est.01505 .0568(1 − .0568)] of this investment in the next quarter.59 Test Restriction Lags RH O = 4 T cnst χ 2 Tests χ 2 df 4. The results show that the restriction is not rejected at the one percent level. and T tests.0568 in Table 5. RHO=4. D831834.0398 .0568 . This is as expected since the above specification does not call for a constant term.1976 .50 3.22 5. The first χ 2 test in Table 5.87 t-stat.15 4. of course. The equation passes the lags.0568)2 ] in the next quarter. Regarding the use of the two dummy variables. but the equation itself is only a rough approximation to the way that capital consumption allowances are actually determined each period. The coefficient estimate of .36 percent [. . if CCF is larger than usual in the two subperiods.0812 SE R2 DW Estimation period is 1954:1–1993:2. This is. have been added to the equation to account for this.0174 . was made to try to account for this in equation 21.

69 0. bond rate.6 Financial Sector Equations The stochastic equations for the financial sector consist of an equation explaining member bank borrowing from the Federal Reserve.06 T1 1970:1 T2 1979:4 λ 2.332 2. Equation 23. mortgage rate The expectations theory of the term structure of interest rates states that long term rates are a function of the current and expected future short term rates. These rates are assumed to be determined according to the expectations theory.22 Equation 22 LHS Variable is BO/BR Equation RHS Variable Est.0039 . bank borrowing from the Fed The variable BO/BR is the ratio of borrowed reserves to total reserves.75 Break 1972:4 χ 2 Tests χ 2 df 6. RM. The coefficient estimates of RS and RD in Table 5.01989 . The equation passes the lags.5. The two long term interest rates in the model are the bond rate.01 26. and a negative function of the discount rate.0000 . and it fails the RHO=4 test.0034 . BO. RB.80 3 4 1 t-stat. 5.22 are positive and negative. as expected. cnst (BO/BR)−1 RS RD SE R2 DW Stability Test: AP 6.19 . where the current and past values of the short term . RM.1111 . This ratio is assumed to be a positive function of the three month Treasury bill rate. and the mortgage rate.3698 Estimation period is 1954:1–1993:2.6 FINANCIAL SECTOR EQUATIONS 125 Table 5.3170 . T .23 2. RB. RS.99 4. Equation 22. two term structure equations. RD. and an equation explaining the change in stock prices.05 -1. and stability tests. The estimated equation also includes a constant term and the lagged dependent variable.37 Test Lags RH O = 4 T p-value . respectively. Equation 24.0062 -. 0.

Equations 23 and 24 are the two estimated equations. (This restriction is imposed by subtracting RS−2 from each of the other interest rates in the equations. a constraint has been imposed on the coefficient estimates. for example.04 2.78 -5.95 8.33 2.47 Test Restriction Lags RH O = 4 T Leads +1 Leads +4 e p4 e p8 χ 2 Tests χ 2 df 1. which implies a fairly complicated lag structure relating each long term rate to the past values of the short term rate.) Equation 23 (but not 24) is estimated under the assumption of a first order autoregressive error term.51 7. .958 2.6939 . RHO=4.2438 .2963 -.0316 . a sustained one percentage point increase in the short term rate eventually results in a one percentage point increase in the long term rate.89 2.23 Equation 23 LHS Variable is RB − RS−2 Equation RHS Variable Est.23 and 5. The lagged dependent variable is used in each of these equations. cnst RB−1 − RS−2 RS − RS−2 RS−1 − RS−2 RH O1 .8813 . The test results show that the coefficient restriction is not rejected for either equation. The short rates are significant except for RS−1 in equation 24.23 3. My experience with these equations over the years is that they are quite stable and reliable. During most periods they provide a very accurate link from short rates to long rates. 5. RS) are used as proxies for expected future values.93 4.04 T1 1970:1 T2 1979:4 λ 2.2180 .25359 . and stability tests. interest rate (the three month bill rate. respectively.24.2856 .0499 SE R2 DW Stability Test: AP 6.75 Break 1979:4 Estimation period is 1954:1–1993:2.85 1 2 3 1 1 4 1 1 p-value .0474 .5658 . Both equations pass the lags. This means that.126 5 US STOCHASTIC EQUATIONS Table 5.34 42.0682 . The sum of the coefficients of the current and lagged values of the short term rate has been constrained to be equal to one minus the coefficient of the lagged long term rate. T .33 3.1688 .62 0.25 .2019 t-stat. The results for both term structure equations are thus strong. In addition. The results for equations 23 and 24 are presented in Tables 5.

Equation 25.65 2.897 2. and the χ 2 tests show that the led values are not significant at the one percent level.20 1.0995 . . are not significant in the equations.6 FINANCIAL SECTOR EQUATIONS 127 Table 5.92 4.1298 cnst RM−1 − RS−2 RS − RS−2 RS−1 − RS−2 SE R2 DW Stability Test: AP 5.16 This is thus at least slight evidence against the bond market having rational expectations with respect to the short term interest rate.12 12. was used as a proxy for changes in expected future interest rates.70 7.34387 . In the theoretical model the aggregate value of stocks is determined as the present discounted value of expected future after tax cash flow.0169 t-stat.77 2. RB. In the empirical work the change in the bond rate.03 33. 6.2322 .8418 .2597 -.02 T1 1970:1 T2 1979:4 λ 2. (CF − T F G − T F S). CG. The theoretical model thus implies that CG should be a function of changes in expected future after tax cash flow and of changes in the current and expected future interest rates. and the change in after tax cash flow.1039 . was used 16 Collinearity problems prevented Leads +8 from being calculated for equation 23.4749 .79 2. capital gains or losses on corporate stocks held by h The variable CG is the change in the market value of stocks held by the household sector.0128 .72 . .29 1 2 4 1 1 4 2 1 1 p-value .71 2.27 Test Restriction Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 e p8 χ 2 Tests χ 2 df 2.1835 . p4 and p8 .24 5. The variable for which led values were tried was the short term interest rate (RS).0273 .75 Break 1979:4 Estimation period is 1954:1–1993:2. The e e test results also show that the inflation expectations variables.3093 .24 Equation 24 LHS Variable is RM − RS−2 RHS Variable Equation Est. the discount rates being the current and expected future short term interest rates.1457 .5.72 -0.

1530 -68.25 is the estimated equation. The variables for which led values were tried are the change in the bond rate and the change in after tax cash flow. The explanatory variables are 1) per capita real disposable income. which provides some link from interest rates to stock prices in the model.63 4.5451 t-stat. 2) the private.9413 .0076 . 23.25 Equation 25 LHS Variable is CG Equation RHS Variable cnst RB (CF − T F G − T F S) Est. and it fails the RHO=4 test.22 89. however. The values led one and four quarters are not significant. I M.1158 1. Imports The import equation is in per capita terms and is in log form. and the cash flow variable is not significant.22 -2. For the final χ 2 test RS. The fit of equation 25 is not very good.0286 . The equation passes the lags.60 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 RS χ 2 Tests χ 2 df 2. T .79 0.147 1.75 Break 1979:1 Estimation period is 1954:1–1993:2.47 0. and stability tests. This is thus slight evidence in favor of there being rational expectations in the stock market. nonfarm price deflator (a price deflator for domestically produced goods) relative to the import price deflator.92 0.0006 . the change in the short term rate.31 13.3942 SE R2 DW Stability Test: AP 2. 3) the long term after tax interest rate lagged one . and it is not significant.73 3 4 1 2 8 4 1 p-value .7 The Import Equation Equation 27. 5.12 9. The change in the bond rate is significant.27602 .5204 . but the values led eight quarters are.128 5 US STOCHASTIC EQUATIONS Table 5.26 19. Equation 25 in Table 5.96 T1 1970:1 T2 1979:4 λ 2. as a proxy for changes in expected future after tax cash flow.3172 . 3. was added under the view that it might also be a proxy for expected future interest rate changes.

and so the restriction is rejected. The coefficient estimate for the long term interest rate is negative as expected.03204 . which is a test of the restriction that the coefficient of log P F is equal to the negative of the coefficient of log P I M. -.94 30. and 5) four dummy variables to account for two dock strikes.0000 . a log(AA/P OP ) .4533 . -4.5.17 22.49 2. The short run income elasticity of imports is .0001 .91 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 e p8 log P F log[(X − F A)/P OP ] Othera Spread χ 2 Tests χ2 9.9171 .77 4. and the long run elasticity is 2.995 1.64 4. both fairly high. 4) the lagged dependent variable.0025 .80 Estimation period is 1954:1–1993:2. Although not shown in the table.47 [.0871 .3172 .1478 -.7 THE IMPORT EQUATION 129 Table 5.1183 . The coefficient estimate for the relative price term is positive as expected. log[Y N L/(P OP · P H )].08 9. Z. and the led values are not significant.00 26.12 0. The variable for which led values were tried is disposable income.55 35.0027 -. when log P F is added to the equation.0277 . Many χ 2 tests were performed for the import equation. but (although not shown in the table) their coefficient estimates are of the wrong expected sign.0365 -.0863 .0000 SE R2 DW .0943 t-stat. −1 quarter.0000 . p4 and p8 .0890 and the coefficient for log P F is .0001 .0000 .44 14. The next test in the table adds log P F to the equation.3127/(1 − .52 -2. the coefficient for log P I M is −. The inflae e tionary expectations variables. although it is not significant.19 1.3172.8716)].27 Equation 27 LHS Variable is log(I M/P OP ) Equation RHS Variable cnst log(I M/P OP )−1 log[Y D/(P OP · P H )] log(P F /P I M) RMA−1 D691 D692 D714 D721 Est.8716 .07 df 4 4 1 1 4 2 1 1 1 1 4 4 p-value .68 2.21 24.99 24.50 -1. The results thus suggest that the level of imports responds .58 10.90 0.0561 .27. log(W A/P H ).1844. but it also is not significant. The results are in Table 5. but fails the RHO=4 and T tests. are highly significant. The log P F variable is highly significant. It passes the lags test.32 -3.4447 .

8 Government Sector Equations There is one stochastic equation for the state and local government sector. The equations for U B and RS are discussed in this section. Therefore. real wage. I NT G.3. Experimenting with the import equation reveals that it does much better in the tests if the relative price restriction is not imposed. but it fails the others. As was the case for equation 6. The “other” variables that were added for the next test. are highly significant. the spread values are highly significant. labor constraint. On the other hand (not shown in the table). The χ 2 values is not significant.4. the equation does much better. one explaining interest payments. U B. There are two stochastic equations for the federal government sector. So in summary. In other words. 5. and nonlabor income variables. Finally. and one explaining the three month Treasury bill rate.77) than the t-statistic on the disposable income variable (0. even though it is strongly rejected by the data. The sales variable thus dominates the disposable income variable in this sense. and the equation for I N T G is discussed in the next section. the import equation passes the lags and leads tests. but (not shown in the table) they all have coefficient estimates of the wrong expected sign. It is clearly an equation in which future research is needed. the relative price constraint was imposed on the equation. the t-statistic on the sales variable was higher (1. The next test adds the level of per capita nonfarm firm sales—log[(X − FA)/P OP ]—to the equation to see if it better explains imports than does disposable income. It does not seem sensible that in the long run the level of imports rises simply from an overall rise in prices. which are asset. explaining unemployment insurance benefits. The equation explaining RS is interpreted as an interest rate reaction function of the Federal Reserve. The stability test was not performed for the import equation because of the use of the dummy variables. .35). RS.130 5 US STOCHASTIC EQUATIONS more to the domestic price deflator than to the import price deflator. when log P F is added. this is contrary to what one expects from theory. The consequences of using the sales variable in place of the disposable income variable are examined in Section 11. and so on this score the sales variable does not have independent explanatory power.

three month Treasury bill rate A key question in any macro model is what one assumes about monetary policy.57 10.25∗ . and it fails the stability test.8164 .5. In one respect.06586 . The inclusion of the nominal wage rate is designed to pick up the effects of increases in wages and prices on legislated benefits per unemployed worker.28 Equation 28 LHS Variable is log U B Equation RHS Variable Est. Equation 30. say. Equation 28.27 3. cnst log U B−1 log U log W F RH O1 SE R2 DW Stability Test: AP 12. unemployment insurance benefits Equation 28 is in log form and contains as explanatory variables the level of unemployment. U B. and T tests. and the lagged dependent variable. Since the Fed is run by a relatively small number of people. and in the empirical work an equation like this is estimated.995 2.71 Test Lags RH O = 4 T p-value .19 T1 1970:1 T2 1979:4 λ 2.2565 . trying to explain the behavior of the household or firm sectors. there can be fairly abrupt changes .68 5.0339 . RHO=4.1751 Estimation period is 1954:1–1993:2.57 14.28 show that the coefficient estimates are significant except for the estimate of the constant term. 0.3459 . RS.8 GOVERNMENT SECTOR EQUATIONS 131 Table 5. The equation is estimated under the assumption of a first order autoregressive error term.2349 1.75 Break 1975:1 χ 2 Tests χ 2 df 8.1168 .1220 . trying to explain Fed behavior is more difficult than. In the theoretical model monetary policy is determined by an interest rate reaction function.84 3 3 1 t-stat. The results in Table 5. The equation passes the lags. This equation is interpreted as an equation explaining the behavior of the Federal Reserve (Fed).89 1.71 4. the nominal wage rate.

Abrupt changes are less likely to happen for the household and firm sectors because of the large number of decision makers in each sector. This was modeled by adding the variable D794823 · P CM1−1 to the equation. The signs of the coefficient estimates in Table 5. 2) the degree of labor market tightness. and money supply growth. 3) the percentage change in real GDP. What seemed to happen between 1979:4 and 1982:3 was that the size of the coefficient of the lagged money supply growth increased substantially. which was between 1979:4 and 1982:3. As just noted. and the percentage change in real GDP. The right hand side variables in this equation are variables that seem likely to affect the target rate. however. and the led values are not significant. Having said this. even this change appears capable of being modeled. . The variables that were chosen are 1) the rate of inflation. The variables for which led values were tried are the inflation variable. and 4) the percentage change in the money supply lagged one quarter. as will be seen. labor market tightness.30 are as expected. The estimated equation also includes the lagged dependent variable and two lagged bill rate changes to pick up the dynamics. The results thus seem good for this equation. the results show that the weight given to money supply growth in the setting of the bill rate target was much greater in the 1979:4–1982:3 period than either before or after. where D794823 is a dummy variable that is 1 between 1979:4 and 1982:3 and 0 otherwise. and. Equation 30 is a “leaning against the wind” equation in the sense that the Fed is predicted to allow the bill rate to rise in response to increases in inflation. real growth.132 5 US STOCHASTIC EQUATIONS in behavior if the people with influence change their minds or are replaced by others with different views. only one abrupt change in behavior appeared evident in the data. This treatment is based on the assumption that the Fed has a target bill rate each quarter and achieves this target through manipulation of its policy instruments. the labor market tightness variable. Equation 30 is the estimated interest rate reaction function It has on the left hand side RS. The stability test was not run because of the use of the dummy variable. and the equation passes all of the tests.

2964 .5565 .5. and the following is a discussion of the procedure used here. interest payments—g I N T F is the level of net interest payments of the firm sector.2245 . -15.14 3.87 47.91 5. 5. -5.2033 -.970 2.0684 15.0777 . and AG is the same for the federal government.25 Test Lags RH O = 4 T Leads +1 Leads +4 Leads +8 e p4 e p8 χ 2 Tests χ2 df 9.8923 .7861 .26 3.6098 .36 -5. The difference |AG| − |AG−1 | is the net change in the value of securities of the federal government between the end of the previous quarter and the end of the current quarter.07 1. The current level of interest payments depends on the amount of existing securities issued at each date in the past and on the relevant interest rate prevailing at each date.01 t-stat. It depends on past issues and the interest rates paid on these issues.7411 .50 5.96 10. and they consist of both short term and long term securities.50945 . A number of approximations have to be made in trying to model this link.5096 Estimation period is 1954:1–1993:2.34 3. Data on both of these variables are NIPA data. AF is the level of net financial assets of the firm sector.44 0.76 3.5074 . I NT F .0196 . interest payments—f. AF and AG are negative because the firm sector and the federal government are net debtors.9 INTEREST PAYMENTS EQUATIONS 133 Table 5.5085 .07 5.5806 .01 0. I N T G.30 Equation 30 LHS Variable is RS Equation RHS Variable cnst RS−1 100[(P D/P D−1 )4 − 1] JJS P CGDP R P CM1−1 D794823 · P CM1−1 RS−1 RS−2 SE R2 DW Est.43 6 4 1 3 12 6 1 1 p-value .08 9. and I N T G is the same for the federal government. The link from AF to I N T F (and from AG to I N T G) is thus complicated.9 Interest Payments Equations Equation 19. Consider the federal government variables first.5116 .28 0. Equation 29.1552 . The value of new securities issued by the federal government during the current quarter is this difference plus the value of old securities that came due during . Data on both of these variables are FFA data.

and an interest rate RB − η on its long term securities. Let BG denote the value of long term securities issued during the current quarter by the federal government. It is next assumed that the government pays an interest rate RS on its short term securities.) Given the above assumptions. The other securities have the relevant bond rate multiplying them. λ. the aim of the estimation work is to find values of k. of |AG|−|AG−1 |) consist of long term issues. Then the above assumptions imply that : BG = λ(|AG| − |AG−1 |) + BG−k (5. where the short term is defined to be one quarter.41. (RB−1 − η)BG−1 is thus part of I N T G until the securities expire after k quarters. RB. where k is to be estimated. In addition. it is assumed that the long term securities that expire during the quarter are replaced with new long term securities. and AG.e. where RB is the AAA bond rate and η is a constant parameter to be estimated. For example. and the relevant interest rate for these securities is RB−1 − η. where RS is the three month Treasury bill rate.41 being close to the actual values. It is next assumed that λ percent of the net change in the value of securities in a quarter (i.40) λ is assumed to remain constant over time. which is taken to be of length k quarters. for the firm sector BF0 is the same as BF . the interest payments of the federal government are:17 0 I NT G = i=−k 1 1 (RBi − η)BGi + RS(1 − λ) |AG| 400 400 (5. This work takes as given the actual values of RS.. the value of short term securities is (1 − λ) |AG|. Similarly. and so RS multiplies this. with the rest consisting of short term issues. 17 In . Using equations 5. The the notation in this equation BG0 is the same as BG. BG−1 is the value of long term securities issued last quarter. It is first assumed that the government issues two kinds of securities.134 5 US STOCHASTIC EQUATIONS the current quarter. and a “long term” security. that lead to the predicted values of I NT G from equation 5.e. a “short term” security. (I NT G is at a quarterly rate.. i. The estimation period was 1952:1–1993:2.40 and 5. η that lead to a good fit. η is subtracted from RB because the government generally pays less than the AAA bond rate on its bonds. which is the period for which data on AG exist. Given these assumptions.41) The interest rates are divided by 400 in this equation because they are at annual rates in percentage points and they need to be at quarterly rates in percents.

These two equations are presented in Table A. and η chosen. λ. which produced a RMSE of . First. η was taken to be zero. equation 5. . fairly flat over a number of values. The entire procedure can then be repeated for a different set of values of k. λ.40 and the given values of k and λ.) Third. λ = .3 in Appendix A using the values of k. and 40 after 1991:2. given the values of k and λ and the above computations and given a value for η.72. a dummy variable. Second. the value of BG for 1952:1 was taken to be equal to (1/k)λ |AG1952:1 |. No values of k and λ could be found that gave sensible fits for this period. a value of k of only 10 quarters seemed small. which produced a RMSE of 1. λ. 2 in 1981:4. Equation 5.43.063. The term γ T I was then added to the equivalent of equation 5.40. and η and print out the RMSE for each set. Given these values. they were taken to be equal to the 1947:1 value. and the set that was chosen for the . I NT F .41 can be used to obtain predicted values of I NT G for the 1952:1–1993:2 period. A program was written to search over different sets of values of k. and equation 5. values of BG for 1952:2 through 1993:2 can be generated using equation 5. Second. the k − 1 values of BG before 1952:1 were taken to be equal to this value.41 is equation 29. . and η = .41 for the firm sector. The set of values that gave the smallest RMSE was k = 52. which means that firms are assumed to pay interest rate RB on their long term securities. the objective function was fairly flat over a number of values. (Values before 1947:1 are needed if k is greater than 20 quarters.66. The objective function was. . and η. For the first set of values. and η = . which produced a RMSE of . if values of RB for the computations in equation 5. with two differences. however. .40 above is equation 56 in the model. λ = . from which a root mean squared error (RMSE) can be computed. The searching procedure for the firm sector thus consisted in searching over values of k. λ = . T I .9 INTEREST PAYMENTS EQUATIONS 135 estimation procedure was as follows. and γ . where γ is a coefficient to be estimated. was constructed that was 0 through 1981:2. and another RMSE computed. and so k was increased somewhat for the final set even though this resulted in some increase in RMSE. 1 in 1981:3.836. The set that gave the smallest RMSE was k = 10.5. given a value for k and a value for λ. A similar procedure was followed for the interest payments of the firm sector. which is the first quarter for which data on RB exist. To account for this unexplained growth.964.4. and γ = .41 were needed before 1947:1. λ. between 1981:3 and 1991:2 I NT F grew faster than seemed consistent with the values of the interest rates and AF . First.5. Again. 40 in 1991:2. In addition. and the set that was chosen for the model is k = 16.

and so k was decreased somewhat for the final set even though this resulted in some increase in RMSE.) With respect to the notation for the model in equation 4. Finally.1 in Chapter 4. one version of equation 29 had log I NT G regressed on a constant. sensible in the sense of being consistent with the predicted values of interest rates and security issues. For example. both equations 19 and 29 have in general nonzero values of uit . which produced a RMSE of 1. it does tie changes in these variables to changes in interest payments in a way that is not likely to deviate substantially in the long run from the true relationship. they are still stochastic equations in the sense that the predicted values from the equations do not in general equal the actual values. λ = .3 using the values of k. These types of equations provide a slightly better fit than the procedure discussed above.41.144. but they have poor dynamic properties. In other words. Equation 5. 18 In . and γ chosen. Since households hold much of the debt. and γ = .41 for the firm sector is equation 19.7 that equation 29 has an effect on the effectiveness of monetary policy in the model.136 5 US STOCHASTIC EQUATIONS model is k = 40. log(−AG). as will be seen. λ. These two equations are also presented in Table A. The equations were usually in log form and usually included the lagged dependent variable.40 above for the firm sector is equation 55 in the model. The variance of the error term in equation 19 previous specifications of equations 19 and 29 the interest payments variables were regressed on interest rates and the value of securities. and equation 5.40. although equations 19 and 29 have not been estimated in a usual way. For 3SLS and FIML estimation and the stochastic simulation work below. is now offsetting more of the substitution effect of a change in interest rates than it did earlier. interest payments change as interest rates change and as AG and AF change in a way that seems unlikely to drift too far from the truth. and log RB. log RS. Although the above specification is obviously only a rough approximation of the links from interest rates. As the size of the federal government debt (|AG|) increases. (Equations 55 and 56 are. the predicted interest payments from the equations tend to drift away from sensible values. the change in interest payments of the federal government for a given change in interest rates increases in absolute value. AG. A value of k of 52 quarters seemed large. identities because BF and BG have simply been constructed using the equations. and AF to interest payments.18 It will be seen in Section 11. however. the error terms for equations 19 and 29 have been used after adjusting for heteroskedasticity. With no restrictions imposed. where an estimate of the covariance matrix of all the errors in the model is needed. log I N T G−1 . the change in interest revenue of the household sector for a given change in interest rates is getting larger in absolute value as the size of the debt increases. This income effect on households is thus increasing over time and.

19 This means that uncertainty from equations 19 and 29 is taken into account in 3SLS and FIML estimation and in the stochastic simulation work even though they are not estimated in a traditional way. but only slight evidence 19 AF is close to zero for the first few quarters of the estimation period. 4) Leads +1 in equation 13. This means that equation 19 is divided through by |AF + 10| and that equation 29 is divided through by |AG| before computing the error terms to be used in estimating the covariance matrix.5. age distribution has an effect on U. The wealth variable is significant in two of the four household expenditure equations.10 Additional Comments The following is a discussion of some of the results that pertain to sets of equations. and this is the reason for adding 10 to it.10 ADDITIONAL COMMENTS 137 was assumed to be proportional to (AF + 10)2 . 5) Leads +8 in equation 13. 7) Leads +8 in equation 24. 3) Leads +8 in equation 12. 20 This same conclusion was also reached in Fair and Dominguez (1991). 5. 3. At least some of the led values are significant in three of the four household expenditure equations and in one of the four labor supply equation. There is thus some evidence that the rational expectations assumption is helpful in explaining household behavior.S. . They are not significant at the one percent level in any of the other equations in which they were tried except for Leads +4 in the employment equation 13 and for Leads +8 in the capital gains equation 25. macroeconomic equations. This is thus evidence that the U. 2) Leads +1 in equation 11. 6) Leads +4 in equation 24. Changes in stock prices thus affect expenditures in the model through their effect on household wealth. The age variables are jointly significant at the five percent level in three of the four household expenditure equations. 1. and the variance of the error term in equation 29 was assumed to be proportional to AG2 . the age variables were also significant in the equation explaining I H H . and the sign patterns are generally as expected.S.20 2. In Fair and Dominguez (1991). They are significant at the five percent level in eight other cases: 1) Leads +4 in equation 5. and 8) Leads +4 in equation 27. contrary to the case here.

5. and their coefficient estimates have the wrong expected sign in the import equation. 18 of 28 passed the RHO=4 test. 13 and 14. The labor constraint variable (Z) is significant or close to significant in the four labor supply equations. where led values were significant in three of the four household expenditure equations and in two of the four labor supply equations. 7.138 5 US STOCHASTIC EQUATIONS that it is helpful in explaining other behavior. The overall results thus suggest that the specifications of the equations are fairly accurate regarding dynamic and trend effects. The nominal adjustment specification is equation 5.12. but the excess capital variable is not significant in the investment equation 12. all but 3 of 23 passed the T test. All but 3 of 28 equations passed the lags test. 9. . 4. 6.6. In all three of the money demand equations the nominal adjustment specification dominates the real adjustment specification. Also. interest income is part of disposable personal income (Y D). The excess labor variable is significant in the employment and hours equations. The inflation expectations variables are not significant in the four expenditure equations. but in almost none of the other equations. but less so regarding the serial correlation properties of the error terms and stability. Therefore. Table 1. the economic consequences of the rational expectations assumption are examined in Section 11. Given the number of equations that failed the stability test. and 14 of 23 passed the stability test.21 As noted previously. 8. Either the short term or long term interest rate is significant in the four household expenditure equations. an increase in interest rates has a negative effect on household expenditures through the interest rate variables 21 This general conclusion is consistent with the results in Fair (1993b). suggesting that there is a discouraged worker effect in operation. it may be useful in future work to break some of the estimation periods in parts. The evidence suggests that nominal interest rates rather than real interest rates affect household expenditures and imports. which is significant or nearly significant in the four equations. but in general it seems that more observations are needed before this might be a sensible strategy.

More will be said about this in Chapter 11. and the investment equation 12. four appear in equation 27 to pick up the effects of two dock strikes. This result is consistent with the results of estimating highly disaggregate price equations in Fair (1993a). 10. it appears to come through the effects on household behavior rather than on firm behavior. They are not significant in the housing investment equation 4. The spread values are highly significant in the consumption of durables and import equations. One would expect there to be important nonlinearities in the behavior of prices as the economy moves into very high activity levels (and very low levels of unemployment). This is perhaps an area for future research. the data have little to say about the behavior of prices in very high activity periods. One appears in equations 13 and 14 to pick up a structural break. If there is an effect of the spread values on the economy. and one appears in equation 19 to pick up an unexplained increase in interest payments of the firm sector. and the change form is strongly rejected. The acceptance of the level form over the change form has important implications for the long run properties of the model. but this effect cannot be picked up in the data. The evidence is thus mixed. where the stress is on the effects of the spread on investment behavior. 11. This is not necessarily what one would expect from the discussion in Friedman and Kuttner (1993). The real wage constraint in the wage equation is not rejected. one appears in equation 30 to pick up a shift of Fed behavior between 1979:4 and 1982:3. They are significant at the five but not one percent level in the consumption of nondurables equation. On the other hand. The level form of the price equation is not rejected. 12.10 ADDITIONAL COMMENTS 139 and a positive effect through the disposable personal income variable. the production equation 11. the significance of log P H in equation . only on the price level. A permanent change in demand in the model does not have a permanent effect on the rate of inflation.5. There is a fairly small use of dummy variables in the equations. which is as expected. 13. where the level form gave somewhat better results. Four of the most serious negative test results are the highly significant time trends in equations 1 and 3. and so the data suggest that the real wage rate is not a function of the level of prices or nominal wage rates. two appear in equation 21 to pick up an unexplained increase in capital consumption.

Future work is needed on these equations.140 5 US STOCHASTIC EQUATIONS 6. . and the significance of log P F in equation 27.

1a and 6. As noted in Chapter 5. with up to 15 equations estimated per country.1b through 6. The selection criterion for the first stage regressors for each equation was the same as that used for the US model. The three exceptions are the interest rate. The “a” part of each table presents the estimates of the “final” specification. pp. the choice of first stage regressors for large scale models is discussed in Fair (1984). and tested in this chapter.5. and other variables were added to this set for each individual equation. exchange rate.6 The Stochastic Equations of the ROW Model 6.15b. and they were briefly discussed in Section 3.3. The estimation periods were chosen based on data availability. The 2SLS technique was used for the quarterly countries and for equations 1. The empirical results are presented in Tables 6. the main predetermined variables in each country’s model were chosen to constitute a “basic” set. The OLS technique was used for the other equations for the annual countries. the periods were chosen to use all the available data. one pair of tables per equation. and 3 for the annual countries. and forward rate equations. 215–216. 2. The equations are listed in Table B. Stochastic equations are estimated for 32 countries. This chapter does for the ROW model what Chapter 5 did for the US model. With three exceptions. Briefly. The 2SLS technique had to be used sparingly for the annual countries because of the limited number of observations.3.1 Introduction The stochastic equations of the ROW model are specified.15a and 6. estimated. 141 . and the “b” part presents the results of the tests.

Because of this. This means that no moving average process of the error term has to be accounted for since the leads are only one period. where appropriate. The earliest starting quarter (year) for these periods was 1972:2 (1972). Specification In Section 3. This variable was only added to the equations in which an interest rate was included as an explanatory variable in the final specification. This measure of the expected rate of inflation will be denoted p e . 2) estimating the equation under the assumption of a fourth order autoregressive process for the error term.2. The extra theorizing that is discussed at the beginning of Chapter 5 is also relevant here. For example. One of the additional variables added. subject to data limitations. the searching procedure was the same as that used for the US equations.142 6 ROW STOCHASTIC EQUATIONS where the estimation periods were chosen to begin after the advent of floating exchange rates. The tests are similar to those done for the US equations. the basic tests are 1) adding lagged values. This is a guide for the theory behind the model and in particular for the theory behind the linking together of the countries. and 6) testing for structural stability. As discussed in Chapter 5. To repeat from Chapter 5. The estimation periods used for the leads test were one period shorter than the regular periods because of the need to make room at the end of the sample for the led values. For the quarterly countries the expected rate of inflation was taken to be the actual rate of inflation during the past four quarters. 5) adding additional variables. 3) adding a time trend. and for the annual countries it was taken to be the inflation rate (at an annual rate) during the past two years.1 that is behind the specification of the US model is relevant here. the specification of the ROW equations follows fairly closely the specification of the US equations. the notation “RHO+” instead of “RHO=4” is used in the tables in this chapter to denote the autoregressive test.5 the equations of the econometric model were matched to the equations of the theoretical model of Section 2. this is a test of the nominal versus real interest rate specification. Also.3. was the expected rate of inflation. The led values were one quarter ahead for the quarterly countries and one year ahead for the annual countries. and so the theory in Section 2. 4) adding values led one or more periods. For the annual countries the autoregressive process for the error term was taken to be third order rather than fourth order. Lagged dependent variables were used extensively to try .

Entering the variable in ratio form in the equations allows the error to be approximately1 absorbed in the estimate of the constant term. This procedure is. is in A and not in A/(P Y · Y S). of course.6. Also. All the coefficient estimates in an equation are presented in a table if there is room. where many of the coefficient estimates of the output term seemed too large. Also. the rejection rate was high for the investment equation (equation 3). the asset variable has been divided by P Y · Y S before it was entered as an explanatory variable in the equations. say A. As discussed in Section 3. as will be seen. and the autoregressive assumption was retained if the estimate of the autoregressive coefficient was significant.) This was done even for equations that were otherwise in log form. ¯ and A/(P Y ·Y S) is not constant. some equations for some countries were initially estimated and then rejected for giving what seemed to be poor results. but at least it somewhat responds to the problem caused by the level error in A. and explanatory variables were dropped from the equations if they had coefficient estimates of the wrong expected sign. . Because much of the specification of the ROW equations is close to that of the US equations. The Tables The construction of the tables in this chapter is as follows. the asset variable is off by a constant amount.1 INTRODUCTION 143 to account for expectational and lagged adjustment effects. Only the differences are emphasized. then including the latter ¯ ¯ ¯ variable in the equation means that it is not A but A/(P Y · Y S) that is part of the equation. Both current and one quarter lagged values were generally tried for the price and interest rate variables for the quarterly countries. One difference between the US and ROW models to be aware of is that the asset variable A for each country in the ROW model measures only the net asset position of the country vis-à-vis the rest of the world. and the values that gave the best results were used.3. and so taking logs of the variable is not appropriate. it does not include the domestic wealth of the country. For example.3. If there is 1 If the level error. The equations were initially estimated under the assumption of a first order autoregressive error term. the specification discussion in this chapter is brief. and the reader is referred to Chapter 5 for more detail regarding the basic specifications. Data limitations prevented all 15 equations from being estimated for all 32 countries. crude. (P Y is the GDP deflator and Y S is potential GDP. This is what is meant by the error being only approximately absorbed in the estimate of the constant term.

Many of the values of λ for the annual countries are 1. and so this previous material does not have to be read. Third. an equation will be said to pass a test if the p-value is greater than . The following discussion tries to give a general idea of the results. The model was cut in size to lessen problems caused by poor data. This was done because of the short sample periods. The OECD has better NIPA and labor data than is available from the IFS data. quarterly data were constructed for all the countries by interpolating the annual data. and labor force equations were added to the model (equations 12–15). which means that only one possible break point was specified. Chapter 4. and the trade share matrix is now 45 × 45 rather than 65 × 65.0. as with the US model. and the capital stock variable was used in the investment equation. For the stability test the AP value is presented along with the degrees of freedom and the value of λ. the number of countries (not counting the United States) for which structural equations are estimated is now 32 rather than 42. the present discussion of the model is self contained. First. annual data were used for countries in which only annual NIPA data existed. OECD data were used whenever possible rather than IFS data. Previous Version of the ROW Model The previous version of the ROW model is presented in Fair (1984). Again. Some of the main changes are the following.) To save space. As in Chapter 5. but the reader is left to pour over the tables in more detail if desired. employment. a few more coefficient constraints were imposed. Fifth. In the previous version. More changes have been made to the ROW model since 1984 than have been made to the US model. and when this happens it is discussed in the text. Second. The R2 values is also not presented if there is limited space.01. In a few cases other coefficient estimates are also not presented because of space limitations. wage. . The sample period that was used for the estimation of each equation is presented in the b tables except for equation 10. where the sample periods are presented in the a table. (There is no b table for equation 10.144 6 ROW STOCHASTIC EQUATIONS a space constraint. The AP value has a ∗ in front of it if it is significant at the one percent level. which means that the equation fails the stability test. as for the US model. There are obviously a lot of estimates and test results in this chapter. only the p-values are presented in the b tables for the χ 2 tests. estimates of the capital stock of each country were made. the estimate of the constant term is not presented. and it is not feasible to discuss each estimate and test result in detail. Fourth. Finally.

6 in Fair (1984).1a and 6. 2) the import variable includes all imports in equation 27 but only merchandise imports in equation 1. The expected inflation variable is relevant for 9 countries. and the lagged dependent variable. asset variable was not significant in equation 27 and so was dropped from the equation. The three main differences between the equations are 1) the U. The explanatory variables include the short term or long term interest rate.2 Equation 1. Table 6. 5 fail the RHO+ test. and some of the discussion in the following sections is similar to the discussion of the previous version in Sections 4. although it is not significant. The results in Tables 6. and 14 fail the stability test. and it is significant at the one percent level in only 1 case. Four of the 32 equations fail the lags test (at the one percent level).2 and it is only significant for 1.S. 6.3 Equation 2: C: Consumption Equation 2 explains real per capita consumption.6. The variables are in logs except for the interest rates and the asset variable. The explanatory variables include price of domestic goods relative to the price of imports.1b show that reasonable import equations seem capable of being estimated for most countries. however. 6.2 EQUATION 1. log P Y was significant) in 7 of the 23 cases.2. 2 Remember . the lagged that the expected inflation variable is relevant if an interest rate appears as an explanatory variable in the equation. The constraint was rejected (i. Only for Switzerland (ST) is the coefficient estimate for income of the wrong expected sign. To save space.2. the lagged value of real per capita assets. the short term or long term interest rate. M: MERCHANDISE IMPORTS 145 The basic structure of the ROW model has.1a does not include the estimate of the constant term. and 3) the income variable is disposable personal income in equation 27 and total GDP in equation 1. per capita income. 9 fail the T test. per capita income. The led value of the income variable was used for the leads test.e.5 and 4.. For the countries in which the relative price variable was used. M: Merchandise Imports Equation 1 explains the real per capita merchandise imports of the country. Equation 1 is similar to equation 27 in the US model. the log of the domestic price level was added to test the relative price constraint. remained the same between the previous version and the current version.

80 a Variable lagged once.516 (0.22) SW – GR IR PO SP NZ SA VE CO JO SY ID MA PA PH TH .41) .0384 .19) . .21) .22) .34) .229 (1.960 (2.723 (13.79 2.989 (23.0815 2.53 2.0369 .65) .117 (3.035 (1.05 1.96 2.21) .0072 (-2.03) – .74) – – 1.0268 .32) .953 (1.82) a -.78 2.28) – – -.08) 1.19) .93) .670 (4.593 (3.94) .21) .0574 .116 (3.51) – .84 1.82) a .20 1.651 (2.50) .0767 .486 (3.1596 .52) .474 (6.100 (2.893 (2.1a log(M/P OP ) = a1 + a2 log(M/P OP )−1 + a3 log(P Y /P M) + a4 (RSorRB) +a5 log(Y /P OP ) + a6 [A/(P Y · Y S)]−1 a2 a3 a4 a5 a6 ρ SE DW Quarterly CA .538 (1.273 (0.404 (4.0054 (-4.24) FI .0024 (-0.166 (2.1487 .16) .492 (3.205 (2.831 (4.93) SO .16) GE .638 (4.17) .90) .744 (2.0257 .37) – .58) – .199 (4.68) .53) .08 1.59) 1.941 (8.280 (1.80) – .66 2.74) .206 (3.282 (1.356 (2.91 1.77) – – – -.51) .441 (2.82) -.47) .184 (1.996 (6.673 (6.05) – – – – – – – – – – – – – – – – – – – – .193 (2.49) – – – – – – a -.70) – .18) 1.51) .146 6 ROW STOCHASTIC EQUATIONS Table 6.096 (4.13 1.401 (3.42) – – -.1047 .592 (7.69) .0830 .07) .21 1.87) .04) DE .0661 .213 (2.78) – .16) .25) – – – – – – – – – – – – .96) .79) 1.328 (2.2217 .199 (4.72) UK .70 2.87) 1.283 (-1.0533 .48) 1.92 2.735 (4.49) 1.470 (1.0783 .688 (5.622 (3.98) IT .78 1.303 (1.053 (2.1115 .95) .65) .52) – 1.035 (1.830 (16.49) – – – b -.0073 (-4.94) 1.282 (1.0767 .867 (11.765 (14.86 2.0556 .59) .67) 1.18) .0540 .327 (3.01 1.274 (3.73) .80) .830 (23.0392 .141 (0.659 (1.634 (4.94) .0395 .274 (2.77) .073 (2.0080 (-1.195 (2.478 (2.0767 .265 (1.24) .253 (1.485 (5.63) .055 (1.872 (18.11) .37) ab -.744 (0.13 2.29) – – .12) Annual BE .250 (1.83) .488 (3.70) .893 (7.306 (5.565 (7.82) .126 (1.14) – .402 (-2.101 (4.1043 .0280 .655 (2.0026 (-1.68) 2.137 (3.37) – – a .40) 1.86) – – – – -.514 (2.83) .57) b -.12) .0422 .866 (8.81) .298 (3.01) 1.92) .68) .90 2.56) .180 (1.55 .75) ST .67) .05 1.18) NO .01 1.229 (1.09) -.37) .39) a -.0168 (-1.125 (1.21 1.25) NE .82 1.98 2.85) KO .47) .219 (2.79 1.0016 (-0.14 1.06) .24 1.022 (0.289 (3.212 (2.55) .33 1.31) .06) – – .44) AS .066 (0.911 (6.961 (2.758 (5.61) -.324 (2.59) 1.020 (3.0018 (-1.173 (2.96 2. b RB rather than RS.64) a .29) .0459 .32) 1.0291 .0590 .21 1.0244 .169 (1.1900 .29) .208 (-2.1002 .682 (5.639 (10.744 (2.622 (2.56) FR .272 (2.82) .51) AU .912 (5.25) .57) .465 (1.381 (5.243 (-1.72) .0346 .282 (3.648 (7.004 (2.18) – .83) JA .200 (0.

005 .805 CO .003 .679 a .325 FI .392 .889 .948 .004 UK .736 .269 .521 Annual BE .00 ∗ 7.434 .189 .094 JA .92 ∗ 10.127 .357 .110 .132 .214 .873 .00 Sample 1966:1–1992:3 1967:3–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1971:1–1991:4 1966:1–1992:3 1976:1–1991:4 1971:1–1992:2 1962:1–1991:2 1964:1–1991:4 1969–1990 1969–1990 1974–1990 1969–1990 1963–1990 1969–1990 1962–1990 1969–1990 1962–1990 1970–1989 1963–1991 1972–1991 1971–1991 1965–1990 1962–1989 1972–1987 1972–1991 1962–1991 1962–1990 a log P Y −1 used rather than log P Y . The results in Tables 6.2a.406 .300 GR .3 EQUATION 2: C: CONSUMPTION Table 6.023 .009 .001 – .286 .000 .00 2. Equation 2 is similar to the consumption equations in the US model. 3 fail the RHO+ test.806 Stability AP (df) λ 5.83 (5) 1.204 147 Lags p-val Quarterly CA .721 .049 NZ .790 .007 .926 .22 (5) 1.42 (3) 1.014 – – .066 .000 a .363 MA .002 .70 (5) 2.202 .03 ∗ 9.000 – – – – – – – – – – – – log P Y p-val .00 ∗ 18.000 .061 .748 . and so interest rate and wealth effects on consumption have been picked up as well as the usual income effect.007 NE .94 (6) 4.003 .369 .068 .794 . As in Table 6.153 SW .434 SY .381 JO .88 (4) 2.31 (5) 1.054 .391 .545 .653 .51 (4) 1.57 ∗ 9.35 (4) 1.00 ∗ 12.1b.100 .17 (3) 1.479 .00 ∗ 12.003 .1a and 6.48 (5) 1.097 FR .225 KO .994 .721 .97 (4) 1.734 .2b are of similar quality to the results in Tables 6.1a.070 .125 .39 (4) 1.237 SO .011 – – a .020 .654 .145 PH .73 (4) 1.439 .001 .00 3.537 .785 .150 TH .387 .446 PA . The two main differences are 1) there is only one category of consumption in the ROW model compared to three in the US model and 2) the income variable is total GDP instead of disposable personal income.00 1.974 .57 (5) 1.532 .000 . Most of the tests in Table 6.2a and 6.065 .230 .000 . the estimate of the constant term is not presented in Table 6.289 .33 (5) 2.733 .081 .88 (5) 3.666 .00 ∗ 9.320 .00 4.2b are passed.891 .21 6.02 (7) 4. and the lagged dependent variable.113 ID .3 and 9 fail the stability 3 Multicollinearity problems prevented the leads test from being computed for the UK.251 AU .00 2.006 .572 VE .00 ∗ 12.004 .015 .06 (6) 1.113 .07 (5) 1. ∗ Significant at the one percent level.1b Test Results for Equation 1 RHO+ T Leads p-val p-val p-val .206 .427 .052 IT .154 .00 5. .00 1.099 – .00 ∗ 28.971 – .57 (5) 1.199 NO . The variables are in logs except for the interest rates and the asset variable.54 (4) 3.00 3.6.315 .362 .73 (4) 1.25 5.490 .285 .11 (5) 1.119 .001 .052 .952 .379 .072 .62 (5) 1.94 (4) 2.365 DE .008 – .841 – – – .531 .093 .060 ST .184 – – – – – – .94 (5) 1.467 .704 .302 . The interest rate and asset variables appear in most of the equations in Table 6.00 5.00 ∗ 7.2a.341 – – .081 .00 ∗ 13. 5 fail the leads test.001 .422 .17 5.000 .865 .47 (5) 1.518 – . value of real per capita assets.508 .699 .588 SA .681 .087 .309 .022 .876 .375 .381 .00 6.000 .026 .57 4.00 3.997 .786 .002 SP .352 .276 .044 PO .884 IR .51 ∗ 10.00 (6) 2.398 .095 .017 AS .014 .69 ∗ 42.03 5.71 (5) 1.858 .000 pe p-val .005 .03 (6) 1.77 (5) 1.00 1.177 . Eight of the 32 equations fail the lags test.031 GE .

518 (-3.55) .383 ( 3.257 ( 2.0007 (-2.09) .189 ( 3.22 a Variable lagged once.0244 .25) b -.65 2.337 (-1.87) – – -.253 (-2.88 1.03) .73) .0123 .836 (15.842 ( 5.965 (20.016 ( 4.455 ( 3.99) .52) – .689 (13.0022 (-3.0017 (-4.53) .0078 .882 (19.46) .38) .91) .07) .80) NE -.899 (12.359 (-3.49 1.013 ( 2.37) .905 (21.04) .186 ( 1.18) .017 ( 1.63) Annual BE -.0013 (-5.49) – – a1 Quarterly CA -.70) UK -.013 ( 0.254 ( 2.709 ( 3.530 ( 5.96) – .81) .59) SO -.21) .85 2.148 6 ROW STOCHASTIC EQUATIONS Table 6.19) GE .436 (-1.61) .96) .183 (-2.698 ( 4.0069 .34 1.477 ( 3.499 ( 3.21) .540 ( 5.10) – -.65) -.190 ( 0.268 ( 2.464 ( 2.509 ( 2.21) .92) -.39) .59) .92 2.77) – -.21) – -.87 2.732 ( 8.051 ( 1.51 2.50 1.286 (-2.0029 (-2.198 (-0.83 2.050 ( 1.19) NZ -.29) .57) .027 (-0.537 ( 6.36 1.454 ( 5.568 ( 5.85) – – – – .272 ( 2.02 1.0129 .17) .354 ( 1.39) .0112 .48) .106 ( 2.07 2.86) SP -.27) .78) .35) .42) VE -.62) SW -.0082 .85) .497 ( 5.45) -.69) -.76) .63) .593 ( 2. .31 2.24) ab -.99) .54) . b RB rather than RS.34) DE -.159 ( 2.0179 .96 1.720 ( 4.70) .0135 .78) .42) .520 (-2.243 ( 1.962 ( 2.821 (15.30) -.0029 (-4.309 ( 6.1439 .0051 (-2.96) .625 ( 5.0036 .666 ( 5.861 (18.22 1.021 ( 2.059 (-0.246 ( 1.97) – .404 ( 2.087 ( 0.24) .098 (-1.93) .81) – -.94) – .99) .980 ( 1.468 ( 5.011 ( 3.0313 .29 1.64) – .052 ( 0.48) -.123 ( 4.13) IR .90) AU -.20) .73) .08) – a -.92) .088 (-2.45) .60 2.73 2.0254 .196 ( 0.50) – -.151 ( 2.481 (-0.61) .44) .58) FR -.213 ( 3.04 2.99) .87) .357 ( 3.41) .75) .0319 .160 ( 1.89) – – – -.00) .73) .506 ( 6.136 ( 1.67) GR .64) NO .340 (-2.026 ( 4.31) .0272 .04) .0091 (-3.233 (-4.0019 (-2.35) KO .023 (-0.11 1.0266 .495 ( 5.60) .05 1.11) .166 (-1.093 ( 2.0010 (-2.46) .338 ( 3.100 ( 1.69) SA -.113 (-0.51) PO -.427 ( 3.0357 .1059 .617 ( 3.36) b -.82) .48 1.0176 .97 2.68 1.937 (-1.0504 .98) -.102 ( 2.0048 (-2.23 1.451 ( 3.0011 (-4.12 1.164 ( 1.53) .802 ( 7.302 ( 1.396 ( 0.231 ( 2.74) .0555 .94) b -.75 1.57) .72) .43) .43) .0248 DW 1.35) JA .47) -.85) .68) -.524 ( 5.0062 .0060 .0139 .64) .26) b -.05) .622 (-6.0093 .0588 .396 (-0.910 ( 1.98) a -.0017 (-3.57) .0009 (-0.127 (-1.297 ( 3.910 (-1.018 ( 0.43) AS -.2a log(C/P OP ) = a1 + a2 log(C/P OP )−1 + a3 (RSorRB) +a4 log(Y /P OP ) + a5 [A/(P Y · Y S)]−1 a2 a3 a4 a5 ρ .91) .032 ( 1.842 ( 6.68) – – – – – – – – – – .00) .36) .22) .022 ( 0.0342 .70) .157 ( 3.0085 .120 ( 2.75) IT -.263 (-2.895 (22.0024 (-1.975 (23.634 ( 4.853 ( 9.68) .88 1.20) – – – .08) – b -.21) .0110 .89) – – – -.422 ( 2.01) .14) ST .0101 .166 ( 1.12) .54 2.0012 (-0.91) – .36) FI -.48) CO – JO SY ID MA PA PH TH -1.398 ( 3.467 ( 4.227 (-1.079 ( 0.0044 (-1.32) .70) .25) .557 ( 5.519 ( 2.16) – – .78) – – – – b -.05) .27 1.529 (-2.268 (-4.65) -.37) SE .923 (34.079 ( 0.0142 .449 ( 4.

022 .98 (5) 1.241 .193 .281 .008 .030 – – – .59 (4) 1.003 .218 .4 EQUATION 3: I : FIXED INVESTMENT Table 6.2b Test Results for Equation 2 Lags p-val Quarterly CA . and it is significant in only 5 cases.003 .828 .66 3.6.940 .070 – – – – RHO+ p-val .047 .909 .00 ∗ 13.4 and it is only significant for 2. the current 4 Multicollinearity problems also prevented the χ 2 test from being performed for the UK for the p e case.00 4.747 .507 .067 .26 (4) 1.001 GE .63 (4) 1.160 .324 Stability AP (df) λ ∗ 19. the lagged value of the capital stock.001 .08 Sample (6) 4.00 ∗ 9.003 .547 GR .872 .818 .00 1966:1–1992.150 .00 2.304 – .00 0.506 .361 .00 2.270 .471 .108 NZ .00 ∗ 13.016 .519 – .172 .19 (4) 1.125 .106 – .980 .892 .905 .000 .20 (6) 1.03 (3) 2. test.886 .007 – .021 FI .06 (5) 2.905 SO .836 – .288 DE .331 .99 (5) 1.139 .578 .98 (2) 1.444 .248 4.783 .178 .085 IT .101 TH .020 UK .893 .00 7.099 SY .048 .616 .585 .236 .00 0.39 (6) 3.102 .773 .00 1.00 ∗ 9.001 ID .002 AU .35 (3) 1.024 .085 .17 ∗ 11.513 – .01 3. The expected inflation variable is relevant for 16 countries. .000 . 6.155 .69 ∗ 13.92 5.000 .076 .786 JO .150 .941 .033 .755 .398 .241 .205 VE .017 .779 SP .839 PO .07 (5) 1.106 CO .61 (4) 2.048 AS .968 .119 MA .288 .000 .000 .00 2.24 (3) 1.047 .003 .27 (4) 1.048 IR .00 5.613 .000 .755 .00 6.21 6.000 PH .015 .002 .48 (3) 1.3 1967:3–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1971:1–1991:4 1966:1–1992:3 1976:1–1991:4 1971:1–1992:2 1962:1–1991:2 1964:1–1991:4 1969–1990 1969–1990 1974–1990 1969–1990 1963–1990 1969–1990 1962–1990 1969–1990 1962–1990 1970–1989 1963–1991 1972–1991 1971–1991 1965–1990 1962–1989 1972–1987 1973–1991 1962–1991 1962–1990 at the one percent level.00 0.695 .46 (6) 2. The led value of the income variable was used for the leads test.114 .188 .299 .00 6.531 .09 (4) 1.590 .00 3.448 PA .892 ST .383 .110 .41 (6) 3.118 NE .510 .006 . It includes as explanatory variables the lagged value of investment.298 .292 SA .698 .317 .03 (6) 2.57 (4) 1.003 .00 5.44 (5) 1.306 .005 SW .621 .006 JA .921 .659 .244 .258 .001 KO .000 .87 (5) 1.56 (5) 3.57 3.077 NO .00 ∗ 6.18 (4) 1.968 .083 .28 ∗ 8.007 .209 Leads p-val .4 Equation 3: I : Fixed Investment Equation 3 explains real fixed investment.114 .03 2.00 2.357 – – – – .191 ∗ Significant 149 pe p-val .104 Annual BE .75 (4) 1.390 .392 .000 FR .99 (5) 2.71 (4) 1.292 T p-val .03 7.85 (6) 1.708 .28 (5) 1.189 .

66) IT .25) Annual BE .042 ( 4.69) .121 ( 4. and the short term or long term interest rate.8421 1.1122 (-4.66) SO .220 ( 4.90) -.86) a -50.5534 263.93 1.46) SA .93) a4 .18) UK .82 1.44) .54) .02) JA .798 ( 22.10 SE 464.66 1.0392 (-2.136 . Equation 3 differs from the investment equation 12 for the US model.688 ( 4.06 (-2. did not produce sensible results for most countries.22 (-3.178 ( 5.3807 216.97) ab -54.83 (-2.57) .72) .92 (-4.043 α1 14.41) -.30) -.3839 538.799 ( 7.27 1.081 .928 ( 20.788 ( 20.04) -.1724 (-5.2) (6.52) -.39) SW .17) .56) -.06) .87 1.75 – – – – -1. Typically.4) .117 ( 5.946 ( 34.00) b -85.96 2. The use of equations 5.3035 343.04 13.0086 (-4.59) – 29.1195 (-4.115 ( 5.9770 2.0976 (-3.887 ( 27.545 .73) -.485 .6964 8. Equation 3 instead is based on the following simpler set of equations: K ∗∗ = α0 + α1 Y + α2 RB K ∗ − K−1 = β(K ∗∗ − K−1 ) I = K − K−1 + δK−1 I − I−1 = λ(I ∗ − I−1 ) ∗ ∗ (6.346 .7454 1.1172 6.095 .468 ( 5.75) ab -122.25 2.619 .55) -.23– 5.50 10.26.21) b -8.084 .677 ( 7.1) (6.66) .182 2.096 ( 3. the coefficient estimate of the current change in output term seemed much too large.0089 (-4.1083 (-3.53) .9135 9.08) -.0019 (-1.77) GE .62 (-1.150 6 ROW STOCHASTIC EQUATIONS Table 6.60) -.70) a Variable a3 -.329 ( 4.86) GR . value of output.786 ( 7.95) lagged once.525 ( 5.45 1.97) a5 ab -38.56 2. b RB rather than RS.271 ( 4.18) PH .64 (-0.799 ( 21.54 1.61 10.99 2.25 in Chapter 5.0044 (-2.059 .17) .59 3.267 ( 3.77 19.47) IR .57 β .59) -.384 .339 ( 4.50 9.0139 (-5.32) -.52 .0238 (-1.3a I = a1 + a2 I−1 + a3 K−1 + a4 Y + a5 (RSorRB) a2 Quarterly CA . which lead to the estimated equation 5.8812 41.13) -.3504 379.80 (-2.23 (-2.69 (-4.26 1.77) .63 2.38 1.8804 4.3) (6.6513 DW 1.44 3.745 ( 7.91) .055 .51) -.0094 (-5.152 ( 2.57 2.84 9.98) .0032 (-1.041 .071 ( 2.52) DE .50) FR .62 2.858 ( 17.46) -.607 ( 5.

40 Sample (6) 4.1 is the capital stock that would be desired if there were no adjustment costs of any kind.000 .46 2.113 DE . where K ∗ is the capital stock that would be desired if there were no costs of adjusting gross investment.021 SO .) By definition. (As discussed in Chapter 3.016 JA .74 (5) 2.4.655 .6. and the interest rate.001 T p-val .842 .003 .046 .00 0. two types of partial adjustment are postulated.00 ∗ 9. which is equation 6.68 at the one percent level.430 .320 .28 (5) 1.566 .222 .757 .00 2.00 (4) 1.054 Annual BE .205 PH .005 .812 .018 .255 . where δ is the depreciation rate.5) Gross investment is thus a function of its lagged value.98 ∗ 14.3b Test Results for Equation 3 Lags p-val Quarterly CA .059 .003 ∗ Significant 151 pe p-val .186 .947 . It is taken to be a function of output and the interest rate. “desired” gross investment.06 for the annual countries. I ∗ .69 (4) 1.008 .975 .1–6.453 .72 (5) 2.197 – – – – – .008 .92 5.00 3.755 . is determined by equation 6.000 Leads p-val .66 ∗ 8. As was the case for durable consumption and housing investment in Chapter 5. equation 6.99 (4) 1.3.98 4.4 yields: I = (1 − λ)I−1 + λ(δ − β)K−1 + βλα0 + βλα1 Y + βλα2 RB (6.129 SA . Combining equations 6.000 SW .630 .000 FR .03 ∗ 8.00 1966:1–1992:3 1967:3–1992:3 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1966:1–1992:3 1962:1–1991:2 1969–1990 1969–1990 1969–1990 1963–1990 1969–1990 1970–1989 1962–1991 ∗ 26.001 .079 .3 is the same equation for the desired values. current output. .671 IR .359 .251 .000 IT .47 (5) 1.003 UK .63 (5) 3.328 GR .000 Stability AP (df) λ ∗ 11.00 (5) 1. the lagged value of the capital stock. Given K ∗ . δ is .570 . As was the case for the stock of durable goods and the stock of housing in the US model in Chapter 5. K ∗∗ in equation 6. the two partial adjustment equations are a way of adding both the lagged dependent variable and the lagged stock to the equation.155 .693 .967 .014 . and equation 6.42 (4) 1.60 (4) 1.883 .4 EQUATION 3: I : FIXED INVESTMENT Table 6.259 .2.25 (5) 3.012 .066 RHO+ p-val .948 .948 .368 .59 ∗ 32.13 (5) 3.919 .005 .293 .603 .577 . The second type of adjustment is an adjustment of gross investment to its desired value.002 .015 for the quarterly countries and . The first.004 GE .00 6.026 . is an adjustment of the capital stock. I = K − K−1 + δK−1 .490 .

068 .67) .466 .126 ( 2.28 390.89) .47) .58) a3 .78) .636 .80) .3838 (-6.628 ( 4.15) .47) UK .22) – .66 DW 2.3826 (-2.408 ( 3.255 .158 .261 ( 2.76) .78) -.0940 (-1.92 1.245 ( 3.69) -.495 ( 3.027 (13.66 252.632 ( 2.305 ( 4.95) – – .85) .063 ( 0.326 ( 5.39) FR .10 417.29) JA .84 2.57 13.882 ( 5.05 1.874 (17.34 6.1849 (-3.289 ( 1.26) .22 1148.2461 (-2.91 1.5022 (-1.57 1.025 .81 1.717 (12.955 (17.965 (10.72) – SE 562.871 ( 7.08) .61) .50) -.0765 (-1.2232 (-1.32) .97 1.966 (21.09) .056 .315 ( 7.19) -.38) .78) -.762 (10.474 .933 (19.050 .83 .100 β .53) GE .118 .00 1.16 1.76 1.109 .20 1.80) .750 1.32 7.95 2.71) .731 ( 7.30 36.553 .06 2.0622 (-1.373 .722 ( 6.173 .420 ( 1.1330 (-1.04 1.11) – – .67 3.271 .031 .38 6.446 ( 4.60) -.302 .07) – – – – – .99 1.37) .0186 (-1.82) .096 .20) .36) AS .39) .461 2.03 1.21 2.64) -.680 .123 .556 2.452 .725 (15.0901 (-5.50) .01) ρ .16) -.156 .74) .0229 (-0.171 ( 3.090 .87 2.4a Y = a1 + a2 Y−1 + a3 X + a4 V−1 a2 Quarterly CA .03 2.89) .96 1.33) 1.60 177.0822 (-1.94) a4 -.20 24.98) -.81) .1578 (-0.68) AU .09 6.0984 (-1.64) IT .88) -.177 ( 0.039 -.85 2.95) FI .209 .33) -.192 .09) NE .246 .666 1.319 ( 3.71) .28) -.44) -.91 3.006 .32) .318 -.1728 (-3.594 ( 7.531 ( 4.78) – – .62 2.363 .91) -.724 ( 9.70 544.509 ( 4.03 17.212 ( 3.24) -.002 (-0.07 2.94) -.765 ( 7.705 ( 7.189 ( 3.152 6 ROW STOCHASTIC EQUATIONS Table 6.47) .856 (12.99) KO .442 ( 6.17) 1.112 .667 (10.0219 (-1.284 ( 5.96) .12) -.55) -.99 α .871 .96 1194.645 ( 6.665 ( 6.485 .410 ( 2.119 .0593 (-1.2416 (-1.298 .41) 1.351 ( 0.50) .0645 (-5.78 53.48) -.888 ( 0.469 ( 4.1560 (-0.428 ( 4.801 (11.624 ( 5.1030 (-1.264 .195 .075 .019 (59.88) .348 .60) -.57) -.009 (50.083 ( 1.91) -.01) .80) .08) .242 .58) -.198 ( 1.01 1.408 .585 .480 .380 ( 3.71 1.3792 (-2.04) .13) -.62 1.33 100.88) Annual BE – DE SW GR IR SP SA VE CO JO PA PH TH – .

the RHO+ test in 2 cases.406 .35 (5) 1.026 .00 5.217 .13 (4) 1.639 .00 ∗ 9.66 ∗ 23.15 (4) 1.00 2.04 (4) 1.500 KO .826 PA .297 . For the quarterly countries λ (i.191 .03 5. In the case of CA both rates gave about the same results and . and the leads test in 6 cases.94 (5) 1.000 .061 Stability AP (df) λ ∗ 21.041 for CA to .003 IT .87 (4) 1.412 .3b the equation fails the lags test in 6 of the 14 cases. the T test in 3 cases.492 SP .69 3.806 .452 .. gave better results than did the long term rate.619.024 .92 ∗ 11.141 .00 ∗ 9.001 .579 JO . The estimate of λ is one minus a2 in the table.068 .66 (4) 2.58 ∗ 9.834 .851 .540 .331 .340 IR .00 1966:1–1992:3 1967:3–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1966:1–1992:3 1976:1–1991:4 1971:1–1992:2 1964:1–1991:4 1969–1990 1969–1990 1969–1990 1963–1990 1969–1990 1969–1990 1970–1989 1963–1991 1972–1991 1971–1991 1972–1991 1962–1991 1962–1990 at the one percent level.39 (3) 1.54 (4) 1. For the annual countries λ ranges from . the short term interest rate.00 ∗ 8.437 TH .751 .113 . If 4 of the 9 countries for which interest rates were significant. In Table 6.00 1.356 .74 (5) 2.353 .108 .890 .026 .815 FR .57 4.056 Leads p-val .049 .4 EQUATION 3: I : FIXED INVESTMENT Table 6.00 1.227 .063 .095 for JA.071 .73 (5) 2.31 (4) 1.481 .059 .03 2.56 (5) 1.008 .00 2.121 .55 (4) 1.00 4.018 UK .962 . RS.748 .810 .53 (3) 1.025 .42 (5) 1.005 .148 PH .136 to .014 .566 .00 6.97 Sample (5) 4.286 SW .21 ∗ 10.027 SA .001 .220 . The estimate of the constant term is not presented in Table 6.015 .130 .452 .076 .000 .89 (5) 1.001 .00 0.082 .016 JA .141 .001 .303 .6. In none of the 9 relevant cases is the price expectations variable significant.686 .058 .00 4.113 GR .962 .495 .945 .584 .00 ∗ 10.63 (4) 1.012 . The led value of output was used for the leads test.667 .023 .4b Test Results for Equation 4 Lags p-val Quarterly CA .718 FI .054 for JA to .686 GE .3a.008 Annual BE .15 (5) 1. Also presented in the table are the implied values of β and α1 .e.002 .25 (5) 3. 1 − a2 ) ranges from .532 and β ranges from .89 (4) 2.021 AU .174 AS .929 .149 CO .024 .009 NE .00 ∗ 11.736 ∗ Significant 153 RHO+ p-val .212 for UK and β ranges from .004 VE .862 .796 DE .68 (5) 3.00 5. and so it was used.47 (4) 2.987 T p-val .201 to . Equation 3 fails the stability test in 7 of the 14 cases.160 .238 .639 .778 .

0629 .604 ( 15.58) .72 1.64) AS .194 ( 2.67 2.359 ( 8.36) – – – – – – – – – – – – – – – .32) .093 (-4.0108 .449 (-4.172 ( 2.26) .84) NO .0319 .337 (-1.926 ( 11.706 ( 15.73) .166 ( 3.36) – – .99) IR .97) a -.934 ( 33.54) – – .20 1.67) .121 (-1.01) GE .34) a .02 2.0062 .52) .090 ( 5.0439 .03) .266 (-2.83 1.0110 .0239 .88) – .91 1.52) TH .62) .34) .962 ( 74.270 ( 5.70 1.024 ( 2.25) a -.62) .251 ( 3.0761 .87 1.29) PA .140 ( 2.0075 .13) SP .061 ( 0.668 ( 13.019 ( 5.905 ( 4.006 ( 0.0195 .24) .070 ( 1.79) -.179 ( 4.23) – a -.582 ( 25.78 2.327 ( 2.55) .308 (14.48) AU .029 ( 0.010 ( 2.01 2.019 ( .799 ( 16.32) JO .73) .079 ( 1.52) – – .0271 .847 ( 15.03 1.33) .68) .379 ( 1.115 (-1.87) PH .0115 .30 1.68) – – – – .455 ( 2.20) -.89) – – a2 Quarterly CA .98) a .12) – – -.259 ( 3.81) Annual BE .339 (-1.689 ( 2.0086 .26) – – – – .40) SE .300 ( 2.023 ( 1.56) – -.027 ( 2.59) .762 ( 14.34) .25 .19) DE .829 ( 23.62) .27) .43) .071 ( 4.490 ( 4.088 ( 1.89 2.89) .125 (-1.63 2.58) .682 ( 12.41) ST .82) .61) – .378 ( 1.0066 .082 ( 4.066 (-0.296 (-1.34) .17) .54) a -.17 1.0115 .374 ( 4.67) -.39) – – – – .42) – – – – – – – – – -.16) a .107 ( 2.842 ( 13.06) .75) .022 ( 0.29 2.70) JA .96) – – – – – – a .0257 .39) SY .79 1.87 .42) – – .246 ( 3.97) .154 6 ROW STOCHASTIC EQUATIONS Table 6.0097 .0055 .80) .015 ( 1.0119 .69) .165 ( 4.802 ( 14.613 ( 6.546 ( 12.699 ( 28.19) – .985 (219.978 ( 48.101 (-2.01) MA .186 ( 6.0266 .79) IT .184 (-4.821 ( 16.05) SW .220 ( 4.54) CO .782 ( 18.088 ( 2.0033 .10 1.199 ( 2.26 1.13 1.856 (-1.217 (10.638 ( 5.84 2.30) .54) NE .258 ( 7.16) NZ .36) .0090 .948 ( 25.24) .65) – – – – -.033 ( 3.0428 DW 2.38) .060 (-0.501 ( 4.97 1.0202 .493 ( 5.79) SO .78) GR .90 a Variable lagged once.42) .0556 .028 ( 1.602 ( 6.0147 .53) .877 ( 33.24) a .583 ( 4.81) FR .06) .531 ( 5.0085 . .025 ( 2.0408 .38) – – – – -.45) – -.73) UK .181 ( 4.83 2.390 (-1.78) FI .012 ( 0.67) .92) .92 1.591 ( 6.063 ( 3.02) PO .73) .101 ( 3.5a log P Y = a1 + a2 log P Y−1 + a3 log P M + a4 log W +a5 ZZ + a6 J J S a3 a4 a5 a6 ρ a .24 1.89) – – – a -.204 ( 3.086 ( 1.090 ( 1.64 2.67) – -.78) KO .90) .076 ( 4.132 ( 1.0053 .

5 The .007 .000 .30 (5) 2.00 3.00 2.992 .25 (4) 3.408 .03 6.150 IR .002 . The coefficient estimate for RS was -38.41 (5) 1.634 FR .000 .276 .606 .003 MA .000 . both were used.787 NO .57 5.003 .001 .138 .134 .394 .076 GR .000 .562 .238 .93 (5) 1.00 0.236 .023 . with a t-statistic of -1.088 .3a for CA.000 RHO+ p-val .00 2.00 4.489 TH .000 .03.298 NZ .013 . This is an important area for future work.00 ∗ 8.000 Leads p-val .27 (4) 1.000 .000 .292 .00 ∗ 8.50 (5) 1.03.203 .001 .74 (4) 1.000 .75 ∗ 10.000 .001 .000 .000 .057 . The main problem.000 .000 .23 (3) 1.627 .116 .041 .836 .69 ∗ 12.66 ∗ 27.027 T p-val .460 – – . which seemed to exist for any specification tried.587 .442 .92 (5) 2.006 .000 .254 .00 ∗ 18.074 .00 1.90 (5) 1. The overall results for equation 3 are thus weak in that the results for over half of the countries did not appear sensible.263 .001 .470 JA .34 (3) 1.15 (4) 1. there still seems to be a substantial amount of simultaneity bias.92 ∗ 8.083 .00 2.098 PO .107 .000 .45 (3) 1.00 ∗ 9.002 Stability χ 2 (df) λ ∗ 16.286 .793 . is that the coefficient estimate of the current output term is too large.022 .256 JO .290 .000 .42 (5) 1.00 1.279 .747 .000 .000 .015 .98 (4) 1.870 .00 5.002 .04 (6) 2.91 (4) 1.039 .66 (5) 2.052 UK .206 .136 – .053 .078 .949 SP .001 .764 ST .398 FI .000 .070 .95 (5) 1.534 .858 .786 .023 .000 .93 (6) 2.981 AS .51 ∗ 20.091 .398 .627 .000 .49 Sample (6) 4.001 CO .03 (6) 1.000 NE .99 (3) 1. coefficient estimate for RB is presented in Table 6.035 .000 .538 .195 .050 .00 ∗ 6.000 .339 .146 SW .891 IT .082 DE .00 1.382 .284 .281 .042 .000 ∗ Significant Lags p-val .409 .002 .00 ∗ 7.78 (4) 1.046 .09 (3) 4.088 KO .000 .117 .005 .126 .27 (6) 3.21 ∗ 8.00 2.164 .39 1966:1–1992:3 1967:3–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1971:1–1991:4 1966:1–1992:3 1976:1–1991:4 1971:1–1992:2 1962:1–1991:2 1964:1–1991:4 1969–1990 1969–1990 1974–1990 1969–1990 1963–1990 1969–1990 1962–1990 1969–1990 1962–1990 1972–1991 1971–1991 1965–1990 1972–1987 1972–1991 1962–1991 1962–1990 at the one percent level.000 .004 .32 (5) 1.036 PH .6.000 .17 ∗ 20.600 PA .98 (4) 1.114 . Even though the 2SLS technique is used.006 .13 (4) 1.098 . p-val .573 .5b Test Results for Equation 5 Level p-val Quarterly CA .935 .486 .423 SO .219 .918 .00 5.000 .971 .98 2.000 .000 Annual BE .128 .172 .128 .000 .173 – – – – – – – Chg.004 .171 – .036 .100 SY .001 .000 AU .089 GE .4 EQUATION 3: I : FIXED INVESTMENT 155 Table 6.263 .194 .322 .603 .035 .891 .028 .5 The reason that equation 3 was estimated for only 14 countries is that the results for the other countries were not good.214 .

and β are presented in equations 5. 1 − a2 ) ranges from . In particular.333 to . the wage rate. and so with hindsight the construction of the excess capital variable for each country that was described in Chapter 3 was not needed. and 3 fail the leads test. Equation 4 does well in the tests in Table 6.22 in Chapter 5.6 Equation 5: P Y : Price Deflator Equation 5 explains the GDP price deflator.937 and α ranges from . One. The parameters λ. It is the same as equation 11 for the US model. The equation fails the stability test in 10 of the 24 cases.4b except for the stability test. the coefficient estimates of equation 4 are consistent with the view that firms smooth production relative to sales. and the lagged stock of inventories. is the percentage gap between potential and actual output ((Y S − Y )/Y S). the current level of sales. and a demand pressure variable.19–5.4a. denoted J J S.. The same tests were performed for equation 5 as were performed for equation 10 in the US model. It includes as explanatory variables the lagged level of production. Also presented in the table are the implied values of α and β. denoted ZZ. Four of the 24 equations fail the lags test.e.531 to . which is equation 5. the level . and so these results add support to the production smoothing hypothesis.707 and α was . Data permitting. and the other. As was the case for equation 11 in the US model. α and λ are adjustment parameters. For the United States λ was . It is the same as equation 10 for the US model.473.156 6 ROW STOCHASTIC EQUATIONS The specification of equation 3 that was finally chosen does not use excess capital as an explanatory variable.553. the price of imports.829 and α ranges from . 6.21. For the annual countries λ ranges from . is the ratio of jobs per capita to its peak to peak interpolation (J J /J J P ). For the quarterly countries λ (i. 6. The led value of sales was used for the leads test. The estimate of the constant term is not presented in Table 6. 3 fail the RHO+ test. α.050 to .680. The estimate of λ is one minus a2 in the table. It includes as explanatory variables the lagged price level. two demand pressure variables were tried per country.5 Equation 4: Y : Production Equation 4 explains the level of production. 3 fail the T test.100 to .

The demand pressure variables were not included in 13 of the 29 cases because they did not have the expected sign. and so this hypothesis continues its winning ways. It is the same as equation 9 for the US model.7 Equation 6: M1: Money6 Equation 6 explains the per capita demand for money.6. The level specification is rejected over the more general specification in only 6 of the 29 cases. As was the case for the US results. The wage rate appears in 18 equations. 6. The estimate of the constant term is not presented in Table 6. on the other hand. and of these 18 equations. and 8 fail the stability test. usually with very large χ 2 values. Seven of the 29 equations fail the lags test. none fail the RHO+ test.7 EQUATION 6: M1: MONEY 157 specification was tested against the more general specification and the change specification was tested against the more general specification. Table 6.) The results for the demand pressure variables are thus not as strong as the results for import prices. is rejected in 24 of the 29 cases. Only 1 of the 19 equations fails the lags test. The change specification. The same nominal versus real adjustment specifications were tested here as were tested for US equation 9 (and for the US equations 17 and 26). The estimates in Table 6. Equation 5 does fairly well in the tests in Table 6.5b except for the stability test and possibly the T test. the change specification is strongly rejected by the data. and income. and the results in this section are essentially an update of these earlier results. The results in the table show that the price of imports is significant in most of the equations. it was almost always insignificant.5a. Equation 6 includes as explanatory variables one of the two lagged money variables. The nominal versus real (NvsR) test results in the table simply show that adding the lagged money variable that was not chosen for demand equations are estimated in Fair (1987) for 27 countries. 4 fail the T test. and 14 fail the stability test. only 3 fail the leads test. 4 fail the RHO+ test. (When a demand pressure variable had the wrong sign. The led value of the wage rate was used for the leads test. Import prices thus appear to have important effects on domestic prices for most countries.6b shows that the equation does well in the tests.6a show that the nominal adjustment specification won in 13 of the 19 cases. 6 Money . depending on which adjustment specification won. 10 fail the T test. the short term interest rate.

16) .98 1.0636 .55) .80) .851 (12.69) GE -.917 (53.40) PO .0300 .034 ( 1.211 ( -1.0117 (-3.0046 (-2.05 1.0010 (-0.129 ( 2.46) – .806 ( 1.18) .151 ( 5.16 1.95) ρ – -.35) .382 ( 1.42) – – a3 .87) – – – – – .65) NE -.24) – .0043 (-6.80) JA -.290 (-1.74) -.0472 .0238 .35) .0146 .087 ( -0.08 2.112 ( 5.80) IT -.620 ( 4.106 ( -2.87) AS -.21) AU .53) – – .804 ( -4.060 ( 0.0572 (-5.225 ( 1.42) .0045 (-5.191 ( 1.89) -.0756 DW 1.43) .183 ( 1.900 (30.12) FI -1.0118 (-3.08 1.545 ( 1.53) VE .645 ( 7.89) – .65) -.75) .074 ( 0.0314 .069 ( 3.0065 (-4.731 ( 3.6a log[M1/(P OP · P Y )] = a1 + a2 log[M1/(P OP · P Y )]−1 +a3 log[M1−1 /(P OP−1 · P Y )] + a4 RS + a5 log(Y /P OP ) a1 Quarterly CA -.22) -.84) .0097 .20) a -.0023 (-1.690 ( 5.740 (12.86) -.0220 .38 1.68) SO -.576 ( -3.19) .0055 (-4.495 ( 2.32) a4 -.92 2.0262 .98 1.297 ( 2.12 2.845 (19.44) .98 2.37) .84) Annual BE .55) – – .654 ( 7.22) PH 1.60) .35) a -.77 2.61) SW .0049 (-8.0065 (-1.0067 (-3.53) .09) ID -.879 ( -2.0218 .0049 (-1.93) – -.27 2.881 ( -3.0241 .69) a -.209 (-2.28) – – – – – – SE .25 1.069 ( 9.593 ( 4.0041 (-3.34) – -.57) .0112 (-3.66) – – .669 ( 9. the final specification does not produce a significant increase in explanatory power.317 ( -0. .559 ( 3.0471 .68 1.344 ( 2.914 (55.06 1.161 ( -2.65) FR .445 ( 3.10) .05) -.0155 .49) – .06) a Variable a2 – .48) – .75) .70) -.0631 .27) .06) -.92) – -.0357 .379 ( -3.17) PA 6.13) -.074 ( 1.210 ( 4.32) – .74) -.627 ( -2.49) – .67) – -.90) UK -.05) .23 1.92 2.399 ( 1.632 ( 6.0168 .840 ( 2.83) -.16 lagged once.17) .748 ( 5.158 6 ROW STOCHASTIC EQUATIONS Table 6.06) DE -1.75) – .95 1.223 ( 2.86) .915 ( 4.223 ( 3.0021 (-1.573 ( 3.47) .221 ( 0.0342 .68) a5 .47) .275 (-2.1027 .762 (10.885 (14.33) -.62) .930 (37.916 (24.20) .011 ( 0.265 ( 2.213 (-1.

