This action might not be possible to undo. Are you sure you want to continue?


....
Theory and Applications with EViews
Economet cs .
Ben Vogelvang
FT Prentice Hall
Additional _ IUppoIt III •• wop FNPTn1co···.... c I Ie
•
+
with EViews
Harlow, Engtand • London • New W>rk • Boston· San Francisco· Toronto Sydney ' Tokyo • SInppore • Hong Kong • Seoul· Taipei · New Deihl
lerdam . Munich·
Pearson Education Limited
Edinburgh Gate
Harlow
Essex CM20 2IE England
and Associated Companies throughout the world
Visit us on
th~
World Wid" Web oJ;
www.pearsoned.oo.uk
First published 2005 C Pearson Education Limited 2005 The right of Ben VogeJvang to be identified as author arthis work has been asserted by him in accordance with the Copyright, Designs and Patents Act 1988.
;
AU rights reserved. No pan of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic. mechanical. photocopying, recording or otherwise, wilhout either the prior wrillen pennission of the publisher or a licence permitting restricted copying in the United Kingdom issued by the Copyright Licensing Agency Ltd. 90 Tottenham Court Road. London WIT 4LP. All trademarks used herein are the property of their respective owners. The use of any trademark in this text does not vest in the author or publisher any uadcmark ownership rights in such trademarks, nor does the use of such trademarks imply any affiliation with or endorsement of this book by such owners.
ISBNIO: (}"273683748 ISBN13: 9780273683742
British Library CataloguinginPublication Data
•
A cataJogue record for this book is available from the British Library Library of CongrtSS CatalogtnglnPubUcation Data A catalogue record for this book is available from the Library of Congress
10987654 1009080706 Typeset in 101 12 pt.1imcs by 59 Printed and bound in MaJaysia
The publisher'S policy is to use paper numufacturedfrom sustainabfejorests.
Dedicated to my fine small family: Yvonne, Astrid and Valentijn, Michiel
•
Contents
Preface Acknowledgements
x,•
xv
Part I Preparatory Work
Introduction to Part I Chapter 1 Basic Concepts of Econometric Models
1. 1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 Scientific quantitative economic research Economic and econometric models Economic data Variables of an economic model Structural and reducedform equations Parameters and elasticities Stochastic models Applied quantitative economic researc h
1
2
5
5
6
8 8 II 12 14 15
Chapter 2
Description of the Data Sets and Introduction to the Cases
2 .1 2.2 2.3 2.4 Introduction Data set I: commodity prices Data set 2: macroeconomic data Data set 3: oil marketrelated data
19
19 20 21 22
•
VIII
Contents
2.5 Data set 4: money market 2.6 Data set 5: crosssection data 23 24
Chapter 3
Basic Concepts of EVlews and Starting the Research Project
3.1 3.2 3.3 3.4 3.5 Introduction The creation of a workfile in EViews Viewing variables and some procedures Some basic calculations in EViews Case I: the data analysis
25
25 26 32 41 47
Part II The ReducedForm Model
Introduction to Part II Chapter 4 Description of the ReducedForm Model
4. I 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 Introduction The assumptions of the classical regression model The ordinary least squares (OLS) estimator Relationship between disturbances and residuals Estimation of the variance of the disturbance term Desirable statistical properties of estimators The distribution and some properties of the OLS estimator Maximum likelihood estimation Regression anal ysis with EViews
49
50 51
5I 52 55 6I 63 64 67 79 80
Chapter 5 Testing the Deterministic Assumptions
5. 1 5.2 5.3 5.4 5.5 5.6 Introduction Relationships among variables of the linear model The coefficient of determination Survey of some useful distributions and test principles Distribution of the residual variance estimator Interval estimates and tbe Itest to test restrictions on parameters 5.7 The Wald Ftest to test restrictions on parameters 5.8 Alternative expressions for the Wald Ftest and the LMtest
85
85 86 89 92 95 96 103 108
Chapter 6 Testing the Stochastic Assumptions and Model Stability
6. I 6.2 6.3 6.4 6.5 6.6 6.7 Introduction A normality test Tests for residual autocorrelation Tests for heteroskedasticity Checking the parameter stability Model estimation and presentation of estimation results Case 2: an econometric model for timeseries data
113
I 13 I 15 I 17 125 132 135 137
8 Introduction The generalised least squares (GLS) estimator in theory The SUR model Case 4: a SUR model for the macroeconomic data Autocorrelated disturbances Heteroskedastic disturbances Case 5: heteroskedastic disturbances Introduction to ARCH and GARCH models 171 171 172 174 178 179 182 191 192 Chapter 9 Models with Endogenous Explanatory Variables 9.2 9.6 215 2 15 2 16 221 226 226 229 230 235 235 236 239 242 247 .3 10.7 Two specific multipleequation models Chapter 11 Qualitative Dependent Variables 11.3 7.4 7.1 10.6 Part III Introduction Specification errors Prediction Multicollinearity Artificial explanatory variables Case 3: exercises with the estimated model IX 139 139 139 143 155 159 166 Specific Structural Models Introduction to P a rt III 167 168 Chapter 8 Estimation with More General DisturbanceTerm Assumptions 8.3 9.1 11.4 Introduction The linear probability model Probit and logit models Estimation of binary dependent variable models with EViews 11 .Contents Chapter 7 A Collection of Topics Around the Linear Model 7. 1 8.3 8.1 7.2 11.2 10.5 Introduction The instrumentalvariable (IV) estimator The twostage least squares (2SLS) estimator IV and 2SLS estimation in EViews Case 6: consistent estimation ofstructuraI models 199 199 201 204 2 10 213 Chapter 10 Simultaneous Equation Models Introduction Mathematical notation of the SEM Identification Case 7: identification of the equations of a SEM Estimation methods Case 8: simultaneous estimation of a macroeconomic model 10.1 9.2 8.4 10.5 10.7 8.2 7.3 11.4 8.6 8.5 8.5 Case 9: modelling a qualitative dependent variable 10.4 9.5 7.
5 Trends and unitroot tests 12.6 Examples of theoretical ACFs and PACFs for some TS models 14.4 Models with a lagged dependent variable Chapter 14 Univariate TimeSeries Models 14.2 A finite distributed lag model 13.7 Case 10: cointegration analysis Chapter 13 Distributed Lag Models 13. Unit Roots and Colntegration 12.4 The autocorrelation function (ACF) 14.3 Infinite distributed lags 13.uklvogelvBDIL For students • • Empirical data sets for practical learning.7 The Box.2 Timeseries models and some properties 14.x Contents Part IV Timeseries Models Introduction to Part IV Chapter 1.1 Introduction 12.2 Lag operators. notation and examples 12.2 Dynamic Models.5 The partial autocorrelation function (pAC F) 14.3 BoxIenkins procedures 14.4 Errorcorrection models 12.Ienkins approach in practice 14.8 Case II: BoxIenkins procedures References Index 249 250 253 253 254 259 265 278 292 303 305 305 306 3 11 3 16 321 32 1 324 327 328 330 333 340 352 353 357 Companion Website and Instructor resources Visit the Companion Website at www.3 Longrun effects of shortrun dynamic models 12. as described in Chapter 2 of the book Assignments and Checklists for each chapter for a better understanding of the discussed subjects For lecturers • • • Downloadable Case Notes to assist in using the data sets to carry out the case exercises described in the book Exercises with solulions for each chapter for use in teaching and class testing Solutions 10 the Assignments and answers 10 the Checklisl questions .pearsoned. 1 Introduction 13.6 Relationships between nonstationary variables 12.co.1 Introduction 14.
It is very important to recognise which assumptions have been made by an author. whether these assumptions have been tested in a correct way.Preface '. but econometrics clearly belongs to the economic discipline. la first course in econometrics' is presented that meets a11 these requirements. In other words. In this course.:t' ". Therefore they should have knowledge about quantitative economic methods: they should know how models arise. In this book. many exact derivations are given. References to the econometrics literature. its application and the use of econometric software distinguishes this book from existing econometric texts. the aim of this econometrics course is to make students aware of the underlying statistical assumptions of the methods they want to use or that others have used. With this approach to teaching econometrics. Econometrics is not mathematics or statistics only. etc.~ •  '. students will get a clear understanding of econometric practice and the underlying econometric theory. The integration of econometric theory. what the underlying assumptions are. this startingpoint is obviously recognisable in the text. • About this first course in econometrics Economists will regularly be confronted with results of quantitative economic research. econometric theory is combined with econometric practice by showing and exercising its use with software package EViews. . and in what way estimates of parameters or other economic quantities are computed. It is not imperative that students can prove all the underlying statistical theorems that are necessary to derive estimators and statistics. In thi s book. Basic knowledge of econometric theory is necessary to understand what may (or may not) be done in applied quantitative economic research. are given in throughout the text and are listed at the end of this book. but in some cases it is sufficient just to explain what is going on in that particular situation. for more detailed information on any subject.
Thomas (1997). to estimate their parameters in a correct way by using economic data and econometric software (EViews). the use of EViews is discussed in detail in connection with the econometric theory. A very short but practical introduction to get a first impression of econometrics is Franses (2002). A case does not concern the solving of a mathematical or statistical problem. In the mentioned textbooks. They will be able to specify simple linear economic models. For this reason. it uses real economic data to solve an economic question. to test simple hypotheses about economic behaviour. after which the course can be extended with chapters from Parts III and IV. which is to learn what can be done with econometrics. together with the included examples. to test the underlying assumptions. This book is suitable for selfstudy because of the extensive and integrated discussion of econometric methods and their use with EViews. and why econometrics is a compulsory subject for an economist. and to evaluate the computer results. the theory is discussed and illustrated with empirical examples and exercises that are often mathematical or statistical calculations. The objective of this book is learning 'to correctly perform quantitative economic research' by means of showing elaborated examples from econometric practice and by working through cases. With empirical data sets students act as if they were doing a quantitative economic project. Greene (2000). Most of the cases can be considered as a 'simulation' of performing quantitative economic research. Features and objectives of this book The main feature of the book is the strong integration of econometric theory and matching software to apply the theory. Parts I and II have to be discussed and studied completely. Verbeek (2000) and Wooldridge (1999). so the text has been written in a way that various selections of chapters can be made for a desired level of the course in question. Earlier obtained results will be reconsidered in following chapters. Stock and Watson (2003). such as Dougherty (2002). It is the auther's experience that this approach to teaching econometrics is very effective in understanding the use of econometric methods. This distinguishes this book from existing econometric texts. a number of cases have been included in this book. The total number of subjects can be too much for one COllIse. as a broad number of econometric topics is introduced and discussed with their applications. Therefore. The cases will help to realise the target of this first course. Hendry (1995). Johnston and DiNardo ( 1997). Students see the benefit of theoretically explained methods more clearly when they are directly applied in an empirical example. For this purpose. to forecast economic variables and to simulate an economic policy. Such a project is evaluated throughout the book because the same data will be used in successive chapters. Gujarati (2003). Studenmund (200 I). However. they will be acquainted with a variety of econometric models and methods. This book offers a complete text for a 'first course'. After this ' first course in econometrics' students will have the following skills. data and exercises. the use of this book is not limited to university courses in .xii Preface The best way to learn about doing quantitative economic research is to study econometrics where theory and empirical work alternate in an integrated way: immediately applying a discussed procedure on economic data by using econometric software. They will have learned to compute elasticities empirically. FinaJJy. Maddala (2001). Hamilton ( 1994).
However.1 Student Version at a very low price. The appearance of a window can cliffer slightly in various versions of EViews. EViews is published by Quantitative Micro Software (QMS). its application with EViews will be demonstrated by showing the specific EViews windows. For that purpose. will be described on the website related to this book. the use of EViews 4. Such compact c hapters provide adequate knowledge to study this introductory course in econometrics. EViews has been chosen for this course as it is the author's experience that it is the most suitable software for realising the objectives stated above. The discussion of EViews in this book concerns the basic procedures for analysing economic data. who have an interest in quantitative economic research .1 has been discussed. it makes hardly any difference which level of EViews is used. then you are advised to read the statistical and/or mathematical chapters that are often found in various available econometrics textbooks. Students can buy the EViews 3. provided they can submit a proof of academic affiliation. estimating simple models and testing the underlying statistical and economic assumptions. The use and output from EViews is extensively described in this text. Microfit.eviews. OxMetrics (which has a number of known modules like PcGive. Part I consists of three preparatory chapters. will be described. . RATS and Stata. The way to use the software for estimating parameters of economic models and for testing hypotheses. It can very well be studied by economists in trade and industry. Each time an estimator or a statistic is discussed. or government institutions. Some examples of wellknown packages that can be used under Windows are EViews. Organisation and contents This book has four parts. Prerequisites for this econometrics course Basic knowledge of matrix algebra and statistics (distributions. PcGets. For detailed information see the 'QMS Home Page' at http://www. • EViews (Econometric Views) Various econometric software packages are available.Preface xiii econometrics. In the first c hapter some basic concepts of economic and econometric models are briefly summarised. which are discussed in this book. estimation and test principles) is required. Finishing this manuscript coincided with the release of EViews 5 by QMS. lf you do not have this knowledge and cannot take a course in statistics or mathematics. but this is not confusing. the application of methods in EViews is not an automatic affair: each time a student has to take decisions him or herself about the next step that has to be done. STAMP and TSP/GiveWin). by simply 'clicking' on the procedure names. Interesting aspects of updates of procedures. In this book.coml.
These chapters concern the mode l with a quaJjtati ve dependent variable. This di scussion is important for knowing in which way correct dynamic model specificati ons for timeseries data can be obtained. twohour lessons would be needed. it is possible to do sucb a course in about 14 or 15 twohour lessons. a number of specific models are discussed. In Part m. and many statistical tests concerning the validity of the assumptions. It will take about seven. Part II concerns the ' reducedform model'. Chapter 13 discusses some 'distributed lag models'. Applications with EViews are presented whenever possible and useful. To give an idea of timing. If the COurse should be extended with all the subjects of this timeseries block. With a selection of the subj ects in Part m the basic course can be extended. different mode ls with their own specific properties will be introduced. which are the last dynamic causal mode ls discussed in thi s book. netlvogelvang By using a website for additional information. This model is discussed in detail in Chapters 4 to 7. 'cointegration' and 'errorcorrection' models. Chapter 12 discusses the estimation of coherent shortrun and longrun models. Parts I and II form already a minimal but complete basic econometrics course. different estimators are introduced in Chapters 8 and 9. from Part II. data is easily updated when new and interesting information becomes available. eight. and two multiple equation models: the seemingly unrelated regression model (SUR) and the simultaneous equation model (SEM). Various types of dynamic models will be discussed. all the subjects of Part Part IV pays allention to the dynamic bebaviour of economic models for timeseries data. For these reasons. Website All the data that are used in the examples and cases in the book can be found on the related website: www. In Chapter 3.xiv Preface Chapter 2 introduces the cases and describes the data tbat will be used throughout the book. Subjects m that come up are ' unitroot tests'. EViews is introduced and extensively discussed. Finally. All these topics belong in a contemporary basic course in econometrics. Chapter 14 presents the univariate timeseries models (ARIMA models). 10 and II . In Chapters 8. . All the necessary assumptions that have to be made for estimating the parameters of the model. twohour lessons to discuss in an extended course. come up for di scussion. It concerns issues like what to do when some of the assumptions concerning the reducedform model. have been rejected or when they are not valid.booksites. Other information that is relevant after the book has been printed and live updates of EViews will also be placed on the site.
Especially. This resulted in many comments and suggestions. Jonneke Bolbaar and Job Zinkstok. with whom I share an office at our Department of Econometrics. for always having time to answer all my varied questions.which also resulted in many improvements. They always quickly and adequately answered my email questions. . His detai1ed too comments and suggestions have been of inestimable value. I am very indebted to the guidance of Justinia Seaman. Tim Parker and Janey Webb. I appreciate the informal but constructive way in which everything was discussed. which have clearly contributed to the many text improvements. and for his sharp eye in judging the empirical oriented topics in the book which yielded many improvements. The final text has benefited greatly from aIJ the mentioned contributions. who have 'voluntarily' spent a lot of their time in judging the book from the point of view of potential users . Koos Sneek and Job Zinkstok who have all read the complete manuscript. I thank Kees van Montfort for providing me with the interesting crosssection data set about the Dutch finns that is used in the cases in Chapters 8 and II. Janey was very helpful during the first part of the project in particular. In alphabetical order. I am very grateful to Jonneke Bolhaar. I am grateful to Hidde Smit for his ideas about a clear order in di scussing a number of subjects. I thank Bernd Heidergott. I want to thank various Pearson Education personnel for their guidance during the entire project. Koos Sneek has very carefully read the text.The text of thi s book has benefited very much from the comments and suggestion s that I have received from colleagues and exquantitativeeconomics students at the Vrije Universiteit Amsterdam. Hidde Smit.the students . I thank two of my exstudents. and guarded me against formulating a number of statistical statements in a < casual' and therefore sometimes nonexact way.
Ben Vogelvang . I am thankful to Quantitative Micro Software for their support and for giving me permission to include aU the Ey!ews screen shots in this book. Only I am responsible for the remaining errors in the text.xvi Acknowledgements Finally.
" .:::.':: '. • . '."i.... ..
In fact. .Introduction to Part I In this part some necessary basic concepts concerning econometric research are discussed. Different phases in an econometric research can be distinguished. These aspects determine the contents of Part n. These phases are schematically presented at the end of Chapter I. this scheme indicates the way empirical research is done. The chapter contains a discussion of important aspects of the linear economic and econometric model. The different types of data that can be collected and the different kinds of economic variables and equations that are distinguished will also be discussed. The cases and the available data are introduced in Chapter 2. The graphical output from EViews can be directly inserted into a text editor like Word or Scientific WorkPlace (LaTex). the basic concepts of econometric models are introduced. These basic EViews operations are extensively discussed with many examples of the windows that are opened by EViews. The relationship between the parameters of the model and elasticities is emphasised. In Chapter 3. Before a serious start can be made with learning econometrics you need to know about basic aspects of economic and econometric models. In the last section of Chapter 3. This exercise corresponds with the first phase in an empirical study. The procedures in EViews have more interesting and useful aspects thanjust their easy use for computational work. this has been summarised. the last chapter of Part I. In Chapter I. The data that have been used in the examples in the text are also available and this makes it possible for the reader to reproduce the examples. from the beginning to the end. Information is given about what can be done in those windows. the first exercise is formulated as the start of an empirical econometric research project. namely a data analysis. . a number of basic concepts of EViews are introduced. Each phase consists of a number of topics or procedures that have to be done or can be done. about the sorts of economic data that exist and about the use of a software package like EViews that is discussed in this book.
the 'reducedform model '.Introduction to Part I 3 The knowledge obtained from this part will be sufficient to start you on the discussion of the first econometric model. that is introduced in Part n. and the empirical applications using EViews. .
.
with the reality of the economic world. For example. Empirical quantitative economic research consists of confronting a hypothesis. like an estimated model. Research must be objective and reliable. and test statistics to test hypotheses about economic behaviour. so that other people can check the reliability of the results. Theoretical economic research in general concerns the development of economic theories.1J Scientific quantitative economic research In economic research. The results have to be capable of being reproduced by other researchers. the emphasis is on correctly performing applied econometric research and on the notion th~t some tOOls can be used whereas others yield incorrect results. when a research paper is submitted to a journal for publication. estimators to estimate unknown parameters of economic models. The research has to be done in . In this text. Another methodological aspect is to keep the finally obtained result. as parsimonious as possible to present a transparent analysis of the investigated economy. fannulated from the economic theory. Many methods have been made operational in software packages. When doing research it is important to adhere to a number of methodological rules. as in other disciplines of science. it also concerns the development of statistical methods to test the validity of economic theories. These methods fill the toolbox for econometricians. The toolbox mainly consists of statistical and econometric methods.Chapter 1 Basic Concepts of Econometric 1. you may be asked to put the used economic data at the editor's disposal. In quantitative economics or in econometrics. the distinction can be made between theoretical and empirical research. The hypothesis is tested by using real economic data.
it is also important to write a research paper in a clear. In obtaining an econometric model for timeseries data. The model consists of mathematical equations with economic variables and unknown parameters. Statistical assumptions about the model will be made and tested afterwards. After the unknown parameters have been estimated by using economic data for the variables and by using an appropriate econometric estimation method. Hypotheses can be formulated from the econonUc theory about the economic behaviour of economic agents in the real world.just every possible combination of explanatory variables can be used . Data nUning has to do with estimating many regression equations in a nonstructural manner . In its most simple form. lfyou want to prove an economic theory then you have to find a parsimonious econometric model in a structural way. Attention to this stage of a research project will be given in Section 6. Often it is possible to derive properties of the model a priori. consumption is determined by income only. the approach of David Hendry will be used: structural modelling from general to specific. These properties are relevant for the evaluation of the estimation result later on. (See. That model can be used to test the hypothesis. In the literature it is customary to use the notation C for consumption and Y for income. and also Charemza and Deadman (1992) for a discussion on this topic.l) C = f3t + f3 2 Y.and then choosing subjectively the result that confirms the idea that the researcher wants to prove.) Naturally.6 Chapter 1. Hypotheses can also be formulated on the basis of people's own observations. in fact the final result will be worthless in many cases. A number of rules exist to help you write in a systematic way. Hendry ( 1995).2 Economic and econometric models A quantitative economic research project starts with an economic theory.1) is specified as a linear economic model by writing: (l. U that happens. before the first case that simulates a research project. which does not reject the econonUc and statistical assumptions that have been made to obtain this model. 'Data nUning' as a method of obtaining results predetermined by the researcher is potentially an unethical method of empirical research. Case 2.6. Basic Concepts of Econometric Models an ethically wise manner. 1. The model of the economic behaviour that has been derived from a theory is the economic model. the result deserves the title 'How to lie with statistics' ! Correctly found scientific evidence to support the original hypothesis stemming from the econonUc theory has not been found .34) a simple clarifying example is given with a macroeconomic consumption function that is shown below. is introduced. one has obtained an econometric model. Then the consumption function is written as a function of income: C=f(Y ). pp. for example. In Stewart and Wallis ( 1981.2) . by looking at the theory only. With this approach the data is analysed and modelled in a structural and objective way. Equation ( 1. If such a model cannot be found then the conclusion should be drawn that the hypothesis is probably not valid and bas to be rejected. ( 1. transparent and systematic way.
yields: 1 dC C Y dY < y2 ' dC C dY < Y ' YdC C dY < 1. . ( 1. These inequalities will be used in the 'economic evaluation' of the econometric model after the parameters of the model have been estimated. > a and C > a a.4) is the mathematical expression of theineorne elasticity. Ifboth variables are expressed in identical units then these restrictions imply: dC a < dY < l.6 (and the 'hat' above the C in Section 4.3).. but consumption increases less than income does.2): a <f32 < 1. Further. The parameters (31 and f32 are unknown and have to be estimated. you may expect a positive value for the intercept: f31 > a. It is also known that the part of the income that will be consumed will decrease when income rises.Economic and econometric models 7 It is common to use Greek characters to denote the parameters. which is the econometric model. The correct way of specifying a causal model for economic variables will be introduced in Section 4. The last inequality states that the income elasticity is less than I.4) The lefthand side of inequality (1.2. with the assumption that Y < .81. or formulated in mathematical terms: d ('.81Y.4 and f32 = a. then the model is written as  C = 15.3) results in a restriction on the parameters f32 of the consumption equation ( 1. Notice that this equation has not been correctly specified: an exact relationship between consumption and income does not exist... For example: consumption increases as income increases.dY Differe ntiation of the quotient C /Y gives: a. C. (1.c / Y ) < .4  + a. The usual notation for pubUshing an econometric model is discussed in Section 6. This implies for equation ( 1. Suppose that the parameters have been estimated with the resulting values f31 = 15.3) Inequality (1. d ( C/ Y)_ ~dC _ £ dY .Y dY y2 Elaboration of the inequality. 1) that the first derivative of consumption with respect to income is positive and sma11er than 1. It is possible to look already at some properties of this function that are known from the economic (Keynesian) theory.
The nature of economic variables can be endogenous. The collected data form the sample. Each observation is: • a realisation of a variable at a certain moment. crosssection data and panel data. Timeseries data are historical data. for people. but aiternatively it is also possible to analyse endoftheperind data. called the dependent variable.3 Economic data The unknown parameters of a proper economic mndel can be estimated if sufficient data on the variables are available. quanerJy data. for example. Now it is often possible to get the data in electronic form. or futures prices on commndity exchanges. but the data concern the same panel of subjects or objects in every perind. from international orgartisations by using the internet. (See Chapter 2 for some examples. The data that are necessary to do the exercises included in this book can be downloaded from the internet (see Chapter 2). Data are published in national and international statistical perindicals. Variables on the righthand side of the equation are the explanatory variables. country's GOP. persona] mcome. for example. In this book.8 Chapter 1.) An important difference between timeseries data and crosssection data is that timeseries data have a fixed ordering in time. This can sometimes be observed for highfrequency data such as weeldy or daily data: for example. stock market data. for example. from low to high income. published timeseries data concern average values of the variables during a year or a quaner. Basic Concepts of Econometric Models 1. In most studies. for example. company's prnduction. • Three main types of economic data can be distinguished: timeseries data. Crosssection data are data collected during one perind. the researcher determines a logical ordering. Crosssection data do not have a fixed ordering.g.4 Variables of an economic model In an economic mndel there is one variable on the lefthand side of the equation. exogenous or lagged . The difference between timeseries data and crosssection data is also seeD in the notation. attention is only paid to the analysis of timeseries data and crosssection data. Panel data are crosssection data that are collected at various points of time. Economic data can be collected in different ways. etc. then they have to be manually input to a computer file. Crosssection data are then collected typically in onc month or one quaner. large crosssections are found collected for only a few points of time. If data have to be collected in this way. Historical data can be observed at different frequencies. companies or countries. like annual data. which is not a convenient way of imponiog data. or a realisation for a cenain object or subject such as the sales of firms in the same business or the ownership of a PC by a student. Often. (See.) Sometimes data is free to download but sometimes you have to pay for data. 1. and Yi) . The data are considered as drawn from populations. Gt and Yd . whereas for crosssection data the subscript i is used (G. etc. for example. Baltagi (20(H) for an extensive discussion about mndelling panel data. For timeseries data the subscript t is added to the variable (e.
In general notation. In the macroconsumption equation. If more equations have been specified to determine other endogenous variables of the system then the system is called a simultaneous equation model (SEM) . It is an open door. = Ct + It. especially of the explanatory variables. an equation for Y. can be added in the form of an identity. . + f32 Yt . where It is the variable for the investments. but an equation for income was not considered yet as so far the 'research interest' concerned consumption only. The model is not necessarily a model of only one equation. . then that system of equations is called complete.2 A model for~ consumption In a country In Vogelvang ( 1988) an econometric model forthe world coffee market has been developed.2). . but it is clear that lagged variables only can occur in timeseries models. For example. are endogenous variables. because they are determined in the same macroeconomic model. Therefore the dependent variable is always an endogenous variable. For example. Knowledge about the nature of the variables. One equation can often be considered as part of a larger model where other variables are determined too.Variables of an economic model 9 dependent. In fact they are endogenous. A variable is an endogenous variable if the variable is determined in the model in question. If the number of equations is identical to the number of endogenous variables. The exogenous variables are determined outside the model. which gives the following most simple Keynesian model as an example of a complete SEM: Ct = f3. Example 1. but their values have already been realised. annual. Lagged dependent (or lagged endogenous) variables are ' predetermined' in the model. it is always important to judge the nature of all the explanatory variables of that equation. etc. This example concerns a quarterly model for the consumption of coffee in a coffeeimporting . A complete model can be solved for the endogenous variables. (See the next section for more discussion on thi s topic. is of importance for the choice of an estimator for the unknown parameters and for knowing the properties of the estimator. model (1. one can cbange the percentage of any tax variable and compute the effect on consumption. both the variables C t and Y. In sucb a model it is possible to simulate any policy for a more or less completely specified economy.) Although it is not necessary to specify a complete model for all the endogenous variables. for example. n. if you are only interested in the study of one of the equations of that model. Y. which is considered as an exogenous variable in this model specification. Below two examples are given to highlight these points. This is necessary if one wishes to perform simulations with the model as mentioned above. The subscript t indicates that the variables are observed as timeseries with. quarterly or monthly observations. this model is the subject of Chapter 10. That is the reason why many examples in this book relate to the coffee market. . t runs from 1 to n: t = I .
y. In other words. the shortrun price response is {32 and the longrun price response is /32 + /35 + /36' (A detailed disc ussion about shortrun and longrun models is given in Part IV. + (32P. > 0. (3. For example: C~oJ = (3. the prices from previous periods. but this is just an example to illustrate the di scussed topics of an economic model. are also specified in the equation. Other explanatory variables can be specified with lags too. (33 > 0. For that reason.1) influences the coffee consumption in the present period (t ).) The coffee consumption equation can be written as: C~oJ = (3."~r When the price increases with one unit in period t the consumption of coffee has decreased with /32 in period t. Of course more economic variables can be relevant in explaining coffee consumption. Although that depends on the frequency of the data. just as the variables .10 Chapter 1. The equation is called a dynamic model when lagged variables have been specified. with f32 + /35 in period t + 1. For example. Opposite signs of a lagged variable indicate a correction on the adjusunent in the previous period(s)..'o{ + /36 P. + (32P"oJ + (33 Pfea + /3. which is clearly not unrealistic from an economic point of view. the quantity of consumed coffee from the previous period (t . and income (Y. In this model. That can sometimes be considered as a strong restriction on the model. When the price increases with one unit in period t the consumption of coffee decreases with (32 in period t . p" o J The models are static models in the examples so far. C~of and p tCo / are two endogenous variables because both variables will be determined in a complete model for the world coffee market. but for monthly data this can be different. Y. The variable C~O{ is a lagged endogenous variable. Pt eo and Y. This variable C'l represents the habit formation or the (stochastic) trend in consumption. as tea can be a substitute for coffee. Notice that the expected signs of the parameters are: (32 < 0. are two exogenous variables because they are not determined in a model for the coffee market. as in the following specification: In thi s model.the price of tea (Pt ea ) . called lagged variables. Such a change can be unrealistic for economic reasons. then a new price needs three periods before the adjustment in consumption is complete and the longrun price respon se is larger than the shortrun price response. It is also possible that lags oftbe dependent variable are specified as explanatory variables. No a priori restrictions exist on the constant term (31.) . for annual data it is not a problem. If /35 and (36 are negative. the signs of /35 and /36 are not known a priori. and with /32 + /35 + (36 in period t + 2. The implication of a static model is that a change of an explanatory variable in period t is fully reflected in the dependent variable in the same period t.2 below. Basic Concepts of Econometric Models country.) The parameter /32 is negative. An equation explaining is not specified as only the consumption of coffee is 'studied ' here. consumption can be based on the current price but also on the price from the (recent) past. it is possible that the longrun price response is less than the shortrun price response. It is assumed that coffee consumption • depends on the coffee price (C~Of) (prJ) . + /35 P. In that case.c°J + (33Pfeo + (3. (See also Remark 1.