354 .096 PA .172 . and it is similar to equation 30 in the US model.03 3.996 Stability AP (df) λ 5.093 .059 .100 .696 T p-val . short term interest rate.09 (4) 1.052 . the two asset variables were included on the view that monetary authorities may be influenced in their policy by the status of their balance of payments.48 (4) 1.88 (4) 1.00 ∗ 7.018 .00 ∗ 7.215 .041 .69 (5) 2. 4) the first two lagged values of the asset variable for the quarterly countries and the current and one year lagged value of the asset variable for the annual countries.00 4.212 .S.17 (4) 1.376 .8 Equation 7: RS: Short Term Interest Rate Equation 7 explains the short term (three month) interest rate.189 PO .080 SO .112 . and so putting in the two asset variables is similar to putting in the balance of payments.20 (5) 1.119 .000 .21 2.45 (3) 1.00 6. It is interpreted as the interest rate reaction function of each country’s monetary authority.675 . Similarly.048 .013 ID .01 ∗ 7.710 .299 DE .52 (5) 1.40 ∗ 11.580 .390 .500 IT .8 EQUATION 7: RS: SHORT TERM INTEREST RATE Table 6.874 .69 ∗ 25.811 Annual BE .111 RHO+ p-val .217 .706 GE .893 .258 . 2) the two demand pressure variables.212 .333 .87 (4) 1.242 .154 AU .00 at the one percent level.210 .079 .302 .060 .055 .664 . interest rate was included on the view that some monetary authorities’ decisions may be influenced by the Fed’s decisions.839 . .S.47 (4) 1.76 (5) 2. The change in the asset variable is highly correlated with the balance of payments on current account.191 UK .00 4. The explanatory variables that were tried (as possibly influencing the monetary authority’s interest rate decision) are 1) the rate of inflation.92 3.51 ∗ 17.31 ∗ 10.036 .281 . 3) the lagged percentage growth of the money supply.00 5.262 .36 (5) 3.225 .57 (5) 2.017 .970 VE .427 .00 2. The U.011 .00 2.075 PH . and 5) the U.16 (3) 2.6.22 Sample 1968:1–1992:3 1967:3–1992:3 1971:1–1991:2 1979:1–1992:2 1969:1–1990:4 1972:1–1991:4 1978:2–1991:4 1966:1–1992:3 1976:1–1990:4 1971:1–1992:2 1962:1–1991:2 1969–1990 1969–1990 1971–1990 1962–1990 1963–1991 1962–1989 1972–1991 1962–1991 (4) 1.032 AS .207 FR .37 (5) 3.99 (4) 4.41 ∗ 11.063 .424 .44 (4) 3.641 FI .955 .001 .066 NE .139 .314 .000 .199 .6b Test Results for Equation 6 NvsR p-val Quarterly CA .00 2.001 .918 . 6.002 .295 SW .34 (4) 1.379 .266 JA .253 .058 .748 ∗ Significant 159 Lags p-val .

061 ( 1.49) -13.62) KO .62 ( 0.19 1.2926 2.748 ( 13.84 1.292 ( 2.31) -25.53) .111 ( 1.79 ( 1.17) – – 16.23) .71) .89) – -34.49 ( 2.333 ( 4.266 ( 1.65) -21.737 ( 10.028 ( 0.76) NZ .04) .68) -25.81 1.19) .70) VE .7a RS = a1 + a2 RS−1 + a3 P CP Y + a4 (ZZorJ J S) + a5 P CM1−1 +a6 [A/(P Y · Y S)]−1 + a7 [A/(P Y · Y S)]−2 + a8 RSU S a2 Quarterly CA .26 (-1.53 (-0.94) SO .48 (-1.143 ( 2.0927 1.17 2.781 ( 17.66) – -8.28) .89 1.19 2.13) – .42) -17.56) – – a 16.44) – .07) -31.00 (-5.75) -13.81) – – .68 (-2.05) SW .00 (-2.94) – – – – – – rather than ZZ.92) – -26.120 ( 1.92) -8.03) – – -20.836 ( 13.20) .123 ( 1.0078 2.47 2.04) UK .97 (-3.68 1.024 ( 3.92) – -23.87) Annual BE .47) 24.09 ( 0.012 ( 1.95 1.91 1.35) FI .00) PO .43 (-4.34 (-1.137 ( 1.33 (-3.5174 1.345 ( 2.7961 1.87 1.16) – – – 42.51) .92) .67) IR – ( .195 ( 1.57 (-2.59) ST .7804 1.9518 1.03) NE .43) .99 2.30) .17 ( 0. .13 2.72) NO .072 ( 1.65) -18.8892 .39) 13.564 ( 3.143 ( 2.9740 1.78 2.13 1.0748 .73) 19.40) IT .59) 19.60) JA .713 ( 8.22) 11.759 ( 4.61 ( 1.67) – .33) SE .13) PH .43 (-1.26) -30.2022 .42) .84) aJ J S a3 – .857 ( 7.34 2.50) – 6.58) 11.786 ( 11.8879 .1642 1.154 ( 3.706 ( 9.52 (-2.099 ( 0.951 ( 25.782 ( 4.0225 2.04 (-1.494 ( 2.38) 26.97 ( 2.849 ( 17.195 ( 0.44 (-1.93) AS .262 ( 3.7352 .85) 7.0081 2.81 (-0.94 ( 1.41 ( 0. P CP Y = 100[(P Y /P Y−1 )4 − 1]. P CM1 = 100[(M1/M1−1 )4 − 1].1857 DW 2.7384 .96) -12.10) – .35) .51) – a7 5.43 ( 1.86) .154 ( 3.491 ( 1.94 (-3.43) .774 ( 6.74 ( 0.21 ( 1.98) .590 ( 4.70) -32.28) .05 1.534 ( 5.67 ( 1.88) 28.035 ( 0.248 ( 3.91) PA .50 ( 2.52 1.901 ( 11.171 ( .55) AU .406 ( 3.27) .18 2.119 ( 2.62) .32) .160 6 ROW STOCHASTIC EQUATIONS Table 6.70) – – – – – .69 ( 3.8364 5.15) FR .18) .130 ( 0.169 ( 1.76 (-2.46) – -23.95) – -28.58 a6 -8.81 (-1.71 2.62 ( 1.756 ( 4.89) GE .99 2.764 ( 5.6879 .09) -21.94) 21.590 ( 6.426 ( 1.42) – -7.34) a4 a 18.1976 .01) – a8 .26) .89) – 5.71 2.13) DE .332 ( 1.69 (-3.47) .18 ( 2.16) .555 ( 2.077 ( 3.4647 2.18) – – – -37.74 (-0.83 (-1.49 (-0.35) .8082 3.70) .346 ( 2.700 ( 2.

398 .020 . Four equations included the lagged money growth variable. a demand pressure variable in 10 cases.96 (6) 1.63 (5) 1. The estimates and t-statistics are: IT: .051 .173 .079 JA .550 AS . and so there is little evidence in favor of this variable.067 Stability AP (df) λ ∗ 11. and the serial correlation coefficient are not included in Table 6.89 (5) 1. The estimates and t-statistics are: CA: .392 UK .06 (6) 1.040 GE .461 .767 .631 NE .122 .339 (2.7a show that the inflation rate is included in 13 of the 23 cases.306 .438 (2.90 Sample (7) 1.15 (7) 1.00 3.771 T p-val . rate in 21 cases.90 (5) 1.6.02 (7) 1.005 .42 (4) 1.155 (1.55 (6) 1. PA: .031 (1.35 5. PH: . The monetary authorities of other countries do not appear to be influenced in their setting of interest rates by the lagged growth of the money supply.85 (4) 1.132 .49 (5) 1.697 PA . IT: .91 (4) 1.019 AU .002 .267 .75 1.331 .81 ∗ 11.011 .78 6.644 (4.193 .81). The lagged money growth variable.815 NZ .074 FI .8 EQUATION 7: RS: SHORT TERM INTEREST RATE Table 6.199 .7b Test Results for Equation 7 Lags p-val Quarterly CA .968 SO .022 IR .491 .19 (3) 1. and the balance of payments.007 .81 8.16).083 .408 (2.396 .075 . is not significant in any of the 4 cases it is included (see footnote 7).171 KO .301 . There is thus evidence that monetary authorities are influenced by inflation.567 .202 IT . JA: .26).63 (5) 1.100 DE .64 4.050 PO .612 .116 .957 NO . FR: .000 VE .110 ST .056 (1.043 ∗ Significant 161 RHO+ p-val .00 3.S.863 .00 2.90 (7) 1.183 .61).851 .29 (6) 1.00 2.571 . VE: .58 3.7a because of space constraints.66).56 (5) 1.120 (1.309 Annual BE . SO: .00 3. demand pressure.66).00 5.34 (6) 1. .707 (7.05 1.00 1.373 . The estimates of the constant term.023 .7 The results in Table 6.549 .999 .007 .000 PH .51 (8) 1.07).00 1972:2–1992:3 1972:2–1992:3 1972:2–1991:2 1972:2–1992:2 1972:2–1991:4 1972:2–1991:4 1978:2–1991:4 1972:2–1991:4 1972:2–1992:3 1976:1–1991:4 1972:2–1992:2 1972:2–1991:2 1972:2–1991:4 1972–1990 1972–1990 1974–1990 1972–1990 1972–1990 1972–1990 1972–1990 1972–1991 1972–1991 1972–1991 at the one percent level.51).060 .049 .112 . the coefficient of the lagged money growth variable.00 1. The signs of the coefficient estimates of the asset variables (negative for the first and 7 Five equations were estimated under the assumption of a first order autoregressive error.046 .561 SW .00 1.00 ∗ 8.15 (7) 1.33).229 .151 .65 2.11 (4) 1.007 .58 6.004 .027 FR . on the other hand. the asset variables in 15 cases. and the U.998 .00 ∗ 19.47 (6) 1.010 .00 4.

95 2.163 ( 1.932 (32.82) .29) PA .101 ( 1.64) .99) -.35) .96) – – – – – – – – ρ – -.972 0.295 ( 2.3845 .76) UK .32) .66) NE .865 0.905 ( 1.44) IR .66) – .953 0.074 (-1.484 ( 5.5561 .51 1.65) .090 (-1.49) -.332 ( 3.88) .299 ( 3. positive for the second) suggest that an increase (decrease) in the balance of payments has a negative (positive) effect on the interest rate target of the monetary authority.955 (14.472 ( 5.91) .18) .946 0.72) .01) .78) -.4621 .514 ( 4.94 2.28) .325 ( 1.065 ( .722 ( 9.901 (25.108 ( 2.781 0.5498 .07) -.25) AU .63) a4 -.4579 .078 ( .944 0.10) JA .405 ( 5.96) .108 ( -.05 1.6368 1.549 ( 3.738 0.14) -.376 (-4.48) .91) -.86) -.91) . .76) SW .27) .46 2.42) .38) .75) .25) – – – – – – – – -.4599 .142 ( 2.054 (-0.161 ( 2.844 ( 2.73) Annuala BE 1.83) .144 ( 1.111 (-1.33) .435 ( 4.410 (-1.151 ( 1.496 ( 7.344 ( -1.208 ( 3.87) NZ .99) SO .153 ( 1.954 0.934 0.65) SE .2500 .195 ( 2.29) .929 (25.204 ( 0.340 ( 2.4177 .11) .0446 .74) .37) GE .87) AS .98) NO -.311 ( 5.916 (31.86) .25 annual countries a4 is zero and RS−1 rather than RS−2 is subtracted from the other variables.71) .919 ( 6.849 0.35) .69) .2338 .112 ( .162 6 ROW STOCHASTIC EQUATIONS Table 6.052 ( .957 (50.116 ( 1.864 (16.53 1.8289 1.4154 .232 ( 5.92 2.934 0.14) .84) – .168 (-1.3445 .32 2.17 2.00 1.34) .29) PO .5894 .97) .470 ( 7.67 2.918 0.03 1.01) .287 ( 5.085 (-0.79 1.51 1.48) .20) .95 1.926 0.706 (11.164 (-1.971 0.7544 R2 0.08) .4322 .53) IT -.36) -.843 0.898 (29.894 ( 6.813 (10.840 (12.91) a3 .332 ( 4.535 ( 4.55) -.740 0.75) -.30) .185 (-2.106 ( 1.573 ( 4.97 1.00 2.50) .13) DE .01 2.803 DW 2.945 (24.90) .91) ST .947 0.009 ( .20) a For a2 .429 (-5.962 (38.432 ( 2.14) -.8a RB − RS−2 = a1 + a2 (RB−1 − RS−2 ) + a3 (RS − RS−2 ) +a4 (RS−1 − RS−2 ) a1 Quarterly CA .75) .21) FR .6569 1.252 (-3.405 ( 2.3988 .00) .636 (-3.646 ( 5.972 0.336 ( 2.702 (-4.143 ( 2.

19 (3) 1.079 IR .26 (5) 1. 6.581 .00 4.774 PO .386 .253 .78 (4) 2.061 .177 .700 .714 NO .00 .638 . For the annual countries the explanatory variables include the lagged dependent variable and the current and one lagged short rates.068 .020 .447 Annual BE .16 (3) 1.718 .00 a RS −2 added for quarterly countries.92 3. RS−1 added for annual countries.448 .57 2.333 .12 (5) 2.041 RHO+ p-val .705 .316 AU .49 (5) 2.21 3. Equation 7 does well in the tests.970 .000 .042 .626 .439 . and 4 fail the stability test.210 .048 . namely that the coefficients on the short rate sum to one in the long run.727 .483 .00 3.98 (5) 3. The test results in Table 6.8b Test Results for Equation 8 Restra p-val Quarterly CA .805 .54 (4) 4.474 . The equation does very well in the other tests.427 .853 .667 .308 PA . 4 fail the RHO=4 test.46 (3) 1.734 .256 .093 .552 . For the quarterly countries the explanatory variables include the lagged dependent variable and the current and two lagged short rates.25 2.065 .567 .335 T p-val .66 (4) 1.85 (4) 1.033 .83 ∗ 8. Two of the 19 equations fail the lags test.391 .066 .028 JA . Two of the 23 equations fail the lags test. The same restriction was imposed on equation 8 as was imposed on equations 23 and 24.291 .69 1.00 6.093 .9 EQUATION 8: RB: LONG TERM INTEREST RATE Table 6.732 .076 .000 .382 FR .270 .893 . 2 fail the T test.303 GE .87 163 Sample 1966:1–1992:3 1967:3–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1971:1–1991:4 1966:1–1992:3 1971:1–1992:2 1962:1–1991:2 1969–1990 1969–1990 1974–1990 1969–1990 1969–1990 1962–1990 1962–1990 1972–1991 (3) 1.703 .98 (3) 1.8b show that the restriction that the coefficients sum to one in the long run is supported in 17 of the 19 cases.845 .006 .315 NE .218 .00 . and 2 fail the stability test.018 .987 .013 .011 . 3 fail the T test.007 UK .000 . ∗ Significant at the one percent level.761 .551 .60 (3) 1.75 (5) 2.88 (5) 2.135 .247 .924 .03 6.00 1. and it is not .66 ∗ 9.038 .044 .078 SO .429 .15 (5) 2.59 (4) 3.007 .364 .03 0.005 IT . The led value of the short term interest rate was used for the leads test.221 .656 .693 SW . 1 fails the RHO=4 test.854 .396 AS . It is the same as equations 23 and 24 in the US model.264 .6.00 2.00 .17 5.490 Leads p-val .341 .834 Lags p-val .236 DE .070 .9 Equation 8: RB: Long Term Interest Rate Equation 8 explains the long term interest rate.281 ST .015 NZ .054 .144 .575 .082 .84 (3) 1.294 Stability AP (df) λ 3.

and RSU S is the U. Consider first the non European countries.164 6 ROW STOCHASTIC EQUATIONS significant at the one percent level in any of the 19 cases. my experience with term structure equations like equation 8 is that they are quite stable and reliable. Korea. The second set includes Canada. The coefficient on the relative price level is constrained to be one. which 8 RS and RSU S are divided by 100 because they are in percentage points rather than percents. Two types of countries are assumed for the estimation. and the Philippines. As noted in Chapter 5. South Africa.S.S.10 Equation 9 E: Exchange Rate Equation 9 explains the country’s exchange rate.8b support. Japan. dollar.7. and so α1 is multiplied by . E ∗ = e α0 ( 1 + RS/100 . The first set includes all the European countries. and an increase in E is a depreciation of the country’s currency relative to the dollar. which the results in Table 6. which means that in the long run the real exchange rate is assumed merely to fluctuate as the relative interest rate fluctuates. India.7 is a partial adjustment equation. Equation 6. P Y U S is the U. the . New Zealand. P Y /P Y U S. 6. For the annual countries. E ∗ .25 is not used. Also.6 states that the long run exchange rate. .6 and the discussion of reaction functions in Section 2. E. A country’s exchange rate is relative to the U. Jordan. Germany is taken to be the leader among the European countries in this respect. The theory behind the specification of this equation is discussed in Chapter 2. P Y is the country’s domestic price deflator. See in particular the discussion of the experiments in Section 2. domestic price deflator (denoted GDP D in the US model).8 Equation 6.7) E is the exchange rate. (1 + RS/100)/(1 + RSU S/100).25α1 P Y ( ) ) 1 + RSU S/100 P YUS E∗ λ E =( ) E−1 E−1 (6.S.2.6) (6. The exchange rate for these countries is based on the following two equations. Equation 9 is interpreted as an exchange rate reaction function. Australia. short term interest rate (denoted simply RS in the US model). The first are those countries whose exchange rate is assumed to be at least partly tied to the German exchange rate. the interest rates are at annual rates. RS is the country’s short term interest rate. and the relative interest rate. The second are those whose exchange rate is assumed not to be tied to the German rate. depends on the relative price level.2.25 to put the rates at quarterly rates.

25) log[(1 + RS/100)/(1 + RSU S/100)] + λ[log(P Y/P Y U S) − log E−1 ] (6. the use of the relative price level in it is in effect based on the assumption that the monetary authority goes along with the forces on the exchange rate and allows it to change in the long run as the relative price level changes.6 is interpreted as an exchange rate reaction function. δ is the weight on the German exchange rate. the use of the relative interest rate in equation 6. Similarly.6 is consistent with the theoretical model.6 is consistent with the theoretical model in Chapter 2. where a fall in the relative interest rate led to a depreciation.6. The use of the relative price level in equation 6. one is back to the case of the non European countries. Consider now the European countries (except Germany).2. Because equation 6. then the exchange rate of the country relative to the German rate fluctuates in the long run merely as the relative interest rate fluctuates. Equation 6.9b.) In other words. If δ is zero. Equations 6. For δ less than .25α1 P Y 1−δ ) ) EGE δ ( 1 + RSU S/100 P YUS (6. (See experiments 1 and 2 in Section 2. In this model a positive price shock led to a depreciation of the exchange rate.6.8) The restriction that the coefficient of the relative price term is one can be tested by adding log E−1 to equation 6. The exchange rate for these countries is based on the adjustment equation 6.7 and on the following equation: E ∗ = e α0 ( 1 + RS/100 . this variable should have a nonzero coefficient.2.9 differs from equation 6. where the weights sum to one.6 in that the relative price term (with the coefficient of one) is replaced with a weighted average of the relative price term and the German exchange rate. If δ is one.6 is that the monetary authority goes along with the forces on the exchange rate from the relative interest rate change. there are forces in the theoretical model that put downward pressure on a country’s currency when there is a relative increase in the country’s price level. This is one of the tests performed in Table 6. If the coefficient is other than one. (See experiments 3 and 4 in Section 2.9) EGE is the German exchange rate.6.7 imply that log(E/E−1 ) = λα0 + λα1 (.10 EQUATION 9 E: EXCHANGE RATE 165 says that the actual exchange rate adjusts λ percent of the way to the long run exchange rate each period. the assumption in equation 6.8.) Again.6 and 6.

03 2.965 (75.846 (13.973 (15.38) .982 (88.94 1.0726 .236 ( 2.856 ( 4.41) .0757 .872 (17.073 ( 9.24 1.991 (42.39) λ .50) – – – – -1.770 ( 3.413 ( 3.96 1.955 (14.21) NZ .798 .1048 .25) Annual BE 3.49) λδ – – .707 ( 1.0259 .97) – .359 ( -0.930 (23.72 1.23) .81 1.25) .206 (-0.88 2.21) .210 ( 0.0436 .96 2.23) .14) .664 ( 6.666 ( 6.0302 .989 1.85) .69 δ – – 1.75 1.0706 .02) .51) .25) log[(1 + RS/100)/(1 + RSU S/100)] a1 Quarterly CA .940 (17.0919 .154 ( -1.07) .71) 1.272 ( 1.03 1.001 .0241 .950 (41.0335 .932 (18.559 ( 3.51) .71 1.114 ( 1.35) .0483 .166 6 ROW STOCHASTIC EQUATIONS Table 6.81 1.34) .11) .633 ( 4.87) -.22) GE -.888 (25.003 (14.14) .0051 .715 .67) .679 ( 3.31 1.672 ( 2.92 2.760 ( 5.91) .0378 .29) -1.002 ( -0.23) .0144 .41) .088 ( 1.78) UK 3.826 – – – 1.70) .864 ( 3.033 ( 2.0685 DW 1.13) .95) .0052 .945 (14.40) .950 (28.95) .631 ( -2.23 1.010 (69.92 1.557 (-1.74) .23) PH -2.83) DE 1.771 ( -3.67) .57) .415 ( 6.052 ( 1.580 ( -1.63) PO 4.89) .098 ( 2.21) – – – – – -.52) .011 ( 1.674 ( 5.892 ( 22.625 ( 7.692 ( 7.28) JO -.76) .81) .60 2.37) .22) .0270 .71) 1.63 1.40) NE .29) KO -.66) .45) IR 5.35) .964 (50.25) .966 (27.56) .53) ST -.68) .068 – – – – .06) .0506 .0514 .98) .35) ID -1.35) – .606 ( 4.020 .11) AU 1.191 ( 0.0422 .60) .84) .970 – .63) .093 ( -1.039 ( 7.144 ( 1.0293 .758 ( 6.948 (16.166 ( 1.754 (-2.10) SE .552 ( 7.090 ( 1.880 (12.29) .975 (15.07 1.874 ( 8.365 ( 1.371 ( 3.514 ( 2.651 (11.32) FI 5.0499 .047 ( 0.425 (-0.37 1.998 ( 9.99) – – – – – – – – – ρ .36 1.147 1.0253 .16) .03) .814 ( 5.406 ( 10.117 (-0.08) AS .007 ( 0.232 1.60) – -2.71) .846 1.58) JA -.002 .06) .932 (31.64) – – – 1.98) SO .55) IT 4.823 (-1.844 ( 8.971 (94.52) SP 3.0213 .699 ( 7.52) – – – – a4 -.97) .16) .07) .404 ( 2.58 2.607 ( 7.94) FR 1.92 0.932 1.039 ( 1.01) .93) .28) SW 1.936 ( 0.0436 .175 ( 0.37) .930 (22.00 2.09) GR 10.79) .021 .970 .010 (14.821 (16.990 (60.9a log E = a1 + λ[log(P Y /P Y U S) − log E−1 ] + λδ[log EGE − log(P Y /P Y U S)] +a4 (.19) NO .335 ( 2.626 ( 8.712 ( 4.

78 ∗ 14. Under the assumption that Germany is the dominant country in the EMS.009 NE .07 (5) 1.824 .026 .186 AU .036 Stability AP (df) λ 0.117 Lags p-val .00 3.10) . and this is the reason for the use of the German rate in equation 6.005 ID .17 (4) 1.695 GR .461 FR .9 and 6.033 .00 3.095 .267 FI .58 (4) 1.00 2.010 SP .896 .718 JA .111 .000 .75 (3) 1.390 .083 .595 .67 (4) 1.00 1.279 .999 DE .139 .41 1.81 (4) 1.45 (4) 1.299 SO .61 (5) 1.004 .772 .00 2.775 Annual BE .236 .654 IT .579 AS .017 .563 IR .85 (3) 1.059 .81 3.16 (4) 1.00 3.11 (3) 1.594 NO . ∗ Significant at the one percent level. the relative interest rate.00 (4) 1.693 JO .631 UK .6.259 .577 .102 .60 (4) 1.160 .00 0.204 .150 .018 .94 (4) 1.097 . the German rate will pick up some of the effects of the EMS agreement.65 3.041 PH .018 .663 .35 1.996 T p-val .72 (3) 1.00 5.976 .000 .14 (4) 1.9b Test Results for Equation 9 Restra p-val Quarterly CA .369 .7 imply that log(E/E−1 ) = λα0 + λα1 (.00 2.73 (4) 1.024 .587 .160 .396 .784 .170 .00 2.40 (5) 1.002 .044 .404 .11 (4) 1.466 .00 1.574 .984 .00 Sample 167 1972:2–1992:3 1972:2–1992:3 1972:2–1991:2 1972:2–1992:2 1972:2–1991:4 1972:2–1991:4 1978:2–1991:4 1972:2–1991:4 1972:2–1992:3 1976:1–1991:4 1972:2–1992:2 1981:1–1991:2 1972–1990 1972–1990 1974–1990 1972–1990 1972–1990 1972–1990 1972–1990 1972–1990 1972–1990 1972–1991 1972–1989 1972–1991 a log E −1 added.001 .266 .039 .96 (3) 1.838 .153 SW .007 .550 .804 .654 .9.793 .00 1.983 .572 .00 (4) 1.744 .906 .914 .93 (4) 1. Equations 6.00 ∗ 12.976 RHO+ p-val .58 ∗ 12.033 PO .136 .822 .25) log[(1 + RS/100)/(1 + RSU S/100)] +λ[log(P Y/P Y U S)−log E−1 ]+λδ[log EGE −log(P Y /P Y U S)] (6.052 .752 .863 .58 4. The monetary authorities of other European countries may be influenced by the German exchange rate in deciding their own exchange rate targets.75 3.00 3.82 (4) 1. and the German rate fluctuate. the exchange rate fluctuates in the long run as the relative price level.009 .858 .442 .594 .99 (2) 1. This specification can also be looked upon as an attempt to capture some of the effects of the European Monetary System (EMS).458 .011 ST .929 GE .048 – .81 4. one and greater than zero.10 EQUATION 9 E: EXCHANGE RATE Table 6.041 .000 NZ .

for many countries δ is close to one in Table 6. 10 Multicollinearity problems prevented the RHO+ test from being performed for the UK.10 2 fail the T test. and for these countries the exchange rate effectively just follows the German rate in the long run. It will be seen in Chapter 12 that some of the properties of the model are sensitive to the inclusion of the relative interest rate in the exchange rate equations.9a along with the other results. It is included for only 6 countries and is only significant for 1 of these (Austria). and 3 fail the stability test. Regarding δ.10. Two of the countries for which the variable is included are Japan and Germany. but this was not done here. this variable should have a nonzero coefficient. Equation 9 does well in the tests. This means that the adjustment to the long run value is estimated to be very rapid and thus that the exchange rate follows closely the German rate even in the short run. The results do not provide strong support for the use of this variable in the exchange rate equations. There is considerable first order serial correlation in the error terms in the exchange rate equations for most countries.168 6 ROW STOCHASTIC EQUATIONS The restriction that the weights sum to one can be tested by adding log E−1 to equation 6. Two of the 25 equations fail the lags test. This is one of the tests performed in Table 6. This is discussed more in Chapter 12. This is contrary to the case for the European countries (except Germany). Consider first the relative interest rate variable.9a (δ is in fact slightly greater than one in a few cases9 ). say. the estimates of λ are small. the properties that are sensitive to the inclusion must be interpreted with considerable caution.9b. Doing this would have had little effect on the model because the estimates that are greater than one are in fact quite close to one. It is encouraging that so few equations fail the 9 δ could have been constrained to be one when its estimate was greater than one. 3 fail the RHO+ test. where the adjustment to a weighted average of the relative price level and the German exchange rate (with most of the weight on the German rate) is estimated to be quite rapid. Exchange rate equations were estimated for 25 countries. The variable had the wrong sign (and was almost always insignificant) for the other countries. The implied value of δ is presented in Table 6. which means that it takes considerable time for the exchange rate to adjust to. . For many of these countries the estimates of λ are also close to one. If the weights do not sum to one. Given that the relative interest rate is not significant in either the Japanese or German equation. which are important countries in the model. a relative price level change. The restriction discussed above that is tested by adding log E−1 to the equation is only rejected in 4 of the 25 cases. For Germany and for most of the non European countries. and so in this sense the relative interest rate variable is important.

and given this.14! . especially considering that some of the interest rate data are not exactly the right data to use. is 4. The standard error for Japan. The test results suggest that most of the dynamics have been captured and that the equations are fairly stable.6.25 =( ) EE 1 + RSU S/100 In equation 10. and the standard error for Germany is 5.11 Equation 10 F : Forward Rate Equation 10 explains the country’s forward exchange rate. F . the standard errors are roughly in percentage terms. the results in Tables 6.9b do not seem too bad. A better way of examining how well these equations fit is to solve the overall model.25 log(1 + RS/100)/(1 + RSU S/100).9a and 6. many of the key coefficient estimates have t-statistics that are less than two in absolute value. This equation is the estimated arbitrage condition. The key German exchange rate equation passes all the tests. If the arbitrage condition were met exactly. but this is because of the inclusion of the German rate. and this is done in Chapter 9. 6. and although it plays no role in the model. log F is regressed on log EE and . Note the t-statistic for France of 5586. the coefficient estimates for both explanatory variables would be one and the fit would be perfect. The arbitrage condition in this notation is F 1 + RS/100 .11 EQUATION 10 F : FORWARD RATE 169 stability test. However. it is of interest to see how closely the quarterly data on EE. and U SRS match the arbitrage condition. and there is substantial serial correlation of the error terms. Since equation 9 is in log form. The standard errors for a number of the European countries are quite low.10a show that the data are generally consistent with the arbitrage condition.14 percent.83 percent. RS. F . whose rate is not tied to the German rate. The results in Table 6. Exchange rate equations are notoriously hard to estimate.

then its export prices would just be the world prices of the export goods. In practice.12 Equation 11 P X: Export Price Index Equation 11 explains the export price index.62) 1.0105 .148 (-1.902 (10.0056 . If a country is a price taker with respect to its exports.62) . to the export price index.45) NE .0041 R2 .479 ( 5. only some of which are exported.9917 ( 315.75 2.10a log F = a1 log EE + a2 (.75) .168 ( 4.0005 ( 5425.04) UK 1.75) -.946 ( 6.0004 ( 4375.999 . P X.174 ( 6.170 6 ROW STOCHASTIC EQUATIONS Table 6. Export prices are needed when the countries are linked together (see Table B. .82) JA 1.0036 .43) AU 1.9976 ( 471. a country produces many goods.999 .889 ( 9.57 2.23) ST 1.0099 .14) FR 1.25) log[(1 + RS/100)/(1 + RSU S/100)] a1 Quarterly CA .612 ( 4.0071 .77) – SE . where the weights sum to one. P Y . of course.03) AS 1.79) .52) 1.0021 .02 2.06) ρ .616 ( 5.999 .436 ( 4.0069 .11) P W $ is the world price index in dollars.0110 . Equation 6.999 .213 (15.11 thus takes P X to be a weighted average of P Y and the world price index in local currency.0010 ( 807.9894 ( 155.12) 1.27) FI .999 .98) 1.82 1.383 ( 2.618 ( 6.137 ( 1.168 ( 9.53) IT .95 2. To try to capture the in between case where a country has some effect on its export prices but not complete control over every price.31) 1.21 Sample 1972:2–1992:3 1972:2–1992:3 1972:2–1991:2 1972:2–1991:1 1972:2–1991:4 1978:1–1991:4 1978:2–1990:4 1972:2–1991:4 1972:2–1984:4 1976:1–1989:3 1977:1–1992:2 6.61) .999 .323 ( 5.999 .9999 ( 3268. If a country produced only one good.0064 .60) 1.0063 . then the export price would be the domestic price and only one price equation would be needed.998 .09) – – . the following equation is postulated: P X = P Y λ (P W $ · E)1−λ (6.0044 ( 237.05 2.392 ( 3.0004 ( 363.17) 1. and so P W $·E is the world price index in local currency.04 2.78) a2 .999 DW 2. It provides a link from the GDP deflator.17) – . Equation 11 was not estimated for any of the major oil exporting countries.998 .4 in Appendix B).05 1.14) GE 1.0007 ( 5586.08 1.998 .0003 ( 5017.267 ( 8.16 2.06) .54) 1.

) Equation 6. Some of the estimates of λ in Table 6. equation 16 of the US model does not include any demand pressure variables because none were significant. and ZZ).12) The restriction that the weights sum to one and that P W $ and E have the same coefficient (i. The restriction discussed above is rejected in 10 of the 30 cases.6.e. J J S. The equation fails the RHO+ test in 3 cases. Equation 11 does reasonably well in the tests. DE. Multicollinearity problems prevented the stability test from being performed for 5 countries (FR. Equation 16 of the US model included three further lags of the wage rate and price level. 6. For these countries. the restriction is imposed on the coefficients in equation 12 so that the implied real wage equation does not . It includes as explanatory variables the lagged wage rate. there is essentially a one to one link between P Y and P X.12. (See equations L-4 in Table B. therefore. NE. FI. that their product enters the equation) can be tested by adding log P Y and log E to equation 6. the current price level. one of three possible measures of labor market tightness (U R. these variables should be significant. the equation fails the stability test in 4 of them.13 EQUATION 12: W : WAGE RATE 171 and so P W $ was constructed to be net of oil prices. If more disaggregated data were available. Also. which equation 12 does not. and the estimates of the autoregressive parameters are generally large. This is one of the tests performed in Table 6. It is similar to equation 16 for the US model. Of the 25 remaining cases. the lagged price level.11b. and GR). Given the coefficient estimates of equation 5. If this restriction is not met. where some goods’ prices would be strongly influenced by world prices and some would not.11 was estimated in the following form: log P X − log(P W $ · E) = λ[log P Y − log(P W $ · E)] (6.13 Equation 12: W : Wage Rate Equation 12 explains the wage rate. one would want to estimate separate price equations for each good. The same restriction imposed on the price and wage equations in the US model is also imposed here.4. It should be kept in mind that equation 11 is meant only as a rough approximation. and a time trend.11a are close to one (a few are slightly greater than one). Equation 11 was estimated under the assumption of a second order autoregressive error. This type of disaggregation is beyond the scope of the present work.

636 (177.0308 0.17) (-0.11a log P X − log(P W $ · E) = λ[log P Y − log(P W $ · E)] λ ρ1 ρ2 SE R2 Quarterly CA .85 1.217 .68) ( 3.83 1.892 ( 10.38) (11.44) ( 1.42) CO 1.007 .145 -.493 ( 92.480 ( 4.468 .549 .135 .864 .014 .92 2.561 .087 .605 (139.179 .82 1.743 1.88) PA 1.004 .06) IT .62) ( 1.0188 0.1828 0.0203 0.0294 0.884 ( 8.98 2.0444 0.092 .24) ( 0.394 .20) (-2.924 .075 .10) ( 1.891 ( 3.82) SW .31) ( 4.13) (-3.35) ( 0.148 -.846 -.81 1.33) (-1.292 .20 1.24) (11.0139 0.79) ID .03) Annual BE .936 ( 7.95 1.0325 0.78) DW 2.96 2.005 .86) ( 4.626 1.968 ( 17.21) FI .26) (-0.93) ( 8.62) ( 4.69) GE .0817 0.62) (12.83) GR .95 2.0356 0.52) FR .49) AS .744 (153.971 .03) ( 4.85) (-1.266 -.584 ( 10.56 1.86 1.1812 0.976 1.837 -.84 1.43) (13.900 -.877 .963 1.55) ( 4.20 1.062 .57) PO 1.77) ( 7.1378 0.496 .661 .87) MA .267 .36) ( 7.881 ( 7.0325 0.629 ( 85.918 ( 63.591 ( 59.87) (-1.34) (-3.66) ( 4.869 ( 2.11) ( 7.60) AU .049 -.413 1.476 -.0737 0.02) (-1.00 1.551 1.075 .046 .62) (-2.06 1.972 ( 17.817 .476 .0240 0.19) JO .43) ( 5.140 .821 ( 28.59) (-2.64) ( 0.0655 0.07) (-0.1391 0.74) ( 6.167 -.875 ( 7.234 .05 1.90) ( 1.93) ( 6.974 (238.00 1.378 .46) ( 0.94 1.029 .966 -.965 1.039 .667 ( 11.11) SP .214 -.959 .199 -.014 .19) NE .970 ( 7.982 .988 1.0145 0.187 -.90) ( 6.91) ( 4.03 1.42) ( 5.0545 0.58) (-0.346 .64) (-2.35) SY 1.0479 0.92 1.55) ( 0.303 -.0254 0.18) (12.90 2.93) ( 0.84 .380 1.69 1.49) KO .215 .41) SO .0352 0.0184 0.026 1.81) (-1.88) ( 9.49) (-2.12) ( 7.983 ( 23.975 ( 13.98 1.261 -.433 .939 .98 1.819 1.07) NO .93) JA .0167 0.93) ( 5.144 .911 ( 9.307 -.710 1.08) (-2.275 .99) (12.091 1.63) ( 5.14 1.983 .795 ( 69.179 1.262 .601 -.172 6 ROW STOCHASTIC EQUATIONS Table 6.906 ( 3.514 .62) IR .74 1.0361 0.194 .453 .169 -.27) TH 1.0898 0.053 -.945 ( 8.03 2.277 -.16) (-0.300 .422 1.939 ( 10.0154 0.458 .919 .39) NZ 1.0105 0.772 ( 21.129 .58) ( 6.97 2.0108 0.98) (-1.906 -.14) UK .74) PH 1.866 .92) ST .695 .84) DE .354 .71) ( 3.0608 0.61) (-1.015 .27) (-0.

011 IT .00 ∗ 17.567 SW .000 SY .00 2.96 (3) 1.61 (3) 1.091 GR . Sample 173 1969:1–1992:3 1976:1–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1971:1–1991:4 1966:1–1992:3 1976:1–1991:4 1971:1–1992:2 1962:1–1991:2 1964:1–1991:4 1969–1990 1969–1990 1974–1990 1969–1990 1965–1990 1969–1990 1962–1990 1969–1990 1962–1990 1972–1991 1971–1991 1965–1990 1962–1989 1972–1987 1972–1991 1962–1991 1962–1990 a log P Y and log E added.95 (3) 1.147 FI .193 .508 Stability AP (df) λ 3.008 KO .155 NE .) The estimate of the constant term is not presented in Table 6.286 ST .496 . and 5.677 .73 (3) 3.839 .023 .69 (3) 1.00 – 2.57 1.075 RHO+ p-val .295 .583 .912 .00 1. 5.02 (3) 1.11b Test Results for Equation 11 Restra p-val Quarterly CA .000 Annual BE .012 .266 NZ . The test results in Table 6.35.957 SP .919 .44 (3) 1. ∗ Significant at the one percent have the real wage depend on either the nominal wage rate or the price level separately.001 TH .050 ID .138 .635 IR .218 CO .002 .992 .044 .89 (3) 1.60 (3) 1.144 UK .75 (3) 2.00 ∗ 14.609 JA .011 .00 2. 7 fail the .875 .000 FR .54 (3) 1.21 (3) 3.05 – 1.37 in Section 5.885 .517 .00 1.005 MA .56 (3) 5.00 0.11 (3) 1.00 – 1.10 (3) 1. Three of the 18 equations fail the lags test.681 .287 JO .00 level.329 PA .001 SO .57 3.696 .009 PH .00 ∗ 5.897 .03 1.92 – 4. One of the variables appears in 12 of the 18 equations.778 AS .37 (3) 4.51 ∗ 11.6.00 4.509 .69 1.069 .39 (3) 1.66 1.09 (3) 1.522 PO .13 EQUATION 12: W : WAGE RATE Table 6.00 1.86 (3) 2.003 GE .17 4.00 2.70 (3) 1. although in half of the 12 the variable is not significant.00 3.000 AU .524 . The results show that there is a scattering of support for the labor market tightness variables having an effect on the wage rate.92 (3) 1.430 .43 (3) 1.457 .001 .676 DE .12a.95 (3) 2.4.602 NO .97 – 2.95 (3) 1. (See the discussion of equations 5.944 .07 (3) 1.006 .36.12b show that the real wage restriction is rejected in 5 of the 18 cases.

51) 1.08) c -.02 1.07) 1.04) -.0154 .22) AU .0135 .00278 ( 4.56) – – – – -.03) -.20) .74 2.220 -.00101 ( 3.24 2.553 -.425 (-1.0253 .82) lagged once.27) a5 – – c .955 (30.99) – – -.990 ( 4.687 ( 4.27) 1.51) IR .373 (-1.659 ( 2.00039 ( 1.12a log W = a1 + a2 T + a3 log W−1 + a4 log P Y +a5 (U RorJ J SorZZ) + a6 log P Y−1 a2 Quarterly CA .290 ( 5.00009 ( 0.840 (-4.04361 ( 5.92 2.0184 .0117 .17) SP -.552 ( 7.477 ( 4.277 (-1.37) .14 1.72 1.213 ( 0.41) .793 ( 3.30) -.80) .004 -.029 (-0.87 1.377 (-1.169 (-1.62) JA . c ZZ rather than U R.55) -1.905 ( 6.26) .185 -.21) .97 1.560 (-3.902 (20.20 1.711 .407 (-1.0107 .73 1.72 2.0073 .986 -.025 (10.69) GE .74) .82) .67) .64) 1.153 ( 1.0110 .54 2.32) SE .824 (13.06) b .03361 ( 2.19) DE -.034 ( 9.95 1.809 (13.07) .444 ( 1.340 -.49) .96) UK .890 ( 3.00035 (-0.147 (15.00786 ( 7.36) .406 ( 3. The overall test performance is thus only modest.84) FI .069 (-0.22) Annual BE -.00157 ( 7.937 .0403 .589 ( 4. .34) -.0265 .096 ( 6. b J J S rather than U R.916 (23.19) . RHO+ test.00331 ( 1.109 ( 2.88) .210 -.10) -5.24) .49) GR .00010 (-0.08) .80) c -.989 -1.89 1.51) .927 ( 3.46 a6 -.704 ( 4.0143 .18 2.0186 .53) 1.00098 ( 3.68) 1.74) a -.933 -1. and 8 fail the stability test.174 6 ROW STOCHASTIC EQUATIONS Table 6.63) .0365 .65) KO .36) .53) FR -.79) IT .37) .789 (12.03) .582 ( 8.52) – .815 ( 4.444 -.87) .08) .003 ( 8.54) – – – – -.01438 (-0.01) a4 1.108 ( 1.111 -.241 (-1.64) – .934 .0095 .535 (-2.00175 ( 0.943 (25.81) .0307 DW 1.628 ( 2.049 (-0.00302 ( 0.347 (-4.09) .00036 ( 1.147 ( 0.579 ( 2.001 (22.45) -.064 .00273 (-1.0232 .493 ( 2.55) – a -.18 1.239 .901 (26.44) 1.574 ( 2.879 -.86 2.371 -.0093 .45) SW .24) NZ .829 ( 3.493 (-1.41) c -.89) – ρ .00032 ( 2.57) NO .40) NE .68) a Variable a3 .36) .013 (-4.0179 .

829 . which equation 13 does not.03 ∗ 13.04 (6) 1.218 .262 . which is at least indirect support for the theory that firms at times hold excess labor and that the amount of excess labor on hand affects current employment decisions.076 .34 (6) 1.49 175 Sample (6) 4.481 . which equation 13 does not.003 KO .22 (5) 1.00 ∗ 31.64 (6) 1. and a time trend.199 IR .14 Equation 13: J : Employment Equation 13 explains the change in employment. ∗ Significant at the one percent level.21 ∗ 21.000 . The equation fails the lags test in 4 of the 15 cases.00 4.00 7.00 ∗ 7. Most of the change in output terms are also significant.11 The led value of the change in 11 Multicollinearity problems prevented the leads test from being performed for the UK.341 SW .00 ∗ 13.000 .746 .010 DE .869 .000 .070 UK .80 (5) 2.46 (5) 2. Equation 13 for the US model does not include the lagged change in output because it was not significant.486 .465 .004 .66 7. the lagged change in output.041 Stability AP (df) λ ∗ 29.348 SP .038 .193 .094 .824 .000 Annual BE .528 .002 AU .215 FI .666 .140 GE .13a.494 NO .207 NZ .000 .00 6. It passes the RHO+ test and the leads test in all cases.92 ∗ 15.369 . the change in output.081 .002 .744 Lags p-val .20 (5) 2.000 .11 (4) 1. Most of the coefficient estimates for the excess labor variable are significant in Table 6.95 (5) 2.63 (5) 2.035 .000 .067 FR .008 .12b Test Results for Equation 12 Restra p-val Quarterly CA .241 .031 .077 RHO+ p-val .56 (4) 3.836 . and it is similar to equation 13 for the US model. On the other hand.517 GR .23 (4) 2.49 (5) 1.297 NE .00 3. 6.26 (6) 1.600 .00 7.03 4.000 IT .21 ∗ 20.313 .08 (6) 1.25 4.00 1966:1–1992:3 1971:1–1992:3 1971:1–1991:2 1971:1–1992:2 1969:1–1991:4 1972:1–1991:4 1978:2–1991:4 1966:1–1992:3 1976:1–1991:4 1964:1–1991:4 1969–1990 1969–1990 1974–1990 1969–1990 1964–1990 1969–1990 1972–1990 1962–1990 a log P Y −1 added. It is in log form.00 5.6.90 (5) 1. It includes as explanatory variables the amount of excess labor on hand.209 .850 JA .000 .14 EQUATION 13: J : EMPLOYMENT Table 6. and it includes the lagged change in employment.01 6. US equation 13 includes terms designed to pick up a break in the sample period.33 (5) 1. .

08 1.085 ( -5.000004 ( -0.36) IR -.482 ( 4.15a.202 ( 3.96 2.009 ( -2.0070 . The explanatory variables include the real wage.05) NO -.75) .67) AS .51) .78) SP -.0056 .0072 .47) -.22) DE -.07) .144 (-3.000006 ( -0. When the real wage appeared .04) Annual BE -.47) -.174 ( 4. the labor constraint variable.739 ( 8.91 output was used for the leads tests.274 (-1.069 (-1.11) .72) .556 ( 5.15 Equation 14: L1: Labor Force—Men.601 ( 3.53) – – .381 ( 2. respectively.116 ( 1.03 2.135 ( 2.715 (-4.0125 .48) .004 ( -0.0034 .372 ( 3.145 (-4.01) .229 (-5.002402 ( 5.36) -.02 1.274 ( 2.40) -.28 2. The labor constraint variable is significant in most cases in Tables 6.92 2.10 2.05) . a time trend.99) – .322 (-6.68) .001236 ( 2.67) .26) -.411 (-3.007 ( -0.0046 .79) -.000224 ( 0.96) log Y + a5 ρ .001 ( -0.282 (-4.274 ( 2.006 ( 2.0021 .072 ( 1.25) .001245 ( 2.32) a2 -. The equation fails the stability test in 4 cases. Equation 15: L2: Labor Force—Women Equations 14 and 15 explain the labor force participation rates of men and women.036 ( -3.059 (-2.035 ( 3.156 ( 3.30 1.93) -.0038 . which provides support for the discouraged worker effect.59) – .91) .38) .83) .61) a4 .61 1.000299 ( 0.11 1.03) .000060 ( 3.004 ( 1.82) -.65) -.26) .152 ( 1.0099 .0056 .13a log J = a1 + a2 T + a3 log(J /J MI N )−1 + a4 a1 Quarterly CA .54) .0098 .41) FI .065 ( 0.05) -.02) -.364 ( 1.27) .75) – .000067 ( 1.32) – – .729 (-4.18) .64) -.203 (-2.03) JA – ( -0.000259 (-3. and 7 in the US model.000042 ( 0. 6.53 1.89 2.176 6 ROW STOCHASTIC EQUATIONS Table 6.50) -.0041 .58) .24) – – log Y−1 SE .000000 (0.37) SW -.83) – .25 2.50) ST -.000023 (-1.006 ( -2.103 ( -0.61) .002 ( 0.01) .14a and 6.47) .45) -.352 ( 4.268 ( 2.50) -. 6.29) GE -.0129 .031 ( 0.038 ( 1.13) .134 ( -0.145 ( 2.34) .21) .00) -.105 ( 0.17) AU -.182 ( 2.004 ( -0. There is only very modest support for the real wage.424 ( 3.210 ( 4.046 (-1.000084 ( 2.124 ( 1.26) – – – . The overall tests results for equation 13 are thus quite good. They are in log form and are similar to equations 5.40) a3 -.67) – .77) IT .46) UK .000023 ( 1.29) -.007 ( 2.0107 DW 1.16) .020 ( -1.160 (-3.378 ( 2.49) .31) -.96) .305 (-2.91) a5 .0045 .83) -.06 2. and the lagged dependent variable.103 ( 0.

00 2.16 The Trade Share Equations As discussed in Chapter 3.11 (6) 4.543 .229 AS .50 (5) 1. A separate equation was estimated for each ij pair.393 Leads p-val .093 .002 SW .066 Annual BE . The log of the price level was significant (and thus the restriction rejected) in 2 of the 7 cases.725 .75 (4) 1.043 . where i runs from 1 to 44 and j runs from 1 to 45. Both equations do poorly in the stability test.469 .011 IR . Equation 14 fails the lags test in 2 of the 14 cases and the RHO+ in 5 cases.48 (5) 1.46 (4) 1. Equation 15 fails no lags tests out of 10 and 3 RHO+ tests.775 .6.62 (5) 1.421 .313 .648 .27 (4) 1. the log of the price level was added to the equation for one of the tests to test the real wage restriction.859 .053 IT .130 JA . in the equation.045 .295 NO . αij is the fraction of country i’s exports imported by j . · · · .823 .013 . T ) . Equation 14 fails the test in 11 of the 14 cases.000 FI .022 .00 3.13) (t = 1.000 UK .000 SP .16 THE TRADE SHARE EQUATIONS Table 6.013 . The equation is the following: αij t = βij 1 + βij 2 αij t−1 + βij 3 ( P X$it ) + µij t 44 k=1 αkit P X$kt (6.00 Sample 1966:1–1992:3 1967:3–1992:3 1971:1–1991:2 1969:1–1991:4 1972:1–1991:4 1971:1–1991:4 1966:1–1992:3 1977:1–1991:4 1971:1–1992:2 1969–1990 1969–1990 1974–1990 1969–1990 1969–1990 1969–1990 at one the percent level.69 1.627 ST .00 6.25 2.98 ∗ 17.034 .918 .66 (5) 3.08 (5) 3.787 .114 GE .19 (5) 2.028 . The empirical work consisted of trying to estimate the effects of relative prices on αij .650 AU .00 0. 6.274 .00 5.51 (6) 1.92 ∗ 8. One would expect αij to depend on country i’s export price relative to an index of export prices of all the other countries.42 (5) 1.051 ∗ Significant 177 RHO+ p-val .86 (6) 2.17 ∗ 10. with observations for most ij pairs beginning in 1960:1.61 (5) 1.66 0.21 6.00 3.023 .64 (6) 2.027 Stability AP (df) λ 8.075 . The data on αij are quarterly. and equation 15 fails in 7 of 10.860 .371 .098 – .13b Test Results for Equation 13 Lags p-val Quarterly CA .57 ∗ 8.380 .826 .973 DE .

002777 (-7.897 (19.85) .012 ( -1.004326 (-1.0045 .97 2.0028 .000229 ( 2.830 (12.827 ( 6.94) UK .64) 0.038 ( 2.78 – – – – – .000559 (-2.55) ST -.37 1.595 ( 2.0034 .0035 .097 ( 3. j running from 1 to 45. where the weight for a given country k is the share of k’s exports to j in the total imports of i.44) .014 ( 0.74) DE -.206 ( -8.87) JA -.009 ( 0.48) .333 ( 6.06) AS .27) .93 1.148 ( -3.0057 .252 ( -3.20) -.01 1.99) NO -.68) -.0039 .000001 ( 0.65) IT -.10 1.0033 .053 ( -1.47 1.68) .46) .940 (19. Almost all of these cases were for the smaller countries.095 ( -0.07) -.241 ( 2.96) .99) -.000766 (-8.044 ( 0.44) a Variable a2 -.97 2. there are 1936 (= 44 × 44) ij pairs.618 ( 2.05) SP -.07) .01) .000165 (-3.000315 (-4.56) 0.56) .05) -.34) .003 ( 0.295 ( 5.89) – .39) .087 ( 1.50) .858 ( 3.064 ( 1. .150 ( 1. and 44 αkit P X$kt is an k=1 index of all countries’ export prices. There are thus 1936 potential trade share equations to estimate.72) -.29) .142 ( 1.69 2.08) – a5 .014 ( 1.14a log(L1/P OP 1) = a1 + a2 T + a3 log(L1/P OP 1)−1 +a4 log(W/P Y ) + a5 Z a1 Quarterly CA -.46) – – – – – – – – – – SE .36) a3 .46) 0.0051 .69) AU -.0035 .63) .148 ( 1.0077 .005 ( -1.578 ( 6. (In this summation k = i is skipped.53) a4 .352 ( 1.019 ( 1.000098 ( 0.0121 DW 1. Data did not exist for all pairs and all quarters.59) – a .109 ( 4.029 ( -2.178 6 ROW STOCHASTIC EQUATIONS Table 6.063 ( -4.11) .54) GE . and not counting i = j .0029 .31) -. the equation was not estimated for that pair.94) – – lagged once.843 (16.52) .21 1.00 1.50) 0.58) . A few other pairs were excluded because at least some of the observations seemed extreme and likely suffering from measurement error.18) .001298 ( -0.764 (11.000097 (-2. In fact.20) -. and if fewer than 26 observations were available for a given pair.64) .384 ( -5.329 ( 3.0062 . only 1560 trade share equations were estimated.0016 .57) a .34) Annual BE -.21 2.052 ( 3.69) .15) .196 ( -2.73) – .045 ( 3.000729 (-4.43) .) With i running from 1 to 44.81) -.972 (61.003189 (-2.46) SW -.050 ( 3.13 1.096 ( 3.27) FI -.000016 ( 1. P X$it is the price index of country i’s exports.87 2.868 (25.08) ρ – – – .

1 percent had the correct sign and a t-statistic greater than one in absolute value.760 IT .98 ∗ 23.642 UK .13 (4) 3. Each of the 1560 equations was estimated by ordinary least squares.113 . αij t .0132. 83. do not obey the property that 44 αij t = 1.2 percent had the wrong sign and a t-statistic greater than two.58 (5) 2. Of the 1560 estimates of this coefficient.745 JA .284 .000 .31 (4) 4.17 ∗ 22.263 ST .755 ∗ Significant 179 log P Y p-val . βij 3 measures the short run effect of a relative price change on the trade share. the sum of total world exports will not equal the sum of total world imports.755 AS .015 – – RHO+ p-val .771 NO .25 ∗ 11.012 .00 4.00 ∗ 8.57 (4) 2.4 percent had the correct sign and a t-statistic greater than two in absolute value.36 4.008 – . 44. The average of the 1299 estimates of βij 3 that were of the right sign is −. say.00 ∗ 13.255 .007 . The main coefficient of interest is βij 3 .001 .000 – .137 . the coefficient of the relative price variable. and 7.57 ∗ 24.50 (5) 2.69 ∗ 10.88 (4) 1.614 .000 FI .713 .006 SP .11 (5) 1. The trade shares for these ij pairs were taken to be exogenous.0580.14b Test Results for Equation 14 Lags p-val Quarterly CA .038 AU . 3.6.82 1968:1–1992:3 1967:3–1992:3 1971:1–1991:2 1969:1–1991:4 1972:1–1991:4 1971:1–1991:4 1968:1–1992:3 1976:1–1991:4 1971:1–1992:2 1971–1990 1969–1990 1974–1990 1969–1990 1972–1990 (5) 1.000 . The long run effect is βij 3 /(1 − βij 2 ).513 DE . The overall results are thus quite supportive of the view that relative prices affect trade shares.66 ∗ 20. The trade share equations with the wrong sign for βij 3 were not used in the solution of the model. It should be noted regarding the solution of the model that the predicted values of αij t .60 (5) 1.21 ∗ 12.28 (3) 1.000 .292 – – – – – – .362 Annual BE . For solution purposes each αij t was divided by ˆ 44 αij t = 1. and this adjusted figure was used as the predicted trade share.85 Sample (5) 3.18 (4) 1. and the average of the 1299 values of this is −.00 at the one percent level.957 Stability AP (df) λ ∗ 24.231 .5 percent had the wrong sign and a t-statistic greater than one.738 GE .36 (4) 1. Unless ˆ i=1 ˆ this property is obeyed.00 ∗ 7.00 2. ˆ i=1 .16 THE TRADE SHARE EQUATIONS Table 6.21 (4) 1. and 68.3 percent (1299) were of the expected negative sign.731 SW .

and V 1 (equation 4)—equation 3 is by far the weakest.41) .40) -.97) . 2.75) .012422 (12.116 ( 1.779 (18.071 ( 9.167 ( -3.60) .81) – a4 – – .101 ( 2.31) FI -.71 1.701 ( 7.96) . C (equation 2).0101 .97 2.049 ( -1.199 ( -5.001040 ( 2.065 ( 2.876 (20. It may be that the construction of the capital stock series is too crude to allow good results to be obtained.46 1.000067 ( 2.86) . As discussed in point 11 in Section 5.053 ( -2.12) a2 .58) .33) .726 (10.144 ( 1.551 ( 5.035 ( -0.30) . Of the equations explaining the components of GDP—M (equation 1).05) DE -.28 1.57 1.276 (-2.0079 .0129 .000015 ( -0.47 1.81 1.22) .0207 DW 2.938 (34.019 ( 0.0057 .72) .0048 .23) .000988 ( 3.50) – – – SE .13 were adjusted to satisfy the requirement that the trade shares sum to one. 6.61) AU -.06) -.180 6 ROW STOCHASTIC EQUATIONS Table 6.40) SW -.249 ( 5. this has important implications for the long run properties of the model. 1.17 Additional Comments The following are a few general remarks about the results in this chapter.0031 .846 (16.41) ρ – -.918 (14.000157 (-1.80) Annual BE -.96) . The significance of the import price index in the price equations is also .17) -.909 ( -3.69) – – a5 – .25) .003224 ( 1.5b is an important result.662 ( -5.40) – .237 ( 3.558 ( 2.21) .689 ( 3.97) a3 .97 1. The strong rejection of the change form of the price equation in Table 6. the values predicted by the equations in 6.467 ( -74.64) SP -1. or it may be that the sample sizes are too small to allow the simultaneity issue to be handled well.0111 .15a log(L2/P OP 2) = a1 + a2 T + a3 log(L2/P OP 2)−1 +a4 log(W/P Y ) + a5 Z a1 Quarterly JA -.91) . I (equation 3).002545 ( 2.14) .60) AS -.89) .000207 ( 5.0081 .19) IT -.81) ST -.150 ( -2.000124 ( 0.72) .904 (18.429 ( -1.712 ( 3.10.16) .69) .492 (-2.20) .35 2.43) – – – .0046 .18) – – – .77 In other words.040 ( 0.

6.15b Test Results for Equation 15 Lags p-val Quarterly JA . The results of estimating the demand for money equations in Table 6. 3. As was the case for the US model.692 – – – . 8.00 at the one percent level.21 (5) 1.000 .410 IT .253 – – RHO+ p-val .548 Annual BE . The led values are significant .168 .320 SW . This shows how price levels in different countries affect each other. 7. A key question for the exchange rate equations in Table 6.46 ∗ 21. important.15 ∗ 22.371 . In very few cases is the inflation expectations variable significant.095 AU .S.025 FI . The verdict is not yet in on this question.17 ADDITIONAL COMMENTS Table 6.00 ∗ 8.7a.17 (5) 1.33 (4) 1.156 AS . The excess labor variable is significant in most of the equations in Table 6.00 4. 6.09 (6) 1.888 Stability AP (df) λ ∗ 14. interest rate is significant in 11 of the interest rate reaction functions in Table 6.00 4.243 .35 Sample (3) 3. See also point 5 in Section 5. the results support the use of nominal interest rates over real interest rates.420 SP . The U.000 .64 (4) 1.645 ST .131 ∗ Significant 181 log P Y p-val – – .534 .42 1967:3–1992:3 1971:1–1991:2 1972:1–1991:4 1971:1–1991:4 1976:1–1991:4 1971:1–1992:2 1971–1990 1969–1990 1969–1990 1972–1990 3.167 DE . 5. This is evidence that the Fed influences the economies of other countries by influencing other countries’ interest rates. There is little support for the use of the led values and thus little support for the rational expectations hypothesis. It will be seen in Chapter 12 that this is an important link. 4.000 . which adds further support to the theory that firms at times hoard labor.64 (3) 1.57 (5) 2.10.79 ∗ 39.6a provide further support for the nominal adjustment hypothesis over the real adjustment hypothesis.214 .28 ∗ 9.17 ∗ 11.13a.242 .501 .9a is whether one can trust the inclusion of the relative interest rate variable in the equations.00 (5) 1.98 (4) 2.

The results are not as good for the stability test. For the lags test there are 45 failures out of 274 cases. More observations are probably needed before much can be done about this problem. These results suggest that the dynamic specification of the equations is reasonably good.182 6 ROW STOCHASTIC EQUATIONS at the one percent level in only 18 of the 153 cases in which they were tried. and for the RHO+ test there are 43 failures out of 304 cases. . for the T test there are 44 failures out of 217 cases. where there are 105 failures out of 299 cases. T. The equations in general do well for the lags. and RHO+ tests. 9.

Finally. . ut denotes the m–dimensional vector (u1t . n). umT ) .7 Estimating and Testing Complete Models 7. G denotes the k × m · T matrix:      G =    G1 0 . . . The model will continue to be written as in 4. . . Jt denotes the n × n Jacobian whose ij element is ∂fi /∂yj t . . . . . 0 Gm          where Gi is defined in Section 4. 183 . u denotes the m · T –dimensional vector (u11 . (i. . . j = 1. um1 . . u1T . . umt ). . 0 0 G2 . . . . Some additional notation is needed from that used in Chapter 4 to handle the complete model case. .1. . . .1. . . The additional notation is as follows. . and denotes the m × m covariance matrix of ut .1 Notation This chapter discusses the estimation and testing of complete models. . . .

3. 7. Stochastic simulation requires that an assumption be made about the distribution of ut .4) where the derivatives are evaluated at the optimum.5. is usually estimated from the 2SLS estimated residuals.2) ˆ where G is G evaluated at the 3SLS estimate of α. for a more detailed discussion of stochastic simulation.3. 3SLS estimates of α are obtained by minimizing S = u [ ˆ −1 ⊗ Z(Z Z)−1 Z ]u = u Du (7. which is done for the computational work in the next chapter. An estimate of the covariance matrix of ˆ the 3SLS coefficient estimates (denoted V3 ) is ˆ ˆ ˆ V3 = (G D G)−1 (7.184 7 TESTING COMPLETE MODELS 7. 2 See Fair (1984). ). 6. Section 7. An estimate of the covariance matrix of the FIML coefficient ˆ estimates (denoted V4 ) is ˆ V4 = − ∂ 2L ∂α∂α −1 (7. Sections 6. . FIML estimates of α are obtained by maximizing L=− T T log | | + log |Jt | 2 t=1 (7.3 Stochastic Simulation2 Some of the methods in this chapter and in Chapter 10 require stochastic simulation.3. and 6.3.2 3SLS and FIML1 Two full information estimation techniques are three stage least squares (3SLS) and full information maximum likelihood (FIML).3) with respect to α. It is usually assumed that ut is independently and identically distributed as 1 See Fair (1984).3. and so it will be useful to give a brief review of it. for a more detailed discussion of the 3SLS and FIML estimators. 6.1) with respect to α.5.2.4. Under the assumption that ut is independently and identically distributed as multivariate normal N(0. where ˆ is a consistent estimate of and Z is a T × K matrix of predetermined variables.

one is interested in predicted values more than one period ahead. denoted αi . denoted σit . ). is ˜ µit = ˜ Let 1 J j y J j =1 it j (7. t This can be done as many times as desired.7) Given the data from the repetitions. From each repetition one obtains j a prediction of each endogenous variable. Given u∗ and given αi for all i. Stochastic simulation also requires that consistent estimates of αi be available for all i. Alternative assumptions simply change the way the error terms are drawn. in predicted values from dynamic simulations.7. i. although other assumptions can clearly be used. The variance of σit .8) In many applications. the normality assumption has always been used. Let yit denote the value on the j th repetition of variable i for period t. one can solve the model ˆ t for period t. For J repetitions.” Another repetition can be made by drawing a new set of values of u∗ and solving again. Call this simulation a “repetition. can be computed as fi (yt . The covariance matrix can then be ˆ ˆ ˆ ˆ ˆ estimated as (1/T )U U . ˜ ˜2 ˜2 2 ). it is also possible to compute the variances of the stochastic simulation estimates and then to examine the precision of the estimates.5) ˜ σit = (yit − µit )2 2j (7.6) The stochastic simulation estimate of the variance of variable i for period t. One simply draws . αi ). The variance of µit is simply σit /J . For the results in this book. is denoted var(σit ˜ var(σit ) = ˜2 1 J 2 J j =1 (σit − σit )2 ˜2 2j (7. is then ˜2 σit = ˜2 1 J 2j σ J j =1 it (7. denoted µit .3 STOCHASTIC SIMULATION 185 multivariate normal N(0. xt .e. consistent estimates of uit . Given these estimates. ˆ ) distribution. where U is the m × T matrix of the values of uit . ˆ ∗ denote a particular draw of the m error terms for period t from the Let ut N (0.. the stochastic simulation estimate of the expected value of variable i for period t. The above discussion can be easily modified to incorporate this case. This is merely a deterministic simulation for the given values of the error terms and coefficients. denoted ˆ uit .

ˆ ˆ An important conclusion that can be drawn from stochastic simulation studies using macroeconometric models is that the values computed from deterministic simulations are quite close to the mean predicted values computed from stochastic simulations. When coefficients are drawn. each repetition yields eight predicted values. material in this section is taken from Fair (1994a). Section 7. say.5. although in practice the possible nonexistence of moments is generally ignored.4. for each endogenous variable. each repetition consists of a draw ˆ ˆ of the coefficient vector from N (α. Each repetition is one dynamic simulation over the period of interest.3. V ) and draws of the error terms as above. Alternative measures of dispersion that are robust to the nonexistence problem give very similar results to those obtained using variances. It has been known since the work of Orcutt (1948) and Hurwicz (1950) that least squares estimates of lagged dependent variable (LDV) coefficients are biased even when there are no right hand side endogenous variables.186 7 TESTING COMPLETE MODELS values for ut for each period of the simulation. Table IV) for 10. . an estimate of the distribution of the coefficient estimates is N(α. In other words. For example. Section 8. 20.9804 (for small values of the coefficient). Given V ˆ mality assumption. the results in Orcutt and Winokur (1969.5. Macroeconometric model builders have generally ignored this problem. V ). for 100 observations the ratio of the expected value of the LDV coefficient estimate to the true value was . an eight quarter period. the 2SLS estimate of all the coefficients in the model. suggest that the possible nonexistence of moments is not an important problem for macroeconometric models. It is also possible to draw coefficients for the repetitions.4 Median Unbiased Estimates4 The estimator considered in this section will be called the “median unbiased” (MU) estimator. one per quarter. ˆ ˆ say. for references. and let V denote ˆ and given the northe estimate of the k × k covariance matrix of α. the bias that results from using deterministic simulation to solve nonlinear models appears to be small.3 It may be the case that the forecast means and variances do not exist. For. perhaps because they feel that the bias is likely to be small for the typical number of observations that are used. However. Hurwicz’s estimates of the bias in an equation with only the LDV as an explanatory variable were small after about 100 observations. 7. and 40 observations show biases larger than those of Hurowitz 3 See 4 The Fair (1984). Let α denote. Results in Fair (1984).

for 100 observations and a true coefficient of . . Constrain α1i to be equal to α1i and reestimate ˆ ∗ denote this estimate of α the other elements of αi by 2SLS. . The procedure for obtaining median unbiased estimates of the α1i coefficients (i = 1. more complicated than the equations just discussed. Typical macroeconomic equations are. of course. from the distribution t N (0. Add b1i to α1i to obtain a ˆ ˆ ∗ first estimate of the true value of α1i . denoted b1i .7. . say. . 100 observations. Let α1i denote the coefficient of interest in equation i. Furthermore. The interest here is on the coefficient of the lagged dependent variable. . Let α1i denote this estimate: ∗ ∗ α1i = α1i + b1i . .9388 in the equation with the constant term and time trend added. and lagged dependent variable. . Estimate each equation i by 2SLS. defined as the difference between the base estimates and the MU estimates.8. T . ∗ ). Guess the bias of α1i . Let α1i denote the 2SLS estimate of ˆ α1i . . the ratio of the median of the LDV coefficient estimate to the true value is . It is important to know how the size of the biases for these types of equations compare to those estimated for simpler equations. the error terms are sometimes serially correlated. . and the equations may be nonlinear in both variables and coefficients. 2. some of the explanatory variables are likely to be endogenous. Draw T values of the vector u∗ . Let αi i (i = 1. Use the estimated residuals from these constrained regressions to estimate the covariance matrix . They have more explanatory variables. 3. . For the work here the 2SLS estimates will be taken to be the base estimates. time trend. can be computed. . The following stochastic simulation procedure provides a way of obtaining median unbiased estimates in macroeconometric models. Andrews (1993) has recently shown that the bias is further increased when a time trend is added to the equation. . m) using the 2SLS estimator is as follows: 1. From these estimates the bias for a coefficient. The procedure requires that one coefficient per stochastic equation be singled out for special treatment. t = 1. Let ∗ denote this estimate of . . For example. but other coefficients could be considered. . Use these values and the values αi∗ (i = 1. . m). . suggesting that the bias in this case is also larger for. This procedure is an extension of Andrews’ (1993) idea of computing exact median unbiased estimates in an equation with a constant term. m) to solve .4 MEDIAN UNBIASED ESTIMATES 187 for the case in which there is a constant term in the equation.

. Again. Take the median unbiased estimate of α1i to be α1i . . T . Keep doing steps 3 and 4 until convergence is reached and one branches to step 6..e. . and take the other ∗ coefficient estimates to be those in αi∗ (i = 1. . the solution values of the endogenous and lagged endogenous variables). . . This is a dynamic simulation of the model over the entire estimation period using the drawn values of the error terms and the coefficient values αi∗ . . . m). If this condition is not met. m). ∗ Let Zit . . Do a second repetition by drawing another T values of u∗ . . for example. Given the new data (i. . If this condition is met. go to step ˆ 6. and then find the median α1i of the J values of α1i (j = 1. m 4. Now repeat step 3 for these new values. .188 7 TESTING COMPLETE MODELS the model dynamically for t = 1. the median 2SLS estimates are within a prescribed tolerance level of the original estimates based on ∗ the historical data. m). . Let ∗ denote this estimate of . m). . . (1) estimate each equation by 2SLS. This is one repetition. . . and record the estimate of α1i as α1i (i = 1. . use the estimated residuals from these constrained regressions to estimate the covariance matrix . . Let αi∗ denote this estimate of αi (i = 1. . (i = 1. If for each i α1i is within a prescribed tolerance level of α1i . . t = 1. . T . . . . . The lagged endogenous variable values in xt in 4. . it means that for the particular coefficient values used to generate the data (the αi∗ ’s). m). J ). Then constrain α1i ˆ ∗ and reestimate the other elements to be equal to this new value of α1i of αi by 2SLS using the historical data. After this solution. ∗ 6. update Zit to incorporate the new lagged endogenous variable values (if lagged endogenous variable values are part of Zit ). and recording the estimate of α1i as α1i (i = 1. . J ). Do (j ) m this J times. denote this update. For a 90 percent confidence interval. using the new data to estimate each equation by (2) 2SLS. take the new value of α1i m to be the previous value plus α1i − α1i for each i. . 5. ˆ ˆ m Confidence intervals for α1i can be computed from the final set of values (j ) of α1i (j = 1. . . α1i is the median unbiased estimate in that it is the value of α1i that generates data that lead to the median 2SLS estimate equaling (within a prescribed tolerance level) the 2SLS estimate based on the historical data. .1 are updated in the solution process. The estimated bias ∗ of α1i is α1i − α1i . . . . . using these values and the values αi∗ t to solve the model. .

7. and compute ˆ using these estimates. The lagged endogenous variable values in 4. Second. it focuses on just one coefficient per equation. If some variables are not stationary. this is a dynamic simulation of the model over the entire estimation period. In fact. Take an estimator. much of the recent literature in time series econometrics has been concerned with the consequences of nonstationary variables. of course. Take these coefficient estimates. the asymptotic approximations may not be very good. ˆ ) distribution (assuming the normality assumption is used). the material in this section is taken from Fair (1994a). denoted α. The model in 4. ˆ From the N(0.1 are 5 As in the previous section. as noted above. or FIML. this procedure does not require the normality assumption. As in ˆ step 3 of the previous section. No other coefficient estimate in an equation necessarily has the property that its median value in the final set of values is equal to the original estimate. As noted above. and estimate the model. Other distributions could be used to draw the u∗ values. Other estimators could be used. The procedure does. the basic t estimator need not be the 2SLS estimator. there is no guarantee that the procedure will converge. solve the model for the entire period 1 through T . Given these error terms and α. . 3SLS. Remember that overall convergence requires that convergence be reached for each equation. The focus. convergence was never a problem. as will be seen. but it must be on one particular coefficient per equation. The procedure proposed here for examining asymptotic distribution accuracy is similar to the procedure of the previous section. say 2SLS. draw a vector of the m error terms for each of the T observations. In this case is a scalar and the “solution” of the model simply consists of solving the particular equation (dynamically) over the sample period. and achieving this much convergence could be a problem. however.5 Examining the Accuracy of Asymptotic Distributions5 It is possible using stochastic simulation and reestimation to examine whether the asymptotic approximations of the distributions of estimators that are used for hypothesis testing are accurate.5 ASYMPTOTIC DISTRIBUTION ACCURACY 189 5 percent of the smallest values and 5 percent of the largest values would be excluded. First. have two limitations.1 can also consist of just one equation. Also. need not be on the coefficient of the LDV. 7. however. as the base values. For the results in the next chapter.

say. the matrices of first stage regressors. Remember that these coefficient estimates are the ones used to generate the data. as for 2SLS. and 5 percent values and for both left and right tails. 10. ˆ ) distribution and that the coefficient vector used in the solution is always α. (Remember that the draws of the errors are always from the N(0. One can then compute the actual percent of the coefficient estimates from the exact distribution that lie above this value and compare this percent to 20 percent.190 7 TESTING COMPLETE MODELS updated in the solution process. . this distribution of the J values will be called the “exact distribution. The predicted values from this solution form a new data set. and record each set of estimates. The asymptotic distribution can then be compared to this exact distribution to see how close the two distributions are. are updated to incorporate the new lagged endogenous variable values if the matrices are used in the estimation.” although it is only an approximation because is estimated rather than known.84 times the estimated asymptotic standard error. the median of the exact distribution for a coefficient was first compared to the coefficient estimate from the technique in question. which is 2SLS in this case. Next. and record the set of estimates. For the empirical work in the next chapter. For the work in the next chapter. For ease of exposition.) ˆ If J repetitions are done. Repeat the draws. as discussed in the previous section. Zit . For 20 percent. estimate the model by the technique in question. Also. and estimation for many repetitions. this value is the median plus 0. solution. are likely to be biased downward. Given this data set. The coefficient estimates of the lagged dependent variables. which are likely to be a good approximation of the exact distribution. 20 percent of the coefficient estimates should lie if the asymptotic distribution is correct. There are a number of ways to examine the closeness of the asymptotic distribution to the exact distribution. for example. given the median from the exact distribution and given the estimated standard error of the coefficient estimate from the asymptotic distribution. This is one repetition. one can compute the value above which. It will be seen that the exact and asymptotic distributions are generally quite similar regarding their tail properties. defined as the difference between the median and the coefficient estimate. One can then examine the bias of a coefficient estimate. this comparison was made for 20. one has J values of each coefficient estimate.

AC models do not have the problem.6 VAR AND AC MODELS 191 7. GDP is then determined from the GDP identity. each of the seven variables is taken to be a function of the constant. as VAR models do. These are the same variables used by Sims (1980) with the exception of RS. There are thus 30 coefficients to estimate per each of the seven equations. for a total of 19 coefficients per equation. and 7) the bill rate. and the first four lagged values of each of the other variables. For the work in Fair and Shiller (1990) three sets of Bayesian priors were imposed on VAR4. a time trend. It is possible to decrease the number of unrestricted coefficients to estimate in VAR models by imposing various priors on the coefficients. As will be seen in the next chapter. which has been added here. as the sum of the components. denoted VAR4. 5) the log of the money supply. For the second VAR model. a time trend.6 VAR and AC Models for Comparison Purposes When testing complete models. “autoregressive components” (AC) models appear to be better benchmarks than VAR models in the sense of being more accurate. although no priors were imposed for the work in this book. its first five lagged values. its first four lagged values. and the first two lagged values of each of the other variables. An AC model is one in which each component of GDP is regressed on its own lagged values and lagged values of GDP. log GDP D. 2) the log of the GDP deflator. denoted VAR5/2. Therefore. Two VAR Models Two seven variable VAR models are used in the next chapter for comparison with the US model. each of the seven variables is taken to be a function of the constant. of adding large numbers of parameters as the number of variables (components in the AC case) is increased. U R. if the interest is in GDP predictions. it is useful to have benchmark models to use for comparison purposes. For the first VAR model. log M1.7. 3) the log of the wage rate. The results using these versions were similar to the results using VAR4. the results using VAR4 are likely to be close to the results that would be obtained using priors. 4) the log of the import price deflator. log W F . Vector autoregressive (VAR) models provide useful benchmarks. RS. and very little gain seemed to result from the use of the priors in terms of making the VAR models more accurate. log P I M.6 6 Sims (1993) considers a nine variable VAR model with five lags and imposes an elab- . log GDP R. The seven variables are (in the notation of the variables in the US model) 1) the log of real GDP. however. 6) the unemployment rate.

10 The method discussed in this section was briefly outlined in Section 1. raw data variables and are defined in Table A. 7 AC models were first proposed in Fair and Shiller (1990). Chapter 8. EX. The final equation of the AC model is the GDP R identity. I KG. CS. P ROG and P ROS are combined in the US model in such a way that they do not appear as separate variables. and the AC model used in the next chapter consists of estimated equations for each of these components.. say. This estimate is presented in equation 7. 8 The 19 components in alphabetical order are CD. and whether or not lagged values of GDP R are included). I M. however. I H F . CN. I KF . and the first two lagged values of GDP R. the length of the lag. VAR4. but at the present time it would be extremely difficult to try to duplicate Sims’ procedures. The results in Fair and Shiller (1990) show that going from a few components to 17 improves the accuracy of the AC model. then the uncertainty from both of these sources has been accounted for. I KH . however. 7. 9 The number of components in the US model at the time of this work was 17. I V F . P ROS.3. .8 Each component is taken to be a function of the constant. alternative choices of number of components. COG. P ROG. a time trend.2. the results are not highly sensitive to different versions of the AC model (i. One might think that forecast error variances computed in this way could orate set of priors on the coefficients. and ST AT P . They are. I KB.7 except that a k subscript has been added to denote the length ahead of the forecast.192 The AC Model7 7 TESTING COMPLETE MODELS There are 19 components of GDP R in the US model (counting the statistical discrepancy ST AT P ).e.4.7 Comparing Predictive Accuracy10 As discussed in Section 7. As with different versions of the VAR model. stochastic simulation allows one to compute forecast error variances.9 but that going beyond this does not. If the estimated variance is based on draws of both the error terms and coefficients. In future work it would be interesting to see how well this model does compared to. Let σitk denote the stochastic simulation estimate ˜2 of the variance of the forecast error for a k period ahead forecast of variable i from a simulation beginning in period t. and it is discussed in more detail in Fair (1980a) and in Fair (1984). I H H . The results also show that adding lagged values of GDP R (versus not having GDP R in the equations at all) leads to a slight improvement in accuracy. its first five lagged values. I H B. COS. I V H . hence 17 instead of 19 components were used.

have no exogenous variables. From this simulation one also obtains an estimate of the expected value of the k period ahead forecast of variable i. The first is controlling for different sets of exogenous variables across models (VAR and AC models. Let the prediction period begin one period after the end of the estimation period. where again k refers to the length ahead of the forecast. two additional problems. then ˆisk ˜ in equation 7. µisk . the normality assumption to draw exogenous variable values for the stochastic simulation. and call this period s.5. Another way is to estimate in some manner the forecast error variance for each exogenous variable (perhaps using past errors made by forecasting services in forecasting the variable) and then use these estimates and. One therefore ˆisk isk 2 has two estimates of σisk . in ˜2 equation 7. Consider stochastic simulation with both error terms and coefficients drawn. for example. The expanded model can then be stochastically simulated to get the variances.7 COMPARING PREDICTIVE ACCURACY 193 simply be compared across models to see which variances are smaller.9 is a sample draw from a distribution with a known mean of zero 2 2 and variance σisk . say. One is to estimate autoregressive equations for each exogenous variable and add these equations to the model. is thus under this assumption an unbiased estimate of σ 2 . A model may have small estimated variances of the structural error terms and small estimated variances of the coefficient estimates (which leads to small forecast error variances from the stochastic simulation) because it has managed to spuriously fit the sample well. There are. The expected value of the difference between the two estimated variances for a given variable and period is zero for a correctly specified model. A further step is needed to handle this problem. however. The expected value is not in general zero for a misspecified model.9) If it is assumed that µisk exactly equals the true expected value. yis+k−1 . The difference between ˜ this estimate and the actual value. whereas a model like the US model has many). The second problem is the possibility of data mining.7. and this fact can be used to adjust the forecast error variances for the effects of misspecification. where σisk is the true variance. This can be done in a variety of ways. σisk . one computed from the mean forecast error and one . denoted ˆisk : ˆisk = yis+k−1 − µisk ˜ (7.7. 2 . is the mean forecast error. in equation 7. The expanded model in effect has no exogenous variables. which is to compare variances estimated from outside sample forecast errors with variances estimated from stochastic simulation. From a stochastic simulation beginning in period s one obtains an estimate of the variance of the forecast error. The square of this error.