Trends in a country's population migbt also influence the total coffee consumption.Y. Remark 1.1 Warning Be aware of specifying strange restrictions in the model. like 'holes' in lag patterns. A shortterm price change will have less effect on consumption when the trend is strong (/35 close to I). Remark 1_2 Is the specification of the consumption equation realistic? Income (Y. Note that the specification of lags decreases the sample length that is used to estimate the parameters. A model for variables measured in monthly averages will have more dynamics than a model for variables with an annual frequency. This point will discussed in more detail later on.Structural and reducedform equations 11 P.":{.. and the variable is mindlessly deleted. are called the structural equations. For example. mathematically seen. this can happen wben it is observed that the inlIuence of Y. Many lags can be expected in a model for highfrequency data. otherwise the difference equation will explode. + (33Y.5 Structural and reducedform equations The equations of the economic model as specified in the economic theory.. a difference equation. The parameter (35 has the restriction 0 < (35 < 1.2 has a remarkable restriction: consumption in period t is determined by income from the periods t and t . 1. but in this equation C . A model is just valid in a certain sample and within a certain range of the variables. or will show an unrealistic alternating influence of C~o{. The utmost consequence of this specification is that enormous quantities of coffee are consumed at very high income levels! That is actually not very realistic. A complete model of a number of structural equations is the structuraL form of the model. + (3. A model with a lagged dependent variable is. as will be explained in Chapter 4. is not significantly different from zero after the parameters bave been estimated. an equation like: Ct = (3.2. One has to be cautious with computing predictions for an endogenous variable for extreme values of the explanatory variables by using an estimated econometric model. Then it is better to specify consumption an~ncome per capita.11 That is clearly not a realistic model. or just a few or no lags are expected in a model for lowfrequency data."'{ is called a lagged dependent variable_ The specification of lagged endogenous variables does have its own statistical problems. The frequency of the data is impoI1ant for the specification of the dynamics in a model.l has been linearly specified in the coffee consumption equation. like the price of soft drinks as one more possible substitute for coffee. Different types of structural equations can be distinguished. For example. but not by income from period t . An equation . The same is true for longterm model simulations. Other variables can be relevant in explairting coffee consumption.
Other assumptions about the parameters can be made. as seen in Example 1. Solving that model for Ct and Yi gives the two reducedform equations: Or in general notation: Ct = ?Tu + 7r12It l Yi = "21 + "22 I t· The reduced form will be used to simulate a policy or to compute forecasts for the endogenous variables. the coefficients of the variables. then it is also a specification for the longrun model.12 Chapter 1. implying that the economic structure of the model is assumed to be constant in the sample. so they will be found as explanatory variables in reducedform equations. Lagged endogenous and lagged exogenous variables in a structural model are predetermined. which are free of units. Without lagged dependent variables the reducedform equations show the longrun effect too. Below it is shown that restrictions on . Other relationships are. Therefore.2. The SEM in Example 1. A complete SEM can be solved for the endogenous variables. In a first econometrics course it is usual to assume that parameters are constant in the model. it is often better to check the val idity of the estimates by computing and evaluating the magnitude of elasticities. • 1. The reducedform model consists of reducedform equations. some of the reducedform equations will be difference equations that can be solved to compute longrun effects. When lagged endogenous variables are specified in a structural form. Basic Concepts of Econometric Models like the consumption equation is called a behavioural equation .1). but not in this book. The solution is called the reducedform model. The following terminology is used. The magnitude of the parameters depends on the units of the variables. of course after the estimated reduced form parameters have been inserted.technical relationships (like production functions) and identities (as in Example 1. in more detail. for example. It is important to be aware of the following point. The values of the estimated parameters can hardly be interpreted when the units of the various variables are different.6 Parameters and elasticities It is useful to look at the parameters.1 is a complete structural form with two equations for the two endogenous variables C t and Y" The investment variable It is supposed to be exogenous. Unknown parameters do not occur in an identity. The equations for timeseries data that have been considered so far are models that describe the behaviour of the economic variables in the short run. If a static specification is a correct model for the short run. The parameters are estimated by an estimator and the result is called an estimate.
of Y with respect to X is: BY. eyx = BX. The log transformation is a popular transformation in econometric research : because it removes nonlinearities to some extent. The elasticity is conslant in the sample as the parameter f32 is assumed to be constant. An example of the computation of an income elasticity in EViews is given in Section 3. This is a restriction! Be aware . What you want to compute depends on what has happened in the sample period. and X. and Yi. The quantities that can be reponed in a paper are. So. when using the bivariate model.5) with respect to X t gives the elasticity e. Thea it is interesting to look at a graph of the elasticity. for example. x of Y with respect to X: This means that the coefficient f32 of model (1.4. and Yi have been transformed in their (natural) logarithms. This can be done in EViews in a simple way. for example. the maximum. X. The following specifications can be distinguished and are illustrated with a simple linear bivariate model. Regularly. you may expect that the price elasticity with respect to demand has not been very constant. But in a period with heavy price changes on a commodity market. . and because the coefficients can easily be interpreted as elasticities. minimum and mean value of the elasticity in the sample period.. The variables X.5) has become an elasticity and this elasticity no longer varies with any variable. The elasticity can be computed for all the points of time of the sample period.Parameters and elasticities 13 elasticities are imposed by the log transformation of the variables. but presenting only an average elasticity can be less informative. The model is: In (Yi) = f31 + f32 ln (X t ) (1. the elasticity ey x is computed as: • The elasticity varies in the sample with both the variables Y. Yi . If the economy was rather quiet in that period it can be sufficient to compute the mean elasticity only. you find that only the sample means are substituted for X.5) Differentiating (1. it has a dimming influence on changes of the variance of the variables. Then it is more informative to publish more statistics than just a mean. • The model has been linearly specified in the observed two variables: Yi The definition of the elasticity eyx = f31 + f32X .
At the righthand side of the equation. then it is not a problem but if this is not the case then the log transformation should not be applied. Economic relationships are not exact relationships so they need to be written in a different way. or to vary with only one of the variables. ) = {3. is now computed as: and this elasticity varies with X. The elasticity e yx + f32 X. • The model is: In (Y. A disturbance term will be included at the righthand side of the equation to complete the equation. 1. in that particular sample period? U the answer is yes. the models were disc ussed as exactly determined mathematical equations. two parts of the specification will be di stinguished : the systematic part. A di sturbance term is a stochastic variable that is not observed. • The model is: 10 this situation the elasticity e yx is: This elasticity only varies with Y. to restrict the elasticity to be constant. That is possible in the 'natural sciences' where exact relationshjps are derived between physical variables. which concerns . Is it reali sti c. which are the following sernilogarithmic models. it is not the case in economics.. Conclusion The conclusion is that one has to be aware of the consequences when transforming variables to their logarithms. only. However.14 Chapter 1. Two other possible specifications can be di stinguished. for economic reasons.. Basic Concepts of Econometric Models before the application of the log transformation whether or not this restriction is an economically realistic restriction in the sample period that is analysed.7 Stochastic models Until now.
the systematic part bas not been correctly specified and has to be respecified. All endogenous variables. ( 1. the specification will be more generally di scussed. By using the estimates. The stochastic behaviour of the disturbance term implies that the endogenous variable is a stochastic variable too..6) with only one explanatory variable is rather limited in explaining the behaviour of an economic variable. This model can be used to compute elasticities. + u. + /32X. This process can be characterised by the foll owing scheme of stages in an empirical research project. is the disturbance term representing the nonsystematic part of the equation. the validity of the assumptions will be tested. the variable u. X. the aim of empirical econometric research is the analysis of an economic phenomenon by specifying an economic model (tbe economic theory). The deterministic assumptions The deterministic assumptions concern the specification of the economic model. First a number of assumptions of the model have to be formulated after which the unknown parameters will be estimated. Y.8 Applied quantitative economic research In this section. whether they are dependent or explanatory variables. = /3. When that is not the case. etc. to test economic hypotheses. It will be clear that model ( 1. and the nonsystematic part. the process of an econometric research project is schematically summarised. a linear bivariate model is specified in general notation as follows. These assumptions are introduced in Section 4. to simulate policies. to make it a valid equation. the disturbance term u. 1. For example.Applied quantitative economic research 15 the specification of variables. All variables that are known to be relevant for the explanation of the endogenous variable have to be specified in the model.6) The variable Y. which is the remaining random Donsystematic variation. In a wellspecified model. is the dependent (endogenous) variable. is an exogenous explanatory variable. This concern s detenninistic assumptions as weU as stochastic assumptions.2. When introducing the general linear model in Section 4. The disturbance term is a stochastic variable and assumptions will be made with respect to its probability distribution. the difference between the systematic part and the endogenous dependent variable will be random without any systematic behaviour. to compute forecasts. estimating its parameters and testing the assumptions that have been made by using common sense. which is the formulation ofthe null hypothesis about the relationship between the economic variables . cannot be observed. are stochastic variables. As mentioned above. If some of them are rejected then the model has to be reformulated. The resulting model will be an econometric model. A s introduced earlier. economic knowledge and statistical methods. All the points mentioned will be discussed in foll owing chapters. but then the equation between the variables is still not an exact relationship. based on the economic theory.2 and will be considered in more detail in Chapter 6.
Distinction can be made between linearity in the parameters and linearity in the variables. The basic specification of the model originates from the economic theory. If no sufficient data are available at the desired time frequency. The choice of which variables have to be included in the model stems also from the economic theory. The mathematical form of the model has to be determined. For example.like the one in Example 1. The parameters have to be estimated by a nonlinear estimation method. which is in accordance with the research objective. Is it possible to assume that the model is linear or is it necessary to specify a nonlinear model? A linear model is more convenient to analyse. whereas longterm economic cycles can be analysed with a macroeconomic model that has been estimated by using annual data.J + (33 In (Z. + (32 In (X . When we are talking about a linear model this concerns a model that is linear in the parameters. The availability of data can influence the assumptions that have to be made about the specification of the equation. An important decision is made about the size of the model.16 Chapter 1. at micro or macro level.J = In ((3tl In (Y. .J + u. The nonlinear model: Y. = (3. A coffee consumption equation. ) + u. What is important regarding tbe choice of an estimator for the parameters is that you have knowledge about the nature of G .J = + (32 In (X. Basic Concepts of Econometric Models of interest. This can be shown with an example of the following nonlinear model (e. then it is possible that the research objective has to be reformulated. In other words. The following nonlinear model cannot be linearised by a simple log transfonnation: Specific reasons must exist to specify an additional disturbance term here. will probably adequately be estimated for monthly or quarterly data. The parameters of this model can be estimated with a linear estimator like the ordinaryleast squares estimator. whether one or more equations have to be specified. can be linearised by the log transformation: In (Y. an adequate choice of the time frequency must be made. a model that has to explain shortrun stock market prices should be estimated by using daily or weekly data (average or closing prices).g. etc. does the analysis concern a single equation model or a multiple equation model? A different question is: at what economic level bas the model to be formulated. but it has to be based on a realistic assumption.J + (33 In (Z. This model is nonlinear in the variables but it is linear in the parameters. or (3.2. For a timeseries model. x f' Zf'e u . a CobDouglas production function). The stochastic assumptions and estimation of the parameters Statistical assumptions need to be formulated about tbe stocbastic variables in the model with respect to the di stributions andlor their moments.
Applied quantitative economic research 17 the explanatory variables: are these variables endogenous or exogenous. Then the unknown parameters are estimated and test statistics are calculated to evaluate the results. heteroskedasticity tests. for example when looking at the sign and magnitude of the parameters or elasticities. the elimination of systematic bebaviour from the variables or the choice of a specific estimator for the parameters. A first evaluation is obtained by using common sense and economic knowledge. autocorrelation tests. as described in the next stage below. like seasonal patterns. it is relevant to evaluate the forecasting power or simulation properties of the model by simulating the estimated model for either the entire sample period or for only a part of the sample period. Looking at a plot of the residuals can be very informative about cycles or outliers that have not been observed before. This is followed by testing the stochastic assumptions that have been made by using a normality test. But when the evaluauon is satisfactory. The continuation of the project The rejection of some of the preceding assumptions may give rise to returning to the specification or estimation stage. If the stochastic assumptions have not been rejected. present in the timeseries data? Is the variance of the disturbance term in a model with crosssection data constant or is the variance proportional to the variance of one or more of the explanatory variables? Conclusions from such an analysis give rise to action: for example. the 'reducedform model ' will be introduced and discussed. The evaluation of the results takes place in various ways. then it is time to write the final version of the researcb report. It provides information about the behaviour of the economic variables in the sample period. as was explained in Section 1. Attention must be paid to the hypothesis that the structure is linear and constant in the sample period. Conclusion The scheme given above will be elaborated in Chapter 4. This concerns the unsystematic part as well as the systematic part of the model. the deterministic assumptions can be tested by using statistical tests to test for restrictions on parameters: for example. Do the variables correlate as expected? Are trends or other systematic patterns.4? A data analysis is important at the beginning of the research project. the ttest and Ftest can be performed. An estimation method will be chosen on the basis of these assumptions. In Chapter 4. the coefficient of determination R2 can be interpreted. economic fluctuations or cycles. In some projects. Often the systematic part ofthe model has to be respecified during tbe process of achieving a final parsimonious model. etc. This model encompasses 'the classical regression model' which is a linear static model with exogenous explanatory variables only. Evaluation of the estimation results The evaluation concerns the verification of the validity and reliability of all the assumptions that have been made. etc. .
'structural models'. Basic Concepts of Econometric Models The 'reducedform model' can be a more general and dynamic model with lags of the dependent variable as explanatory variables too. Later on. These three types of models have different properties. are di scussed (see Chapter 9). which are models that have also endogenous explanatory variables.18 Chapter 1. It will be shown that the choice of an estimator that has to be used is dependent on the stochastic properties of the model specification. . Their parameters have to be estimated in a different way and the used estimators have different properties.
The cases have been formulated in a way that lets the student perform empirical econometric research. by using the provided real economic data. timeseries or crosssection data are analysed and modelled. which is in contrast to the mentioned exercises. The reader can reproduce these examples . It could be more interesting or exciting to use data that has been collected from your own research instead of the provided data sets. but they give rise to economic questions that can be answered with econometric modelling and statistical testing. it is not necessary to use the provided data. A number of files are offered that will be sufficient to use on many aspects of the discussed econometric themes. Then you can kill two birds with one stone: studying applied econometrics and starting your own research at the same time! In the next section.uklvogelvang. In an applied econometric research project. not yet in an autonomous way but according to the questions or assignments fonnulated in the case. The cases do not concern mathematical or statistical exercises with one solution. a description is given of the available data that will be used in the cases. The aims of this book are not achieved by including a huge number of data sets.Chapter 2 Description of the Data Sets an Introduction to the Cases Download the data sets at www. Of course. The data files are Microsoft Excel files that can be downloaded from the accompanying website. 2. In fact it is simulated empirical work that is guided by the points that are mentioned in the cases. the data set that contains the timeseries data that has been used in the examples throughout the text will be introduced.co. Both types of data come up in the relevant chapters. These cases are an introduction to perfonning empirical econometric research.pearsoned. Sometimes a choice between a couple of sets can be made.1 Introduction In this chapter. The result of econometric research is not one unique answer.
What to do with the crosssection data set is formulated in the relevant Cases 4 and 9.2 Data set 1: commodity prices The timeseries data that will be used in the examples originate from former research on international agricultural commodity markets as done by the author in the past. 2.2. coffee and tea: • the coffee price is the Composite Indicator Price (according to the definition in the International Coffee Agreement 1976) in US$ cents per pound. . Each research project has its own work:fiIe. therefore the reader is advised to use the examples given in the text when studying tbe first six chapters. The data set has been updated with recent data and data are available in the tile pcoccoftea. price movements on the cocoa market were observed that were not caused by circumstances in the world cocoa market. the procedure for importing data in an EViews ' workfile' is described. • • The data are from the UNCTAD. The complete first exercise starts in Case 2 at the end of Chapter 6. Description of the Data Sets and Introduction to the Cases by using the data which will contribute to a better understanding of the text. to a certain extent. cocoa (beside other uses) and tea are beverages that may be considered. In Section 3. Coffee. for example. Subsequently. When studying the price formation of these commodities. This data set concerns the following monthly data on the price of cocoa.20 Chapter 2. the cocoa price is the Cocoa NY!London 3month futures price also in US$ cents per pound and for the sample period 1960(01)2002(09). Another important feature is that coffee and cocoa bave been involved in an international commodity agreement (w ith varying success). This is explained in Remark 3. An important . for the sample period January 1960 to September 2002. it is wellknown that some of the commodities are often traded on the same futures exchange. In sections with timeseries data an introductory guidance is given conceming possible models that can be estimated. If you want to include any variable from a different worktile in the work:fiIe that is currently being used. one of the data'sets will be selected for a data analysis as formulated in Case I. That is quite a way ahead. Monthly Commodify Price Qul/etin. Vogelvang (1994). It is possible to exchange variables between various workfiles of EViews. like the New York Board of Trade. The following considerations have played a role in analysing a relationship between tbese prices. after the specific models for these cases have been discussed. see.5.6 in Section 3. as substitutes for each other. In Chapter 3. Further. So you are not limited to using only variables from one of the offered data files that are discussed in thi s chapter. which is also a reason for price interrelationships. the tea price is the London price in US$itonne for the sample period 1960(01 )1998(06).property of these commodities is that they are all commodities that can be stored.xls. The idea behind these data is as follows. Tea is traded unrestrictedly. but probably by economic developments in . then that is no problem. In an easy and convenient way variables can be copied from one EViews workfile into another. the accompanying data sets are introduced in consecutive sections.
3 Data set 2: macroeconomic data Macroeconomic data sets are offered for the USA. They concern quarterly data for sample periods between 1980 and 2003. For example. The data on the commodity prices show more or less similar developments in some subperiods. sample period: 1980(0 1)2003(0 1). The files with OECD data are defined as: • • • • • Macro _USA. sample period: 1979(04)2003(0 1). and the Bureau of Economic Analysis (BEA).oecd. http://www. 2.xls. MacroJapan. sample period: 1980(0 1)2003(0 1).orgl.xls. The data are seasonally adjusted. More facts about the prices are given in the examples. a simple relationship can be used to illustrate many of the items discussed in several chapters.Data set 2: macroeconomic data 21 the world coffee market. The selected variables. The data files are Excel files. . These data have been collected from the website of the Organization of Economic Cooperation and Development (OECD). sample period: 1980(0 1)2003(0 1). Australia and Japan. government final consumption expenditure (gov).xls. The real data of the USA are expressed in the currency value of 1996 and for the other countries in the 1995 value (A ustralia 2000(0 I The units of the OECD data are millions of the national currencies. These variables contain the nominal values in domestic currencies. are: • • • • gross domestic product (gdp).1 + Ut2 I'll Yt =Ct+ lt +G t . Macro _UK. These data sets can be used individually to estimate simple models for the economy of one country. http://www. the high prices of all the commodities in 1976 and 1977 were caused by occurrences on the coffee market (see Figure 3.3. but they can also be used together to estimate one model (a SUR model) for a number of countries simultaneously. France. but for Japan are in billions of yen. Where real data are also provided.20). of course beside specific price movements of the particular commodities themselves in other periods. etc. MacroAustralia. Ct = It = + Pll yt + Uti 1'21 + 1121 Y t + 1'22Yt. sample period: 1979(04)2003(01 ). For instance. r _ cons.xls. the UK. with their names in parentheses as they are shown in the files. MacroIrance. investments (private and government) (inv). For these reasons. private final consumption expenditures (cons) .bea. as will be introduced in Section 8.doc. as a bas is model a simple SEM can look like: ».gov/.xls. you always have to create your own EViews workfiles. for the USA and Japan the units are in billions. the names are preceded with an r as in " _gdp.
'price_defLpc': price deftator for personal consumption expenditures. 1947(01 )2003(02). The data is used again in Case 8 where system estimation methods are applied to estimate the parameters of the SEM simultaneously. An exercise to model one of the equations is formulated in Case 2. It concerns quarterly data (at annual levels) for the same and for some more variables observed in the period 1946(01 )2003(0 I).. The variables are measured in billions of US doUars. 'price_gas_oil': price index (1996 = 100) of gasoline and oil.xls. AU the data have been obtained from the NIPAtables. Dynamic consumption and investment equations can be estimated with the provided data.xls contains monthly data on crude oil prices and retail gas prices. One more GDP variable is found in this file: gdp.xls. 1947(0 1)2003(02). Description of the Data Sets and Introduction to the Cases Parameters of endogenous variables are denoted by f3ij S. These data are obtained from the National Income and Product Accounts tables (NIPAtables) of the BEA. The data will also be used in other more specific exercises. and Oil. and parameters of exogenous or predetermined variables by l'ij S. 1947(0 1)2003(02).. They are seasonally adjusted. With this more extensive data set it is possible to refine the macro model and to investigate this long sample for the presence of structural breaks. One more data file concerning the USA has been included for a long sample period: Macro_USAJong.xls. and measured in billions of US doUars. 'quanLvehicles': quantity index (1996 = 100) of vehicles and parts. The file oil. More information about the data.gas. 1947(01 )2003(02). 'price_vehicles': price index (1996 = 100) of vehicles and parts. The net exports of goods and services are also included.nso'] representing the GDP that is not seasonally adjusted and measured at quarterly levels. 2. Case 5 is an exercise with a SUR model.goyjny) investment.USA.goSJlso. can be found at the mentioned websites. for the following variables: • • • • • • • 'exp_vehicles': expenditures on vehicles and parts.xls. such as the computation of index figures and the seasonal adjustment. 'quanLgas_oil' : quantity index (1996 = 100) of gasoline and oil. 1946(0 1)2003(02). Two separate variables are inc1uded for the investments representing private (grossprivareJnv) and government (gross. and of course more data. at annual rates.USA.xls contains quarterly observations that are seasonally adjusted at annual rates..22 Chapter 2. 'price_de fLdg' : price deftator for durable goods. The file Vehicles. These data have been obtained from the website of the Energy Information Administration . Also included are the GDP and the real GDP from the file Macro _USA _long. With this data set the expenditures on vehicles in the USA can be analysed. A 'discrepancy variable' has to be added to make the identity balanced as the involved variables are not all the variables in the 'national accounts ' that determine Yi . 1947(0 1)2003(02). 1947(01)2003(02).4 Data set 3: oil marketrelated data Two data files related to the oil market in the USA are provided: Vehicles_USA.
but as the GDP was only encountered as quarterly data.Data set 4: money market (EIA). an economic model has to be formulated and these data can be analysed. Ln cents per barrel : 1973(10)2003(04). (p_ unlead_ prem i um). III cents per barrel: 1973(10)2003(04). In Case I. an interest variable. A lot of economic data with long timeseries can be found in that database. The original monetary data from FREDll have a monthly frequency. in tenth cent per gallon: 1976(0 I )2003(05). not seasonall y adjusted (mI. ' leaded regular gasoline. Louis that can be found at http://research.5 Data set 4: m oney market A file with monetary data from the USA has been included to analyse the demand for money. GDP. after which an econometric model will be estimated in Case 2.orgl. M2 and M3. whereas in Case 2 an econometric model is determined. The following variables are included in the Excel file: • • • • • • • 23 'crude oil domestic first purchase price' in cents per barrel: 1974:(01)2003(04).gov/. (pJ eadTeg). US city average retail price'. 'unleaded premium gasoline. With the data from this file a similar analysis can be done as given in the examples. the same variable as before.B. The following variables are avai lable for the period 1959(0 I )2003(02). US city average retail price'. all the data has been rendered into quarterly data. US city average retail price'. 'landed cost of crude oil imports'.doe. (p_ crudeoil _ . • • • • • the money stock MI. seasonal adj usted (price_deJLgdp) . (landed_ imp) . cost of crude oil imports' . but now concerning a relationship between world market prices of crude oil and relail prices of gasoline. 'all types of gasoline. US city average retail price' . price index of the GDP. For comparing these prices be aware of the different units of the variables: I barrel = 42 US gallons. If these data will be used. (p _ all). http://www. m2 . 2. The money stock will be used as an approximation for the demand for money that can be explained by the GDP and the interest rate. The data have been obtained from the economic database FREDll of the Federal Reserve Bank of St. not seasonal adjusted (cpi ). 'unleaded regular gasoline.eia. then the formulation of an economic model for the money demand is done in Case I together with a data analysis.stlouisfed. m3). us) ' EO. The selected monetary data for perfomtiog a case concerning the American money demand is available in the file: moneyJtsa. consumer price index. in tenth cent per gallon: 1973(10)1991(04). whicb bas been seasonally adjusted at annual levels (gdp). in tenth cent per gallon: 1981(09)2003(05). in tenth cent per gallon: 1978(01)2003(05). the 3Month Treasury Bill market rates (tbr).x/s. (p_unlead_reg). . (fob_imp).
industry) . The dummy variable ' i ncentr' is equal to one if the firm is located in a central region and otherwise equal to zero.s. The dummy variable ' extech' is equal to one if the firm acquires external techno· logical knowledge and otherwise equal to zero.s ). The dummy variable 'rdcoop' is equal to one if the firm is engaged in R&D co· operation and otherwise equal to zero. The files are: pcdata. That case is an application of the subjects that are discussed in Chapter 8.xls are described in Section 11. the variable 'percentage of sales of products new to the industry' (new.xls and business. the probability can be estimated that a firm needs external technological knowledge given a number of characteristics of that firm.5 for a model with a qualitative dependent variable.24 Chapter 2. 1\vo cases are formulated in Part III.firm). a topic which is discussed in Chapter II. they only use the data in more simple formulated cases. number of employees in full·time employment (emp/). Description of the Data Sets and Introduction to the Cases 2.indus/7·y) will be explained by a causal model. cross·section data 1\vo data files with cross·section data are used. • • • • • percentage of sales of products that are new to the firm (new. where some of the variables from the original file will be used. The following variables will be used. In Case 9 (Chapter II) we will analyse whether the need for external technological knowledge (the variable ext ech) can be explained by variables from this data set.s). R&D costs in proportion to sales (rd. A number of dummy variables are included in the data set: • • • • The dummy variable 'modern' is equal to one if the firm is a modern company and equal to zero if that is not the case. The cases are not corresponding to the original study. In Case 4.6 Data set 5. percentage of sales of products that are new to the industry (new. With such a model. Modelling of the dummy variable ext ech concerns the problem of estimating a model for a qualitative dependent variable. The other data file consists of businesseconomic data that will be used for cases in Section 8. sales growth in percentages between 1990 and 1992 (gr .s. .xls. Ridder and Kleinknecht (2002).4. The simulated data in the file pcdata. The business economic data originates from the study by Van Montfort. One file concerns simulated data that is used in an example in Chapter 11. The concept of dummy variables is explained in Section 7.s.5. The file consists of cross·section data from 2886 firms in the Netherlands that were collected in 1992.6 (heteroskedasticity) and in Section 11 .
. a preliminary data analysis will be done in this chapter. We will see how graphs can be made. A relationship between these variables yields a simple but useful 'model ' to illustrate applications from econometric theory.Chapter 3 Basic Concepts of EViews ~Starting the Research P . correlations can be computed. A complete description of all the procedures in EViews is found in its help file. etc. As this book offers a first course in econometrics with the accent on showing the use of econometric methods in the economic practice of doing research. elasticities can be calculated. especially pictures.1 Introduction In thi s chapter the use of a number of general basic procedures from EViews is introduced as a start to learning the principles of thi s software package. only the most important or relevant procedures are discussed in detail. However. In fact. The help function in EViews is very instructive and is identical to the EViews 4 User's Guide that is also available as a pdf file under the help button. we will look at the output. as introduced in the preceding chapter. Besides the use of EViews. you will be able to use EViews autonomously. Many examples of EViews procedures will be given by using these data. data can be seasonally adjusted. After becoming acquainted with EViews in thi s and following chapters. that can easily be imported in a text editor. this book is Dot an EViews manual. will be used. 3. For thi s purpose the data on worldmarket commodity prices.1 of EViews is used to describe the procedures in thi s book. Version 4 .
start again by clicking on 'File'. so a workfile has to be made for each data set. tables. The names of the variables can precede the series in the data file or can be given when importing data in EViews. which will be equal to one and 'End observation'. ' Undated or irregular' concerns the import of crosssection data.26 Chapter 3 . It is also possible to double·c1ick on the name of the workfile in a filemanagement program like Windows Explorer or Total Commander (Ohisler (2003). g raphs. this is shown in Figure 3. An existing workfile will be opened in following EViews sessions to continue the research project. start EViews in Windows. but additional data can be imported at any time. Then EViews starts with that workfile. Various data sets will be used in thi s book.coml)if theextension(. The data will be imported in the workfile in the first EViews session.wfl) has been associated with the EViews program. If cross· section data has been selected the range changes in ' Start observation'. The names of recently used workfiles are visible in the history list and one can select directly a workfile from thi s list. For every research topi c or project a workfile is created. the data are organised by observation. which is equal to the number of the last observation..ghisler. like series. The data files can be organised by variable or by date. So the first thing to do in EViews is to create the workfile and to import the data that creates the first objects: the variables as a number of series.. groups oj variables. 1.2) with questions about the data: the frequency and the sample period. step one: the new object . Basic Concepts of EViews and Starting the Research Project 3. In the example below. EViews can read various types of data files. g Figure 3. A window appears (see Figure 3. First. see: http: //www. The following points bave to be done to start working with EViews. the names are in the file. Most of the options concern time·series data. All these objects are part of the workfile.. equatiolls. Next click on 'File' and 'New'. In a next EViews session.2 The creation of a workfile in EViews The basis in EViews is the workfile.1: Creating a workfile. you have to know how the data has been organised in the original data file that has to be read. and select 'Workfile' . 8. To import the data.. An EViews workfile contains objects. and the data are in an Excel file. 1 that a history list exists with the names of recently used workfiles. etc. Notice at the bottom of Figure 3..
two variable names appear: a vector (c) that will contain estimation results.The creation of a workfile in EViews 27 F~~~~~~~~ r ltmi . • fle_ 5"". and a series with the name resid that .... WOIkfie{".. and click OK. o. In this window. ~lfnh.2: Creating a workfile... . Urrl .. . lL~ ~ Dalf IS .. step two: defining the frequency For timeseries data.wfll Figure 3. The window in Figure 3.a .) dew.. step three: saving the file in a directory . Three series of monthly data as mentioned in the introduction will be imported. I Figure 3.....'.3: Creating a workfi le. mark the correct frequency and fill in the 'Start date' and ' End date ' ..J · 14. So select 'Monthly' and type the ' Start date' and the 'End date' .3 shows the new workfile. .uel.
. If the data are not neatly organised in a matrix. this window is shown in Figure 3. Click on ' Procs' and select ' Import' and then select the type of data file: a database file. the data can be imported in the workfile. one for reading an Excel file in Figure 3. do not mark 'File laid out as rectangle'. coffee and tea. That is usually easier than browsing through the directory structure to find the workfile every time you continue working with EViews. If the data file contains the names of the variables.6. Next ctick OK. among other things. so that it is possible to check whether the correct data will be imported and to see that the data are organised by observation. an ASCU file or an Excel fiJe (which is a convenient way to import the data) as shown in Figure 3. It is convenient to check the box 'Update default directory' at the bottom of the 'Save as' window : then EViews starts in the same directory in a followup session when the program is started in general without tinking to a file.5. and change the name 'Untitled' to a suitable name.6.5 and one for reading a text file in Figure 3. The vector c will be used for testing restrictions on parameters or for specifying equations.4: lmporting data in the workfiJe wiU contain the residuals after the parameters of a model have been estimated. EViews 'asks' whether the data file has been organised by observation or by series (in 'Columns' or in 'Rows' ). Basic Concepts of EViews and Starting the Research Project Figure 3. In Figure 3. In Figure 3. and browse for the data file that has to be imported. In this window one has to fill in similar information as mentioned before concerning Figure 3.5 we see that the data are in sheet2 of the Excel file. Two screen shots are given.wfl . In this example. the data that will be imported are three series with the monthly worldmarket prices of cocoa. So select 'Read TextLotusExcel '. The residuals in resid will be used to test statistical assumptions. organised in columns. The workfHe gets the extension . Ctick 'Save as' from the file menu. There is a preview window. At this stage.28 Chapter 3.5 .4. The names are in the fi rst row. When the data are in a text file a different window is opened. so the first data cell is cell 82 (the first column contains the dates).3 the name cofcoctea has been chosen. only the number of variables in the file has to be entered. The data on the commodity prices are read from an Excel data file as shown in Figure 3.