It results in one value of disk for each variable i and each length ahead k. Assume. where the estimation period ends in period s − 1 and the prediction period begins in period s. Therefore. the expected value of disk is zero for a correctly specified model. it is not in general true that the expected value of disk is zero.e. In general. The procedure described so far uses only one estimation period and one prediction period. It is possible for misspecification to affect the two estimates in the same way and thus leave the expected value of the difference between them equal to zero.. Misspecification has two effects on disk . under the two assumptions of no error in the stochastic simulation estimates. and so in general one would not expect the expected value of disk to be zero for a misspecified model.194 7 TESTING COMPLETE MODELS computed by stochastic simulation. The estimated variances computed by means of stochastic simulation will thus in general be biased. for example. Because of the common practice in macroeconometric work of searching for equations that fit the data well (data mining). more observations must be generated. First. the effect on disk is ambiguous. Let disk denote the difference between these two estimates: 2 disk = ˆisk − σisk ˜2 (7. ˜2 ˜2 then disk is the difference between the estimated variance based on the mean forecast error and the true variance. If the model fits the data well within sample. This can be done by using successively new estimation periods and new prediction periods. the stochastic simulation estimates of the variances will be small because they are based on draws from estimated distributions of the error terms and coefficient estimates that have small (in a matrix sense) covariance matrices. Since one observation is obviously not adequate for estimating the mean of disk . that one has data from period 1 . If the model. this should result in large outside sample forecast errors. Second. and ˆisk large outside sample forecast errors. it seems likely that the estimated means of disk will be positive in practice for a misspecified model. this does not seem likely. If a model is misspecified. The estimated mean of disk is thus likely to be positive: σisk is small ˜2 2 is large because of because of small estimated covariance matrices. is in fact misspecified. however. Since misspecification affects both estimates. σisk = σisk ). the estimated variances computed from the forecast errors will in general be biased estimates of the true variances. the estimated covariance matrices that are used for the stochastic simulation will not in general be unbiased estimates of the true covariance matrices. if the model is misspecified. although fitting the data well.10) 2 If it is further assumed that σisk exactly equals the true value (i.

denoted as dik . . The assumption made for the empirical work in the next chapter is that the mean is constant across time.11) Since the procedure in arriving at σitk takes into account the four main sources ˆ2 of uncertainty of a forecast. A variety of assumptions are possible. ˜2 Assume that the period of interest begins in period t. and k. . 150). The model can be estimated through. .8 COMPARING INFORMATION IN FORECASTS 195 through period 150. it can be compared across models for a given i. . it is possible to estimate the total variance of the forecast error. Chapter 8. . period 100. is the ˆ2 ¯ sum of the stochastic simulation estimate plus dik : ˜2 ¯ ˜ σitk = σ itk + dik ˆ2 (7. It focuses on the 11 The material in this section is taken from Fair and Shiller (1990). ¯ Given dik . and so on. and exogenous variables. . with the prediction beginning with period 101. for the two period ahead forecast (k = 2) it will yield 49 values of dis2 . misspecification is assumed to affect the mean in the same way for all s. say. This process can be repeated through the estimation period ending with period 149. 149). Given this assumption. ¯ the mean. can be estimated by merely averaging the computed values of disk . and let σ denote the ˜ itk stochastic simulation estimate of the variance based on draws of error terms. . Stochastic simulation for this prediction period will yield for each i and k a value of di102k . In other words. . coefficients. The model can then be reestimated through period 101. Stochastic simulation for the prediction period will yield for each i and k a value of di101k in equation 7. 7. which are discussed in Fair (1984). . (s = 101. The final step in the process is to make an assumption about the mean of disk that allows the computed values of disk to be used to estimate the mean. with the prediction period now beginning with period 102. For the one period ahead forecast (k = 1) the procedure will yield for each variable i 50 values of dis1 (s = 101.7. The total variance. Note that calculating the individual disk values that are needed to ¯ calculate dik is computer intensive in that it requires estimating and stochastic simulating many times.10.8 Comparing Information in Forecasts11 Introduction This section discusses an alternative way of comparing models from the method of comparing variances in the previous section. denoted σitk . t.

one cannot perform such tests with most large scale models because. Davidson and MacKinnon (1981). In other words. See also Nelson (1972) and Cooper and Nelson (1975) for an early use of encompassing like tests. 13 See. VAR models are typically much smaller than structural models and in this sense use less information. or is most of the information extraneous? The same question can be asked of AC models versus structural models. How should one interpret the differences in forecasts? Does each model have a strength of its own. carries different information from another’s can be examined by regressing the actual change in the variable on the forecasted changes from the two models. as noted above. which is discussed below. if the RMSEs are close for two forecasts. for example. Even ignoring this point. are the a priori restrictions of large scale models useful in producing derived reduced forms that depend on so much information. First. is related to the literature on encompassing tests12 and the literature on the optimal combination of forecasts. These restrictions might be wrong in important ways and yet the model contain useful information. One cannot answer this question by doing conventional tests of the restrictions in a structural model. Estimation can proceed effectively only because of the large number of a priori restrictions imposed on the model. for example. and Mizon and Richard (1986). Chong and Hendry (1986). The above question with respect to VAR models versus structural models is thus whether the information not contained in VAR models (but contained in structural models) is useful for forecasting purposes.196 7 TESTING COMPLETE MODELS information contained in each model’s forecast. The information set used in a large scale macroeconometric model is typically so large that the number of predetermined variables exceeds the number of observations available for estimating the model. Hendry and Richard (1982). little can be 12 See. and so their forecasts are not perfectly correlated with each other. restrictions that do not work out to be simple exclusion restrictions on the reduced form equation for the variable forecasted. This procedure. real GDP.13 This procedure has two advantages over the standard procedure of computing root mean squared errors (RMSEs) to compare alternative forecasts. Granger and Newbold (1986). The question whether one model’s forecast of a variable. Econometric models obviously differ in structure and in the data used. so that each forecast represents useful information unique to it. or does one model dominate in the sense of incorporating all the information in the other models plus some? Structural econometric models make use of large information sets in forecasting a given variable. for example. however. there are not enough observations to estimate unrestricted reduced forms. .

404). p. so that only information through period t − 1 should be used for the forecasts. Second. One way to handle the exogenous variable problem is to estimate.7. as McNees notes. this may bias the results against the US model. it may still be that the forecast with the higher RMSE contains information not in the other forecast. even if one RMSE is much smaller than the other. The coefficient problem can be handled by doing rolling estimations for each model. The first is if actual values of the exogenous variables for periods after t − 1 are used in the forecast. The third is if information beyond t − 1 has been used in the specification of the model even though for purposes of the tests the model is only estimated through period t − 1. This method of dealing with exogenous variables in structural models was advocated by Cooper and Nelson (1975) and McNees (1981). such as revised seasonal factors and revised benchmark figures. The procedure requires that forecasts be based only on information available prior to the forecast period. noted that the method handicaps the model: “It is easy to think of exogenous variables (policy variables) whose future values can be anticipated or controlled with complete certainty even if the historical values can be represented by covariance stationary processes. 1981. It should be stressed that the current procedure does not allow one to discover whether all the variables in a model contribute useful information for forecasting. For the work in the next chapter autoregressive equations have been estimated for each exogenous variable in the US model.” (McNees. not that it contains no extraneous variables. The fourth is if information beyond period t − 1 has been used in the revisions of the data for periods t − 1 and back. There is no way to test for this using the RMSE framework. say. although. Assume that the beginning of the forecast period is t. an autoregressive equation for each exogenous variable in the model and add these equations to the model. With the current procedure one can sometimes discriminate more. The second is if the coefficients of the model have been estimated over a sample period that includes observations beyond t − 1. however. If.8 COMPARING INFORMATION IN FORECASTS 197 concluded about the relative merits of the two. to do so introduces superfluous errors into the model solution. There are four ways in which future information can creep into a current forecast. McNees. For the forecast for period t. say. for example. the model can be . The expanded model effectively has no exogenous variables in it. It can only be said that the large model contains all the information in the smaller models that it has been tested against. the regression results reveal that a large model contains all the information in smaller models plus some. it may be that the good results for the large model are due to a small subset of it.

If one were interested in adjusting for generated regressors. there would be a generated regressor problem. the model with the true coefficients). It is extremely difficult to try to purge the data of the possible use of future information. By “model” in this case is meant the model inclusive of any exogenous variable equations. If the purpose were to evaluate the forecasting ability of the true model (i. However.. The data revision problem is very hard to handle.14 The if the model is not misspecified. For the work in the next chapter nothing has been done about this problem either. To handle the data revision problem one would have to try to construct data for period t that are consistent with the old data for period t − 1.e. Quasi ex ante forecasts may. however. and so on. and so this potential problem has been ignored here. then the rolling estimation forecasts (where estimated parameters vary through time) may carry rather different information from forecasts estimated over the entire sample. the model can be estimated through period t. although the changes in the model that have been made since 1976 are fairly minor. have different properties from forecasts made with a model estimated with future data. This may bias the results in favor of the US model. It is not enough simply to use data that existed at any point in time. parameters change through time).198 7 TESTING COMPLETE MODELS estimated through period t − 1. They are not true ex ante forecasts if they were not issued at the time.g. Models are typically changed through time. a version of the US model was used that existed as of the second quarter of 1976. for the forecast for period t + 1.. If the beginning observation is held fixed for all the regressions. and this is not straightforward. and almost no one tries. For the work in Fair and Shiller (1990). and model builders seldom go back to or are interested in “old” versions. the sample expands by one observation each time a time period elapses. because data for period t are needed to compare the predicted values to the actual values. the correction discussed in Murphy and Topel (1985) could not be directly applied here because 14 Even . but they are forecasts that could in principle have been issued had one been making forecasts at the time. the interest here is in the performance of the model and its associated estimation procedure. If the model is misspecified (e. This rolling estimation was followed for the work in the next chapter. and all the predictions were for the period after this. Forecasts that are based only on information prior to the forecast period will be called “quasi ex ante” forecasts. of course. say period t − 1. estimated parameters will change through time due to sampling error. The third problem—the possibility of using information beyond period t − 1 in the specification of the model—is more difficult to handle. For the work in the next chapter the current version of the US model has been used.

then the estimates of β and γ should both be zero. Murphy and Topel require a single covariance matrix. In the extreme case where there were so many parameters in a model that the degrees of freedom were completely used up when it was estimated (an obviously over parameterized model).12) If neither model 1 nor model 2 contains any information useful for s period ahead forecasting of Yt . these two forecasts should be quasi ex ante forecasts. One can guard against this degrees of freedom problem by requiring that no forecasts be within sample forecasts. s > 0.15 The Procedure ˆ Let t−s Y1t denote a forecast of Yt made from model 1 using information available at time t − s and using the model’s estimation procedure and forecasting ˆ method each period. Note that the estimation procedure used to estimate a model and the model’s forecasting method are considered as part of the model. 16 If both models contain “independent information” in the present terminology. Let t−s Y2t denote the same thing for model 2. say. 15 Nelson (1972) and Cooper and Nelson (1975) do not stipulate that the forecasts be based only on information through the previous period. but the information in. In this case the estimate of the constant term α would be the average s period change in Y . their forecasts will not be perfectly correlated.) The parameter s is the length ahead of the forecast. It should also be noted that some models may use up more degrees of freedom in estimation than others.8 COMPARING INFORMATION IN FORECASTS 199 focus here is on quasi ex ante forecasts. Lack of perfect correlation can arise either because the models use different data or because they use the same data but impose different restrictions on the reduced form. . and with varied estimation procedures it is often very difficult to take formal account of the number of degrees of freedom used up. then β and γ should both be nonzero. model 2 is completely contained in the covariance matrix of the coefficient estimates used to generate the forecasts changes through time because of the use of the rolling regressions. The procedure is based on the following regression equation: ˆ ˆ Yt − Yt−s = α + β(t−s Y1t − Yt−s ) + γ (t−s Y2t − Yt−s ) + ut (7. which is true of quasi ex ante forecasts proposed here. If both models contain information. no account is taken of these procedures here. (In the notation above. If both models contain independent information16 for s period ahead forecasting. it would be the case that the forecast value equals the actual value and there would be a spurious perfect correspondence between the variable forecasted and the forecast.7.

. and White and Domowitz (1984) for the estimation of asymptotic covariance matrices. and in general forecast errors are heteroskedastic. it does not make sense in the current setup to constrain β and γ to sum to one. but the setup and interests are somewhat different. H1 is the hypothesis that model 1’s forecasts contain no information relevant to forecasting s periods ahead not in the constant term and in model 2. If. α = 0.200 7 TESTING COMPLETE MODELS model 1 and model 1 contains further relevant information as well. say that the true process generating Yt is Yt = Xt + Zt . of course. If. Also. this procedure bears some relation to encompassing tests. however.12 for different models’ forecasts and test the hypothesis H1 that β = 0 and the hypothesis H2 that γ = 0. be serially correlated even for the one period ahead forecasts. Say that model 1 specifies that Yt is a function of Xt only and that model 2 specifies that Yt is a function of Zt only. and so it has been assumed to be zero here. Both forecasts should thus have coefficients of one in equation 7. define X as the T × 3 matrix of variables. then the estimates of β and γ would not generally be zero when the mean of the dependent variable is nonzero. as is usually the case for encompassing tests. for example. ˆ 17 If both models contain the same information. This can be done using the procedure given by Hansen (1982). For example. for example. if k period ahead forecasts are considered. Let θ = (α β γ ) .12. Such serial correlation. where k > 1.12.12 were estimated without a constant term. does not appear to be a problem for the work in the next chapter. It is also not sensible in the current setup to assume that ut is identically distributed. It is likely that ut is heteroskedastic. As noted above. If both models’ forecasts are just noise. and β and γ are not separately identified. and H2 is the hypothesis that model 2’s forecasts contain no information not in the constant term and in model 1.17 The procedure is to estimate equation 7. Also. both models’ forecasts were noise and equation 7. then β but not γ should be nonzero. and let ut = Yt − Yt−s − Xt θ. this introduces a k − 1 order moving average process to the error term in equation 7. and so in this case β and γ would sum to two. It also does not make sense in the current setup to constrain the constant term α to be zero. and Obstfeld (1983). ut is simply the forecast error from model 1. whose row ˆ ˆ ˆ t is Xt = (1 t−s Y1t − Yt−s t−s Y2t − Yt−s ). 18 The error term in equation 7. β = 1. then the forecasts are perfectly correlated. the estimates of both β and γ should be zero. and γ = 0. where Xt and Zt are independently distributed.12 could. Cumby. Also.18 Both heteroskedasticity and the moving average process can be corrected for in the estimation of the standard errors of the coefficient estimates. Huizinga.

12 the level of Y could be regressed on the forecasted levels and a constant. Say that model 1 does rolling regressions of Yt − Yt−1 on Xt−1 and uses these regressions to forecast. the levels version of 7.12 regressions the two forecasts tend to be increasingly collinear as time goes on. and the covariance matrix is simply White’s (1980) correction for heteroskedasticity. Say that model 2 always takes the forecast to be bXt−1 where b is some number other than a. If Y is an integrated process.12 can be used. Adding a large number of observations does not cause the regressions to single out the first ˆ model.15) t=j +1 ˆ where θ is the ordinary least squares estimate of θ and s is the forecast horizon. it only has the effect of enforcing that β + (γ b)/a = 1. then any sensible forecast of Y will be cointegrated with Y itself. When s equals 1. In equation 7.13) (7. It should finally be noted that there are cases in which an optimal forecast does not tend to be singled out as best in regressions of the form 7.12. Note that as an alternative to equation 7. This is straightforward once the stochastic simulation has been set 19 The material in this section is taken from Fair (1993c) .7. so that model 2 remains forever an incorrect model.14 is zero. For variables that are not integrated. Thus. running the levels regression with an additional independent variable Yt−1 (thereby estimating β and γ without constraining their sum to one) is essentially equivalent to the differenced regression 7.12. V (θ). essentially they are collinear after the first part of the sample. the estimates of β and γ tend to be erratic. ˆ 7. Say the truth is Yt − Yt−1 = aXt−1 + et . If Y is an integrated process.9 ESTIMATING EVENT PROBABILITIES ˆ ˆ The covariance matrix of θ. is ˆ V (θ) = (X X)−1 S(X X)−1 s−1 201 (7. In the level regression. even with many observations.14) S= 0+ j =1 T ( j + j) j = ˆ ˆ (ut ut−j )Xt Xt−j (7. and one would in effect be estimating one less parameter.9 Estimating Event Probabilities19 Stochastic simulation can be used to calculate the probability of various events happening. the sum of β and γ will thus be constrained in effect to one. the second term on the right hand side of 7.

its estimated probability would be 15 percent. however. of course. If it occurred. which is the assumption upon which the estimation is based. the true coefficients are fixed. (This is assuming. The only extra work for each extra event is keeping track of how often each event occurs in the repetitions. In estimating probabilities by stochastic simulation. Pt can be compared and Watson (1989) do present. 21 I am indebted to Gregory Chow for suggesting to me that one may not want to draw coefficients when estimating probabilities. It is possible for a given event to compute a series of probability estimates and compare these estimates to the actual outcomes. For each repetition one can record whether or not this event occurred. within the context of their leading indicator approach. T . Although coefficient estimates are uncertain. the reason that economic events are stochastic is because of stochastic shocks (error terms). To see how good a model is at estimating probabilities. . there are T values of Pt and Rt available. Many events. it seems best to draw only error terms (not also coefficients).202 7 TESTING COMPLETE MODELS up and the event defined. can be considered. only error terms are drawn. Government policy makers and business planners are obviously interested in knowing the probabilities of various economic events happening. . Model builders who make forecasts typically do not directly answer probability questions. such as two consecutive quarters of negative growth out of five for the period beginning in quarter t. and let Rt denote the actual outcome of At . 20 Stock . Let Pt denote a model’s estimate of the probability of At occurring. Consider an event At . They are consistent with the probability structure of the model. not because the coefficients are stochastic.20 An advantage of estimating probabilities from stochastic simulation is that they are objective in the sense that they are based on the use of estimated distributions. Assume that 1000 repetitions are taken. Consider a five quarter prediction period and the event that within this period there were two consecutive quarters of negative real GDP growth. say. of course. where each value of Pt is derived from a separate stochastic simulation. which is 1 if At occurred and 0 otherwise. .)21 For the estimation of probabilities in the next chapter. that the true coefficients are fixed. In the real world. This procedure for estimating probabilities can also be used for testing purposes.” If probabilities are assigned to the scenarios. They typically present a “base” forecast and a few alternative “scenarios. . they are subjective ones of the model builders. If one computes these probabilities for t = 1. 150 times out of the 1000 repetitions. estimates of the probability that the economy will be in a recession six months hence.

17) where T is the total number of observations. Other solution methods for rational expectations models are summarized in Taylor and Uhlig (1990). If model 1 has lower values than model 2. where Hansen’s method was described. has come to be widely used for deterministic simulations of rational expectations models.24 for example. the Federal Reserve.10 RATIONAL EXPECTATIONS MODELS 203 to Rt for t = 1. . The basic solution method. with 0 being perfect accuracy.3.16) and the log probability score (LP S): T LP S = −(1/T ) t=1 [(1 − Rt ) log(1 − Pt ) + Rt log Pt ] (7. the Canadian Financial Ministry. Chapter 11. 7. Two common measures of the accuracy of probabilities are the quadratic probability score (QP S): T QP S = (1/T ) t=1 2(Pt − Rt )2 (7. It is also possible. called the “extended path” (EP) method. this is evidence in favor of model 1. with 0 being perfect accuracy. Larger errors are penalized more under LP S than under QP S. 24 For example. . . QP S ranges from 0 to 2. however. . These other methods do not yet appear practical for medium size models and up. the extended path method has been programmed as part of the TROLL computer package and is routinely used to solve large scale rational expectations models at the IMF. T . 23 The 22 See. material in this section is taken from Fair and Taylor (1990). The testing procedure is thus simply to define various events and compute QP S and LP S for alternative models for each event.7. to use FIML to estimate models with rational expectations. .10 Full Information Estimation and Solution of Rational Expectations Models23 Introduction The single equation estimation of equations with rational expectations was discussed in Section 4. It has also been used for simulation studies such as DeLong and Summers (1986) and King (1988). and LP S ranges from 0 to infinity. Diebold and Rudebusch (1989). Methods for the solution and FIML estimation of these models were presented in Fair and Taylor (1983) and also discussed in Fair (1984).22 It is also possible simply to ¯ ¯ compute the mean of Pt (say P ) and the mean of Rt (say R) and compare the two means. and other government agencies.

19) where yt is an n–dimensional vector of endogenous variables. . This section has two objectives. and it seems much more promising than was originally thought. yt−p . The model is written as fi (yt . but the preliminary results using the method were mixed. yt−1 . .1. In the process of doing this some errors in the earlier work regarding the treatment of models with rational expectations and autoregressive errors are corrected. αi ) = uit (7. . Et−1 is the conditional expectations operator based on the model and on information through period t −1. . . the full information estimation method has not been tried by others. . more experimenting with the less expensive method has been done. . The lagged values of the endogenous variables are written out explicitly. n) (7. . Since this earlier work. (i = 1. αi is a vector of parameters. This earlier work discussed a “less expensive” method for obtaining full information estimates. and xt is now a vector of only exogenous variables. Chapter 11. Case 1: ρi = 0 Consider solving the model for period s. Et−1 yt . . The first is to discuss the new results using the less expensive method that have been obtained and to argue that full information estimation now seems feasible for rational expectations models. ρi is the serial correlation coefficient for the error term uit . .204 7 TESTING COMPLETE MODELS but probably because of the expense. In what follows i is always meant to run from 1 through n. The Solution Method The notation for the model used here differs somewhat from the notation used in equation 4. something that was only briefly touched on in the earlier work. The following is a brief review of the solution method for this model. and expectations. Et−1 yt+1 . It is assumed that estimates of αi are available. xt . . The second objective is to discuss methods for stochastic simulation of rational expectations models. . however.18) uit = ρi uit−1 + it . that current and expected future values of the exogenous variables . xt is a vector of exogenous variables. parameters. More details are presented in Fair and Taylor (1983) and in Fair (1984). and it is an error term that may be correlated across equations but not across time. Et−1 yt+h . The function fi may be nonlinear in variables.

Given these values. ..25 Given these guesses. . Sometimes information on the steady state solution (if there is one) can be used to help form the guesses. then overall convergence has not been achieved. the entire process above is repeated for k one larger. 25 Guessed values are usually taken to be the actual values if the solution is within the period for which data exist. This solution provides new values for the expectations through period s + h + k—the new expectations values are the solution values. In the above process the expected values of the exogenous variables would be used for all the solutions. Es−1 ys+h were known. otherwise k must continue to be increased until the criterion is met. To check for this. but future predicted values would not affect current predicted values. The model would be simultaneous. . Otherwise. If the expectations Es−1 ys . If increasing k by one has a trivial effect (based on a tolerance criterion) on the solution values for s through s + h. .10 RATIONAL EXPECTATIONS MODELS 205 are available. the last observed value of a variable can be used for the future values or the variable can be extrapolated in some simple way. Es−1 ys+1 . . and so on. then overall convergence has been achieved. . Values of the expectations through period s + h + k + h are first guessed. If the solution values for periods s through s + h depend in a nontrivial way on these guesses. So far the guessed values of the expectations for periods s + h + k + 1 through s + h + k + h (the h periods beyond the last period solved) have not been changed. Es−1 ys+h . The expected future values of the exogenous variables (which are needed for the solution) can either be assumed to be the actual values (if available and known by agents) or be projected from an assumed stochastic process. Es−1 ys+1 . The EP method iterates over solution paths. . the model can be solved again for periods s through s + h + k. This process stops (if it does) when the solution values for one iteration are within a prescribed tolerance criterion of the solution values for the previous iteration for all periods s through s + h + k. the model can be solved for periods s through s + h + k in the usual ways. In practice what is usually done is to experiment to find the value of k that is large enough to make it likely that further increases are unnecessary for any experiment that might be run and then do no further checking using larger values of k. 7. which provides new expectations values. .7. Given these new values. where k is a fairly large number relative to h. the expected values of the exogenous variables being chosen ahead of time. one extra step is needed at the end of the overall solution. This yields values for Es−1 ys . and that the current and future values of the error terms have been set to their expected values (which will always be taken to be zero here). If the expected future values of the exogenous variables are not the actual values.18 could be solved in the usual ways (usually by the Gauss-Seidel technique).

not reduced form errors.206 7 TESTING COMPLETE MODELS 7. Once the expectations are solved for. one could solve the model as above.2 of this paper. where an error term like uis−j +r is computed as ρir uis−j . . no general convergence proofs are available. of course. The following method replaces the method in Section 2. and so there is no guarantee that the particular set found is unique.18 is used to solve for uis−j . for example. The error terms for period t − 1 (uit−1 . it is. including in the engineering literature. Although the certainty equivalence assumption is widely used. be more than one set of solution values. There may. which yields the final solution value ys . the solution is only approximate unless fi is linear. 7. and so the corrections discussed in this section pertain to both sources. Es−1 ys will differ from ys . If convergence is a problem. .18. Case 2: ρi = 0 and Data Before s − 1 Available The existence of serial correlation complicates the problem considerably. based on the assumption that uis−j −1 = 0.2 in Fair and Taylor (1983). To the extent that the expected value of xs differs from ˆ the actual value. The only difference is that the value of an error term like uis+r−1 would be ρir uis−1 instead of zero. . Chapter 11. the set that the method finds may depend on the guesses used for the expectations for the h periods beyond s + h + k. n) depend on expectations that were formed at the end of period t − 2. not always a good approximation. one can sometimes “damp” the solution values to obtain convergence. This assumption is like the linear quadratic approximation to rational expectations models that has been proposed. . the method relies on the certainty equivalence assumption even though the model is nonlinear. Second. of course. the model is solved for period s − j + 1 using the EP method.26 Consider again solving for period s. The actual values of ys−j and xs−j are used for this purpose (although the solution values are used for the expectations) because these are structural errors being estimated. Given the values for uis−j . by Kydland and Prescott (1982).18 is then solved for period s using the actual value of xs . i = 1. ˆ Two points about this method should be mentioned. . where j > 0. Once 26 The material in Fair and Taylor (1983) is also presented in Fair (1984). and so a new viewpoint date is introduced. This case is discussed in Section 2. If there is more than one set. First. The overall solution method first uses the EP method to solve for period s − j . In practice convergence is usually not a problem. but an error was made in the treatment of the second viewpoint date. Since expectations of functions are treated as functions of the expectations in future periods in equation 7. If the values of uis−1 were known.

27 If the absolute values of these errors are within a prescribed tolerance criterion. The overall method begins by guessing values for uis−2 . 27 These are again estimates of the structural error terms. and so on through the solution for period s. If going back one more period has effects on the solution values for period s that are within a prescribed tolerance criterion. and so on until convergence is reached. the new guess for uis−2 is computed as the old guess plus is−1 /ρi .7.19 can be used to solve for is−1 . Given these values. 7. one can solve for period s using the EP method.10 RATIONAL EXPECTATIONS MODELS 207 the expectations are solved for. The errors computed in step (iii) should be the structural error terms. .18 is used to solve for uis−j +1 . and the solution is finished. which can be used in the solution for period s − j + 2. From the solution values for the expectations. where uis−2 is the estimated value on the last iteration (the value consistent with is−1 being within a prescribed tolerance criterion of zero). convergence has been achieved. The solution problem is to find the values of uis−1 that are consistent with this assumption. then overall convergence has been achieved. From the solution for period s. Again.18 and 7. This type of an assumption is usually made when estimating multiple equation models with moving average residuals. which can then be used in the solution for period s + 1 using the EP method. The solution for period s is based on the assumption that the error terms for period s − j − 1 are zero. Given the values of uis−1 . At the point of convergence uis−1 can be computed as ρi uis−2 . To see if the solution values for period s are sensitive to this assumption. the values of uis can be computed. the model can be solved for period s − 1 using the EP method and the fact that uis+r−2 = ρir uis−2 . not the reduced form error terms. It should be noted that once period s is solved for. Step (iii) on page 1176 in Fair and Taylor (1983) is in error in this respect. the entire process is repeated with j increased by 1. otherwise j must continue to be increased. in practice one usually finds a value of j that is large enough to make it likely that further increases are unnecessary for any experiment that might be run and then do no further checking using larger values of j . The model is solved again for period s − 1 using the new guess and the EP method. Case 3: ρi = 0 and Data Before Period s − 1 not Available This case is based on the assumption that is−1 = 0 when solving for period s. period s + 1 can be solved for without going back again. Otherwise. 7.

is of the dimension of the number of stochastic equations in the model. they are predetermined from the point of view of taking derivatives for the Jacobian. given the expectations. which compares to only 5 when there are no expectations. If.208 Computational Costs 7 TESTING COMPLETE MODELS The easiest way to think about the computational costs of the solution method is to consider how many times the equations of a model must be “passed” through. Since the expectations have viewpoint date t − 1. The objective function that FIML maximizes (assuming normality) is presented in equation 7. as noted above. k is 30. For Case 2 above the number of passes is increased by roughly a factor of j if overall convergence is not checked. j is usually a number between 5 and 10. the number of passes is increased by a factor of q +1. Note for both Cases 2 and 3 that the number of passes is increased relative to the non serial correlation case only for the solution for the first period (period s). h is 5. Then the total number of passes for convergence is N2 N1 (h + k + 1). say. and Jt is of the dimension of the total number of equations in the T model. N1 is usually some number less than 10 when the Gauss-Seidel technique is used. FIML Estimation Assume that the estimation period is 1 through T . no additional passes are needed over those for the regular case.3 above and is repeated here for convenience L=− T T log | | + log |Jt | 2 t=1 (7. If k is increased by one to check for overall convergence.250. the total number of passes is slightly more than doubled. If q is the number of iterations it takes to achieve convergence for Case 3 above. In practice q seems to be between about 5 and 10. then the total number of passes needed to solve the model for one period is 11. N2 is 15. Let N2 be the number of iterations it takes to achieve convergence over these periods. The ij element of is (1/T ) t=1 it j t .20) is the covariance matrix of the error terms and Jt is the Jacobian matrix for period t. this check is not always done. Let N1 be the number of passes through the model that it takes to solve the model for one period. although. If period s + 1 is to be solved for. and N1 is 5. Checking for overall convergence slightly more than doubles the number of passes. The EP method requires solving the model for h + k + 1 periods. and so no additional problems are involved for the .

125. In the 25 coefficient problem below. The computed derivative is then the difference in the two solution values of Et−1 yt+r divided by the change in the first element of α. however. The trick is to compute numerical derivatives of the expectations with respect to the parameters and use these derivatives to compute (and thus L) each time the algorithm requires a value of L for a given value of α.817 evaluations of L to converge. There is. however. computing the error terms requires knowing the values of the expectations. say. 28 Note . Therefore. In the serial correlation case α also includes the ρi coefficients. the method gives an approximation of the likelihood function. One simply solves 7. both solutions using the EP method. In the standard case computing for a given value of α is fairly inexpensive. Hence.7. which would be over 3 trillion passes if done this way. a way of cutting the number of times the model has to be solved over the estimation period to roughly the number of estimated coefficients. which themselves depend on α. then 1. and in fact FIML estimates of the US model are presented in the next chapter.250 passes through the model are needed to solve the model for one period and if T is 100. In what follows α will be used to denote the vector of all the coefficients in the model. and this algorithm will not be discussed here. What any algorithm needs to do is to evaluate L many times for alternative values of α in the search for the value that maximizes L. the Parke algorithm required 2. to compute for a given value of α one has to solve for the expectations for each of the T periods. This is only one pass through the model since it is the structural error terms that are being computed. 11. To compute that these solutions of the error term it are only approximations when fi is nonlinear. In the rational expectations case. FIML estimation of moderate to large models is expensive even in the standard case.19 for the it error terms given the data and the value of α.10 RATIONAL EXPECTATIONS MODELS 209 Jacobian in the rational expectations case.28 It should be clear that the straightforward combination of the EP solution method and FIML estimation procedures is not likely to be computationally feasible for most applications. An algorithm that can be used for large scale applications is discussed in Parke (1982).000 passes are needed for one evaluation of and thus one evaluation of L. If. One can first solve the model for a given value of α and then solve it again for the first element of α changed by a certain percent. Suffice it to say that FIML estimation of large scale models is computationally feasible.18 and 7. Consider the derivative of Et−1 yt+r with respect to the first element of α. and some tricks are needed to make the problem computationally feasible.

30 Once these derivatives are computed. Once the maximization algorithm has found the new optimum. derivatives of uit−1 with respect to the coefficients are also needed when the errors are serially correlated. Since one has (from the derivatives) an estimate of how the expectations change when α changes. one does not have to solve the model any more to get the expectations. Convergence is achieved when the coefficient estimates from one iteration to the next are within a prescribed tolerance criterion of each other. when the maximization algorithm changes α and wants the corresponding value of L. one coefficient change per solution. These derivatives can also be computed from the K + 1 solutions. K · T (h + 1) derivatives are computed and stored for each expectations variable. and so no extra solutions are needed in the serial correlation case.29 One solution is for the base values. Once the K+1 solutions of the model have been done and the maximization algorithm has found what it considers to be the optimum. This is the second “iteration” of the overall process. a new iteration performed. although no systematic procedure of trying to find the optimal percentage size was undertaken. they can be used in the computation of for a given change in α. Assuming that the solution method in Case 3 above is used for the FIML estimates. This 29 In the notation presented in Section 7. and so on. and the K solutions are for the K changes in α. For the work here two sided derivatives seemed unnecessary. which are then used in the computation of .1 k rather than K is used to denote the dimension of α. From these K + 1 solutions. At this point the new solution values (not computed using the derivatives) can be used as new base values and the problem turned over to the maximization algorithm again. K is used in this section since k has already been used in the description of the EP method.” “Two sided” derivatives would require an extra K solutions. the model can be solved again for the T periods using the optimal coefficient values and then L computed. since the derivatives are only approximations. one derivative for each length ahead for each period for each coefficient. new base values can be computed. where each coefficient would be both increased and decreased by the given percentage. Five percent seemed to give slightly better results than one percent. This value of L will in general differ from the value of L computed using the derivatives for the same coefficient values. 30 Derivatives computed this way are “one sided. where K is the dimension of α. In other words. .210 7 TESTING COMPLETE MODELS all the derivatives requires K + 1 solutions of the model over the T number of observations. the derivatives are first used to compute the expectations. For the results below each coefficient was increased by five percent from its base value when computing the derivatives. and no further solutions of the model are needed.

The two further iterations for Model 1. Table 7.21 is assumed to be serially correlated: u1t = ρ1 u1t−1 + 1t (7. Because this model is very small and linear. Three solutions of the model over the 50 periods were needed for the derivatives for the first iteration.0260125 and α21 = −. If the coefficients change substantially on this iteration. The results for Model 1 show that convergence was essentially achieved after one iteration. This may improve convergence.23) . The third iteration in particular was unnecessary. and α21 = α22 = α23 . which compares to 61 that would have been needed had the derivatives not been used. For Model 2 the error term in equation 7. Data for this model were generated using normally distributed serially independent errors with zero correlation between equations. led to very small changes. Values of α15 and α21 of . is a version of the wage contracting model in Taylor (1980): y1t = α11 y1t−1 + α12 y1t−2 + α13 Et−1 y1t+1 + α14 Et−1 y1t+2 + α15 Et−1 y2t + α16 Et−1 y2t+1 + α17 Et−1 y2t+2 + u1t y2t = α21 y1t + α22 y1t−1 + α23 y1t−2 + u2t (7. α12 = α14 = 1/6. The coefficient estimates computed this way are α15 = .1 reports the results of estimating three models by FIML using the derivatives. There are two free parameters to estimate.7.1 shows the results using the “derivative” method discussed above.333333 were used for this purpose.21) (7. Fifty observations were generated. ˆ ˆ Table 7.10 RATIONAL EXPECTATIONS MODELS 211 procedure can be modified by recomputing the derivatives at the end of each iteration. a factorization procedure can be used to evaluate L exactly.22) with the restrictions that α11 = α13 = 1/3. At a minimum. This procedure can in turn be used in the maximization of L using an algorithm like DFP. The first model. α15 and α21 . The DFP algorithm was used for this problem since the model was not large enough to require the Parke algorithm. but it obviously adds considerably to the expense. α15 = α16 = α17 .0333333 and −. then overall convergence has not in fact been achieved. and so the method worked quite well. Model 1. which were based on recomputing the derivatives. one might want to recompute the derivatives at the end of overall convergence and then do one more iteration. The difference between L computed using the derivatives and L computed from the full solution after the first iteration is very small.3916.

0450 189. of func.0252994 .0335672 .673 .0466 61 50 37 No.1248 531. Model 2: Taylor Model. α15 ˆ Start Iteration: 1 2 3 4 5 6 .391609 -. The Parke algorithm was used for Model 3.83545.0467 509.0260117 -.0289718 .0903430 ρ1 ˆ . derivatives solution evals.322183 -.212 7 TESTING COMPLETE MODELS Table 7.32211.0178 532.745 .0784 2817 1103 538 258 The DFP algorithm was used for Models 1 and 2. Serial Correlation ˆ L using derivatives ˆ L using full solution 501.635 .837 .322699 -.6022 509.878 .322646 Model 3: Six Equation Model.0200000 .333333 -. of using using full func. The coefficient estimates using the factorization routine and the DFP algorithm are α15 = . This set of values will be called the “exact” ˆ . Start Iteration: 1 2 3 4 170.0467 509.3424 532. evals.0333333 . where ρ1 was set equal to .0470 509.391612 509. 25 Coefficients ˆ ˆ L L No.3100 189.321324 -.321878 -.0467 ˆ L using full solution 508.0462 509.9918 531.9557 531.0784 184.1 FIML Results for Three Models Model 1: Taylor Model.0886905 .1676 532.200000 -.0297 189.1670 189. Derivatives were recomputed after each iteration for Models 1 and 2.7 to generate the data.2047 189. and α21 = −.0260233 .8590 531.0778620 .9032 77 166 103 103 96 90 No.5016 532.1740 531. of func. No Serial Correlation ˆ L using α21 ˆ derivatives α15 ˆ Start Iteration: 1 2 3 .0098 189. but not for Model 3.210860 -.391662 -.0495646 .8234 505.9346 531.600 .889 α21 ˆ -. ρ1 = ˆ ˆ . evals.0738367.7876 531.3381 189.

with coefficient estimates fairly close to the exact answer. it is not meant to be a good approximation of the economy. but not quite. 100[(Mt−1 /Mt−2 )4 − 1] Ct + It + Qt 213 answer. M is the nominal money supply. RHO cnst. The largest value of L occurred after the fourth iteration. P is the GNP deflator. log P Mt . The model is meant for computational exercises only. 531. Y S is an estimate of potential GNP.10 RATIONAL EXPECTATIONS MODELS Table 7. R is the interest rate. Yt = cnst.2 (C is consumption. Yt − Yt−1 . to the exact answer. Model 2 is much harder to solve than Model 1 in requiring a much larger value of k and many more iterations of the solution paths to converge. The method thus moved from L equal to 501. especially with respect to α15 and ρ1 .31 Model 3 is a simple six equation macroeconomic model with 25 coefficients. Et−1 (Yt+1 − Yt ). log Ct−1 . it moved away from it slightly. and one other local optimum was found in the course of this work. but it could not go the rest of the way to 532. (Y St − Yt )/Y St . Rt 6.7.0333. Rt−1 . The equations are shown in Table 7. log(Mt /Pt ) 4. 100[(Yt /Yt−1 )4 − 1]. Y is GNP. log(M/P )t−1 . On iterations 5 and 6. Q is government spending plus net exports.2 Model 3: Six Equations 1. log Yt . The value of L using the exact coefficient estimates (not reported in the table) was 532.9918.1 show that the derivative method got close.0333. When the method was started off from the exact answer. Et−1 100[(Pt+2 /Pt+1 )4 − 1]. log Pt−1 . t.1. log Pt 5. t is the time although not reported in Table 7. It − It−1 3.8234 to L equal to 531. 31 Also. like the case for iterations 5 and 6 in Table 1. Rt cnst. The likelihood function is fairly flat near the top. P M is the import price deflator. Et−1 log Yt+2 . however. The derivatives were computed after each iteration for this problem. . the method moved slightly further away from the answer. I is investment. This basically seems to be a hard computational problem. The results in Table 7. Et−1 [(Y St+1 − Yt+1 )/Y St+1 ]. log Ct 2. Rt . It−1 cnst.9918. Rt cnst. one of which is a serial correlation coefficient.

4. and t. Model 2 is probably extreme in the degree to which future predicted values affect current predicted values.32 The results in Table 7.3. 2. Y St . V4 for Model 3 is used for the stochastic simulation results discussed next. Stochastic Simulation For models with rational expectations one must state very carefully what is meant by a stochastic simulation of the model and what stochastic simulation is to be used for. Future expected values are in equations 1. In other words. and C. In the present case stochastic simulation is not used to results for Model 3 in Tables 7. and this may be one of the reasons results are not as good for it. Since Model 3 is not part of the US model and is not used for any of the work in the following chapters.4. and 4.1 that the use of the derivatives worked quite well. After the first iteration the difference between L computed using the derivatives and L computed from the full model solution is fairly large (189. ˆ The FIML covariance matrix of the coefficient estimates (V4 ) was estimated for each model using the formula 7. These covariance computations are feasible because the expectations derivatives can be used in calculating the derivatives in 7. The estimation period was 1954:1–1984:4. and 5. They have not been updated for present purposes.1 are the same as those in Fair and Taylor (1990). Convergence had been achieved after iteration 4. no further solutions of the model are needed to compute ˆ ˆ V4 in 7. and the longest lead length is 2.1670 − 184. Model 3 is probably more representative of models likely to be used in practice than is Model 2. but the differences are quite small for iterations 2. and Q are in real terms): The exogenous variables in the model are P Mt . I .1 for Model 3 are based on only one set of calculations of the derivatives. 3. for a total of 124 observations. It can be seen in Table 7. 4. The good results for Model 3 are encouraging. The model was solved 26 times for the 124 observations to get the derivatives for the 25 coefficients. there was no need to update.214 7 TESTING COMPLETE MODELS trend.3381). Y S. Y . RH O means that the error term in the equation is first order serially correlated. The Parke algorithm was used for the maximization. 32 The . Qt . Also. The Hansen estimates were then used as starting values for the FIML calculations.4. the results for Models 1 and 2 in Table 7. where the derivatives are evaluated (numerically) at the optimum.1 and 7. The equations were first estimated using Hansen’s method discussed in Section 4.3 are the same as those in Fair and Taylor (1990).

where p is the length of the period. and uis is computed as ρi uis−1 plus the draw for is .7. Assume that the errors are distributed N (0. For present purposes stochastic simulation begins once the expected values have been solved for. however. and so on. each repetition merely requires solving the model for period s. This is one repetition. on the other hand. This is the first . α is the FIML estimate of α and V ˆ ˆ In this case each draw also involves the vector of coefficients. results for more than one period are needed and the simulation is dynamic. ˆ ).5 and 7. V4 ). not the u errors. One can also use this approach to analyze the effects of uncertainty in the coefficients by assuming that the coefficients are distributed N (α. where ˆ is the FIML estimate of from the last subsection. the EP method must be used p times for each repetition. stochastic simulation is performed for period s. The problem is now no different from the problem for a standard model because the expectations are predetermined. If. available from the solution of the model to get the expectations (see Case 2 in the previous subsection). then an estimate of uis−1 is needed for the solution for period s.7 in Section 7. From this distribution one can draw a vector of error terms for period s. Given these draws (and the expectations). If uit is serially correlated as in 7. (Note that the errors are drawn.3. This estimate is. These solution values are the agents’ expectations. 3) using the existing set of coefficient estimates of the model. Given the expected values for periods s through s + h. and so no further work is needed. The estimate of uis−1 is simply taken as predetermined for all the repetitions. and then 4) solving the model for periods s and beyond. This way of solving for the expected values can be interpreted as assuming that agents at the beginning of period s form their expectations of the endogenous variables for periods s and beyond by 1) forming expectations of the exogenous variables for periods s and beyond.) Stochastic simulation is quite inexpensive if only results for period s are needed because the model only needs to be solved once using the EP method.10 RATIONAL EXPECTATIONS MODELS 215 improve on the accuracy of the solutions of the expected values. 2) setting the error terms equal to their expected values (say zero) for periods s and beyond. Another repetition can be done using a new draw of the vector of error terms. Consider the multiperiod problem. Once the expectations are obtained. The expected values are computed exactly as described above—using the EP method. The means and variances of the forecast values can be computed using equations 7. where ˆ ˆ ˆ4 is the estimated covariance matrix of α.19. the model can be solved for period s in the usual ways. the expectations with viewpoint date s − 1 can be solved for and then a vector of error terms and a vector of coefficients drawn to compute the predicted value of yis . As above.

i. then uis+1 is equal to ρi2 uis−1 plus the draw for is+1 . yis+2 is different with viewpoint date s than with viewpoint date s − 1. The process is repeated through the end of the period of interest.. Then given the solution value of yis and the actual value of xs .33 A new set of expectations must thus be computed with viewpoint date s. this is one repetition. Now go to period s + 2 and repeat the process. Given these expectations. They do not perform stochastic simulation themselves. This requires a second use of the EP method. and so on. agents are assumed to solve the model for their expectations for periods s + 1 and beyond. Note that only one coefficient draw is used per repetition. the number of solutions for the expectations would have been one. Agents are assumed to use the original set of coefficients (not the set that was drawn) and to set the values of the error terms for periods s + 1 and beyond equal to zero. and V4 from the previous subsection were used for the draws. as is common with most macroeconometric models. The estimated standard deviations are presented in may also be that the actual value of xs differs from what the agent expected it to be at the end of s − 1.216 7 TESTING COMPLETE MODELS step.e. At the end. and 100 repetitions were performed. In particular. The length of the prediction was taken to be four. Again. If equation i has a serially correlated error. An agent’s expectation of. The FIML estimates of . The results show.3. that the stochastic simulation estimates of the means are quite close to the deterministic simulation estimates. had the length been taken to be one. where another use of the EP method is needed to compute the new expectations. The real use of stochastic simulation is to compute standard deviations or variances. the value of yis is in general different from what the agent at the end of period s − 1 expected it to be (because of the error terms that were drawn for period s). α. per dynamic simulation. The deterministic simulation estimates are simply based on setting the error terms to zero and solving once for each period (as the agents are assumed to do). The overall process is then repeated for the second repetition. where there are now means and variances for each period ahead of the prediction. Now go to period s + 1. After J repetitions one can compute means and variances just as above. 33 It . Stochastic simulation results for Model 3 are presented in Table 7. a vector of error terms for period s + 1 is drawn and the model is solved for period s + 1. say. This meant that the number of times the model had to be solved for the expectations was 400. Also note that agents are always assumed to use the original set of coefficients and to set the current and future error terms to zero.

7 217 Investment Money Supply Price Level Interest Rate Real GNP a = predicted value from deterministic simulation b = mean value from stochastic simulation c = standard deviation from stochastic simulation The results are based on 100 trials.7 billion.94 9.5 556.0 for the price level.7 521. Stochastic simulation has also been used to evaluate alternative international monetary systems using the multicountry models in Carloyzi and Taylor (1985) and Taylor (1988).8 543.8 23.1 8.0046 8. For this work values of it were drawn.5 274.0293 1. .0293 .4 272.0435 1.1 8.6 17.7 532. investment.0 29.10 RATIONAL EXPECTATIONS MODELS Table 7.8 billion. The vector of coefficients α was taken to be fixed.2 269.0437 .8 10.9 3244.09 3273.57 8.1 268.4 2094.1 6.0595 .0110 8.2 3199.0 13. For real GNP.0587 1. Units are billions of 1982 dollars for consumption.5 2149.9 1. for example. Stochastic simulation is in fact likely to be cheaper than even FIML estimation using the derivatives.0762 . 1982=1.7.74 1. It seems that stochastic simulation as defined above is computationally feasible for models with rational expectations.2 533.6 2130.01 1. and real GNP.1 4 2146.4 3309.4 1.9 2113.3 554.8 23.40 .1 9.0125 8.75 8.0751 1.2 543.1 259. which is about one percent of the mean value of $3309.8 35. but not values of the coefficients.5 1.39 8.5 521.4 12. row c in the table.1 3275.3 3 2129.2 264.28 .7 1.3 Stochastic Simulation Results for Model 3 1983 1 Consumption a b c a b c a b c a b c a b c a b c 2095.79 3201. and percentage points for the interest rate.96 3243. the estimated standard deviation of the four quarter ahead forecast error is $35.6 2 2111.1 5.21 3305. billions of dollars for the money supply.0 17.3 259.0 11.2 28.0083 8.9 264.

FIML estimation is computationally feasible using the procedure of computing derivatives for the expectations. on the other hand. . the results in this section are encouraging regarding the use of models with rational expectations.218 7 TESTING COMPLETE MODELS If. and stochastic simulation is feasible when done in the manner described above. FIML estimation is particularly important because it takes into account all the nonlinear restrictions implied by the rational expectations hypothesis. for example. Conclusion To conclude. the FIML estimation period is 100 observations and there are 25 coefficients to estimate. For a stochastic simulation of 8 periods and 100 repetitions. the model has to be solved using the EP method only 800 times. FIML estimation requires that the model be solved 2600 times using the EP method to get the derivatives. It is hoped that the methods discussed in this section will open the way for many more tests of models with rational expectations.

A brief summary of the results is presented in Section 8. Some of the tests in this chapter require a version of the US model in which there are no hard to forecast exogenous variables. and the 2SLS asymptotic distribution is compared to the exact distribution using the procedure discussed in Section 7. VAR5/2.8 estimates event probabilities for the models and compares the accuracy of these estimates across the models using the procedure discussed in Section 7.4 examines the asymptotic distribution accuracy.4.8 to examine the information content of the forecasts from these models. and this chapter applies these techniques to the US model.8 Estimating and Testing the US Model 8.7 uses the procedure discussed in Section 7. and it is discussed in the next section. For the work in this chapter the model has been estimated by 2SLAD. Finally. Section 8. 3SLS. Section 8. and AC models.1 Introduction The previous chapter discussed techniques for estimating and testing complete models. Section 8. 219 . median unbiased (MU) estimates have been obtained for 18 lagged dependent variable coefficients using the procedure discussed in Section 7. In Section 8.9.2. Section 8. 2SLAD is discussed in Section 4.6 the total variances discussed in Section 7.5 compares the five sets of estimates. and 3SLS and FIML are discussed in Section 7. The rest of this chapter is concerned with testing.3 presents the MU estimates.5.9. Also. and FIML in addition to 2SLS. and Section 8. This version is called US+. VAR4.7 are computed and compared for the US.4.

DELD. Those for which autoregressive equations are not estimated are: AG1. and the second block is the 91×91 covariance matrix of the exogenous variable error terms. for variables that were negative or sometimes negative. J J P . and in this sense it is comparable to the VAR and AC models discussed in Section 7.2. P 2554. The US+ model thus has no hard to forecast exogenous variables. the error terms in the structural equations were assumed to be uncorrelated with the error terms in the exogenous variable equations. T AU G. D811824. MU H .6. Each of the additional equations explains an exogenous variable and is an eighth order autoregressive equation with a constant term and time trend added. D714. T I . In other words. D721.2 US+ Model The US+ model is the US model with an additional 91 stochastic equations. T . H M. H F S. the variables created from peak to peak interpolations.220 8 TESTING THE US MODEL 8. and W LDS. and variables that are constants or nearly constants. LAM. This assumption is consistent with the assumption in the US model that the structural error terms are uncorrelated with the exogenous variables. The first block is the 30×30 covariance matrix of the structural error terms. I H H A. . DELH . W LDG.8 that this treatment of the exogenous variables may bias the results against the US model. however. and for variables that were sometimes close to zero. I KF A. D691. Logs were not used for ratios. Many of the exogenous variables may not be as uncertain as the autoregressive equations imply. DELK. D692. AG3. AG2. CDA. Remember. D794823. The covariance matrix of the error terms in the US+ model is 121×121. from the discussion in Section 7. All the exogenous variables in the model are listed in Table A. DD772. Excluding these variables left 91 variables for which autoregressive equations are estimated. D831834. and for purposes of the stochastic simulation work it was taken to be block diagonal. the dummy variables. T AU S. which have no exogenous variables other than the constant term and time trend. Logs were used for some of the variables. Equations are estimated for all the exogenous variables in the model except the age variables. T XCR. The estimation technique was ordinary least squares.

the same draws of the error terms were used for each iteration. 2 To lessen stochastic simulation error.1 are the 90 percent confidence values.3 MU Estimates of the US Model1 The procedure for obtaining median unbiased (MU) estimates of a model is explained in Section 7. median unbiased estimates of the lagged dependent variable (LDV) coefficients were obtained for 18 of the 30 stochastic equations. defined as the difference between the 2SLS estimate and the MU estimate. Also presented in Table 8. Starting from these values. The second MU confidence value is minus the difference between the median estimate and the estimate at which five percent of the estimates are above it.4 were done 3 times).. The second 2SLS confidence value is the absolute value of the first value. The first MU confidence value for each coefficient is minus the difference between the median estimate and the estimate at which five percent of the estimates are below it.e.1. the value of J in step 3 in Section 7.1.2 The results for the LDV coefficient estimates are presented in Table 8. The first 2SLS confidence value for each coefficient is minus 1. The MU values are computed using the coefficient estimates from the 500 repetitions on the last iteration. and the results are reported in this section. and lagged dependent variable and with the LDV coefficient equal to the 2SLS coefficient estimate presented in the table. These biases are interpolated from Table III in Andrews (1993). for a total of 158 observations.000 = (500 repetitions)×(30 stochastic equations)×(158 observations). and each of the 18 equations is estimated for each repetition. time trend. The number of repetitions per iteration (i. The estimation period was 1954:1–1993:2. The estimates for the other 12 equations were fixed at their 2SLS values.4) was 500. The model is solved dynamically over the estimation period for each repetition.3 MU ESTIMATES OF THE US MODEL 221 8. 1 The material in this section is taken from Fair (1994a). The number of errors drawn per iteration is 2. This procedure was carried out for the US model.. The results in this paper are the same as those in Table 8. the largest difference between the successive estimates of any LDV coefficient was less than .e. The bias for each coefficient estimate.001 in absolute value. The starting point was the set of 2SLS estimates in Chapter 5. is presented in the table. .370. After 3 iterations (i. Convergence thus occurred very quickly.4. The “Andrews bias” in the table is the exact bias for an equation with a constant term.8. after steps 3 and 4 in Section 7.645 times the 2SLS estimate of the asymptotic standard error of the LDV coefficient estimate.

The .057 .035 -. 5.064 .082 -. and so the results suggest that the bias of a typical macroeconometric equation is on average less than the bias of an equation that includes only a constant term.033.003 -.020 -.776 . CS 2.034 -.042 first number for 2SLS is minus 1.032 -.020 -. The 2SLS and MU confidence values in Table 8. IM 30. The average of the left tail values is −.054 -.892 Bias -.016 -.017 .104 .036 -.084 . The average Andrews bias.532 .048 .077 -.025 -.017 -.027 .003 -.091 -.034 -.055 . MF 23.047 .033 -.050 .957 . CD 4.043 -.081 -. CUR 27.027 -.104 -.036 -.071 -.052 -.025 .053 .008 -.020 -.070 -.034 .025 -. PF 11.018.018 .042 .036 -.082 . b The first number for MU is minus the difference between the median estimate and the estimate at which five percent of the estimates are below it.575 . LM 9. 1.074 -.094 -. L2 7.083 -.048 -.022 .040 -.104 -.027 -. CN 3.051 for 2SLS. IKF 17.1 show that the estimated biases are zero to three decimal places for 2 of the 18 coefficients and negative for the rest. Y 12.059 -.003 -.890 .051 -.070 .222 8 TESTING THE US MODEL Table 8.008 .1 are fairly similar.067 -.002 -.031 .074 -.079 .057 for MU and −. L1 6.010 . The average bias across the 18 estimates is −. and 9. The second number for MU is minus the difference between the median estimate and the estimate at which five percent of the estimates are above it.011 . In only four cases in the table is the Andrews bias smaller in absolute value— equations 2.035 -. time trend.035 -. IHH 5. and lagged dependent variable.842 .904 .091 .002 -.040 -.040 .059 -.074 .034 . RM 26.053 .012 -.012 .078 .620 . MH 10.003 -.033 90% Confidence Values MUb 2SLSa -.012 -.035 -. The results in Table 8.987 . RB 24.049 -. The second number for 2SLS is the absolute value of the first number.104 -.863 .027 -.042 .872 .943 .025 -.010 -.881 .051 -.1 Estimated Bias of 2SLS Lagged Dependent Variable Coefficient Estimates Eq.048 -.896 .042 -.919 .033 -.293 -.059 .051 .010 -. 7.029 -.055 -.027 .036 -. L3 8. is −.018 -.009 .064 -.055 .031 -. on the other hand.000 .018 Andrews Bias -.022 -.031 -.054 .034 -.052 .050 -. RS Average a The 2SLS .060 .000 -.645 times the 2SLS estimate of the standard error of the LDV coefficient estimate.

1 are quantitatively important regarding the properties of the model. 8. This is something that most model builders probably know from experience.1. Again. and summary results are presented for all 166 coefficients.3.042 for MU and .051 for 2SLS.5 and 11. Detailed results are presented for the same 18 coefficients that were examined in Table 8.2 than in Table 8.5 the sensitivity of the multiplier properties of the model to the use of the estimates is examined. that the 2SLS estimates of the LDV coefficients are biased downwards. It will be seen that the use of the MU estimates has little effect on the predictive accuracy of the model and on its multiplier properties.4 Asymptotic Distribution Accuracy3 The procedure for examining the accuracy of asymptotic distributions was discussed in Section 7.2. With hindsight. The left tail of the distribution is thus thicker than the right tail. What they basically say is that if one changes a LDV coefficient estimate by about half of its estimated standard error and then reestimates the other coefficients in the equation to reflect this change. This question is examined in Sections 8. the present results are perhaps not surprising.8.4 ASYMPTOTIC DISTRIBUTION ACCURACY 223 average of the right tail values is . In Section 8. It is carried out in this section for the US model.5. 4 The bias estimates are slightly different in Table 8.5 the sensitivity of the predictive accuracy of the model to the use of the MU estimates is examined. The bias results for the 18 coefficients show.3. although the differences are fairly minor. namely the LDV coefficients of the 18 equations. although work with other models should be done to see if the present results hold up. There are 166 coefficients to estimate in the model. An interesting question is whether the biases in Table 8. For all the coefficient estimates the right tail value is less than the left tail value in absolute value. It is clear that the MU confidence interval is not symmetric around the median estimate. The results in this paper are the same as those in Table 8.1 because they are . These results thus suggest that macroeconometric model builders have not missed much by ignoring the Orcutt and Hurwicz warnings 40 years ago.2. the fit and properties of the equation do not change very much.1.4 with the average 3 The material in this section is also taken from Fair (1994a). and in Section 11. For the present results the US model was simulated and estimated 800 times. and so the results from this exercise consist of 800 values of 166 coefficients.5. the 2SLS estimates in Chapter 5 were used as the base estimates. as in Table 8. A summary of these results is presented in Table 8.

836 .0 11. CD 4.4 12.9 5. one can compute the value above which k percent of the coefficient estimates should lie if the asymptotic standard error is accurate.4 2.954 .016.0 10.957 .1 17.3 24.0 21.5 4.4 4. and for k equal to 5 the multiplier is 1.1 8.8 10.1 18.0 22.4 13.842 .850 . 2SLS .6 14.5 2. Given for a particular coefficient estimate the 2SLS estimate of its asymptotic standard error.004 -.3 2. .3 5 0.919 .6 11.1 5.882 .2SLS -.2 is to compare the left tail and right tail estimated probabilities to the values implied by the asymptotic distribution. L2 7.64.001 -.3 20.4 4.9 17.003 -.863 .4 9.9 5.8 7.4 4.9 19.1 4.9 7. L3 8. Let pik be the estimated probability for coefficient i for the asymptotic value of k percent.1 were not done for the results in Table 8.846 . LM 9.9 1.595 .006 -.5 3.3 5.5 13.5 11.022 -.8 15.1 11.8 5.025 -.4 8.892 Med.5 19.0 6.889 Med.1 18.4 Right Tail 10 20 0.620 .8 4. Remember from Section 7.8 23.4 2.984 .9 17.0 14.5 6.6 3.6 9.1 8.8 22.026 -.040 .0 26.0 4.554 .839 .987 .8 23. MF 23.292 -.575 .3 3.1 3.8 8.9 3. CN 3. IM 30.1 20.3 14.5 13. From the 800 coefficient estimates one can compute the actual percent of the coefficient estimates that lie above this value.6 19.2 3. RB 24.6 0.8 1.8 10.4 2. For k equal to 20.6 2.0 2.045 -.6 15.919 . Y 12.293 -.4 15.1 20.012 -.4 3.3 18.532 .6 12.0 3.5 17.3 14. PF 11.8 4.5 4.8 18.931 .4 9.1 8. For k equal to 10 the multiplier is 1. The main point of Table 8.0 2.731 .8 8. RS MEAN(18) MAE(18) MEAN(166) MAE(166) bias being −.3 9.003 -.2.9 9.904 .003 -.8 1.7 3.4 10.9 13. MH 10.4 8. this value is the median plus 0.9 19.6 8.224 8 TESTING THE US MODEL Table 8.9 9.5 6. L1 6.016 -.28. IHH 5.8 12.896 .4 27.9 25.046 -.5 5.016 5 0. .021 -.1 1. CUR 27.0 5.943 .4 6. IKF 17.6 7.4 17.890 .000 -. RM 26.84 times the estimated asymptotic standard error. CS 2.4 8.856 .877 .0 11.881 .6 1.8 Left Tail 10 20 1.3 7. This is as expected.5 how these percentages are computed.8 4.1 3.6 22.0 22.8 6.6 13.4 15.3 4.516 .872 .3 8.2 Asymptotic Distribution Accuracy Eq.2 18.5 2. These are the right tail based on 800 rather than 500 repetitions and because the iterations done for the results in Table 8.9 4.5 16.0 1.4 4.034 -.6 18.4 21.024 -.039 .6 7.9 24.776 .001 .

it means that the unit root problems that have received so much attention in the econometric literature are not likely to be of much concern to macro model builders. The corresponding right tail means are 4. 3. although the mean absolute errors reveal that there is some dispersion across the coefficients. 2SLAD. While the existence of unit roots can in theory cause the asymptotic distributions that are relied on in macroeconometrics to be way off. The means of the 5.3.5| across the 18 coefficients divided by 18. and FIML estimates. and 18. with mean absolute errors of 2. one can thus compute values of pi5 .5 A Comparison of the Estimates Section 8. In general the asymptotic distribution seems to be a good approximation.0. The first step for the results in this section was to compute the 2SLAD. 3SLS. This section compares the closeness of the 2SLS.8. 10.3. with mean absolute errors of 2. .4. In addition. If this result holds up for other models.6. There are some computational tricks that are needed to obtain these estimates. 8. and pi20 . where 5.5 A COMPARISON OF THE ESTIMATES 225 percents. Values of these probabilities for each tail are presented in Table 8.2. For example. the mean absolute errors around the means are presented for the 18 and 166 coefficients. These tricks are discussed in Fair (1984).8.9. respectively. fairly small.3 examined the closeness of the 2SLS and MU estimates.7. Consider the results for the 166 coefficients in Table 8. and so on average the asymptotic distribution has thicker tails than does the exact distribution. however.5 is the mean.2. and 4. 3SLS. in practice the asymptotic distributions seem fairly good.2 for the 18 LDV coefficient estimates. and 20 percent left tail values are 5. respectively.1.3. 8. 9. and 17.4. A similar procedure can be followed for the left tail percents. These differences are. The overall results suggest that the use of the asymptotic distribution is not in general likely to give misleading conclusions. Also reported in the table are the means of the probabilities across the 18 coefficients and across the 166 coefficients. pi10 . the mean absolute error for the left tail pi5 for the 18 coefficients is the sum of |pi5 − 5. These mean values are less than the asymptotic values (except for the equality for the 5 percent left tail value). It also compares the predictive accuracy of the model for all five sets of estimates. and 4. For each tail and each coefficient i. The closeness of the asymptotic distribution to the exact distribution is an important result. 3. and FIML estimates.

0 2 32 3. The main conclusion from this comparison is that the estimates are fairly close 2SLAD computational problem is discussed in Section 6.28 = 0. and these are also listed in Table A. 137 were estimated by 3SLS and FIML.5 Of the 166 coefficients. with the remaining coefficients being fixed at their 2SLS values.3.5.4. 5 The . The Parke (1982) algorithm was used for the 3SLS and FIML estimates.3 Comparison of 2SLS.5 4 51 2. (Remember that the coefficients for equations 19 and 29 were obtained in the manner discussed in Section 5. and FIML Estimates Number of estimates greater than . 18.3.5.81 and this discussion will not be repeated here. 19. the 3SLS problem in Section 6.7 in Appendix A.9 rather than by 2SLS. for a discussion of sample size requirements and the estimation of subsets of coefficients. 6 The equations whose coefficients were fixed for 3SLS and FIML are 15. Section 6.3.) In addition.4. and 29.6 All 166 coefficients were estimated by 2SLAD. 21. the following other coefficients were fixed: the two autoregressive coefficients in equation 4.0 standard errors away from the 2SLS estimates 137 Coefficients: .0.0. 3SLS.0 22 70 1. and the four dummy variable coefficients in equation 27. 7 The choice of first stage regressors for 3SLS is discussed in Fair (1984).5. 25. Section 6.3.5. The first stage regressors that were used for 3SLS are listed in Table A.0 0 13 2 6 Number of sign changes from 2SLS estimates 166 Coefficients: 2SLAD 62 16 4 2 1 3 Average ratio of 2SLS standard error to 3SLS standard error (137 coefficients) Average ratio of 3SLS standard error to FIML standard error (137 coefficients) = 1. 28. These coefficients were fixed to lessen potential collinearity problems. and 3. 20. See Fair (1984).7.2 in Fair (1984). 1. and the FIML problem in Section 6. 2SLAD.226 8 TESTING THE US MODEL Table 8. A comparison of the four sets of estimates is presented in Table 8. the coefficients of T and DD772 · T in equations 13 and 14.5 3SLS FIML 69 101 1.7 The same first stage regressors were used for 2SLAD as were used for 2SLS. 1. 2.5.

This examination is presented in Table 8. the 2SLS standard errors are on average 28 percent larger than the 3SLS standard errors. The average ratio of the 2SLS standard error to the 3SLS standard error across the 137 coefficients is 1. This discussion will not be repeated here.8. The RMSE procedure ignores exogenous variable differences and possible misspecifications. since one would not expect for a correctly specified model that the use of different consistent estimators would result in large differences in the estimates. six. One. Of the 137 3SLS estimates. pp. There are no exogenous variable differences except for the fact that different coefficients multiply the same exogenous variables across versions.3 shows the efficiency gained from using 3SLS over 2SLS. four. 8 If . In other words.3 shows that the 3SLS standard errors are on average smaller than the FIML standard errors. Another way to compare the different sets of coefficient estimates is to examine the sensitivity of the predictive accuracy of the model to the different sets. The second to last result in Table 8. the 3SLS standard errors are on average 19 percent smaller than the FIML standard errors. There are also no specification differences. 16 were greater than one standard error away from the 2SLS estimate. where each of the 158 different models were being compared. 157 two quarter ahead predictions. and so misspecification effects differ only to the extent that misspecification is differentially affected by the size of the coefficients across versions. The average ratio of the 3SLS standard error to the FIML standard error across the 137 coefficients is . three. The smaller 3SLS than FIML standard errors is a typical result. The last result in Table 8.8 There are 158 one quarter ahead predictions. These problems are less serious when it is simply different estimates of the same model being used. and eight quarter ahead RMSEs are presented for four variables for each set of estimates. with the FIML estimates being the farthest apart. two. 6 for FIML. For the FIML estimates. and so on through 151 eight quarter ahead predictions. and 2 were greater than two standard errors. only 22 were greater than one 2SLS standard error away from the 2SLS estimate.5 A COMPARISON OF THE ESTIMATES 227 to each other. and 3 for 2SLAD. The prediction period is the same as the estimation period. the use of RMSEs in the manner done here would not be appropriate and one should use a method like the one in the next section. There were 2 sign changes for 3SLS. In other words. 245–246. and 32 were greater than two standard errors.4. The closeness of these estimates is encouraging. namely 1954:1–1993:2. and a possible reason for it is discussed in Fair (1984).28. 70 were greater than one standard error away from the 2SLS estimate. and only 2 were greater than two standard errors. These predictions are all within sample predictions. Of the 166 2SLAD estimates.81.

04 1.29 1.98 1.55 1.40 0.82 1.56 1.90 0.70 0.06 1.09 0.55 1.71 1.46 1. which are noticeably less accurate than the others. and in general the differences are quite small.91 1.81 1.68 0.58 1.76 0.40 0.54 1.55 0.4 RMSEs for Five Sets of Coefficient Estimates for 1954:1–1993:2 for the US Model 1 2SLS 2SLAD 3SLS FIML MUE 0.30 1.62 1.05 1.63 0.79 0.06 1.90 0.02 1.95 1.57 0.75 0.29 0.52 1. The results in Table 8.52 1.56 0.28 0.54 0.11 1.60 0.33 1.73 0.77 1.80 1.53 1.20 1.68 Number of Quarters Ahead 2 3 4 6 GDP R: Real GDP 1.02 1.89 1.16 0.42 1.40 0.05 1.54 0.40 1.61 1.4 show that the RMSEs are very similar across the five sets of estimates.87 0.78 0.75 0.02 1.90 0.59 1.40 1.98 1.07 1.58 0.53 U R: Unemployment Rate 2SLS 2SLAD 3SLS FIML MUE 0.45 1.67 1.64 0.30 0.29 1.228 8 TESTING THE US MODEL Table 8.58 1.97 0.52 0.28 1.32 0.20 1.42 1.30 0.78 0.60 1.39 1.30 0.60 0.34 2.72 1.52 0.97 1.57 0.36 1.15 1.71 Errors are in percentage points.33 1. simulations is based on a different starting point. My experience with the FIML estimation of macroeconometric models is that FIML estimates are the most likely to differ in large ways from other estimates and that when they do differ they generally lead to a poorer .69 0.01 0.02 1.24 1.42 1. No one set of estimates dominates the others.03 1.68 1.28 1.72 1.00 1. The largest differences occur for the FIML predictions of the price deflator.40 0.03 1.21 1.78 1.60 0.69 0.10 RS: Bill Rate 2SLS 2SLAD 3SLS FIML MUE 0.58 2.54 8 GDP D: GDP Deflator 2SLS 2SLAD 3SLS FIML MUE 0.28 1.27 1.78 0.68 0.52 1.62 0.

The results for the four models are presented in Table 8.7. 554. The estimated covariance matrix of the error terms. and the bill rate. and most of this section is a discussion of this table. Again. They are slightly worse than the 2SLS estimates for the price deflator.7 to compare the US model to the VAR5/2.5 for four variables: real GDP. only the structural error terms were drawn.4 is consistent with the properties of a simple equation with only the lagged dependent variable as an explanatory variable.8. and 1000 repetitions were made for each row.5. In the notation in Section 7. these differences are small. the GDP deflator.6 Predictive Accuracy This section uses the method discussed in Section 7. however. The variances can thus be compared across models.4 is again encouraging.4 do in fact quite well. The fact that the MU results are similar to the others in Table 8. the unemployment rate. ˆ . and they tend to lead to a better fitting overall model. each value in a b row is the square root of σitk . VAR4. Consider the a and b rows for the US model first. say yt = αyt−1 + t . The closeness of the results in Table 8. since one would not expect there to be large differences of this sort for a model that is a good approximation of the economy. and AC models.6 PREDICTIVE ACCURACY 229 fitting model. Standard errors rather than variances are presented in the table because the units are easier to interpret. The 3SLS estimates in Table 8. ˜2 The 2SLS estimates in Chapter 5 were used for this work. Malinvaud (1970). Remember from the discussion at the end of Section 5. and for the b row. prediction errors seem to be little affected by coefficient estimation bias.6. but slightly better for the other three variables. For the a row. is 30×30. The method computes forecast error variances for each variable and period ahead that account for the four main sources of uncertainty of a forecast. both the structural error terms and the coefficients were drawn. The latter three models are discussed in Section 7.9 that equations 19 and 29 are taken to be . The simulation period was 1991:1–1992:4. shows for this equation that the expected value of the prediction error is zero when the distribution of t is symmetric even if the estimate of α that is used to make the prediction is biased. 8. The present results show that even for much more complicated models. For example. 3SLS estimates are generally closer to 2SLS estimates than are FIML estimates. There are considerable computations behind the results in Table 8. p.

73 1.87 1.41 2.48 .86 1.05 .62 1.23 1.99 1.07 2.40 1.55 1.20 1.82 .73 US: a b c d GDP D: GDP Deflator .93 1.91 1.01 .60 1.49 .74 2.61 .77 VAR5/2: a .40 .08 1.91 2.18 1.64 2.36 1.14 1.75 b .31 d .14 1.78 1.73 .31 1.42 2.70 1.80 .44 .01 1.22 1.08 AC: a .35 1.53 1.92 .58 .44 1.63 .64 2.99 2.93 2.43 .17 1.70 .80 b .56 .84 2.78 1.52 1.62 .29 1.86 .14 1.17 1.97 2.33 1.07 2.36 .47 1.80 1.95 2.69 1.52 2.98 2.32 1.71 2.88 1.67 .53 .53 2.96 VAR4: a .18 1.32 3.94 3.53 1.81 1.14 .64 .98 1.91 1.84 .49 3.22 2.12 3.86 .27 b .51 b .52 d .40 .99 1.00 1.97 1.42 1.15 1.80 .65 .64 1.87 1.04 2.05 2.44 .65 2.52 3.34 1.17 1.39 1.42 2.56 1.64 1.21 2.57 2.49 1.75 .96 1.77 2.69 .36 1.61 1.97 1.98 1.24 1.60 .230 8 TESTING THE US MODEL Table 8.79 1.64 2.49 .20 1.75 1.15 1.89 1.51 2.55 1.86 2.34 .82 1.14 .29 VAR4: a .89 .81 2.82 d 1.10 1.05 1.32 2.74 2.5 Estimated Standard Errors of Forecasts for Four Models 1991 1 US: a b c d 2 3 4 1 2 GDP R: Real GDP .20 2.85 1992 3 4 VAR5/2: a .62 1.64 2.89 .18 1.32 1.51 .38 2.92 .33 .58 .18 1.38 1.70 1.58 1.48 3.49 1.36 1.74 2.93 2.45 .90 1.74 .03 1.72 .83 d .30 b .27 d .36 .

23 .18 1.74 1.45 .32 2.56 .96 1.76 .65 .54 1992 3 4 231 VAR5/2: a .72 3.67 2.52 .45 1.33 2.56 3.87 .30 .34 1.75 3.70 .62 1.77 1.28 .90 1.80 .14 a = Uncertainty due to error terms.17 1.71 .79 .87 .37 1.57 .40 1.63 .00 .15 VAR4: a .34 US: a b c d RS: Bill Rate .46 1. and exogenous variable forecasts.21 1.91 4.31 .46 1.63 b .89 .01 1.30 1.02 1.69 .81 .11 2.59 1.12 2. d = Uncertainty due to error terms.02 1.87 4. Errors are in percentage points.42 .60 .44 .31 .07 1.24 .24 1.53 3.58 1.47 . and the possible misspecification of the model.66 .14 1.35 1.84 1.31 .13 1.83 .87 1.67 b .71 1.28 1.46 . exogenous variable forecasts.65 2.07 1.95 4.55 .37 2. .35 1.11 1.01 1.63 2.30 .90 1.05 1.63 1.14 1.57 1.47 1.24 d .44 .85 1.72 .25 d .03 1.8.88 1.91 1.77 .47 2.27 .08 .30 .53 1.52 1.23 b .75 .37 1.24 b .07 1.30 1.03 1.65 d 1.58 .27 .66 d 1.74 3.63 1.02 4.49 . c = Uncertainty due to error terms.53 1.70 . b = Uncertainty due to error terms and coefficient estimates.79 1.5 (continued) 1991 1 US: a b c d 2 3 4 1 2 U R: Unemployment Rate .31 VAR5/2: a .87 1.06 5.98 1.02 1.24 1.86 .23 1.99 1.22 .21 1.6 PREDICTIVE ACCURACY Table 8.03 1.64 .91 1.96 1.58 .93 1.55 .27 .85 3.11 1.82 .08 1. coefficient estimates.00 4.54 .51 1.29 VAR4: a .52 1.54 .62 .96 1.03 .60 3. coefficient estimates.75 1.

3. and again 1000 repetitions were made.232 8 TESTING THE US MODEL stochastic for purposes of computing ˆ even though their coefficients are not estimated in a traditional way. The dimension of V2 is thus 166×166 rather than 175×175. Data for H O prior to 1956:1 were constructed in the manner discussed in Section 3. . ∗ a xiT = xiT + vi1 + vi2 + · · · + viT . Also. These are the same equations that are used for the US+ model discussed in Section 8.2. However. Remember also that equation 19 is divided through by |AF + 10| and that equation 29 is divided through by |AG| before computing the error terms to be used in computing ˆ . for purposes of computing ˆ . but only 166 freely ˆ estimated. The estimation period for equation 15 begins in 1956:1 rather than 1954:1. an eighth order autoregressive equation with a constant and time trend was estimated for each of 91 exogenous variables. is 166×166. There are thus a total of 175 coefficients in the model. First. ˆ The estimated covariance matrix of the coefficient estimates. The formula for this matrix is given in equation 4. Then for prediction period 1 through T . the period beginning in 1954:1 was used for equation 15. V2 . The estimation period was 1954:1–1993:2. the values for xit for a given repetition were taken to be ∗ a xi1 = xi1 + vi1 ∗ a xi2 = xi2 + vi1 + vi2 . the coefficients in equations 19 and 29 were taken to be fixed. Let si denote the estimated standard error from the equation for exogenous variˆ able i. and exogenous variable errors were drawn.5 in Chapter ˆ 4. Let vit be a normally distributed random variable with mean zero and a variance si2 : vit ∼ N(0. For purposes of computing V2 . structural errors. which explains H O. This is the estimation period used for estimating all the equations except 15. For this row. Consider next the c row for the US model. Finally. There are five of these coefficients. four of the coefficients in the wage equation 16 are constrained and thus not freely estimated. si2 ) for all t. coefficients. let xit ∗ given repetition. The estimation period for ˆ was 1954:1–1993:2.2 except that all the equations here are linear whereas many of the equations for US+ are in logs. Let xit be the actual value of exogenous ˆ ˆ ∗ be the value of variable i used for a variable i for period t. The procedure that was used for the exogenous variable errors is the following.

where the beginning was 1956:1). say. Dummy variables were thus added when appropriate as the length of the estimation period increased. Dummy variables whose nonzero values begin after 1976:2 obviously cannot be included in the version of the model estimated only through 1976:2. The estimation technique was 2SLS. In computing the disk values. Given the 68 sets of estimates. For each estimation period the covariance matrix of the structural error terms. The variable D811824 in equation 21 was added for the first time for the period ending in 1981:1. Turn next to the d row for the US model. the error vi2 is carried along from quarter 2 on. the first period will persist throughout the forecast period.7. All estimation periods began in 1954:1 (except for equation 15. Given that the simulation period is 8 quarters in length and given that there are 91 exogenous variables.8. V2 . Finally. . and so forth. For this work V2 was taken to be block diagonal. the variables involving DD772 in equations 13 and 14 were added for the first time for the period ending in 1983:1. Both structural error terms and coefficients were drawn . were estimated along with the coefficients. the model was estimated and stochastically simulated 68 times. The error vi1 is carried along from quarter 1 on. The first estimation period ended in 1976:2.6 PREDICTIVE ACCURACY 233 ˆ where each vit (t = 1. This row required by far the most computational work. If this is true. · · · .7. and the covariance matrix of the coefficient estimates. This treatment implies that the errors are assumed to pertain to changes in the exogenous variables. 728 exogenous variable errors are drawn for each repetition. si2 ) distribution. The variable D794823 · P CM1−1 in equation 30 was added for the first time for the estimation period ending in 1979:4. The first simulation period began in 1976:3. where dik is the mean of the disk values discussed in Section 7. each value in a d row is the square root of σitk . the second in 1976:3. and so on through 1993:1. Given the way that many exogenous variables are forecast. 68 stochastic simulations were run. and so on through 1993:2. the square of each d row value is equal ˆ2 ¯ ¯ to the square of the c row value plus dik . the assumption that the errors pertain to the changes in the variables may be better than the assumption that they pertain to the levels. All simulations were outside the estimation period. Each simulation period was of length 8 quarters subject to the restriction that the last quarter for predictions was 1993:2. it may be that any error in forecasting the level of a variable in. T ) is drawn from the N (0. In the notation in Section 7. by extrapolating past trends or taking variables to be unchanged from their last observed values. the second in 1976:4. Put another way. and the variable D831834 in the same equation was added for the first time for the period ending in 1983:1.