.5: A~.00 27.19 32.rll T mel IlllpOII CI'I.75 32.t.14 1244.72 33.60 27..".221265.76 1391..581350.••lbofare dU: pcccoa ptea 32. flllChir . Figure 3.The creation of a workfile in EViews 29 ""'Figure 3.6: Importing data from an ASCII file .69 28.60 .90 27.aI h! ... ..mp~ T6ondal_ arl Importi ng data from sheet2 from an Excel file EJ tnomed I .54 32.26 1281.40 26.
Remark 3. Statistical and graphical procedures can now be used to analyse the variables. if you are interested in estimating a model for only part of the sample and forecasting the dependent variable for the rest of the sample then you have to adjust the sample. Remark 3. etc.1 Remember that the price of tea is given until June 1998 only. Basic Concepts of EViews and Starting the Research Project pcocoa pcoffee plea rssid Figure 3. Some remarks are appropriate at this stage. This means that the last part of the series is filled with a code for missing ('Not Available') observations: NA. for example. but that depends on what has to be done. are included as objects in the workfile.7: The workfile with the imported data If all has been done in a correct way the names of the variables appear in the workfile window as shown in Figure 3.30 Chapter 3. it is still possible to import more data later on. The sample can be redefined by clicking the button 'Sample'. otherwise wrong names like serOJ. These objects have to be deleted and a new attempt must be undertaken. But. The range can be changed (under 'Procs') or new variables can be imported. In the displayed window they are equal.7. ser02. The range is the entire sample that has been imported into the workfile. After the data have been successfully imported.2 Observe that a 'Range' and a 'Sample' has been defined. .
Select 'Store to DB' under 'Object' in the 'Series window' . where data from a number of workfiles have to be combined. One word of advice: do not wait to save the workfile until the end of the session. That is done by opening more than one workfile inside the same EViews window. For example. so that everything that has been done (transformations of variables.8. You always recognise the conversion in the series sheet as shown in Figure 3.The creation of a workfile in EViews Remark 3.6 In Section 2. This can be solved in the following way. Browse for the correct directory and click 'Yes'.DB? files' in the 'Store window' that has appeared. This variable can be imponed in the quarterly workfile by selecting 'Itnpon' and 'Fetch from DB'. see 'Conversion Options' under 'View' in the 'Series window' . Click with the righthand mouse button on the variable that has to be copied. EViews automatically computes quarterly averages of the monthly data.3. Remark 3. Impon the money stock data into the monthly workfile. it is useful to create a real database for many economic variables instead of individual variables stored in any . specification of equations. Remark 3. The space for this window can be adjusted by dragging the bottom line of the command window up or down. The way to display a series is discussed in Section 3. but dick regularly on the save·button of the workfiJe.5 At the end of the session the workfile has to be saved. monthly and quarterly data. The money stock has been observed monthly. for example. Make a work/ile defined for quarterly data and another for monthly data.7 It is possible that you have data with different frequencies.4 It is useful to click on the button 'Options' to instaU or redefine default options. Select 'Individual. whereas the other variables have quarterly data. as it would be a pity if your work gets lost when your PC unexpectedly hangs up! Remark 3. the option 'Graphics Defaults' is used to make a default installation for the appearance of figures. this has been done with Data set 4 conceming the money market. for example. More con version options are available. etc. and select 'Copy' . When EViews is used in a more advanced manner. where commands can be typed in.db. and select 'Paste'.3 The blank space on top of the EViews window is the command window. in which case EViews can conven the data to a common frequency.) is available in the workfile in the next session. maybe formations of groups. Each result (object) can be sent to the printer for a printout by cUcking on 'Print' in the toolbar of that particular • window. This might be a useful action in Case 5. For example. Examples are given later on.1 it was noted that variables can easily be exchanged between different workfiles. 31 Remark 3. The variable is saved with the extension . Next activate the other workfile by cUcking with the righthand mouse button in an empty space of that window.
the option to make a 'Graph' of the cocoa price series has been selected. you will see what these options result in. Before using the 'View' button see the 'Edit + j.3 Viewing variables and some procedures In the upperleft corner of the window of any EViews object two important buttons are found. We will look at the results of some possible options that can be chosen in thi s window. and its window is opened by double clicking on the series name. the buttons 'View' and 'Procs' . With the options under 'View' you can look at various representations of the object.9. Of course it is Dot the intention to discuss all the options that are available. More specific options that are not discussed in this section can be encountered in following chapters. The options under ' Procs' provide a number of procedures that can be applied on that particular object. Basic Concepts of EViews and Starting the Research Project Figure 3.like a spreadsheet. Using 'View' in the 'Series window' options is useful for data analysis involving graphs. where part of the spreadsheet is visible. In the 'Series window' procedures like seasorial adjustment are found that can be used when quarterly or monthly data have been collected showing a seasonal pattern.9. .' button in the series window. After clicking on that button the series window will look like an Excel sheet where the series can be edited. which is useful when errors in the data occur. Their use will be illustrated by applying some of the options on the imported commodity prices. etc. In this section. First.8: Example of frequency conversion when importing data directory. A variable (series) is selected by one click. for example. but only those which are most frequently used in a research project. descriptive statistics. whereas an explanation of the remaining options can be found in the EViews User's Guide. See the window in Figure 3.32 Chapter 3. In that figure. 3. After double clicking on the series name the series is seen in a new window (the 'Series window'). or when new data become available. More detailed information about the use of a database can be found in the EViews User's Guide. the available options are considered that are found under 'View' and these are shown in Figure 3.
.. Figure 3. The figure can also be included as a graph object in the workfile by clicking on 'Freeze' . the 'Unit Root Test . ~~~::'·~. The appearance of this graph can be changed by double clicking on any part of the graph you would like to change. The graph is shown in Figure 3. LaTex or Scientific WorkPlace... as shown in Figure 3. However."'Z=3~.9: Preparing a graph Also look at the other possibilities under the 'View' button... Figure 3.Viewing variables and some procedures . The meaning of every tab speaks for itself.11 the tab 'Legend' has been selected. namely 'Cocoa price' instead of PCOCOA. 10..~. where among others the nameofthe variable has been replaced by a suitable name that is understandable in a publication. others will be discussed later on.. .. r. 1M for O"ccic!'he St.10 does not have the appearance of a figure an author would like to have in his article.. like the labels. or the graph itself.I I 33 v" . When you are satisfied with the result. lltalsldan lIriIR••U . Some of them are obvious. the graph can be imported into a text editor like MS Word..11. ' will be introduced in Chapter 12... for example. IDS b.... In that window a number of tabs appear corresponding to the various parts of the graph. or the axis.~· . In Figure 3.. After such a doubleclick the 'Graph options window' opens...
II.PCOCOA Figure 3. • ...J~t!!'51E~~l!'er:5"~E5··· 200~r. 12 where a graph object can be found.. 160 120 80 40 O~~~~~~~~~~~~~ 60 65 70 75 80 85 90 95 00 .1 2 . The picture can be saved on disk to be imported into the text file that contains your paper afterwards.11: Adjusting the graph .. .34 Chapter 3.' 1'1 !II 11. ConIoo . In Word it can also be copied via 'cut and paste' .... 1S I ~52 IIiItt X1008 Figure 3..J Select "*" lo d: 8.\ W"r~I... ' " "..· I 1111 II[ 11" Rril'3 L.j.. r Loft . see Figure 3..10: A graph of Pt DC and giving il a name with 'Name'. Basic Concepts of EViews and Starting the Research Project . This yields a window like the olle in Figure 3. To save the file on disk you can use both the series window and the graph window. Click 'Edit' in the upper toolbar and select 'Copy'. ~ Ago C II.
J. and compare Figure 3.. ..._ . ••• . ...Viewing variables and some procedures 35 25°Tll~r~~~~~~~ .13 with Figure 3.___L_. The label name bas been changed into a clear name for the text.~. Gridlines have been plotted and text has been added at the bottom of the picture..14.wmf file.wmf file in a directory Click OK and browse the directories to select the directory where the file has to be saved. . The Yaxis has been manually scaled.•+ . the label has been dragged into the figure and bas got a box. + ~. Figure 3. is given in Figure 3. .. _l_.. ii' ! . ! Cocoa ++.12: Saving the graph as a ..13 all the changes that bave been made. .13. In a similar way. the procedures that are found under ' Procs' can be used. i •.J. ! . Observe in Figure 3. which is a ..1 D. The procedures that can be applied on a variable are shown in the window of Figure 3.•. • • : I I _.1. An example of an adjusted and imported figure of the cocoa price..
]1 · .36 Chapter 3. 1960.13: The graph of pr e as imported in this text Figure 3... .::I==c==:::r:::._ c+'! +_L 95 00 60 65 70 75 80 85 90 The cocoa price. ~~_+.9 Figure 3.2002 ..I IT. Basic Concepts of EViews and Starting the Research Project I ~ oc?a price 1 '+!+'r ' 200y44f.14: Procedures in the 'Series' window .~.1 . I i i II 250 1 1 ! 1 !1 . 150 +~~_+ 100 +__~~_+ 50~ I I' .
Various objects can be opened as a gro"p of objects. for example. In Figure 3... For example. itore 10 DB ••• Figure 3. [t is a rather useless option for this variable. We have seen the available options that can be selected under 'View' and 'Procs' in the 'Series' window."'c has no seasonal pattern. as we have seen in the graph . but also a group of graphs can be created. which becomes a new graph. procedures can also be found to analyse the data. Different options appear by clicking with the righthalJd lIIo"se b"llOIJ on the names of one or more marked variables. here p[oc The procedure that has been selected is seasonal adjustment. we have created the object 'graphcocoa' in the example shown in Figure 3. Displayed in that example is the way to delete an object from the workfile. Also other objects are created in a worlcfile.Viewing variables and some procedures 37 l. Seasonal adjustment is one of the topics that will be discussed in Chapter 7. The contents under buttons like ' View' and 'Procs' is not identical in the windows of each object: it depends on which procedures make sense for the data in that particular window.!Pd"" &am DB. etc.15. The 'Group object' is a very useful object. as P. In an 'Equation' window other views and procedures are avaiJable than those in a 'Series' or a 'Group' window. The use of a group .15 an example is given of a rightclick on the cocoa price. 17 later.15: Options for an object 'Series' . Under 'Quick' in the upper toolbar of the worlcfile. compare the options of 'View' in the 'Series' window with those under 'View' of the 'Group' window in Figure 3. As an example. a group of variables.
.16. For example. It will become a 'Group' by selecting 'Open' (under 'View'). it is not necessary to save a group with a name because a group is very easily created again and the workfile will become too disorderly and large by including every object that has been created.. 08··· Objocl_ . to compute group statistics. sample correlations.b .. i . a series) is selected with the lefthand mouse button. Notice in Figure 3. Another convenient and fast way to create a group is clicking with the righthand mouse button on the selection of variables. and selecting 'as Group'.16: Creating a group with PtOc and p~oJ of variables is very convenient during an EViews session. See also the other possibilities in Figure 3.g... In many cases. If an equation is specified in this manner the order of the selection is important.38 Chapter 3 ... The first variable that is selected is the dependent variable in the regression equation.... To create a group of objects keep the 'Ctr\' button pushed down and click with the lefthand mouse button on the relevant object names. The first option 'Group Members' is useful in order to alter the members of the group. How is a group created? Only one object (e. DB.17.16 the 'Open' option 'as Equation . The different views of the group are various types of graphs. etc.... ' .. However. A display of the options that are avai lable under 'View' in the groups window is given in Figure 3. after wruch the same options appear. existing members . descriptive (group and series) statistics. this is a subject that will be discussed in Chapter 4. Selection of this option yields a static regression equation in the 'Equation Specification' window. Basic Concepts of EViews and Starting the Research Project J. etc. _""iph 1m lSI Figure 3. it is possible to make all kind of multiple graphs....
. ". before including it in a paper. Figure 3. Co. caused by the devastation of colfee trees in Brazil as a result of very cold weather..':w. Figure 3. '1*'. The variances of the two prices are much higher at high prices than at low prices.~n % .18.Viewing variables and some procedures 39 . Jusllike the corrections made for the graph of PtCDC . HoW_TeM .~ •. ~ . 1. Cui ' . 17 a scalier diagram with a regression line has been selected. The resulting scatter diagram is given in Figure 3. The next option that is selected in 'View' is 'Multiple Graphs'.. Furthermore you can observe a clustering of points at low prices and an increasing variance around the regression line when prices are high. $&Itt T_of.121 " Coil". from an economic perspective. The high prices (and heavy price changes) correspond 10 the disturbed and uncertain market during the I 970s. 18 clearly shows that a linear relationship between the two prices exists.T• . The interpretatio n and the consequences of this last observation will be considered in Part Il.. The choice 'Line' in 'Multiple Graphs' yields a picture with two separate graphs in one figure. These observations are. After manual .QeecI. In Figure 3.17: The contents of 'View' in the 'Group window' and the selection of a 'Scatter with Regression' can be deleted from the group and new members can be added 10 the group. '%"I(1) ~ ag.. not unexpected.E.. it is better to make cbanges to this picture 100..
! ! i i1Ii'  I .20 clarifies the idea behind the relationship between p.'''~='C==C==~ 300 i . It is also possible to plot both variables in one graph. It will be clear from the appearance of Figure 3. 200 160 g a.1. PCOFFEE 2~1.' t . I ! . . 250 'j"!' ' ' ' 200.1 i i ': I: 1 I I I! J_.~ . rI' I ..'++'_1_1 250 .. 150 100 "t' 00 " I 0. 300'r~.eoe and Pto! that is used in the examples in this book..i":. ~~~~~1~~'~~'~~~~~ 70 75 80 85 90 95 00 60 6S I' " . .1 ' I 300 l ' .'~hlo. (. . " _. ..IIJlIIII 11/\    pii~IW I .40 Chapter 3.20 that adjustments have been made to the initial picture.ce 1 300 rf~'i .CoHee pr. 19.18: A scatter diagram of ?foc and p~J scaling and some other adjustments the graph has been imported into this text as shown in Figure 3..''''JI' Ilfllllllll W..~1 " 00 Figure 3.+It : 95 04_"_''<:.'.) ~ 120 00 ~ O'+r'~r~~~~ o ~ 00 120 160 200 PCOFFEE 2~ 200 320 Figure 3. _LI'_ )_. . · · 150 100 'tr+'. t ' • ' : : ' · I .' . Select the option 'Graph' in the 'Group' window and both variables are plotted in one graph as shown in Figure 3.20.19: Multiple graphs of p{oe and Pfo f made from a group. .1! .: 'i !t.r ! 11. ~ra~~~~ . Figure 3.LJJ t::::j::::J'~l~~' =~·l 60 65 70 75 80 85 90 i !I " .. ' ~ "'1" ii''S PCOCOA .1 . ." r=='======~ _l".ce 1 1.r" ~. 1.:::'~~·'F i Ii! iii j L :. It is obvious to see that the two prices have a similar pattern. I i i i 2OO~_. Both variables have the same range (0350 $ctllb) by manual scaling of the Yaxis.+1.l'''''l'''''+4 . identically scaled . '1'+1S::. '[ 1. Basic Concepts of EViews and Starting the Research Project _I. I 50 i I O\I j~.Cocoa pr..~.iW __~ '++!+I . ~_ _ _ 1_1'0 __ . ! '1 ' + f lI ..+ ' .
the workfile sbows that the three price variables bave been selected as a group after which the correlation matrix and the group statistics (select 'Descriptive Stats') have been computed. it frequently happens that variables have to be transformed. is to select one of the various buttons and to try out the options. +·+IH 200. The best way to learn the various possibilities of a procedure.20: A combined graph of from a group p[oc and p tCo / made whereas the 'basics' for the price fannatian of both commodities are not identical. In Figure 3. The structure of successive windows is logical and.J. When this correlation matrix is computed for a number of subsamples of the range. or that some number (a scalar) like an elasticity has to be calculated. remarkable changes in the values of the correlation between the price of tea and those of coffee and cocoa can be observed.·······f150 + /.22.21 .~c~o~c~o~a£~ric~e:J P 250~+. Finally two more examples of EViews output are given. 3. the help function is adequate. or to see the possible view options.4 Some basic calculations in EViews When performing empirical econometric research. The prices of cocoa and coffee are rather strongly correlated. After these examples from the 'View' options. Review Section 2. we look at the procedures under 'Procs' in the 'Group' window. from the group statistics the common sample has 462 observations. See these possible procedures in 'Procs' on your own PC. Figure 3. Variables can be transformed or generated with the command 'genr' or by using a . The total sample has 5 13 observations and. as seen in Figure 3. whereas the correlation between the coffee and cocoa price remains stable.2 again for some arguments as to why these prices move together. Because the 'common sample' has been used it was not a problem to include the price of tea in the group too.Some basic calculations in EViews 41 350Tl!~_r~~~~~~ Coffee price 3001··!t·tl==. Clicking on ' Procs' has made these group procedures available. This is useful knowledge regarding the analysis of the relationship between these prices. as mentioned earlier. the price of tea correlates less with the other prices. if necessary.
These transformations are easily performed in EViews. Later on in the course. A scalar can be computed with the command 'scalar' . A wellknown transformation is the log transformation. For example. because the command can be adjusted to tran sform another variable. or genT dlpcocoa = lpcocoa .21: The correlation matrix selecled in the 'Group' window 'Genrbutton'. Examples are given in this section. A second possibility to compute transformations is by clicking on the .42 Chapter 3 .lpcocoa (1) . Basic Concepts of EViews and Starting the Research Project Figure 3. models with variables specified in first differences will also be encountered. which can be an advantage if one has to do a number of similar tran sformations for more than one variable. two new variables lpcocoa or dlpcocoa are generated by typing: genT lpcocoa = log (pcocoa) . The command window keeps a record of the commands by maintaining a history list during an EViews session . Transformations of variables can be specified in various ways. The command for a transformation can be typed into the command window at the top of the EViews window.
Next the window 'Generate Series by Equation' opens. The specification of log(pcocoa ) or d(pcocoa ) . Both possibilities of creating the log of the cocoa price are sbown in Figure 3. For example in an 'Equation Specification' window operators can be used. The log and difference operators are used in the following way.22: Series statistics of aU the variables in the group button 'Genr' in the workfile. All these transformations enlarge the contents of the workfi le.Some basic calculations in EViews 43 Figure 3. Often it is possible to use an operator within a window. using the same example as above. In that window only lpcocoa = log(pcocoa) has to be specified. It is not always necessary to transform variables in advance by creating new objects. In an equation specification window the model is specified and an estimator is chosen to estimate the parameters (discussed in Part 11). But in this way no history Ijst is created.23.
and the righthand mouse button can also be used. CONS'_ I' (3. As mentioned earlier.23: lllustration of two possibilities to generate In (PtCOC) in the window implies that the logarithm or the first differences of the cocoa prices are used in the model.1). The computation of an income elasticity will serve as example.44 Chapter 3. = e( l ) + e(2). the estimation result is explicitly written in the following way. + e(3 ). because in EViews the name C is used for a vector with estimation results. The name CONS is used instead of C. as immediately seen below in equation (3. Y. CONS. In the notation of EViews. the option 'Delete'. The tran sformations d(log(pcoeoa)) or dlog(pcocoa) yield the first differences of the logarithms of the cocoa price in the model. Not only transformed variables but also scalars can be computed. Basic Concepts of EViews and Starting the Research Project gem Ipcocoa=log(pcocoa) Figure 3. And it is even possible to combine the two operators. 1) . ' Delete' is also found in the toolbar of the workfile. without generating a new variable in the workfile. Suppose that the parameters of the following dynamic consumption equation has been estimated. Other manipulations with objects can be found under ' Objects' in the upper toolbar: for example.
which yields the variabl. Doublecticking on the scalar name results in showing its value at the bottom of the EViews window.6) is: eLy = (:h . 197. (See the EViews User's Guide for other available statistics. Y (3.3292« 0. Y SA CONSSA(·l) c 1842. rrr . (3.24 a part of the estimation result of the consumption equation (3. e(2) and e(3) contain the numerical values of the estimated parameters.:11964:4 Dependent Variable: CONSSA Method: Least Squares Date: 01121i04 Time: 10:22 Sample(adjusled): 1974:21984:4 Included ObS8M1lions: 43 after adjusting Variable eoaftielenl SId.. So the sample mean of the series Y is known as @m ean (Y ) and the sample mean of the series CO N S is known as @mean (CO N S). . The two variables cons.) Compute the income elasticity by typing the following expression in the command window: scalar eLy = c (2) • @m ean(y) /@mean (cons).4). with the symbol # as prefix.9.593563 Figure 3. It is simple to compute an income elasticity with this result. In Figure 3. TableAlIl. and y.Some basic calculations in EViews 45 This means that the vector elements e(I). bave been seasonally adjusted. ~T '.es conssa and ysa in the workfile.229 0. The formula to compute an income elasticity (eLy) with respect to the sample means of consumption and income (see Section 1. This elasticity can be computed in EViews by using the facility that many computational results are available with the prefix @ . ··~! V " .3) The object eLy is included in the workfile. . . The estimation procedure will be discussed in Section 4. .. .1) is shown in an ' Equation' window (using a small data set from Thomas (1997): Appendix ill.' elJ = c(2)"'@mean(ysa)l@mean(conssa) .24: Example I of the computation of an income elasticity .2) where Y and CONS are the sample means.
. Another possibility to compute income elasticities is given by the following example. 'Stats Table' ...6. doubleclicking on the object eLy in the workfile shows the result at the bOllom of the window in Figure 3. . given in Figure 3.25: observation Example 2 of the computation of an income elasticity (e_y) for every In the command window the computation ofthe income elasticity (3.7 E Y Figure 3..6555 and the elasticity varies between 0.0 .. the suggestion was made to publish the mean...9 0. These statistics are immediately found in the summary under 'View' . maximum and minimum values of the elasticity in the sample period.'Descriptive Stats'. For example.. .24: eLy = 0. see the command window.26 where the values are shown.see Figure 3. genr e_y = c (2) • ysa/conssa. Basic Concepts of EViews and Starting the Research Project 1.. after which the series e_y is shown as a graph.6556..8 0.6323 and 0.6924.25. where the income elasticity has been computed for all the quarters in a new series (e_y)..46 Chapter 3.. The mean of the elasticity is 0..... 0. In Section 1.3) is demonstrated.
These points are combined in thi s first case.26 is different than the value of eLy in Figure 3.J = c(2)·ysalconssa Figure 3. hecause the formulae are different: the quotient of means versus the mean of quotients! In this section enough of an introduction to the use of EViews has heen provided to make a viable start with the econometrics course in Part II. will he applied in Case I. More general and additional information can he found in the EViews User's Guide. and making a start with an econometric research project. Case 1: the data analysis In this exercise the accent is • • on learning the use of EViews. This concerns a data analysis that precedes the estimation stage. The EViews introduction.26: Series statistics of e_y Notice that the mean elasticity in Figure 3. as discussed in thi s chapter. .24.Case 1: the data analysis 47 genr 8.
which are candidates for your draft paper. Perform a number of exercises on the basis of the examples in this chapter. Generate transformed variables when you expect that they will probably be used in the project too. make a workfile and import the variables. You can also look at the formulation of Case 2. and import the data in the workfile. Basic Concepts of EViews and Starting the Research Project Learning the use of EVlews The best way of learning to use EViews is to reproduce some of the examples from this chapter and to invent some additional exercises around these examples. Make a start with writing the second section of your paper that wi ll focus on the data analysis. Adjust the appearance of the graphs. create a workfile. Some concrete directions concerning the formulation of a model for the accompanying data have been given in Chapter 2 together with the introduction of the data. Do this for al l the relevant variables. The relevant timeseries data. Download the file 'pcoccoftea. Often it is useful to model the logs of the variables as di scussed in Chapter I. data check whether seasonal patterns exist in the series and decide what to do if they are present. such as deflating variables when you would like to model the economic variables in real terms. where the first 'research project' is elaborated in more detail. . • Choose a 'research project' and select the accompanying data set. In case the data are monthly or quarterly. Make a group of relevant variables. Formulate your research objective(s) with the chosen data set. etc. Examine the variables by making graphs.xls'. Download the data into your computer. Decide which graphs you like to see of these variables: multiple or combined graphs. Start EViews on your PC. have been described in Chapter 2.48 Chapter 3. Check whether the data have been correctly imported and save the workfile. scatter diagrams. by using any word processor. • • • • • • . The start of the research project Part of the first phase in an econometric research is form ulating the economic problem and writing an introduction for the research paper. 3 or 4. Write an introduction with the formulation of the economic problem that will be analysed. in a way that they are suitable to import into your paper. The following assignments can be used as directions for the exercise. The economic data is collected and a data analysis is done. Data set 2. Look at 'Descriptive Statistics' and 'Correlations' and decide which information is useful for the next section in your paper concerning data analysis. Of course it is also possible to use timeseries data that you have collected for your own research.
Part II .
because the properties of this estimator will be known. • . In the Preface. with exogenous and lagged explanatory variables (for timeseries data). The first two stages concern the detenninistic and stochastic assumptions that have to be made for the model. These stages are discussed in Chapter 4 where the model is defined and the ordinary least squares (OLS) estimator with some of its properties is introduced.Introduction to Part II In Part II. In Chapter 7.8 where the scbeme with the stages of an econometric research project was given. This assumption is tested in Sections 6. You will also be able to estimate and test si mple linear equations. The modelling of structural breaks is a subject that will come up at the end of Chapter 7. the last chapter of Part n. this will be demonstrated. It concerns the linear model for one endogenous variable that is explained by exogenous and/or lagged variables (when the sample concerns timeseries data).5 and 7. Remember Section 1. in a correct way. One of the deterministic assumptions is that the structure of the model is constant in the sample period. Part n providesa lot of insight into many aspects of thi s linear model. a number of different specific topics around the linear model are discussed and the consequences of specification errors are considered. General to specific modelling as a procedure to estimate a model in a structural way and all the other discussed items are applied in the cases. We will also look at forecasting problems. it has already been indicated that Parts I and n can be seen as a complete basic course concerning the linear regression model. After these two parts have been studied you will know whether arnot OLS may be applied as an estimator of the parameters.5. Chapters 5 and 6 concern the evaluation phase where statistical tests are introduced to test the stochastic and deterministic assumptions that have been made with respect to the specification of the model. the general linear model is di scussed in detail. Which additional assumptions have to be made so that the estimated mode l can be used for outofsample forecasting? Multicollinearity is the issue of high sample correlation between the explanatory variables that can bring about inaccurate or wrong estimates.
In the reducedfonn equations the endogenous variables are expressed in teons of the exogenous and lagged variables.1 Introduction An economic model is the result of economic theoretical considerations or stems from ideas based on empirical research. In Chapter I. it is a restrictive model as all the explanatory variables have to be exogenous. It is a clear model to start with for didactical reasons.5.Chapter 4 Description of the ReducedForm M o d~ ~:. Such a model is called the classical reg ression model. The following three specifications of the linear model are distinguished and discussed in detail in Parts [J and ill of this text. This model is regularly discussed as one of the first topics in econometric textbooks. this showed relationships between endogenous and exogenous variables. That model can be solved for the endogenous variables when the number of endogenous variables is equal to the number of equations. to observe changes in the properties of an estimator. all models are linear models. In this chapter. Such a model is called a structural model. we considered a structural model that consists of more than one equation. In Section 1.' el " 4. the structural economic model was introduced. It is true that it is good to start with discussing the classical regression model for didactical reasons. for example. as is done in this chapter also. Redllcedjorm models are more general because lagged dependent (and lagged exogenous) explanatory variables are involved allowing for dynamic economic behaviour in the model. . as it describes the structure of the economic phenomenon that is studied. A special reducedform model is the model with only exogenous explanatory variables. The solution for the endogenous variables is the reducedform model.. but it is also useful to discuss deviations from that model immediately where relevant.
3. The first explanatory variable is: Xtl = 1 for all t.1 ) (4. First. The distinction is not only important for theoretical reasons.44. are introduced. we will look at the reducedform model with exogenous and lagged dependent explanatory variables for the case where timeseries data are modelled. the assumptions concerning the specification of the equation. Description of the ReducedForm Model • The classical regression model In the beginning of Part II. for every t . That makes it necessary to introduce other estimation methods like the twostage least squares (2SLS) estimator.52 Chapter 4. in Section 4. The specification of the linear model is: • Y. with k = 1. The OLS estimator will be extensively discussed. . with additional comments are given. First. independently distributed. is the endogenous. but knowledge about this di stinction is clearly important when interpreting empirical results. In general it will be assumed that a constant term is included in the model. In Section 4. Y. Finally. . .8 the principle of maximum likelihood estimation is introduced. This is followed by the introduction of the OLS estimator in Section 4.2) means normally. How properties of the classical regression model and the OLS estimator change by the inclusion of a lagged endogenous explanatory variable will be clarified at various places in this part of the book. . The structural model In Part ill. dependent variable.. X'2 + .. • • • Y.7. Its parameters are estimated by using the ordinary least squares (OLS) estimator. the deterministic assumptions. all the assumptions of the classical regression model. the classical regression model with only exogenous explanatory variables is considered. 4.. which are models that can also have endogenous explanatory variables. The reducedform model Also in Part n. K . • • The di stinction between these three types of models is of importance when choosing an estimator for the parameters and for the statistical properties of the estimator. structural models are considered. = (31 + (J. ~ N I D (0. After that some statistical properties of estimators in general and of the OLS estimator in particular are discussed in Sections 4.2) NlD in (4. All the assumptions concerning this model will be explained below. is explained by K linearly independent exogenous explanatory variables X'k . O'~) . This will be made clear in this chapter. + U" (4.2 The assumptions of the classical regression model In this section. as indicated above.9 the applications with EViews are discussed. + (3K X'K U. This is the constant term. all the assumptions of the classical regression model are given in the next section.
The variance ()~ is constant.s .. (31 + (32 X t2 + . An exact linear relationship exists between the expectation of Yi and the explanatory variables.. + (3K X 'K + Ut.. .) = + . The valiable Ut is assumed to be normally.The assumptions of the classical regression model • • 53 • • The number of observations in the sample period is equal to n : t = 1.) = E (Y. J nj for an accurate estimation result it is desirable that n is clearly larger than K .. as will be explained in Chapter 6.. The moments of the di sturbance term in model (4. independently di stributed . The estimation results generally improve as n . .n a2 if t = s. K) are linearly independent and form the systematic part of the model. E (UtU. .K .S that no auto· correlation is present. . t.. 2 + . as Ut may have no systematic influence.. . The autocovariances are zero. + (3KX tK + E (u. The difference between the number of observations and the number of explanatory valiables n .) = { OU. is the number of degrees of freedom of the model. of course. The disturbance term is not observed. Tbe length of the sample period must be chosen such that the economic structure bas not cbanged in that period.1) have to be constant. Then we also know the expectation of the endogenous variable YL ) conditional on the explanatory valiables.s=l " .). + (3K X tK . (31 + (32 X. . this is relevant for timeseries models only. Next consider the stochastic assumptions. In Chapter 8. this bas consequences for the cboice of n. Assumptions about Ut are made and will be tested after the parameters have been estimated. The structure is constant in the sample period. .. The normality assumption is not necessary to estimate the parameters with the ordinary least squares estimator (see Section 4.. identical • normal distributions. There the general least squares (GLS) estimator instead of the OLS estimator can be used to estimate the parameters.. the linear model with different di sturbanceterm assumptions will be discussed.n. or in other words the disturbance term is homoskedastic or not heteroskedastic. or in other words all the parameters (3k are constant in the sample period. Yt = (31 + (32 X t2 E (Y. variance and autocovariance of Ut are (with E as the expectations operator): E(Ut) = O.' for all t and s. which is the nonMsystematic part of the model. This assumption means that all the n disturbances bave been drawn from independent. • The variable Ut is the disturbance term. Often these tests will be interpreted as modelspecification tests. which meaI].3) but it is required when assumptions or bypotheses are to be tested. t = l . In general. The K explanatory valiables X tK (k = 1.. if t::j:...K increases. The mean.