Once these calculations have been done and the d row values computed. As was the case for the US model.5. the US model does substantially better. VAR5/2 is more accurate for all variables and all quarters except for the one quarter ahead prediction of the bill rate. Given these means and given the c row values in Table 8. and so on. The same procedure was followed for the other three models except that the other models have no exogenous variables and so no c row values are needed. where the two models essentially tie. the covariance matrices of the coefficient estimates were taken to be block diagonal. however. there were 67 values of disk computed. For the GDP deflator (GDP D) the US model is worse than VAR5/2 for the first five quarters and better for the remaining three. the bill rate. the d row values could be computed. For the bill rate (RS) the US model is better than both VAR models for all quarters. As discussed in Section 7. For the one quarter ahead prediction (k = 1). and consider first the US model versus the two VAR models. For real GDP (GDP R) the US model is better than VAR5/2 for the first four quarters and slightly worse for the remaining four. each model is on an equal footing with respect to the d row values in the sense that the four main sources of uncertainty of a forecast have been accounted for. For the unemployment rate (U R) the US model is better than VAR5/2 for the first four quarters and essentially tied for the remaining four. Remember. The US model is better than VAR4 for all quarters. The estimation technique was ordinary least squares. one can compare the models. that the present results are based on the use of the autoregressive equations .234 8 TESTING THE US MODEL for these simulations (using the appropriate estimates of and V2 ). what conclusion can be drawn about the US model versus VAR5/2? For the first three variables the models are generally quite close. from which the mean ¯ dik was computed. The US model is worse than VAR4 for the first three quarters. Comparing VAR5/2 and VAR4. For these models the number of repetitions per stochastic simulation was 1000 even for the 68 stochastic simulations involved in getting the disk values. tied for quarter four. Turn now to the d row values in Table 8. and better for the remaining four quarters. but only slight. The US model may thus have a slight edge over VAR5/2. Using VAR5/2 as the better of the two VAR models. For the two quarter ahead prediction.7.5. and the number of repetitions per each of the 68 stochastic simulations was 250. The US model is better than VAR4 for all eight quarters. For the fourth variable. and one might call it a tie. The d row values can thus be compared across models. these calculations allowed 68 values of disk to be computed for each endogenous variable i.

8.6 PREDICTIVE ACCURACY

235

for the 91 exogenous variables. As discussed earlier, these equations may exaggerate the uncertainty of the exogenous variables and thus bias the results against the US model. Turning next to the AC model, it does very well in the GDP R predictions. It has the smallest d row values in the table. There clearly seems to be predictive power in the lagged components of GDP R that is not captured in the US and VAR models. Comparing the a and b rows in Table 8.5 shows that coefficient uncertainty contributes much less to the variances than does the uncertainty from the structural error terms. In other words, the a row values are large relative to the difference between the b row and a row values. For the US model the differences between the c row values and the b row values are generally larger than the differences between the b row and a row values, which says that exogenous variable uncertainty (as estimated by the autoregressive equations) generally contributes more to the total variance than does coefficient uncertainty. The differences between the d row and c row values are measures of the misspecification of the model not already captured in the c row values. On this score, the worst specifications for the models are for the bill rate and the best are for the unemployment rate. Again, the differences between the US model and VAR5/2 regarding misspecification are close except for the bill rate, where the US model is much better. Outside Sample RMSEs From the 68 stochastic simulations that are used for the disk calculations, one has for each endogenous variable i, 68 one quarter ahead outside sample error terms, 67 two quarter ahead outside sample error terms, and so on. (These errors are denoted ˆisk in Section 7.7.) From these errors one can compute RMSEs, and the results of doing this for four variables are presented in Table 8.6. Remember, however, that comparing RMSEs across models has problems that do not exist when comparing the d row values in Table 8.5 across models. Exogenous variable uncertainty is not accounted for, which affects the comparisons between the US model and the others but not between the other models themselves. Also, the fact that forecast error variances change over time is not accounted for in the RMSE calculations. The RMSEs in Table 8.6 are, however, all outside sample, which is a least a crude way of accounting for misspecification effects. For what they are worth, the results in Table 8.6 show that the US model is noticeable better than the VAR models for real GDP and the bill rate. The

236

8 TESTING THE US MODEL
Table 8.6 RMSEs of Outside Sample Forecasts for Four Models for 1976:3–1993:2 1 US VAR5/2 VAR4 AC US VAR5/2 VAR4 US VAR5/2 VAR4 US VAR5/2 VAR4 .79 1.07 1.15 .79 Number of Quarters Ahead 2 3 4 6 GDP R: 1.39 1.89 2.05 1.23 Real GDP 1.95 2.33 2.51 3.06 2.61 3.22 1.64 1.95 2.64 3.84 4.33 2.48 2.32 1.98 2.01 8 2.74 4.57 5.15 2.99 3.21 2.89 3.02 1.61 1.68 1.84 3.61 5.10 5.56

GDP D: GDP Deflator .34 .58 .82 1.23 .31 .61 .87 1.18 .33 .62 .88 1.18 U R: .31 .32 .36 .80 1.18 1.17

Unemployment Rate .61 .89 1.16 1.51 .65 .94 1.20 1.52 .74 1.04 1.29 1.68 RS: Bill Rate 1.61 1.91 2.29 2.08 2.52 3.07 2.15 2.56 3.12 3.03 4.13 4.53

1. The results are based on 68 sets of coefficient estimates of each model. 2. Each prediction period began one quarter after the end of the estimation period. 3. For U R and RS the erors are in percentage points. For GDP R and GDP D the errors are expressed as a percent of the forecast mean (in percentage points).

results are fairly close for the GDP deflator and the unemployment rate. The AC model is about the same as the US model and noticeably better than the VAR models. Therefore, as expected, the US model does better relative to the other models when exogenous variable uncertainty is not taken into account. This completes the comparison of the models using the d row values. The next two sections compare the models in two other ways, and the final section summarizes the overall comparison results.

8.7 COMPARING INFORMATION IN FORECASTS

237

8.7

Comparing Information in Forecasts9

Section 7.8 discussed a method for comparing the information in various forecasts, and this section uses this method to compare the forecasts from the US, US+, VAR5/2, VAR4, and AC models. The results of comparing the US and US+ models to the other three are presented in Table 8.7, and the results of comparing the AC model to the two VAR models are presented in Table 8.8. The rest of this section is a discussion of these two tables. When using the method in Section 7.8, the forecasts should be based on information only up to the beginning of the forecast period. In other words, they should be “quasi ex ante” forecasts. The 68 sets of estimates that were used for the results in the previous section are used here to generate the forecasts. As was the case in the previous section, each forecast period begins one quarter after the end of the estimation period. There are 68 one quarter ahead forecasts, 67 two quarter ahead forecasts, and so on. All these forecasts are outside sample, and so they meet one of the requirements of a quasi ex ante forecast. The other main requirement of a quasi ex ante forecast is that it not be based on exogenous variable values that are unknown at the time of the forecast. The VAR and AC forecasts meet this requirement because the models have no exogenous variables, but the forecasts from the US model do not. The 68 sets of forecasts that were computed for the US model are based on the actual values of the exogenous variables.10 The US+ model, on the other hand, has no hard to forecast exogenous variables, and so it meets the exogenous variable requirement. Both the US and US+ models were used for the present results to see how sensitive the results for the US model are to the treatment of exogenous variables. For this work the US+ model was also estimated 68 times, including estimation of the 91 exogenous variable equations, and these
material in this section is an updated version of the material in Fair and Shiller (1990) (FS). In FS the US model was compared to six VAR models, eight AC models, and two autoregressive models, whereas for present purposes only two VAR and one AC model are used. In addition, the version of the US model that was used in FS was the version that existed in 1976, whereas the current version of the model is used here. Finally, only the results for real output were discussed in FS, whereas results for the GDP deflator, the unemployment rate, and the bill rate are also discussed here. The forecasts examined in this section are all quasi ex ante. The information content of actual ex ante forecasts for a number of models is examined in Fair and Shiller (1989) using the present method, but this material is not presented here. 10 Remember that the actual values of the exogenous variables were used in computing the disk values in the previous section. Exogenous variable uncertainty was handled through the c row calculations.
9 The

238

8 TESTING THE US MODEL
Table 8.7 US Model Versus Three Others: Estimates of Equation 7.12

Other Model

One Quarter Ahead Forecast US Other cnst β γ SE

Four Quarter Ahead Forecast US Other cnst β γ SE

VAR5/2 VAR4 AC

-.0008 (0.45) -.0008 (0.50) -.0020 (1.11) -.0002 (0.10) -.0000 (0.02) -.0020 (0.90)

.781 (5.30) .756 (5.35) .620 (3.48) .678 (3.90) .613 (3.02) .289 (1.51)

VAR5/2 VAR4 AC

GDP R: Real GDP US Model -.051 .00691 -.0025 (0.34) (0.41) -.003 .00692 -.0021 (0.03) (0.36) .324 .00681 -.0101 (1.45) (1.56) US+ Model .006 .00825 .0069 (0.04) (0.53) .064 .00823 .0053 (0.52) (0.43) .758 .00770 -.0116 (4.14) (1.45) GDP D: GDP Deflator US Model .416 .00260 .0079 (2.95) (1.36) .387 .00264 .0082 (2.67) (1.54) US+ Model .454 .00284 .0073 (2.41) (1.10) .428 .00282 .0071 (2.82) (1.14) U R: Unemployment Rate US Model .398 .00278 .0385 (2.84) (3.96) .230 .00288 .0409 (1.89) (4.89) US+ Model .392 .00279 .0373 (2.67) ( 3.06) .225 .00288 .0399 (1.73) (3.67) RS: Bill Rate US Model -.027 .795 (0.20) -.054 .795 (0.37) US+ Model -.027 .822 (0.19) -.047 .821 (0.31)

.753 (4.87) .767 (4.86) .505 (3.74) .381 (1.01) .417 (1.08) .335 (2.13)

-.103 (0.72) -.112 (0.84) .578 (2.30) .153 (0.60) .124 (0.52) .911 (3.37)

.01727 .01722 .01629

.02121 .02123 .01866

VAR5/2 VAR4

.0023 (3.22) .0027 (3.71) .0024 (3.08) .0027 (3.53)

.454 (3.49) .461 (3.49) .394 (2.26) .407 (2.96)

.519 (2.54) .489 (2.39) .261 (1.02) .307 (1.43)

.341 (1.59) .377 (1.89) .582 (2.41) .556 (2.94)

.01000 .00981

VAR5/2 VAR4

.01050 .01021

VAR5/2 VAR4

.0018 (0.97) .0030 (1.68) .0011 (0.54) .0021 (1.10)

.579 (4.25) .730 (6.23) .595 (4.14) .748 (5.83)

.689 (2.82) .761 (3.20) .556 (2.30) .625 (2.49)

-.200 (0.85) -.305 (1.48) -.071 (0.28) -.176 (0.75)

.00909 .00892

VAR5/2 VAR4

.00996 .00990

VAR5/2 VAR4

-.31 (0.88) -.32 (0.94) -.35 (0.92) -.36 (0.99)

1.069 (6.55) 1.097 (6.82) 1.073 (6.20) 1.093 (6.54)

1.69 (0.92) 1.63 (0.86) 2.28 (1.34) 2.22 (1.27)

.588 (1.74) .662 (1.83) .501 (1.55) .575 (1.64)

.184 (1.21) .121 (0.75) .186 (0.95) .123 (0.60)

2.180 2.209

VAR5/2 VAR4

2.223 2.247

8.7 COMPARING INFORMATION IN FORECASTS
Table 8.8 AC Versus VAR5/2 and VAR4 Other Model One Quarter Ahead Forecast AC Other cnst β γ SE GDP R: Real GDP VAR5/2 VAR4 -.0010 (0.54) -.0010 (0.54) .916 (5.28) .881 (4.81) .106 (0.81) .120 (1.18) .00778 .00774 -.0038 (0.46) -.0048 (0.60) .938 (3.78) .954 (3.79) .204 (2.28) .181 (2.22) Four Quarter Ahead Forecast AC Other cnst β γ SE

239

.01863 .01873

68 sets of estimates were used. All the forecasts for the US+ model were also outside sample. Again, remember from the discussion in Section 7.8 that the treatment of the exogenous variables as in US+ may bias the results against the model. Many of the exogenous variables may not be as uncertain as the autoregressive equations imply. Both one quarter ahead and four quarter ahead forecasts are examined in Table 8.7. In the estimation of the equations, the standard errors of the coefficient estimates were adjusted in the manner discussed in Section 7.8 to account for heteroskedasticity and (for the four quarter ahead results) a third order moving average process for the error term. Equation 7.12 was used for real GDP and the GDP deflator, where both variables are in logs, and the level version of equation 7.12 was used for the unemployment rate and the bill rate. Turn now to the results in Table 8.7, and consider the forecasts of real GDP first. Also, ignore for now the results for the AC model. The results show that both US and US+ dominate the VAR models for real GDP. The estimates of the coefficients of the VAR forecasts are never significant, and the estimates of the coefficients of the US and US+ forecasts are significant except for the four quarter ahead forecasts for US+, where the t-statistics are about one. It is thus interesting to note that even though the standard errors of the forecasts in Table 8.5 (the d row values) are fairly close for real GDP for the US and VAR models, the results in Table 8.7 suggest that the VAR forecasts contain no information not already in the US forecasts. In this sense the method used in this section seems better able to discriminate among models. The results for the GDP deflator show that both the US (and US+) forecasts and the VAR forecasts contain independent information. In most cases both coefficients are significant, the exceptions being US versus the VAR models for the four quarter ahead forecasts, where the VAR forecasts are not quite significant, and US+ versus the VAR models for the four quarter ahead forecasts, where the US+ coefficients are not quite significant.

240

8 TESTING THE US MODEL

For the unemployment rate US and US+ dominate the VAR models with the exception of the one quarter ahead forecasts from VAR5/2, which are significant in the US and US+ comparisons, although with t-values smaller than those for the US and US+ forecasts. The results for the bill rate show that US and US+ dominate the VAR models for the one quarter ahead forecasts. For the four quarter ahead forecasts the US and US+ forecasts have larger coefficient estimates and larger t-values than do the VAR forecasts, although collinearity is such that none of the tvalues are greater than two. The results of these comparisons are thus encouraging for the US model. Only for the GDP deflator is there much evidence that even the US+ forecasts lack information that is contained in the VAR forecasts. Consider now the AC model, where there are only results for real GDP. The US and US+ comparisons in Table 8.7 suggest that both the US or US+ forecasts and the AC forecasts contain independent information. There clearly seems to be forecasting information in the lagged components of GDP that is not captured in the US model, and this is an interesting area for future research. The VAR versus AC comparisons in Table 8.8 show that the VAR forecasts appear to contain no independent information for the one quarter ahead forecasts, but at least some slight independent information for the four quarter ahead forecasts. As did the results in the previous section, these results suggest that the AC model may be a better alternative than VAR models for many purposes.11
a few exceptions, the results for real GDP here are similar to those in Fair and Shiller (1990) (FS). The US+ version is closest to the version used in FS, and so the following discussion focuses on the US+ results. The one quarter ahead results for US+ in Table 8.7 have the US model dominating the VAR models, which is also true in Table 2 in FS. For the four quarter ahead results neither the US+ nor the VAR forecasts are significant in Table 8.7 and both are significant in Table 2 in FS. However, in both tables the US forecasts have larger coefficient estimates and larger t-values than do the VAR forecasts. Regarding US+ versus AC, the results in Table 8.7 are more favorable for AC than they are in Table 2 in FS. In Table 2 in FS the US model dominates the AC models, whereas in Table 8.7 the AC model has a large and significant coefficient estimate for both the one quarter ahead and four quarter ahead forecasts for US+ versus AC. Finally, the VAR versus AC comparisons in Table 8.8 are similar to those in Table 3 in FS. In both tables the AC forecasts dominate the VAR forecasts for the one quarter ahead results and both forecasts are significant for the four quarter ahead results.
11 With

8.8 ESTIMATING EVENT PROBABILITIES

241

8.8

Estimating Event Probabilities12

The use of event probability estimates to compare models was discussed in Section 7.9. This comparison is made in this section for two events and five models. The five models are the US, US+, VAR5/2, VAR4, and AC models. The two events, labelled A and B are: A = At least two consecutive quarters out of five of negative real GDP growth. B = At least two quarters out of five of negative real GDP growth. Event A is a recession as generally defined. Event B allows the two or more quarters of negative growth not to be consecutive. The first 64 sets of estimates of each model that were used for the results in the previous section were used here. (Only 64 rather than 68 sets of estimates could be used because each forecast here has to be five quarters ahead.) There were 64 five quarter ahead outside sample stochastic simulations performed. The number of repetitions per five quarter forecast was 250 for US and US+ and 1000 for the VAR5/2, VAR4, and AC. Regarding the US+ model, this is the first time that stochastic simulation of the model is needed. For the results in the previous section only deterministic outside sample forecasts were used. As discussed in Section 8.2, when stochastic simulation was performed using US+, the covariance matrix of all the error terms, which is 121×121, was taken to be block diagonal. For the results in this section this matrix was estimated 64 times, each estimate being used for each of the 64 stochastic simulations. The covariance matrices of the coefficient estimates are not needed for the work in this section because coefficients are not drawn. From the stochastic simulation work one has five sets of values of Pt (t = 1, · · · , 64) for each of the two events, one for each model, where Pt is the model’s estimate of the probability of the event for the period beginning in quarter t. One also has values of Rt for each event, where Rt is the actual outcome—one if the event occurred and zero otherwise. Given the values
material in this section is an updated and expanded version of the material in Section 3.3 in Fair (1993c). In Fair (1993c) only within sample forecasts were used and the only comparisons were to the constant model and a fourth order autoregressive model. In this section all the forecasts are outside sample and comparisons are made to two VAR models and an AC model in addition to the constant model. Also, no coefficients are drawn for the present results, whereas they were drawn in the earlier work. (See the discussion in Section 7.9 as to why coefficients were not drawn here.)
12 The

242

8 TESTING THE US MODEL

Table 8.9 Estimates of Probability Accuracy Event A (Actual p = .188) ¯ Model p ¯ QPS LPS Constant US US+ VAR5/2 VAR4 AC .188 .175 .173 .310 .264 .154 .305 .310 .310 .496 .518 .324 .483 .477 .472 .844 .972 .510

Event B (Actual p = .234) ¯ Model p ¯ QPS LPS Constant US US+ VAR5/2 VAR4 AC
∗ LPS

.234 .211 .238 .416 .358 .237

.359 .290 .306 .514 .521 .363

.545 .438 .465 * * .537

not computable.

of Rt , another model can be considered, which is the model in which Pt ¯ ¯ is taken to be equal to R for each t, where R is the mean of Rt over the 64 observations. This is simply a model in which the estimated probability of the event is constant and equal to the frequency that the event happened historically. This model will be called “Constant.” The results for this model are not outside sample because the mean that is used is the mean over the whole sample period. The summary statistics are presented in Table 8.9. In two cases (both for the VAR models) the LP S measure could not be computed because either Pt was 1 and Rt was 0 or vice versa. This is a limitation of the LP S measure in that in cannot handle extreme errors of this type. It, in effect, gives an infinite loss to this type of error. The results in Table 8.9 are easy to summarize. Either US or US+ is best for both events for both error measures except the case of the constant model and event A, where the QPS for the constant model is slightly smaller. This is

8.9 SUMMARY OF THE TEST RESULTS

243

thus strong support for the US model. The results in Table 8.9 also show that the AC model completely dominates the VAR models. This is in keeping with the results in the previous two sections, which generally show the AC model out performing the VAR models. Figures 8.1 and 8.2 plot the values of Pt and Rt for the US+ and VAR5/2 models for event A for the 64 observations. It is clear from the plots why US+ has better QPS and LPS values in Table 8.9. VAR5/2 has high probabilities too early in the late 1970s and comes down too fast after the recession started compared to US+. Note that both models do not do well predicting the 1990– 1991 recession. No model seems to do well predicting this recession.

8.9

Summary of the Test Results

Overall, the results in Tables 8.5, 8.6, 8.7, and 8.9 are favorable for the US model. Even after correcting for exogenous variable uncertainty that may be biased against the model, the model does well in the tests relative to the VAR and AC models. The GDP deflator results are the weakest for the US model, and this is an area for future work. Also, the results in Table 8.7 show that there is information in the AC forecasts of real GDP not in the US forecasts, which suggests that the US model is not using all the information in the lagged components of GDP. Aside from the GDP deflator forecasts, there does not appear to be much information in the VAR forecasts not in the US forecasts. The AC model generally does as well as or better than the VAR models. This suggests that there is useful information in the lagged components of GDP that the VAR models are not using. From another perspective, if one wants a simple, non structural model to use for forecasting GDP, an AC model would seem to be a better choice than a VAR model.

244

81:1-82:3
89:4-90:4

1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0
Figure 8.1
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
Quarter
8 TESTING THE US MODEL

Probability

Estimated Probabilities for Event A for

81:1-82:3
89:4-90:4

1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0
Figure 8.2
245

8.9 SUMMARY OF THE TEST RESULTS

Probability

77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
Quarter

Estimated Probabilities for Event A fo

9

Testing the MC Model
9.1 Introduction

This chapter is concerned with testing the overall MC model. It is the counterpart of Chapter 8 for the US model. There are, however, many fewer tests in this chapter than there are in Chapter 8. The main problem here is the short length of many of the sample periods. Many of these periods do not begin until the 1970s, which effectively rules out, for example, the successive reestimation that was done in Sections 8.6, 8.7, and 8.8 for the US model. Therefore, the main testing of the MC model must rely on within sample predictions, although some outside sample results are reported in this chapter. The size of the MC model is discussed in Section 9.2, and the solution of the model is explained. Section 9.3 then presents the model to which the MC model is to be compared. This model, called ARMC, replaces each of the stochastic equations of the ROW model with a fourth order autoregressive equation for the quarterly countries and a second order autoregressive equation for the annual countries. The within sample test results are discussed in Section 9.4, and the outside sample test results are discussed in Section 9.5.

9.2

The Size and Solution of the MC model

The US model, which is part of the MC model, includes 30 stochastic equations plus one more when it is imbedded in the MC model. This additional equation is discussed below. There are 32 countries in the ROW model and up to 15 stochastic equations per country. If each country had all 15 equations, there would be a total of 480 (32 × 15) stochastic equations in the ROW 247

17) and −. In the US case log P XU S − log P W $U S is regressed on log GDP D − log P W $U S . however. The equation. and the US variables (but not the trade shares). which. This identity replaces identity 32 in Table A. The way in which the US model is imbedded in the MC model is explained in Table B. which depends on P MPU S . Feeding into Table B. P I M. EX.4. In addition.4 from the US model are P XU S and M85$AU S .26.55). P I M depends on P MU S . Counting these variables. The standard error is .0125. Because of data limitations. which depends on I M. which is numbered 132 is: log P XU S − log P W $U S = λ(log GDP D − log P W $U S ) (132) This equation is estimated under the assumption of a second order autoregressive error for the 1962:1–1992:3 period. as just noted. The estimate of λ is . and there are in fact 315 stochastic equations in the ROW model. Table B. there are about 4000 variables in the MC model. There are a total of 1541 unrestricted coefficients in these equations. The estimates (t-statistics) of the two autoregressive coefficients are 1. 4 variables determined when the countries are linked together. various transformations of the variables that are needed for the estimation. EX depends on X85$U S .956 with a t-statistic of 24. Not counting the trade share coefficient estimates. all the coefficient estimates for the US model are presented in Tables 5. which is also determined in Table B.4. feeds into Table B. which is determined in Table B.5: P EX = DEL3 · P XU S . This is shown in Table B. counting the autoregressive coefficients of the error terms.5. P XU S is determined is the same way that P X is determined for the other countries. there are 1299 estimated trade share equations. namely by equation 11.1a–6. not all countries have all equations. Given the predicted value of P XU S from equation 132. These are the estimates that were used for the within sample results below. I M is determined by equa- . and all the coefficient estimates for the ROW model are presented in Tables 6. and the price of imports. there are thus 346 stochastic equations in the MC model. and 22 exogenous variables per country. depends on MU S .30. as discussed in Section 6.15a.1 shows that there are in the ROW model 19 variables determined by identities.248 9 TESTING THE MC MODEL model.16. The two key variables that are exogenous in the US model but become endogenous in the overall MC model are exports. M85$AU S .5.3 in the US model.1–5. Given the 31 stochastic equations in the US model. respectively.50 (19.4.51 (−6. P EX is determined by the identity listed in Table B.

assume that year 1 is the first year to be solved. The solution for the quarterly countries for the four quarters of year 1 is a dynamic simulation in the sense that the predicted values of the endogenous variables from previous quarters are used.) The fourth task is to compute X85$. and P W $ for each country using equations L-2.9. all the stochastic equations and identities are solved. Because some of the countries are annual. . and then quarter 4 k1 times. the overall MC model is solved a year at a time. Finally. and M85$A from the annual countries have to be converted to quarterly values first.13 in Section 6. Given from the solution in step 1 values of E.4. Equation 27 is thus the key equation that determines the U. and M85$A for each country.4.2 THE SIZE AND SOLUTION OF THE MC MODEL 249 tion 27 in the US model. Since all the calculations in Table B. then quarter 2 k1 times. This is done in the manner discussed at the bottom of Table B.4 are quarterly. P MP . and P W $ for all four quarters of year 1 for each quarterly country and for year 1 for each annual country. the third task is to compute the values of αij from the trade share equations—see equation 6. when relevant. Monetary policy is not exogenous because of the use of the interest rate and exchange rate reaction functions. L-3. For the quarterly countries “solved” means that quarter 1 of year 1 is passed through k1 times. In other words. The second task is to compute P X$ using equation L-1. P X. P X.4 can be performed. then quarter 3 k1 times.16.S. A solution period must begin in the first quarter of the year. for the annual countries the quarterly values of these three variables are then converted to annual values by summing in the case of X85$ and averaging in the case of P MP and P W $. import value that feeds into Table B. In the following discussion. For the annual countries “solved” means that the equations are passed through k1 times for year 1. Given the values of P X$. The overall MC model is solved as follows: 1. Given values of X85$. The main exogenous variables in the overall MC model are the government spending variables (G). The procedure in effect takes the distribution of the annual values into the quarterly values to be exogenous. fiscal policy is exogenous. the calculations in Table B. 2. This solution is also dynamic in the sense that the predicted value of αij for the previous quarter feeds into the solution for the current quarter. P MP . and L-4. where k1 is determined by experimentation (as discussed below). the annual values of E. (Remember that the lagged value of αij is an explanatory variable in the trade share equations. in the solution for the current quarter.

the covariance matrix of the error terms is 346×346 and the covariance matrix of the coefficient estimates is 1541×1541. when relevant. and 3 for year 2. there are not enough observations to estimate the covariance matrix of the error terms unconstrained. however. some of the equations are estimated using quarterly data and some using annual data. The computer time needed to solve the MC model once is still large enough to make stochastic simulation costly in time. stochastic simulation would not be routine using. Some of the problems that arise in dealing with these matrices are the following. At the end of this. say. Third. Continue then to year 3. 2. to do so.250 9 TESTING THE MC MODEL 3. First. be eased considerably with the next generation of chips. Stochastic Simulation of the MC Model in the Future Although no stochastic simulation experiments using the MC model were performed for the present work. The best way to handle these problems is probably to take the covariance matrices to be block diagonal. and so in a few years it should be possible to perform the same type of calculations for the MC model as were performed in Chapters 8 and 11 for the US model. In some cases one may also want to take the covariance matrix of the coefficient estimates within a country to be block diagonal. Second. and these are the values of k1 and k2 that have been used for the results in this chapter and in Chapter 12. Repeat steps 1. This computer restriction should. P MP . Even using the block diagonal matrices just discussed. even if the periodicity of the data were the same. I have found that going beyond k1 = 4 and k2 = 7 leads to very little change in the final solution values. 4. the estimation periods generally differ across countries. it should be possible in future work. one block per equation. . and so on. Keep repeating steps 1 and 2 until they have been done k2 times. with a few adjustments. If the solution is meant to be dynamic. one block per country. 486 computers. repeat step 1 and then step 2. declare that the solution for year 1 has been obtained. and P W $ from step 2. in the solution for year 2. Given the new values of X85$. Since the MC model has 346 stochastic equations and 1541 unrestricted coefficients. use the predicted values for year 1 for the annual countries and the predicted values for the four quarters of year 1 for the quarterly countries.

with log P X rather than log P X − log(P W $ · E). It differs from the AC model in that 1) regressions are not performed for components that are not determined by stochastic equations in the MC model. and 5) equation 12. A model that has some similarities to the AC model was used as the basis of comparison for the MC model. For the annual countries only the first two lagged values are used. 2 With the exception noted in the previous footnote. where the left hand side variable is V 1 rather than Y . The US model is the same for both models. one needs a basis of comparison.3 THE ARMC MODEL 251 9. This model will be called the “ARMC” model. 1 With . Each equation of the ARMC model was estimated over the same sample period as was used for the corresponding equation for the MC model. because V 1 is used for the left hand side variable for equation 4. the identity I-4 is changed from V 1 = Y − X to Y = X + V 1. with RB rather than RB −RS−2 . and the first four lagged values of the variable.9. Two VAR models and an AC model were used for this purpose for the US model. The model is solved in the same way as the MC model. not just the GDP identity. 3) equation 9. with log E rather than log(E/E−1 ). the left hand side variable for the ARMC model for each equation is the same as that for the MC model.2 and the trade share calculations are the same. 2) equation 8.3 The ARMC Model In order to decide how good or bad the MC model predicts the data. Each of the stochastic equations for the quarterly countries (except the US) is replaced with an autoregressive equation in which the left hand side variable is regressed on a constant term. and 3) all the identities are used. The exceptions are 1) equation 4. 4) equation 11.1 The MC and ARMC models differ only in this treatment of the stochastic equations. the following five exceptions. 2) regressions are performed for all the variables determined by stochastic equations. The ARMC is like the AC model in that the components of GDP are regressed on their lagged values and the GDP identity is used. not just the components. Also. with log W rather than the left hand side variable used to account for the coefficient restriction. It is simple to describe. a linear time trend. all the identities are the same.

4 Remember that each prediction period begins in the first quarter of the year for the quarterly countries. root mean squared errors (RMSEs) were computed. rather than by taking weighted averages of the individual ratios. where a new prediction period began each quarter. the following countries were dropped in the solution of the model for the following periods: NE: 1972–1976. and for the annual countries ratios are presented for the one year ahead and two year ahead RMSEs. and MA: 1987–1990. On average 3 In order to use the complete 1972–1990 period. The second row.252 9 TESTING THE MC MODEL 9. For the quarterly countries there were 19 one through four quarter ahead forecasts and 18 five through eight quarter ahead forecasts. and so on through 1990.S. . This is contrary to the case for the US model in Chapter 8. The first prediction began in 1972. Consider first the results for the exchange rate in Table 9. and eight quarter ahead RMSEs. The following discussion will concentrate on the weighted averages. and RMSEs were computed. the longest period3 over which the MC model could be solved was 1972–1990. The weight used for a country is the ratio of its GDP in 1985 in U. four quarter ahead. labelled II. The same forecasts were made using the ARMC model. A ratio less than one means that the MC model is more accurate. is the weighted average of the eight quarter ahead results for the quarterly countries and the two year ahead results for the annual countries. NO: 1972–1973. Data on some of the variables did not exist for these countries for the respective periods. A series of two year (eight quarter) ahead predictions were run over the period. the second one in 1973. The results from this work are presented in Table 9. dollars to the total for all the countries.4 Given the forecast values.1.1 for each variable is a weighted average of all the results. Presented at the top of Table 9. labelled I.1. FI: 1972–1976. and a ratio greater than one means that the ARMC model is more accurate. 5 GDP in U. Presented in the table for each of 17 variables for each country is the ratio of the MC RMSE to the ARMC RMSE. For each endogenous variable this results in 19 one year ahead forecasts and 18 two year ahead forecasts for the annual countries. These summary results were obtained by taking weighted averages of the individual RMSEs and then computing the ratio of the weighted averages.5 The first row. is the weighted average of the four quarter ahead results for the quarterly countries and the one year ahead results for the annual countries. For the quarterly countries ratios are presented for the one quarter ahead.4 Within Sample RMSEs Given the data availability. dollars is (P Y · Y )/E.S.

9. For the one year ahead results the weighted average ratio is 1. much of the increase for real GDP is due to the effect of errors made in predicting exports. . and for the two year ahead results it is 1. There is very little change for the unemployment rate and the bill rate.2.05. and the results in Table 9. the outside sample exchange rate results favor MC over ARMC. as will be seen in the next section. consumption (C). It is well known that structural exchange rate equations have a hard time beating autoregressive equations. and the unemployment rate (U R)—the MC model is more accurate than the ARMC model. the balance of payments (S).16 percent for the eight quarter ahead forecast. The RMSEs in the rows labelled “MC” are from the same MC solutions used for the results in Table 9.1 are examples of this.2 show how the accuracy of the US model changes when it is imbedded in the MC model.11 percent for the eight quarter ahead forecast. On average the result seem reasonably good for the MC model. Similarly. For real GDP. investment (I ). For the remaining variables—the price deflator (P Y ).4 WITHIN SAMPLE RMSES 253 the MC model does not do quite as well as the ARMC model.69 percent for the one quarter ahead forecast to 8. On average. The model appears to have explanatory power beyond that contained in the lagged values and the time trend.1.29 percent for the one quarter ahead forecast to 6. For the GDP deflator the increases (after the first quarter) are between 18 and 66 percent. The RMSE for the price of imports ranges from 2. the interest rate (RS). Other variables for which the ARMC model is more accurate than the MC model are the price of exports (P X) and the price of imports (P M). and the wage rate (W ). The results in Table 9. The results for the United States are presented in Table 9.17. Much of the increase for the GDP deflator is due to the effect of errors made in predicting the price of imports. The RMSEs in the rows labelled “US” are from the solutions for the US model alone. Also. For the US model alone exports (EX) and the price of imports (P I M) are exogenous. the results are only slightly worse for the structural model. The same 19 prediction periods were used for the US model alone as were used for the MC model. The results are mixed for GDP (Y ). however. imports (M). The RMSE for exports ranges from 2. the RMSE increases between 7 and 27 percent when the US model is imbedded in the MC model. exports (X85$).

10 1.02 1.80 .02 .95 .10 1.83 1.42 1.18 .19 1.95 .22 .01 .97 1.80 .01 1.82 .74 1.08 1.86 1.95 .78 .07 1.67 .81 .20 .05 1.02 1.62 – – X85$ .96 .89 .13 1.72 .33 1.95 1.97 .41 1.82 .34 1.83 .86 .85 .01 1.06 1.54 1.00 1.94 1.23 .38 1.33 .02 1.86 .00 1.49 .08 .67 .62 .13 .84 – – – .95 .70 1.44 1.93 .06 1.76 .88 .97 .03 1.97 .07 .00 1.70 .07 .98 .48 1.22 – – – – – – 1.91 .91 .26 1.11 1.83 .04 1.05 .83 1.85 1.01 1.11 1.74 .97 .67 .91 1.12 1.10 1.98 1.05 – – – 1.85 .03 .62 .62 .08 .89 .94 1.90 1.06 1.91 .10 .66 .93 1.28 .09 1.99 1.82 1.37 .95 1.82 .67 – – – – – – 1.90 .76 .05 .81 .03 1.06 1.79 .78 .06 .31 1.06 .14 .52 1.08 .11 1.31 1.11 1.57 .75 .75 .05 .90 .83 1.09 1.93 1.84 .19 1.41 1.98 1.95 1.84 .96 .99 .95 1.05 – – .91 1.08 1.94 .08 1.97 .76 .73 .24 1.10 1.10 .14 .92 1.94 1.01 .82 1.75 .83 .95 1.86 1.02 .85 .88 .90 1.03 .01 .86 .96 1.24 .68 .08 1.96 1.94 .11 1.00 1.90 .13 .41 .49 .20 .82 .76 .96 .11 – – 1.96 .90 1.57 .85 .36 1.94 .83 1.78 1.04 .22 1.33 1.75 1.05 1.74 .74 .96 .68 .16 – – .71 .69 .25 .88 .06 1.12 1.11 .85 .77 .19 1.62 .00 1.91 1.72 .93 .48 1.89 .78 .99 1.08 1.54 1.09 1.61 .86 .99 .87 .18 .06 1.97 .00 .95 – – 1.88 .82 .16 1.95 1.75 1.03 .71 1.88 .19 1.20 .88 1.89 .42 .63 1.42 .05 .06 1.89 1.93 1.64 .14 1.07 1.80 .02 1.88 .82 1.54 .27 .86 1.02 1.07 .18 .95 .84 .65 .83 1.02 .84 1.80 .93 .03 .37 .93 .18 .05 1.16 1.98 – – – .55 1.42 .92 .92 .92 .04 .65 .61 .03 1.76 .16 .11 1.64 .81 .02 1.87 1.02 1.82 1.91 1.18 .97 .96 1.83 .25 .94 .99 .80 .90 .94 .87 .00 .87 1.90 .12 1.13 .93 .81 .91 – – .27 .98 .97 .17 1.93 1.92 1.63 .69 .66 .68 .68 .98 .97 1.05 1.71 .58 .84 .13 .03 1.08 .94 .87 .02 .95 .96 1.35 .80 1.99 .07 1.75 .96 1.88 .26 1.95 – – – .21 1.84 1.91 .15 1.00 1.11 .04 .12 1.57 .65 .89 1.02 .98 1.90 .77 .59 .98 .23 1.78 .71 1.79 1.78 1.90 .95 .20 .76 .79 .97 .00 1.87 .95 1.69 .12 – – – – – – .08 1.89 .19 2.87 1.93 .69 .77 .75 .93 .41 1.87 1.84 .92 .96 .98 .81 1.90 1.80 .10 .79 1.04 1.48 1.254 9 TESTING THE MC MODEL Table 9.85 .92 .93 .11 1.20 1.59 .83 .99 1.87 1.02 .01 .82 1.87 1.09 1.87 .37 1.23 .81 1.17 .01 .08 1.18 1.1 Ratios of Within Sample RMSEs 1972-1990: AC/ARMC Y PY RS E M C I V 1/Y I II 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 All All CA CA CA JA JA JA AU AU AU FR FR FR GE GE GE IT IT IT NE NE NE ST ST ST UK UK UK FI FI FI AS AS AS SO SO SO KO KO KO BE BE DE DE NO NO SW SW GR GR IR IR PO PO SP SP NZ NZ .95 1.02 1.11 .76 .41 .10 .66 1.68 1.85 .95 1.84 1.68 .86 1.66 .12 .96 .82 .47 .08 .87 .66 1.98 1.01 .99 1.06 1.30 1.88 .01 .96 1.11 .74 .80 .57 1.74 1.97 .96 .08 1.97 .04 .73 .69 1.65 .84 .89 .08 .21 1.90 1.16 .03 1.97 .

54 .77 .13 1.91 .81 .83 .04 1.06 .11 .65 .07 1.58 .00 1.07 .58 .92 .78 .95 1.92 1.50 .34 1.04 1.34 1.97 .81 .95 1.04 1.35 1.24 1.02 .10 .69 .97 1.78 .1 (continued) S W J .97 1.88 .9.10 1.69 1.07 1.74 1.03 – – .03 .89 .89 – – – – – – .04 1.95 .62 1.88 1.23 1.12 1.63 1.88 .06 .99 1.45 .97 .61 .14 .17 1.91 1.81 1.96 1.95 2.32 1.80 .23 1.84 .84 .73 .94 .05 – – – 1.46 .18 1.86 .00 .00 1.91 .03 1.06 1.88 .46 1.84 .38 1.46 – – – – – – .73 .85 .68 .75 .91 1.04 – – – – – – 1.93 1.10 .78 .04 1.23 .57 .13 .89 .87 .98 .80 1.85 .15 1.03 1.58 – – L1 1.05 1.13 1.67 .01 – – 1.90 .83 .03 1.09 1.97 – – – 1.08 1.14 1.04 .40 .98 .17 .15 1.03 1.04 1.13 .18 1.27 1.15 1.08 .04 1.24 – – – 1.15 1.27 1.03 1.95 .19 1.98 1.01 1.56 1.23 .13 1.06 .03 .70 1.01 1.50 .11 1.19 1.96 1.06 1.88 1.49 1.85 .82 .78 .84 .91 1.69 .89 .02 1.28 1.24 .65 .05 – – – 1.91 .98 1.87 .03 1.55 .76 1.94 .20 Table 9.14 .69 1.95 1.81 .01 1.89 .95 .57 1.09 .79 .19 1.00 PM 1.51 1.66 .00 .53 .45 1.72 .15 – – 1.25 1.90 .94 – – – 1.82 1.57 .98 .74 .67 – – – 1.92 .81 .76 1.97 .06 1.16 .17 1.65 .96 1.36 .86 – – – – – – .75 .87 1.70 .01 1.96 1.04 .02 1.00 1.76 .38 .05 .71 .20 1.19 1.09 1.78 .84 .63 .11 1.94 .03 .91 .06 .89 1.12 .93 .71 .40 1.05 1.98 1.81 .09 1.10 – – – 1.62 .93 1.27 1.13 .05 .35 1.09 .09 1.91 .03 1.07 1.61 1.89 1.29 1.63 1.82 1.66 – – .28 – – – .92 .88 .63 .98 – – .85 1.88 1.83 .15 1.46 1.02 1.86 .00 .77 .56 .91 .27 .15 .16 .99 – – 255 .13 1.00 1.48 1.12 1.01 .15 .97 .61 .97 1.90 .46 .75 .26 – – – – – – .96 1.91 1.02 .70 .92 .78 .71 .69 .31 1.98 1.96 – – – .93 .97 1.96 1.03 1.84 .4 WITHIN SAMPLE RMSES PX I II 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 All All CA CA CA JA JA JA AU AU AU FR FR FR GE GE GE IT IT IT NE NE NE ST ST ST UK UK UK FI FI FI AS AS AS SO SO SO KO KO KO BE BE DE DE NO NO SW SW GR GR IR IR PO PO SP SP NZ NZ 1.87 1.54 1.01 – – – – – – 1.91 .93 1.49 1.13 .87 .98 1.74 1.03 .93 1.08 – – – .92 1.08 – – 1.95 1.83 .65 .87 1.90 .94 1.12 1.85 .22 1.07 .72 .75 .18 1.88 1.86 – – L2 .16 .02 .06 1.03 1.86 1.11 .49 – – UR .99 1.02 .23 1.97 .95 .03 .08 1.05 1.04 1.18 1.03 .62 .02 1.04 .88 .87 1.46 – – – – – – .31 1.02 1.56 .76 .61 1.80 1.84 .34 – – – – – – .69 1.87 .00 .04 1.

74 1.02 1.83 .55 .29 1.12 1.04 1.03 PM 1.92 .99 1.26 .54 1. .59 1.83 .04 1.08 .05 1.08 1.12 PX – – – – 1.23 1.97 – – – – – – .61 .80 .05 1.77 1.3.43 .3 are based on only 4 observations compared to 19 in Table 9.16 1.49 . The outside sample results in this section are thus very preliminary.07 1. L2.1 and in this sense are less reliable.01 .02 1.14 1.00 1.256 Y 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 SA SA VE VE CO CO JO JO SY SY ID ID MA MA PA PA PH PH TH TH .00 .14 1.96 .93 . for example.59 .02 1.02 1.47 1.13 1. and U R are not part of the model for countries SA–TH.75 .94 . the MC model does much better for the exchange rate in Table 9.24 RS – – .95 .02 1.08 S .34 – – 9 TESTING THE MC MODEL Table 9.79 .98 .83 .02 1.30 .90 .34 1.73 .36 – – E – – – – – – 1.95 .66 – – – – – – – – – – .90 1.04 X85$ . and the results are presented in Table 9.08 1. and 1990. These coefficient estimates were then used to predict 1987.62 .03 1.87 1.30 1.31 . The same weighting scheme was used in Table 9.32 Variables W .13 1. each of the non US stochastic equations of the MC and ARMC models was estimated through 1986.08 1.91 .97 .03 .00 1.3 and vice versa.75 1.03 1.94 1.63 .33 1.73 .30 – – .41 1.76 .05 1.75 .92 1.65 – – V 1/Y 1.5 Outside Sample RMSEs As noted at the beginning of this chapter.1 (continued) M C I 1.89 1.74 1.13 1.96 .82 .1.84 .65 1.00 1.98 .15 .91 .09 1.05 1.97 .71 .89 .69 . J .56 . Each number is the ratio of the MC RMSE and the ARMC RMSE.89 .30 – – 1. When comparing Tables 9. 1988.92 .07 1.10 .28 1.13 .61 1.10 1.75 1.89 . RMSEs were computed for these forecasts.23 1.99 1.74 .06 1.3 differ in that some variables for which the MC model does better in Table 9. The results in Tables 9.21 1.1 and 9.00 .91 1. L1.85 1.13 1.58 .99 1.91 .11 1.89 1.30 1.20 1.00 1.15 1.86 1.95 PY – – – – 1.40 1.12 1.1 and 9. 1989.28 1.16 – – – – 1.95 .77 .22 1. 9.07 1.97 .26 1. The outside sample results were obtained as follows.13 1.99 1.04 .14 1. the one year ahead results in Table 9.18 1.82 1.76 .60 1.27 1.76 .94 .01 .45 . This gave for the annual countries 4 one year ahead forecasts and 3 two year ahead forecast.30 1.24 1. short sample periods for many countries limit the amount of outside sample work that can be done.06 1. The following discussion of Table 9.00 .3.21 1.1 do worse in Table 9.17 .4 for the quarterly countries and 1986 for the annual countries.03 .3 remember that.87 .98 1.3 will concentrate on the weighted results.04 1.86 .69 1.79 .49 .45 .3 as was used in Table 9.61 – – – – – – – – – – – – – – . For example. First. For the quarterly countries there were 4 one through four quarter ahead forecasts and 3 five through eight quarter ahead forecasts.61 1.44 .65 .18 1.

14 5.45 .5 OUTSIDE SAMPLE RMSES Table 9. It does better.05 1.17 8 1.99 1.12 1. the MC model appears to do somewhat better relative to the ARMC model in Table 9.82 . The other variables for which the MC model does worse are consumption.07 GDP D: GDP Deflator Alone .75 MC . Again.98 1.04 8.46 1.38 .13 1.86 .33 MC . sometimes considerably better.04 1.68 .95 . and so they are very tentative.3 one would conclude that structural exchange rate equations dominate autoregressive ones. however.98 1. where the ratios are 1.75 1.45 . these results are based on only 3 or 4 observations.44 .16 1. This suggests that the ARMC equations may be somewhat more subject to within sample data mining problems than are the MC equations.92 1.66 .57 .28 U R: Unemployment Rate Alone .20 1.15 1.50 1. and this is left to the reader. Overall.39 .52 .03 P I M: Import Price Deflator MC 2. One can get from Tables 9.46 .77 1.72 1.72 1.14 1.3 an idea of the accuracy of the model for the individual countries.75 1. From the results in Table 9.52 1.11 6.16 1.66 5.54 1. On the other hand.93 1.1.27 .69 4.82 1.52 MC÷Alone 1. When more data become available in the future. it will be interesting to put the MC model through more tests.95 RS: Bill Rate Alone .00 1.05 7.22 .64 4.57 for both the one year ahead and two year ahead results.07 1.09 1.99 1.77 . employment.29 3.52 and 1.9.04 .98 . In .62 1.09 5.22 . The ratio is .85 1. the MC model does worse for the interest rate.20 3.41 .2 US RMSEs: US Alone Versus US in MC Model 1 Alone MC MC÷Alone 2 3 4 6 1.71 1.61.52 1.1 and 9.76 .18 1.50 1.52 1.38 MC÷Alone 1.48 .36 .3 than in Table 9.54 MC .01 .13 1.44 EX: Exports MC 2.47 Errors are in percentage points. This completes the discussion of the RMSE results.16 257 GDP R: Real GDP . and the unemployment rate.96 MC÷Alone . for the other variables.

80 .43 2.06 1.69 .52 1.67 .20 1.50 .17 1.49 2.30 .21 .36 1.59 .42 3.32 2.00 .48 .34 .21 1.97 .97 3.98 2.24 – – – .3 Ratios of Outside Sample RMSEs 1987-1990: MC/ARMC Y PY RS E M C I V 1/Y I II 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 All All CA CA CA JA JA JA AU AU AU FR FR FR GE GE GE IT IT IT NE NE NE ST ST ST UK UK UK FI FI FI AS AS AS SO SO SO KO KO KO BE BE DE DE NO NO SW SW GR GR IR IR PO PO SP SP NZ NZ .67 – – – 1.27 .11 .93 .28 3.70 .59 .57 .86 1.48 .06 .46 1.98 1.61 .76 2.30 .20 1.83 .81 1.01 .90 1.12 1.88 3.10 1.75 .72 .23 .51 .93 .96 1.71 1.52 .11 2.62 .51 .95 2.26 2.64 1.30 2.47 .96 3.11 – – .258 9 TESTING THE MC MODEL Table 9.31 1.76 8.66 .68 .66 – – 2.79 .10 1.61 .15 1.45 2.64 .31 1.62 .31 .96 .42 1.85 .04 4.80 .29 .53 .82 3.43 .66 4.37 .28 3.63 2.72 .64 .10 .01 1.43 1.58 .43 – – – .22 1.80 .79 4.13 8.38 1.62 1.44 .47 .26 – – .27 1.65 5.13 .06 .97 1.64 4.73 .60 4.06 .48 .58 .84 4.19 2.90 .74 .18 1.29 1.00 1.50 .17 .99 1.45 .55 .61 .78 .72 .18 1.95 – – – 2.47 2.08 .79 1.60 .60 .44 .86 4.33 .35 1.42 .02 1.83 3.09 9.52 1.90 .50 .70 1.57 .25 .55 .15 1.94 3.56 .49 .56 .61 .52 2.69 2.11 .10 .68 .28 .54 – – – – – – .38 .94 4.20 1.56 1.96 .22 .73 1.32 .95 1.27 .39 1.46 2.58 1.69 .20 .84 .64 .54 1.06 5.21 .96 .04 .00 .64 1.48 – – – – – – .20 1.45 .52 .07 1.58 1.67 .61 .80 .60 .57 .53 .71 1.69 .90 – – .02 .21 1.94 .77 .99 .67 .87 .81 .48 .16 .82 .36 .64 .41 .17 3.52 1.54 .73 .96 6.83 .20 1.15 1.30 – – X85$ .36 1.76 .42 .13 2.05 1.89 1.79 .63 .66 .24 1.03 1.56 .31 1.55 .05 .75 .53 3.33 1.57 .98 15.84 .51 .17 .70 1.35 2.22 .74 .20 .84 .69 .25 .50 .11 .17 1.29 .95 .44 2.83 3.44 .27 .73 .95 .52 2.15 2.01 .85 .55 .89 .18 3.85 .38 2.50 3.60 .14 2.81 1.77 .51 .99 1.81 .18 .15 1.52 2.49 .50 .35 .70 1.70 1.52 .43 .97 4.74 .40 1.50 1.64 .28 .58 .59 .94 2.49 .49 .62 .74 1.91 .11 3.94 .22 .36 1.05 .69 1.30 1.38 4.39 3.61 1.74 .29 .54 .39 .47 1.74 .27 .24 .47 2.57 .10 2.89 1.94 1.16 1.38 7.89 2.24 .58 2.29 2.33 .06 1.81 1.50 .71 1.02 .80 .51 .26 .31 2.45 1.22 .19 .54 .64 1.21 .24 .53 1.12 2.72 1.09 1.51 .64 5.24 .14 1.42 .30 2.88 .95 .60 1.88 1.73 .13 1.38 .72 1.57 .67 1.51 1.92 .36 .98 1.55 1.86 .05 – – 1.93 .85 .89 .85 .60 1.15 1.07 .55 1.56 .28 .11 .46 .93 .69 .46 .88 2.64 .49 1.40 .33 .33 .82 2.88 1.70 .73 .01 .48 .69 .21 .30 .99 4.54 1.91 1.01 3.95 1.25 .18 11.80 .15 1.25 1.90 .43 .41 1.10 2.57 1.49 .27 3.98 .22 .02 5.35 .23 .57 .32 .71 .27 .02 .79 .55 .19 .63 .36 5.44 .80 .06 .59 .69 1.46 .29 .82 1.52 .07 .99 .41 2.30 – – – – – – 1.32 .92 1.29 2.24 .70 1.03 1.51 .38 .72 1.88 .86 .94 .86 .89 .71 .72 .59 1.23 .60 .56 2.57 1.71 1.50 .75 .27 .59 .

13 .05 – – – 1.47 .81 .92 1.09 .13 .76 .26 .62 .45 1.21 1.81 1.53 .09 .17 .56 .84 – – – – – – .77 .41 .43 3.18 – – – – – – 1.97 .94 2.57 3.19 2.45 1.19 1.09 .65 .09 .78 .11 .57 .52 .51 .70 1.00 .32 .96 – – 2.08 5.93 – – .50 .88 1.85 3.81 .92 .14 .00 .52 – – – – – – 2.35 – – UR 1.14 .33 .43 .69 .01 .10 .54 .18 .50 4.61 .46 1.82 – – – 1.73 .70 – – – 1.32 .40 .81 .06 1.89 1.79 1.75 .24 1.88 6.85 .96 1.54 .69 2.13 .13 1.75 .25 – – – 5.60 .42 .28 .95 1.51 .18 .85 .44 .93 .27 .92 .92 .59 1.17 .25 .42 .34 .42 1.10 1.18 1.67 .37 .94 Table 9.25 2.69 1.88 .65 .85 2.96 2.66 .75 .88 2.44 .32 3.59 .19 .21 .14 – – L2 .78 .61 1.49 .32 .23 .78 .34 – – – – – – 2.05 .92 .37 .29 .01 1.77 1.06 1.12 1.92 .48 .87 .39 1.21 4.21 2.61 1.42 .31 1.34 1.95 – – – – – – .75 .06 – – – – – – 4.52 1.98 – – – 1.88 .82 2.53 1.57 .00 .65 1.53 .55 .18 .58 .40 .70 1.23 .86 10.06 .31 .91 .93 .47 2.43 1.60 .14 .67 1.03 2.50 1.57 .34 1.22 .21 .37 .30 – – – – – – 1.33 1.98 .53 .55 .65 .51 .68 .52 .88 2.47 .94 4.92 .42 1.09 .31 1.31 7.05 1.69 .14 .65 .29 .32 1.02 1.16 .53 1.69 1.53 .49 .76 .12 .11 – – – 3.43 .30 3.44 1.23 .48 1.82 .57 .77 1.49 2.77 – – 259 .14 .88 1.12 .19 .31 1.61 – – – 1.15 1.20 .61 2.82 – – – 1.64 3.47 .33 1.71 1.3 (continued) S W J .16 .29 2.67 .48 PM .64 1.17 1.51 .21 .60 .9.23 .43 .31 .77 1.66 .60 .06 3.08 1.15 1.16 .87 .74 .15 .03 – – – – – – .42 .41 1.07 .15 – – 2.02 .46 1.16 – – – 2.14 – – – 3.61 .68 1.94 .09 1.33 .53 7.60 .09 .33 – – L1 1.50 1.10 2.73 .66 .72 1.41 .66 1.52 3.42 .59 1.41 1.13 1.00 .94 .14 1.5 OUTSIDE SAMPLE RMSES PX I II 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 4 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 All All CA CA CA JA JA JA AU AU AU FR FR FR GE GE GE IT IT IT NE NE NE ST ST ST UK UK UK FI FI FI AS AS AS SO SO SO KO KO KO BE BE DE DE NO NO SW SW GR GR IR IR PO PO SP SP NZ NZ .69 .58 .40 .82 4.46 .85 – – .48 .15 1.31 4.46 .75 1.11 .97 4.16 4.31 .43 .34 .37 1.92 .07 1.51 .23 1.29 .65 .82 .16 .83 .34 1.50 1.39 1.10 1.77 1.80 .23 1.77 1.16 1.52 – – .67 .06 2.19 .74 1.97 .69 1.29 2.51 .17 .23 .27 2.41 .60 1.69 .84 1.99 1.39 .07 1.99 .60 .73 .39 .50 1.25 2.39 .73 .79 2.42 1.85 1.78 .39 .78 – – 1.56 3.61 .77 1.63 .10 2.58 .65 1.93 .38 .56 .38 .13 .71 .16 1.

74 1.78 1.13 .89 .20 1.54 .31 1.83 2.25 .67 .13 1.59 1.09 – – E – – – – – – 1.07 .59 – – V 1/Y 1. J .55 .17 .59 1.95 .81 .93 .26 .68 .14 – .62 .12 1.07 .2.78 2.97 – – Table 9.69 .59 – – – – – – – – – – – – – – .87 .26 .52 1.79 – – .18 .93 1.22 1.04 .77 1.90 2.97 1.16 – – .13 1.68 – – – – – – 4.03 .50 5.37 1.54 .63 2.04 .3 (continued) M C I .20 – – – – – – – – – – 2.60 S 1.37 2.10 .01 7.10 1.99 2.25 PM .98 1.58 .18 – – .06 1.74 .63 .02 1.42 2.09 1.37 .74 – – 1. Other testing techniques will also become available when stochastic simulation of the MC model becomes practical.42 .58 3.77 .01 .99 1.43 1.17 1.59 .72 .59 1.91 3.89 – .57 .70 .79 4.75 6.76 .31 .33 2.20 .68 .01 .10 1. Each number is the ratio of the MC RMSE and the ARMC RMSE.63 1.85 .31 . and U R are not part of the model for countries SA–TH.04 .86 .39 .32 1.48 X85$ 1.68 . L1.86 .41 2.71 3.15 . which if true would suggest that the MC model has not handled all the lags right.29 1.96 .65 2.96 3.84 PY – – – – .81 .20 1.05 2.29 1.89 1.17 Variables W .94 .04 1.66 – – – – 1.68 .75 .73 1.70 .79 .45 10.77 .14 1.74 .75 PX – – – – 1. The possible future use of stochastic simulation of the MC model is discussed in Section 9.46 1.96 RS – – 1.50 – – – – .260 9 TESTING THE MC MODEL Y 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 SA SA VE VE CO CO JO JO SY SY ID ID MA MA PA PA PH PH TH TH .23 1.41 1.43 .85 1.62 . it will be interesting to see if the ARMC forecasts contain information not in the MC forecasts.68 .60 1. .95 1.02 1.82 .72 3.11 1.56 2.51 3.71 1.48 .32 1. particular.28 .09 1.01 1.78 .75 .50 .54 1. L2.77 1.66 .58 8.99 1.52 .39 1.36 – .53 4.39 .85 2.00 1.04 1.

the fact that an econometric model has properties that are consistent with the theory is in no way a confirmation of the model. Therefore. the final version of a model is likely to have multiplier properties that are similar to what one expects from the theory. The methods discussed here are not tests of models. These methods are then applied in Chapter 11 to the US model and in Chapter 12 to the MC model. However. theoretical) grounds. A model that does not do well in tests is not likely to have properties that accurately reflect the way the economy works.e. Models must be tested using methods like those in Chapters 4 and 7. It is sometimes argued with respect to the testing of models that if a particular model has properties that seem reasonable on a priori (i. They are meant to be used after one has some confidence that the model being analyzed is a reasonable approximation of the economy. this is evidence in favor of the model.. 261 .1 Introduction This chapter discusses various methods for analyzing the properties of macroeconometric models. not by examining the “reasonableness” of their multiplier properties. Essentially one does not stop until this happens. because of the back and forth movement between specification and results.10 Analyzing Properties of Models 10. that occurs in macro model building. including multiplier results.

The form of the output that is examined depends on the nature of the problem.1 Deterministic Simulation The procedure that is usually used to compute multipliers is based on deterministic simulation. there is no obvious thing to divide the change in the endogenous variable by. . it can merely be. The procedure described in the present section uses Hall’s suggestion. and let yitk denote the predicted value from the ˆb b simulation that uses xt+k−1 (k = 1.2. Hall pointed out to me that a more straightforward procedure is simply to draw both error terms and coefficients at the same time. the change or percentage change in the endogenous variable itself. In private correspondence in 1984. . Section 9. and let xtb denote an alternative set. the predicted values of the endogenous variables can be computed. if more than one exogenous variable has been changed. T ) for the exogenous variable values. although the first stochastic simulation could be avoided if one were willing to assume that predicted values from deterministic simulations are close to mean values from stochastic simulations. In fact. Let xta denote a “base” set of exogenous variable values for period t. which is generally the case in practice. Given 1) the initial conditions as of the beginning of period t. . .3. for example. and 4) values of the error terms for the entire period. denoted δitk . Let yitk denote the k period ahead predicted value of endogeˆa a nous variable i from the simulation that uses xt+k−1 (k = 1.2 Computing Multipliers and Their Standard Errors1 A useful way of examining the properties of a model is to consider how the predicted values of the endogenous variables change when one or more exogenous variables are changed. 1 The . S. and thus the word “multiplier” should be interpreted in a very general way. The output that one examines from this exercise does not have to be the change in the endogenous variables divided by the change in the exogenous variable. 2) the coefficient estimates. Section 9. 10. T ). . . is an estimate of the effect on the endogenous original discussion of the procedure discussed in this section is in Fair (1980b). Assume that the prediction period begins in period t and is of length T . The original procedure required that a stochastic simulation with respect to the error terms be done within a stochastic simulation with respect to the coefficients. 3) a set of exogenous variable values for the entire period.3. although the use of the word “multiplier” is somewhat misleading. The difference between the two ˆ predicted values. This exercise is usually called multiplier analysis.G. This avoids any stochastic simulations within stochastic simulations.262 10 ANALYZING PROPERTIES OF MODELS 10. . This section is thus a replacement for Fair (1980b) and Fair (1984). . It was also discussed in Fair (1984).

for example. Once the residuals are added to the equations.1) Obvious values of the error terms to use in the deterministic simulations are their expected values. ˆb 10. This does . a model in which inflation responds in a very nonlinear way to the difference between actual and potential output: inflation accelerates as output approaches potential. the model predicts the actual data before any policy change is made. The same set of values is used for all experiments. they are never changed. In other words. but this will not necessarily be the case if the model is predicting a much lower level of output than actually existed. Consider now a period in which output is close to potential. By “actual” in this case is meant the values of the estimated residuals that result from the estimation of the equations. which is simply to use the actual (historical) values of the error terms rather than zero values. If the actual values of the error terms are used. since the error terms are assumed to be uncorrelated with the exogenous variables. If these values are used and if the actual values of the exogenous variables are used. This solution will be called the “perfect tracking” solution. There is an easy answer to this problem if the simulation is within the period for which data exist.1 is not an unbiased estimate of the change because the predicted values are not equal to the expected values. the inflationary consequences of the policy change will not be predicted very well. This experiment should be quite inflationary. if the model is predicting that output is not close to potential when in fact it is.2 Stochastic Simulation ˆ For nonlinear models δitk in 10. For linear models it makes no difference what values are used as long as the same values are used for both simulations. With the use of the actual residuals. Consider. and in this case the choice of zero values has some problems.10.2. For nonlinear models the choice does make a difference. which are almost always zero. and consider an experiment in which government spending is increased. and thus a simulation is ˆa only needed to get yitk . yitk is simply the actual value of the variable. The use of the actual values of the error terms has the advantage that only one simulation needs to be performed per policy experiment. Note that this procedure is not inconsistent with the statistical assumptions of the model.2 MULTIPLIERS AND THEIR STANDARD ERRORS variable of changing the exogenous variables: ˆ ˆb ˆa δitk = yitk − yitk 263 (10. the problem regarding the response of inflation to output does not exist. the simulation will result in a perfect fit.

. however. solve the model again using the alternative set of exogenous variable values bj b ˜ (xt+k−1 . . denoted ¯ ¯ ˜2 δitk . and so if one were only interested in estimates of the changes. and the variance. since deterministic predictions are generally quite close to the mean values from stochastic simulations. compute the mean. . of δitk : ¯ δitk = (1/J ) J J j =1 j ˜j δitk (10.2) 4. 3. For the same set of error terms and coefficients as in step 1. T ). and the square root of sitk is its standard error. k = 1. denoted sitk . where J is the desired number of repetitions. . Draw a set of error terms and coefficients and solve the model using a the base set of exogenous variables values (xt+k−1 . T ). where α is the vector of coefficient ˆ ˆ ˆ ˆ is the estimated covariance matrix of α. V ) distribution. Given the values from the J repetitions. . . k = 1. estimates and V ˆ 1. ˆ ) distribution if the “base” values of the error terms are taken to be zero and from the N (ut . Compute bj aj ˜j ˜ ˜ δitk = yitk − yitk aj (10. Let yitk denote the k period ahead predicted value of variable i from this solution. .3) sitk = (1/J ) ˜2 j =1 ˜ ¯ (δitk − δitk )2 (10. 5. seem to be an important problem in practice. Let yitk denote the k period ahead predicted value of variable i from this ˜ solution. 2. ˆ ) distribution if the ˆ historical values of the error terms are used for the base values. Repeat steps 1 through 3 J times. it seems unlikely that stochastic simulation would be needed.264 10 ANALYZING PROPERTIES OF MODELS not.4) ¯ ˜2 δitk is thus the multiplier. where ut is ˆ the vector of historical errors for period t. The error terms are assumed to be drawn from the N (0. The main reason ˆ for using stochastic simulation is to compute standard errors of δitk . . . The coefficients are assumed to be drawn from the N(α. The stochastic simulation procedure is as follows. .

3 Sources of Economic Fluctuations2 There has been considerable discussion in the literature about the ultimate sources of macroeconomic variability.” such as changes in the demand for produced goods relative to services. and it is used in the next chapter in the application of the present approach. Economies seem complicated. that one has a model like US+ to work with and assume that the variable of interest is real GDP.3 Assume. 3 When 2 The . This was done for the stochastic simulation work in the next chapter. As discussed in Section 7. (See Shiller (1987) for more references. including the error terms in the exogenous variable equations if such equations are added. one may want to take the covariance matrix of the error terms to be block diagonal. one can estimate the variance of GDP by means of stochastic simulation. it is possible using stochastic simulation to estimate the quantitative importance of various sources of variability from a macroeconometric model. as it was for the probability calculations in Section 8. to “unusual structural shifts. to changes in desired consumption for Hall (1986). to oil price shocks for Hamilton (1983). sometimes only one. to unanticipated changes in the money stock for Barro (1977). Let σit denote the estimated variance ˜2 of real GDP (endogenous variable i) for period t. Shocks to the stochastic equations can be handled by a straightforward application of stochastic simulation. Since by definition exogenous variables are not modeled. as discussed in Section 8. and it may be that there are many important sources.10. where the estimated variance is based on draws of all the error terms in the model.3. it is not unambiguous what one means by an exogenous variable shock. As discussed in this section. given the estimated covariance matrix of the error terms. Macroeconometric models provide an obvious vehicle for estimating the sources of variability of endogenous variables.2. Shocks to the exogenous variables can then be handled by stochastic simulation of the expanded model. The sources vary from technology shocks for Kydland and Prescott (1982).8. Now consider material in this section is taken from Fair (1988a). Shiller (1987) surveys this work.) Although it may be that there are only a few important sources of macroeconomic variability. this is far from obvious. using a model like US+. for Lilien (1982). One approach is to estimate an autoregressive equation for each exogenous variable in the model and add these equations to the model. The US+ model is a model like this. There are two types of shocks that one needs to consider: shocks to the stochastic equations and shocks to the exogenous variables. Shocks to the exogenous variables are less straightforward to handle.3 SOURCES OF ECONOMIC FLUCTUATIONS 265 10. where he points out that a number of authors attribute most of output or unemployment variability to only a few sources. therefore.

This is done in Section 11. In the second edition of Industrial Fluctuations. It is important to realize what is and what is not being estimated by this procedure. one can examine how sensitive the results are to the effects of the correlation of the error terms across equations to see how to weigh the results. would reduce the amplitude by about a quarter. . What is being estimated is the contribution of the error term in the exogenous variable equation to the variance of GDP. compute the variance of GDP. the standard deviation) of industrial fluctuations. any procedure is somewhat crude. Pigou (1929).e. say. the variance of GDP falls by 5 percent when the error term for equation k is not drawn. Two exogenous variables can have the same 4 Regarding the use of this difference as an estimate of an error term’s contribution to the variance of GDP. Let σit (k) denote the estimated variance ˜2 of real GDP based on fixing the error term in equation k at its expected value. then fixing. Another way to estimate this contribution would be to draw only the error term for equation k. Likewise. one can examine the effects of the error term correlation on the results. See Shiller (1987) for more discussion of this. these two procedures are not the same. where it will be seen that the main conclusions using the US+ model are not sensitive to the effects of the correlation. If the error terms across equations are correlated. There is no right or wrong way of estimating this contribution. He thought that the removal of “autonomous monetary causes” would reduce the amplitude by about half. Fortunately. and because of the correlation. and compare this variance to the variance when all the error terms are drawn. two error terms one at a time and summing the two differences is not the same as fixing the two error terms at the same time and computing the one difference.266 10 ANALYZING PROPERTIES OF MODELS fixing one of the error terms at its expected value (usually zero) and computing the variance of GDP again. The difference between σit and σit (k) is an estimate of how much the error ˜2 ˜2 term in equation k contributes to the variance of GDP. This contribution is not the same as the multiplier effect of the exogenous variable on GDP. the removal of “psychological causes” would reduce the amplitude by about half. Removal of “real causes. Robert Shiller has informed me that Pigou had the idea first. after grouping sources of fluctuations into three basic categories. Consider an exogenous variable shock.4. Again. one can say that equation k contributes 5 percent to the variance of GDP. say.” such as harvest variations. In this way one can examine the contribution that various sectors make to the variance of GDP. however. In this case the stochastic simulation is based on draws of all but one of the error terms.4 If. gave his estimate of how much the removal of each source would reduce the amplitude (i. If the error term in equation k is correlated with the other error terms in the model. One can fix the error terms in a subset k of the equations at their expected values and draw from the remaining equations. In the above discussion k need not refer to just one equation.

Let δit (k) be the difference between the two estimated variances: ˜ ˜2 ˜2 δit (k) = σit − σit (k) (10. The formula for the variance of σit is presented in equation ˜2 ˜2 ˜2 7. If one exogenous variable fits its autoregressive equation better than does another (in the sense that its equation has a smaller estimated variance). k equal to the first and second equations ˜it (k) for k equal to the first equation plus δit (k) for k equal to ˜ is the same as δ the second equation.8 in Chapter 7. and this can also be estimated. Third. The notation σit will be used to denote the estimated variance of endoge˜2 nous variable i for period t based on draws of all m + q error terms. is the variance of ˜ δit (k). to use measures of exogenous variable shocks other than error terms from autoregressive equations.10. It is possible. as noted above.6) . Ignoring the first two reasons. where ˜ subset k can simply be one equation. however. they are based on the use of an estimated covariance matrix of the error terms rather than the actual matrix. Second. What is of more concern here. Let j 2j 2j δit (k) = σit − σit (k) (10.5) ˜ In the US application in the next chapter. they are based on the use of estimated coefficients rather than the true values. it is not likely to be the same as the size of the multiplier. Because of the correlation of the error terms across equations. it is not ˜ in general the case that δit (k) for. other things being equal. First. values of δit (k) are computed for the one through eight quarter ahead predictions of real GDP and the private nonfarm price deflator for a number of different choices of k. say. it is possible to estimate the variances of σit and σit (k). then. Also. it will contribute less to the variance of GDP. Computational Issues For a number of reasons the stochastic simulation estimates of the variances are not exact. they are based on a finite number of repetitions. of course. it can turn ˜ out that δit (k) is negative for some choices of k. In other words. it is possible to estimate the precision of the stochastic simulation estimates for a given number of repetitions. The notation σit (k) will be used to denote the estimated variance when the error ˜2 terms in subset k of the equations are fixed at their expected values. but whatever measure is used.3 SOURCES OF ECONOMIC FLUCTUATIONS 267 multiplier effects and yet make quite different contributions to the variance of GDP.

the covariance between them ˜2 ˜2 will be increased. To conclude. . Fortunately. The trick is to make the covariance high. which can be done by using the same draws of the error terms for the computation of both σit and σit (k). This will be seen in Table 11. there is an easy trick available. all the above values can be computed.8) Given values of yit and yit (k). is then J 1 j ˜ ˜ var[δit (k)] = ( )2 [δ (k) − δit (k)]2 J j =1 it j j (10.4 Optimal Choice of Monetary-Policy Instruments5 Over twenty years ago today Poole (1970) wrote his classic article on the optimal choice of monetary-policy instruments in a stochastic IS–LM model. Any ˜2 ˜2 one equation of a model. If these same 8000 numbers are used to compute both σit and σit (k). When this trick is used. 10. it was ˜2 ˜2 ˜ not enough to make the variance of δit (k) anywhere close to being acceptably ˜ small. 5 The material in this section is taken from Fair (1988b). j = 1. The advent of inexpensive computing has made applications like this routine and thus has greatly expanded the questions that can be asked of such models.6.268 2j 10 ANALYZING PROPERTIES OF MODELS j where σit is defined in equation 7. but without any tricks. from the stochastic simulations.10 in the next chapter.7) ˜ ˜ The estimated variance of δit (k). 1000 repetitions leads to variances ˜ of δit (k) that are acceptably small. The estimated mean of δit (k) across the ˜ J repetitions is δit (k) in equation 10. One thousand repetitions was enough to make the variances of σit and σit (k) acceptably small. for example. The variance of δit (k) is 2 plus the variance of σ 2 (k) minus twice the covariequal to the variance of σit ˜ ˜ it ance. estimating sources of economic fluctuations in macroeconometric models is an obvious application of stochastic simulation. J . · · · .5: ˜ δit (k) = 1 J j δ (k) J j =1 it (10. Stochastic simulation error turned out to be a bigger problem than I originally thought it would be. denoted var[δit (k)]. requires 8000 draws of its error term for 1000 repetitions for a forecast horizon of 8 quarters.

Most people would probably agree that between about October 1979 and October 1982 the Fed put more emphasis on monetary aggregates than it did either before or after. Otherwise. such as those of Turnovsky (1975) and Yoshikawa (1981). This is the question that Poole examined. This is the policy where the Fed behaves according to the equation M = α + βr.6 Other studies that have extended Poole’s work. The closest study before the present work was that of Tinsley and von zur Muehlen (1983). a la the Poole analysis. Poole also showed that there is a combination policy that is better than either the interest rate policy or the money supply policy. the interest rate has seemed to be the Fed’s primary instrument. It is interesting to ask if the use of the interest rate can be justified on the basis of the Poole analysis. and parameters are such as to lead. Is the economy one in which the variances. the covariance of the two error terms. the variance of the error term in the LM function. They used this instrument to target a particular variable. Interestingly enough. These are the two “instruments” of monetary policy. Poole showed that the choice of the optimal instrument depends on the variance of the error term in the IS function. The intermediate targets they tried are the monetary base. covariances. and parameters in the model such as to favor one instrument over the other. The unemployment rate and the inflation rate are the “ultimate” targets. to the optimal instrument being the interest rate? Stochastic simulation can be used to examine this question using a macroeconometric model. but they did not analyze the same question that Poole did. called an “intermediate” target. where the their stochastic simulation experiments. In the present case the aim is to see how well the interest rate does when it is used as the policy instrument in minimizing the squared deviations of real output from its target value compared to how well the money supply does when it is used as the policy instrument. For each of these target choices. and the Federal Funds rate itself. have been primarily theoretical. three definitions of the money supply. and the size of the parameters in the two functions.4 OPTIMAL MONETARY INSTRUMENTS 269 Poole assumed that the monetary authority (henceforth called the Fed) can control the interest rate (r) or the money supply (M) exactly. covariances. in particular the interest rate over the money supply? The purpose of this section is to show that stochastic simulation can be used to examine Poole like questions in large econometric models. nominal GNP. Are the variances. Tinsley and von zur Muehlen always used the interest rate (the Federal Funds rate) as the policy instrument. If the aim is to minimize the squared deviation of real output from its target value. they examined how well the choice did in minimizing the squared deviations of the unemployment rate and the inflation rate from their target values. 6 In . Poole’s analysis had not been tried on an actual econometric model prior to the work discussed here.10.

The solution of optimal control problems using large scale models turns out to be fairly easy.5 Optimal Control Optimal control techniques have not been widely used in macroeconometrics. . Finally. and vice versa if the variance is smaller when the money supply is fixed. . . . but if they improve in the future.4 for the US model. . If the variance is smaller when the interest rate is fixed. A general specification of the objective function is W = h(y1 .7 It is possible through repeated stochastic simulation to find the optimal values of α and β for an econometric model. and let σit (M) denote the same thing when the money supply is the policy ˜2 instrument. .9) also Tobin (1982) for a discussion of this. xT ) 7 See (10. x1 . The following is a brief discussion of optimal control. . Chapter 10. . . . The first step in setting up a problem is to postulate an objective function. optimal control techniques are likely to become more popular. Assume that the period of interest is t = 1. T . 10. fix the money supply path and perform a stochastic simulation to get the variance of real GDP. The issue is then to compare σit (r) to σit (M) for i equal to real ˜2 ˜2 GDP to see which is smaller. Let σit (r) denote the stochastic simulation estimate of the variance of en˜2 dogenous variable i for period t when the interest rate is the policy instrument. If the horizon is more than one quarter ahead. so that. . then variances are computed for each quarter of the simulation period. A more complete discussion is in Fair (1984). The Procedure The procedure is a straightforward application of stochastic simulation. The simulations are dynamic. Second. the computed variance for the fourth quarter is the variance of the four quarter ahead prediction error. this is evidence in favor of the interest rate. for example. The variance of real GDP for a given period corresponds to Poole’s loss function if one takes the target value of GDP for that period to be the mean value from the stochastic simulation. First. compare the two variances. . yT .270 10 ANALYZING PROPERTIES OF MODELS parameters α and β are chosen optimally. Models may not yet be good enough to warrant the use of such techniques. and this procedure is also done in Section 10. . fix the interest rate path and perform a stochastic simulation to get the variance of real GDP.

. which are needed by the algorithm. . In the nonlinear case this does not lead to the exact answer because the values of W that are computed numerically in the process of solving the problem are not the expected values. . . the constrained maximization problem is transformed into the problem of maximizing an unconstrained function of the control variables: W = (z) (10. which means that 10. and let z be the k · T –dimensional vector of all the control values: z = (z1 .1 are not zero. values of W can be obtained numerically for different values of z. xT −→ W . .1 are all set to zero. The model can be taken to be the model presented in equation 4. . .10. . stochastic simulation has to be done. Let zt be a k–dimensional vector of control variables. For each iteration. Consider now the stochastic case. An algorithm like DFP is generally quite good at finding the optimum for a typical control problem. . zT ). .10) where ht (yt . xT . . xT . . . . . the problem can be turned over to a nonlinear maximization algorithm like DFP. given values for x1 . yT and then using these values along with the values for x1 . xT to compute W in 10. Stated this way. a scalar. xt ) (10. yT . . .1 in Chapter 4. In most applications the objective function is assumed to be additive across time. where the error terms in 4. . Nevertheless. . By substitution. can be computed numerically. . For nonlinear models it is generally not possible to express yt explicitly in terms of xt . is the value of the objective function corresponding to values of the endogenous and exogenous variables for t = 1. .11 explicitly as a function of x1 . In order to compute the expected values correctly. . Consider first the deterministic case where the error terms in 4. . xt ) is the value of the objective function for period t. .9 can be written T W = t=1 ht (yt . . For each value of z one can compute a value of W by first solving the model 4. . Given this setup. where zt is a subset of xt . . T . In this case . . which means that it is generally not possible to write W in 10. It is possible to convert this case into the deterministic case by simply setting the error terms to their expected values (usually zero).10. the optimal control problem is choosing variables (the elements of z) to maximize an unconstrained nonlinear function.5 OPTIMAL CONTROL 271 where W . the derivatives of with respect to the elements of z.11) where stands for the mapping z −→ y1 . . The problem can then be solved as above. x1 . .1 for y1 . . .

12) This procedure increases the cost of solving control problems by roughly a factor of J . then the optimal policy for the US model will generally be to achieve the output target almost exactly unless the weight on the inflation loss is very high. The US model has the following problem regarding the application of optimal control techniques to it. no optimal control experiments were performed in the next chapter. If the aim is to minimize a loss function that has in it squared deviations of output from some target value and inflation from some target value. Without some nonlinearity in price behavior at high levels of output. . and thus the bias in computing the expected value of W using deterministic simulation is likely to be small. . Steps 1 and 2 are repeated J times. Because of this problem. The bias in predicting the endogenous variables that results from using deterministic rather than stochastic simulation is usually small. . Let W j denote this value. the optimal control solution is likely to correspond to the output target being closely met unless the weight on the inflation loss is very high. A set of values of the error terms in 4. the expected value of W is the mean of these values: ¯ W = (1/J ) J j =1 ˜ Wj (10. yT and the value of W corresponding to this solution is computed from ˜ 10. .1 is drawn from an estimated distribution. ˜ 4.e. . 3.272 10 ANALYZING PROPERTIES OF MODELS each function evaluation (i. Reliable estimates of the behavior of the price level at very high output levels cannot be obtained in the sense that the data do not appear to support any nonlinear functional forms. The difficulty pertains to the demand pressure variable in the price equation 10. In this sense the optimal control exercise is not very interesting because it all hinges on the form of the output variable in the price equation. . about which the data tell us little. which was discussed in Chapter 5. 2.. Given the J values of W j (j = 1. where J is the number of repetitions. each evaluation of the expected value of W for a given value of z) consists of the following: 1.10. the model is solved for y1 . and it is probably not worth the cost for most applications. . . J ). . Given the values of the error terms.

however. The procedure for doing this is straightforward. the draws of the error terms are around their estimated historical values. If one is merely interested in the mean paths of the variables. These kinds of experiments are useful ways of teaching students macroeconomics. These properties can then be compared. In other words. economy would have been like in the 1980s had tax rates been higher and interest rates lower than they in fact were.6 Counterfactual Multiplier Experiments It is sometimes of interest to use a model to predict what an economy would have been like had something different happened historically. and experiments like multiplier experiments performed.7. then stochastic simulation is not likely to be necessary because mean values are usually quite close to predicted values from deterministic simulations. From this base the model can then be solved using either deterministic or stochastic simulation. the US model is used to predict what the U. then the estimated residuals are added to the model before solving it. if desired. An example of this is in Section 11. to the estimated properties of the actual economy. stop with the predicted economies.6 COUNTERFACTUAL MULTIPLIER EXPERIMENTS 273 10. for example. In Section 11. These economies can be treated like the actual economy. If one wants to use the shocks (error terms) that existed historically. Generating predictions in this manner is a way of answering counterfactual questions. One need not. If stochastic simulation is used.S.10. One chooses the exogenous variable changes to make and solves the model for these changes. where the effectiveness of monetary policy is examined in the predicted economy with higher tax rates and lower interest rates.7. one can examine the properties of the predicted economies using methods like the ones discussed in this chapter. The monetary-policy properties in this economy are compared to those estimated for the actual economy. .

.

which the results in Chapter 8 suggest may be the case. If the model is a reasonable approximation of the actual economy. and then a monetary-policy experiment is performed 275 . Section 11.4.2 contains a general discussion of the properties of the model.2.3.5 examines the choice of the optimal monetary-policy instrument in the model using the method discussed in Section 10.3 are the two main sections in the book to read to get an understanding of the US model.2 and 11. Sections 11. Section 11. The model is first used to predict what the economy would have been like had tax rates been higher and interest rates lower in the 1980s. Section 11. It uses as an alternative version of the model the equations discussed in Chapter 5 with the values led eight quarters added. then these two sections also provide insights into how the actual economy works.7 examines the question of whether monetary policy is becoming less effective over time because of the growing size of the federal government debt.6 examines the sensitivity of the properties of the model to the rational expectations assumption.3 using the method discussed in Section 10. Multipliers and their standard errors are computed in Section 11. This is background reading for the multiplier analysis to come. and this chapter applies these techniques to the US model.1 Introduction The previous chapter discussed techniques for analyzing the properties of models.11 Analyzing Properties of the US Model 11. Section 11. Section 11.4 examines the sources of economic fluctuations in the US model using the method discussed in Section 10.

The firm sector responds by increasing production (Y ): Equation 11. The results of this experiment are then compared to the results of the same experiment performed using the actual economy. 5. 2. output and employment will increase. How much output increases relative to the price level depends on how close actual output (Y ) is to potential output (Y S). if government spending is increased in the US model. households spend more. 11. and 4. The main way in which the economy expands in the US model from an increase in. 2. Employment increases. 4. The level of sales of the firm sector (X) increases because of the increase in government purchases of goods: Equation 60. Therefore. If there is disequilibrium in the theoretical model in the sense that the labor constraint is binding on households. and hours per job (H F ): Equations 12. The increase in investment and household expenditures increases X. As Y approaches a value 4 percent greater than Y S. and so on. government spending will result in an increase in output. say. jobs (J F ). As can be seen from the demand pressure variable in equation 10. the qualitative properties of the two models are similar. then an increase in. the predicted price level approaches infinity. which effectively bounds Y below a value greater than 4 percent of Y S. 3. The exercises in Sections 11. which leads to an increase in household expenditures: Equations 1. the closer is Y to Y S.8 to estimate what the economy would have been like in 1978 and 1990 had the Fed behaved differently. say.2 A General Discussion of the US Model’s Properties Because the theoretical model in Chapter 2 was used to guide the specification of the US model. The increase in Y leads to an increase in investment (I KF ). Similarly. Finally. employment increases further.8 are examples of counterfactual experiments discussed in Section 10. and 14. . which leads to a further increase in Y . 1. the model is used in Section 11. the labor constraint becomes less binding on households.7 and 11. the discussion of the properties of the theoretical model in Chapter 2 is of relevance here.6. government purchases of goods is as follows. 13. and so on. the more will the price level rise for a given change in Y . 3.276 11 PROPERTIES OF THE US MODEL using this economy. The increase in jobs and hours per job leads to an increase in disposable income (Y D).

· · ·) MF = f17 (RS. others indirectly affect the decision variables by influencing variables (through identities) that in turn influence. These are: MH = f9 (RS. the decision variables. . · · ·) BO/BR = f22 (RS − RD. which explains RS.2 GENERAL DISCUSSION OF THE PROPERTIES Fiscal Policy Variables 277 The main federal government fiscal policy variables in the US model are the following: COG Purchases of goods D1G Personal income tax parameter D2G Profit tax rate D3G Indirect business tax rate D4G Employee social security tax rate D5G Employer social security tax rate JG Number of civilian jobs JM Number of military jobs T RGH Transfer payments to households Some of these variables appear as explanatory variables in the stochastic equations and thus directly affect the decision variables. · · ·) CU R = f26 (RS.11. The effects of changing each of these variables (except J M) in the model are examined below. directly or indirectly. · · ·) BR = −G1 · MB 0= MB + MH + MF + MR + MG + MS − (9) (17) (26) (22) (57) CU R (71) 0 = SG − AG − MG + CU R + (BR − BO) − Q − DI SG (77) M1 = M1−1 + MH + MF + MR + MS + MDI F (81) The other key equation is the interest rate reaction function. Monetary-Policy Options To see the various monetary-policy options in the model. equation 30. it will be useful to list a subset of the equations in the US model.

the federal government budget constraint. can be matched to its stochastic equation. SG. Instead of treating AG as endogenous. This leaves equation 77. 57. In other words. MDI F . MF .) This means that AG is the variable that adjusts to allow RS to be the value determined by equation 30. demand deposit and currency holdings of the foreign sector. MH . and 26. gold and foreign exchange holdings of the federal government. is determined by an identity elsewhere in the model (equation 76). can be matched to the stochastic equations that explain them.e. One can also consider the case in which equation 30 is dropped from . M1 can be matched to its identity.. the reserve requirement ratio. The question then is what endogenous variable is to be matched to equation 77. the following variables in the above block are then exogenous: the discount rate. Bank borrowing. by open market operations. The federal government savings variable. and total bank reserves. that this is done only for expositional convenience. MS. and CU R. Q. BR. The model is simultaneous. the target bill rate is assumed to be achieved by the purchase or sale of government securities. and the federal government sector. the discrepancy term. DI SG. the obvious variable to match to equation 77 is AG. G1. The demand for money variables. 22. i. Remember. If AG is taken to be endogenous. (AG will be called the “government security” variable. it will be convenient to match variables to equations. Consider the matching of variables to equations in the block given above. equations 9. 17. If equation 30 is included in the model (and thus RS matched to it).278 11 PROPERTIES OF THE US MODEL In considering the determination of the variables in the model for the various monetary-policy options. which states that the sum of net demand deposits and currency across all sectors is zero. 81. BO. and the variable that is involved in the definition of M1. RD. This would mean that the target bill rate was achieved by changing the discount rate or the reserve requirement ratio instead of the amount of government securities outstanding. can be matched to its identity. it seems better to treat AG as endogenous rather than RD or G1. and so it is not a candidate. the state and local government sector. and nearly all the equations are involved in the determination of each endogenous variable. the net financial asset variable of the federal government. MR. one could take it to be exogenous and take either RD or G1 to be endogenous and match the one chosen to be endogenous to equation 77. however. Remember that AG is negative because the federal government is a net debtor. and MG. MB can be matched to equation 71. Since the main instrument of monetary policy in practice is open market operations.

The main effects of interest rates on the real side of the economy are the direct and indirect effects on household expenditures (equations 1. What this means is that the three instruments of monetary policy—AG. In the exogenous monetary-policy case. and the response of the model to changes in AG can be examined. Using all three instruments is essentially no different from using one with respect to trying to achieve. they affect the economy by affecting interest rates. For a given change in fiscal policy. that changes in RD and G1 have virtually no effect on the real side of the economy. there are a variety of assumptions that can be made about monetary policy. which in turn change real variables. and BR is solved using equation 77. and another is the level of nonborrowed reserves. 3. Both of these are common variables to take as policy variables in monetarypolicy experiments. namely. It is also possible in the exogenous monetary-policy case to take some variable other than AG to be exogenous. and the indirect effects are from interest revenue being a part of disposable income and disposable income appearing in the expenditure equations.2 GENERAL DISCUSSION OF THE PROPERTIES 279 the model.) When equation 30 is dropped.1 To return to fiscal policy for a moment. It also means in the endogenous monetary-policy case. it should be obvious that fiscal policy effects are not independent of what one assumes about monetary policy. RS is solved using equation 9. MB is solved using equation 57. The direct effects are from interest rates appearing as explanatory variables in the equations. MH is solved using equation 71. 1 The . say.11. where AG is endogenous and RD and G1 are exogenous. the main way in which monetary policy affects the economy is by changing interest rates. Any effects that they might have are simply “undone” by changes in AG in the process of achieving the target interest rate implied by equation 30. M1. If either of these is taken to be exogenous. but yet both RS and M1 remain endogenous. One possible choice is the money supply. and 4) and on nonresidential fixed investment of the firm sector (equation 12). some real output target. and G1— all do the same thing. 2. Changes in AG change interest rates. BR − BO. In this case RS is matched to equation 77 and AG is taken to be exogenous. monetary policy is exogenous. The interest rate is “implicitly” determined in this case: it is the rate needed to clear the asset market given a fixed value of AG. RD.8 in Appendix A. The main possible assumptions are 1) equation way in which the model is solved under alternative monetary-policy assumptions is explained in Table A. AG must be endogenous. (In the numerical solution of the model in this case.

Consider first the links from output to the unemployment rate. 6. There is no reason to expect Okun’s law to hold in the sense of there being a stable relationship between output and the unemployment rate over time. The third link is that when the number of jobs increases. AG is endogenous. 4) nonborrowed reserves exogenous. Also. At any one time the relationship depends on such things as the amount of excess labor on hand and the value of the labor constraint variable. the relationship between output and the unemployment rate varies over time. How much the number of people in the labor force changes in any one period also depends on the value of the labor constraint variable. According to this equation. how much the number of jobs changes in any one period depends in part on the amount of excess labor on hand. For example. and 7).280 11 PROPERTIES OF THE US MODEL 30 included in the model and thus monetary policy endogenous. This means that the number of new people employed increases by less than the number of new jobs (equation 85). The second link is that when the number of jobs increases. the number of people in the labor force increases (equations 5. The size of these links in the model is such that the unemployment rate initially drops less than the percentage change in output. 3) the money supply exogenous. Various Relationships To conclude this general discussion of the model’s properties. the number of people holding two jobs increases (equation 8). The first link is that when output increases. which also varies over time. because the links vary in size over time. This means that the unemployment rate falls less than it otherwise would for a given increase in the number of new people employed (equations 86 and 87). Because this relationship is not constant. In all but assumption 5. it will be useful to consider the relationships in the model between certain endogenous variables. Although the percentage increase in jobs is less than the percentage increase in output. and 5) government securities outstanding. The sensitivity of fiscal policy effects to the first three of these assumptions is examined below. The relationship between output and the price level is also not necessarily . exogenous. the number of jobs increases (equation 13). How much the number of people holding two jobs changes in any one period depends in part on the value of the labor constraint variable. the relationship between jobs and output is not constant across time. which varies over time. 2) the bill rate exogenous. the initial percentage increase in the number of jobs is less than the percentage increase in output. the variables do not obey Okun’s law. AG.

In equation 10 other things affect the price level aside from output. the unemployment rate. Results are presented for real GDP.1. in particular the price of imports.11. the bill rate. and when these other things change. where one policy variable was changed per experiment. J F · H F changes by less than this percentage in the immediate quarter. the variable presented in Table 11. The number of repetitions for each stochastic simulation was 250. The buffer for this is the amount of excess labor held: as output falls.2 The values in the 0 rows are the estimated effects from the deterministic simulations. and vice versa.1 is −SGP . excess labor builds up. the price level will change even if output does not. the last 16 quarters of the sample period. the private nonfarm price deflator. is procyclical in the model.3 MULTIPLIERS AND THEIR STANDARD ERRORS 281 stable over time in the model. Similar considerations apply to the amount of excess capital held. For the stochastic simulations the error terms were drawn from the N(ut .3. Y /(J F · H F ). which is −SGP . and the results are presented in Table 11. where ut is the vector of historical errors for ˆ ˆ is easier to discuss the government deficit as a positive number. A tight relationship is even less likely to exist between the price level and the unemployment rate because of the many factors that affect the labor force and thus the unemployment rate but not necessarily output. Productivity defined as output per paid for worker hour. SGP is in nominal terms.2.3 11. 2 It . thus making the base solution the perfect tracking solution. Consequently. For the deterministic simulations the historical errors were added to the equations and treated as exogenous. ˆ ) distribution. Consider finally the relationship between output and employment. Eight experiments were performed. Other things being equal. The 2SLS estimates in Chapter 5 were used for these results. excess labor is gradually eliminated because it has a negative effect on the demand for employment and hours. The simulation period was 1989:3–1993:2. and the federal government deficit. and the values in the b rows are the estimated standard errors computed from the stochastic simulations. 11. Excess capital is gradually eliminated because it has a negative effect on investment. the values in the a rows are the estimated effects from the stochastic simulations.1 Computing Multipliers and Their Standard Errors Fiscal Policy Variables Multipliers and their standard errors were computed in the manner discussed in Section 10. The first set of experiments concerns the fiscal policy variables. When Y changes by a certain percentage.

13 .31 . V ) distribution.32 .34 .37 .02 .46 .91 .00 1.87 .69 .07 .90 .36 .06 .40 .06 .22 .00 . .13 -.21 1.03 1.76 .05 .27 .27 .09 .04 .84 .18 1.31 .24 16 .11 .25 . where α ˆ ˆ ˆ ˆ is the estimated covariance matrix is the vector of coefficient estimates and V ˆ of α.20 .40 .33 .91 .89 .62 1.76 .20 .33 .04 .59 .22 GDP R: 1.01 1.09 .06 2 Number of Quarters Ahead 3 4 8 12 Real GDP 1.63 .38 .53 .13 .44 .89 1.37 .33 1.34 .13 -.25 -.31 1.03 . ˆ For the first experiment COG was increased from its historical value each quarter by an amount equal to one percent of the historical value of GDP R in that quarter.14 .72 .56 .32 .00 .22 .75 1.69 .91 .91 .83 .69 .03 .68 .21 .02 .25 -.1 Estimated Multipliers and Their Standard Errors for Eight Fiscal Policy Experiments 1 COG ↑ 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 1.26 1.07 .28 . The units in Table 11.24 .34 .04 .27 .30 .01 1.20 .11 D1G ↓ D2G ↓ D3G ↓ D4G ↓ D5G ↓ JG ↑ T RGH ↑ period t.00 .56 .11 .70 1.60 .06 .36 .11 1.17 .21 .11 .70 .30 .11 .05 -. The coefficients were drawn from the N (α.77 1.11 .01 1.15 .11 .64 .24 -.33 .32 .18 .36 .11 1.04 .15 .84 .46 .32 .15 .11 .36 .10 .34 .11 .87 .01 .00 1.23 . The dimension of ˆ is 30×30.16 1.24 .29 .73 .03 1.49 .18 .92 .05 .02 1.33 1.32 .73 .28 1.01 .69 .282 11 PROPERTIES OF THE US MODEL Table 11.07 .15 .09 .08 1.76 .29 .77 1.32 .63 .93 .37 .50 . For GDP R and P F .01 .32 1. and the dimension of V is 166×166.04 .11 .29 .31 .1 are as follows.02 1.03 .

RS.01 . where the units are in percentage points for U R and RS and in billions of dollars for −SGP .18 .06 .05 -.10 .11.10 .49 .10 .70 .00 .01 .00 .37 .92 .43 .00 .99 1.10 .00 .80 .06 .06 .00 .46 .11 .84 .45 .45 .00 -.42 .26 283 P F : Price Deflator .12 .11 .80 .20 -.08 -.99 1.17 .49 .18 . and −SGP a number in the 0 row is simply δitk .82 .50 .02 .18 .31 -.46 .04 .18 .10 . a number in the 0 row for GDP R or P F is the percentage change in the variable (in percentage points) that results from an exogenous increase in GDP R of one percent.90 .00 .01 .42 .00 .03 . and −SGP .02 .36 .18 . a number in the a row .01 -.47 .19 . where δitk is defined in equation 10.31 -.00 .94 .34 .00 .01 -. ˜j ˜ aj A number in the a row for GDP R and P F is the mean of 100(δitk /yitk ) ˜j across the J repetitions (J equals 250 for each experiment).47 .17 .36 -. For U R.03 .35 .00 .33 .00 .08 D1G ↓ D2G ↓ D3G ↓ D4G ↓ D5G ↓ JG ↑ T RGH ↑ ˆ ˆ a number in the 0 row is 100(δitk /yitk ).18 .01 .00 .64 .39 .09 .98 .1 (continued) 1 COG ↑ 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b .03 .10 .38 . For ˆ U R.19 .41 .01 .18 .00 .07 .40 .82 .38 .2.41 .00 .05 .00 .3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.03 .04 1.28 .08 .22 .40 -.01 .00 .16 .18 16 .10 -.20 .20 .00 .01 .00 .27 .03 .02 1.17 .00 2 Number of Quarters Ahead 3 4 8 12 .43 .04 -.43 .14 -.37 . ˆa a values are the actual values because the base run is the perfect tracking The yitk ˆ solution.92 .05 .29 -.00 .01 .00 .03 .40 .04 .40 .43 .00 .07 .63 .04 -.11 .1.57 .49 .04 .00 .10 . where δitk is defined in equation 10.02 .30 .01 .18 -.14 .21 -.03 .00 .00 .00 .39 .00 .21 . RS.44 .28 .71 .80 . Since COG was changed by one percent of GDP R.04 .02 .44 .18 -.19 .

11 -.18 .00 .02 .08 .10 -.01 -1.22 -.13 -.32 .40 .14 -.50 .00 .26 -.93 -1.23 .59 .49 -.04 . which is the square root of sitk in equation 10.42 -.66 -.09 -.03 -. and −SGP .27 .22 -.73 .27 .06 -.08 -.05 .22 -.64 -.284 11 PROPERTIES OF THE US MODEL Table 11.13 -.28 -.39 -.52 .01 -. which is denoted δitk in equation 10.31 . is equal to [D1G + (T AU G · Y T )/P OP ]Y T . The aim is to change D1G so that the decrease in personal income taxes in real terms is equal to the change in COG.17 .05 -.40 -.00 -.39 -.49 .17 -.14 .12 D1G ↓ D2G ↓ D3G ↓ D4G ↓ D5G ↓ JG ↑ -1.90 -.64 .14 -.16 -.42 -1.28 .02 -.00 -1.05 .84 .29 .02 2 Number of Quarters Ahead 3 4 8 12 -.24 .41 .00 . A number ˜j ˜ aj in the b row for GDP R and P F is the standard deviation of 100(δitk /yitk ) from the J repetitions.13 Unemployment Rate -. The changes for the other policy variables in Table 11.08 . where Y T is taxable income.05 -.01 .07 -.06 -.89 -1.12 -.00 .02 .67 -.85 . a number in the b row is the ˜j ˜2 standard deviation of δitk .05 T RGH ↑ ˜j ¯ is simply the mean of δitk .20 .37 .67 -1.16 .01 .04 -.36 .72 -.05 .30 -.05 -.02 -.52 .05 -.36 .50 -.19 .07 .32 -.01 -.06 .26 .09 -.04 -.10 .1 (continued) 1 COG ↑ 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b U R: -.01 -.00 -.00 .34 -.06 .30 .56 -.12 .17 .24 .16 -.30 -.10 .18 -.52 .09 -.09 -. for example.01 .08 -.22 -.32 -.13 -. Let COG denote the change .06 .89 .14 . the change in D1G.1 were made to be comparable to the change in COG with respect to the initial injection of funds into the system.05 -. For U R.04 -.03 .03 -.10 .01 -.00 -.3.06 .20 -.02 -.02 .52 -1.24 .09 .18 -.43 -.03 -. T H G.04 -.27 -. Consider.57 -1.18 -.89 -.25 .39 .19 .11 .01 . RS.06 .01 .09 -.02 .04 .15 -.4.3).08 -.01 -1.08 -.87 -.38 .17 .06 .14 16 -.18 -. From equation 47 in the model (see Table A. the variable for personal income taxes.

60 .60 .10 -.41 .11 .01 1.37 .00 .10 . The other matchings are as follows: D3G .10 .21 -.16 1.46 .03 -.01 .13 .17 .99 .87 .79 .04 1.08 .34 .95 .64 .1 (continued) 1 COG ↑ 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b .85 .15 .32 .66 .20 .52 .11 .63 .16 .3 is 49.24 .21 .21 .37 1. The values that were used for P G and Y T for these calculations are the actual values.00 .48 .46 .53 . of course.52 .24 -.12 .02 -. The changes in the other policy variables are similarly done.08 .11 .3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.96 .08 .23 1.55 .40 .08 -.13 .11 .03 .19 . For D2G the relevant tax variable is T F G.32 .53 .53 .03 -.00 . not the predicted values.34 .04 .01 .34 .06 .05 1.65 .13 .01 .55 .12 1.46 .49 .12 1.52 .96 .00 .98 .09 .57 .08 -. The aim is to decrease D1G in such a way that the decrease in T H G is equal to P G · COG.94 .15 . the level of corporate profit taxes. where P G is the price deflator for COG.16 .43 .10 .14 .02 1.15 1.04 .00 .23 . The predicted values are.07 .05 1.31 .08 .39 1.25 .06 .10 -.01 .59 .01 -.11 .11 .59 .26 . All this procedure does is to change D1G by an amount that would lead personal income taxes to decrease by the historical value of P G · COG if nothing else happened.01 .90 .03 .24 .02 .38 .13 -.09 -. and the relevant equation in Table A.13 .45 . affected by the change in D1G.31 .14 1.83 .03 1.18 1.29 .11.03 2 Number of Quarters Ahead 3 4 8 12 1.19 .30 .59 .90 .23 285 RS: Bill Rate .15 .00 .01 -. The change in D1G for the given quarter is thus −(P G· COG)/Y T .03 .04 .07 .60 .79 .53 .25 .15 .32 .23 .06 1.15 .31 .57 .18 -.32 .08 -.12 .01 1.58 .43 .14 .13 D1G ↓ D2G ↓ D3G ↓ D4G ↓ D5G ↓ JG ↑ T RGH ↑ in COG for a given quarter.35 1.20 16 1.55 .84 .42 .57 .54 .13 .55 .04 -.

7 D1G ↓ D2G ↓ D3G ↓ D4G ↓ D5G ↓ JG ↑ T RGH ↑ 0 = Estimated effects from deterministic simulations.2 13.6 10. however. T XCR.9 1.6 1.5 16.1 18.5 11.3 9.7 12. J G to W G · J G · H G (no separate equation).1 12.3 .4 18.4 18. b = Estimated standard errors of a row values.5 .3 13.1 14.7 16.1 .4 11.0 .6 2.6 11.2 −SGP : Federal Government Deficit 9.1 11.2 12.5 1.3 13.1 18.6 13.4 2.1 11.1 2.2 .1 10.9 16.8 18.2 13.5 11.0 12.8 11.3 11.5 9.4 .4 10.7 11.5 11. No attempt was .6 .3 12.6 16.0 .6 13.2 11. a = Estimated effects from stochastic simulations.3 .0 16.1 11.8 15.6 3.5 8.9 12.4 1. A multiplier experiment could thus be run in which T XCR was changed.4 .8 15.1 9. is also a fiscal policy variable.2 11.4 .2 .8 8.0 .1 (continued) 1 COG ↑ 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 2 Number of Quarters Ahead 3 4 8 12 15.4 11.5 11.5 1.3 11.1 11.3 .1 1.2 11. each of the policy variables 3 The tax credit dummy variable.8 . The units are percentage points except for −SGP .5 .4 1.5 16.4 18.6 12. and T RGH to itself (no separate equation).0 11.4 .4 11.4 13. It appears in the investment equation 12.3 17.0 3.7 8.2 1.8 8.9 1.2 11.3 1.4 .6 13.6 1. D4G to SI H G and equation 53.0 12.3 9.4 .3 To repeat.4 . one would also have to estimate how much profit taxes would be changed by the T XCR change and then adjust D2G accordingly.2 .6 .9 17.8 21.5 11.7 15.9 12.4 18.5 13.7 21.4 11.5 16.8 11. which are billions of dollars.286 11 PROPERTIES OF THE US MODEL Table 11.0 .3 11.8 8.9 16 17.1 10.4 13.5 11.0 .4 13.8 13.3 19.1 .8 8.7 13.4 16.6 .3 10.4 .7 8.5 12.5 19.5 15.5 17.4 1.3 4.0 12. D5G to SI F G and equation 54.4 11.3 13.4 9.5 16.3 1.2 9.5 12.1 10. to I BT G and equation 51.4 8.8 9.1 .3 16.2 .4 . then.6 19.2 11.7 . In doing this.2 12.3 .7 11.0 .7 11.2 2.7 2.1 .3 11.4 .

The table shows that the government deficit (−SGP ) rises in response to the COG increase. In this sense all the experiments in Table 11. Even though macroeconometric models are nonlinear. considerably higher by the end of the sixteen quarter period. COG Increase Table 11. the price deflator.3 MULTIPLIERS AND THEIR STANDARD ERRORS 287 was changed each quarter by an amount that based on the historical values of the variables in the model led to the same injection of funds into the economy as did the COG increase. The output multiplier reaches a peak of 1. If such an experiment were run.5 billion higher. This is due in large part to the increase in government interest payments that results from the higher interest rates. The deficit increases are. . The reasons for the increase in output were discussed above. The results of the individual experiments will now be discussed.77 in the third quarter and declines after that. The relationship between government spending changes and changes in the government deficit is examined in more detail in Section 11. and the bill rate and to a decrease in the unemployment rate.7 below. however. the effects on real GDP would be small because the estimate of the coefficient of T XCR in equation 12 is small. The government deficit rises. and this is certainly true for the results in the table. The price deflator rises because of the effects of the increase in output on the demand pressure variable in the price equation 10. their predicted values based on deterministic simulations are generally close to the means from stochastic simulations. this increase is less than the increase in nominal government spending (P G·COG).1. and this discussion will not be repeated here. The unemployment rate falls because employment rises as a result of the output increase. After four quarters the deficit is $8. Turning now to the results in Table 11.1 are of the same size. This is because of the endogenous increase in tax revenue as a result of the expanding economy. made to run a T XCR experiment for the present results.1 shows that the increase in COG leads to an increase in output (real GDP).3.11. it is first immediately clear that the 0 and a rows are very close. Although not shown in the table. The Fed responds (equation 30) to the output and price increases by raising the bill rate. Part of the reason for the decline after three quarters is the higher value of the bill rate due to the Fed leaning against the wind.

over time interest payments by firms drop because they need to borrow less due to the higher after tax profits. and in fact by quarter 12 the unemployment rate is higher for the D1G experiment than it was in the base case. It also increases the after tax interest rates (RSA and RMA). other things being equal. which has a positive effect on the labor supply variables L2 and L3 and on the number of people holding two jobs (LM). which has a positive effect on household expenditures. which have a negative effect on household expenditures. the government needs to borrow more. These positive effects on the unemployment rate thus offset some of the negative effects from the increase in jobs as a result of the output increase. Table 11. is probably not trustworthy. and so interest payments by the government rise. On the other hand. however. The difference between the unemployment rate effects for the COG and D1G experiments is a good way of seeing why Okun’s law is not met in the model. and in fact.) Other things being equal. an increase in the labor force leads to an increase in the unemployment rate. this leads to an increase in after tax profits and thus to an increase in dividends paid by firms. D2G Decrease Table 11. A number of things affect the unemployment rate aside from output. since an increase in the number of people holding two jobs means that the total number of people employed rises less than the total number of jobs. This is a standard result. . The decreases in the unemployment rate are much smaller for the D1G decrease than for the COG increase.1 shows that the decrease in D2G. Also. (It also has a negative effect on L1. smaller than the effects of the COG increase. When the profit tax rate is decreased.1 shows that the net effect on the economy is expansionary. but this is more than offset by the positive effect on L2 and L3. the profit tax rate. increases disposable income (Y D). This is because the decrease in D1G increases the after tax wage rate (W A). This in turn leads to an increase in disposable income. as just noted. The same is true of an increase in LM. the personal income tax parameter. The effects of the tax rate decrease are. by quarter 12 the net effect is positive. has little effect on output. This result. Tax rate decreases generally have smaller effects than government spending increases in models because part of the decrease in tax payments of households is saved. The way in which the profit tax rate affects the economy in the model is the following. because it is receiving less in profit taxes. however.288 D1G Decrease 11 PROPERTIES OF THE US MODEL The decrease in D1G.

It is likely. they suggest that a very effective way to decrease the federal government deficit would be to raise the profit tax rate. This would raise revenue and have little negative impact on real output. This in turn leads the unemployment rate to fall less than otherwise. In fact. as was the case for the D1G experiment. which leads to only small changes in household expenditures and thus in output. it may be that firms pass on an increase in profit tax rates in the form of higher prices. which decreases the overall price deflator for the household sector P H (equation 34). and 37). which is not part of the model. D3G Decrease A decrease in D3G.3 MULTIPLIERS AND THEIR STANDARD ERRORS 289 The net effect of these interest payment changes on disposable income could thus go either way. however. the fall in the unemployment rate is less for the D3G experiment than it is for the COG experiment. This is because the fall in P H raises the real wage. The results . I tried adding D2G in various ways to the price equation 10 to see if an effect like this could be picked up. but with no success. which has a positive effect on labor force participation. 36. the net effect on disposable income is small.11. Y D/P H . and P CD (equations 35. It may be that profit tax rates are not changed often enough for reliable results to be obtained. After four quarters the output effect is slightly larger than it is for the COG experiment. the indirect business tax rate. The results in Table 11.1. decreases the consumption price deflators P CS. the employee social security tax rate. If the output multipliers from a profit tax rate increase are in fact as low as in Table 11. and 4. is similar to a decrease in D1G in that it increases disposable personal income.1 show that the decrease in D3G has a positive effect on output. the model is probably not trustworthy regarding the effects of D2G changes on the economy. P CN. At any rate. which has a positive effect on household expenditures in equations 1. Since dividends are slow to respond to after tax profit changes and since the interest payment changes nearly cancel each other out. 2. D4G Decrease A decrease in D4G. that changes in D2G affect the economy in ways that are not captured in the model. On the other hand. The decrease in P H raises real disposable income. the unemployment rate is higher than it was in the base case by the end of the 16 quarter period. 3. For example.

J G Increase An increase in J G. the initial decreases in the unemployment rate are larger for the J G experiment. the number of civilian jobs in the federal government. however. whereas for the COG experiment. fairly small. One small difference between the two experiments is that the after tax interest rates RSA and RMA are affected by changes in D1G but not D4G (equations 127 and 128). however. has a positive effect on household expenditures. This in turn leads to an increase in household expenditures. the level of transfer payments to the household sector. Again. this experiment suggests that an effective way to lower the government deficit would be to increase D5G. In particular. falls less for the D1G experiment than it does for the T RGH experiment. which. leads to an increase in employment and thus disposable personal income. among other things. The unemployment rate. much of the initial increase in output is produced by firms lowering the amount of excess labor on hand and increasing hours worked per worker rather than increasing jobs. firms may pass on changes in D5G in the form of wage changes. which increases household expenditures. The overall effect on real GDP is. This is because of the labor . increases disposable personal income. The lower price level leads to an increase in real disposable income. This would lower the deficit without having much effect on output. D5G Decrease A decrease in D5G. and like the D2G experiment. T RGH Increase An increase in T RGH . There are likely to be other firm responses to a change in D5G that are not captured in the model. There is a large immediate change in jobs for the J G experiment. which has a negative effect on the price level (equation 10). On the other hand. The output increases for the J G experiment are somewhat below the output increases for the COG experiment (except for the first quarter). the employer social security tax rate. The social security tax is only a tax on wage income. and this is not part of the model. The output effects of this experiment are similar to those of the D1G experiment.290 11 PROPERTIES OF THE US MODEL for this experiment are thus similar to those for the D1G experiment. lowers the cost of labor to the firm sector. these results are probably not trustworthy.

06 .7 -2.00 .08 P F : Price Deflator .1 -6.11.13 -.3 .5 shows that misspecification contributes some to the total variances of the forecasts from the US model.02 .1 See notes to Table 11.02 .22 .00 2 Number of Quarters Ahead 3 4 8 12 .00 -.00 .00 .1 -5.01 -.24 .09 -.10 .32 .1.14 .2 Estimated Multipliers and Their Standard Errors for a Decrease in RS of One Percentage Point 1 RS ↓ 0 a b 0 a b 0 a b 0 a b -.6 .15 . The assumption of correct specification is the key restriction in the present context. force increase in the D1G experiment caused by the lowering of the tax rate.22 -.0 -2.7 -2.12 16 .0 -2.01 .01 -.49 .00 .01 .24 .17 .00 .34 . Table 8.2 .01 .04 . There is no such tax rate effect at work for the T RGH experiment.6 291 GDP R: Real GDP .00 -.09 -.14 -7.14 . This result is consistent with the results in Table 8.4 -1.00 .01 . multiplier uncertainty will be small.01 -.06 . provided that the model is correctly specified.48 .0 1.00 .3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.24 .18 -.24 .27 . The Estimated Standard Errors The estimated standard errors (the b row values) in Table 11.1 are thus encouraging regarding the accuracy of the properties of the model.1 -5.24 . and so it should be taken .1 .05 .22 .7 1. Most of the uncertainty of multipliers comes from the uncertainty of the coefficient estimates.17 .03 .01 RS ↓ RS ↓ U R: Unemployment Rate .05 .19 .05 -. The results in the b rows in Table 11.1 in general seem fairly small. which show that the contribution of the uncertainty of the coefficient estimates to the total uncertainty of the forecast is generally relatively small.46 .5.04 -.1 .44 .5 -3.12 .00 .0 -6.7 -8.5 -3. and if the effects of coefficient uncertainty are small.10 -.29 .09 RS ↓ −SGP : Federal Government Deficit -1.01 .

49 percent. It will be seen in Section 11. Another way of trying to get a sense of how much confidence to place on the multiplier results is to examine their sensitivity to alternative specifications.2 A Monetary-Policy Experiment: RS Decrease The most straightforward way to examine the effects of monetary policy in the model is to drop the interest rate reaction function (equation 30). After four quarters the rise in GDP R is .6 regarding the use of the rational expectations assumption. The federal government deficit is $5.3.7 that the model suggests that the effect on output of a decrease in RS would be slightly larger if the federal government debt had not grown so much during the 1980s.1 billion lower after eight quarters.3. After two years the percentage rise in GDP R is thus about half of the percentage point decrease in RS. and then compute multipliers for a change in RS. which is in part because of the lower government interest payments and in part because of higher tax receipts caused by the more expansionary economy. where the alternative specifications are supported by the data as much or nearly as much as the original specification. The results in Table 11. and after eight quarters the rise is . take the bill rate RS to be exogenous. and the simulation period is the same as that for the fiscal policy experiments: 1989:3–1993:2. Remember from earlier discussion that a change in RS has both a substitution effect and an income effect on the economy.24 percent. The units in Table 11. . The experiment is a sustained decrease in RS of one percentage point. but the income effect is negative.2.2 are the same as those in Table 11.4 below regarding the specification of the import equation and in Section 11. 11. The substitution effect from a decrease in the bill rate is positive.1. This is done in Section 11.292 11 PROPERTIES OF THE US MODEL into account in the estimation of the standard errors of the multipliers. It is an open question as to how this can be done. the present estimates of the standard errors must be interpreted as only lower bounds. Given that it was not done here. The results of doing this are reported in Table 11.2 show that the overall effect on real GDP from the one percentage point decrease in RS is moderate.

The results from three experiments are reported in the table. that neither variable was significant.3. RS.1 used the interest rate reaction function as the monetary-policy assumption. U R.3.3 presents results from making two other assumptions. however. 11. According to the interest rate reaction function. say. the monetary-policy behavior that is reflected in the estimated interest rate reaction function is less accommodating than the behavior of keeping M1 unchanged in the wake of an increase in government spending. In other words. This is because. eight quarters the output differences across the three experiments are modest.3 MULTIPLIERS AND THEIR STANDARD ERRORS 293 11. namely an increase in COG with the interest rate reaction function used. along with the results for GDP R. The results for M1 are also presented in Table 11.7 that the level of nonfarm firm sales dominated disposable income in the equation in the sense of having a higher t-statistic when both variables were included in the equation. (See the results for M1 for the first experiment in Table 11. P F . which results in the first experiment being less expansionary than the third. which leads output to increase less for these two experiments than for the second experiment where the bill rate is kept unchanged. The bill rate rises more in the first experiment than in the third. which is used for the first experiment.3). In both the first and third experiments the bill rate rises in response to the COG increase. the Fed leans against a COG increase by actually having the money supply contract. and the third experiment is the change in COG with M1 held unchanged from its base period values.4 Sensitivity of Fiscal Policy Effects to the Specification of the Import Equation It was seen in the discussion of the import equation in Section 5. When the money supply is constrained to be unchanged in the third experiment. and −SGP . and Table 11.3 Sensitivity of Fiscal Policy Effects to Assumptions about Monetary Policy All the experiments in Table 11. the lean in terms of higher interest rates is thus not as great. It is possible to make other assumptions. the effects of a change in RS on output are modest. The second experiment is the change in COG with RS held unchanged from its base period values. The first experiment is the same as the first experiment in Table 11.11. After eight quarters the increase in RS is 1. as seen above. (Disposable income was chosen for the final specification of the equation . Collinearity was such.12 percentage points in the first experiment and .40 percentage points in the third experiment. After.1.3.

22 .40 1.14 1.12 .33 1.61 .85 1.31 .05 because this is consistent with the use of disposable income in the household expenditure equations.32 -.11 .44 .21 -1.47 .89 -1.45 .28 .32 .91 -1.63 -.00 .11 .07 1.19 .10 1.00 2 Number of Quarters Ahead 3 4 8 12 Real GDP 1.00 . the less confidence can be placed on them .34 -.49 .65 .00 .04 .26 1.09 1.41 .23 .15 1.84 1.77 1.19 COG ↑RSex.07 1.65 1. GDP R: 1.30 1.11 .07 1.05 .64 -.82 .84 1.24 1.91 -1.01 .00 .50 .65 -.82 .00 .62 1.18 .11 .87 .46 -.64 1.49 Unemployment Rate -. U R: -.12 .17 -.11 .09 .11 .20 .28 .07 .17 .61 .22 .12 .89 -.19 -.87 .77 1.80 .12 .00 .35 1.30 .30 COG ↑RSex.07 .82 1.40 .07 .75 -.93 -1.11 .89 .11 . COG ↑M1ex.11 .21 -.30 .63 .44 -.99 1.04 -1.29 1.76 .11 .88 .90 .45 .10 .04 .) This is thus a case in which the data do not discriminate well between two possible variables.21 16 .94 1.11 .05 -.04 -.04 .84 .75 1.33 1.19 -.18 -1.50 .00 .37 -.00 .30 COG ↑RSex. The more sensitive the properties are.95 -1.09 1.50 1.53 .02 .30 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 0 a b 1.11 -.89 .18 1.10 1.54 .29 .92 .50 .13 .50 -.32 -.66 -.67 -.15 -.65 -.294 11 PROPERTIES OF THE US MODEL Table 11.17 1.10 .65 .32 -.21 -.67 -. and it is of interest to see how sensitive the properties of the model are to the use of the two variables.00 .16 -.15 .33 .31 1.13 1.21 1.15 COG ↑Eq.32 .67 -.90 1.14 -.11 .11 -1.87 .08 . COG ↑M1ex.92 .99 .11 1.11 .91 .90 1.00 1.28 . P F : Price Deflator .07 .46 .86 1.09 COG ↑Eq.3 Estimated Multipliers and Their Standard Errors for an Increase in COG under Three Monetary Policy Assumptions 1 COG ↑Eq.77 .10 . COG ↑M1ex.28 1.96 -1.64 1.65 .36 .46 .32 .11 .50 -.

00 . COG ↑M1ex.18 .3 a 9.79 .3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.00 .00 .0 6.13 -1.54 .00 .00 . The results for the first version are the same as those in Table 11.6 1.33 .95 .3 6. See notes to Table 11.6 8.1 7.00 295 COG ↑RSex.43 -.9 0 9.5 8.00 .12 1. The COG. and the results for the second version are for the sales variable .11.42 -.6 8.9 6.00 .30 0 a b 0 a b 0 a b .04 .20 .57 -.00 .11 . The results are reported in Table 11.8 8.2 a 9.9 11.8 8.31 . Eq.3 .10 .00 .21 .08 .05 2 Number of Quarters Ahead 3 4 8 12 .00 COG ↑Eq.00 .62 .19 .69 .8 8.0 1. = M1 exogenous.62 .8 8. Otherwise.3 .0 8.00 .00 .5 -1.00 .8 10.38 . and T RGH experiments in Table 11.00 .00 .5 1.6 12.00 .00 -1.00 .08 .26 .00 . −SGP : Federal Government Deficit 0 9.4 8.16 .00 .4 6.09 -.14 .1 12.12 1.38 .1.30 COG ↑RSex.1 7.7 7.6 .1 were performed with the sales variable replacing the disposable income variable in the import equation. RS: Bill Rate .26 .3 7.20 .11 .9 6.3 .04 . COG ↑M1ex.01 .50 -1.5 12.35 .9 b .11 .39 .25 -.13 15.00 .43 .39 .3 8.00 M1: Money Supply -.00 .15 .16 .9 17.00 .9 0 9.03 . M1ex.09 .2 .09 .4 .08 .00 .3 6.00 .00 .00 16 1.90 .19 .6 6.25 -.00 .3 .41 .16 17.84 .1 -1.5 .0 7.5 1.3 .00 .40 .94 .15 1.11 .4 .32 .9 a 9.17 .25 . COG ↑M1ex.4 for GDP R.0 1.00 .21 .26 .79 .8 9.0 12.30 COG ↑RSex.26 .03 .09 .00 .3 7.9 8.30 = Equation 30 used.00 .1 b .= RS exogenous.43 . RSex.3 5.00 .00 .7 1.00 .3 (continued) 1 COG ↑Eq.60 .67 -1.90 . D1G.95 .00 .15 1.9 b .9 0 a b 0 a b 0 a b -.08 .93 .10 .00 .05 .14 . because the data do not discriminate between the two variables.14 1.7 15.00 .6 .1. everything else was the same.10 COG ↑Eq.43 .56 .2 .00 .82 .

24. The reason for this is the following.29.296 11 PROPERTIES OF THE US MODEL Table 11. .21.21 .58 16 .73 .68 .72 .69 versus .59 1 = Income variable in the import equation (regular version).96 for the first.90 1.96 1.37 . Results are from deterministic simulations. whereas disposable income only indirectly increases (as income expands due to the expanding economy).4 Estimated Multipliers for Three Experiments and Two Versions of the Import Equation 1 COG ↑ D1G ↓ T RGH ↓ 1 2 2 2 2 2 1.4 are the following for the three experiments: 1.69 . replacing the disposable income variable in the import equation.25.29 . In the other two experiments in Table 11. imports respond more quickly in the second version because they are directly affected by sales.69 . the level of sales directly increases. whereas sales only indirectly increase (as the disposable income increase induces an increase in sales).4 the output response is greater in the second version than in the first. disposable income directly increases. The differences are thus slightly less than one standard .33 .93 for the third.77 1. 2 = Sales variable in the import equation.62 1.37 Number of Quarters Ahead 2 3 4 8 12 GDP R: 1.77 Real GDP 1. The eight quarter ahead multipliers in Table 11.63 .20 compare to the eight quarter ahead estimated standard errors in Table 11.4. Therefore. .87 .00 1. and . . The reason for this is the following.63 . imports respond more quickly in the first version.1 of .93 . When D1G or T RGH increase.89 .90 . When COG increases. Therefore.56 . In the first version they respond only as disposable income responds. These differences of .16 1.34 .62 1.91 .21 versus . Since imports respond more in the second version.73 versus . and .36 . and . Only the results from deterministic simulations are presented in Table 11.75 .83 . the output response is less because imports subtract from GDP. contrary to the case in the first experiment.08 . respectively.33 .23.96 .11 1.32 .75 1. resulting in a smaller output increase.62 .03 1.53 .90 for the second.06 . The results for the COG experiment show that the output multipliers are larger for the first version than for the second.

13 .4 were to be taken as estimates of multiplier uncertainty due to possible misspecification of the import equation.11 .32 1.29 297 GDP R: Real GDP 1. This is.06 1.66 -.11.95 1.43 1. If quite different multipliers are obtained for different sets of con- .02 .28 1.77 1.91 1.56 1.13 U R: Unemployment Rate -.00 .87 1.63 1.59 -.45 1.36 1.30 .98 .55 .26 .79 .82 -.80 .16 -1.16 .83 . only a very crude way of trying to estimate multiplier uncertainty due to misspecification.07 -.69 -.19 .66 -.81 .50 -.38 1.88 1.32 .00 .61 -.81 .07 .76 1.44 .84 .10 1.54 -. the total multiplier uncertainty would be about double this.90 .68 -.11 .33 -.06 .39 .17 .82 .93 -1. 11.26 .21 1.90 .07 .72 .43 .11 .61 1.27 1.00 .91 error.00 .82 .11 1. but it may be suggestive of the likely size of this uncertainty.36 -.79 RS: Bill Rate .96 .80 1.96 -1.5 Sensitivity of the Multipliers to Alternative Coefficient Estimates It is straightforward to compute multipliers for different sets of coefficient estimates.09 1.50 16 .75 1.97 -.28 .28 .62 1.04 .81 1.11 -.93 -1.35 .68 . of course. if the differences in Table 11.32 2 Number of Quarters Ahead 3 4 8 12 1.80 .5 Estimated Multipliers for a COG Increase for Alternative Sets of Coefficient Estimates 1 2SLS 2SLAD 3SLS FIML MU 2SLS 2SLAD 3SLS FIML MU 2SLS 2SLAD 3SLS FIML MU 2SLS 2SLAD 3SLS FIML MU 1.34 -.67 1.91 1.82 .43 1.72 -.07 .18 1.11 1.12 1.26 -.3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.86 .05 .19 1.98 .41 1.82 -.43 .14 1.62 1.53 .94 .10 .52 -.11 -.89 1.00 .30 -.43 -.03 1.41 -.89 -.10 -.15 .32 -.35 .12 1.46 -.3.35 .02 .11 -.65 -. Therefore.75 P F : Nonfarm Price Deflator .81 1.53 -.85 1.87 1.34 1.44 .

Multipliers are presented in Table 11. After the second quarter. and when the economy . P I M. It was increased by 10 percent in each of the quarters of the simulation period (from its base period values). 11. The results show that an increase in P I M is contractionary and inflationary. The multipliers are for the COG experiment with the interest rate reaction function used. the differences in Table 11. and MU.5 are probably the least trustworthy. 2SLAD.5 for the five sets of coefficient estimates that have been obtained for the US model—2SLS.7 The Deficit Response to Spending and Tax Changes When the economy expands.1. which is the output effect dominating the price effect in the interest rate reaction function (equation 30).6 are the same as those in Table 11. the overall FIML results stand out somewhat from the rest. tax revenues increase and some government expenditures like unemployment benefits decrease. 3SLS. which are less than half the size of the others. This in turn leads to a fall in household expenditures. however.77 percent higher in response to the 10 percent increase in import prices. The largest difference is for the FIML multipliers for the bill rate.5. This closeness of the results complements the closeness of the results in Section 8. These results are based on deterministic simulations. this may be a cause of concern since one does not expect this to be true in a correctly specified model. The results are presented in Table 11. After eight quarters output is 1.3. FIML.6. which is another reason for the fall in expenditures.1: 1989:3–1993:2. The units in Table 11. the FIML results in Table 11.6 Multipliers from a Price Shock: P I M Increase The next experiment examined is an increase in the import price deflator.3.5 are not large enough to have much economic significance. The Fed responds to the initial change in prices by increasing the bill rate. It can be seen that the multipliers are quite close across the different estimates. The simulation period was the same as in Table 11. This experiment is the best example in the model of a situation in which real output and the price level are negatively correlated. the domestic price level increases (equation 10). the bill rate is lower. say 2SLS versus 3SLS. which leads to a fall in real disposable income. As in Section 8.27 percent lower and the price level is 1.5 regarding the predictive accuracy of the model. Otherwise. When P I M increases. 11. Given that the FIML forecasts are on average not as accurate as the others in Table 8.298 11 PROPERTIES OF THE US MODEL sistent estimates.4.

22 -.24 -1.18 . tax revenues decrease and some government expenditures increase.56 .35 .9 1.02 .6 299 GDP R: Real GDP -. is the fact that the Fed may lower interest rates in contractions and raise them in expansions. interest payments of the government fall.0 1.65 .08 .12 -1.37 .02 .94 .25 .33 . the government deficit decreases when the economy expands and increases when the economy contracts. which increases the deficit. if one is contemplating lowering the deficit by decreasing government spending or raising taxes.2 1.51 .20 .79 1.23 -.7 .2 .05 .9 -6.10 .1.1 .26 -.90 -.18 .41 2.3 2. contracts.07 -.43 2. then.42 .45 -.2 .85 -. which will presumably affect the economy.6 Estimated Multipliers and Their Standard Errors for an Increase in P I M of 10 percent 1 PIM ↑ 0 a b 0 a b 0 a b 0 a b 0 a b -. that to lower the deficit by $10 billion takes more than a $10 billion cut in government spending.2 1.7 1.34 -.41 -6.29 PIM ↑ PIM ↑ −SGP : Federal Government Deficit 1.38 1.34 .77 -1.23 -.48 .59 -.20 .03 .35 -3.48 .34 . On this score.9 2.06 .77 .33 .75 -1.5 .3 .35 .93 2.3 MULTIPLIERS AND THEIR STANDARD ERRORS Table 11.9 2. for example.07 -.3 16 -.8 1.11 PIM ↑ PIM ↑ U R: Unemployment Rate .15 2 Number of Quarters Ahead 3 4 8 12 -1.10 -1. In particular. It may be.1 See notes to Table 11. and as interest rates rise.27 . It is of obvious interest to policy makers to know how the deficit responds to changing economic conditions.89 1.07 .09 .63 .3 2. it is important to know how the changes in the economy will affect the deficit.90 1.45 .27 .90 . As interest rates fall.24 .20 .14 .87 .22 RS: Bill Rate .54 . which decreases the deficit.19 -.14 .64 .35 .06 .17 .26 P F : Price Deflator .40 2. .10 .11.33 -.23 . interest payments rise. on the other hand.51 -.23 -.7 -3. Working in the opposite direction.05 -.06 .30 -.

The estimates in this section were obtained as follows. If there were no response of the economy to the fiscal policy changes. the Fed keeps the bill rate unchanged. D1G was increased each quarter so as to make the nominal increase in personal income taxes (T H G from equation 47) $10 billion. Of the three assumptions examined in Section 11.3 are close to being symmetrical for positive and negative fiscal policy changes. P G is an endogenous variable in the model. For the third fiscal policy change. of course.3. Three fiscal policy changes were then made for two monetary-policy assumptions. and the purpose of this section is to provide such estimates. For the first fiscal policy change. would be −$10 billion in each case. (The results in Table 11. −SGP . A similar situation holds for D1G in the second fiscal policy change. 1989:3–1993:2. 4 Since . The two monetary-policy assumptions used were equation 30 and the policy of keeping the bill rate unchanged. Although not shown in the table. Also. the bill rate falls the most when equation 30 is used. These estimates will.3—the Fed behaves according to the interest rate reaction function (equation 30).300 11 PROPERTIES OF THE US MODEL It is easy in a model like the US model to estimate how much the deficit changes as government spending or taxes change. Its value each quarter is whatever is needed to make P G · COG $10 billion less than its base value.4 For the second fiscal policy change. the residuals from the estimation of the stochastic equations were first added to the stochastic equations. which results in a perfect tracking solution for the model when the actual values of the exogenous variables are used. The key question then is how much the changes in −SGP deviate from −$10 billion. The simulation period was. T RGH . resulting in six experiments. COG was decreased each quarter so as to make the nominal decrease in government spending (P G · COG) $10 billion. The results for the six experiments are presented in Table 11. The results in Table 11. COG is in effect an endogenous variable in this experiment (with P G · COG being exogenous).3 show that the bill rate rises the most (in response to the government spending increase) when equation 30 is used. depend on what is assumed about monetary policy. and the Fed keeps the money supply unchanged—the one that mitigates the effects of a government spending change or a tax change the most is the use of the reaction function. which is in nominal terms.7. when there is a government spending decrease. as for the multiplier experiments above. the change in the federal government deficit. since fiscal policy effects in the model depend on the monetary-policy assumption. as above.) The least mitigating assumption is when the bill rate is kept unchanged in response to the fiscal policy change. was decreased by $10 billion each quarter.

9 -4. The D1G results are presented next in Table 11.1 that changes in D1G have smaller impacts on GDP than do changes in COG. The decrease in GDP was.2 -6.3 MULTIPLIERS AND THEIR STANDARD ERRORS 301 Table 11. In this case the decreases in the deficit never fall below $6. Even with the bill rate held constant. Consider the COG results first.5 -9. It is known from Table 11.7 -6. the fall in the deficit is $7.7 billion in the second quarter. This is because of the lower interest rates. The decreases in the deficit are larger for the D1G increase than for the COG decrease.11.30 RS ex.7 -4.2 -11. and then rises to a little over $6 billion in the fourth year.9 -8.5 -9.0 -10. government interest payments are noticeably lower in this case after about two years even with the bill rate held .8 percentage points with equation 30 used. The results are much different if the Fed behaves according to equation 30.5 -8. Units are billions of current dollars.1 -6.4 billion.6 -8. the decrease in the deficit is greater than $10 billion after 12 quarters.8 -9. -9.5 -7.8 -9.5 -8.4 -6.6 -5. When the bill rate is unchanged.5 billion. also less in this case because of the stimulus from the lower interest rates.4 billion in the fifth quarter.7 Estimated Effects on the Federal Government Deficit of Six Fiscal Policy Experiments 1 COG ↓ D1G ↑ T RGH ↓ 2 Number of Quarters Ahead 3 4 8 12 -11. of course. = RS exogenous. RSex. The fall reaches a low of $4.1 -13. and the results in Table 11.0 -8.3 -9. Results are from deterministic simulations.30 = Equation 30 used.7.3 Eq.30 RS ex.5 -9.5 -9.7 -12.0 Eq. and by the twelfth quarter the decreases are greater than the $10 billion fall in spending.4 -7. By the end of the period the decrease in federal government interest payments (I NT G) was $6.9 -9.5 -11. -7.7 -9.7 −SGP : Federal Government Deficit Eq. the fall in RS after a few quarters was a little over . The government thus loses about 40 cents of each one dollar cut in spending in terms of the impact on the budget if the Fed responds by keeping interest rates unchanged.2 -5.5 -9.0 RS ex.4 -9. Although not shown in the table.0 -10.4 -8.8 -6. Eq.2 billion in the first quarter and $5. Although not shown in the table.1 -8.30 -7.2 -14.2 16 -12.3 -10.7 reflect this.

For example. of course. never rule out the possibility that the Fed would behave in a more expansionary way in response to some deficit reduction plan than would be implied by its historical behavior. then a policy of reducing the deficit by contractionary fiscal policies will be much more successful than if the Fed does not allow the interest rate to fall.0 billion by the end of the period. The deficit reductions are. in all cases for the experiments in Table 11. by the sixteenth quarter interest payments were $2. The results also show that government tax and transfer changes are better tools than government spending changes on goods for lowering the deficit because they have smaller impacts on output.1 that the impact of a change in T RGH on GDP is generally in between the impacts for the COG and D1G changes. but close to the D1G results. even larger when equation 30 is used. and so deficit reduction is not without some costs even for the most optimistic case in the table. These payments change when interest rates change and when the government debt changes due to current and past deficit changes. To conclude.302 11 PROPERTIES OF THE US MODEL constant because the government debt is lower due to smaller past deficits. although closer to the impacts for the D1G change. the present results show that the Fed plays a large role in deficit reduction issues.7 are for the T RGH decrease. One can. However. The change in government interest payments caused by various policy changes is now a non trivial part of the overall change in the government deficit. the Fed would have to behave in a more expansionary way than that implied by equation 30. . Again.7 that use equation 30 as being based on historical Fed behavior. It can be seen from Table 11. This lower level of interest payments is the primary reason for the deficit reductions greater than $10 billion. The results are in between the COG and D1G results.7 reflect this. the results in Table 11. In order to make the output costs zero. although not shown in the table. If the Fed leans against the wind as equation 30 specifies. The final results in Table 11. In this case the table shows that the fall in the deficit is $14.7 the effects on output were negative. One should thus think about the results in Table 11. of course.4 billion lower.

and the second total is the sum of the nine individual values. The Results for Real GDP The results in Table 11. 3) supply shocks. and this section uses this procedure to examine the sources of economic fluctuations in the US+ model. imports (I M). but it is useful for organizing the discussion. 2) financial shocks. and 5) shocks from the interest rate reaction function. Each number in Table 11. For this model the covariance matrix of the error terms is 121×121. Only error terms were drawn for the stochastic simulations (not also coefficients).9. and exports (EX). and the results for the price deflator are presented in Table 11. and it is taken to be block diagonal. H F . ˜2 The results in Table 11. This grouping is somewhat arbitrary. labor demand (J F . each number ˜ is 100[δit (k)/σit ].8. The results for real GDP are presented in Table 11.8 is the value computed from the tenth simulation.4 Sources of Economic Fluctuations in the US Model5 Section 10. H O). In addition. 4) fiscal shocks. The first simulation was one in which none of the equations’ error terms was fixed.3 discussed a procedure for examining the sources of economic fluctuations in macroeconometric models. Each of the other 29 simulations consisted of fixing one or more of the error terms in the 121 equations. which can be interpreted as monetarypolicy shocks. Nine demand shocks were analyzed: three types of consumption (CS. The first block is the 30×30 covariance matrix of the structural error terms. one equation’s error term was fixed except for the simulation regarding labor demand. Remember that the US+ model is the US model with the addition of 91 autoregressive equations for the exogenous variables. Consider the demand shocks first. three types of investment (I H H . The first total presented for the demand shocks in Table 11. and the second block is the 91×91 covariance matrix of the exogenous variable error terms. I V F . a tenth simulation was run in which the error terms in all eleven equations were fixed. For each of the nine simulations.8 are divided into five categories: 1) demand shocks. where three equations’ error terms were fixed. CD).11. . CN . In terms of the notation in Section 10. I KF ).8 are based on 30 stochastic simulations of 1000 repetitions each.3. The difference between 5 The material in this section is an updated version of the material in Section IV in Fair (1988a).8 is the difference between the two variances as a percent of the overall variance (in percentage points).4 SOURCES OF FLUCTUATIONS IN THE US MODEL 303 11.

7 39.0 .7 5.0 .8 are financial shocks. depending on which total is used.8 2.9 .3 4.0 .4 3.9 .3 73.4 68.3 -1. H F .9 2.1 .2 1.9 37.0 2.5 1. If each of the eleven error terms were uncorrelated with all the other error terms in the model. the two totals would be the same.304 11 PROPERTIES OF THE US MODEL Table 11.9 -.0 .6 1.7 -.1 percent of the variance of GDP for the first quarter.6 57.2 57.3 4.5 5.5 3.0 .0 35.1 1.5 1. The results in the table show that the demand shocks account for between 75.5 .2 4.7 4.0 2.5 38.3 1.8 Variance Decomposition for Real GDP 1 Demand Shocks: CS CN CD IHH I V F (eq.7 51.8 5. however.1 .0 . It simply means that the effects of the shocks to the plant and equipment investment equation are relatively small.8 2.0 .6 .8 5.0 1.2 10.6 6.1 .3 1.4 6.0 . The household sector’s variables—CS.1 75. The contribution declines to between 51. that a result like the one that plant and equipment investment (I KF ) shocks have a small effect on the variance of GDP does not mean that plant and equipment investment is unimportant in the model.3 1.8 .8 31.6 . MF. CU R RB.7 the two totals is an indication of how much the correlation of the error terms across equations matters.4 1.9 65.0 .9 4.3 .5 1.6 4.6 -1.6 .5 1.1 .9 .5 64.0 1.9 1. RM CG Totala Totalb 2 Number of Quarters Ahead 3 4 5 6 7 8 GDP R: Real GDP 5. Three .9 81.0 4.8 1.7 1.0 4.0 1.2 65.0 1. 11) I KF J F.3 -.5 1. I KF .7 percent for the eighth quarter.6 4.2 60. The next type of shocks presented in Table 11.6 61.5 1.8 6.1 .1 1.0 1.1 .2 8.0 .8 55. CN .4 2.9 .2 1. J F .6 .8 1.0 3. and I H H —contribute more than do the firm sector’s variables— I V F .6 6. Remember.2 2.1 1. with import shocks the next most important.5 65.0 1.3 1.9 37.8 1.0 and 53.9 53.1 1.6 .1 6. CD.3 1.1 . H F.0 1.6 62. H O IM EX Totala Totalb Financial Shocks: MH.9 35.5 8.5 6.9 -. and H O—to the variance of GDP.4 3.2 7.0 .7 and 81.6 1. Export shocks contribute by far the most to the total.7 45.2 1.

2 .7 .5 3.1 10. H G T RGH COS S&L tax rates JS T RSH Totala Totalb .1 8.3 5.2 3.9 -.1 18.1 . shocks to long term interest rates (RB. Almost all of this contribution is from the price of imports.8 1.2 .6 .8 -1.8 .5 22.4 SOURCES OF FLUCTUATIONS IN THE US MODEL Table 11.7 .0 .3 .6 2. J M.1 -.1 18.8 -2.3 -.8 (continued) 1 Supply Shocks: PF WF PIM P OP 1.3 12.4 6.0 .8 2.9 -.9 .5 .7 . The results show that the supply shocks account for a rising proportion of the variance across the horizon.8 -.7 -. P OP 2.9 9. The effects of supply shocks are presented next. financial shocks were analyzed: shocks to money demand (MH .0 .9 .6 . MF .6 . federal .4 .7 -.7 -2.3 -.1 . Four supply shocks were analyzed: shocks to the aggregate price level (P F ).3 14.0 -. b Sum of the individual values.7 from stochastic simulation with all the relevant error terms set to zero at the same time.7 2.9 Number of Quarters Ahead 3 4 5 6 .2 6.1 .4 .5 4.2 12.0 .0 16. CU R).2 1.6 .9 20. 3 Totala Totalb Fiscal Shocks: COG Fed tax rates J G.0 -.6 .5 1. and shocks to stock prices (CG). 2.3 1.9 23.5 2.0 16.7 305 5.2 .8 10.4 -.2 2.2 4.2 -.6 3.2 11.1 20.1 4.5 . shocks to the aggregate wage rate (W F ).7 .1 1.9 15. The results show that the effects of these shocks are quite small.5 20.1 2.1 5.3 and 8.2 8 1. and shocks to population (P OP 1.2 .4 2 .4 7.7 -2.9 7 1.9 -.5 -. shocks to the price of imports (P I M).3 -.1 19.5 .5 .0 -1.0 .2 .1 .4 .5 1.6 5.4 1.7 2.0 14.2 3. RM).3 -1.2 .1 1.1 6.6 -.7 -1.2 2.4 -2.3 8.7 .6 21.8 -. reaching between 8. P OP 3).11. Eight fiscal shocks were analyzed: shocks to federal government purchases of goods (COG).9 .1 1.2 .9 -.4 -.4 13.0 -.4 2.4 -2. The effects of fiscal shocks are presented next.1 a Computed -.5 1.3 14.1 -1.0 4.1 4.9 Federal Reserve Shocks: RS .7 percent by the eighth quarter.2 11.

9 -. The overall results for real GDP thus show that demand shocks contribute the most to the variance of real GDP. state and local jobs (J S).7 23.8 -.3 6.306 11 PROPERTIES OF THE US MODEL Table 11.2 .5 14. H G).1 .8.1 5.2 .0 .0 .5 .0 . The effects of the shocks to the interest rate reaction function are presented last in Table 11. D3G.4 .3 5. The results show that the fiscal shocks are the second largest contributor to the variance of GDP.5 1.0 .0 .5 6. H F.2 11.6 3.1 .0 32.2 .8 tax rates (D1G.4 .0 1.2 -.4 .0 .1 . For the first quarter the contribution is between 13.4 -1.7 -2.2 -1. state and local government purchases (COS).0 .0 . D5G).9 and 19.1 .9 -.0 . federal transfer payments to persons (T RGH ).4 3.5 .0 .2 -.9 9.3 . D2G.4 .3 12.0 -2.5 3. and for the eighth quarter the contribution is between 14.3 1. MF.2 . RM CG Totala Totalb 2 Number of Quarters Ahead 3 4 5 6 7 8 P F : Price Deflator .0 .8 -. D4G.6 -. The largest contributor to the effects of the fiscal shocks is T RSH .4 percent.0 .0 1.1 1.9 -1. D2G.9 -2.0 . Supply shocks are of growing importance over the horizon. with fiscal shocks contributing the next most.9 -.3 -1.3 .0 .8 4.4 .7 8.4 .0 -.0 . H O IM EX Totala Totalb Financial Shocks: MH.1 .8 -1. federal jobs and hours (J G.0 -.0 .0 2.5 1.2 -.1 8.0 .3 and 14.0 2.0 .3 -1.3 31.8 -.5 17.6 1.2 -2.5 3.1 1.1 -.3 -1.1 .9 Variance Decomposition for the Nonfarm Price Deflator 1 Demand Shocks: CS CN CD IHH I V F (eq.8 3. CU R RB.6 . The tax rates and labor variables contribute very little.0 percent. J M.0 23.0 -.8 -1. but still account for less than 10 percent of the variance after eight quarters.4 -2.1 -.4 .1 -.0 .9 -.2 -1.0 .0 .1 .1 .8 1.0 .5 1.3 .0 .0 .2 -. The effects of financial shocks and shocks to the interest rate reaction function are .0 .3 -.6 2.7 19.8 47.0 .5 -1. and state and local transfer payments to persons (T RSH ). 11) I KF J F.0 .3 -.7 24.0 .2 . state and local tax rates (D1G. D3G). The results show that these effects are very small.0 .

0 .4 2.0 .9 30.1 .4 .5 8.4 and 21.6 The Results for the Price Deflator The results for the private nonfarm price deflator are presented in Table 11.1 -.0 100.3 .0 97.7 .7 -2.1 .4 88.8 5.9 -1.7 12.4 4. The results show that most of the variability for the first few quarters is due to shocks to the price equation.1 -5. They are based on the same stochastic simulations as those used for the GDP results.1 . 6 This .0 .1 .4 8.0 .0 . J M.3 307 .9 (continued) 1 Supply Shocks: PF WF PIM P OP 1.2 1.1 .0 . and fiscal shocks account for between 7.2 .1 .4 Federal Reserve Shocks: RS .2 .1 .4 1.3 77. 3 Totala Totalb Fiscal Shocks: COG Fed tax rates J G.8 7.6 2.0 .1 -.5 7.3 68.2 93.7 -.2 60.2 .0 a Computed . In quarter 8 demand shocks account for between 31.8 23.0 -.4 .8 and 47.1 6.0 .1 .8 2.6 38.5 .9 .0 .5 -.4 14.8 1.2 .1 81.3 Number of Quarters Ahead 3 4 5 6 81.5 -2.1 .0 96.1 .7 2.0 99. but after about four quarters other shocks begin to matter.4 -1.0 -4.0 .4 1.0 .4 SOURCES OF FLUCTUATIONS IN THE US MODEL Table 11.5 8 20.7 -1.0 .5 1.1 -.6 -3.2 1.4 -.1 -2.4 3.0 .0 66.5 57. 2.3 38.5 -4.3 percent of the variance.8 -1.6 38.0 -.1 36.0 -3.0 .9 3.8 percent. H G T RGH COS S&L tax rates JS T RSH Totala Totalb 97.2 -.8 15.5 -3.1 .0 .1 13.3 -2.1 .0 . b Sum of the individual values.0 2 91.1 .6 73.1 -.1 7 27.1 .3 -7.0 1.5 -5.6 -.9 -3.11.0 .4 50. very small.1 73.0 .0 .9 from stochastic simulation with all the relevant error terms set to zero at the same time.1 .9.2 21.5 2.6 -1.1 -1.1 -.1 -.8 -5.0 98.3 61.2 -5.0 91. general conclusion is the same as the one reached from the results in Table I in Fair (1988a) based on earlier data and estimates.

and 8. the general conclusion from both tables is the same. There are thus large shocks to the import price deflator in the stochastic simulations. contributing 38.7 and 51. The second is that the autoregressive import price equation has a fairly large variance. Otherwise. the eight quarter ahead totals for the demand shocks are 31. The first is that the import price deflator has a large effect on the domestic price level in the domestic price equation. the eight quarter ahead totals for the demand shocks are 53. and results in Table II in Fair (1988a) for the price deflator attributed less to fiscal shocks and more to supply shocks for quarters 6. For example. At the same time the estimated effects of the individual components should be discounted somewhat because their sum differs so much from the other total. Using a four equation model.9. shocks to the price of imports grow in importance over time. supply. The “non summation” totals for the demand and fiscal shocks should probably be used in Table 11.2 percent after 8 quarters. on the other hand. The two totals for the demand shocks and fiscal shocks in Table 11.9. fairly close. 8 BW actually examine the variance of GNP. since these at least take into account the correlation of the error terms within groups.8 and 47. 7. BW provide estimates of the percent of the variance of GDP8 due to four shocks: demand.8 are close to each other.0. For example. but for ease of exposition GDP will be used in the present discussion. The two totals for the supply shocks in Table 11. 7. are noticeably different for quarters 6. which have a large impact on the variance of the GDP deflator through the price equation.9.3.308 11 PROPERTIES OF THE US MODEL Within the category of supply shocks. In this sense the correlation of the error terms across equations is not a problem. however. not GDP. money supply. There are two reasons for the importance of the shocks to the import price deflator. and none of the major conclusions from the results depend on which total is used. Comparison with Other Results The present results can be compared to those of Blanchard and Watson (1986) (BW).9 are.7 The Effects of the Error Term Correlation Across Equations The two totals for each type of shock in Table 11. which are quite close. For these quarters the totals based on summing the individual values are greater than the other totals for the demand shocks and smaller for the fiscal shocks. and so the estimated effects of the individual supply components are probably more trustworthy. 7 Earlier . and 8 than the results in Table 11.

These differences are fairly small. The fiscal shock comparisons are 19.8 (demand shocks plus financial shocks).9 = 65. Table 5.3.8 for the one quarter ahead forecast and 15. Fiscal shocks account for 12. BW attribute about three fourths of the variance to supply shocks and about one fourth to demand shocks for one quarter ahead.0 in Table 11.0 versus 14. 74. Table 2.8. The relevant number in Table 11.8. 133.9 Their demand shocks are probably closest to the first two categories of shocks in Table 11.7 percent in Table 11.8 and 11. compared to 8. The present results thus attribute more of the variance to supply shocks. in this case the shocks to the domestic price equation versus shocks to the import price deflator.4 percent.1 percent.4 for four quarters ahead.1 + 0.0 for one quarter ahead and 16.9 for each category of shock will be used. Remember.9 + 0.7 for the four quarter ahead forecast.8 is 81.11.9 masks important individual differences. however. which compares to 64.0 versus −0.4 SOURCES OF FLUCTUATIONS IN THE US MODEL 309 fiscal.3 percent and fiscal shocks 1. For the four quarter ahead forecast.3 in Table 11.7 percent. with the US+ model attributing slightly more to fiscal shocks and slightly less to supply shocks than does Bernanke’s model. The import price deflator is not a variable in the BW model. (The effects of the other shocks are minor. which compares to 53.3 percent of the variance of output is attributed to demand shocks eight quarters out. and supply shocks account for 12.0 versus 0. For the “Money-Credit” model. The main differences in these results is that four quarters out the US+ model has more contribution from the demand and fiscal shocks and less from the supply and monetary-reaction shocks.8 in Table 11. the BW estimate is 54. Finally.9 100 percent of the variance is attributed to supply shocks one quarter out and 91. The BW money supply shocks are closest to the shocks to the interest rate reaction function here.8. p. that the total for the supply shocks in Table 11. The comparisons are 4. Regarding the variance of the price deflator.8. . BW estimate that 74.0 for BW versus 0. 9 The 10 The results cited here are taken from Blanchard and Watson (1986). The supply shocks are 3.0 versus 1.4 for one quarter ahead and 16.0 for four quarters ahead.0 percent. results cited here are taken from Bernanke (1986).4 percent four quarters out.6 = 81. (For the following comparisons the first total in Tables 11.8 percent.10 53. In Table 11. compared to 19.0 versus 22.3 in Table 11.) For the one quarter ahead forecast.) The values four quarters ahead are two thirds and one third. Bernanke (1986) has employed the BW methodology to estimate a number of small models and then to provide estimates of the decomposition of the variance of output. p.0 percent of the variance of GDP is due to demand shocks. Four quarters out demand shocks account for 6.

From an examination of results like those in Table 11. Without this trick.10 except that the differences for the stock price equation fixed are larger both absolutely and relative to their standard errors in Table III than in Table 11. and the estimated standard errors of the difference values are presented in the second of the two rows. Accuracy of the Stochastic Simulations Results are presented in Table 11. The units are billions of 1987 dollars. as just noted. results in Table III in Fair (1988a) are similar to those in Table 11. the differences themselves are quite small. the standard errors are much too large for anything meaningful to be done with the difference values. The variance estimates are fairly precise.10. except perhaps for the predictions seven and eight quarters ahead. with estimated standard errors less than 5 percent of the variance estimates. the standard errors of the difference values in general seemed small enough to allow meaningful comparisons to be made. The first row in Table 11.11 Remember that these estimates are based on the trick of using the same draws for both simulations. The values of the difference are presented in the first of the two rows. The next two rows pertain to the stochastic simulation in which the error term in the export equation is fixed. There are thus no simple stories to be told about the sources of output and price variability. The same two rows are then presented for the simulation in which the error term in the stock price equation is fixed. For stock prices the differences are small.8. The results show that for exports the standard errors are around 10 percent of the difference values.10 for all the variables. It is clearly not the case that only one or two shocks dominate. at least not within the context of a macroeconometric model like the one used here. The results for the US+ model show that there are number of important contributors to the overall variance. 11 The .10 that help give one an idea of the precision of the estimates based on 1000 repetitions. although they were still fairly far from zero.310 Conclusion 11 PROPERTIES OF THE US MODEL The procedure used in this section allows one to get a good idea of the quantitative contribution of various shocks to the variance of endogenous variables like real GDP and the price deflator. although. These are the results used for real GDP in Table 11.10 presents the estimates of the variance of real GDP. For quarters two through six the standard errors are large relative to the differences. which gives a reasonable amount of precision. and the second row presents the estimated standard errors of the variance estimates.

2 (56. and the 91 exogenous variable equations form the second block.0) 559.35 (4. For present purposes the nominal 13 When 12 The . and this section is an application of this procedure.1) 7 1592. the covariance matrix of the error terms was taken to be block diagonal.5 (5. the money supply is the policy instrument.7 371.2) Error Term in the Export Equation Fixed 152.003) 2 387.60) Units are billions of 1987 dollars.1) -10. namely the same draws of the error terms were used for the computations of both σit (r) and σit (M).4 discussed a procedure for comparing the use of different monetary-policy instruments.0) (24.3 259. This question does not arise in Poole’s analysis because the price level is exogenous.. The paths were material in this section is an updated version of the material in Fair (1988b).0) (50.0) . when the money supply (M1) is the policy variable.0) 62.07 . 11.9 (44. and so the covariance matrix of the error terms is 120×120 rather than 121×121 as in the previous section. a path for it is needed.6 (64.7) 1208.1) 3 Number of Quarters Ahead 4 5 6 964.10 Estimated Precision of the Stochastic Simulation Estimates for Real GDP 1 σ2 ` [var(σ 2 )]1/2 ` ` δ(k) ` {var[δ(k)]}1/2 ` δ(k) ` {var[δ(k)]}1/2 136. Likewise.13 -3. A similar trick was used here as was used in the previous section for the stochastic simulations.11.4) (35. The procedure requires stochastic simulation. and the number of repetitions per stochastic simulation was 1000.8 (14.4 (72.45 -2.e.8 (71. As discussed below.38) (1.028 (.8) Error Term in the Stock Price Equation Fixed . and the US+ model was used for the stochastic simulations. Estimates are based on 1000 trials. ˜2 ˜2 When the bill rate (RS) is the policy variable (i.4 (17. equation 30 is dropped from the model for the results in this section.13 a path for it is needed.7) (44.9 (55.0 500. The first 29 equations form the first block. The simulation period was 1970:1–1971:4. the question arises as to whether it is the nominal or the real money supply that is the instrument. As in the previous section.79) (3. exogenous for the stochastic simulation).11) (1.98) (2.35 -.5 455. and only error terms were drawn for the repetitions (not also coefficients).8 (54.5) -7.2) 1422.1 (6.7 (31.90) 693.5 OPTIMAL MONETARY INSTRUMENTS IN THE US MODEL 311 Table 11.5 Optimal Choice of Monetary-Policy Instruments in the US Model12 Section 10.21) 8 1688.4) 538.87 (.44 (4.