X t1 . the residuals e. .4) The vector y is an (n x 1) vector wi th observations on the dependent variable. Y2 Y3 y = 1 1 1 X .  ~ ~ ~ (J.fl. bold face and lower case letters for vectors. The disturbances are assumed to be uncorrelated (covariances are zero) and homoskedastic. The parameter vector {3 is a (K x 1) vector with K unknown parameters. The explanatory variables are linearly independent. the vectors and matrices are shown below.3 X 23 X 33 X . The matrix I n is the (n x n ) identity matrix.B + u u ~ N (O .fl KX.{3• • • • 1 U = . Description of the Reduced·Form Model • After the parameters flK have been estimated. can be computed. . This linear model and its assumptions can be written in matrix notation. Y. they aU have the same variance u~. To illustrate the notation in more detail. • The linear model (4. . Notice that aU the assumptions have been specified in the matrix notation as introduced above.3) . = Jot .2) are written in matrix notation as follows. In general. The vector u is an (n x 1) vector with n nonobserved disturbances. which means that the rank r (X ) = K. The residuals are defined as: e. which is a clear and simple notation and more convenient for analytical purposes. X t 2 1 . the vector y is a column vector and y' is a row vector. the following notation will be used: • • bold face and capital letters for matrices.54 Chapter 4.1) and the stocbastic assumptions (4.2 X 22 X 32 X. (4. X '2 .X • • . fl2 U2 U3 . and X is an (n x K) matrix with observations on al l the explanatory variables. • • • • • • fl K The collected data on the economic variables yt .C1~q (4. 1 XtK are in the columns of the vector y and the matrix X. Therefore the (n x n ) variancecovariance matrix of .K X 2K • X 3K Ul fl.K · When the economic model is correctly specified there is no significant autocorrelation or heteroskedasticity found in the residuals by the statistical tests that will be used to test the assumption that Ut is random . ... y = X.
. .. . . . and the predicted Y.E (u )]' = E uu' =E Un Eu. .. If estimates {3" {3" . et = Yt  ¥t .. . as endogenous explanatory variables occur. .E (u)[[(u . . ..K..={3I +{32 X .2+ .als et are computed as: .... . All the assumptions will be tested when estimates of the unknown parameters have been obtained and the residuals have been calcul ated. do not sati sfy the assumptions of the classical regression model. . If (some of) the assumptions are rej ected then thi s rej ection will be interpreted as an indication that the model has been misspecified ..n and the residu. The notation is clarified below... The implication is that the syste matic pan of the mode l has to be respecified in most cases.. Var (u) = E [( u ... t  = 1...3 The ordinary least squares (OlS) estimator The least squares principle yie lds an estimator that minimises the squared differences between the observed Y.. ... 4. . from the estimated model. 13K and <T~ are unknown and have to be estimated from the data..The ordinary least squares (OLS) estimator 55 the disturbances Var(u ) has been written as cr~In in (4 . 13K have been computed then predicted values Y t are computed as:    Y.. +{3KX. EU2U I EU I U 2 EU I U3 EU2U3 2 E u3 U n) EU I U n Eu~ EU3U 2 EU3 U l • • EUn Ul 2 Eu n 2 cru a a a cr 2 a a a cr 2 u u a a a a 2 cru = a~In .... . n . t = l .. These differences are called the residuals. a a a a The K + 1 parameters {31. the macroeconomic consumption equation and the coffeeconsumpti on equation. Notice tb at the examples discussed in Chapter I..4).
. the predicted values Yi.fh. . Define the function of the squared residuals as: n  f ( 131.2 t= 1 and determine the minimum of f with respect to 131. is given in Figure 4. and the residuals e. the regression line. Y. Figure 4.. .fh. The residuals are a function of the parameter estimates. The sum of the squared residuals is used to avoid cancelling of positive and negative residuals. Y.. after which we look at the OLS estimator for the general multivariate model in matrix notation. That hest fit is obtained by using an estimator for the unknown parameters that ensures that we get a model with residuals e.56 Chapter 4. For didactical reasons the OLS estimator will first he derived for the bivariate regression model in a rather extensive way. Y.··· . The bivariate regression model The following bivariate regression model is used to derive the OLS estimator: Y. Y.13K) = ' " e" L.. + u.  For illustration.· ·· .1: lllustration of the bivariate regress ion Yt = {JI + /hX t .1 . Description of the Reduced·Form Model . The problem is solved by minimising the sum of the squares of the residuals as a function of the parameter estimates. a picture with the observations tt . We want to have a model with the hest fit to the observed data.13K. that are as small as possible. X. = 131 + 132 X.
= n73. Y with X and Y as the sample means of X and Y (notation: X = ~   t= l I: n X " etc.y. and fh yields the following first order con (4. f (73.! . and f32 can be calculated from (4.X) 1 (32 ._ _ _ __ ' n 2 n I: (X. Two wellknown expressions for (3.. + 732 L t= l n X.X) (3. = . n (4.) t= l n 2 Differentiating ditions.t= 1 I: (Y.   .8): . (4. .5) (4 . .8) Solving the normal equations for the unknown parameter estimates (3.X . . L t= l X."":.The ordinary least squares (OLS) estimator 57 The function that has to be minimised is: = L (Yi . + 732L t= 1 X~ .Y) (X. .6) These two equati ons are rewritten and called the normal equations: L t= 1 n Y.(32X . and (32 that can be computed from the data. and (32 results in ex pressions for (3.).7) n n L t= l Y.(32) with respect to 73. ..7) and (4. = 73.
7): (4.Y) =L x.nYX = iJ.9) and (4.X.' t=l t= l n n n L Y. .nX' t= l t= l Next substitute the equations (4. + iJ. LX.58 Chapter 4. n +iJ. . 1I) ~ ~ Substitute (:J. = nY X . = (Y iJ.X.Y.10) are obtained by rewriting the summations at the lefthand side of the equations. .9) L (X.10) The two relationships (4. X) (Y.Y) =iJ. t= l t= l n n (4. LX.X) ' =L X. : L Y. which gives: L (X. = Y L t= l t= l X. i1zx L t= l X. . etc.X) (y.' t= l n t= l t= l n n n n n n L y.X) LX. and by using substitutions of the mean in the following way: LX t= l n t = nX . Description of the ReducedForm Model The following two relationships between variables in deviation form (in deviation from their sample means) and their levels will be used in the derivation that is given below.9) and (4.8) and solve the eq uation for (:J. _X)2 t= l t= l n n . .nX'.nXY.x. in (4. LX. . . L (X.x.ni1zx' + i1z LX. t= l t= l n n (4.10). L (X. . Rewrite equation (4.' t= l L y.
.'"'1. Write the predicted values and residuals as (n x 1) vectors y and e : y = xjj. the point determined by the sample means of X and Y.. (4. Then it is not necessary to write all the sums of squares and products of variables for solving the minimisation problem to get the OLS estimator. other expressions can be written for {31 and {32' Values for (4. As before.13) and (4..   The multivariate regression model For the multivariate case it is more convenient to proceed with the model in matrix notation (4. and X..3): y = X(3 + u . Expression (4. see also Section 5. .. the sum of the squared residuals has to be minimised as a function of the parameter estimates defined as: n  f{ f or in matrix notation: ( {31.11):   {31 = Y . ._ n L (X.{3K ) =L.2./h.1 2) can be calculated when economic data for Y.{32X. 14) . and e = y .X) t= 1 n 2 (4.e" t= 1 f (13) = e'e. '" 2 ..The ordinary least squares (OLS) estimator Thus the solution for /h has been obtained: • 59 1h = L (X.13) Of course. Y) .y.    (4. 13) implies that the regression line passes through (X . .X) (Y._ _ _.12) The estimate {31 can be computed by substitution of the estimate {32 in equation (4..Y) ". have been collected.
Because the explanatory variables X'k are linearly independent we know for the rank of X that r(X) = K . (See. .60 Chapter 4. which can be written in matrix notation as:  (4.2{3' X' y  + {3'X'X{3. by deriving the secondorder   conditions: at__ (S) a{3a{3' • = 2XX. 15) Notice the similarity with the normal equations (4.15) by (X'X) .7) and (4. 14). Johnston and DiNardo (1997) p. Therefore the inverse matrix (X'X) .12) for the parameter {3. and the OLS estimator can be written as:  S = (X'X) .2X'y + 2X'XS = o. Multiply equation (4. 464 for some rules concerning matrix differentiation. t= l LX.xS) = y' y . (4. L xl t= l t= l n { h {32   The equations (4. Description of the Reduced·Form Model First. Obtain the firsto rder conditions by differentiating t (S) at (S) = . Solving for (3 immediately gives the normal equations:  X'y = X'X{3. The matrix X'X is a positive definite matrix. This is a quadratic function in with respect to {3. 14) as an explicit function of the estimates: t (S) = e'e = (y  y)' (y .16) Notice the similarity of this expression in matrix notation for the entire parameter vector {3 and expression (4.1 on both sides. it is necessary to check whether this solution is a minimum indeed. and r (X'X) = K. for example. n n L¥i t= l n L ¥iX. . Before going on. rewrite the function (4. so the least squares solution determines the minimum of the function (4.1 X'y. .8) for the bivariate model.y) = (yxS)' (y . 15) consist of K linear equations in K unknown parameters {3k . in the bivariate model.' exists.)  S. t= l n n LX.
For this reason the notation M x for this expression . ~ ~ ~ 4. e = Mxu .X (X'X ).9. The vector {3 contains the point estimates of the parameters {3. The subscript X is used to show clearly that M x is an expression in the matrix X .. So. M x =In . after which the computation of the OLS estimator in EViews will be considered in Section 4. Interval estimates will be determined in Chapter 5.17) (4. The vector {3 is printed as the first column in the EViews regression output. But first a matrix will be introduced for notational convenience that is often used in analytical considerations of relationships among economic variables. which is verified by computing the product.4 later. In Section 5.lX'. Mx Mx = M x . the relationship between the unobserved variable u and the calculated residuals e will be derived. the matrix P x is a symmetric idempotent matrix. M x and the explanatory variables are orthogonal: Mx X = 0 . The 1 matrix expression In . and its properties are: • • • M x is symmetric and idempotent: M x = MSc. the definition of the (n x n) matrix M x is: Mx = In. some properties of the OLS estimator are discussed. Just like M x .X (X'X r X' is frequently encountered in all kinds of analytical derivations. M x is singular.4 Relationship between disturbances and residuals In this section. x in the P x = X (X'X ). In the following sections.P x· Sometimes it is convenient to use P x in analytical derivations. a number of relationships among variables of the model and the OLS estimator will be considered. it has rank equal to n . which can easily be checked. which is important in analysing the stochasti c properties of {3. 18) . These concern relationships that we use at various stages later in the course or which are useful to know for empirical work. Examples that make use of this matrix are also found in this section.K (see the next section). see Figure 4. The matrix M x has no economic relevance. (4.Relationship between disturbances and residuals ~ 61 • Notice that {3 is linear in the endogenous (stochastic) variable y.2.which is standard notation for this matrix in the econometrics literature . The second term in the definition of M is frequently denoted as the matrix P econometrics literature. The following two interesting relationships can simply be derived: e = MxY .l X' .is introduced and the properties of M x are determined.
The relationships (4.X{3 = Y . the residuals e are heteroskedastic and autocorrelated. E (e ) = E (Mxu ) (n x l ) = MxE (u ) = OJ Var (e) = Eee' (n xn) = E M x uu' M 'x = Mx (Euu' ) M 'x = Mx (a~q M'x  2M x· Uu This implies the following sampling distribution for the residual vector e: e ~ N (O.1 7) and (4. Therefore. The residuals wiU be used to test the disturbanceterm assumptions.62 Chapter 4. It is a degenerated distribution as the variancecovariance matrix is singular. The disturbances are not observed themselves. we need to know the distributio n of the residual vector e. 18) means that the residuals are a linear combination of the unknown disturbances. Description of the ReducedForm Model Equation (4.X (X'X ). e= y. Under the assumption that u is normally distributed. e is normally distributed too. if the covariances of the disturbances • are zero and the variances are identical then this is not valid for the residuals! .l X') Y = MxY = Mx (X{3+ u ) = Mx u. that under the assumption that u is homoskedastic and notautocorrelated.l X'y = (In . Notice that this result has been obtained by using all the assumptions of the classical regression model.X (X'X ).18) are obtained as follows. Notice also. In the steps to obtain expressions for the vector E (e) and the matrix Var (e) the assumption that all the explanatory variables are exogenous is used explicitly. because of the linear relation (4. So.y = y . a~Mx) .18). but a linear combination of the disturbances is observed.
Rewriting the result .the variance of the disturbance term CT~. an unbiased estimator of CT~ is derived by using the OLS residuals.483). as X is exogenous = tr (Mx (u~In)) ' the disturbanceterm assumptions are substituted = CT~. which is equal to the sum of the elements of the main diagonal of M . = CT~ (tr (In) . tr (M x) . using its definition ') tr ((X'X)' X 'X)) . [n this section.5 Estimation of the variance of the disturbance term One parameter of the linear model has still to be estimated . The following result will be proved: • is an unbiased estimator of q~ : E (u~) = CT~.: Johnston and DiNardo (1997). p. The notation tr (M ) means the trace of the matrix M . II E (e'e) = E (u'M x u ) = E (tr (u'Mxu )) .tr (X (X 'Xr CT~ (tr (In)  = X ) .K.K). This also proves that r (M x) = n . Next the expectation E (e' e ) is determined by using straightforward matrix algebra.g. as Mx is an idempotent matrix (see e.Estimation of the variance of the disturbance term 63 4.tr (I K )) = CT~ (n . because r (Mx ) = tr (Mx ).1 8) it follows that: L e~ = e'e t=1 n = u'Mxu. cyclical change in the multiplication 1 = CT~ (tr (In) . as u'Mxu is a scalar = E (tr (M x uu')) . cyclical change in the multiplication = tr (Mx (Euu' )) . Proof From (4.
familiarity with basic statistics for econometrics is essential.6 Desirable statistical properties of estimators this section. of regression). X. The introduction will be done for the parameter !i. but it will be clear from the text    what is meant. Description of the ReducedForm Model gIves: Define: .e'e n. is identical to drawing a number from tltis distribution. the usual notation will be introduced. in the simple bivariate linear model ( 1.K 1 1 = nK L e~. An estimator is a stochastic variable. but some statistical concepts will be outlined for a better understanding of the econometric methods and their properties.5 for a simpler proof of the unbiasedness of a~. [n the regression output of EViews.. 4.6): Y. The estimator (32 has a probability distribution which is written as fe . = (3.4 later. because one has to know about the interpretation and quality of the estimation results. see Figure 4. The estimator i1~ is an unbiased estimator of the clisturbance variance u~. (32 for f3" based on a random sample. The computation of a point estimate ".64 Chapter 4. Knowledge of these properties is desirable when doing empirical research.:2 U = .c. The estimator of (32 is written as f3. • + f3. after the distribution of a~ has been determined. we look at some properties of estimators in general.. [n fact. in the following section. the value of au is printed and it is called the standard error of regression (S.. This book does not discuss statistics. Also. + u. which are considered as desirable. The notation for an estimator and an estimate is identical. These concepts [n are important for reviewing estimation results and writing a research paper correctly. See Section 5. .E. we check whether the OLS estimator has these properties. Then. t= l n then it follows that: E (a~) = O'~ .
and becomes degenerate at the point rh .2. the class of linear unbiased estimators). or the mean of the sampling errors should be close to zero. However. are only realised once.) = ~2 A Figure 4. Consistency The estimator fi2 is a consistent estimator of fi2 if the distribution iii collapses at the point fi2 into a straight line when the sample size n goes to infinity.. '" E(Il."" (h.n) = fi2. the data. This me'ans that. it is the estimator with the smallest variance. The usual notation to show that an estimator is consistent is: P lim n . If it is possible to take many draws from this distribution. .) = fJ2·  The sampling distribution Iii of fi2 ma~ look as shown in Figure 4.. becomes narrower   and narrower when n increases. observations on economic variables. then the mean value of these draws should be in tbe neighbourhood of fi2 . the variance of fJ2 goes to zero. in some welldefined class of estimators (e.Desirable statistical properties of estimators 65 . is equal to the unknown parameter 132. like the one in Figure 4. The difference between tb'e estimate (the draw from the sampling distribution) and the mean fi2 is called the sampling bias or sampling error.. The distribution of the estimator. then the estimator is called an unbiased estimator. so only one estimate of fJ2 is obtained. h Efficiency An estimator is an efficient estimator if.2: The sampling distribution of ih : fii" with E(ih) = (h Unbiased ness If an estimator fh has the property that the mean E (f3.2. and in cases where the estimator is biased.) of its distribution function iii. for increasing n. This   is written as: E ({J.g. the bias also goes to zero.