7 13.4 16. Once these paths are chosen.0 percent larger than the variance under the interest rate policy.6 2. the predicted values of the money supply were taken as the values for the money supply path.5 16.11 for selected variables in the model.8 399.6 8.2 .9 .0 Number of Quarters Ahead 3 4 5 6 19.8 219.7 1.6 4.6 12. The differences for some of the other variables in Table money supply is taken to be the policy instrument.3 2.1 8 . .2 9.2 .6 2 13.11 Percentage Difference Between the Variance Under the Money Supply Policy and the Variance Under the Interest Rate Policy 1 GDP R CS CN CD IHH I KF IV F IM CG P CGDP D UR P I EF . ˜2 ˜2 ˜2 Remember that for Poole’s loss function i is equal to real GDP.3 chosen as follows.4 10.6 21.7 7. All the simulations are done without equation 30 in the model.5 4. A dynamic simulation was first run over the eight quarter period with the error terms set to zero and the interest rate reaction function (equation 30) included in the model.0 12.3 76.0 .9 16.6 19.2 9. Likewise.2 23.8 386.3 1.4 61.9 493.1 524.4 13. although for quarters 1.2 2.8 27.5 7. and 8. where the variance under the money supply policy is 21.7 7.9 4.1 7 2.6 16.7 9.6 5.6 9.3 5.5 53.4 46.4 413.4.5 .7 -1.4 1.6 12. and so the results in Table 11.0 91.7 5.312 11 PROPERTIES OF THE US MODEL Table 11.0 4.5 13.2 4.5 38.5 3.3 10.7 1.2 2.9 7.8 . each number in the table is 100[σit (M) − σit (r)]/σit (r). In terms of the notation in Section 10.1 428. The results for real GDP show that the interest rate policy is better for all eight quarters.1 3.5 454. The predicted values of the bill rate from this simulation were then taken as the values for the interest rate path.1 1.7 14.0 10.5 86.3 .9 .6 .1 14.1 14.2 7.0 36.5 26. The percentage differences between the two variances are presented in Table 11.9 2. The largest difference is four quarters ahead.3 62.7 15.8 107.5 9.7 22. the differences are very small. equation 30 is then dropped from the model.7 30.4 9.7 -9.2 14. 7.8 .11 for real GDP are the percentage differences between the two loss function values.9 2.8 2.3 10.3 40.

0 163.1 1.5 21.0 .3 5.1 -.9 230.4 13. 22.11.4 -5.6 2.8 39.3 2.5 6.3 199. In particular. The results in Table 11.8 -3.5 .6 36.3 -.2 2.2 -2. An interesting case to consider next is one in which there are no shocks to the money equations.1 4.11.6 94.2 5.0 2. If in Poole’s model there are no shocks to the LM function.4 . and it is of interest to see if something similar holds for the US model.7 -.8 7.8 3.9 2.0 9.4 .2 43.1 3. which is as expected.3 Number of Quarters Ahead 3 4 5 6 9. and 26) to zero across all repetitions and running the stochastic simulations again.2 2 6.0 -6.7 .2 .0 -1.1 .7 .5 .6 1.7 -5.5 194.6 11.11. This can be done by setting the error terms in the four money equations (equations 9.7 1.9 8 -8.3 1. which means that the variances for these three variables are considerably larger under the money supply policy than under the interest rate policy.4 2.1 4.8 2.2 8. .7 9.1 .3 .12 Percentage Difference Between the Variance Under the Money Supply Policy and the Variance Under the Interest Rate Policy: No Shocks to the Money Equations 1 GDP R CS CN CD IHH I KF IV F IM CG P CGDP D UR P I EF .0 2.0 -2. the money supply policy is better.7 173.11 are quite large.6 13.8 . 17.3 -1.2 8.2 7 -5.12 for real GDP and its components.4 20.6 -3.8 8. I H H .5 .9 32. the differences for CS.5 6.3 -6. The results of doing this are presented in Table 11.2 -4.4 6.6 3.9 7.3 42.3 1.9 . and CG are large and positive. The overall change in results is thus what one would expect from Poole’s analysis: the money supply policy does better relative to the interest rate policy when there are no shocks to the money equations. For quarters 7 and 8 the variances of real GDP under the money supply policy are smaller.2 5.9 2.12 are more favorable for the money supply policy than are the results in Table 11.3 4.4 25.9 1.1 202.5 9.2 .7 1.8 4.2 3.2 .5 OPTIMAL MONETARY INSTRUMENTS IN THE US MODEL 313 Table 11.8 226.3 5. and for the other six quarters the variances under the money supply policy are closer to the variances under the interest rate policy than they are in Table 11.5 .4 -1.8 2.4 -1.

8 -.13 is that the optimal policy is very close to the interest rate policy.8 7 -2. Equation 10.7 . and a new stochastic simulation was run.3 -.2 3.1 .5 -.9 8 -2.6 -.5 -.4 -.0 -.7 2 .8 -.6 . The value that was chosen as the optimal value was . and the variances of GDP for the eight quarters were recorded.0 .314 11 PROPERTIES OF THE US MODEL Table 11. Another value of β was chosen. values of the money supply and the interest rate from the base path (values that do not change from repetition to repetition) and β is the parameter to be determined.4 -4. respectively.1 -.3 -.0 -2.1 -.2 .4 -3.1 .3 -.1 -.2 -.8 -2.9 -1.2 1.1) where M ∗ and r ∗ are. and the value of β that led to the smallest variances of GDP was taken to be the optimal value.0 . where the numbers are the percentage differences between the variance under the optimal policy and the variance under the interest rate policy.4 1.3 -5.5 .0 -.1 -1.6 -1.4 .8 .5 .13. The main conclusion to be drawn from the results in Table 11.0 -.13 Percentage Difference Between the Variance Under the Optimal Policy and the Variance Under the Interest Rate Policy 1 GDP R CS CN CD IHH I KF IV F IM CG P CGDP D UR P I EF .9 . A stochastic simulation of 1000 repetitions was run.0 3.9 was added to the model and a particular value of β was chosen.1 -1.1 -.1 -.1 -1.9 .7 -1.6 1.4 . The percentage .4 -.4 .2 -2.7 -6.2 -.7 -1.1 .4 -3.0 Number of Quarters Ahead 3 4 5 6 -.0 .5 -.0 .2 .0 1.2 -1.2 1.0 The Optimal Policy The optimal policy is defined here to be the policy where the Fed behaves according to the equation log M = log M ∗ + β(r − r ∗ ) (11.1 -.9 .025. The results using this value of β are presented in Table 11.0 -.7 .5 .2 -. This process was repeated for a number of values of β.3 .1 2.2 .8 1.3 -.9 .5 .7 -.8 -1.1 2.1 .8 -1.0 .0 .6 . The optimal value of β was determined as follows.9 -.5 .0 .1 -1.7 .1 4.1 4.

14 11. and so the specification of equation 4 was changed from second to first order and the specification of equation 11 was changed from third to first order.6 Sensitivity of Multipliers to the Rational Expectations Assumption in the US Model15 Section 4. Clearly. and so there is a little more support here for the interest rate policy.1–5. not much is to be gained by using the optimal policy over the interest rate policy. 15 The material in this section is an updated version of the material in Section 5 in Fair (1993b). It has no led values in it. The first version consists of the basic equations in Table 5. and 5 in Fair (1988b).30 in Chapter 5 with three modifications. namely that it is quite close to the interest rate policy.11. and Chapter 5 carried out these tests for the US model.6 MULTIPLIERS AND RATIONAL EXPECTATIONS 315 differences in the table are very small.7 percent less than under the interest rate policy. 14 The . In the first version the interest rate policy gains relative to the money supply policy.1–5. one with more interest sensitive expenditures imposed on the model and one with rational expectations in the bond market imposed on the model. This version of the model will be called Version 1. This was done because collinearity problems prevented the Leads +8 specification from being estimated under the assumption of a first order autoregressive error for equation 23. The solution program for models with rational expectations used here can only handle first order autoregressive errors. The interest rate policy does a little better in the present results than in the earlier results.11.5 discussed how the RE hypothesis can be tested within the context of a macroeconometric model. and 23. 2. The three modifications concern the treatment of serial correlation of the error terms.30 in Chapter 5 with the equations estimated for the Leads +8 results. respectively. The third modification was that the specification of equation 23 was changed from a first order autoregressive error to no autoregressive error. 11. With the exception of equations 4.12. In Fair (1988b) two versions of the US model were analyzed that were not analyzed here. 10. and 11. For example. and in the second version money supply policy gains relative to the interest rate policy. These equations have values led 1 through results in Tables 11.13 are similar to those in Tables 1. is the same for both sets of results. How much difference to the properties of a model does the addition of the led values make? Two versions of the US model are examined here. 11. the eight quarter ahead variance of real GDP under the optimal policy is only 2. The main conclusion about the optimal policy. the second version replaces the basic equations in Tables 5. This section considers how the economic significance of the RE hypothesis can be examined.

88 .63 2.50 .25 1.07 3.01 .58 .27 -.01 .59 .01 -.31 1.10).26 -.64 .67 2. unanticipated changes.19 1.74 . This is not likely to be an important exception because the led values were not significant in equation 10 (see Table 5.21 1.09 1.98 1 = Non RE Version.10 1.70 1.21 .25 -.11 2.59 1.30 1. The COG increase was 1 percent of real GDP.70 .94 – – – – – – – – 1. Equation 10 is an exception because the basic equation was used instead of the Leads +8 version. This was done to preserve the restrictions that are imposed on the coefficients in equations 10 and 16.00 2.08 1.19 .21 -. 2a = RE Version.22 -.77 .54 .41 .34 1.51 1.92 .90 -.32 .55 .19 .41 .84 .01 .59 1. Equation 11 is an exception because the order of the autoregressive error was dropped from two to one. and equation 23 is an exception because the order was dropped from one to zero. with the coefficients for each variable constrained to lie on a second degree polynomial with an end point constraint of zero.46 3.11 .80 .79 – – – – – – – – .68 1.05 1.09 -.01 .09 2.96 2a . Finally.26 . 8 times in them.93 2.87 1.00 .46 1. anticipated changes. The increase began in 1970:1.10 .57 -.03 1. equation 4 is an exception because the equation with a first order autoregressive error and Leads +8 did not have sensible coefficient estimates.30 .87 -1.24 -. The equation used in this case is the same as .70 – – – – – – – – .70 1.69 1.84 .00 -.14 Estimated Multipliers for the US Model with Rational Expectations Sustained COG Increase GDP R PF 1 2 1 2 2a 1968:1 1968:2 1968:3 1968:4 1969:1 1969:2 1969:3 1969:4 1970:1 1970:2 1970:3 1970:4 1971:1 1971:2 1971:3 1971:4 1972:1 1972:2 1972:3 1972:4 – – – – – – – – 1.98 1.59 .82 .10 1.75 .68 1.65 .61 .53 .03 -. 2 = RE Version.16 -.19 1.37 1.00 .316 11 PROPERTIES OF THE US MODEL Table 11.

72 .6 MULTIPLIERS AND RATIONAL EXPECTATIONS Table 11.26 .31 – – – – – – – – .03 -. the one for Version 1.41 .58 .53 .15 .00 . Also.73 .03 . The decrease began in 1970:1.06 .51 .02 -.54 .03 . all the equations for both versions were estimated only through 1990:4.00 -.19 1 = Non RE Version.76 .29 -. The RS decrease was 1 percentage point.23 . namely the equation with a first order autoregressive error (as opposed to second in Table 5.01 .05 .04 -.12 . .06 -.13 . and so to make both versions comparable.06 .48 .60 .11. where the first stage regressors are all lagged once. The first stage regressors are the ones listed in Table A.00 -.74 – – – – – – – – -. Any equations in Chapter 5 for which no led values were tried are the same for both versions.08 . The estimation techniques were 2SLS and Hansen’s method.24 .11 .02 .41 -. the identities are the same for both versions. 2a = RE Version. anticipated changes.03 .09 .41 .20 .36 .4) and no led values. unanticipated changes. This version of the model will be called Version 2.75 .00 .16 .77 .11 .03 .00 .00 .69 .26 .77 .55 .19 . 2 = RE Version.27 -.06 .02 .7 except for the equations with led values and a first order autoregressive error (equations 11 and 14).00 .15 .28 .10 .45 .32 .01 .65 .04 -. The estimation period for any equation with led values had to end in 1990:4 rather than 1993:2.14 (continued) Sustained RS Decrease GDP R PF 1 2 1 2 2a 1968:1 1968:2 1968:3 1968:4 1969:1 1969:2 1969:3 1969:4 1970:1 1970:2 1970:3 1970:4 1971:1 1971:2 1971:3 1971:4 1972:1 1972:2 1972:3 1972:4 – – – – – – – – -.02 -.14 -.15 .04 .02 -.71 .10 .67 .09 .53 .62 .01 .00 -.39 – – – – – – – – .59 .28 317 2a .03 .02 .07 .01 .17 .

The policy variable was then changed and the . For experiment 1 the change is unanticipated. The specification is the same (with the three exceptions noted above). The only restrictions imposed before estimation are that there are eight leads and the coefficients of the led values lie on a second degree polynomial. The results for each experiment were obtained as follows.318 11 PROPERTIES OF THE US MODEL It should be noted that because of the reestimation only through 1990:4. This shorter estimation period was. Four policy experiments were performed. The second two experiments are a sustained decrease in the bill rate (RS) beginning in 1970:1. but the estimates are not. Version 2 is solved under the assumption that expectations are rational in the Muth sense. Both the anticipated and unanticipated results are the same for Version 1 because future predicted values do not affect current predicted values—Version 1 is not forward looking in this sense. It should also be noted that Version 2 has fewer restrictions imposed on it than are imposed on most RE models. In many RE models at least some of the coefficients are chosen a priori rather than estimated. the RE version of the model in Fair (1979b) simply imposes rational expectations in the bond and stock markets without estimation. the coefficient estimates for Version 1 are not the same as the coefficient estimates in Tables 5.10. The results for real GDP and the private nonfarm price deflator are presented in Table 11. used for the Leads +8 tests in the tables. The first two are a sustained increase in federal government purchases of goods in real terms (COG) beginning in 1970:1. For the second two experiments the interest rate reaction function (equation 30) is dropped and the bill rate is taken to be exogenous. and for experiment 2 the change is anticipated as of 1968:1. For experiment 3 the change is unanticipated. and for experiment 4 the change is anticipated as of 1968:1. These two assumptions imply that agents’ expectations of the future values of the endogenous variables are equal to the model’s predictions of them. In particular.1–5. more restricted RE models to have very different properties from those of their non RE versions. Version 2 is solved using the EP solution method discussed in Section 7. The version was first solved using the actual values of all exogenous variables. These solution values are the “base” values.30. For example. and so the results for these tests in the tables are precisely the comparison of the equations of Version 1 versus those of Version 2 (with the three exceptions noted above). It is thus quite possible for Version 2 to have properties similar to those of Version 1 and yet for other.14. it is assumed that agents use the model in solving for their expectations and that their expectations of the exogenous variables are equal to the actual values. however.

as discussed above. not all signs for the led values were what one might expect. The initial changes in investment are thus negative because of the positive future output changes. Although not shown in Tables 5.14. and so this section has remained agnostic about the signs. Likewise. Consider first the results for the unanticipated government spending increase. which is very close to the $117 billion for Version 1. It is not necessarily sensible. and this effect is large enough to make the initial changes in output negative.30. The results for the price deflator for the COG increase parallel fairly closely those for output. for current investment to be a negative function of future output.1–5. of course. the sum of the output increases over the 20 quarters is $115 billion. For this case Version 2 has slightly higher multipliers than Version 1. The aim of the exercise in this section is not to test theories. Although not shown in Table 5. regardless of what their coefficient estimates might be. Before discussing the results. but the changes are noticeably larger than in the unanticipated case once the policy action is taken in 1970:1. The three quarter ahead multiplier is 2. The results are presented in Table 11. but to see how much difference the addition of the led values makes to a model’s properties. Although not shown in the table. all the equations estimated using Leads +8 in Chapter 5 were used regardless of the significance levels of the led values. Aside from the exceptions mentioned above. The difference between the predicted value of a variable from this solution and the predicted value from the first solution is the estimate of the response of the variable to the policy change.6 MULTIPLIERS AND RATIONAL EXPECTATIONS 319 version was solved again. The effects on real GDP are thus fairly similar for Versions 1 and 2 for the unanticipated case. the sum of the output increases over the 12 quarters is $117 billion for Version 1 and $139 billion for Version 2.11. the point of this section is not to worry about signs. the coefficient estimates for the future output changes are negative for the Leads +8 results. It may be that other theories would imply different signs.70 for Version 1.12. Although not shown in the table. it should be noted that the experiments were performed without concern about possible wrong signs of the coefficient estimates of the led values. but. with Version 2 being slightly more expansionary. no concern was given as to whether the led values were statistically significant or not.11 for Version 2 versus 1. For the anticipated spending increase the changes in real GDP for the first six quarters after the announcement are negative for Version 2. The reason the output increases are negative for the first few quarters after the announcement for Version 2 has to do with the investment equation 12. which is as expected since output appears in the .

5 in 1993:2.14 show fairly modest differences in the policy properties of the model from the addition of the led values. Again. It is negative because the federal government is a net debtor. −AG. that household income (Y D). One consequence of the increasing size of the government debt is that the size of the income effect of interest rate changes on demand is increasing relative to the size of the substitution effect. is drawn from both sets of results. This means.S. With the exception of three household expenditure equations. which is an increase in AH in the model. The larger is the debt. The differences between the two versions are also fairly modest for the bill rate decrease. 11.14. These results are presented in the second half of Table 11. The same general conclusion.14. where there is about a year’s delay after the change in the bill rate is implemented before positive effects on output begin to appear. is falling more over time for a given fall in interest rates because of 16 The results in Table 11. the larger is the change in interest payments of the government (I N T G) (and thus the interest receipts of those holding the debt) for a given change in interest rates.14 are similar to those in Table 2 in Fair (1993b) based on earlier data and estimates. the results in Table 11. the largest differences occur for the anticipated case. which is the federal government debt variable in the model. The main difference is that the differences between Versions 1 and 2 for the bill rate decrease are somewhat larger in Table 2 than in Table 11. 17 The material in this section is taken from Fair (1994b). Overall. which includes interest receipts. rose as a percent of GDP from 16. which means that it is multiplied by four from the variable in the model. Also.) Much of this increase in the government debt was financed by U.320 11 PROPERTIES OF THE US MODEL demand pressure variable in the price equation 10.7 Is Monetary Policy Becoming Less Effective?17 It is well known that the federal government debt as a percent of GDP has risen substantially since 1980. 18 For these calculations GDP is taken to be at an annual rate. and so one would not expect them to contribute in important ways to the properties of the model.9 in 1980:1 to 45.16 This conclusion is perhaps not surprising given the results in Chapter 5. For ease of exposition. For example. most of the led values in Chapter 5 are not significant. however. the output increases are smaller for the rest of the horizon in the anticipated versus unanticipated case. for example. households. . −AG will be referred to as the government debt.18 (Remember that AG is the value of net financial assets of the federal government.

which offsets at least some of the positive substitution effect. and it would not be clear how much of the difference in results to attribute to government debt differences as opposed to other differences. the interest rate reaction function. The US model is used to examine this question. Then various paths of D1G and RS were tried. stimulate the economy by lowering interest rates may thus be decreasing over time due to the increasing size of the income effect relative to the substitution effect.19 This results in a perfect tracking solution when the actual values of the exogenous variables are used. The procedure used here controls better for other differences. Call this economy the “alternative” economy. The problem with this procedure. . is that other things would be different as well between the two periods. but different values of D1G and RS. It turned out that a sustained 19 Adding the estimated residuals to the equations before solving the model assumes that the shocks that occurred in the actual economy also occur in the alternative economy. The model is then used to run the same monetary-policy experiment for both the actual and alternative economies. The residuals from the estimation of the stochastic equations were first added to the equations and taken as exogenous. The two economies have the same shocks. however.7 IS MONETARY POLICY BECOMING LESS EFFECTIVE? 321 the increasing holdings of government debt by households. The beginning quarter for the changes was 1980:1. A fall in income from a fall in interest rates has a negative effect on demand. and the changes were sustained through 1990:4. An alternative procedure to that followed here would simply be to run the monetary-policy experiment for an earlier period when the government debt was not as large and compare these results to those for a later period. was dropped from the model so that RS could be treated as an exogenous policy variable. The difference in results for the two economies is an estimate of how much the effectiveness has been changed as a result of the rise in the government debt. For this work equation 30. The purpose of this section is to try to estimate how large this decrease in the effectiveness of monetary policy has been since 1980. say.11. The ability of the Federal Reserve to. The model is first used to estimate what the economy would have been like between 1980 and 1990 had the federal government debt not risen so much. The Alternative Economy In creating the alternative economy the aim was to raise the personal income tax rate (D1G) to generate more tax revenue and thus lower the deficit from its historical path while at the same time lowering the bill rate (RS) to keep real GDP (GDP R) roughly unchanged from its historical path.

8 -737.5 Dif.1 -12.9 898.6 -11.4 4454.0 3049.9 3386.7 1839.5 -451.8 32.3 497.5 1357.4 -32.1 4678.2 -29.5 490.1 12.9 -18.9 46.4 682.5 3315.5 4250.8 -147.4 1196.8 7.1 3411.9 -105.4 32.4 -23.5 6.1 -157.2 -16.3 869.2 -71.9 471. 3.2 -19.1 -17.5 25.8 -43.3 38.9 −SGP Act.8 811.8 45. 446.2 -308.5 -18.6 -93.0 54.4 4407.4 524.3 3656.8 589.3 -55.2 -12.3 1471.4 16.8 28.9 2802.6 2745.2 Alt.3 2883.6 38.8 4425.0 2763.6 -14.0 4816.2 34.0 -766.0 1010.0 1414.8 5045.0 35.6 33.3 900.4 3845.0 -33.5 -5.0 4244.2 4581.5 2826.0 34.1 -253.9 8.0 -79.7 489.8 4315.9 5.8 -27.5 33.9 38.9 3411.0 539.1 -571.6 1328.9 -429.5 -521.1 1711.5 28.1 14.7 -162.6 2764.4 -21.0 12.8 27.4 16.2 -256.4 -17.2 2027.9 1134.8 27.9 2530.2 1926.0 -171.0 -26.9 -15.6 -432.5 2673.0 32.1 2074.8 1512.0 1554.4 5.5 -546.4 Dif.4 2109.3 30.5 1167.1 1255.7 -240.1 -224.3 1625.2 3534.7 5294.2 -340.7 3726.9 3370.4 33.1 3613.6 -347.6 -20.2 -193.7 17.6 552. 1980:1 1980:2 1980:3 1980:4 1981:1 1981:2 1981:3 1981:4 1982:1 1982:2 1982:3 1982:4 1983:1 1983:2 1983:3 1983:4 1984:1 1984:2 1984:3 1984:4 1985:1 1985:2 1985:3 1985:4 1986:1 1986:2 1986:3 1986:4 1987:1 1987:2 1987:3 1987:4 1988:1 1988:2 1988:3 1988:4 1989:1 1989:2 1989:3 1989:4 1990:1 1990:2 1990:3 1990:4 Alt.2 743.5 2877.7 2826.0 -707.3 −AG Act.2 -133.0 -651.7 -119.9 10.7 464.0 4503.0 1395.1 12.8 1087.0 -272.5 -25.0 3442.9 -6.2 -288.5 -17.8 -359. -6.8 Dif.3 635.1 36.4 37.0 -20.2 451.5 27.2 47.4 40.9 39.7 -117. 440.4 47.1 17.4 -26.3 5125.8 -271.6 38.6 -186.5 1587.6 -16.322 11 PROPERTIES OF THE US MODEL Table 11.6 23.1 4428.2 -497.1 955.2 -367.3 850.5 -8.6 -202.3 15.7 34.0 36.5 -22.9 1052. -5.1 2782.0 620.6 37.5 3204.3 3444.6 3217.4 1401.5 1237.3 730. 9.8 -413.3 -14.9 44.0 -323.1 1415.4 3319.5 12.9 13.3 575.4 -5.15 Comparison of the Actual and Alternative Economies Quar.9 -26.9 10.6 -474.0 -624. .5 3789.8 1405.2 39.5 -218.6 5251.1 1.8 5617.1 5044.5 10.4 4710.6 4625.2 1234.0 1179.3 1205.7 2799.6 -178.9 5020.4 4731.4 4245.4 32.1 -8.5 2697.0 3913.2 -22.5 20.2 -92.1 36.8 27.1 9.9 -19.5 -28.5 1187.3 2.5 5149.6 -82.7 2496.8 1124.9 1799.0 34.3 4336.5 -61.3 -452.4 18.0 475.7 491.9 27.6 45.7 1877.0 707.6 24.4 41.9 33.6 3519.0 4940.7 1657.2 -51.3 1376.2 -472.1 -26.4 4822.5 963.2 3052.9 -9.2 2634.5 1981.4 933.5 46.5 increase in D1G of 1 percentage point and a sustained decrease in RS of 2 percentage points over the 1980:1–1990:4 period led to little change in GDP R from the base path and a substantial decrease in the deficit (and thus the debt).5 21.7 5414.6 29.3 1752.1 -679.1 -28.9 46.6 1305.0 4231.4 5.5 51.6 3386.9 725.2 9.8 45.7 514.2 611.1 18.6 2750.8 -143.0 21.1 28.8 2743.1 783.3 -210.8 42.6 26.1 -23.5 25.2 1293.8 1088.8 5601.8 -24. 2312.8 56.5 -67.0 35.6 2789.1 4.3 -11.9 5458.5 2141.0 10.4 3719.6 3388.8 787.1 30.8 -18.4 997.3 3.9 -236.3 39.1 -11.8 -394.4 837.8 Alt.5 4119.6 1044.5 -129.7 -3.9 501.7 3440.1 40.3 1278.1 -22.9 1353.6 4501.8 3343.5 -491.5 -10.8 4060.9 44.2 1299.3 -327.5 14.4 12.7 -12.5 -407.4 -105.8 AH Act.1 873.2 670.7 2664.5 -12.6 1144.5 4874.7 -42.5 14.1 2724.9 2180.5 549.0 -10.7 4312.2 4903.5 -387.9 -13. 34.0 4152.8 -306.6 -598.3 5477.4 533.7 24.6 4731.2 -289.2 4950.9 4475.1 -376.6 -7. 2346.2 46.

3 and 15.6 -12.9 -13.2 10.0 -.8 16.9 1063.0 48.1 9.3 -10.5 -19.5 3.4 1209.1 1020.2 32.6 937.8 2.2 30.4 -.9 1075.2 19.5 23.8 26.9 44.6 -9.9 16.2 -.0 1225.4 1209.1 16.3 -11.8 41.3 37.6 -19.6 -15.7 953. 23.6 1216.2 -12.0 960.7 -20.4 14.9 14.5 1176.8 -22.5 24.2 42.4 1209.0 -17.0 33.2 917.3 -6.2 1215.3 -11.0 21.7 -14.2 17.1 -5.8 10.3 32.5 37.3 -1.1 1213.7 -19.8 1209.8 16.4 -14.9 1204.4 1056.7 27.2 1018.9 .8 34.4 19.9 1043.7 21.1 31.5 12.2 15.3 -2.4 -21.5 -18.7 1139.0 -.4 .7 1163.2 1103.11.1 42.3 41. -5.5 979.6 1034.3 19.9 1.7 IS MONETARY POLICY BECOMING LESS EFFECTIVE? Table 11.9 30.6 928.3 -15.8 -16.5 2.4 1106.8 -7.0 933.2 26.5 3.6 .4 920.6 21.8 33.4 23.8 11.1 31.5 323 Units are billions of 1987 dollars for real GDP and billions of current dollars for the others.1 4.3 GDP R Act.3 -16.2 38.3 -18.6 26.8 40.7 13.4 941.8 24.8 9.0 41.4 -14.3 1044.0 31.6 17.0 13.3 1164.6 12.4 1138.3 -13.8 21.5 -1.4 34.0 1056.0 23.0 -11.0 28.8 -19.2 13.15 (continued) Quar.7 32.3 -5.7 17.5 13.4 952.2 1094.3 1156.0 994.4 11.0 -15.8 43. The flow variables are at quarterly rates.2 -15.3 16.2 3.6 -3.6 -1.6 1194.4 -8.3 -16.0 9.5 -1.4 28.3 26.8 1184.2 I NT G Act.7 -17.4 1129.3 956.8 937.9 31.2 19.1 11.3 1096.9 28.2 47.4 20.6 26.8 1155.4 35.4 -1.9 1082.7 37.3 1.8 21.0 -8.4 30.9 -8.3 49.0 -14.4 .3 22.0 21.4 20.8 938.4 45.8 -11.3 19.1 percent during this period.3 36.5 929.1 16.1 1115.1 1195.8 Dif.2 -13.0 25.4 21.9 20.9 21.0 -5.6 -14.3 27.4 -10.2 1.9 1062.7 -22. 18.7 -6.0 32.8 1214.8 -2.2 9. Dif.0 20.0 931.0 -10.3 36.3 27.8 -12.6 -.2 -6.9 -13.3 32.0 -5. 942.1 -3.4 -17.3 -.9 17.0 .1 38.2 21.8 1176.2 19.5 41.5 950.6 1.3 1082.4 .4 Alt.8 -12.6 -18.0 37.2 18.2 -14.0 -19.8 Dif.4 54.4 -.3 1226.4 -.2 -17.3 -12.4 16.8 1103.1 -18.1 25.2 1048.6 1229.2 26.1 2.4 19.3 1223.6 962.0 14.1 43.4 23.4 -19.6 3.3 -. 11.6 38.4 35.6 27.2 1076.6 31.6 35.5 26.9 22.3 percent for this .1 -7.5 -11.6 -2.1 .6 29.5 48.4 3.7 -16.0 17.5 1129.2 1048. 12.9 -10.3 1204.7 34.1 1106.5 40.7 -12.5 20.8 27. The actual value of the bill rate ranged between 5.7 41.2 .0 -21.6 20.2 956.9 -1.9 29.8 34.1 -1.3 12.0 33.7 32.4 49.4 Alt.6 .5 -9.6 19.8 16. 1980:1 1980:2 1980:3 1980:4 1981:1 1981:2 1981:3 1981:4 1982:1 1982:2 1982:3 1982:4 1983:1 1983:2 1983:3 1983:4 1984:1 1984:2 1984:3 1984:4 1985:1 1985:2 1985:3 1985:4 1986:1 1986:2 1986:3 1986:4 1987:1 1987:2 1987:3 1987:4 1988:1 1988:2 1988:3 1988:4 1989:1 1989:2 1989:3 1989:4 1990:1 1990:2 1990:3 1990:4 Alt.9 941.5 32.9 40.6 .5 962.4 964.4 -18.8 1096.5 32.5 -.0 -4.4 20.3 22.6 15.9 36.4 SH Act.2 1183.3 37.8 22.0 15.6 921.8 -15. -2.0 938.3 13.2 7.2 1036.5 .1 52.3 18.3 -9.0 947.3 26.0 977.0 -1.0 -2.1 48. 940.1 13.2 45.5 . and so the lowest value of the bill rate was 3.3 .2 -1.5 20.2 -15.0 -7.5 1228.7 33.0 -15.3 19.1 -14.5 20.6 20.4 18.3 1115.0 20.2 996.5 919.8 -14.5 26.6 37.7 26.1 20.7 -15.3 29.4 -11.1 -11.4 939.8 34.5 933.2 -3.6 -4.3 41.2 1224.5 2.5 -5.7 1095.8 20.

8 percent higher in the alternative economy than the actual economy. In 1990:4 nonresidential fixed investment was 1.324 11 PROPERTIES OF THE US MODEL simulation. which was caused in part by the lower interest rates.0 percent. lower interest rates with output held constant led to more private investment. nonresidential fixed investment of the firm sector (I KF ) is higher in the alternative economy than the actual economy. the simulation period used in this section was taken to end in 1990:4. More investment means a larger capital stock (KK). net financial assets of the household sector (AH ). Output is roughly the same in both economies. and household saving variables are at quarterly rates in billions of current dollars. the federal government debt (−AG). The variables −AG and AH are stock variables in billions of current dollars. Household saving was $19. I am reluctant to push the model into values that are too far outside the range used in the estimation.15 shows that by the end of the simulation period the federal debt was $766. which was due to the lower past levels of household saving.0 percent higher in the alternative economy. as expected.4 billion less.15 are interesting in their own right in that they show that a 1 percentage point increase in the average personal income tax rate the data used in this book go through 1993:2. The six variables are the federal government deficit (−SGP ). Although not shown in Table 11. which at an annual rate is $115. Note that the real GDP path is similar to the actual path.2 billion less than the actual (historical) value. household saving (SH ). The results in Table 11. which is extremely low by historical standards.4 billion less.15. 20 Although . but the bond rate is lower in the alternative economy. The level of net financial assets of the household sector was $491. raising the personal tax rate and lowering the bill rate led to less government dissaving and less household saving.8 billion less at a quarterly rate.15. and this is the reason for stopping in 1990:4. interest payments of the federal government (I N T G). The level of federal interest payments was $22. Remember that the deficit.5 billion less by the end of the simulation period. and so investment is higher in the alternative economy. Thus. and real GDP is at a quarterly rate in billions of 1987 dollars. Table 11. By the end of 1992 the actual value of the bill rate was down to 3.20 The actual and predicted values of six variables from this simulation are presented in Table 11. interest payments. and a 2 percentage point drop in the bill rate would have lowered it to 1. and real GDP (GDP R). and by 1990:4 KK was 3. as expected. which was the aim of the simulation. The federal deficit was $28.0 percent.2 billion. In equation 12 I KF depends positively on output and negatively on the bond rate. Therefore.

This resulted in a larger fall in disposable income in the actual economy—$19. that the difference between the real GDP increases in Table 11.1 billion in the actual economy versus $68. a difference of 29. and the perfect tracking solution is the solution that reproduces the data for this economy. This means that when the model is solved using the actual values of the exogenous variables. The sum of the changes across the 16 quarters is presented for some of the variables. For each experiment the bill rate was lowered by 1 percentage point in each of the 16 quarters and the model solved.27 percent in the actual economy.16 that the government 21 The smallest value of the bill rate for these experiments was 2. The experiment runs through 1990:4. The residuals are the same for both economies.1 billion in the alternative economy. Note finally from Table 11.35 percent in the alternative economy compared to . It is also the case.4 billion versus $3. The Monetary-Policy Experiment Given the alternative economy.21 For these experiments the residuals were added to the stochastic equations and taken to be exogenous. which resulted in less household demand and thus smaller real GDP increases.16 grows larger as the number of quarters ahead increases. 22 The experiment for the actual economy is the same as the one done for the results in Table 11.9 billion versus $34.2. The monetary-policy experiment is a sustained decrease in the bill rate of 1 percentage point beginning in 1987.11.5 percent in 1987:1 for the alternative economy.3 percent.6 percent. the next step is to run a monetary-policy experiment for the two economies and compare the results. a difference of 13.7 IS MONETARY POLICY BECOMING LESS EFFECTIVE? 325 and a 2 percentage point decrease in the bill rate beginning in 1980:1 would have remarkably changed the debt structure of the U. The increase in real GDP over the 16 quarters is $60. economy by the end of the 1980s while having only trivial effects on real GDP.16 for selected variables. The results in Table 11. For the alternative economy the “actual” values of the bill rate and D1G are the values relevant for this economy.22 A comparison of the results for the two economies is presented in Table 11. The actual values are thus the “base” values.6 billion. .2 except that the starting quarter here is 1987:1 as opposed to 1989:3 in Table 11.1. The (negative) effect from the fall in income is thus larger in the actual economy. for a total of 16 quarters. By the 16th quarter the change in real GDP from the base value is .S. a perfect tracking solution results. which is a useful summary statistic.9 billion over the 16 quarters. however.16 show that government interest payments fell more in the actual than in the alternative economy—$51.

4 -.18 .16 . Act. .27 .00 2 Number of Quarters Ahead 3 4 8 12 .9 -4.9 -19.1 68.1 – – -51.00 .6 -2.01 . Alt.3 -2. These cuts replaced the income tax increase.38 -5.1 -1.6 -1. two other alternative economies were generated.6 -3.6 -68.00 .47 . -.8 -2.4 -1.38 .01 -.5 -.6 -4. Alt.35 16 .25 .1 Sum 60.7 Act.8 -4. Alt.9 -1.2 -1.4 -3.00 .02 RS ↓ RS ↓ I NT G: Federal Government Interest Payments Act. Act.9 -57.7 -1.0 -2.0 -1.2 Alt.51 . Two Other Alternative Economies To examine the robustness of the results to the use of different fiscal-policy tools to generate the alternative economy. Values are percentage changes (in percentage points) from the base values for GDP R and P F and absolute changes (in billions of current dollars at a quarterly rate) from the base values for the others. namely a decrease of 2 percentage points in the bill rate from its base value each quarter.3 -.4 -1.16 Estimated Multipliers in the Actual and Alternative Economies for a Decrease in the Bill Rate of One Percentage Point 1 RS ↓ Act.3 -1.4 -6.5 RS ↓ RS ↓ −SGP : Federal Government Deficit -1.00 . This is primarily due to the larger drop in government interest payments in the actual economy.6 -1.8 Act.33 .1 -1.27 P F : Price Deflator .6 -1. Y D: Disposable Personal Income -2. = Actual economy. The bill rate change in both cases was as above.9 -2.326 11 PROPERTIES OF THE US MODEL Table 11.9 .02 .3 -1. -1.1 -2.0 .00 . Alt.15 .8 -1. -1.44 .06 . = Alternative economy. deficit decreases more in the actual economy than in the alternative economy.35 .6 -2.2 -3. and for the second the level of government purchases of goods (COG) was cut.1 -1.5 -5.4 -5.9 -34.17 .4 -1. For the first the level of transfer payments from the federal government to households (T RGH ) was cut.7 -1.4 GDP R: Real GDP .30 .05 . Alt. Sum = Sum of the effects across the 16 quarters.

For the second of the two other alternative economies. This decrease is comparable in size to the 1 percentage point increase in the average personal income tax rate above.16 for the alternative economy.7 IS MONETARY POLICY BECOMING LESS EFFECTIVE? 327 For the first of the two other alternative economies. changing government purchases of goods has more of an impact on GDP in the model than does changing transfer payments or changing personal tax rates. which compares to $766. changing the level of transfer payments in the model has very similar effects to changing the personal income tax rate.2 billion.3 billion lower than in the actual economy.5 billion in Table 11. The results for the monetary-policy experiment were very similar to those in Table 11. This is 11.2 billion in Table 11.6 billion lower than in the actual economy in 1990:4. since the alternative economy in the current case has a larger government debt (and a smaller level of household net financial assets) than does the alternative . the government debt was much less. The change in the 16th quarter was .16. which compares to 29. These slightly smaller percentages are as expected. which compares to $68.3 percent more in Table 11. As also seen in Table 11. and the results using transfer-payment decreases were quite similar to those using tax-rate increases.16 for the alternative economy.35. Real GDP in the alternative economy was little changed from that in the actual economy.15.15. and the level of net financial assets of the household sector was much less. The results for the monetary-policy experiment in this second case were similar to those in Table 11. The quarterly decreases ranged from $4.8 percent more than in the actual economy. which is the same as in Table 11. and the change in the 16th quarter was .33. The sum of the real GDP increases across the 16 quarters was $67. the level of government purchases of goods was decreased each quarter by exactly the amount needed to keep real GDP unchanged from its base value. which compares to 13.2 percent more than in the actual economy. The level of household net financial assets was $426.16. The sum of the real GDP increases across the 16 quarters was $68. and so the government debt decreased less. the decrease in expenditures on goods needed to keep real GDP unchanged in light of the bill rate decrease was less than the decrease in transfer payments needed or the increase in personal taxes needed. The same conclusions clearly hold when transfer-payment decreases replace tax-rate increases.5 billion.1. In 1990:4 the government debt was $593. As seen in Table 11. which compares to $491. This is 22. the level of transfer payments was decreased each quarter from its base value by 1 percent of the historical value of taxable income (Y T ).6 percent more in Table 11. The federal government deficit thus decreased less in this case.11.1.16.1 billion in Table 11. Therefore.6 billion at quarterly rates.8 to $11.16.

48) for each quarter of the period and the model was solved for this set of values. A similar conclusion is reached for two other alternative economies. economy would have been considerably different by 1990 had the average personal income tax rate been 1 percentage point higher and the bill rate 2 percentage points lower beginning in 1980.16. The differences are.) The first thing to note from the table is that the bill .15 show that the financial asset and liability structure of the U. in the actual economy than it would be if the economy were instead the alternative economy in Table 11. and the estimated residuals were added to the other stochastic equations and taken to be exogenous.6. The 1978:3–1983:4 period was considered. an interesting question to consider is what the economy would have been like had the Fed not had such a tight monetary policy during the 1978–1983 period. For example. The solution values from this simulation are estimates of what the economy would have been like had the Fed not tightened and had the same shocks (estimated residuals) occurred. The bill rate was then taken to be equal to its 1978:2 value (6.16 show that the effectiveness of monetary policy in changing real GDP is between about 13 and 30 percent less.15.328 11 PROPERTIES OF THE US MODEL economy used for the results in Table 11. one can use a model to ask what the economy would have been like had some government policy been different. depending on the measure used. 11. This question was examined using the US model. The government would have dissaved less and the household sector would have saved less. resulting in a substantially lower government debt by 1990 and a substantially smaller level of net financial assets of the household sector. and the same basic conclusion holds here as holds in the other two cases.17. Conclusion The results in Table 11. The results in Table 11. one generated by cutting transfer payments instead of increasing taxes and one generated by cutting government purchases of goods instead of increasing taxes. Equation 30 was dropped from the model. (The differences for GDP R and P F in the table are percentage differences rather than absolute differences. however.8 What if the Fed had Behaved Differently in 1978 and 1990? As discussed in Section 10. The results are presented in Table 11.S. fairly modest.

7 982.48 6.39 10.76 3.32 -6.9 2.1 -15. − Act.89 12.4 942.76 6.72 1.84 -1.0 972.0 960.5 27.96 -.5 -2.0 994.62 -1.09 12.0 933.59 1.8 938.766 .71 -2.48 6.2 46.88 5.20 -2. -.41 8.46 10.8 978.40 -.3 956.45 6.2 -17.6 921.4 6.3 -14.0 21.788 .42 -1. .4 -1.28 1.48 6.39 7.79 5.75 -7.8 -14.32 5.79 UR Act.48 6.63 11. 7.5 -18.48 6.37 14.48 6.1 Pdif. = Est.35 8.88 -2.60 329 Dif.83 1.8 45.5 14. 915.33 -1.20 -.575 .1 -17.17 .48 6.6 941.72 2.575 .685 .98 -3.68 .48 6.815 .53 Dif.775 .1 -2.3 5.4 .89 -3.647 .42 7.51 5.4 16.84 5.80 -1.67 1.9 44.659 .57 -2.3 10.4 GDP R Act.702 .5 929.5 -13.19 8.25 -1.49 7.1 3.80 13.6 963. .43 .60 .757 .9 9.55 -1.01 .32 -.3 -5.37 9.89 -8.3 Est.0 8.48 1.5 12.24 13.794 .32 7.803 .65 -1.690 .70 8.7 4.48 6.67 10.60 2.7 -2.23 8.741 .2 9.42 6.05 9.31 Dif.41 -5.708 .5 927.99 . 1978:3 1978:4 1979:1 1979:2 1979:3 1979:4 1980:1 1980:2 1980:3 1980:4 1981:1 1981:2 1981:3 1981:4 1982:1 1982:2 1982:3 1982:4 1983:1 1983:2 1983:3 1983:4 Est.86 5.48 6.563 .3 -7.03 .95 1.784 .7 Est.6 1.601 .93 6.733 .4 18. 6. .675 .94 10.86 5.716 .9 46.83 15.48 Est.88 -3.0 -7.32 2.8 29.06 -1.601 .66 -.797 .71 1.6 983.55 .1 -1. .02 5. 6.5 933.3 15.71 3. 6.747 .00 .61 -.68 .48 6.1 36.10 9.9 19.71 5.07 9.75 -1.4 32.17 Estimated Economy if the Fed had not Raised Interest Rates in 1978:3–1983:4 Quar.58 3.00 .1 -8.17 2.54 -6.5 -16.4 981.94 -2.83 -2.7 999.4 920.778 . 5.46 5. .22 -1.11.05 3.6 -.5 2.)/Act.6 24.01 -.769 .586 .4 928.616 .0 -16.81 -1.3 -16.68 7.87 5.36 1.48 6.5 944.00 -.0 992.8 958.61 -5.87 3.48 6.71 7.662 .25 RS Act.802 .627 .32 8.790 .00 .36 9.48 6.28 .8 -14.27 .42 9.39 7.9 938.6 962.40 3.8 6.616 .2 −SGP Act.8 28.59 8.08 8.645 .8 950.3 -1.93 8. − 1].0 931.586 .35 -8.1 970.9 945.06 .4 952.14 8.00 5.811 Pdif.48 6.5 958.0 977.563 .50 1.3 29.0 947.4 -3.824 PF Act.5 4.2 986.28 Est.06 6.63 2.02 12.71 14. Pdif.51 Dif.02 5.48 6.4 32.67 1.15 -5.13 9.808 .42 9.723 .7 928.45 -1.2 961.8 DIFFERENT FED BEHAVIOR IN 1978 AND 1990 Table 11.4 939.48 6.628 .757 . = 100[(Est.23 -7.08 .08 -.84 9.7 927.01 1.25 6.8 937.8 42.68 9.8 -12.83 -1.48 6.9 935. − Act.6 928.8 962.3 1.9 10.29 7. 5.48 6.5 7.0 -10.48 6.03 7.671 .70 1. .36 9.17 1.48 6.05 9. -. 1978:3 1978:4 1979:1 1979:2 1979:3 1979:4 1980:1 1980:2 1980:3 1980:4 1981:1 1981:2 1981:3 1981:4 1982:1 1982:2 1982:3 1982:4 1983:1 1983:2 1983:3 1983:4 Quar.7 1. 915.60 -1.76 3.

97 .24 1.5 -10.5 54.5 −SGP Act.051 1.1 39.7 1247. 5.09 Dif.5 1213.90 6.00 3.36 1.22 .00 1.58 -.05 -2.26 6.1 -7.49 -4.62 -.0 58.4 43.99 2.38 1.058 1.10 .50 6. The actual bill rate peaked at 15.42 -.37 .105 1.041 1. 1.8 63.99 6.00 3.065 1.11 Est.48 -. Whether this is a policy that one thinks should have been followed depends on the weights that one attaches to the price level and output.23 .5 1217.071 1.38 6.6 1274.00 3.91 -.059 1.115 1.2 67.3 41.54 7. By the end of the period the price level would have been 1.00 3. and this point cannot be .02 .1 1249.72 3.05 5. -4.21 .00 3. the unemployment rate lower.2 47.5 -11.47 7.9 58.5 1267. 1226.1 -11.05 .59 5.54 -.031 1.116 Pdif.8 36.083 1. 5.32 6. as expected.120 PF Act.64 -.5 1239.72 -.74 6.1 57.1 1261.58 -.6 percent higher. 1226.39 .072 1.1 59.092 1.32 7.35 .57 5. and the price level higher had the Fed not tightened.13 -.00 .041 1.27 -.089 1.1 1233.3 1236.2 -8.7 Est. 7.48 is much lower than the actual rates that occurred.1 1269.7 Pdif.9 1230.00 3.08 2.4 61.00 3.49 7.41 .17 show.26 7.13 3.41 4.59 -2.051 1.02 -3.8 1209.02 Dif.74 6.8 49.1 1226.02 6.7 1275.01 6.41 See notes to Table 11.1 -6.110 1.73 6.6 1216.19 -.1 27.5 1275.67 .62 6.7 -9. The results in Table 11.7 66.99 7.96 Dif.0 -10.91 3.6 70.58 5. 36.29 .64 1.00 3.41 -1.00 Est.0 72.080 1.08 .3 GDP R Act.4 Est.1 percent in 1981:3. 28. .092 1. that output would have been higher.066 1.00 3.17.9 55.00 6.3 1254. 1.85 7.2 1219. 3. rate value of 6. .1 -8.00 3.05 -. In 1981 and 1982 output would have been over 3 percent higher and the unemployment rate would have been over 1.0 50.1 -10.16 .01 -.3 51.01 .85 RS Act. -8.34 -.101 1.3 1219.18 Estimated Economy if the Fed had Lowered Interest Rates in 1990:3–1993:2 Quar.330 11 PROPERTIES OF THE US MODEL Table 11.01 .00 3. -.031 1.94 6.5 percentage points lower.41 6.096 1.00 .0 65. At some point the price level is likely to rise rapidly as output increases. Is the gain of the added output greater than the loss of a higher price level? Remember that in using the US model to consider this question the data do not discriminate well among alternative forms of the demand pressure variable.9 1218.98 UR Act. 1990:3 1990:4 1991:1 1991:2 1991:3 1991:4 1992:1 1992:2 1992:3 1992:4 1993:1 1993:2 Quar.3 1276.7 -8. 1990:3 1990:4 1991:1 1991:2 1991:3 1991:4 1992:1 1992:2 1992:3 1992:4 1993:1 1993:2 Est.58 3.

If the tradeoff estimates are to be trusted.4 percent higher.11.6 percentage points lower and output is 1. and so the tradeoff estimates in this case may not be too bad. only reaching 3. Having said this. The results are presented in Table 11. Again. where there was a recession followed by sluggish growth. equation 30 was dropped from the model. The question examined here is what this period would have been like had the Fed lowered the bill rate to 3. The price level at the end is . The format of this table is the same as that of Table 11. where it lowered the bill rate gradually. they suggest that one would really have to hate higher prices not to think that the Fed overdid it a bit during this period.0 percent by the end of 1992. as expected.0 for each quarter of the period and the model was solved for this set of values.17 does not push the economy into extreme output ranges. Again. The solution values from this simulation are estimates of what the economy would have been like had the Fed kept the bill rate at 3. and the estimated residuals were added to the other stochastic equations and taken to be exogenous.8 DIFFERENT FED BEHAVIOR IN 1978 AND 1990 331 pinned down. however. Therefore. At the peak difference the unemployment rate is .4 percent higher.0 percent and had the same shocks (estimated residuals) occurred. one has to be cautious in using the model to consider tradeoffs between output and the price level. The 1990:3–1993:2 period was considered. and the price level higher had the bill rate been 3. These results thus suggest that had the Fed followed this policy it could not have completely eliminated the sluggish growth that occurred during this period.0 percent right at the beginning of the recession (in 1990:3) and kept it there.18. Another interesting period to consider is the period starting with the recession of 1990–1991.17. the unemployment rate lower. The bill rate was then taken to be equal to 3. the experiment in Table 11. . This contrasts to its actual behavior. the results show that output would have been higher.0 percent.

.

Table 12.3. Table 12. The detailed results are presented and discussed in Section 12. none of the experiments that require stochastic simulation are performed.1 Introduction The properties of the MC model are examined in this chapter.4.12 Analyzing Properties of the MC Model 12. government spending (an increase in COG). and a summary of the major properties of the model is presented in Section 12. Section 12. In particular.3 presents results from 32 multiplier experiments.2 contains a general discussion of the properties of the MC model.2 shows how sensitive the properties of the US model are to its being imbedded in the MC model. The main exogenous variable for each country is the level of government spending. so that all the effects on and from the other countries are taken into account. To save space. since no stochastic simulation of the MC model is done in this study (see the discussion in Section 9. Each experiment is done within the context of the complete MC model.2).1 shows the effects on all the countries of an increase in U. The experiment for each country is a change in the country’s government spending. Table 12.2 for the US model in that it tries to give a general idea of the properties of the model before examining the multiplier experiments. It is similar to Section 11. As was the case with Chapter 8 versus 9.S. The rest of the chapter is a discussion of multiplier experiments. only the effects on the own country are 333 . however. this chapter contains fewer different types of experiments than did Chapter 11. Finally. This chapter is the counterpart of Chapter 11 for the US model. one per country (except the United States).

In short other countries’ economies are stimulated by the increase in imports of the given country. whereas they can in the US model. The effects on and of a country’s exchange rate and balance of payments are also discussed. which increase sales (X—equation I-3). This means that tax and transfer changes cannot be analyzed in the other models. and investment (M—equation 1. This in turn increases the exports of other countries (X85$—equation L-2). the level of merchandise imports (M—equation 1) increases. consumption.2 A General Discussion of the MC Model’s Properties The properties of each country’s model by itself are similar to the properties of the US model in the sense that. In general. fixed investment is disaggregated into residential and non residential in the US model but not in the other models. they do not take away from the fact that the general features of the other models are similar to those of the US model.3. the specification of the US equations was used to guide the specification of the other countries’ equations. a statement like “variable x affects variable y” is not precise. whenever possible. This in turn increases production (Y —equation 4). Also.3. Be aware. C—equation 2. . I —equation 3).2 is thus relevant here. When. Keep in mind in the following discussion that because most variables are endogenous and because the model is simultaneous. The main focus of the discussion in this section is on how the countries are linked together—how they affect each other. 12. this increases the import variable that feeds into the trade share calculations (M85$A—equation I-8). Another difference is that consumption is disaggregated into three categories in the US model but not in the other models. Much of the discussion of the US model in Section 11. which increases imports.334 12 ANALYZING PROPERTIES OF THE MC MODEL presented in Table 12. for whatever reason. but it is sometimes helpful to focus on partial effects—effects through only one channel—in explaining the model’s properties. and this discussion will not be repeated. Trade Links The MC model has standard trade links. however. These three tables are discussed in Section 12. The first link that will be discussed is the standard trade link. While these differences are not trivial. everything affects everything else. that an important difference between the US model and the other models is that the other models do not have an income side. which increases the export variable that is part of the national income accounts (EX—equation I-2).

dollar) leads to depreciations of the other countries’ exchange rates through these equations. Therefore. The German exchange rate equation is important in the model because of the effect that the German exchange rate has on the other European exchange rates. U. other things being equal. Therefore. An increase in the given country’s domestic price level has thus led to an increase in the other countries’ domestic price levels.12. which increases their domestic price level (P Y —equation 5).S. which further increases its imports. the domestic price level (P Y —equation 5) increases. to an increase in the given country’s exports. other things being equal. for whatever reason. which increases the import price variable that appears in their domestic price equations (P M— equation I-19). has a negative effect on other countries’ output.2 GENERAL DISCUSSION OF THE PROPERTIES 335 This is not. of course.S. This in turn increases the import prices of other countries (P MP —equation L-3). There is thus a “trade feedback effect” in this sense. which increases other countries’ exports.S. this increases the country’s export price level (P X—equation 11) and the export price variable that feeds into the trade share calculations (P X$—equation L-1). This is also not the end of the story because the increase in the other countries’ domestic price levels leads.S. This stimulates the given country’s economy. Interest Rate Link The U. which then further increases the given country’s domestic price level. . German Exchange Rate Link The German exchange rate appears as an explanatory variable in the exchange rate equations of the other European countries. and so on. a depreciation of the German exchange rate (relative to the U. an increase in the U.S. interest rate increases the other countries’ interest rates through these equations. Price Links There are also price links in the model. an increase in the U. to an increase in the given country’s import price level. among other things. When. interest rate. the end of the story because the increase in imports of the other countries leads. There is thus a “price feedback effect” in this sense. short term interest rate (RS) appears as an explanatory variable in most of the interest rate reaction functions of the other countries. Therefore. This in turn has a negative effect on demand and output in the countries.

S. This in turn leads to an increase in its domestic price level (equation 5). It could thus be that the country’s exchange rate depreciates in response to the U. on the other hand.S. a fall in the country’s interest rate relative to the U. and so the properties of the model that depend on the inclusion of this variable in the equations must be interpreted with considerable caution. then there will be an appreciation of the country’s currency relative to the U.) Regarding the relative interest rate variable. Austria. Finland.9: the equations for Canada. expansion if the relative interest rate is in the exchange rate equation. A depreciation is thus inflationary. Whether or not the relative interest rate is in the exchange rate equations affects the properties of the model in the following way.S.336 12 ANALYZING PROPERTIES OF THE MC MODEL Effects on Exchange Rates The two effects on a country’s exchange rate through its exchange rate equation are the relative price effect and the relative interest rate effect. other things being equal. it is an important variable in the model. Germany. Assume in this expansion that the U. For Germany and for countries whose exchange rates are not directly influenced by the German rate. Because it appears in the Japanese and German equations. the effect on the exchange rate is ambiguous because. price level rises relative to a particular country’s price level and that the U. dollar through the relative price effect. If the increase in the domestic price level is less than the increase in the price of imports. the relative price term and the German rate have coefficients that are constrained to sum to one. (For countries whose exchange rates are directly influenced by the German rate. significant in either of these equations. Exchange Rate Effects on the Economy A depreciation of a country’s currency (an increase in E) leads to an increase in its price of imports (equations L-3 and then I-19). Say that there is an increase in government spending in the United States that results in an expansion. it appears in six equations in Table 6.S. and Norway. interest rate rises relative to this country’s interest rate. If the relative interest rate does not appear in the country’s exchange rate equation. Japan. but only the relative price variable. the relative interest rate is in the equation. It is not. If. the relative price variable is entered to have a long run coefficient of one. however.S. which are the other European countries. so that in the long run the real exchange rate fluctuates according to fluctuations in the relative interest rate. rate leads to a depreciation of the country’s currency. which the domestic price equations imply will be true .

which improves the balance of payments. An increase in net assets increases imports and consumption (M—equation 1. The effect of a change in the exchange rate on the balance of payments is thus ambiguous. A depreciation of the exchange rate has a positive effect on exports and a negative effect on imports. Effects on the Balance of Payments The balance of payments (S) is determined by equation I-6. On the other hand. In other words. then there is a decrease in the demand for imports (equation 1). Depending on the size of the responses and the lags. which worsens the balance of payments. however. This in turn leads through the trade share equations for the amount of the imports of other countries that are imported from the given country to rise as a percent of the total imports of the countries. a depreciation of a country’s currency leads to an increase in the export price index in $ (P X$—equation L-1).2 GENERAL DISCUSSION OF THE PROPERTIES 337 in this case. Balance of Payments Effects on the Economy When S increases. an increase in A leads to a fall in the short term interest rate through the interest rate reaction function (RS— equation 7). Also. The price of exports (P X) and exports (X85$) have a positive effect on S and the price of imports (P M) and imports (M) have a negative effect. A depreciation of a country’s currency may thus be expansionary since imports fall and exports rise. Holding the export price index in local currency (P X) constant. it may be that a depreciation at first worsens and later improves the balance of payments. Offsetting this at least somewhat. a depreciation has a positive effect on the price of imports. which is the J curve effect. The lagged percentage change in the money supply is an explanatory variable in the interest rate reaction functions of only 4 countries. and the only way that the money .12. This then leads to an increase in the country’s exports (X85$—equation L-2). The Role of the Money Supply The money supply (M1) plays a minor role in the model. other things being equal. C—equation 2). net assets vis-à-vis the rest of the world increase (A— equation I-7). is the fact that the monetary authority may raise interest rates (through the interest rate reaction function) in response to the increased inflation. an improving balance of payments leads the monetary authority to lower the short term interest rate.

Doing this increases the price of imports for the other countries (equations L-2 and I-19).3 Computing Multipliers Change in U. S/(P Y · Y ). where the latter seven are countries with trade share equations only.1 As is known from analyzing the US model in Chapter 11. Government Spending For the first experiment U. 12. which leads to an increase in these other countries’ domestic price levels (equation 5). and U R. 1 The .1. and P X is taken to be exogenous for each. A perfect tracking solution was first obtained by adding all the residuals (including the residuals in the trade share equations) to the equations and taking them to be exogenous. real GDP for each of the quarters 1984:1– 1986:4. This is the case in Table 12. Algeria. Effects of Oil Price Changes The oil exporting countries in the MC model are Saudi Arabia. The key link here is the fact that the price of imports is an explanatory variable in the domestic price equation for each country.1.S.S. since these two variables are in percents rather than percentage points.338 12 ANALYZING PROPERTIES OF THE MC MODEL supply affects other variables in the model is through its effect on the short term interest rate in the reaction functions.S. This leads through the price feedback effect to further increases in import prices and domestic prices. An increase in the price of oil can thus be modeled by increasing P X for these countries. the price level. COG was then increased. Venezuela. The following absolute changes for S/(P Y · Y ) and U R were multiplied by 100. The results for selected variables are presented in Table 12. Iran. an increase in government spending leads to an increase in output. The export price index (P X) for each of these countries is essentially the price of oil. An increase in oil prices (or other positive price shocks) can thus lead to a worldwide increase in prices. The variables for which absolute changes are used are RS. Indonesia. Each number in the table is either the percentage change in the variable in percentage points or the absolute change in the variable. RS is already in percentage points. government spending on goods (COG) was increased by one percent of U. Kuwait. and the United Arab Emirates. The difference for each variable and period between the solution value from this simulation and the actual value is an estimate of the effect of the change on the variable for the period. Iraq. and the short term interest rate. Libya.

07 -.07 -.03 .28 .14 .46 .30 .01 .00 .00 .00 .01 – – – – – – -.04 -.03 -.35 -.40 -.30 .96 1.08 .81 .59 .44 .45 .02 -.01 .94 -.49 .43 .00 .14 .07 .07 .19 .29 .14 .03 .86 -1.37 .18 1.01 -.40 -1.04 .08 .06 .52 .00 .22 -.11 .39 .00 -.40 .02 .07 -.22 -.20 .25 .11 -.89 .09 .28 .03 .28 .08 .28 1.63 2.15 .01 -.00 .18 -.14 1.02 .59 .12 -.22 -.00 .00 .84 -.17 -.10 .07 -.46 .60 .03 .30 -.01 -.03 .19 .01 -.23 -.73 -.85 2.13 .01 .20 -.05 -.04 -.23 .55 -.07 .04 .47 .33 .04 -.08 .81 .20 -.04 .57 -1.00 .24 -.02 -.17 .73 .00 .03 .07 1.05 .20 .08 .11 -.00 -.04 .24 .07 -.79 1.01 .01 .06 -.01 -.04 .00 .05 -.02 .01 -.10 .08 -.40 2.11 .03 .72 .01 .02 .01 .21 .01 -.28 .09 .46 -.80 1.00 .10 .01 -.02 .60 -1.15 -.01 .16 -.00 1.16 .10 PM -.00 .01 .03 .01 .00 -.04 -.05 -.24 .08 -.13 -.41 .03 .09 .01 -.19 -.02 -.00 -.17 Sa -.36 .14 -.16 -.18 .70 .63 – – – – – – .05 .14 .18 .15 .70 -.30 .43 .03 -.01 .12 .85 .49 .54 1.65 2.14 1.01 -.05 .52 .09 -.05 -.3 COMPUTING MULTIPLIERS Table 12.88 .00 .02 -.07 .02 -.01 .03 .12 -.39 .74 .79 .07 .01 .16 .11 .01 .43 -.16 .14 .16 .06 .00 .16 .26 .07 -.97 -.00 .03 .60 .01 -.13 -.10 .82 .01 .04 .02 -.22 -.14 .00 -.11 .69 -.16 1.74 .28 – – – – – – – – – – – – .72 .16 -.03 .00 -.00 -.22 .07 -.18 -.03 -.01 .01 -.00 -.06 .01 -.10 -.06 1.77 .32 1.04 -.26 .12 .00 .00 .35 .90 .23 -.02 .08 1.18 -.07 -.64 1.01 .00 -.81 1.32 -.02 -.11 .23 .21 .07 -.95 .16 .01 .02 .15 -.09 .14 .59 .06 -.15 .21 .22 .01 .00 .00 .95 1.99 .05 .29 -2.14 .08 -.01 -.00 .10 .39 .05 .01 .78 2.95 1.09 -.00 -.29 .07 -.22 .56 1.24 .06 -.03 -.00 .32 -.49 1.01 -.04 -.17 -.85 1.17 .38 .19 -.13 -.11 .01 .00 -.12 .45 .28 -.12 .01 -.11 .42 .02 -.36 .28 -.22 .00 .07 .04 1.04 .40 .58 .03 .38 .17 .06 -.22 .15 .02 .02 .05 – – – – – – .24 -.04 -.18 .02 .08 .79 1.67 .06 -.10 -.08 -.29 -1.01 -.33 .14 .09 .04 .08 .05 .05 -.28 .04 .08 -.06 .71 .18 .05 -.05 .09 -.01 .08 .84 1.73 .00 -.18 .1 Multipliers for a U.05 -.01 -.91 2.20 .06 .00 -.52 -.10 .28 .57 .25 -.04 1.32 .89 .00 .60 -1.45 .01 -.02 -.63 .01 .61 -1.08 -.29 .03 -.41 .82 .00 .03 -.15 .01 .02 .49 .00 .10 .82 1.04 .40 .01 .01 .03 .02 -.25 -.97 1.00 -.42 -.01 .04 -.22 .51 .25 .26 .06 -.02 -.04 -.53 -.05 -.31 .56 .10 -.16 .01 .07 -.80 PY .00 .11 .00 -.22 .21 .01 .79 .12 1.26 .29 .18 .01 .00 .01 .02 .00 .10 .22 .02 339 Y 1 US 2 US 3 US 4 US 8 US 12 US 1 CA 2 CA 3 CA 4 CA 8 CA 12 CA 1 JA 2 JA 3 JA 4 JA 8 JA 12 JA 1 AU 2 AU 3 AU 4 AU 8 AU 12 AU 1 FR 2 FR 3 FR 4 FR 8 FR 12 FR 1 GE 2 GE 3 GE 4 GE 8 GE 12 GE 1 IT 2 IT 3 IT 4 IT 8 IT 12 IT 1 NE 2 NE 3 NE 4 NE 8 NE 12 NE 1 ST 2 ST 3 ST 4 ST 8 ST 12 ST 1 UK 2 UK 3 UK 4 UK 8 UK 12 UK 1.00 -.00 .22 UR -.56 .00 .05 .42 -1.01 -.08 .54 .45 .43 .51 -.00 .65 .00 .65 .11 .07 .37 .52 .28 .22 .01 .03 .01 -.14 -.15 .57 .36 .00 .16 .29 .03 .00 .02 .00 -.97 1.10 .01 .02 -.07 .12.02 .08 .10 .28 .01 .04 .07 .03 -.82 1.06 .13 -.02 -.00 -.56 .18 .63 .00 .00 .06 -.08 .10 .29 .11 .01 .22 .36 .43 .12 .00 .02 -.00 .49 1.41 .99 1.13 .35 .02 .13 .01 -.24 .01 -.08 – – – – – – .99 1.81 1.07 -.09 -.09 .28 .01 -.06 -.40 -.88 -.81 .11 .52 .73 2.28 .00 .00 .10 .01 -.24 .00 .01 -.04 -.00 .01 -.47 .48 -.01 -.03 .91 .21 .26 .47 .12 .03 -.51 .42 -.47 .29 .11 -.05 .22 .32 .00 .00 .00 -.04 -.11 -.84 1.01 .02 .74 .00 .28 .04 .65 .10 .01 . Government Spending Increase RS E M C I X85$ PX .13 .02 .11 -.91 -1.05 .10 -.04 .04 .23 -.04 -.47 .00 .12 -1.55 .50 .02 .01 -.03 -.01 -.07 .01 .38 .20 .76 .11 .00 .11 -.08 .55 -1.02 .38 .07 -.02 .04 -.32 -.18 .68 .09 -.03 .19 .61 .04 .18 .27 .01 .03 .99 -.S.14 -.04 .01 -.65 -1.06 .02 .04 .01 .40 .18 -.03 -.77 .

62 .16 -.30 .41 .13 .12 .02 .06 -.05 .51 -.05 .04 – .96 -.06 .03 .06 .02 -.51 .03 .04 -.59 -.24 .09 -.03 .39 .11 -.49 -1.57 -.41 .05 .02 -.36 .08 – -.04 .07 .74 2.02 -.19 .20 .90 .03 -.57 -1.31 .01 -.12 -.13 -.48 .49 .68 .03 -.83 -.01 -.52 1.22 1.06 1.16 .08 .00 .38 .23 .03 .14 -.24 -.00 -.28 .08 -.08 .50 .17 -.15 -.42 .24 .39 .15 .05 -.14 .38 .01 -.04 -.09 -.16 -.09 .15 .89 .69 -2.43 .01 .06 -.04 -.03 -.02 – -.04 .18 -.11 .14 -.03 -.07 -.06 .04 – -.13 .53 1.00 – – – -.31 -1.14 .10 .95 -.30 .07 -.35 -.26 .02 -2.33 .51 .16 .74 -1.04 -.03 .08 .27 .29 1.19 .46 .34 .00 -.15 1.01 -.19 -.96 1.54 .16 -.07 – -.05 .62 .07 .05 -.89 .15 .11 – Sa .03 .79 UR .70 .04 -.13 .03 – -.15 1.10 .05 .02 .22 -.17 -.23 -.36 .27 -.23 -.00 -.06 -.05 -.05 – .45 .14 -.86 .05 .01 – .04 .17 .00 .01 .05 -.02 -.08 – -.13 .12 -.07 -.81 -1.06 .00 .06 -.01 -.08 .02 .25 .45 .20 -.03 .05 -.22 -1.28 .00 -.16 .04 -.29 .26 .02 -.48 .25 .66 .02 -.03 – .04 -.41 -.05 -.02 -.01 -.04 .02 .19 .00 .02 -.02 .05 .15 -.10 .05 .05 – -.71 .30 .04 .62 -.77 .12 -.72 .02 .60 .43 .06 – .10 -.00 -.02 .01 -.91 -2.06 -.48 – .02 – .11 -.06 .37 .02 .05 .05 .02 .16 .15 PM .07 -.00 -.00 – .09 -.01 .02 .06 -.01 – .04 -.01 .02 .20 .12 .04 .04 .31 .05 .89 -.04 -.64 -.01 .01 -.01 -.06 .35 .29 .03 -.11 -.26 -.03 -.03 .04 .01 – .15 1 FI 2 FI 3 FI 4 FI 8 FI 12 FI 1 AS 2 AS 3 AS 4 AS 8 AS 12 AS 1 SO 2 SO 3 SO 4 SO 8 SO 12 SO 1 KO 2 KO 3 KO 4 KO 8 KO 12 KO 1 BE 2 BE 3 BE 1 DE 2 DE 3 DE 1 NO 2 NO 3 NO 1 SW 2 SW 3 SW 1 GR 2 GR 3 GR 1 IR 2 IR 3 IR 1 PO 2 PO 3 PO 1 SP 2 SP 3 SP 1 NZ 2 NZ 3 NZ Y .87 1.02 -.02 .17 -.37 .52 .05 – .04 .44 1.12 .10 -.03 .12 .03 .01 -.01 .15 .01 -.01 .19 .08 -.23 – -.03 -.04 .05 -.06 .46 -.04 -.02 -.40 -.07 -.03 .00 .27 -.04 .66 1.02 .46 -.00 .58 .00 -.08 – .14 -.04 .34 -.05 .17 .10 1.23 .00 .37 .60 .02 .63 .98 -.96 .04 .95 1.01 .01 .41 -.08 -.61 1.07 .91 -.11 .51 .50 -.43 -.05 .45 .00 .05 -.68 .12 – – – – – – – – – – – – .40 .56 .18 .04 .01 .42 – .01 .32 .02 .01 .09 .01 .47 -.08 .43 .01 – -.02 -.11 .11 -.03 .06 .23 .11 .16 RS .01 -.11 .07 .03 .02 -.00 .11 .46 -.07 .03 .31 .02 -.43 – -.08 .52 -.21 .54 1.49 2.28 -1.03 .13 1.03 .14 -.01 -.01 -.53 .45 .03 .01 .00 .14 -.39 -.19 .11 .09 .29 -.13 .07 .03 .17 -.25 -.01 .02 .06 -.00 .01 .02 .20 -.34 .39 .04 – – – .09 -.27 .02 -.04 -.11 .42 .07 .70 1.04 -.07 -.10 .10 -.22 .04 .67 2.06 .02 .02 -.13 .34 -.01 .00 .01 -.12 .00 .05 – .39 .47 .15 .33 -.00 -.58 .01 -.22 -.33 -.340 12 ANALYZING PROPERTIES OF THE MC MODEL Table 12.16 -.02 – .02 -.00 -.07 .08 .00 – – – -.12 -.18 -.00 -.30 -.02 .72 .00 -.08 .04 -.14 -.10 .22 – -.84 1.77 .06 .12 -.13 -.43 -.02 – .22 .08 .01 .89 – – – .00 -.01 .01 -.06 -.13 -.27 – -.15 PY .06 .35 .15 -.96 .09 .01 .20 -.00 .1 (continued) M C I .15 -.06 .13 .04 – -.24 -.35 X85$ .14 -.12 .09 -.06 -.27 -.20 .05 .75 .72 .03 .38 .45 1.07 .04 -.00 .41 .10 .25 – – – .07 -.59 .00 .11 .98 -.01 .15 .06 .59 .31 1.05 -.24 .36 .07 .04 -.41 1.11 -.16 -.12 -.81 -.14 -.26 PX .03 -.22 .12 E .

26 .12 -.01 -.21 2 PH -.23 -.69 -1.26 -.15 -.00 .01 – .18 -.02 -.02 – Sa .21 PM -.03 .32 .05 – -.03 .S.23 -.23 – – – -.19 .06 341 1 SA – 2 SA – 3 SA – 1 VE – 2 VE – 3 VE – 1 CO – 2 CO – 3 CO – 1 JO -.00 .00 – .01 – .35 .09 . Except for Canada.19 -.08 -.60 -1.07 -.08 – . and the relative increase in the U.22 -. rate.70 -.25 -.08 -.08 -. price level.00 -. price level rises not only absolutely but relative to the price levels of the other countries.02 -.12 .22 -.01 .08 -. 1.15 -.02 -.10 -.08 .33 -.98 -1.33 – -.05 . other things being equal.05 -. 3.04 .03 – .05 .02 -.12 -.03 -.31 .12 -. the exchange rates appreciate.01 -.06 .16 .12 .04 -.00 – .03 – . not S itself.34 -.01 -.13 -.79 3 JO -1.02 .10 -.02 -.00 – .00 .02 – -.29 -.03 -.02 .00 .02 .35 -. These are .43 .26 -.18 -.48 -1.00 -.06 -.37 -.16 -3.09 .59 .01 – . the net effect could go either way.S.04 -.04 – . interest rate has a depreciating effect.02 .11 .04 .02 .27 .01 -.03 -.05 -.11 -.01 -.03 .22 -.14 .07 -. the U.05 .21 -1.04 .03 -.10 -. has a contractionary effect on demand and then output in the other countries.45 1 MA – 2 MA – 3 MA – 1 PA – 2 PA – 3 PA – 1 PH -.04 .16 -.15 .32 .03 -. As discussed in the previous section.S.S.07 .1 (continued) M C I -.12 -.47 -.05 -.00 .39 -.29 .S.56 -.01 – -. This.00 .66 -. The relative increase in the U.89 -.04 .07 2 ID -.01 .15 1 TH – 2 TH – 3 TH – a Variable is S/(P Y · Y ).33 -.14 .86 .08 -. Y -.01 -. and so there is a relative increase in the U.94 .59 -.00 – .02 -.S.95 -.65 .00 – .35 -.08 -2.12 -.10 -.02 -.04 -.18 .02 .09 -.22 .04 -.19 PX – – – – – – -.02 -.09 . and so there is a relative increase in the U.00 -.13 .16 -.94 – – – E X85$ -.21 RS – – – .17 -.25 -.38 -.25 3 ID -.23 .08 PY – – – – – – -.10 -.00 -.00 -.03 -.02 -. interest rate leads to an increase in the interest rates of other countries (through the other countries’ interest rate reaction functions).13 -.S.37 -.06 – .40 -.02 -.54 UR – – – – – – – – – – – – – – – – – – – – – – – – – – – – – – discussion will focus on what these changes did to the economies of the other countries.06 -.00 – .21 -.14 -.06 – .14 .07 -.75 -1.14 -.28 .06 -.33 .45 – – – – – – – – – – – – – – – .00 .08 .03 -.09 -.00 -.02 -.08 -.11 -.68 -1.67 .68 . 2.3 COMPUTING MULTIPLIERS Table 12.13 -.01 -.07 -.03 -.06 .47 -.05 -.09 -. The increase in the U.68 3 PH -1.15 .19 .01 – -.79 -1.00 .01 -.12 .16 -. The U.24 2 JO -.36 1 SY – 2 SY – 3 SY – 1 ID -. price level has an appreciating effect on the other countries’ exchange rates.21 -.01 – .12 -.11 .11 .41 -.05 -.03 -. For the non European countries in which the relative interest rate variable is not included in the exchange rate equations.09 -.72 -.06 -.07 -.02 .12 -.01 .12.00 – .14 .17 – .02 -. interest rate rises more than the other countries’ interest rates.01 .04 -.

and the Philippines. of course. 5. on the other hand. as expected since only the relative price effect is operating. South Africa. New Zealand. as noted in item 1. and so the interest . the exchange rates would have appreciated. The net effect on the United States of the exchange rate changes is for the price of imports to fall.342 12 ANALYZING PROPERTIES OF THE MC MODEL Australia. (Canada’s exchange rate appreciates slightly because of a slight rise in Canada’s interest rate relative to the U. In other words.S. If the variable were not included.) The exchange rates of all the European countries depreciate. but. there are more negative changes in output than positive ones. India. This result is.) U. The net effect could thus go either way. dollar on net appreciates.S. For the other countries except Canada. For example. the increase in their interest rates has a negative effect. its inclusion leads to a depreciation of the German and Japanese exchange rates. 6. The exchange rate results show that the relative interest rate variable is important. imports increase partly because of the increased demand in the U. The depreciation of the exchange rates in Table 12.S. imports leads to an increase in the exports (X85$) of most (but not all) countries. The German exchange rate depreciates because the relative interest rate variable is in the German exchange rate equation and this leads to a depreciation of the exchange rates that are influenced by the German exchange rate. This is because of the effect of the German exchange rate.S.S. 4. which is primarily because of the net appreciation of the U. In this sense the U. even though it is not significant in the exchange rate equations of Germany and Japan.1 must thus be interpreted with caution. however. The increase in U. fall. this property is not strongly supported by the data. Korea. on net the relative interest rate effect dominates the relative price effect. exports. rate. The increase in exports of the other countries has a positive effect on their output. Jordan.S. Again. which means that the relative interest rate effect dominates the relative price effect for these countries. (The reason that exports fall for some countries is discussed in the next item. In general.S. U. This property is not strongly supported by the data. dollar. not just those in which the relative interest rate is an explanatory variable in the exchange rate equations. the exchange rates depreciate. and the net effect does in fact vary across countries. economy and partly because of the fall in the price of imports relative to the price of domestically produced goods.

and it is helped (other things being equal) by the rise in the price of exports and the fall in the price of imports. The second set is the same as that in Table 12.06 -0.92 0.40 0.81 0.18 0.84 -0.16 1.55 -0.94 -0.15 0.47 0.84 -0.30 -0. 7.55 -0. which is due to the drop in import demand from other countries.16 1.11 0.81 0.75 1.03 -0.03 0.08 0.2.17 0.00 0.45 0.14 1.95 1.22 -0.13 0. the same experiment just described was done using the US model by itself.73 -0.91 -1.28 C I X85$ 0. The net effect is negative.00 0.35 PX 0.28 -0.33 0.19 1.06 0.12.00 0. this can lead to a fall in its output even if there is no interest rate effect operating.26 0.1.00 0.31 2.74 0. Given the output changes.38 US in MC 0.22 0.S.84 -0.06 -0.73 -0.00 0. The U.34 2.92 1.19 -0.09 -0.35 -0. The results are presented in Table 12.85 0.22 0.25 -0.29 0.34 0. The first set of results is for the US model alone.00 -0.08 0.96 1. The U.12 1.00 0.82 1.55 -0.79 1. and the balance of payments of most of the other countries improves.11 -0. If a country is one in which exports fall.18 0.87 0.89 -0.78 1.12 M 0.08 0.45 0.33 2. The main differences are: .43 rate effect on average dominates the export effect. the effects on employment.00 0.81 0.64 1.53 US Alone 0.28 -0.05 -0.S.38 2.34 0.95 -0.80 1.24 1.S.76 PY 0.46 0.16 0. balance of payments deteriorates. 8.10 -0.52 0.10 0.46 0.07 -0.00 0.07 343 UR -0. and the unemployment rate are as discussed in Chapter 11 for the US model.91 0.00 0.74 S -0.79 1.99 RS 0.52 -0. Government Spending Increase Y 1 US 2 US 3 US 4 US 8 US 12 US 1 US 2 US 3 US 4 US 8 US 12 US 1.32 0. Government Spending in the US Model Alone It is interesting to see how sensitive the properties of the US model are to being imbedded in the MC model. For some countries the level of exports falls.11 1. balance of payments is hurt by the rise in imports and the fall in exports.63 1.S.00 -0.49 1.84 0. Change in U.2 Multipliers for a U. the labor force. and this discussion will not be repeated here.81 0.56 0.20 0.82 0.3 COMPUTING MULTIPLIERS Table 12. and the second set is for the US model imbedded in the MC model.07 0.88 PM 0.19 0.82 0.02 -0.01 -0. The unemployment rate generally rises when output falls and vice versa.89 0.97 1.05 -0.40 0.21 0.79 1.03 -0.01 -0. To examine this.02 -0.04 1.22 1.00 0.22 1.

01 .23 -.22 1.01 .89 .52 .01 .89 1.05 .30 .81 .21 -.18 .10 1.31 -.14 -.82 .56 1.64 -1.00 .36 .13 -.13 .06 .344 12 ANALYZING PROPERTIES OF THE MC MODEL Table 12.42 .05 .01 -.17 -.34 .94 .02 .33 -.00 .42 -.00 .67 -.17 .31 -.04 .09 .01 -.05 -.16 -.25 -.46 -.33 2.17 1.23 -.99 .21 .01 -.16 1.15 -.70 .94 .30 .12 -.83 2.00 -.73 1.50 PM -.66 -1.00 -.44 .01 -.66 1.00 .35 2.21 .58 .59 1.04 -.07 -.03 -.53 .58 .54 -.52 -.73 2.00 .31 .19 .22 .14 .32 -.85 .15 .46 .14 1.09 .29 3.16 .29 .68 -.00 .03 .35 1.36 2.76 1.05 .81 .00 .25 -1.05 -.11 1.56 1.34 UR -.26 .00 .85 2.01 .17 .58 -.93 .54 2.21 1.25 .33 2.57 2.35 -.01 .14 -.01 -.13 .15 2.25 -.03 .44 .06 -.32 .90 .29 .00 -.00 1.43 .06 .20 .33 3.88 .46 -.88 2.01 .45 -.17 -.09 -.14 1.32 .06 .19 1.50 1.21 -.40 .31 .12 .82 1.05 .37 -.02 .90 – – – – – – .48 1.84 -1.24 2.91 -1.64 -.20 .33 -.04 .00 .02 .78 1.02 .06 .08 .01 -.04 -.14 -.89 .26 .38 .58 .22 .07 -.09 -.43 1.38 .11 -.00 .44 .61 -.08 .31 -.45 .37 .44 -.27 1.16 .83 1.20 .46 .81 1.89 .30 .63 -.09 .47 .66 .39 .07 -.26 .87 .74 .73 .55 3.31 -.78 1.00 .19 -.43 .76 -.54 -.05 .00 .08 .22 -.00 .12 .28 1.24 .02 -.35 .09 .07 1.86 1.15 -.51 2.28 1.16 .42 -.00 .24 -.14 .22 -.18 1.24 .89 .72 .49 1.04 .11 .36 .07 .79 1.12 1.02 -.29 -.00 2.12 1.01 .13 .01 .18 .08 -.01 .02 .40 1.34 -.02 -.52 -.03 -.13 -.50 .92 .10 1.12 .06 -.00 .71 1.12 -.04 .70 1.37 .77 .43 .04 .98 .61 .64 – – – – – – .81 3.07 .95 1.06 .00 .52 1.82 .09 .23 4.06 -.06 -.42 .35 1.63 .17 -.00 .02 .12 .03 .03 .00 .00 .16 -.05 -.99 -1.40 2.06 .90 .08 .59 .24 1.41 .21 -.16 .24 .00 .29 -.10 .18 1.26 .17 -.04 .08 -.12 .80 .54 .99 1.62 .25 -.02 .11 -.12 1.09 -.12 -.11 .05 -.02 .39 .33 .11 -.02 -.24 .97 .79 -.52 2.44 1.00 .02 .01 .80 1.94 -.55 .13 .18 .01 .32 .39 .00 .28 .08 1.25 .54 1.11 .07 -.16 .04 -.41 2.46 .01 .00 .03 .52 -.11 – – – – – – – – – – – – .84 1.22 -.00 .50 .69 2.03 .14 .23 -.09 1.46 .83 3.18 -.00 -.02 .03 .99 .00 .22 .38 .60 .27 .15 -.79 1.10 -.14 1.38 .03 -.24 1.17 .81 1.13 .04 -.21 .23 -.10 1.54 .01 .28 -.46 -.31 .14 -.12 -.02 .11 .11 2.95 .00 -.39 .28 -2.67 1.05 2.07 -.68 – – – – – – -.47 -.48 .30 -.00 .19 -.64 1.04 -.50 -1.13 -.78 1.51 .04 .09 -.09 -.04 .02 -.00 -.01 .27 -1.45 .26 .04 -.61 .01 .31 -.91 1.87 1.01 -.37 .17 -.04 .04 .13 -.27 .22 .03 -.90 .65 1.11 .15 1.41 .33 .00 1.53 .26 .77 .00 .40 -.33 -.38 -.03 -.45 2.59 .18 .80 1.29 .55 -.37 .27 -.43 .00 .81 -1.13 1.00 .11 -.15 .13 1.01 -.01 .46 -.01 .30 -.00 .26 .07 .01 .00 -.35 -.33 .66 1.75 1.07 .17 .12 .10 -.17 .09 .15 1.61 1.11 .25 1.23 -.40 .00 -.04 .91 .03 -.63 – – – – – – -.14 .00 .00 .10 .38 7.01 -.00 .83 1.35 .00 .04 .05 .07 .03 -.06 .06 .42 1.61 .87 1.48 .31 1.00 .00 .55 -.01 -.99 1.11 .58 -.46 .68 .79 1.07 .06 -.94 .01 -.05 -.32 .83 1.40 Y 1 US 2 US 3 US 4 US 8 US 12 US 1 CA 2 CA 3 CA 4 CA 8 CA 12 CA 1 JA 2 JA 3 JA 4 JA 8 JA 12 JA 1 AU 2 AU 3 AU 4 AU 8 AU 12 AU 1 FR 2 FR 3 FR 4 FR 8 FR 12 FR 1 GE 2 GE 3 GE 4 GE 8 GE 12 GE 1 IT 2 IT 3 IT 4 IT 8 IT 12 IT 1 NE 2 NE 3 NE 4 NE 8 NE 12 NE 1 ST 2 ST 3 ST 4 ST 8 ST 12 ST 1 UK 2 UK 3 UK 4 UK 8 UK 12 UK 1.44 .08 .43 .05 .46 .01 .38 -.16 -.07 PY .39 -.00 .10 -.30 .35 -.06 -.04 .16 .48 -.50 .15 Sa -.04 .45 -.06 .01 .37 -.30 .21 1.35 .01 .03 .16 -1.51 .29 .48 .97 -3.18 1.20 .03 .23 .07 .13 .3 Multipliers for Own Government Spending Increases RS E M C I X85$ PX .00 .00 .42 -.95 4.09 .50 1.27 .02 .02 .00 .01 1.46 1.06 .56 .12 .31 .06 -.87 1.38 .01 .79 .18 .02 -.00 .49 5.26 .29 1.00 -.55 .66 .76 .11 .49 3.53 .07 .99 1.09 -.96 1.04 .02 -.04 .26 -.05 .01 .24 -.49 .06 -.03 -.06 .02 -.84 -.27 -.23 .00 -.22 .92 2.79 1.66 .33 -.02 .80 1.04 -.37 .65 1.38 .75 .86 -.73 -.01 -.22 -.12 2.53 -.59 .

00 .29 .35 .01 -.07 .02 .36 .79 .85 1.19 .14 -.91 .00 .00 .30 1.25 .01 .99 2.21 .03 -.67 -.59 .69 1.25 1.41 .82 2.28 -.01 .19 -.00 .03 -.47 -.08 .39 E .50 -.45 -.01 -.20 -.29 .61 2.80 1.00 .50 3.08 1.00 .22 -.06 -.13 .04 -.03 .00 .36 .00 .39 1.00 .59 .06 -.06 2.00 .01 -.91 .44 1.40 -.18 .14 1.19 -.11 -.32 .00 .07 .56 .24 .12 -.08 -.60 -.86 1.13 .30 .50 .48 .07 -.42 .28 .00 .01 .33 .03 -.92 .18 .09 1.03 -.19 .45 3.04 -.00 .30 .18 – – – .23 -.68 .06 1.69 – – – .06 .00 .02 .00 .07 .54 3.05 -.01 .07 .12 1.41 .07 1.00 .00 .38 .62 1.41 .59 .02 -.43 1.31 -.84 1.02 .00 .21 .00 .18 .80 .20 -.40 -.12.26 .23 -.14 UR -.13 .05 .01 .61 .18 .05 .05 -.82 .94 1.35 -.00 .96 1.08 .86 2.14 -.60 .50 .35 .33 9.00 6.28 .05 .07 .00 .14 .17 .01 5.41 .71 .25 1.64 1.11 .03 -.88 .34 .41 .93 .00 .21 .02 .18 .00 -.02 -.21 -.03 -.02 -.00 .68 .00 .02 .05 -.01 1.81 .02 .31 -.77 – – – – – – – – – – – – .07 .14 -.67 -.01 -.09 .77 .59 -.35 .08 .77 .16 1.64 .93 .87 2.03 -.76 1.08 1.04 -.31 .07 .01 -.45 1.22 .96 .71 -.62 – – – – – – – – – Sa -.09 .13 .06 -.23 -.00 -.67 -.26 .19 -.73 .59 .98 2.81 -.00 .00 .06 1.27 -.23 -.01 -.10 -.13 1.00 .28 1.00 -.21 -.28 .02 .00 .06 -.35 .09 -.28 .75 1.00 .13 .14 -.98 PY .08 -.19 1.25 -.22 -.36 .00 .03 -.41 .75 -.05 -.99 .05 .35 -.79 2.49 .00 .40 .31 .14 2.17 -.47 1.14 1.82 .17 1.17 .00 .07 .00 -.90 1.11 -.01 -.08 .03 .14 .37 1.24 -.16 1.13 .01 -.14 1.02 .00 .09 .00 .13 -.48 RS .78 2.07 -.12 .17 -.16 .32 .96 -1.41 .50 .00 .13 .99 .34 .60 1.77 .00 .89 .38 -.33 1.04 -.06 .31 -.07 .59 1.67 -.36 -.01 -.01 .06 1.42 -.01 -.01 .37 -.93 2.03 .03 -.00 .74 -.22 345 Y 1 FI 2 FI 3 FI 4 FI 8 FI 12 FI 1 AS 2 AS 3 AS 4 AS 8 AS 12 AS 1 SO 2 SO 3 SO 4 SO 8 SO 12 SO 1 KO 2 KO 3 KO 4 KO 8 KO 12 KO 1 BE 2 BE 3 BE 1 DE 2 DE 3 DE 1 NO 2 NO 3 NO 1 SW 2 SW 3 SW 1 GR 2 GR 3 GR 1 IR 2 IR 3 IR 1 PO 2 PO 3 PO 1 SP 2 SP 3 SP 1 NZ 2 NZ 3 NZ .15 -.08 .00 -.32 2.59 .15 -.21 .16 -.01 .23 .30 .00 -.01 .83 1.03 – – – -.27 .55 .76 1.01 -.03 1.06 -.68 .51 .02 .01 .39 -.02 -.28 -.02 .3 COMPUTING MULTIPLIERS Table 12.45 -.55 -.09 – – – – – – 1.43 .01 .01 -.13 .08 1.00 .04 .00 .13 .55 .64 .55 .00 -.13 .11 .17 -.60 .54 1.00 .44 -.62 .00 -.39 .45 -.19 .40 -.09 .18 .82 1.21 .36 1.83 3.02 -.00 .17 1.09 -.66 2.03 .97 .76 1.09 .10 .40 .00 .45 3.46 .06 4.05 -.78 .08 .63 .01 1.87 1.09 .84 1.00 .46 .72 .35 .60 .96 .14 X85$ .48 PM .21 -.00 .99 .01 -.34 -.12 -.36 .23 3.00 .11 2.09 -.47 .05 -.16 .00 .03 -.33 .00 .74 .09 -.80 -1.07 -.49 -.04 .00 .33 1.30 .26 -.02 -.67 1.19 -.00 .52 .67 1.04 -.28 2.01 .17 .10 -.31 .64 1.05 .33 1.93 1.09 -.00 .56 .05 .29 .23 -.09 -.76 .00 .15 2.01 .92 – – – .32 – – – – – – – – – – – – -.21 -.70 – – – .18 .38 .3 (continued) M C I .03 .14 .01 .02 1.84 .09 -.27 .04 -.06 .20 1.04 .03 -.35 .95 .01 1.16 -.01 .18 1.20 .07 PX .73 .01 -.23 -.17 -.48 1.61 .02 -.25 -.01 .15 .07 1.43 -.54 1.43 .00 .02 .17 2.00 .48 -.08 .61 .06 2.33 -.38 .21 .18 .10 .48 1.12 .07 .35 1.91 .32 -.74 .00 .18 -.00 .48 .00 -.72 – – – -.00 .09 -.

00 .32 .346 12 ANALYZING PROPERTIES OF THE MC MODEL Table 12.15 .S.00 .00 .44 .00 .00 – – – X85$ .00 .17 .00 .S.31 2.03 Y 1 SA 2 SA 3 SA 1 VE 2 VE 3 VE 1 CO 2 CO 3 CO 1 JO 2 JO 3 JO 1 SY 2 SY 3 SY 1 ID 2 ID 3 ID 1 MA 2 MA 3 MA 1 PA 2 PA 3 PA 1 PH 2 PH 3 PH 1 TH 2 TH 3 TH .10 2.04 -.01 -.47 .56 .23 .91 1.19 PY – – – – – – .00 .00 .12 -.54 .00 .3 (continued) M C I .44 1.22 1.37 1.24 -.30 1.00 .38 1.06 -.93 1.00 .91 1.00 .58 .00 . not S itself.06 -.14 .96 1.00 .01 -.00 .00 .00 .00 .14 .00 RS – – – .11 .99 2.00 .64 .00 – – – .66 1.00 .28 1.08 .00 .27 .00 .00 1.78 1.77 . alone case in terms of increasing output are the facts that imports rise less and exports do not fall.03 .41 1.00 .01 PX – – – – – – . Exports remain unchanged in the U.00 .38 – – – – – – – – – – – – – – – – – – – – – 1.00 .35 .00 . The domestic price level rises more in the U.00 .73 1.00 . Imports rise less in the U.04 .02 . This means that the results in Chapter 11 for the US model alone would not likely change .00 .00 .83 1.67 1.06 .05 – – – E – – – – – – – – – .39 1.07 1.95 2.01 -.45 -.23 4.00 .00 .64 1.46 .24 -.00 .00 .07 .2 are fairly small.00 .00 .11 .00 . The domestic price level relative to the price of imports rises less in the U.00 .50 -.44 .00 .50 .79 .76 1.05 -.26 .00 . In general the differences in Table 12.06 .00 .00 .00 .47 .54 1.96 1.00 .00 1.00 UR – – – – – – – – – – – – – – – – – – – – – – – – – – – – – – a Variable is S/(P Y · Y ).06 .00 .29 – – – – – – – – – – – – – – – .00 .14 -.61 1.85 .12 -.00 .16 .00 .56 .07 .49 1.00 . Working against it are the facts that the interest rate rises more and the price level rises more relative to the nominal wage.52 .00 .00 .00 – – – – – – .04 1.01 .28 1.00 .76 .00 .58 . 1.24 -.64 .01 .06 .01 .17 -.02 -.00 .00 .00 .73 1.00 .36 -.10 1.00 .49 .00 .15 1.10 -.00 .00 .37 -.00 – – – .01 .35 .00 .S.61 2.12 1.00 PM . These offsetting forces roughly cancel out with respect to the effects on output.11 1.43 1.00 .75 .01 .19 -.05 1.00 .01 -.70 .00 .00 . The differences in output in the two cases are small.00 .00 .00 .20 . 3.32 .19 .00 .26 -.74 .21 – – – Sa -. alone case instead of falling as in the other case.S. This leads to a slightly larger rise in the short term interest rate (through the interest rate reaction function of the Fed).00 .05 -.14 1.92 1.14 1.00 .14 .10 -.00 . alone case because the price of imports does not fall.01 .00 .03 1.09 .42 .02 -. Working in favor of the U.S. alone case because of the smaller rise in the domestic price level relative to the price of imports.00 .10 1.21 .33 -.49 .00 .30 .09 2.01 -.51 2.03 -.04 1.17 .01 -.00 .00 .55 .86 .43 2.00 . 2.00 .00 .69 .30 -.72 .71 1.24 -.02 -.13 -.23 3.42 . alone case.00 .48 .

3 COMPUTING MULTIPLIERS 347 much if they were done for the US model imbedded in the MC model.3. 1. A perfect tracking solution was first obtained by adding all the residuals (including the residuals in the trade share equations) to the equations and taking them to be exogenous. In almost every case the domestic price level rises. one per country (except the United States).3 presents results from 32 multiplier experiments. Table 12. There were 32 such experiments. so the relative interest rate effect dominates the relative price effect for Germany. The German exchange rate appreciates. Unlike Table 12. This is as expected since.1. The exchange rate results are mixed. . all done using the complete MC model. which is due to the fact that their exchange rates are heavily tied to the German exchange rate.1. The opposite is true for Canada. which is too many tables even for me. Some of the main results in Table 12. Some exchange rates appreciate and some depreciate. there are offsetting effects on the exchange rate. The own results for each country are presented in Table 12. and the changes in imports are generally positive.3 are the following. Change in Each Country’s Government Spending As mentioned at the beginning of this chapter. 3. The changes in exports are generally negative. as discussed above. as does the short term interest rate (if the variables are endogenous). 2. This is the main reason that the balance of payments changes are generally negative.12. The main difference concerns the domestic price level. which does not change very much as the government spending of the other countries changes. 4. where the inflationary consequences of spending increases would be at least slightly larger with the US model imbedded in the MC model. no results are presented for the non own countries. For Japan the exchange rate changes are very small. To do so would have required 32 tables the size of Table 12. In almost every case the balance of payments worsens. G was then increased. Government spending (G) for each country was increased by one percent of its real GDP for each of the quarters 1984:1–1986:4 for the quarterly countries and for each of the years 1984–1986 for the annual countries. The exchange rate changes are also small for many of the European countries.

5. Nevertheless. to see. A rise in consumption. A government spending increase in a country generally leads to the following: 1. the price of imports may decrease or increase. If in practice one were going to use the MC model for policy purposes. A rise in the demand for imports.4 Common Results Across Countries One conclusion that emerges from the results in Tables 12. 4. A fall in exports.3 is that there are few simple stories. A worsening of the balance of payments. 2. policy action would have. The main effects of a government spending increase that vary across countries are 1) the exchange rate may appreciate or depreciate and 2) because of this.1 and 12. the results for each country would have to be examined individually. A rise in the domestic price level. and this chapter will end by listing them. 3.S. 6. a fall in the unemployment rate. . what the effects on other countries some U. employment. A rise in the short term interest rate. and the labor force.348 13 CONCLUSION 5. The changes in the price of imports are generally negative when the exchange rate appreciates and positive when the exchange rate depreciates. investment. say. 12. It is unlikely that a general result across all countries would emerge. there are a few results that are fairly robust across countries. The size and many times the sign of the effects differ across countries.

a multiplier reaction can take place in the model. Again. Households maximize utility and firms maximize profits. and where future research seems important. 349 . If the amount of labor that households want to supply is more than the amount that firms want to employ. what are some of the main macro results that emerge from this work. their expectations are not necessarily equal to the model’s solution values. 1. Unemployment in the model can be considered to be the difference between the amount of labor that households want to supply at current prices and the amount that firms actually employ. The theoretical model provides an explanation for the existence of disequilibrium within a maximizing context. I will review what I think are some of the main theoretical points. Computer technology is now such that much testing and analysis of models can be done that was not practical even a few years ago. The techniques that were used in the book will not be reviewed in this conclusion. and the approach was used to construct the US and MC models. these two amounts can differ in the model because firms may not set prices right because they do not have rational expectations. Instead. and a number of examples of testing and analysis have been presented in this book. In other words. Prices and wages may not be set in such a way as always to have the amount of labor that households want to supply at these prices equal to the amount that firms want to employ. but firms do not have enough information to form rational expectations in the Muth sense.13 Conclusion This book began by arguing for the Cowles Commission approach in dealing with macro data.

8. The theory is further supported by the industry results in Fair (1989). The Lucas critique is simply handled by arguing that it may be quantitatively of small importance and that if it is not it should be picked up in the various tests. and 6.10 and 6.3. and it provides an explanation for the observed procyclical movements of productivity.12 and did not play a role in the investment equations in the ROW model. 5.13. This result has very important implications for the long run properties of models and calls into question the widespread use of the inflation rate over the price level as the variable to be explained by price equations. The rational expectations hypothesis has a scattering of support in Chapter 5.1. The age distribution variables are significant in Tables 5. 3. The results in Tables 5. whereas the level form is generally accepted. the data are weak at choosing the functional form of the demand pressure variable. It appears possible to pick up effects of the changing age distribution of the population in at least some macroeconomic equations. 5. 4. but very little in Chapter 6. and 5. 7.5b show that the change form of the price equation is strongly rejected by the data. This is thus support for the theory that firms at times hoard labor.5a provide strong support for the hypothesis that import prices affect prices of domestically produced goods. There are too few observations at these points for precise estimates to be made. For almost every country this is so. 6. on the other hand. is not significant in Table 5. models that ignore it will be seriously misspecified and should not do well in the tests.10 and 6. Tables 5.2.14. 5. The excess capital variable. The results in Tables 5. The results in Section 11.6 suggest that the use of the rational expectations assumption has only a minor . In other words. On the other hand.350 13 CONCLUSION 2. The theory is further supported by the survey results of Fay and Medoff (1985) and by the industry results in Fair (1969).13a support the use of the excess labor variable in explaining employment demand.4a are consistent with the theory that firms smooth production relative to sales. if the critique is quantitatively important. One cannot pin down at what point prices begin to rise very rapidly as output increases.11 and 6. The results in Tables 5.

4. Faster computers will also ease the computational burden of dealing with the rational expectations assumption. it will be useful to try leads longer than one quarter or one year for the ROW equations in order to provide a stronger test of the rational expectations hypothesis. To conclude. Interest rate and price links are important in the MC model as well as the usual trade links. . The MC model suggests that there are few simple stories that can be told about the effects of one country on another and about the size of the own effects across countries. In future work with more data and faster computers it should be possible to perform many of the tests and experiments for the MC model that were done in this book for the US model. What I have tried to show is that testing can be an important aspect of the approach. 10.S. monetary policy is becoming somewhat less effective over time as the size of the government debt rises. A few of the robust conclusions than can be drawn from the results are listed in Section 12. 11. interest rate is estimated as affecting the decisions of the monetary authorities of a number of countries. and probably further work is warranted even though in the end the economic importance of the assumption may remain small.13 CONCLUSION 351 effect on the properties of the US model. As noted above. The results in Section 11.7 suggest that U. I hope that this book will stimulate more work using the Cowles Commission approach. there is some scattering of support for the hypothesis. In this sense it is of minor economic importance.S. and in the future more work can be done using this assumption. 9. The U. In future work when more data are available. The result that import prices affect domestic prices provides the main link for prices across countries.

.

Appendix A Tables for the US Model .

.

5c. S & L Gov. (g) 6. U. Chartered Commercial Banks (2) Foreign Banking Offices in U. Government (US) Government-Sponsored Enterprises Federally Related Mortgage Pools Monetary Authority (MA) 355 2. Financial (b) 4. Nonfarm Nonfinancial Corporate Business (F) 3a. Household (h) Corresponding Sector(s) in the Flow of Funds Accounts 1a. Farm Business (FA) 1c. (s) 6. 5d. Nonfarm Noncorporate Business (NN) 2. 5b.S. Fed. (3) Bank Holding Companies (4) Banks in U. Households and Nonprofit Organizations (H1) 1b. Commercial Banking (B1): (1) U. Private Nonbank Financial Institutions (B2): (1) Funding Corporations (2) Savings Institutions (3) Credit Unions (4) Life Insurance Companies (5) Other Insurance Companies (6) Private Pension Funds (7) State and Local Government Employee Retirement Funds (8) Finance Companies (9) Mortgage Companies (10) Mutual Funds (11) Closed End Funds (12) Money Market Mutual Funds (13) Real Estate Investment Trusts (14) Security Brokers and Dealers (15) Issuers of Asset Backed Securities (ABSs) (16) Bank Personal Trusts 4. Gov. Foreign (r) 5.TABLES FOR THE US MODEL Table A.S.1 The Six Sectors of the US Model Sector in the Model 1.S. Affiliated Areas 3b. Firm (f) 3. Foreign (R) 5a. State and Local Governments General Funds (S) .S.

Currency held outside banks. Exog. B$. Cash flow. 0 otherwise. 21 Exog. Net financial assets. B$. s. Purchases of goods. Indirect business tax rate. Net financial assets. B$. Percent of 16+ population 66+ minus percent 16–25. B87$. 68 25 2 Exog. 1 in 1969:1. g. Description APPENDIX A Total net wealth. B$. 1 in 1972:1. Capital consumption. Employee social security tax rate. B$. B$. Exog. 0 otherwise. 0 otherwise. Consumer expenditures for durable goods. s. Exog. Exog. 1 in 1971:4.356 Table A. Profit tax rate. 66 75 79 55 56 22 57 Exog. Consumer expenditures for services. g. Net financial assets. Exog. Capital consumption. B$. Purchases of goods. Exog. Percent of 16+ population 56–65 minus percent 16–25. B87$. h. b. Estimated long term bond issues in the current period. Exog. B$. Bank borrowing from the Fed. f. Employer social security tax rate. B87$. Exog. Exog. . 0 otherwise. g. 91 Exog. Percent of 16+ population 26–55 minus percent 16–25. Estimated long term bond issues in the current period. Indirect business tax rate. Exog. f. B87$. Peak to peak interpolation of CD/P OP . B$. Exog. g. r. h. b. s. B$. 89 73 70 77 Exog. Net financial assets. B$. g. Capital gains (+) or losses (−) on corporate stocks held by the household sector. s. 90 Exog. B$. Exog. s. Net financial assets. Marginal personal income tax rate. Marginal personal income tax rate. 0 otherwise. Personal income tax parameter. B87$. Consumer expenditures for nondurable goods. f. g. Exog. 1 26 Exog. Net financial assets. Personal income tax parameter. 1 in 1969:2. 0 otherwise. h. Exog. Capital consumption. g. s. Profit tax rate. B$. B87$. B$. B87$. f. g. g. Total bank reserves. 1 from 1979:4 through 1982:3. 1 from 1981:1 through 1982:4. 3 Exog. B$.2 The Variables in the US Model in Alphabetical Order Variable AA AB AF AG AG1 AG2 AG3 AH AR AS BF BG BO BR CCB CCF CCH CD CDA CF CG CN COG COS CS CU R D1G D1GM D1S D1SM D2G D2S D3G D3S D4G D5G D691 D692 D714 D721 D794823 D811824 Eq.

g. hours per quarter. Average number of overtime hours paid per job. Residential investment. Dividends paid. B$. Discrepancy for h. Total expenditures. B87$. g. B87$. Residential investment. Average number of non overtime hours paid per job. Indirect business taxes. B87$. Exog. B$. B87$. Physical depreciation rate of the stock of capital. Exog. Peak to peak interpolation of H F . B$. B$. 0 otherwise. Discrepancy for b. Physical depreciation rate of the stock of housing. Exog. Average number of hours paid per job. B$. f. f. B$. B$. Exog. Nonresidential fixed investment. Exog. b. hours per quarter. Discrepancy for g. Discrepancy for f. F I U S deflator. . Gross Domestic Product. Exog. Exog. B$. Exog. Exog. g. B$. Total expenditures. 62 15 Exog. Exog. Gross National Product. B$. Peak to peak interpolation of I H H /P OP . Discrepancy between NIPA and FFA data on capital consumption. 0 otherwise. rate per quarter. Deviation of H F from its peak to peak interpolation. Exports. Average number of hours paid per civilian job. 51 52 Exog. B$. hours per quarter. g. Exog. s. B$. Dividends paid. Exog. rate per quarter. Discrepancy for s. 357 1 from 1983:1 through 1983:4. Indirect business taxes. B$. B87$. Exog. f. Gross National Product. Reserve requirement ratio. 4 Exog. GN P chain price index. Average number of hours paid per military job. B87$. h. B87$. hours per quarter. GDP chain price index. b. B$. f. hours per quarter. Exog. Residential investment. Exog. Payments of factor income to the rest of the world. civilian and military. Exog. s. Exog. B$. Total employment. Physical depreciation rate of the stock of durable goods. 106 113 Exog. Exog. B87$. B$. Farm gross product. rate per quarter. B$. nonfinancial corporate business. Receipts of factor income from the rest of the world. s. Exog. hours per quarter. Dividends received by s. F I ROW deflator. f. Exog.TABLES FOR THE US MODEL D831834 DB DD772 DELD DELH DELK DF DI SB DI SBA DI SF DI SG DI SH DI SR DI SS DRS E EX EXP G EXP S FA F I ROW F I ROW D FIUS F I U SD G1 GDP GDP D GDP R GNP GNP D GNP R HF HFF HFS HG HM HN HO HS I BT G I BT S IHB IHF IHH IHHA I KB Exog. 85 Exog. Gross Domestic Product. Average number of hours paid per job. 18 Exog. b. Discrepancy for r. millions. 1 from 1977:2 on. B$. 82 84 83 129 131 130 14 100 Exog. Exog.

millions. f. h. 13 Exog. h. Imports. B$. B$. B$. 96 Exog. f. Number of worker hours required to produce Y. B87$. s. 19 29 Exog. g. millions. Ratio of actual to potential J J . MF MG MH MR MRS 17 Exog. B$. millions. Potential value of J J . B$. Net interest receipts. Net demand deposits and currency. B$. Number of “moonlighters”: difference between the total number of jobs (establishment data) and the total number of people employed (household survey data). Stock of durable goods. 8 APPENDIX A Nonresidential fixed investment. end of quarter.S. Inventory investment. 9 Exog. B87$. B87$. 27 Exog. . Number of jobs. Inventory valuation adjustment.358 I KF I KFA I KG I KH IM I NS I NT F I NT G I NT OT H I NT ROW I NT S IV A IV F IV H IV V H JF JG J H MI N JJ JJP JJS JM JS KD KH KK KKMI N L1 L2 L3 LAM LM 12 Exog. B87$. g. B87$. Nonresidential fixed investment. Stock of capital. Labor force of men 25–54. possessions plus change in demand deposits and currency of private nonbank financial institutions plus change in demand deposits and currency of federally sponsored credit agencies and mortgage pools minus mail float. Exog. g. B$. B$. millions. Net interest payments. Exog. M1 MB MDI F 81 71 Exog. Stock of housing. h. millions. Labor force of women 25–54. millions. 88 Exog. B$. B87$. Labor force of all others. 20 117 Exog. millions. Exog. B$. f. millions. B87$. sole proprietorships and partnerships and other private business. Net interest payments. Demand deposits and currency. millions. Inventory investment. Demand deposits and currency. Money supply. h. Insurance credits to households from g. Number of jobs. B87$. B$. B$. Net interest payments. Net increase in demand deposits and currency of banks in U. Exog. g. Number of military jobs. B$. B$. Inventory investment. r. f. government. Net interest payments. s. B$. U. r.S. Demand deposits and currency. Nonresidential fixed investment. h. f. Exog. B$. Ratio of the total number of worker hours paid for to the total population 16 and over. 16+. 94 95 Exog. Mineral rights sales. Peak to peak interpolation of I KF . Amount of capital required to produce Y . f. b. B87$. B87$. 58 59 92 93 5 6 7 Exog. Number of civilian jobs. Demand deposits and currency. Amount of output capable of being produced per worker hour. g.

Percentage change in GDP D. Ratio of P I H to P D. Ratio of P I V to P D. f. annual rate. B$. Price deflator for inventory investment. Bond rate. Exog. 104 110 31 Exog. Ratio of W M to W F . Price deflator for X − F A. Exog. Purchases of goods and services. b. annual rate. millions. Price deflator for residential investment. s. Gold and foreign exchange. s. Price deflator for EX. Exog. Noninstitutional population of men 25–54. Ratio of P S to P D. percentage points. Exog. B$. B$. Price deflator for COS. 40 34 Exog. 67 Exog. Percent of 16+ population 25–54. Exog. Price deflator for CN. Output per paid for worker hour (“productivity”). Exog. Before tax profits. Purchases of goods and services. Percentage change in GDP R. 42 120 Exog. 118 41 Exog. annual rate. 38 39 Exog.TABLES FOR THE US MODEL MS MU H P 2554 P CD P CGDP D P CGDP R P CM1 P CN P CS PD P EX PF P FA PG PH P I EB P I EF P I EH PIH PIK PIM PIV P OP P OP 1 P OP 2 P OP 3 P ROD PS P SI 1 P SI 2 P SI 3 P SI 4 P SI 5 P SI 6 P SI 7 P SI 8 P SI 9 P SI 10 P SI 11 P SI 12 P SI 13 P SI 14 P UG P US PX Q RB Exog. Exog. Ratio of gross product of g and s to total employee hours of g and s. Price deflator for nonresidential fixed investment. Exog. Exog. Exog. Price deflator for CD. Ratio of W G to W F . Price deflator for X. millions. Before tax profits. Price deflator for CS. Noninstitutional population of women 25–54. Exog. millions. Exog. Price deflator for FA. Amount of output capable of being produced per unit of capital. Ratio of P CS to (1 + D3G + D3S)P D. Before tax profits. Exog. B87$. percentage points. percentage points. Noninstitutional population 16+. percentage points. Noninstitutional population of all others. 23 359 Demand deposits and currency. Ratio of W S to W F . B$. B$. Price deflator for COG. 37 122 123 124 36 35 33 32 10 Exog. millions. Percentage change in M1. Exog. Ratio of I NT ROW to I N T F + I N T G. Price deflator for X − EX + I M (domestic sales). 16+. Ratio of P I K to P D. Price deflator for I M. Exog. Ratio of P G to P D. B$. Price deflator for CS + CN + CD + I H H inclusive of indirect business taxes. g. . Exog. g. Exog. adjusted. Ratio of P EX to P X. h. Ratio of P CN to (1 + D3G + D3S)P D. Ratio of P CD to (1 + D3G + D3S)P D.

f to s. 103 Exog. B$. g. B$. B$. etc.360 RD RECG RECS RET RM RMA RNT RS RSA SB SF SG SGP SH SH RP I E SI F G SI F S SI G SI GG SI H G SI H S SI S SI SS SR SRZ SS SSP ST AT ST AT P SU BG SU BS SU R T T AU G T AU S T BG T BS T CG T CS T FA T FG T FS T HG T HS TI Exog. 0 through 1981:2. Exog. B$. B87$. B$. percentage points. f. 2 in 1952:2. Exog. Employer social insurance contributions. B$. h. Subsidies less current surplus of government enterprises. B$. B$. . B$. Exog. Exog. 49 50 47 48 Exog. Saving. B$. Total employer and employee social insurance contributions to g. B$. B$. Exog. s. h to s. Saving rate. B$. B$. Farm taxes. 1 in 1981:3. Rental income. h. Exog. Three month Treasury bill rate. Ratio of after tax profits to the wage bill net of employer social security taxes. Saving. 30 130 72 69 76 107 65 121 54 Exog. 109 Exog. After tax mortgage rate. Corporate profit taxes. Progressivity tax parameter in personal income tax equation for g. Corporate profit taxes. Current surplus of federally sponsored credit agencies and mortgage pools and of the monetary authority. Employee social insurance contributions. percentage points. B$. Exog. Retirement credits to households from s. Corporate profit taxes. Saving. Subsidies less current surplus of government enterprises. B$. 105 112 Exog. Saving. B$. 24 128 Exog. f to g. 102 108 Exog. B$. r. Corporate profit tax receipts. Personal income taxes. . g. B$. Total receipts. Employer social insurance contributions. 1 in 1952:1. f to g. B$. After tax bill rate. s. g. Progressivity tax parameter in personal income tax equation for s. B$. h. s. h to s. B$. . Employer social insurance contributions. g. B$.40 in 1991:2 and thereafter. Total employer and employee social insurance contributions to s. B$. g to g. B$. NIA surplus (+) or deficit (−). b to s. s. percentage points. Mortgage rate. B$. Exog. Saving. B$. B$. APPENDIX A Discount rate. 74 116 78 114 Exog. Employee social insurance contributions. f to s. s. 53 Exog. . Corporate profit taxes. b to g. Corporate profit tax receipts. b. percentage points. Total receipts. g. percentage points. NIA surplus (+) or deficit (−). Statistical discrepancy. s to s. h to g. . h to g. Employer social insurance contributions. Personal income taxes. B$. B$. 2 in 1981:4. Statistical discrepancy relating to the use of chain type price indices. B$. Exog. B$. Saving. B$.

B$. g to r. (Includes supplements to wages and salaries including employer contributions for social insurance. (Includes supplements to wages and salaries except employer contributions for social insurance. h. Disposable income. f to r. B$. g. Wage accruals less disbursements.) Total sales f. 45 119 46 60 61 11 115 99 98 64 97 Personal income tax receipts. B$. B87$ = Billions of 1987 dollars. Number of people unemployed. Unemployment insurance benefits. Total transfer payments. B$. B$. Labor constraint variable. excluding unemployment insurance benefits.TABLES FOR THE US MODEL TPG T RF H T RF R T RGH T RGR T RGS T RH R T RRSH T RSH T XCR U UB U BR UR V WA WF 101 Exog. B$. B87$. Transfer payments. 86 28 128 87 63 126 16 361 WG WH 44 43 W LDG W LDS WM WR WS X XX Y YD Y NL YS YT Z Exog. B$. Exog. Production. B$. Transfer payments. B$.) Average hourly earnings of civilian workers in g. Exog. B$. Transfer payments. (Includes supplements to wages and salaries except employer contributions for social insurance.) Real wage rate of workers in f. g to s. B87$. s to h. Exog. s. Stock of inventories. Exog. After tax nonlabor income. Transfer payments. f. B$ = Billions of dollars. Unborrowed reserves. Transfer payments. h to r. g to h. B87$. . B$. B$. Dummy variable for the investment tax credit. (Includes supplements to wages and salaries including employer contributions for social insurance. f.) Average hourly earnings of workers in s. After tax wage rate. Total sales. B$.) Average hourly earnings excluding overtime of all workers. Transfer payments. millions. B$. h. g. Average hourly earnings of military workers. s to h. B$. f to h. Potential output of the firm sector. (Includes supplements to wages and salaries including employer contributions for social insurance. 111 Exog. Civilian unemployment rate. Exog. Transfer payments. (Includes supplements to wages and salaries except employer contributions for social insurance. f. Taxable income. h. Exog. B$.) Average hourly earnings excluding overtime of workers in f.) Wage accruals less disbursements. (Includes supplements to wages and salaries except employer contributions for social insurance. B$. Exog.

(CD/P OP )−1 . AG3. log(L3/P OP 1)−1 . log(L2/P OP 2) cnst. I KF−1 − DELK · KK−1 . log(CS/P OP ) Explanatory Variables APPENDIX A cnst. RSA. RMA−1 · I H H A. Y−1 . (I H H /P OP )−1 . log(CN/P OP ) cnst. RSA [Consumer expenditures: services] 2. Y−2 . log[(Y S − Y )/Y S + . CD/P OP cnst. log[Y D/(P OP · P H )]. log(LM/P OP )−1 . log(AA/P OP )−1 . RH O = 1 [Demand deposits and currency—h] Firm Sector 10. AG2. log(CN/P OP )−1 . AG2. log[W F (1 + D5G)]. Z [Number of moonlighters] 9. Y−3 . RH O = 2 [Residential investment—h] 5. log(L2/P OP 2)−1 .3 The Equations of the US Model STOCHASTIC EQUATIONS LHS Var. Y D/(P OP · P H ). T XCR · I KF A. RH O = 1 [Price deflator for X − F A] 11. AG2. cnst. log P F log P F−1 . RH O = 3 [Production—f] I KF (KK − KKMI N)−1 . V−1 . log(W A/P H ). AG1. AG3. (KH /P OP )−1 . RMA · CDA [Consumer expenditures: durables] 4. AG1. log(L1/P OP 1)−1 . Household Sector 1. I H H /P OP cnst. AG3. (KD/P OP )−1 . log(W A/P H ). log(CS/P OP )−1 . 12. . log(L3/P OP 3) cnst. RB−3 · I KF A [Nonresidential fixed investment—f] Y. Y−4 . RMA [Consumer expenditures: nondurables] 3. log[Y D/(P OP · P H )].362 Table A. log(MH /(P OP · P H ) cnst. AG2. log(W A/P H ). AG1. log[MH−1 /(P OP−1 · P H )]. T [Labor force—all others 16+] 8. log(AA/P OP )−1 . Y D/(P OP · P H ). log(LM/P OP ) cnst. AG3. (AA/P OP )−1 . AG1. T [Labor force—men 25–54] 6. Y−1 . log(CN/P OP )−1 . log(L1/P OP 1) cnst. log(W A/P H ). Z. Y cnst. log[Y D/(P OP · P H )]. Z. log P I M. X.04]−1 . Z [Labor force—women 25–54] 7.

T . RS−1 − RS−2 [Mortgage rate] 21. [Average number of hours paid per job—f] 15.41 · T I + 0 1 i=−39 ( 400 )RBi BFi 1 + ( 400 )RS · . log CCF 23. log[CU R−1 /(P OP−1 · P F )].TABLES FOR THE US MODEL 13. log P F−4 . log[(X − F A)/P OP ]. log H F−1 . log W F−4 . I N T F . RB. H F F . RB−1 − RS−2 . D831834 [Capital consumption—f] Financial Sector 22. log(MF /P F ) cnst. BO/BR cnst. (CF − T F G − T F S) [Capital gains or losses on corporate stocks held by h] 26. log J F cnst. log P F−2 . RS. (P X − P X−1 )V−1 . log(J F /J H MI N )−1 . log H F cnst. RH O = 1 [Inventory valuation adjustment] log[(P I K · I KF )/CCF−1 ]. RH O = 1 [Interest payments—f] 20. DD772 · T . log P F−3 . DD772 · log(J F /J H MI N )−1 . log[CU R/(P OP · P F )] cnst. log H O cnst. H F F−1 . RS − RS−2 . RH O = 1 [Average hourly earnings excluding overtime—f] 17. RM−1 − RS−2 . RH O = 1 [Average number of overtime hours paid per job—f] 16. log DF log[(P I EF − T F G − T F S)/DF−1 ] [Dividends paid—f] 19. T . DD772. (BO/BR)−1 . RM − RS−2 cnst. RH O = 1 [Currency held outside banks] 25. RS−1 − RS−2 . T . CG . D811824. log W F−3 .60 · |AF | 363 log Y . RS − RS−2 . T . log W F log W F−1 . RS(1 − D2G − D2S) [Demand deposits and currency—f] 18. RSA. DD772 · log J F−1 . RB − RS−2 24. log P F . log W F−2 . log P F−1 . I V A cnst. log J F−1 . RD [Bank borrowing from the Fed] cnst. log Y [Number of jobs—f] 14. RH O = 1 [Bond rate] cnst. log(J F /J H MI N )−1 . log(MF−1 /P F ). cnst. log(X − F A).

P H = 35. D692. RS cnst.5·H O)+J G·H G+J M ·H M +J S·H S)] [Average hourly earnings excluding overtime of all workers] P SI 10 · W F [Average hourly earnings of civilian workers—g] IDENTITIES 31. W H = 44. P S = 42. I NT G 1 − . RS−1 . 100[(P D/P D−1 )4 − 1]. log[Y D/(P OP · P H )]. log U .364 Import Equation 27. P I K = 40. P I V = 43. D794823 · P CM1−1 . RH O = 1 [Unemployment insurance benefits] 29. log(I M/P OP )−1 . P CN = 37. W G = . P CD = 38.34 · |AG| [Interest payments—g] 0 1 i=−15 ( 400 )(RBi 30.5·H O)+W G·J G·H G+W M ·J M ·H M +W S ·J S · H S−SI GG−SI SS)/(J F (H N +1.4)BGi + ( 400 )RS · . P CS = 36. P D = 34. RS−2 [Three month Treasury bill rate] [P F (X − F A) + P F A · F A]/X [Price deflator for X] P SI 1 · P X [Price deflator for EX] (P X · X − P EX · EX + P I M · I M)/(X − EX + I M) [Price deflator for domestic sales] (P CS ·CS +P CN ·CN +P CD ·CD +P I H ·I H H +I BT G+I BT S)/(CS + CN + CD + I H H ) [Price deflator for (CS + CN + CD + I H H ) inclusive of indirect business taxes] P SI 2(1 + D3G + D3S)P D [Price deflator for CS] P SI 3(1 + D3G + D3S)P D [Price deflator for CN ] P SI 4(1 + D3G + D3S)P D [Price deflator for CD] P SI 5 · P D [Price deflator for residential investment] P SI 6 · P D [Price deflator for nonresidential fixed investment] P SI 7 · P D [Price deflator for COG] P SI 8 · P D [Price deflator for COS] P SI 9 · P D [Price deflator for inventory investment] 100[(W F ·J F (H N +1. P CM1−1 . P G = 41. D721 [Imports] Government Sectors 28. P I H = 39. log(I M/P OP ) APPENDIX A cnst. D691. log W F . log U B−1 . J J S. log(P F /P I M). P CGDP R. P X = 32. RS−1 . log U B cnst. P EX = 33. D714. RMA−1 .

XX = P SI 11 · W F [Average hourly earnings of military workers] P SI 12 · W F [Average hourly earnings of workers—s] [D1G + ((T AU G · Y T )/P OP )]Y T [Personal income taxes—h to g] [D1S + ((T AU S · Y T )/P OP )]Y T [Personal income taxes—h to s] D2G(P I EF − T F S) [Corporate profits taxes—f to g] D2S · P I EF [Corporate profits taxes—f to s] [D3G/(1 + D3G)](P CS · CS + P CN · CN + P CD · CD − I BT S) [Indirect business taxes—g] [D3S/(1 + D3S)](P CS · CS + P CN · CN + P CD · CD − I BT G) [Indirect business taxes—s] D4G[W F · J F (H N + 1. H N = 63. W M = 46.40(AF − AF−1 ) + BF−40 [Estimated long term bond issues in the current period. g] 365 62. T F S = 51. T H S = 49. BG 57. BF 56.TABLES FOR THE US MODEL 45. W S = 47. I BT G = 52.5 · H O) + W G · J G · H G + W M · J M · H M + W S · J S · H S + DF + DB − DRS + I N T F + I N T G + I N T S + I NT OT H − I NT ROW + RNT + T RF H + P I EH − SI GG − SI SS [Taxable income—h] . X = 61. BR = 58. SI H G = 54. SI F G = 55. V = 64. I BT S = 53.66(AG − AG−1 ) + BG−16 [Estimated long term bond issues in the current period. T F G = 50. KH = 60.5 · H O)] [Employer social insurance contributions—f to g] −. KD = 59. Y T = −G1 · MB [Total bank reserves] (1 − DELD)KD−1 + CD [Stock of durable goods] (1 − DELH )KH−1 + I H H [Stock of housing—h] CS + CN + CD + I H H + I KF + EX − I M + COG + COS + I KH + I KB + I KG + I H F + I H B + I V H − P I EB − CCB [Total sales—f] P CS · CS + P CN · CN + P CD · CD + P I H · I H H + P I K · I KF + P EX · EX − P I M · I M + P G · COG + P S · COS + P I K(I KH + I KB + I KG) + P I H (I H F + I H B) + I V V H − P X(P I EB + CCB) − I BT G − I BT S [Total nominal sales—f] HF − HO [Average number of non overtime hours paid per job—f] V−1 + Y − X [Stock of inventories—f] W F · J F (H N + 1.5 · H O)] [Employee social insurance contributions—h to g] D5G[W F · J F (H N + 1. T H G = 48. f] −.

0 = 67. 0 = 78. GDP R = .5·H O)−RN T −T RF H −T RF R− P I EH − CCH + SU BG + SU BS − I N T F − I NT OT H + I N T ROW − CCF − I V A − ST AT − SI F G − SI F S + F I U S − F I ROW [Before tax profits—f] XX − W F · J F (H N + 1. 0 = 76. (determines AG unless AG is exogenous)] T H S + I BT S + T F S + T BS + SI H S + SI F S + T RGS + DRS − P S · COS − W S · J S · H S − I N T S − SU BS − T RSH − U B − RET + SI SS [Saving—s] SS − AS − MS − DI SS [Budget constraint—s. P I EF = 68. SS = 79. SG = 77. SH = APPENDIX A Y T + CCH − P CS · CS − P CN · CN − P CD · CD − P I H · I H H − P I K · I KH − I V V H − T RH R − T H G − SI H G + T RGH − T H S − SI H S + T RSH + U B + I N S + RET [Saving—h] SH − AH − MH + CG − DI SH [Budget constraint—h. GDP = 83. (determines AR)] T H G + I BT G + T F G + T BG + SU R + SI H G + SI F G + MRS − P G · COG − W G · J G · H G − W M · J M · H M − I N T G − T RGR − T RGH − T RGS − SU BG − I N S + SI GG − P I K · I KG [Saving—g] SG − AG − MG + CU R + (BR − BO) − Q − DI SG [Budget constraint—g. 0 = 72. M1 = 82. (determines AF)] MB + MH + MF + MR + MG + MS − CU R [Demand deposit identity. 0 = 80. SR = 75. 0 = 81. 0 = 74. (determines AS)] AH + AF + AB + AG + AS + AR − CG + DI SH + DI SF + DI SB + DI SG + DI SS + DI SR + ST AT [Asset identity (redundant equation)] M1−1 + MH + MF + MR + MS + MDI F [Money supply] XX + P I V (V − V−1 ) + I BT G + I BT S + W G · J G · H G + W M · J M · H M + W S · J S · H S + W LDG + W LDS + P X(P I EB + CCB) [Nominal GDP] Y + P I EB + CCB + P SI 13(J G · H G + J M · H M + J S · H S) + ST AT P [Real GDP] 66.5 · H O) − RN T − T RF H − T RF R − P I EH − CCH +SU BG+SU BS −I N T F −I NT OT H +I N T ROW −P I K ·I KF − P I H · I H F − MRS − SI F G − SI F S + F I U S − F I ROW [Cash flow—f] CF − T F G − T F S − DF [Saving—f] SF − AF − MF − DI SF − ST AT + DI SBA [Budget constraint—f. CF = 69. 0 = 71. SF = 70. (determines AB)] P I M · I M + T RH R + T RGR + T RF R − P EX · EX + F I ROW − F I U S [Saving—r] SR − AR − MR + Q − DI SR [Budget constraint—r. SB = 73.366 65. (determines MB)] P X(P I EB +CCB)−P I K ·I KB −P I H ·I H B −DB −T BG−T BS −SU R [Saving—b] SB − AB − MB − (BR − BO) − DI SB − DI SBA [Budget constraint—b. (determines AH)] XX+P I V (V −V−1 )−W F ·J F (H N +1.

J J = 96. D1GM = 91. Y S = 99. T CG = 103. KK = 93. U = 87. E = 86. H F F = 101. Y NL = 367 100. KKMI N = 94. D1SM = 92. T P G = 102. civilian and military] L1 + L2 + L3 − E [Number of people unemployed] U/(L1 + L2 + L3 − J M) [Civilian unemployment rate] P SI 14(I N T F + I NT G) [Net interest receipts—r] (AH + MH )/P H + KH [Total net wealth—h] D1G + (2 · T AU G · Y T )/P OP [Marginal personal income tax rate—g] D1S + (2 · T AU S · Y T )/P OP [Marginal personal income tax rate—s] (1 − DELK)KK−1 + I KF [Stock of capital—f] Y /MU H [Amount of capital required to produce Y] Y /LAM [Number of worker hours required to produce Y] (J F · H F + J G · H G + J M · H M + J S · H S)/P OP [Ratio of the total number of worker hours paid for to the total population 16 and over] J J /J J P [Ratio of actual to potential J J ] min(0. Z = 98. RECG = GDP /GDP R [GDP chain price index] J F + J G + J M + J S − LM [Total employment. J H MI N = 95. P U G = 105. 1 − J J P /J J ) [Labor constraint variable] LAM(J J P · P OP − J G · H G − J M · H M − J S · H S) [Potential output of the firm sector] [1 − D1G − D1S − (T AU G + T AU S)(Y T /P OP )](RN T + DF + DB − DRS + I NT F + I NT G + I N T S + I NT OT H − I N T ROW + T RF H + P I EH ) + T RGH + T RSH + U B [After-tax nonlabor income—h] HF − HFS [Deviation of H F from its peak to peak interpolation] T HG − T FA [Personal income tax receipts—g] T F G + T F A + T BG [Corporate profit tax receipts—g] SI H G + SI F G + SI GG [Total social insurance contributions to g] P G · COG + W G · J G · H G + W M · J M · H M + W LDG [Purchases of goods and services—g] T P G + T CG + I BT G + SI G [Total receipts—g] . J J S = 97. U R = 88. I N T ROW = 89. GDP D = 85.TABLES FOR THE US MODEL 84. AA = 90. SI G = 104.

EXP G = APPENDIX A P U G + T RGH + T RGR + T RGS + I N T G + SU BG − W LDG [Total expenditures—g] 107. T RRSH = T RSH + U B [Total transfer payments—s to h] 112. P OP = P OP 1 + P OP 2 + P OP 3 [Noninstitutional population 16 and over] 121. RECS = T H S + T CS + I BT S + SI S + T RGS [Total receipts—s] 113.5 · H O)] [Ratio of after tax profits to the wage bill net of employer social security taxes] 122. P U S = P S · COS + W S · J S · H S + W LDS [Purchases of goods and services—s] 111. W A = 100[(1 − D1GM − D1SM − D4G)[W F · J F (H N + 1. EXP S = P U S + T RRSH + I N T S − DRS + SU BS − W LDS [Total expenditures—s] 114. W R = W F /P F [Real wage rate of workers in f] 120.368 106. P ROD = Y /(J F · H F ) [Output per paid for worker hour: “productivity”] 119.5 · H O)] + (1 − D1GM − D1SM)(W G · J G · H G + W M · J M · H M + W S · J S · H S − SI GG − SI SS)]/[J F (H N + 1. U BR = BR − BO [Unborrowed reserves] 126. T CS = T F S + T BS [Corporate profit tax receipts—s] 109. P CM1 = 100[(M1/M1−1 )4 − 1] [Percentage change in M1] 125. SI S = SI H S + SI F S + SI SS [Total social insurance contributions to s] 110. I V F = V − V−1 [Inventory investment—f] 118. SH RP I E = [(1 − D2G − D2S)P I EF ]/[W F · J F (H N + 1. Y D = W F · J F (H N + 1. P CGDP D = 100[(GDP D/GDP D−1 )4 − 1] [Percentage change in GDP D] 124. SRZ = (Y D − P CS · CS − P CN · CN − P CD · CD)/Y D [Saving rate—h] 117.5 · H O) + J G · H G + J M · H M + J S · H S] [After tax wage rate] . SGP = RECG − EXP G [NIPA surplus or deficit—g] 108.5 · H O) + W G · J G · H G + W M · J M · H M + W S · J S · H S + RN T + DF + DB − DRS + I N T F + I N T G + I N T S + I NT OT H − I NT ROW + T RF H + T RGH + T RSH + U B − SI H G − SI H S − T H G + T F A − T H S − T RH R − SI GG − SI SS [Disposable income—h] 116. P CGDP R = 100[(GDP R/GDP R−1 )4 − 1] [Percentage change in GDP R] 123. SSP = RECS − EXP S [NIPA surplus or deficit—s] 115.

TABLES FOR THE US MODEL 127. RMA = 129. RSA = 128. GNP R = 131. GNP D = RS(1 − D1GM − D1SM) [After tax bill rate] RM(1 − D1GM − D1SM) [After tax mortgage rate] GDP + F I U S − F I ROW [Nominal GNP] GDP R + F I U S/F I U SD − F I ROW/F I ROW D [Real GNP] GN P /GN P R [GNP chain price index] 369 Sector definitions: b = financial sector f = firm sector g = federal government sector h = household sector r = foreign sector s = state and local government sector . GNP = 130.

2 1.8 1.9 1.9 1 3 4 5 8 11 12 14 16 17 1 3 4 5 8 11 12 14 16 17 19 22 6 12 13 6 12 13 2 3 5 10 11 Description APPENDIX A Gross Domestic Product Personal Consumption Expenditures. Durable Goods Personal Consumption Expenditures.2 1.1 1.1 1.2 1.4 The Raw Data Variables for the US Model NIPA Data from the Survey of Current Business Real variables are in 1987 dollars Variable R1 GDP R2 CDZ R3 CNZ R4 CSZ R5 I KZ R6 I H Z R7 I V Z R8 I V FAZ R9 EXZ R10 I MZ R11 GDP R or GDP /P Y A R12 CD or CDZ/P CDA R13 CN or CNZ/P CNA R14 CS or CSZ/P CSA R15 I K or I KZ/P I KA R16 I H or I H Z/P I H A R17 I V R18 I V FA R19 EX or EXZ/P EXA R20 I M or I MZ/P I MA R21 P U RG or P U RGZ/P GA R22 P U RS or P U RSZ/P SA R23 FAZ R24 P ROGZ R25 P ROSZ R26 FA R27 P ROG R28 P ROS R29 F I U S R30 F I ROW R31 CCT R32 T RF R33 ST AT Table Line 1.1 1. Nondurable Goods Personal Consumption Expenditures.1 1. Services Real Nonresidential Fixed Investment Real Residential Fixed Investment Real Change in Business Inventories Real Change in Farm Business Inventories Real Exports Real Imports Real Federal Government Purchases Real State and Local Government Purchases Farm Gross Domestic Product Federal Government Gross Domestic Product State and Local Government Domestic Gross Product Real Farm Gross Domestic Product Real Federal Government Gross Domestic Product Real State and Local Government Gross Domestic Product Receipts of Factor Income from the Rest of the World Payments of Factor Income to the Rest of the World Consumption of Fixed Capital Business Transfer Payments Statistical Discrepancy . Nondurable Goods Real Personal Consumption Expenditures.2 1.9 1.7 1. Services Nonresidential Fixed Investment Residential Fixed Investment Change in Business Inventories Change in Farm Business Inventories Exports Imports Real Gross Domestic Product Real Personal Consumption Expenditures.9 1.2 1.1 1.2 1.1 1.2 1.8 1.2 1.2 1.370 Table A.1 1.1 1.7 1.7 1.1 1. Durable Goods Real Personal Consumption Expenditures.1 1.2 1.2 1.8 1.2 1.9 1.

2 3.2 3.2 3. Nonfinancial Corporate Business Proprietors’Income with Inventory Valuation and Capital Consumption Adjustments Rental Income of Persons with Capital Consumption Adjustment Personal Interest Income Government Unemployment Insurance Benefits Interest Paid by Persons Personal Transfer Payments to Rest of the World (net) Personal Tax and Nontax Receipts.16 1.3 3. Nonfinancial Corporate Business Profits Tax Liability. State and Local Government (S&L) Corporate Profits Tax Accruals. Corporate Business Capital Consumption Adjustment.TABLES FOR THE US MODEL R34 W LDF R35 DP ER R36 T RF H R37 COMP T R38 SI T R39 DC R40 CCCB R41 P I ECB R42 DCB R43 I V A R44 CCADCB R45 I N T F R46 CCCBN R47 P I ECBN R48 T CBN R49 DCBN R50 CCADCBN R51 P RI R52 RNT R53 P I I R54 U B R55 I P P R56 T RH R R57 T P G R58 T CG R59 I BT G R60 SI G R61 P U RGZ R62 T RGH R63 T RGR R64 T RGS R65 I N T G R66 SU BG R67 W LDG R68 T P S R69 T CS R70 I BT S R71 SI S R72 P U RSZ 1. Federal Government Transfer Payments (net) to Persons. Federal Government Indirect Business Tax and Nontax Accruals.16 1.14 1. Federal Government Contributions for Social Insurance.16 1.16 1.3 3.2 3. Federal Government Net Interest Paid. S&L Contributions for Social Insurance. Federal Government (see below for adjustments) Corporate Profits Tax Accruals. Federal Government Personal Tax and Nontax Receipts.14 1.14 1. Corporate Business Consumption of Fixed Capital.16 1.2 3.3 17 19 21 2 7 25 2 10 13 15 16 17 20 28 29 31 34 9 12 14 17 28 29 2 6 9 13 15 19 20 21 22 27 30 2 6 7 11 14 371 Wage Accruals less Disbursements Personal Dividend Income Business Transfer Payments to Persons Compensation of Employees Employer Contributions for Social Insurance Dividends Consumption of Fixed Capital.16 1. Federal Government Wage Accruals less Disbursements.16 1.1 2. S&L Indirect Business Tax and Nontax Accruals.1 3. Nonfinancial Corporate Business Dividends.1 2.9 1.1 2.3 3.1 2. Corporate Business Net Interest.2 3.2 3.9 1.2 3.9 1.1 2. Corporate Business Profits Before Tax.16 1. Corporate Business Inventory Valuation Adjustment. Corporate Business Dividends. Federal Government Subsidies less Current Surplus of Government Enterprises. Federal Government Grants in Aid to State and Local Governments.16 1. S&L Purchases of Goods and Services.2 3.3 3.16 2.2 3. Federal Government (see below for adjustments) Transfer Payments (net) to Rest of the World. Federal Government Purchases. Nonfinancial Corporate Business Profits Before Tax. Nonfinancial Corporate Business Capital Consumption Adjustment.2 3. S&L .16 1.

Services.1 7. data for 1959:3–1987:4 Chain type Price Index.3 3. Durable Goods.1 7.1 7. Federal Government Personal Contributions for Social Insurance to the Federal Government.3 3. Residential Fixed Investment. annual data only Chain type Price Index. Personal Consumption Expenditures. Personal Consumption Expenditures. data for 1959:3–1987:4 Chain type Price Index.17 17 18 22 25 8 3 5 6 14 16 17 6 22 30 38 62 86 94 102 118 142 2 3 61 APPENDIX A Transfer Payments to Persons.14 3. S&L Net Interest Paid. Personal Consumption Expenditures.7b 3. data for 1959:3–1987:4 Chain type Price Index. Gross Domestic Product. State and Local Government Purchases.14 3. Federal Government Purchases.14 3.1 7. annual data only Personal Contributions for Social Insurance to the S&L Governments.3 3. S&L Wage Accruals less Disbursements. data for 1959:3–1987:4 Chain type Price Index. data for 1959:3–1987:4 Chain type Price Index.1 7. Imports. annual data only .1 7.372 R73 T RRSH R74 I NT S R75 SU BS R76 W LDS R77 COMP MI L R78 SI H GA R79 SI QGA R80 SI F GA R81 SI H SA R82 SI QSA R83 SI F SA R84 P Y A R85 P CDA R86 P CNA R87 P CSA R88 P I KA R89 P I H A R90 P EXA R91 P I MA R92 P GA R93 P SA R94 F I U SD R95 F I ROW D R96 I NT ROW A 3. data for 1959:3–1987:4 Chain type Price Index. Rest of the World. Military.1 7.14 7.14 3. annual data only Government Employer Contributions for Social Insurance to the S&L Governments.3 3. Nondurable Goods.13 8.1 7. Nonresidential Fixed Investment. data for 1959:3–1987:4 Implicit Price Deflator for Receipts of Factor Income from the Rest of the World Implicit Price Deflator for Payments of Factor Income to the Rest of the World Net Interest. annual data only Government Employer Contributions for Social Insurance to the Federal Government. annual data only Other Employer Contributions for Social Insurance to the S&L Governments. Exports. data for 1959:3–1987:4 Chain type Price Index.14 3.1 7. S&L Subsidies Less Current Surplus of Government Enterprises.1 7. annual data only Other Employer Contributions for Social Insurance to the Federal Government. data for 1959:3–1987:4 Chain type Price Index. data for 1959:3–1987:4 Chain type Price Index. S&L Compensation of Employees.13 7.

NN Capital Consumption. US Net Financial Investment. US Change in Demand Deposits and Currency.TABLES FOR THE US MODEL Flow of Funds Data Variable R97 CDDCF R98 NF I F R99 I H F Z R100 MRS R101 P I EF R102 CCNF R103 DI SF 1 R104 CDDCNN R105 NF I NN R106 I KNN R107 I V NN R108 CCNN R109 CDDCFA R110 NF I FA R111 I KFA R112 P I EF A R113 DFA R114 T FA R115 CCFA R116 CCADFA R117 CDDCH 1 R118 MV CE. Official Foreign Exchange and Net IMF Position Net Financial Investment. FA Corporate Profits. FA Dividends. and Foreign Exchange. NN. FA Capital Consumption Adjustment. SDR’s. H1 Net Financial Investment. NN Net Financial Investment. CCE is the change in the stock excluding capital gains and losses Net Financial Investment. US Insurance Credits to Households. US . Demand Deposits. FA Net Financial Investment. US Discrepancy. Current Surplus = Gross Saving. S Change in Gold and SDR’s. F Net Financial Investment. FA Capital Consumption. FA Tax Accruals. R Change in U. NIPA. F Depreciation Charges. Nonprofit Institutions Discrepancy. S Change in Demand Deposits and Currency. F Change in Demand Deposits and Currency. S Discrepancy. FA Nonresidential Fixed Investment. CCE Code 103020005 105000005 105012001 105030003 106060005 106300005 107005005 113020003 115000005 115013005 115020003 116300005 133020003 135000005 135013003 136060005 136120003 136231003 136300103 136310103 153020005 153064105 Description 373 R119 NF I H 1 R120 I KH 1 R121 DI SH 1 R122 NF I S R123 DI SS R124 CDDCS R125 RET R126 CGLDR R127 CDDCR R128 CF XU S R129 NF I R R130 P I EF 2 R131 DI SR1 R132 CGLDF XU S R133 CDDCU S R134 NGRR R135 I N S R136 NF I U S R137 DI SU S 155000005 155013003 157005005 205000005 207005005 213020005 224090005 263011005 263020000 263111005 265000005 266060005 267005005 313011005 313020005 313011301 313154005 315000005 317005005 Change in Demand Deposits and Currency.S. NN Change in Demand Deposits and Currency. US Net Capital Grants from R. R Change in U. MV CE is the market value of the stock. F Residential Construction. NN Inventory Investment. H1 Net Purchases of Corporate Equities of Households. FA Change in Checkable Deposits and Currency. F Mineral Rights Sales Profits before Tax. S Retirement Credits to Households. NN Nonresidential Fixed Investment. H1 Nonresidential Fixed Investment. F Discrepancy. Also. R Net Corporate Earnings Retained Abroad Discrepancy. R Change in Gold.S.

CA Net Increase in Liabilities.374 R138 CDDCCA R139 N I ACA R140 N I LCA R141 I KCAZ R142 GSCA R143 DI SCA R144 NI DDLB2 R145 CBRB2 R146 I H BZ R147 CGD R148 CDDCB2 R149 NI AB2 R150 NI LB2 R151 I KB2Z R152 DI SB2 R153 CGLDF XMA R154 CF RLMA R155 N I LBRMA R156 NI DDLRMA 403020003 404090005 404190005 405013003 406000105 407005005 493127005 443013053 645012205 656120000 693020005 694090005 694190005 695013005 697005005 713011005 713068003 713113000 713122605 APPENDIX A Change in Demand Deposits and Currency. Multi Family Units. B1 Net Increase in Liabilities. B1 Nonresidential Fixed Investment. Reits Capital Gains Dividend Change in Demand Deposits and Currency. CA Gross Saving. B2 Discrepancy. MA Net Increase in Liabilities. Chartered Commercial Banks Change in Liabilities in the form of Vault Cash of Commercial Banks of the MA Net increase in Financial Assets in the form of Demand Deposits and Currency of Banks in U. U. MA Change in Federal Reserve Loans to Domestic Banks. B2 Net Increase in Financial Assets.S. B2 Residential Construction. MA Change in Vault Cash and Member Bank Reserves. B1 Net Increase in Financial Assets. Government Mail Float. B1 Discrepancy. B2 Change in Reserves at Federal Reserve. B2 Net Increase in Liabilities.S. B2 Change in Gold and Foreign Exchange. U. CA Net Increase in Financial Assets. MA Change in Liabilities in the form of Demand Deposits and Currency due to Foreign of the MA Change in Liabilities in the form of Demand Deposits and Currency due to U. B2 Nonresidential Fixed Investment. Net Increase in Liabilities in the form of Checkable Deposits. B1 Mail Float. MA Discrepancy. Private Domestic Nonfinancial R157 N I DDLGMA 713123105 R158 NI LCMA R159 N I AMA R160 N I LMA R161 I KMAZ R162 GSMA R163 DI SMA R164 CV CBRB1 R165 N I LV CMA R166 NI DDAB1 R167 CBRB1A R168 NI DDLB1 R169 NI AB1 R170 NI LB1 R171 I KB1Z R172 DI SB1 R173 MAI LF LT 1 R174 MAI LF LT 2 713125005 714090005 714190005 715013003 716000105 717005005 723020005 723025000 743020003 753013003 763120005 764090005 764190005 765013005 767005005 903023105 903029205 . CA Fixed Nonresidential Investment.S. Government of the MA Change in Liabilities in the form of Currency Outside Banks of the MA Net Increase in Financial Assets.S.S. CA Net Increase in Liabilities in the form of Checkable Deposits. Possessions Change in Reserves at Federal Reserve. MA Fixed Nonresidential Investment. CA Discrepancy. Foreign Banking Offices in U. MA Gross Savings. MA Change in Member Bank Reserves.

B-4. Aaa Corporate Bond Rate. Total Private Sector. Quarterly average of monthly data. A-33. unpublished. SA in millions.] R182 CL2 R183 T L R184 AF 1 R185 AF 2 R186 P OP R187 P OP 1 R188 P OP 2 R189 J F R190 H F R191 H O R192 J Q R193 J G R194 J H Q . [EE. C-5. [FRB.] Noninstitutional population of females 25–54. All Persons.] Discount Rate. [EE. Rate at F.] Armed Forces of Females 25–54. [EE.” Quarterly average of monthly data. A-33. Quarterly average of monthly data. Quarterly average of monthly data. A8. All Persons. percentage points [FRB.] Noninstitutional population of males 25–54.TABLES FOR THE US MODEL Interest Rate Data Variable Description R175 RS R176 RM 375 R177 RB R178 RD Three Month Treasury Bill Rate (Auction Average). SA in millions.millions. Quarterly average of monthly data. Quarterly average of monthly data.] Civilian Labor Force of Females 25–54. A-36 and A-37. percentage points. SA in millions. Sum of Employed and Unemployed. unpublished. [EE. secondary markets.Y. FHA mortgages (HUD series). “Armed Forces.] Civilian Labor Force of Males 25–54. SA in millions.] Total noninstitutional population 16 and over.] Federal Government Employment. SA. A25. [EE.] Mortgage Rate. millions. [EE. See below for adjustments. See below for adjustments. See below for adjustments. See below for adjustments. [BLS. “Armed Forces. Quarterly average of monthly data.] Armed Forces of Males 25–54. Quarterly average of monthly data. millions. [BLS. SA. [BLS. A-36 and A-37. B-4. Linear interpolation for missing monthly observations. percentage points. SA in millions. “Labor Force Level—Total Noninstitutional Population.] Employment and Population Data Variable R179 CE R180 U R181 CL1 Description Civilian Employment. SA in millions. Quarterly average of monthly data. Quarterly average of monthly data.] Total Labor Force. SA in millions.] Total Government Employment. [FRB. [BLS. Quarterly average of monthly data. millions.” Quarterly average of monthly data. Bank of N. A36. [BLS.] Unemployment.”] Average Weekly Overtime Hours in Manufacturing. millions. “Basic Industry Data for the Economy less General Government. Quarterly average of monthly data.”] Average Weekly Hours. unpublished. Quarterly average of monthly data.” Quarterly average of monthly data. SA in millions. inclusive of any surcharge. “Basic Industry Data for the Economy less General Government. [BLS. Quarterly average. [BLS. All Persons.] Total Government Employee Hours. Females 25–54.] Employment. unpublished. Quarterly average of monthly data. A25. SA in millions of hours per quarter. [EE. Sum of Employed and Unemployed. See below for adjustments. Quarterly average of monthly data. All Persons. [EE. See below for adjustments. unpublished. Total Private Sector. unpublished. unpublished. unpublished. Males 25–54. [FRB. percentage points. [BLS. C-9. See below for adjustments.R.

1970:3 = 0.] R199 SI F S = [SI F SA/(SI F SA + SI QSA)](SI S − SI H S) [Employer Contributions for Social Insurance.00084 1. R199. f to g. 1970:1 = 1.00069 1973:1 1.00056 1.00028 1.775.00042 1.00268 R195 SI H G = 1.825.] Multiplication factors (see the discussion in Section 3. January 1993 Federal Reserve Bulletin.00009 1.775.25.00071 1.00251 0.00547 0.99878 1.675.] R200 SI SS = SI S − SI H S − SI F S [Employer Contributions for Social Insurance.25. MA. f to s. B2.99880 1. 1968:4 = 1.] R197 SI F G = [SI F GA/(SI F GA + SI QGA)](SI G − SI H G) [Employer Contributions for Social Insurance.525. h to g.] R57 T P G = T P G from raw data −T AXADJ R62 T RGH = T RGH from raw data −T AXADJ [T AXADJ : 1968:3 = 1. For the construction of variables R195. and R203.00107 1. 1969:1 = 2.) Variable P OP P OP 1 P OP 2 CL1 CL2 CE Abbreviations: BLS EE FRB SA Notes: Bureau of Labor Statistics Employment and Earnings.2. the annual observation was used for each quarter of the year. 2. R197. US. See Table A.] R196 SI H S = SI G + SI S − SI T − SI H G [Employee Contributions for Social Insurance.1. b to s. S.00078 1. h to s. 1969:3 = 1. NN. g to g.00046 1952:1–1977:4 1. 1970:2 = 1.8.00006 1.725.376 Adjustments to the Raw Data APPENDIX A [SI H GA/(SI H GA + SI H SA)](SI G + SI S − SI T ) [Employee Contributions for Social Insurance. b to g.] R201 T BG = [T CG/(T CG + T CS)](T CG + T CS − T CBN) [Corporate Profit Tax Accruals. R. 1969:4 = 1.] R198 SI GG = SI G − SI H G − SI F G [Employer Contributions for Social Insurance. F.00052 1. H1.00123 1.00375 1952:1–1972:4 1. 1975:2 = −7.1 for abbreviations: B1. 1969:2 = 2. . FA.] R202 T BS = T CG + T CS − T CBN − T BG [Corporate Profit Tax Accruals.] R203 I NT ROW = −[(I N T F + I N T G)/(I N T F annual +I N T G annual)]I N T ROW A [Net Interest Receipts of r. s to s. CA.2. February 1993 Seasonally Adjusted 1951:1–1971:4 1.00297 1.00014 1.

TABLES FOR THE US MODEL The Raw Data Variables in Alphabetical Order Variable AF1 AF2 CBRB1A CBRB2 CCADCB CCADCBN CCADFA CCCB CCCBN CCE CCFA CCNF CCNN CCT CD CDDCB2 CDDCCA CDDCF CDDCFA CDDCH1 CDDCNN CDDCR CDDCS CDDCUS CDZ CE CFRLMA CFXUS CGD CGLDFXMA CGLDFXUS CGLDR CL1 CL2 CN CNZ COMPMIL COMPT CS CSZ CVCBRB1 DC DCB DCBN DFA DISB1 DISB2 DISCA DISF1 DISH1 DISMA No. R192 R173 R174 R100 R118 R98 R110 R119 R105 R129 R122 R136 R134 R169 R149 R139 R159 R166 R168 R144 R157 R156 R155 R170 R150 R140 R158 R160 R165 R85 R86 R87 R90 R92 R41 R47 R112 R101 R130 R89 R53 R88 R91 R186 R187 R188 R51 R27 R24 R28 R25 Variable PSA PURG PURGZ PURS PURSZ PYA RB RD RET RM RNT RS SIFG SIFGA SIFS SIFSA SIG SIGG SIHG SIHGA SIHS SIHSA SIQGA SIQSA SIS SISS SIT STAT SUBG SUBS TBG TBS TCBN TCG TCS TFA TL TPG TPS TRF TRFH TRGH TRGR TRGS TRHR TRRSH U UB WLDF WLDG WLDS 377 No. R184 R185 R167 R145 R44 R50 R116 R40 R46 R118 R115 R102 R108 R31 R12 R148 R138 R97 R109 R117 R104 R127 R124 R133 R2 R179 R154 R128 R147 R153 R132 R126 R181 R182 R13 R3 R77 R37 R14 R4 R164 R39 R42 R49 R113 R172 R152 R143 R103 R121 R163 Variable DISR1 DISS DISUS DPER EX EXZ FA FAZ FIROW FIROWD FIUS FIUSD GDP GDPR GSCA GSMA HF HO IBTG IBTS IH IHBZ IHFZ IHZ IK IKB1Z IKB2Z IKCAZ IKFA IKH1 IKMAZ IKNN IKZ IM IMZ INS INTF INTG INTROW INTROWA INTS IPP IV IVA IVFA IVFAZ IVNN IVZ JF JG JHQ No. R131 R123 R137 R35 R19 R9 R26 R23 R30 R95 R29 R94 R1 R11 R142 R162 R190 R191 R59 R70 R16 R146 R99 R6 R15 R171 R151 R141 R111 R120 R161 R106 R5 R20 R10 R135 R45 R65 R203 R96 R74 R55 R17 R43 R18 R8 R107 R7 R189 R193 R194 Variable JQ MAILFLT1 MAILFLT2 MRS MVCE NFIF NFIFA NFIH1 NFINN NFIR NFIS NFIUS NGRR NIAB1 NIAB2 NIACA NIAMA NIDDAB1 NIDDLB1 NIDDLB2 NIDDLGMA NIDDLRMA NILBRMA NILB1 NILB2 NILCA NILCMA NILMA NILVCMA PCDA PCNA PCSA PEXA PGA PIECB PIECBN PIEFA PIEF1 PIEF2 PIHA PII PIKA PIMA POP POP1 POP2 PRI PROG PROGZ PROS PROSZ No. R93 R21 R61 R22 R72 R84 R177 R178 R125 R176 R52 R175 R197 R80 R199 R83 R60 R198 R195 R78 R196 R81 R79 R82 R71 R200 R38 R33 R66 R75 R201 R202 R48 R58 R69 R114 R183 R57 R68 R32 R36 R62 R63 R64 R56 R73 R180 R54 R34 R67 R76 .

5 Links Between the National Income and Product Accounts and the Flow of Funds Accounts Receipts from i to j : (i. b. r.378 APPENDIX A Table A. f. j = h. s) hh = fh = bh = rh = gh = sh = hf = ff = bf = rf = gf = sf = hb = fb = bb = rb = gb = sb = hr = fr = br = rr = gr = sr = hg = fg = bg = rg = gg = sg = hs = fs = bs = rs = gs = ss = 0 COMP T − P ROGZ − P ROSZ − (SI T − SI GG − SI SS) − SU BG − SU BS + P RI + RNT + I N T F + T RF H + DCBN + DC − DF A − DCB + P I EF A + CCT − CCCB + CCFA + CCADF A − W LDF DCB − DCBN 0 P ROGZ − SI GG − W LDG + T RGH + I N S + I N T G + SU BG P ROSZ − SI SS − W LDS + T RRSH + RET + I N T S + DP ER − DC + SU BS CSZ + CNZ + CDZ − I BT G − I BT S − I MZ − F I ROW − P I ECB + P I ECBN − CCCB + CCCBN − CCADCB + CCADCBN + I H Z − I H F Z − I H BZ + I KH 1 + I KFA + I KNN + I V F AZ + I V N N I H F Z+I KZ−I KH 1−I KF A−I KN N −I KBZ−I KGZ+I V Z−I V F AZ−I V N N I H BZ + I KBZ EXZ + F I U S P U RGZ − P ROGZ + I KGZ P U RSZ − P ROSZ P I ECB − P I ECBN + CCCB − CCCBN + CCADCB − CCADCBN 0 0 0 0 0 I MZ + T RH R + F I ROW T RF R 0 0 T RGR 0 T P G + T F A + I BT G + SI H G T CG − T F A − T BG + MRS + SI F G T BG + GSCA + GSMA 0 SI GG 0 T P S + I BT S + SI H S T CS − T BS + SI F S T BS 0 T RGS SI SS . g.

.4 for the definitions of the raw data variables.TABLES FOR THE US MODEL Saving of the Sectors SH = SF = SB = SR = SG = SS = Checks 0= SH = SF = SB = SR = SG = SS = 0= 0= 0= f h + bh + gh + sh − (hf + hb + hr + hg + hs) hf + ff + bf + rf + gf + sf − (f h + ff + fg + f s + f r) hb − (bh + bf + bs + bg) hr + gr − rf + f r hg + f g + bg + gg − (gh + gf + gr + gs + gg) hs + f s + bs + gs + ss − (sh + sf + ss) 379 SH + SF + SB + SR + SG + SS NF I H 1 + NF I FA + NF I NN + DI SH 1 NF I F + DI SF 1 + ST AT − DI SBA + W LDF NI AB1 − NI LB1 + NI AB2 − NI LB2 + DI SB1 + DI SB2 + DI SBA NF I R + DI SR1 + NGRR NF I U S +N I ACA−N I LCA+N I AMA−N I LMA+DI SU S +DI SCA+DI SMA NF I S + DI SS −NI DDLB1+NI DDAB1+CDDCB2−N I DDLB2+CDDCF +MAI LF LT 1+ MAI LF LT 2 + CDDCU S + CDDCCA − N I DDLRMA − N I DDLGMA + CDDCH 1 + CDDCFA + CDDCN N + CDDCR + CDDCS − N I LCMA CV CBRB1 + CBRB1A + CBRB2 − N I LBRMA − N I LV CMA CGLDR − CF XU S + CGLDF XU S + CGLDF XMA See Table A.

Eq. Base Period=1971:4. Def. Eq. values before 1952. . Eq.039 Sum of CV CBRB1 + CBRB1A + CBRB2. 1965:4. Value=1926.. Eq. 1973:2. Eq.521 Def. 68 MV CE − MV CE−1 − CCE CN P U RG − P ROG P U RS − P ROS CS Sum of NI LCMA. 70.049511 ... Value for 1952.665 Def. Value=−215. 51 Def..329 (CCCB + CCADCB − CCCBN − CCADCBN )/P X.964 Def. 1963:2. Eq. Eq. 77.1 taken to be 0. Eq.. 1988:4.1 taken to be 7..018428 for 1971:1–1980:4.394 Def. Eq. Base Period=1971:4.023068 for 1981:1– 1993:2. Eq. Eq. 55.421 Def. Peak quarters are 1953:1.6 Construction of the Variables for the US Model Variable in US Model AA AB AF AG AH AR AS BF BG BO BR CCB CCF CCH CD CDA CF CG CN COG COS CS CU R D1G D1GM D1S D1SM D2G D2S D3G D3S D4G D5G DB DELD DELH DELK DF DI SB DI SBA DI SF Construction APPENDIX A Def. Value=244. Value=. Base Period=1971:4. 49 Def. 91 Def..006716 . Eq.1 Def. CCCBN + CCADCBN − CCF A − CCADF A CCT − CCCB + CCF A + CCADF A CD Peak to peak interpolation of CD/P OP . Eq. values before 1952. Value=−. DC − DFA − (DCB − DCBN ) DI SB1 + DI SB2 CCNF + CCF A − CCCBN DI SF 1 + W LDF . Base Period=1971:4. Base Period=1971:4. 1978:4. 1985:1. Eq. 1955:3. Value=53. 50 Def..1. Eq.. 89.1 taken to be the same as the value for 1952. Base Period=1971:4. 47 Def. 53 Def. Eq. 1960:2. .0.1 Sum of CF RLMA.... Eq.014574 for 1952:1–1970:4..1 taken to be the same as the value for 1952. Def. Value for 1952. 75. Eq. 90 Def. and 1993:2. 73. 1968:3.977 Def.. Eq. 79. Base Period=1971:4. See below for P X. 48 Def. Base Period=1971:4.310 Def. Value=−108. Eq...380 Table A. 66... 56.. 52 Def. Value=35. Base Period=1971:4. 55 DCB − DCBN .. Value=−230.

Peak quarters are 1957:3. 1964:3.. 1978:3. and 1989:3. or GDP Def.. 1963:4.. Eq. 131 Def. 84 GDP R Def. (I KB1Z + I KB2Z)/(I KZ/I K) (I KZ − I KH 1 − I KFA − I KN N − I KBZ)/(I KZ/I K) Peak to peak interpolation of I KF . Eq. 100 Peak to peak interpolation of H F .. Eq. 130 13 · H F Def. 82. Eq. Constructed values for 1952:1–1955:4. 106 Def... Eq. 129 Def.. 1974:1.TABLES FOR THE US MODEL DI SG DI SH DI SR DI SS DRS E EX EXP G EXP S FA F I ROW F I ROW D FIUS F I U SD G1 GDP GDP D GDP R GNP GNP D GNP R HF HFF HFS HG HM HN HO HS I BT G I BT S IHB IHF IHH IHHA I KB I KF I KFA I KG I KH IM INS INT F INT G I N T OT H I N T ROW INT S 381 DI SU S + DI SCA + DI SMA DI SH 1 DI SR + NGRR DI SS DC − DP ER TL−U EX Def. 1977:2. Eq.. and 1986:3. Eq. 1980:1. J H Q/J Q I BT G I BT S I H BZ/(I H Z/I H ) I H F Z/(I H Z/I H ) (I H Z − I H F Z − I H BZ)/(I H Z/I H ) Peak to peak interpolation of I H H /P OP . 113 FA F I ROW F I ROW D FIUS F I U SD Def. Eq. 1966:1. J H Q/J Q 520 Def. ((I KCAZ + I KMAZ)/(I KZ/I K) (I KH 1 + I KNN + I KF A)/(I KZ/I K) IM I NS I NT F I NT G P I I − I NT F − I NT G − I NT S − I P P + I N T ROW I NT ROW I NT S . Eq.. and 1985:2. 62 13 · H O. 1966:1. Eq. 57 Def.. Peak quarters are 1955:2. The peaks are 1952:4. 1969:3.

Flat beginning and flat end. Base Period=1971:4. Base Period= 1971:4. Eq.219 Def.909 Sum of CDDCU S + CDDCCA − N I DDLRMA − N I DDLGMA. Rate=DELD Def. Value=887. and 1992:4. 31 F AZ/F A P CD P CGNP D P CGNP R P CM1 P CN P CS PD P EX PF P FA . Eq.. 1965:4. 1962:3. Eq. 1985:4.. Value=247. 71. Flat end. 95 Peak to peak interpolation of J J . 93 CL1 + AF 1 CL2 + AF 2 Def. 85 Def. 1978:2.503 MRS Sum of CDDCS. Eq.. Value=6. Eq. Eq. Eq. 1969:1.. and 1992:4.7. Peak quarters are 1953:2. 1977:3. 1969:1. 58. Value=64. The peaks are 1952:4. Peak quarters are 1953:4. Base Period=1971:4. 1973:2. 59. Value=313. 81.114 Peak to peak interpolation of Y/KK. Base Period=1952:4. 1961:4. Value=149.. 1955:4.. Eq. Eq. Def. Dep. and 1990:1.. CDZ/CD Def. 1984:2. Eq. Base Period=1971:4.526 Sum of CDDCH 1 + CDDCF A + CDDCN N. Value=1270.. Value=10. Eq. Dep. 1979:3.. 1955:3.. Dep.610 NI DDAB1 + CDDCB2 + CDDCCA − MAI LF LT 1 Sum of CDDCF + MAI LF LT 1 + MAI LF LT 2. Eq. Def. 123 Def..382 IV A IV F IV H IV V H JF JG J H MI N JJ JJP JJS JM JS KD KH KK KKMI N L1 L2 L3 LAM LM M1 MB MDI F MF MG MH MR MRS MS MU H APPENDIX A IV A I V − I V F A − I V N N/P I V I V FA + I V N N/P I V I V FAZ + I V N N JF JG Def. 92. Base Period=1971:4. Value=−189. Base Period=1952:4... Rate=DELK Def. Eq. 96 T L − CE − U JQ − JG Def. Rate=DELH Def. Eq. 124 CNZ/CN CSZ/CS Def. 1989:2. 122 Def. 1973:3.079 Sum of CDDCR. 33 EXZ/EX Def.. Eq..276. Value=12. 86 Peak to peak interpolation of Y /(J F · H F ). Eq. 1959:3.. Base Period=1952:4.571. 94 Def. Also sum of −N I DDLB1+N I DDAB1+CDDCB2−N I DDLB2. 1959:2. Base Period=1971:4. 1965:4. Base Period=1971:4.

Eq.6. Def..34.1.95.867. 45 Def.0...36. 1962:4 = . 1970:1 through 1970:4 = .31. Eq.355. Eq. 41 Def. 1961:2 = . Eq.4. Eq. Eq.1.288. Eq. 1990:3 = 1. or P I EF 1 + P I EF 2 P I EFA I H Z/I H I KZ/I K I MZ/I M (I V Z − I V F AZ)/(I V − I V FA). Eq.8.. 1983:2 = .. Base Period=1971:4. 40 Def. 39 Def.9. 44 Def. 38 Def... 1980:2 = . 1979:3 and 1979:4 = . 42 Def. Eq. 46 (P ROG + P ROS)/(J H Q + 520 · AF ) Def. Value=30. 1990:1 = 1. Eq. Eq. 1958:3 = ... 37 Def. Eq...... 35 Def. Eq. See below for P X.8. 67. Eq. 118 (P U RSZ − P ROSZ)/(P U RS − P ROS) Def. Eq. 105 Def. 110 or P U RSZ (CDZ + CNZ + CSZ + I H Z + I KZ + P U RGZ − P ROGZ + P U RSZ − P ROSZ + EXZ − I MZ − I BT G − I BT S + I V F AZ + I V N N )/(CD + CN + CS + I H + I K + P U RG − P ROG + P U RS − P ROS + EX − I M + I V F A + I V NN/P I V ) Sum of CGLDF XU S + CGLDF XMA. 112 RET RM Def. 88 Def. 1991:3 = 1. Eq.85.TABLES FOR THE US MODEL PG PH P I EB P I EF P I EH PIH PIK PIM PIV 383 P OP P OP 1 P OP 2 P OP 3 P ROD PS P SI 1 P SI 2 P SI 3 P SI 4 P SI 5 P SI 6 P SI 7 P SI 8 P SI 9 P SI 10 P SI 11 P SI 12 P SI 13 P SI 14 P UG P US PX Q RB RD RECG RECS RET RM RMA RNT RS (P U RGZ − P ROGZ)/(P U RG − P ROG) Def.. Eq. RB RD Def. 34 (P I ECB − P I ECBN )/P X.... 104 or P U RGZ Def.2. 1980:4 = .. 32 Def. Eq. 128 RNT RS . 1981:2 = . 1982:2 and 1982:3 = 1. Eq.. 36 Def. 1959:3 = . 1975:3 and 1975:4 = . Eq.. P OP P OP 1 P OP 2 P OP − P OP 1 − P OP 2 Def. Eq. with the following adjustments: 1954:4 = .

. and 0 otherwise U UB Def. Eq. T BG T BS T CG T CS T FA Def.... 74 Def. Eq. 83 SU BG SU BS GSCA + GSKA Determined from a regression.. Eq. 76 Def. 101 TPS 0 through 1981:2. Determined from a regression. 111 .. 1 in 1981:3. Eq. 2 in 1981:4. 130 Def. Eq. 108 Def. ... Eq. . Eq. . 114 ST AT Def. Eq..2.. 116 Def.. 65 Def. Eq.384 RSA SB SF SG SGP SH SH RP I E SI F G SI F S SI G SI GG SI H G SI H S SI S SI SS SR SRZ SS SSP ST AT ST AT P SU BG SU BS SU R T AU G T AU S T BG T BS T CG T CS T FA T FG T FS T HG T HS TI TPG T RF H T RF R T RGH T RGR T RGS T RH R T RRSH T RSH T XCR U UB U BR APPENDIX A Def. 1. Eq. 102 Def.43 in 1975:2–1986:1.5 in 1962:3–1963:4 and 1971:3.0 in 1964:1–1966:3 and 1967:3–1969:1 and 1971:4–1975:1. TPG T RF H T RF − T RF H T RGH T RGR T RGS T RH R T RRSH Def. 40 in 1991:2 and thereafter..3.. See Section 3. Eq. Eq.2. Eq. 125 . 121 SI F G SI F S SI G SI GG SI H G SI H S SI S SI SS Def. 1.3. See Section 3.. 78 Def.. Eq. Eq. 107 Def. 69 Def. .. Eq. Eq.. 72 Def.

. 43 W LDF W LDG W LDS COMP MI L/[520(T L − CE − U )] Def.. 115 Def. 64 Def.3. 97 Variables in the model in the first column are defined in terms of the raw data variables in Table A. Eq. Eq.. 119 (P ROSZ − W LDS)/[(J Q − J G)(J H Q/J Q)] Def..0 Def. Eq. Eq... Eq. 61 Def. Value=870. 99 Def. 87 Def. 60 Def... 117.. Base Period=1988:4. 98 Def. Eq..4 or by the identities in Table A. Eq. 126 [COMP T − P ROGZ − P ROSZ − (SI T − SI GG − SI SS) + P RI ]/[J F (H F + .TABLES FOR THE US MODEL UR V WA WF WG WH W LDF W LDG W LDS WM WR WS X XX Y YD Y NL YS YT Z 385 Def.. 63 Def.. Eq.. Eq. Eq.. Eq.5 · H O)] (P ROGZ − COMP MI L − W LDG)/[J G(J H Q/J Q)] Def. Eq. Eq.

386 APPENDIX A Table A.7 First Stage Regressors for the US Model for 2SLS and 3SLS Basic Sets Linear Log constant log(AA/P OP )−1 log(COG + COS) log(CD/P OP )−1 log(CN/P OP )−1 log(CS/P OP )−1 log(1 − D1GM − D1SM − D4G)−1 log EX log H F−1 log(I H H /P OP )−1 log(I M/P OP )−1 log(J F /J H MI N )−1 log[(J G·H G+J M ·H M +J S ·H S)/P OP ] log(KH /P OP )−1 log(KK/KKMI N)−1 P CM1−1 100[(P D/P D−1 )4 − 1]−1 log P F−1 log P I M RB−1 RS−1 RS−2 T log[(T RGH + T RSH )/(P OP · P H−1 )] log V−1 log W F−1 log Y−1 log Y−2 log Y−3 log Y−4 log[Y N L/(P OP · P H )]−1 Z−1 U R−1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 constant (AA/P OP )−1 COG + COS (CD/P OP )−1 (CN/P OP )−1 (CS/P OP )−1 (1 − D1GM − D1SM − D4G)−1 EX H F−1 (I H H /P OP )−1 (I M/P OP )−1 (J F − J H MI N )−1 (J G · H G + J M · H M + J S · H S)/P OP (KH /P OP )−1 (KK − KKMI N )−1 P CM1−1 100[(P D/P D−1 )4 − 1]−1 P F−1 PIM RB−1 RS−1 RS−2 T (T RGH + T RSH )/(P OP · P H−1 ) V−1 W F−1 Y−1 Y−2 Y−3 Y−4 [Y N L/(P OP · P H )]−1 Z−1 U R−1 .

DD772−1 · log(J F /J H MI N )−2 log[(Y S −Y )/Y S +. DD772. log P I M−1 . log DF−2 (BO/BR)−1 . log P F−4 . log(L3/P OP 1)−2 . log(AA/P OP )−2 log(W A−1 /P H−1 ). (KH /P OP )−2 . RMA−1 · CDA. (AA/P OP )−2 . log DF−1 . RB−3 . log[Y D/(P OP · P H )]−1 . AG1. log[MH−1 /(P OP−1 · P H )]−1 . log P F−1 . (KD/P OP )−1 . D692. log[CU R−2 /(P OP−2 · P F−1 )] D691. log(L1/P OP 1)−2 . AG3. log W F−2 . RSA−1 . log(MF−2 /P F−1 ). (P CD/P H )−1 . 100[(P D/P D−4 )−1]−1 . log(LM/P OP )−1 . DD772 · T . log(P F /P I M)−2 log U B−1 . log P F−3 . DD772−1 · log(J F /J H MI N )−2 . log(L3/P OP 3)−1 . log U−2 . CF−1 . log(I M/P OP )−1 . AG2. AG2−1 . RS−5 . D714. RD−2 RB−2 . V−4 I KF−1 . (I H H /P OP )−3 log(W A−1 /P H−1 ). DD772 · T . log(AA/P OP )−2 log(W A−1 /P H−1 ). RS−6 . [Y D/(P OP · P H )]−2 . RS−2 (1 − D2G − D2S)−2 log(P I EF − T F G − T F S)−1 . 1−D2G−D2S. AG3. RS−4 . log[Y D/(P OP · P H )]−2 . log P F−2 . RS−3 . I KF−2 . log W F−6 log(MF /P F )−1 . log[Y D/(P OP · P H )]−1 . log[Y D/(P OP · P H )]−2 . RM−2 . log(1 + D5G)−2 . DD772 · log(J F /J H MI N )−1 . log U−1 . [Y D/(P OP · P H )]−1 . J J S−1 . log[(Y S −Y )/Y S +. log(P CS/P H )−1 . log(AA/P OP )−2 . DELK · KK−1 . AG1. RSA−1 . AG2. log(1 + D5G)−1 . AG3. log(AA/P OP )−2 log[MH /(P OP · P H )]−1 . log(L2/P OP 2)−1 . log[(Y S − Y )/Y S + . log(L1/P OP 1)−1 .TABLES FOR THE US MODEL Additional First Stage Regressors for Each Equation Eq. [RB(1 − D2G − D2S)]−1 . log(AA/P OP )−2 log(W A−1 /P H−1 ). log(CN/P OP )−2 . RD−1 . AG3. log[(Y S −Y )/Y S +. V−3 . log(CN/P OP )−3 . log(J F /J H MI N )−2 . AG2. log W F−4 . AG1. RS−4 . AG3−1 . log P F−2 .04]−1 . log(I M/P OP )−2 . T XCR · I KFA. D2G + D2S. AG2. log[Y D/(P OP · P H )]−1 . log(LM/P OP )−2 . RS−1 (1−D2G−D2S)−1 . RMA−1 · I H H A. 100[(P D/P D−4 ) − 1]−1 log(W A/P H )−1 . log(CS/P OP )−2 . DD772 · log(J F /J H MI N )−1 .04]−1 . 100[(P D/P D−4 ) − 1]−1 (CF − T F G − T F S)−1 . 100[(P D/P D−4 ) − 1]−1 (W A/P H )−1 . RSA−2 . AG1−1 . 1 2 Basic Set log log Additional 387 3 4 linear linear 5 6 7 8 9 log log log log log 10 log 11 12 13 linear linear log 14 16 17 18 22 23 24 25 26 27 log log log log linear linear linear linear log log 28 30 log linear log J F−1 . AG1. log J F−2 . (I H H /P OP )−2 . RMA−1 . log W F−3 . DD772−1 · log J F−2 log H F−2 . D721. 100[(P D/P D−4 ) − 1]−1 RM−1 . log P F−5 . 100[(P D/P D−4 ) − 1]−1 (W A/P H )−1 . 100[(P D/P D−4 ) − 1]−1 log(W A/P H )−1 . RSA−1 . RB−3 · I KF A. [Y D/(P OP · P H )]−1 . log P I M−2 . log(J F /J H MI N )−2 . log(L2/P OP 2)−2 . log[(X − F A)/P OP ]−1 . (BO/BR)−2 . (CD/P OP )−2 . DD772 · log J F−1 . DD772. 100[(P D/P D−4 ) − 1]−1 . log P F−3 V−2 . log(D2G + D2S). log W F−5 .04]−2 . (KK − KKMI N)−2 . log W F−2 . RS−3 . RSA−2 log(1+D5G). T F G + T F S log[CU R/(P OP · P F )]−1 . log[Y D/(P OP · P H )]−1 . log(P CN/P H )−1 . RS−3 . log U B−2 D794823 · P CM1−1 . Y−5 . log P F−6 .04]−3 . RMA−2 ·I H H A−1 . (P I H /P H )−1 .

388 APPENDIX A 3SLS First Stage Regressors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 constant (AA/P OP )−1 COG + COS (CD/P OP )−1 log(CN/P OP )−1 log(CS/P OP )−1 (1 − D1GM − D1SM − D4G)−1 EX log H F−1 (I H H /P OP )−1 log(I M/P OP )−1 log(J F /J H MI N )−1 (J G · H G + J M · H M + J S · H S)/P OP (KH /P OP )−1 (KK − KKMI N )−1 P CM1−1 100[(P D/P D−1 )4 − 1]−1 log P F−1 log P I M RB−1 RS−1 RS−2 T (T RGH + T RSH )/(P OP · P H−1 ) V−1 log W F−1 Y−1 Y−2 Y−3 Y−4 Z−1 U R−1 log[Y D/(P OP · P H )]−1 log[Y D/(P OP · P H )]−2 AG1 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 AG2 AG3 (KD/P OP )−1 log(CN/P OP )−2 log(L1/P OP 1)−1 log(L2/P OP 2)−1 log(L3/P OP 3)−1 log(LM/P OP )−1 log(W A/P H )−1 log(W A/P H )−2 log(W A/P H )−3 log(W A/P H )−4 RM−1 RM−2 RS−3 RS−4 RS−5 RD−1 (BO/BR)−1 log[MH−1 /(P OP−1 · P H )] log(MF−1 /P F ) log[CU R−1 /(P OP−1 · P F )] D794823 · P MC1−1 log[(Y S − Y )/Y S + .04]−1 I KF−1 − DELK · KK−1 log(J F−1 /J F−2 ) DD772 DD772 · log(J F /J H MI N )−1 DD772 · log J F−2 DD772 · T J J S−1 (I H H /P OP )−2 T XCR · I KF A RB−3 · I KF A .

9 30 57 71 77 81 125 127 RS Eq. The variable listed is the one that is put on the left hand side of the equation and “solved for.TABLES FOR THE US MODEL Table A. where monetary policy is exogenous. RSA Out BR MB AG MH U BR RS U BR Exog. in the standard version equation 125 is used to solve for the level of nonborrowed reserves (U BR): U BR = BR − BO (125) When. RSA Out MB MH AG M1 BR RS AG Exog.” Eq.30 MH RS BR MB AG M1 U BR RSA RS Exog. No. however. the level of nonborrowed reserves is set exogenously. the equation is rearranged and used to solve for total bank reserves (BR): BR = U BR + BO (125) The following shows the arrangement of the equations for each of the five monetary-policy assumptions. equation 30 is dropped and some of the other equations are rearranged for purposes of solving the model. Under alternative assumptions.8 Solution of the US Model Under Alternative Monetary-Policy Assumptions 389 There are five possible assumptions that can be made with respect to monetary policy in the US model. it is explained by equation 30—the interest rate reaction function. For example. In the standard version monetary policy is endogenous. RSA Out MB MH BR M1 U BR RS . MH Out BR MB AG M1 U BR RSA M1 Exog.

129. 9 1. 121 35. 126 .390 Table A. 2. 76. 73. 115. 31 67. 64. 90. 98. 82. 58. 61. 121. 54. 53. 80 70. 99. 69. 67. 68 34. 3. 115. 83. 60. 60. 67. 126 62 43. 128 17. 126 15 43. 64. 74. 121. 52. 69 66. 95. 115 13. 50. 37. 52. 115 73. 36. 54. 52. 104 60. 115. 72. 80 77. 74. 60. 61. 2. 65. 104. 55. 80 66. 61. 60. 36. 126 Exog 43. 80 64. 99 126. 2. 125 22. 37. 9 80. 99. 68. 77. 122. 68. 68. 80 29. 130 130 67. 4. 64. 14 58 59 12. 127.9 Cross-Reference Chart for the US Model Variable AA AB AF AG AG1 AG2 AG3 AH AR AS BF BG BO BR CCB CCF CCH CD CDA CF CG CN COG COS CS CU R D1G D1GM D1S D1SM D2G D2S D3G D3S D4G D5G D691 D692 D714 D721 Eq. 80 75. 115 86 Exog 27 Exog 27 Exog 27 33. 76. 83. 82. 99 126. 115. 77. 74 107 114 17. 104. 73 70. 7 80 19. 116 71. 51. 110 34. 26. 128 48. 129. 116 3 25. 118 15 100 43. 80 79. 61. 65. 77 47. 95. 113. 129 123 84. Exog Exog Exog Exog Exog Exog Exog Exog 18 Exog Exog Exog Exog Exog Exog Exog Exog 85 Exog 106 113 Exog Exog Exog Exog Exog Exog 82 84 83 129 131 130 14 100 Exog Exog APPENDIX A Used in Equation 30 21 21 64. 116 60. 98. 68. 89 73 70 77 Exog Exog Exog 66 75 79 55 56 22 57 Exog 21 Exog 3 Exog 68 25 2 Exog Exog 1 26 Exog 90 Exog 91 Exog Exog Exog Exog Exog Exog Exog Used in Equation 2. 51. 52 53. 72. 61. 78. 76. 80 34. 130 130 57 84. 56. 3. 61. 125 60. 49. 62. 83 67 65. 67. 9 1. 54 27 Variable D794823 D811824 D831834 DB DD772 DELD DELH DELK DF DI SB DI SBA DI SF DI SG DI SH DI SR DI SS DRS E EX EXP G EXP S FA F I ROW F I ROW D FIUS F I U SD G1 GDP GDP D GDP R GN P GN P D GN P R HF HFF HFS HG HM HN HO Eq. 99. 78. 65. 121 17. 53. 92 64. 100. 130 131 − 131 62. 51 35. 95. 3. 126 10. 89 80 80 19 29 73. 82. 64. 91. 80 1. 51. 127. 61.

30. 104 34 1. 61. 115. 2. 35. 83. 48. 76 I KH Exog 60. 65. 87 94. 64. 121. 98 30 43. 95. 54. 81 93 − 34. 8. 72 21. 95. 61. 121 64. 4. 98. 99 34. 42 32 33. 67. 82. 99 5. 104. 52. 61. 97. 61. 68. 68. 65 14. 78. 115. 68 IHH 4 34. 116 12. 115 INT G 29 64. 118. 33. 68. 52. 92 I KFA Exog 12 I KG Exog 60. 36. 87 86. 61. 104. 90. 115. 9. 110 32 35 36 37 . 68. 61. 60. 82. 105 I BT S 52 34. 74 10 16. 106. 81 68. 60. 99. 67. 81. 61. 95. 4. 89 Exog 60. 67. 126 43. 51. 65 IHHA Exog 4 I KB Exog 60. 72 IHF Exog 60. 65. 26. 27. 98. 82. 74 67. 31. 126 13. 76. 72. 83. 64. 81 71. 95. 38. 79. 87. 99. 6. 65. 67. 27. 65 IM 27 33. 50. 8. 120 − 61. 98 85 M1 MB MDI F MF MG MH MR MRS MS MU H P 2554 P CD P CGDP D P CGDP R P CM1 P CN P CS PD P EX PF P FA PG PH P I EB P I EF P I EH PIH PIK PIM PIV P OP P OP 1 P OP 2 P OP 3 P ROD PS P SI 1 P SI 2 P SI 3 P SI 4 81 71 Exog 17 Exog 9 Exog Exog Exog Exog Exog 37 122 123 124 36 35 33 124 391 57. 115. 119 Exog 31 40 61. 73 81 70. 61. 61. 77 66. 52. 37. 78. 61. 49.TABLES FOR THE US MODEL Exog 43. 83. 51. 71. 116 − 30 30 34. 65. 120 6. 43. 76 71. 115 67 − 60 61. 61. 74 INS Exog 65. 68. 110. 83 67 Exog 38 39 Exog 42 120 Exog Exog Exog 118 41 Exog Exog Exog Exog 18. 82. 98. 65. 126 3 4. 51. 98. 83. 68. 120 7. 7. 110. 71. 76. 115. 14 96. 76 INT F 19 64. 115 I N T ROW 88 64. 5. 113. 75. 53. 82. 68. 99. 47. 68. 76 10. 85. 99. 126 43. 88. 39. 76. 51. 61. 72. 85. 3. 98. 72 I KF 12 21. 78. 97 96. 52. 2. 76. 41. 67. 87 86. 59. 89 71. 95. 95. 61. 78. 64. 9. 27. 88. 26. 82. 91. 116 34. 17. 61. 27. 82. 64. 61. 82 1. 115 HS INT S IV A IV F IV H IV V H JF JG J H MI N JJ JJP JJS JM JS KD KH KK KKMI N L1 L2 L3 LAM LM Exog 20 117 Exog Exog 13 Exog 94 95 Exog 96 Exog Exog 58 59 92 93 5 6 7 Exog 8 64. 99. 64. 60. 3. 126 I BT G 51 34. 40. 85. 89 12 12 86. 7. 112 IHB Exog 60. 61. 115 I N T OT H Exog 64. 61. 85. 61. 76. 78.

78 128 2. 104. 23. 82. 99. 126 65. 28. 115 − − 11. 25. 25. 78. 82. 9. 61. 117 5. 111. 68. 115. 76. 29 22 107 114 65. 82. 78. 26. 8 10. 118 115 1. 115. 104. 94. 68. 101. 78. 68. 77 12. 70. 64. 106 Exog 82. 78. 54. 126 44 43. 115. 76. 90. 76. 106. 126 60 11. 16 47. 64. 82. 76. 76. 103. 115 112 43. 69. 68. 64. 74 65. 78. 115 12 28. 115. 110. 116 99 − 98 10 64 47. 65. 67. 4. 64. 74. 3. 126 43 − Exog 82. 78. 2. 76. 99 97 5. 99. 7. 48. 108 65. 99. 93. 33. 106 67. 12. 121. 43. 76. 78. 99 72. 68. 32. 63. 90. 4. 78. 115 67. 126 119 − 46 43. 25. 119 75. 115 17. 76. 112. 127 1. 78. 7. 29. 13. 17. 6. 109. 20. 69. 103 67. 103. 31. 68. 68. 82. 49. 91. 27 64. 68. 102. 109 105 43. 113 45 43. 67. 102 72. 78. 3. 113 SU R T T AU G T AU S T BG T BS T CG T CS T FA T FG T FS T HG T HS TI TPG T RF H T RF R T RGH T RGR T RGS T RH R T RRSH T RSH T XCR U UB U BR UR V WA WF WG WH W LDG W LDS WM WR WS X XX Y YD Y NL YS YT Z Exog Exog Exog Exog Exog Exog 102 108 Exog 49 50 47 48 Exog 101 Exog Exog Exog Exog Exog Exog 111 Exog Exog 86 28 128 87 63 126 16 APPENDIX A 72. 91. 76. 119. 126 75 − 79 − 67. 115 113 65. 104. 112 65. 27. 6. 53. 26 73 70 77 − 66 − 67. 24. 14. 76 5. 76. 99 48. 102 18. 115 18. 63 61 67. 67. 22. 72. 8 . 115 65. 87 65. 64. 13. 115 74. 82 11 10. 80 83 67. 83. 14. 19. 108 105 112 101. 64. 9. 115 29 105 64. 99. 46. 110. 68. 106 76. 106. 33. 7. 99. 76. 115 65. 67. 44. 19.392 P SI 5 P SI 6 P SI 7 P SI 8 P SI 9 P SI 10 P SI 11 P SI 12 P SI 13 P SI 14 P UG P US PX Q RB RD RECG RECS RET RM RMA RNT RS RSA SB SF SG SGP SH SH RP I E SI F G SI F S SI G SI GG SI H G SI H S SI S SI SS SR SRZ SS SSP ST AT ST AT P SU BG SU BS Exog Exog Exog Exog Exog Exog Exog Exog Exog Exog 104 110 31 Exog 23 Exog 105 112 Exog 24 128 Exog 30 130 72 69 76 107 65 121 54 Exog 103 Exog 53 Exog 109 Exog 74 116 78 114 Exog Exog Exog Exog 38 39 40 41 42 44 45 46 83 88 106 113 20. 78. 115. 115. 111. 109. 45.

Appendix B Tables for the ROW Model .

.

Pounds (mil. PO 22. AO Nigeria Algeria Indonesia Iran Iraq Kuwait Libya United Arab Emirates Israel Bangladish Singapore All Other 395 Quarterly Countries 1. PA 32. PH 33. MA 31. KO United States Canada Japan Austria France Germany Italy Netherlands Switzerland United Kingdom Finland Australia South Africa Korea Annual Countries 15. DE 17. ID 30. UK 11.) Aust. Dollars (mil. IA 37. IN 38.1 The Countries and Variables in the MC Model Local Currency Trade Share Equations Only U.) Phil. TH Belgium Denmark Norway Sweden Greece Ireland Portugal Spain New Zealand Saudi Arabia Venezuela Colombia Jordan Syria India Malaysia Pakistan Philippines Thailand Bel. Rupees (bil.) Pesetas (bil.) Guilders (bil. IS 43. IT 8. NE 9.) Swill Francs (bil.) Won (bil. Dollars (mil. FR 6. Francs (bil. IQ 39.) 34.) Schillings (bil.) Rand (mil.) Jor. BE 16.) Syr. BA 44. Pesos (bil. AU 5. SI 45.) Yen (bil. FI 12.) U.S. CO 27. Dollars (mil. SY 29.) Riyals (bil.) Bolivares (bil.) Den.) Col. NZ 24. Pesos (bil. AL 36.) Baht (bil. UA 42.K. IR 21. Francs (bil. NI 35.) Irish Pounds (mil.) . JO 28. SW 19.) Can.) Lire (bil. SA 25. Rupees (bil. GR 20.) Markkaa (mil. KU 40.Z.) Escudos (bil.) Pak.) N. Kroner (bil. AS 13. NO 18.) Fr. VE 26. Kroner (bil. ST 10. LI 41. Dollars (mil.TABLES FOR THE ROW MODEL Table B.) Ringgit (mil. US 2.) Drachmas (bil. Mark (bil. GE 7. CA 3. JA 4. Kroner (bil. Pounds (mil.) Nor.) D.) Ind. SO 14. SP 23. Dinars (mil.) Swe.

thousands Variables Determined by Identities: I-1. constant lc Balance of Payments.0 Merchandise Exports from the Trade Share Calculations. thousands Employment Population Ratio Peak to Peak Interpolation of JJ Labor Constraint Variable Potential Y Demand Pressure Variable Import Price Index. KMI N I-12. RS 8. end of period. lc Net Stock of Foreign Security and Reserve Holdings.0 Employment. V 1 I-5. V I-6. base year = 1. P Y 6. S I-7. αij P X$ X85$ P MP PW$ Trade share coefficients from trade share equations Export Price Index.0 Variables Determined by the Trade Share Calculations: L-1. I 4. I M I-2. 85 lc Consumption. constant lc Unemployment Rate Minimum Required Employment. base year = 1.0 World Price Index. A I-8. lc per $ Three Month Forward Rate. J J S I-16. constant lc Final Sales. 1985=1. C 3. Y S I-18. 85 $ Exchange Rate.0 Money Supply. Y 5. constant lc Total Exports (NIPA). 1985=1. 1985=1. Z I-17. lc per $ Export Price Index. J MI N I-14. J 14. J J I-15. percentage points Exchange Rate. W 13. lc Merchandise Imports from the Trade Share Calculations. 1985=1. lc per $ Capital Stock. lc Three Month Interest Rate. constant lc Fixed Investment. P M Total Imports (NIPA).0 . RB 9. percentage points Long Term Interest Rate. ZZ I-19. X I-4. E 10. M1 7.0 Nominal Wage Rate. M85$A I-9. L-4. thousands Labor Force—men. L2 Merchandise Imports. M 2. 1985=1. constant lc GDP Deflator. P X 12. EX I-3. EE I-10. constant lc Inventory Investment. 85 $ Import Price Index from the Trade Share Calculations. U R I-13. K I-11. L-3. constant lc Real GDP.396 APPENDIX B Table B.1 (continued) A Brief Listing of the Variables per Country Variables Determined by Stochastic Equations: 1. constant lc Inventory Stock. thousands Labor Force—women. F 11. L-2. constant lc Minimum Required Capital Stock. L1 15.

constant lc Import Discrepancy.TABLES FOR THE ROW MODEL Exogenous Variables: AF DEL EXDS E85 G I MDS JJP LAM MS M85$B MU H P M85 P OP P OP 1 P OP 2 P SI 1 P SI 2 P X85 ST AT T TT XS Notation: lc 85 lc constant lc 397 Level of the Armed Forces. thousands Ratio of (EE + EE−1 )/2 to E Ratio of P M to P MP P X in Base Year divided by P X in 1985 NIPA Statistical Discrepancy Time Trend Total Net Transfers. 85 lc Peak to Peak Interpolation of J J Peak to Peak Interpolation of Y /J Non Merchandise Imports. 85 $ Peak to Peak Interpolation of Y /K P M in Base Year divided by P M in 1985 Population. lc Non Merchandise Exports. thousands Depreciation Rate for the Capital Stock Export Discrepancy. millions Population of men. 85 lc Merchandise Imports from Countries other than the 44 in the Trade Share Matrix. 85 lc local currency 1985 local currency local currency in the NIPA base year . 85 lc per 85 $ Government Expenditures. thousands Population of women. 85 lc E in 1985.

[OECD data or (IFS90C or IFS90N)/P X.3. 1985 = 1.] Minimum amount of employment needed to produce Y in thousands. [J J /J J P . [A−1 + S.] Discrepancy between NIPA import data and other import data in 85 lc.] Consumer price index. [OECD data.3.3.] Depreciation rate for the capital stock (K). BOP data. end of period.0. [I M − P M85(M + MS). [See Section 3..3. I-7 exog 2 none exog 9 I-9 I-2 exog exog 10 exog 3 I-1 exog none 13 I-14 exog I-15 I-13 I-10 I-11 exog 14 15 1 exog I-8 exog Description Net stock of foreign security and reserve holdings. lc per $.] Import price index. and income (debit) in 85 lc. 1985 = 1.] Peak to peak interpolation of J J . [IFS66 or other 66 options. [(IFS64 or IFS64X)/100.2 The Variables for a Given Country in Alphabetical Order Variable A AF C CP I DEL E EE EX EXDS E85 F G I IM I MDS IP J JJ JJP JJS J MI N K KMI N LAM L1 L2 M MS M85$A M85$B Eq.] Difference between total merchandise imports and merchandise imports from the trade share matrix in 85 $ (i.] Exchange rate.] Government purchases of goods and services in constant lc. rate per quarter or year. 85 lc per 85 $.] Peak to peak interpolation of Y/J .] Total exports (NIPA) in constant lc. 1985 = 100.] Capital stock in constant lc. [See Section 3.] Exchange rate.3. [OECD data or IFS98C/P M. services.] Other goods.] Minimum capital stock needed to produce Y in constant lc.0. [OECD data. Base value of zero used for the quarter prior to the beginning of the data. [IFSB.] Employment population ratio. end of quarter.] Ratio of J J to J J P . [OECD data or IFS96F/CP I .3. [OECD data. [See Section 3.No.3.] Merchandise imports (fob) from the trade share matrix in 85 $. in lc. [See Section 3. [(IFS77AED·E)/P M. [OECD data or IFS67. [IFSRF. .] Three month forward rate. [M/E85 − M85$A.] Personal consumption in constant lc. [EX − P X85(E85 · X85$ + XS). [See Table B. [OECD data or (IFS91F or IFS91FF)/P Y .] Gross fixed investment in constant lc. imports from countries other than the 44 in the trade share matrix).] Peak to peak interpolation of Y/K. [IFS34 or IFS34.0.e. [Y/MU H .015 per quarter. [IFS75/100. [IFSAE.] Total imports (NIPA) in constant lc.3.3.] Level of the armed forces in thousands.] Industrial production index. See Section 3.] Labor force of women in thousands. average for the period. [Y /LAM. [IFSRF in 1985.3. [J /P OP .] Money supply in lc.] Total employment in thousands. [OECD data or IFS93/P Y . [See Table B. [IFS71V/P M. 1985 = 1.] Labor force of men in thousands.] Total merchandise imports (fob) in 85 lc. [..3.] Discrepancy between NIPA export data and other export data in 85 lc.3. lc per $.B.] E in 1985.] Import price index from DOT data.398 APPENDIX B Table B.060 per year. lc per $.] MU H M1 PM P MP exog 6 I-19 L-3 .

] Stock of inventories. for annual data. [OECD data or IFS93I/P Y .] Other goods.] Final sales in constant lc.] Export price index.] Merchandise exports from the trade share matrix in 85 $.] Potential value of Y .P or IFS99A. 1 in 1952.] Unemployment rate. [(Y S − Y )/Y S. [V−1 + V 1. 2 in 1952.TABLES FOR THE ROW MODEL P M85 P OP P OP 1 P OP 2 P SI 1 P SI 2 PW$ PX P X$ P X85 PY RB RS S ST AT T TT UR V exog exog exog exog exog exog L-4 11 L-1 exog 5 8 7 I-6 exog exog exog I-12 I-5 399 V1 W X XS X85$ Y YS Z ZZ I-4 12 I-3 exog L-2 4 I-17 I-16 I-18 P M in the NIPA base year divided by P M in 1985. [(E85 · P X)/E.. [See Table B. [For quarterly data.4.P or IFS99B. [IFS74/100.0 in the NIPA base year.] P X in the NIPA base year divided by P X in 1985. and transfers in lc.] Real GDP or GNP in constant lc. [See Table B.] Demand pressure variable. end of period. Population in millions. BOP data. [IFS99Z. NIPA = national income and product accounts. Base value of zero was used for the period (quarter or year) prior to the beginning of the data.] World price index.] Long term interest rate. [min(0.4. . [IFS61 or IFS61A. services. 2 in 1953. percentage points.P.] Three month interest rate. [OECD data or IFS99A. [IFS60 or IFS60B or IFS60C or IFS60X.0.] Inventory investment in constant lc. [(L1 + L2 − J )/(L1 + L2 − AF ).] Labor constraint variable.7. Saving of the country. Balance of payments on current account. [OECD data.] Total net goods. 1 − J J P /J J ).R or IFS99B. $/85$.] Nominal wage rate.] Export price index.] Population of men in thousands.] Time trend.R.] lc = local currency.] Population of women in thousands.2. [See Table B. etc. $/85$. in constant lc. equals 1. and income (credit) in 85 lc. IFSxx = variable number xx from the IFS data.6. [IFS65 or IFS65EY.1. GDP or GNP deflator.] Statistical discrepancy in constant lc. 1985 = 1. 1 in 1952. etc. [(IFS77ADD·E)/P X.] [P M/P MP . [LAM · J J P · P OP . [OECD data or (IFS99B/IFS99B. [OECD data. percentage points. services.] Total net transfers in lc.] [[(EE + EE−1 )/2]/E. [See Table B. [Y − C − I − G − EX + I M − V 1. 1985 = 1.0. [Y − V 1.

log E 10. Z 7. log( P OP ·P Y ) APPENDIX B Explanatory Variables cnst. ZZ or J J S. I-16. log(C/P OP ) 3. . I-6. RSU S : P CP Y = 100[(P Y /P Y−1 )4 − 1] and P CM1 = 100[(M1/M1−1 )4 − 1] cnst. [A/(P Y · Y S)]−1 . I-14.25 · log[(1 + RS/100)/(1 + RSU S/100)] log P Y − log(P W $ · E) cnst. U R or J J S or ZZ. I-3. log P Y−1 . RS 8. RB−1 − RS−2 . X. log[M1/(P OP · P Y )]−1 or log[M1−1 /(P OP−1 · P Y )]. ZZ or J J S cnst. log Y . Y−1 . [A/(P Y · Y S)]−1 cnst. V−1 cnst. 1 − J J P /J J ) . log(M/P OP )−1 . I-4. log Y−1 cnst. log(Y /P OP ). I−1 . Y . . RS−1 . log P Y . RS or RB cnst. log(L2/P OP 2) Identities I-1. I-10. T .25 · log[(1 + RS/100)/(1 + RSU S/100)] log EE. log EGE − log(P Y /P Y U S). log(W/P Y ). log( P WX ) $·E 12. I 4. log W 13. log(L2/P OP 2)−1 . 1. log P Y M1 6. I-13. T . I-11. I-12. I-5. P CP Y . RS − RS−2 . Y 5. I-2. [A/(P Y · Y S)]−1 cnst. log W . log(W/P Y ). log(Y /P OP ). [A/(P Y · Y S)]−2 . log P Y−1 . log(J /J MI N )−1 . log(L1/P OP 1)−1 . log(P Y /P M). I-9. T . Z cnst. log(P Y /P Y U S) − log E−1 . I M = P M85(M + MS) + I MDS EX = P X85(E85 · X85$ + XS) + EXDS X = C + I + G + EX − I M + ST AT V1 = Y −X V = V−1 + V 1 S = P X(E85 · X85$ + XS) − P M(M + MS) + T T A = A−1 + S M85$A = M/E85 − M85$B EE = 2 · P SI 1 · E − EE−1 K = (1 − DEL)K−1 + I KMI N = Y /MU H U R = (L1 + L2 − J )/(L1 + L2 − AF ) J MI N = Y /LAM J J = J /P OP J J S = J J /J J P Z = min(0. log P M. T . RS−1 − RS−2 cnst. log W−1 . RS or RB. RS. log F P 11. log J 14. log(C/P OP )−1 . I-7. log(Y /P OP ) cnst. log(L1/P OP 1) 15. RS or RB. I-15. K−1 . P CM1−1 . cnst. I-8.400 Table B.3 The Equations for a Given Country Stochastic Equations LHS Var. RB − RS−2 9. log(M/P OP ) 2.

Y S = LAM · J J P · P OP ZZ = (Y S − Y )/Y S P M = P SI 2 · P MP 401 Variables Explained When the Countries are Linked Together (Table B.TABLES FOR THE ROW MODEL I-17. I-18. I-19. L-3.4) L-1 L-2. L-4. P X$ X85$ P MP PW$ .

Once the quarterly values have been computed from the trade share calculations. i = 1. · · · . Construction of αij : The raw data are: XX$ij X$i Merchandise exports i to j in $. · · · . j = 1. L-4. · · · . i = 1. This summation also excludes SA and VE. the annual values of X85$i that are needed for the annual models are taken to be the sums of the quarterly values. M85$Ai = X85$i αij = 44 j =1 XX85$j i . · · · .] Total merchandise exports (fob) in $. · · · . j = 1. 44 [DOT data. and all these calculations are quarterly. · · · . 45 Linking of the Annual and Quarterly Data Quarterly data exist for all the trade share calculations. · · · . j = 1. P MPi is not computed if Ei is missing or E85i is missing. i = 1. 45 j =1 XX85$ij . i = 1. · · · . P W $i = ( 33 j =1 P X$j X85$j )/( 33 j =1 X85$j ).4 Equations that Pertain to the Trade and Price Links Among Countries L-1. M85$Ai . . For each of these three variables the predicted value for a given quarter was taken to be the predicted annual value multiplied by the ratio of the actual quarterly value to the actual annual value. i. i = 1.] The constructed variables are: XX$i45 = X$i − i = 1.402 APPENDIX B Table B. 33 XX85$ij = XX$ij /P X$i . P MPi = (Ei /E85i ) 44 j =1 αj i P X$j . P X$i = (E85i /Ei )P Xi . 33 i = 1. Similarly. 33 An element in this summation is skipped if P X$j is missing or X85$j is missing or j=i. · · · . 33 = XX85$ij /M85$Aj . This means in effect that the distribution of an annual value into its quarterly values is taken to be exogenous. 44 L-2. · · · . An element in this summation is skipped if αj i is missing or P X$j is missing. 44. · · · . 44. 33 L-3. Feeding into these calculations from the annual models are predicted annual values of P X$i . which are the oil exporting countries among the 33. · · · . the annual values of P MPi and P W $i are taken to be the averages of the quarterly values. X85$i = 45 j =1 αij M85$Aj . 45 XX85$ij is missing if XX$ij is missing or P X$i is missing. i = 1. and Ei . · · · . 45 i = 1. 33 [IFS70/E. 44 j =1 XX$ij . i = 1.

M. = DEL3 · P XU S . S. P EX is determined as P SI 1 · P X. P M. This is the same as equation I-2 for the other countries. This is the same as equation I-19 for the other countries. = (X85$U S + XSU S + EXDSU S )/1000.2. M85$A. = MU S − M85$BU S . = 1000 · I M − MSU S − I MDSU S . = P SI 2U S · P MPU S . M85$BU S . This is the same as equation I-8 for the other countries. EX is an exogenous variable in the US model by itself.2. I MDSU S is constructed from the data as 1000 · I M − MU S − MSU S and is taken to be exogenous. = P XU S (X85$U S + XSU S ) − P MU S (MU S + MSU S ) + T TU S . = DEL4 · P MU S . The P X variable here is not the same as the P X variable in Appendix A. MSU S .2. P SI 2. DEL4 is constructed from the data as P I M/P MU S and is taken to be exogenous. T TU S . EXDSU S is constructed from the data as 1000 · EX − X85$U S − XSU S and is taken to be exogenous. T T . P MP . MS. Variable X85$U S P MPU S P W $U S P XU S P EX Determination Determined in Table B.TABLES FOR THE ROW MODEL Table B. P W $. This equation is dropped when the US model is linked to the ROW model. EXDSU S .4 Determined by equation 132 in the US model. In the US model by itself. I MDSU S .4 Determined in Table B. P SI 2U S . M85$B. This equation is equivalent to equation 11 for the other countries. XS. and X85$. P I M is an exogenous variable in the US model by itself. which is equation 32 in Table A. P X (= P X$). but endogenous when the US model is linked to the ROW model. EX and P I M are exogenous in the US model by itself.5 Links Between the US and ROW Models 403 The data on the variables for the United States that are needed when the US model is imbedded in the MC model were collected as described in Table B. These variables are (with the US subscript dropped): EXDS.4 Determined in Table B. See the discussion in Section 9. DEL3 is constructed from the data as P EX/P XU S and is taken to be exogenous. P MU S PIM EX MU S M85$AU S SU S Note: The new exogenous variables for the US model when it is linked to the ROW model are DEL3. . and XSU S . This is the same as equation I-6 for the other countries. DEL4. I MDS. This is the same as equation I-1 for the other countries.

The key assumption is that the four quarterly changes within the year are the same: y1t − y4t−1 = y2t − y1t = y3t − y2t = y4t − y3t = { δ2 for t = 1. and so on. λy2 = a2 . 4). 2. one can solve for y40 and δ2 in terms of a1 and a2 : y40 = (13/32)a1 − (5/32)a2 δ2 = (a2 − a1 )/16 Using y40 and δ2 . one can solve for δ3 in terms of a3 and y42 : δ3 = (a3 − 4y42 )/10 Using y42 and δ3 . 2 δt for t ≥ 3 (ii) Given i and ii for t = 1. · · ·. Given y42 from these calculations and given i and ii for t = 3. One can then solve for δ4 in terms of y43 and a4 . Then: y1t + y2t + y3t + y4t = λyt where λ={ 1 for flow variables (at quarterly rates) 4 for stock variables and price variables (i) Assume that the annual data begin in year 1. λy3 = a3 . Note: The annual population data that were collected for the model are mid year estimates.6 The Procedure Used to Create Quarterly Data from Annual Data Let yt be the (observed) average value of the variable for year t. In order to apply the above procedure to these data. one can then construct quarterly data for years 1 and 2 using ii. 3. the assumption was made that the average value for the year equals the mid year value. one can then construct quarterly data for year 3. 2. and let yit be the (unobserved) average value of the variable for quarter i of year t (i = 1. . and let λy1 = a1 .404 APPENDIX B Table B.

2 and from MS$ and XS$ above that MS$ = (P M · MS)/E XS$ = (P X · XS)/E Note also from Table B. When only annual data on X$ were available and quarterly data were needed. where λ is the last observed annual value of X$ /X$. then S$ and T T $ were constructed as: (i) S$ = X$ + XS$ − M$ − MS$ + P T $ + OT $ T T $ = S$ − X$ − XS$ + M$ + MS$ (ii) where S$ is total net goods. for P T $ (where λ is the last observed annual value of P T $/X$). [IFS77AFD.] Merchandise imports (fob) in $. and income (credit) in $.TABLES FOR THE ROW MODEL Table B. and transfers in $ (balance of payments on current account) and T T $ is total net transfers in $. Similarly for MS$.] Merchandise exports (fob) in $.] Private unrequited transfers in $. Similarly for MS$ (where λ is the last observed annual value of MS$/M$. services. [IFS77AGD. from equations ii–vii. When no data on M$ were available. BOP data. BOP data. BOP data. data on S and T T were constructed as: S = E · S$ TT = E ·TT$ • Note from MS and XS in Table B. and income (debit) in $. [IFS77ABD.7 Construction of the Balance of Payments Data: Data for S and T T The relevant raw data variables are: M$ M$ X$ X$ MS$ XS$ PT$ OT $ Merchandise imports (fob) in $. interpolated quarterly data were constructed using M$. services. services. Equations i and ii were then used to construct quarterly data for S$ and T T $. the equation for S can be written S = P X(E85 · X85$ + XS) − P M(M + MS) + T T which is equation I-6 in Table B. BOP data. Similarly for XS$ (where λ is the last observed annual value of XS$/X$). where λ is the last observed annual value of M$ /M$. and for OT $ (where λ is the last observed annual value of OT $/X$). • When only annual data on M$ were available and quarterly data were needed.2 that M$ = (P M · M)/E X$ = (E85 · P X · X85$)/E Therefore. [IFS77AAD. BOP data. [IFS70/E.] Other goods. then X$ was taken to be λ · X$. (v) (vi) (vii) (vii) (iii) (iv) .) When no data on X$ were available. then M$ was taken to be λ · M$.] Merchandise exports (fob) in $. interpolated quarterly data were constructed using X$.] Other goods. [IFS77ADD.3. P T $. Similarly for XS$. • After data on S$ and T T $ were constructed. [IFS77AED. BOP data.] 405 • When quarterly data on all the above variables were available. and OT $.] Official unrequited transfers in $. [IFS71V/E.

.

” Econometrica. . Donald W. 50.Bibliography Allen. Donald W. “Exactly Median-Unbiased Estimation of First Order Autoregressive/Unit Root Models. “A Portfolio Approach to International Capital Flows. 82-93.” Journal of International Economics. and Ray C. Almon.” International Economic Review. 615-640. 33. Andrews. 1994. 67. 1989. Shirley. 30.” American Economic Review. Andrews. 1982.” American Economic Review... “The Two Stage Least Absolute Deviations Estimators. Amemiya. “Time-to-Build and Aggregate Fluctuations: Some New Evidence. 61. 1965. . “The Distributed Lag Between Capital Appropriations and Expenditures. 139-165. “Optimal Tests When a Nuisance Parameter is Present Only Under the Alternative. and Werner Ploberger.” Econometrica. and Herschel I. Barro.” Review of Economic Studies. 101-115. Robert J.. 135-160. Yale University. “Inference in Nonlinear Econometric Models with Structural Change. “Unanticipated Money Growth and Unemployment in the United States. “A General Disequilibrium Model of Income and Employment. Kenneth.” in Cowles Fiftieth Anniversary. K. 3. Grossman. Sumru. “Cowles in the History of Economic Thought.” Econometrica.. 889-920. Polly Reynolds.” Econometrica. Takeshi. Altug. 1991. 1971. 1977. K. Robert J. 178-196. Fair. 55. 1973. Arrow. 689-711. 61. 62. 1993. Barro. 1988.

“Econometric Issues in the Analysis of Equilibrium Models.. 1974. Olivier J. Studies in International Finance. The American Business Cycle: Continuity and Change. 53-67. Pagan. 53. Blanchard.” Brookings Papers on Economic Activity.” Review of Economic Studies. Finn. “Alternative Explanations of the Money-Income Correlation.. Canova.T Press. ed. F. 1986.. “Are Business Cycles All Alike?” in Robert J. Taylor.. “Econometric Evaluation of Linear Macro-Econometric Models. Branson. . Watson. and Martin Eichenbaum. Yock Y. eds. Martson. and A. Quebec City. International Money Markets and Flexible Exchange Rates. “Statistical Estimation and Testing of a Real Business Cycle Model.” Discussion Paper 24. 27-50.” in Jagdeep Bandhari. 1991. and Mark W. Amsterdam: North-Holland. eds.. 671690. “A Simple Macroeconomic Model with a Government Budget Restraint. Nicholas. Black. Carl F.” in Karl Brunner and Allan H. Blinder. Boston: Federal Reserve Bank of Boston. September. ed. 76. Meltzer. 32. Lawrence J. “International Capital Mobility in the Coordination of Monetary Policy Rules. 1986. Gregory C.. M.” paper given at the Canadian Econometric Study Group meeting. “Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations. Exchange Rate Policy under Uncertainty. International Aspects of Stabilization Policies. Chow. Cambridge: M. R. Chong. August 30. 1986.. 1981.R. 1968.” in A. Real Business Cycles. Real Exchange Rates and Actual Policies. Institute for Empirical Macroeconomics. Hendry. “Retail Inventory Behavior and Business Fluctuations.” Journal of Political Economy. and David F. 1991. Princeton: Princeton University Press.” mimeo.I. No. and R. Herring. Alan. Carloyzi. Christ.... William H. 1990. Christiano. and John B. Chicago: University of Chicago Press.. Stanley W. “Stocks and Flows in International Monetary Analysis. Ando. 1985. 443-505. Ben S. Gordon.408 BIBLIOGRAPHY Bernanke.. 1973.

Chicago: Rand McNally & Company. DeLong. 1975. . 1031-1043. 1974.. Phillip. Birth and Fortune: The Impact of Numbers on Personal Welfare. Flexible Exchange Rates and Macroeconomic Equilibrium. James S. Summers. Claassen and P. 369-391. “Scoring the Leading Indicators.” Journal of Money.. Cumby. 333-355. 21. P. Easterlin.” Journal of Business. and Maurice Obstfeld.” in E. Cooper. Brechling. “The Keynesian Counterrevolution: A Theoretical Appraisal. Russell. John Huizinga. 1965. “Several Tests of Model Specification in the Presence of Alternative Hypotheses. Nelson.” in F.. The Theory of Interest Rates. Rudiger. and Glenn D. and Edwin Kuh. NorthHolland Publishing Co. “Is Increased Price Flexibility Stabilizing?”.. Chicago: Rand McNally & Company. Francis X. The Brookings Model: Some Further Results. 1969. Dornbusch. and Banking. American Economic Review. 1981. eds. Lawrence R. Amsterdam: North-Holland. Credit. “The Ex-Ante Prediction Performance of the St. Klein. and Charles R. Hahn and F. 1-32. . . J. Rudebusch.. eds. The Short-Run Demand for Workers and Hours. “Two-Step Two-Stage Least Squares Estimation in Models with Rational Expectations. Diebold. 1965. J. 62. Ballinger Publishing Co. Bradford and Lawrence H. 1989.” Journal of Econometrics. 1976. Robert W.. 1987. 7.BIBLIOGRAPHY 409 Clower. Robert E. “Capital Mobility. 1969. 40. 781-793. Recent Developments in International Monetary Economics. Gary Fromm. MacKinnon. 2nd ed. Duesenberry.. Davidson. Volume I: The Theoretical Model. H. Fair. eds.. London: Macmillian. 1983. 78. Louis and FRB-MIT-Penn Econometric Models and Some Results on Composite Predictions. 1986. Salin. 261-278. The Brookings Quarterly Econometric Model of the United States. Chicago: University of Chicago Press.. Richard A.” Econometrica. and James G. A Model of Macroeconomic Activity. R.. Ray C.

” The American Economic Review. . 25-46. 1979a. 353-370. 1980b. Belongia and Michelle R. 1988a. 1987. 1980a.” Journal of International Economics. 1992.” Econometrica. Federal Reserve Bank of St. The Business Cycle: Theories and Evidence. 69. 21. 1976. 239-245. “Excess Labor and the Business Cycle.” Journal of Monetary Economics.” Econometrica. 539–552. . “The Cowles Commission Approach. . eds. . 313-332. “The Production Smoothing Model is Alive and Well. MA: Harvard University Press. . “Estimating the Uncertainty of Policy Effects in Nonlinear Econometric Models. . 1984.410 BIBLIOGRAPHY . Louis. 1989. 1979b. 1978. “A Model of the Balance of Payments. 9.” International Economic Review. 75. 22. 1993a.” in Michael T. . “Optimal Choice of Monetary Policy Instruments in a Macroeconometric Model.” The Quarterly Journal of Economics. . 355-378. Volume II: The Empirical Model. 23. 75. 1165-1179. 1988b.” Journal of Monetary Economics.” The Review of Economics and Statistics. and Analysis of Macroeconometric Models. Estimation. “An Analysis of a Macro-Econometric Model with Rational Expectations in the Bond and Stock Markets. “Inflationary Expectations and Price Setting Behavior. 8-18. . “The Sensitivity of Fiscal-Policy Effects to Assumptions About the Behavior of the Federal Reserve. 46. “Estimating the Expected Predictive Accuracy of Econometric Models.” The Review of Economics and Statistics. Specification. . 1985. 473-480.. Ballinger Publishing Co. . Real Business Cycles Theories. . “Sources of Economic Fluctuations in the United States.” The American Economic Review. 69. . Cambridge. Garfinkel. 103. 301-315. and New Keynesian Economics. . “International Evidence on the Demand for Money. 1381-1391. 133-147. 48. A Model of Macroeconomic Activity.

“The Informational Content of Ex Ante Forecasts. . “Labor and Output Over the Business Cycle: Some Direct Evidence. Dominguez.” The American Economic Review. and Robert J. “Computing Median Unbiased Estimates and Examining Asymptotic Distribution Accuracy in Macroeconometric Models. Watson.” Oxford Economic Papers. . 75. Fair. 5. 1990.” The Review of Economics and Statistics. Parke. Indicators.” Journal of Applied Econometrics. Fair. 80. Ray C. and John B.BIBLIOGRAPHY 411 . 157-176. 1993c.C. and Kathryn M. Board of Governors of the Federal Reserve System. Business Cycles. Washington. Ray C. 83. . 287-293. 1169-1185. “Full Information Estimation and Stochastic Simulation of Models with Rational Expectations. “The Fair-Parke Program for the Estimation and Analysis of Nonlinear Econometric Models. “Solution and Maximum Likelihood Estimation of Dynamic Rational Expectations Models. 1994b. “Effects of the Changing U. 1993b.. Chicago: The University of Chicago Press.” in James H. D. January. 1989.” Econometrica.” The American Economic Review. 71.. 375-389. . Stock and Mark W.. Jon A.. Flow of Funds Coded Tables.S. “Is Monetary Policy Becoming Less Effective?” mimeo. 81.” mimeo.” The American Economic Review. and William R. 20551. 1993d. 325-331. 1992. Ray C. 638-655. . . 169-190. Fair. 381-392. 1993. 1991.” American Economic Review.” mimeo. “Testing Macroeconometric Models. Age Distribution on Macroeconomic Equations. “Estimating Event Probabilities in Macroeconometric Models using Stochastic Simulation. 1994a. 1985. 51. eds. and James L... Taylor. Medoff. 1276-1294. 1990. “Testing the Rational Expectations Hypothesis in Macroeconometric Models. 1983. Fay. 45. “Comparing Information in Forecasts from Econometric Models. . . and Forecasting. Shiller.

“Generalized Instrumental Variables Estimation of Nonlinear Rational Expectations Models. 45. 472-492. 228-248.” American Economic Review. 1993. Johnson. Lars Peter. Hall. 1978. Prices. 1982.” Journal of Political Economy. 33-61. 1986. and Kenneth Singleton. Forecasting Economic Time Series. 1977. Benjamin M. 674-688.” American Economic Review. Jean-Michel. “Stochastic Implications of the Life-Cycle Permanent Income Hypothesis: Theory and Evidence. and Paul Newbold...412 BIBLIOGRAPHY Frenkel. and Dale W. Granger. Gordon. . 86. . “Oil and the Macroeconomy since World War II. 2nd ed.. Jacob A. 1269-1286. Jacob A. Business Cycles.” Econometrica. New York: Academic Press. Girton. 1982. “Large Sample Properties of Generalized Method of Moments Estimators. Robert E. “The Role of Consumption in Economic Fluctuations.” Econometrica. 1029-1054.. Lance. “Portfolio Equilibrium and the Balance of Payments: A Monetary Approach. “Financial Capital Movements and Central Bank Behavior in a Two-Country Portfolio Balance Model. Henderson. and Harry G. 91. 50. The Monetary Approach to the Balance of Payments. 1986. 65. The American Business Cycle: Continuity and Change.. 82. 1975. Indicators. Kuttner.” Journal of Monetary Economics. 1976.. Toronto: University of Toronto Press. James D. J. Robert E. and Kenneth N.. Hall. eds. 1992.” in Robert J. “Why Does the Paper-Bill Spread Predict Real Economic Activity?” in James H.” Econometrica. ed.” Journal of Political Economy. 535-572. Chicago: University of Chicago Press. Stock and Mark W. Chicago: The University of Chicago Press. and Interest Rates. Hamilton. “Money. 1983.. Rodrigues. Grandmont. 2. and Carlos A. Frenkel. 213249. and Forecasting. 50. Friedman. Watson. Income. 1976. eds. 971-989.. “Temporary General Equilibrium Theory.. Hansen.. Clive W. .

Stephen R. 1984.” Scandinavian Journal of Economics. Leonid. 876-886. Hendry. Prescott. Adrian R. Klein. F. 1950.. “Least-Squares Bias in Time Series. Kouri. . Lawrence R.BIBLIOGRAPHY 413 Heckman. Statistical Inference in Dynamic Economic Models. Denis Sargan. “Time To Build and Aggregate Fluctuations. “Haavelmo and the Birth of Modern Econometrics: A Review of The History of Econometric Ideas by Mary Morgan. Inc. Krane. King.” Econometrica.. K. 20. 10. Kydland. 280-304. 1023-1100.” in Z. 2.. . 3-33. Hurwicz. 90. Amsterdam: North-Holland. Amsterdam: North-Holland. Intriligator. Pagan. ““Sectoral Shifts and Cyclical Unemployment. Koopmans. An Econometric Model of the United States 1929-1952. James J. . and Arthur S.” Journal of Econometrics... C. “On the Formulation of Empirical Models in Dynamic Economics. 40.” 1988. 1342-1370. Cowles Monograph No. “Production Smoothing Evidence from Physical-Product Data.. 1976. D. David F. New York: Wiley. “Is Increased Price Flexibility Stabilizing?: Comment. 78. 1989. and J.. 1992.” Oxford Economic Papers. 365-383. Lilien. Braun. David F. ed.C.. Goldberger. Griliches and M. American Economic Review.” in T.. 19211941. “The Encompassing Implications of Feedback Versus Feedforward Mechanisms in Econometrics.” mimeo. “Dynamic Specifications. 1955. 777-793. 132-149.” Journal of Economic Literature.. and Steven N. eds. Economic Fluctuations in the United States. “The Exchange Rate and the Balance of Payments in the Short Run and in the Long Run. Handbook of Econometrics. 30.E.” Journal of Political Economy. 1982. 50. and Jean-Francois Richard. Penti J. 1950. Cowles Commission Monograph No.. Spencer D. Hendry. 11. David M. 1982. 267-272. 1982.. and E. New York: John Wiley & Sons. 1988.

” Quarterly Journal of Economics.H. Myhrman. T.” Econometrica. “The Methodology of Macroeconometric Model Comparisons. Dale T. Amsterdam: North-Holland. 1985. Topel. “Rational Expectations and the Natural Rate Hypothesis: Some Consistent Results. N. and Robert H. New York: Blackwell. Mizon. eds. 370-379. Julio J. ““Intertemporal Substitution in Macroeconomics. Robert E. Statistical Methods of Econometrics. eds.” in K. . 225-251. 167-211. N. Gregory. Grayham and Richard. Jean-Francois. Murphy. 1970. “The Encompassing Principle and Its Application to Testing Non-Nested Hypotheses. 3..” Journal of Business and Economic Statistics.. Johan. Mankiw.. and Laurence H. 1991.. . 1976. 1970. Kmenta and J. 1976. 44.. 1991. 100. Yale University.” in Cowles Fiftieth Anniversity. and David Romer. McNees. Salin. 397-442. 1981. Mankiw. “A Theory of Wage and Employment Dynamics. The Phillips Curve and Labor Markets. Ramsey. 1976. B. Cambridge: The MIT Press... Models of Business Cycles. S.” in J. Gregory. Amsterdam: North-Holland. Claassen and P. Malinvaud. Phelps et al. Volumes 1 and 2. 43-52. Summers. eds. Edmond. Amsterdam: North-Holland.” in E. 1986. “Balance-of-Payments Adjustment and Portfolio Theory: A Survey. Meltzer. Stephen K. Jr. ““Estimation and Inference in Two-Step Econometric Models. New York: Norton. Brunner and A. Kevin M. Mortensen... “Econometric Methodology at the Cowles Commission: Rise and Maturity. 54. 1985.” Econometrica. 203-237. Recent Developments in International Monetary Economics.. Rotemberg. “Econometric Policy Evaluation: A Critique.414 BIBLIOGRAPHY Lucas. Amsterdam: North-Holland. B. . Microeconomic Foundations of Employment and Inflation Theory. 657-678.. McCallum. New Keynesian Economics. 1987. Large Scale Macroeconometric Models. eds.” in E.

1970. Patinkin. 1970. William R. Orcutt. 1968-70. 48. 62. 77.” Journal of the Royal Statistical Society. Microeconomic Foundations of Employment and Inflation Theory. “Optimal Price Policy Under Atomistic Competition. S.” Econometrica.BIBLIOGRAPHY 415 Nelson.” Econometrica. Edmund S. New York: Harper & Row. 1969. 87-92. et al. Interest. Jr. and Sidney G. Poole.. 37.. Phelps.. Winokur.. Christopher A. New York: Norton. Edmund S. Shiller. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model. 1-48..” Econometrica.. . Parke. 1980.. S. 1-45. Don. London: Macmillian. 81-96. 1970. 309-337. Phelps. 1971. 1929. 2nd ed. Sims.” in E. Series B.” American Economic Review. “The Personal Tax Surcharge and Consumer Demand. Microeconomic Foundations of Employment and Inflation Theory. 1948.. and Herbert S. Okun.. Microeconomic Foundations of Employment and Inflation Theory... 167-204. 124-166. 1970. “First Order Autoregression: Inference. 1919-1932. 902917. Industrial Fluctuations. William.” American Economic Review. Charles R. 1982.” Brookings Papers on Economic Activity. “Ultimate Sources of Aggregate Variability. New York: Norton. Phelps et al. 1956.. . Guy H. “Macroeconomics and Reality.. New York: Norton. Arthur C. “A Study of the Autoregressive Nature of the Time Series Used for Tinbergen’s Model of the Economic System of the United States. 197-216.” in E. Robert J.S. Estimation. 1987. 50. “The Prediction Performance of the FRB-MITPenn Model of the U. Jr. 84. Arthur M.. and Prices. Edmund S.. “Money Wage Dynamics and Labor Market Equilibrium. “An Algorithm for FIML and 3SLS Estimation of Large Nonlinear Models.. Phelps et al. 1-14. Economy. Pigou. Money. and Prediction. Winter. Phelps. 1972.” Quarterly Journal of Economics.

Stoker. Taylor. Tobin. 178D. Credit.” in James H. 1989.. “Optimal Choice of Monetary Instrument in a Linear Economic Model with Stochastic Coefficients. 72.” Journal of Business and Economic Statistics.416 BIBLIOGRAPHY . 1975. James. 1961. P. 1986. “Simple Tests of Distributional Effects on Macroeconomic Equations. 7. “Solving Nonlinear Stochastic Growth Models: A Comparison of Alternative Solution Methods. Business Cycles. “Economic Theory and BEA’s Alternative Quantity and Price Indexes. Thomas M. 1990.. 94. John Fitzgerald Kennedy School of Government.” Journal of Money. 49-52..” Survey of Current Business. and Banking. White. “A Nine-Variable Probabilistic Macroeconomic Forecasting Model. von zur Muehlen. A. “Policy Analysis with a Multicountry Model.” Journal of Political Economy. April. Tinsley. George J. . James H. and Forecasting.. 1980. Watson. Triplett. “The Economics of Information. 69. 8. 763-795. 1939. Tinbergen. 1993. 1982. Turnovsky.. Band 1 (VVW. Stock and Mark W. Stigler..” Mimeo. Harvard University. Banken und Versicherungen. 2. John B. “Conditional Intermediate Targeting. eds.” Discussion Paper No. Statistical Testing of Business Cycle Theories. 1992. 1983. and Mark W. 51-80.” NBER Working Paper..” Geld. . “Output and Price Stability: An International Comparison. S. Indicators. 179204. Watson. 1988..” Journal of Economic Dynamics and Control. 213-225. 109-132. Halbert. Chicago: The University of Chicago Press.” Econometrica.” Journal of Political Economy. and Harald Uhlig. 48. Stock. and P. “New Indexes of Coincident and Leading Economic Indicators. . 817838. Geneva: League of Nations. Karlsruhe).. 1980. “Financial Structure and Monetary Rules. Jack E. 1-21. J. “A Heteroskedasticity Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.

Young. “Nonlinear Regression with Dependent Observations. 1981.. 1984. 541-565.. Allan H. H. “Alternative Monetary Policies and Stability in a Stochastic Keynesian Model. Halbert.” Econometrica. 72.BIBLIOGRAPHY 417 White. 143-161. 1992.” International Economic Review. 32-48. “Alternative Measures of Change in Real Output and Prices. Yoshikawa. 22.” Survey of Current Business. 52. and Ian Domowitz. .

.

. F. 12 event probabilities. Gregory C. M.. Lawrence J. 104. 203 disequilibrium.. 12 Amemiya. 257 Davidson. Phillip. 137 Domowitz. Shirley.. 199 Cowles Commission approach. 8. Alan.. Olivier J. 196 Chow. Russell. 350 Finn. Nicholas. 5. 24 Blanchard. 215–216. Steven N. Kenneth. 223 Cooper. xvii. 102.. 10.. 211–212. 205– 210. J. 1 asymptotic distributions.. Robert W. 50. Robert J. 196 decision equations. Ben S. 80. 217 chi-square tests. 17.. 1 Canova. 97. 75–77. 4. Rudiger. Kathryn M. 64. 11 Cumby. 203 DFP algorithm. 17–18. 201. Carl F.. Sumru. 203.Index age distribution. 40. 241 extended path method. 112 Branson. Polly Reynolds. 71 Andrews. 24 Braun.. J. Ian. 85. 89. 200 balance-sheet constraints.. 18 Christiano. 12 Carloyzi. 24 Duesenberry. 14 data mining. 115. 309 Black. Robert E. 218 Fay. 43–44 Barro. 12.. 4–5. 276 Dominguez. 15. 10. Francis X.. 75. 187. 81 Altug. 71 Chong. Stanley W. 189. Richard A. xvii. 17 Easterlin. 78.. 221 Arrow. 97 Eichenbaum. 78. 16 DeLong... 12 . 200 Dornbusch. 97 Diebold. 12 Clower. 193–194. 104 Allen.. 17. William H. Jon A. 24 Almon. Martin. 71. 24. 7. 196–197. 1. 202 Christ. 271 Diane Macunovich. Yock Y. 265 Bernanke. James S. Bradford.. 308 Blinder. K.. Donald W. 68–70. Takeshi.

12. 73. 223 Johnson. 196–197. Stephen K. 265 Lucas critique.. 24. 14 McCallum. Lawrence R. 31.. G. S. 15 Kuh. David F.. 17 Granger. Adrian R. 229 Mankiw. 288 Okun. 78. 6.. 43. Kenneth N. Alan B.. B. 14 Newbold. 2.. 199 new Keynesian economics.. Lars Peter. 196 Huizinga. Robert E. 208.420 flow of funds. Jean-Michel. 2. 13–14. Dale T. 17. 68–70. T. 72. 67–68. 1 Kouri. 1 full information maximum likelihood. 1. 237 Obstfeld. Grayham.. Guy H. 1. N... 12. James G. 288 Orcutt. Edmond. 14 Pagan. 24 Nelson. 24 Hendry. Johan. 18 Murphy. 7.... 12. K. Gregory. 24 Friedman. 24. 198 Myhrman. 10.. James L. 223 MacKinnon. Edwin. 186.. 265 Gauss-Seidel technique.. 197 median unbiased estimates. Gary.. 139 Kydland. 298 Medoff. 11. 24 King.. Charles R. 72 . Lance.. Penti J. E.. 139 Fromm. 280. 265 Hansen. 298 INDEX Krueger. 7. 68–70. 280. 1 Grandmont. 196 Mortensen. Dale W. 196 information in forecasts. 200 Hurwicz... 12. 24 Goldberger. Maurice. Arthur S. James D. David M. James J. 208 Girton. 4 Henderson. J. 67 McNees. 24 Krane. 24.. 184.. 203 Klein. Harry G. 73. Benjamin M. 195. Clive W. F. 350 Lucas. Kevin M... Stephen R. 40 Frenkel. 200 Heckman. 107. 12. 1 Kuttner. 186. Leonid.. 10. 350 Mizon. Jacob A. Herschel I. 206. 200 Okun’s law.. Arthur M. 205. Spencer D. 221. 75. 196 Grossman.. 12 Hall. Robert E. 196 Malinvaud. 50. 186. 265 Hall. Paul. 15. 262 Hamilton. 17 Lilien.. 225. John... Jr.

Lawrence H. 214. 209. 18 Phillips curve.. 225. 184. 14 Rotemberg. Allan H. J. J. 72 Shiller. 269 Tobin. 14–15 Summers.. 298 two stage least squares. 198.. 27. Julio J.. 164. P. 202. 203–204. 195. P..INDEX Parke algorithm. Carlos A. 14 Rudebusch. Sidney G. 308 White. 203 Uhlig. 30. 10. Robert J. 65. 211. 41 Turnovsky.. 265–266 Sims. Don. 12 stability tests. 41 .. 226 Parke. 196 Rodrigues. Robert H. 203 von zur Muehlen. William R.. 206.. 70. 12. 202 Stoker. 206–207.. 269 Young. Thomas M. 191– 192. Edmund S.. 211.. John B. 315 reaction functions. Laurence H. Harald.. 186 Winter. 75 Stigler. 17 Phelps. 159. 225. 191 Singleton. 214 Stock. 24 Romer. Halbert. 3. 209. 226 Patinkin. 6... 298 rational expectations. 24 Pigou. 11 Richard. 18 stochastic simulation. C. 298 Tinbergen... 225. William. 266 Ploberger. 10.. 203 Watson. Jack E. xvii.. 1 Tinsley.A. 22. 37. 198 Triplett. Werner. 18 Yoshikawa. James H.. 269 two stage least absolute deviations. George J. David. Jean-Francois. E. S. xvii. 240. Arthur C. 269 Sargan. 12. 76–77 Poole. H. 237. Herbert S. Mark W. 268 Prescott. 249 real business cycle approach. James.. 203.. 217 three stage least squares. 12... Denis. xvii. 184. 79 Summers. 65.. Glenn D.. 200–201 Winokur. 131. 265 421 Taylor. 269 Topel. Kenneth. 214. Christopher A.