Only minor attention to asymptotic behaviour of estimators is given in this book. P lim stands for probability limit.) Example 4. The series i32. For a detenninistic variable the limit x of the series Xn is defined as: lim Xn x if for every f: > 0.' . Therefore the concept 'converging in probability' is introduced. and the matri x X is not singular: Plim (:x'y) = Plim (X) · Plim (y ) n +oo n jooo n+oo Plim n +oo (X.i32 ) is to be determined.   P lim (xn n tex> n +oo + Yn) = P lim (xn) + P lim (Yn) n +oo n +oo n +oo P lim (x n . 71. If Xn is a sequence of stochastic variables.P lim (Yn) n +oo P lim (xn .7. Not all rules hold for the expectations operator E. HX) op O.n stands for the OLS estimator of fh applied for a sample with n observations. the estimator must  be rescaled and the asymptotic distribution of Jii (lh . • See the difference between the definition of the concept 'limit' and 'probability limit'. Similar rules can be given for random matrices or vectors. etc. thea the Xn converge in probability to a constant x for n there is a number N such thai forevery n > N : Prob (l xn  t 00 if for every e and 0 > 0. P lim (Yn ) n +oo n +oo n +oo P lim (xn / Yn) n +oo = P lim (xn ) / P lim (Yn ) .I) = (Plim n +oo (X)) . Yn) = P lim (xn) . if P lim (Yn) n +oo n + fl. in notation: P lim • >00 Xn = x.' Some properties of the P lim operator are summarised and then used to consider the (in)consistency of the OLS estimator in Section 4. The number x is called the 'probability limit' (the P lim) of Xn . The consistency property implies that the estimator i32 itself does not have an asymptotic distribution .n . If asymptotic behaviour of the estimator is to be studied.O. In case the variable Xi is a stochastic variable. 1 is given to illustrate the above introduced concept. xl < e ) > 1 . only passive knowledge is needed to understand the practical implications for the estimation result of some model specifications. except when the variables are independent.n converges in probability to fh if n goes to infinity.66 Chapter 4. there is a number N such .x l > e for all n. if the multiplication is possible. . Description of the ReducedForm Model The notation fh . This will not be discussed further. (See other econometric or statistical textbooks for a more theoretical discussion on this topic. The following rules are valid. il can be true that for any f: > 0 with positive probability IXn .+00 = that for every n > N : IX n  xl < E.Yn) = P lim (xn ) . as asymptotic theory is a topic that is beyond its scope.
the di stribution of {3 is derived. then it is wellknown that the sample mean: x = L: Xi is distributed as x ~ N (I'. Define the (m x l ) vector z as a linear transformation of y: z = c+Cy. a2 In) with mean I' and variance a2 In.7 The distribution and some properties of the OLS estimator ~ • In this section. Var(z ) = C ... the results of which will be applied in determining the distribution of {3. ~ A linear transformation of a stochastic vector Let Y be an (n x 1) random vector.1 The sample ~ as esdmator ~ the popuIetion ilI"n In general. then z is also normally distributed. In upcoming analytical derivations the following three transformed distributions will be used directly when the occasion should arise. a 2) . c an (m x 1) vector with constants. 1 Xn is a sample from a normally distributed population with mean J. X2.. accord.1') ~ N (0 . then the first two moments of the linear transformed vector z look like: E(z ) = c+C ·E(y ). But first we consider di stributions of linearly transformed stochastic variables (vectors).l and variance a 2 . or more explicitly: . If Xl. The estimator 0 is said A simple example is the sample mean as estimator of the population mean. Then. and C an (m x n) matrix with constants.e) ~ N (0. (On.ing to the above: ! Vn(X . then it is easy to show that Vn to have an asymptotic normal distribution: (e. q). implying: 4. Var(y ) . . C '. if Bn is a consistent estimator of the parameter () I with sampling distribution: N ql n).The distribution and some properties of the OLS estimator 67 Example 4. If y is normally distributed.
(4.20) (4 .20) of z is written in matrix notation: z ~ N (0'. = £1. Then the equations (4. and cov (x. a.. then z is di stributed as follows: z ~ N (c + C/J. ..23) . we consider the following bivariate example. 2 C2) a. Define the vectors c' and JL': and the covariance matrix n: qY y ax ) 2 . (4.2 1) (4.19) For clarification.21) and (4. Next look at the di stribution of the following linear transformation of these variables: From basic statistics it is known that: z '" N (JJ.Lz = • ( Cl C C2 ) (. a~) .). c'nc) . = (CI ( ax a xy 0'2 y a xy ) ( Cl C2 ) 1 c'nco So the distribution (4.22) for /J% and a~ can be written as follows: /J%= J. CnyC').Lz = CIJ.68 Chapter 4.. IL. Let two dependently di stributed stochastic variables x and y be given by : x ~ N (/Jx.22) + C2 J.::) .ix U z =C10"x (4.1. with J.z..y 222 +C2O"y + CI C2 0'xy' 2 22 These expressions can be written in matrix notation. a~) . y) = a xy . Description of the ReducedForm Model if and z = c + Cy. y ~ N (/Jy..
• The distribution and some properties of the OLS estimator . 19). Y2) . or put together as: ( ZI ) = (CI ) + (CI Z2 C2 0 O)(YI ) . defined as: ZI = CI + CIYI Z2 = C2 + C 2Y 2 . Two linear transformations of the same stochastic vector . with (m x n) matrices C I and C 2 with constants. consider two stochastic (n x 1) vectors YI and Y2 that are dependently distributed as follows: where the (n x n) covariance matrices have been defined as follows. we write the distributi on of Z as shown in (4. We consider two vectors Zl and Z2 that are linear functions of the stochastic vectors YI and Y2. fl y. = VaT (Y2) . = Ow (YI. In the second situation. The transformation are written as follows: Then the distribution of (ZI Z is: 2)' A linear transformation of two dependently distributed stochastic vectors Finally. 69 When this result (4. we have the two (m x 1) vectors Zl and Z2 that are linear functions of the same stochastic (n x 1) vector Y with covariance matri x VaT (y ) = fl y: ZI = CI + CIY Z2 = C2 + C 2 Y. = VaT (YI) .23) is generalised for a multivariate distribution with more than two variables. fl y"y. fl y. C2 Y2 .
The wellknown form used to analyse the properties of fJ is the expression of fJ in fJ and its sampling error as obtained in equation (4.' X'y .70 Chapter 4.Notice! ) (X'X ). the OLS estimator fJ is an unbiased estimator of fJ with exogenous explanatory variables.   the expressions for the mean vector and covariance matrix of f3 are easily obtained.24) the moments of the OLS estimator fJ are derived.  Unbiased ness and variance of the OLS estimator By using the notation introduced above.l X' u (4.' = a~ (<.l X' (X{3 + u ) = (X'X ).24) fJ + (X'X) . By using the di stribution (4. With expression (4. we encounter a linear transformation of the vector u with disturbances wben deriving tbe distribution of the OLS estimator fJ. X (X'X ). The vector {3 bas been written as a linear transformation of the vector u . This distribution is necessary for computing the standard errors for the individual parameter estimates.' X' u . The covariance matrix is: • Var (ii) = Var ((X'X) .' .' X 'XfJ = + (X'X ). As will be see n below. a~In .' X ' .      ii = • (X'X ). E u = fJ + (X'X ).l X'u is the sampling error of the estimator fJ.  E (ii) = fJ + (X'X r ' X' .24).' X' · Var (u )· X (X'Xr ' = (X'X r ' X' . Description of the ReducedForm Model Then the distribution of ( Zl Z2 ) I is: [n the next subsection. .19). 0 = fJ · This proves tbat the OLS estimator fJ is an unbiased estimator of the parameter vector fJ in  the classical regression model. the di stribution of fJ will be determined.' X' u) = (X'X ). = (X'X ). The expectation and the covariance matrix of fJ will be computed. Thedistribution is also used for testing bypotheses concerning possible values of the parameters {3k.24) (X' X). The second term in expression (4.
Later on. The estimated covariance matrix of the parameter estimates .25) ~ Once more.I .6. Then the standard errors of {3k (k = 1.25)! The covariance matrix of {3 shows that the estimates of the parameters are usually not independently distributed.I . K) ~ are written as: wbere se stands for 'standard error' . superscripts have been used to denote that they are elements of the inverse matrix. . in Figure 4.. {3 is an unbiased estimator of {3. contains the standard errors se of the estimates {3k . the distribution of {3 cannot be used directl y to compute interval estimates because it contains the unknown variance O'~ . Sometimes it is necessary to use the covariances of the parameter estimates for a statistical test. we then have asymptotic results for the OLS estimator.is: ~ VaT (i3) = u~ (X'X ). in other words.I contain the estimated variances afthe estimates and the elements outside the diagonal are the estimated covariances between the estimates. .26).4. .25) because a~ is the parameter of the distribution and not u~ . but interval estimates cannot be computed by using the distribution (4. we will see that these sampling results for the parameters of the reducedform model only hold for very large samples or. So. and the sampling distribution of {3 ~ ~ IS: (4. interval estimates are derived by using the Student's tdistribution. In Section 5. The diagonal elements of u~ (X'X ) .25) can be written for an individual parameter estimate {3k as: ~ where Xkk is tbe kth diagonal element of (X'X ). The second column of the EV~ews regression output. when we look at the more general reducedform model. and it is conditional that all the disturbanceterm assumptions are valid. However. The standard errors se are the square roots of the elements on the main diagonal of the estimated covariance matrix (4.26) In EViews this matrix is shown in the 'Equation' window under 'View' with the option 'Covariance Matrix'. in the classical regression model. notice that Utis is the sampling distribution of {3 derived for exogenous X.The distribution and some properties of the OLS estimator 71 ~ Notice that the disturbanceterm assumptions have been substituted above to obtain the covariance matrix of the distribution (4.  (4. The distribution (4. (ifk) (ifk) . By using the estimate u~ for u~. feasible standard errors of the parameter estimates are computed in practice.
befote we can look at the 'quality' of the estimated parameters in a correct way.'!:'at the OLS estimator (3 has a smaller variance than any other linear unbiased estimator (3 of (3. That is the reason that we first need an estimated model where the disturbanceterm assumptions are not rejected. The theorem is proved by showing . (3 = Cy. the standard errors are undetestimated in that case and the estimation result looks bener than it is.1 Inaccurately calculated standard arrors We have seen that the results in the tegression output are only valid if the disturbanceterm assumptions are not violated. then EViews (and every other software package) does not calculate the correct covariance matrix! In fact. in matrix terms: the difference between the covariance matrix of {3 and f3 is a positive definite matrix written   as (Var (i3) . This property implies restrictions on the coefficient matrix C : E(3 = E(Cy) = E (C(X(3  + u )) • = E (CX(3 + Cu) = CX(3 + CEu = CX(3. If the disturbanceterm assumptions have heen rejected. The estimator {3 is unbiased: E(3 = (3.l X' of the OLS estimator.Var (3) ). Ploof Define (3 as a different linear unbiased estimator of (3:  C is a (K x n) matrix wi~ constants. SO EViews always computes the matrix (4. 1be absence of autocorrelation and beteroskedasticity has heen used to get expression (4.72 Chapter 4.just as the (J< x n) matrix (X'X). (3 is an unbiased estimator if  . Your computer does not know whether the bypotheses about the disturbance term u have or have not heen tejected. In the econometrics literature. Description of the ReducedForm Model Remark 4. the efficiency theorem in the linear model is called the Gauss. The efficiency property of the OLS estimator The OLS estimator is efficient in the class of linear unbiased estimators.26) for the variancecovariance matrix of the OLS estimator.Markov theorem. which is a historical name. Or.26). With this propeny the OLS estimator is called BLUE (best linear unbiased estimator).
l X ' + D .l + DD' = (X 'X ). Then the OLS estimator is call ed BLUE: best unear unbi ased estimator.I is the positive definite matrix DD'.6)  (3) .(X 'X ). . Hence: VaT' (. By using CX = IK the following property of the matrix D is derived: (X 'X ).6) = a~ CC' = a~ ( X' X ).l X ' or C = (X 'X ).I X 'X + DX = I K I K+ DX = I K DX = O.6) = Var (CX(3 + Cu) = Var(Cu ) = C ·Var (u) ·C'· = C·u~ In·C' 2 = a u CC'. The OLS estimator is an efficient estimator in that class.I + DD') = Var (3) + a~ DD'. The differe nce Var VaT' is equal to the positi ve defi nite matri x a~ DD'. Define this difference as the (K x n) matrix D : D = C . Remember the expression of the covariance matrix of the OLS estimator Var (3) = a~ (X 'X rl . which means that the OLS estimator has the smallest variance in the class of linear unbiased estimators. Look at the difference of (X 'X r 1 X ' and C to link the two covariance matrices. (.The distribution and some properties of the OLS estimator For the covariance matrix of (3 one finds: 73 Var (. Using th is property yields: CC' = ( X ' X ).I + DD' .l X ' + D ) ( X 'X )I X ' + D )' = (X' X r 1 x 'x (X' X ). So the difference between the (K x K ) matrices CC' and (X 'X ).
74 Chapter 4. To be clear. I:xlK A necessary requirement is that all the elements of the matrix converge to finite constants for increasing n. Description of the ReducedForm Model The consistency property of the OLS estimator Asymptotic theory is not a topic in this book. This distinction is elaborated in Examples 4. (See the statistical literature or any of the references to more theoretical econometric textbooks. the P lim of the inverse matrix is the inverse of the P lim: 1 Plim ( _ XiX n +oo n )1 Ox'x. for a more detai led discussion on this subject. The matrix (X'X) /n converges to a matrix Oxx that has elements equal to population moments. or equal to finite constants in the case of detenninistic variables. I: ". The usual notation for this property is: Plim ('XIX) = Oxx . X'3 .24. the matrix is written in more detail (all summations run from 1 till n): n 'X'X n ='n I:X'2 I:X'3 I:X'K I:X'2 I:x12 I:X'3X'2 .) To prove the consistency of the OLS estimator.. X'2 I: ". n +oo n And because of the properties of the probability limit mentioned in Section 4. .. in the case of random variables.6. but with the concept of consistency and the P lim operator as introduced in Section 4. • I: ". . consider the (K x 1) vector (X'u) /n with sample covariances (with respect zero) of the explanatory variables and the disturbance term . The (K x K) matrix with sample moments of the explanatory variables (with respect to zero) is (X'X) In.4. the following additional assumptions are made about the covariances of the explanatory variables. I:X'KX'2 I:X'3 I:X'2 X'3 I:x1 3 • • I:X'K I:X'2 X'K . = (0 Further.6 we are able to look at the large sample properties of estimators.
4) concerning the large sample properties of the OLS estimator will be given. ) ( . The startingpoint is again the expression /3 = {3 + (X'X ). the reducedform model.!.XX n .X'u) = q . Next compute the probability limit P lim {3n . and the structural form model.. x'x . Example 4. The following result for the OLS estimator is obtained. 0 The OLS estimator is consistent. We clearly see that the crucial prereq uisi te for {3 to be consistent is that the explanatory variables and the disturbance term are independently distributed.28) Three examples (Examples 4. where aU the explanatory variables are exogenous.I X'u . then these covariances converge to zero for increasing n : . n . oo  P lim 71+00 /3n = {3 + P lim 71+00 ((X'X ) . we consider the classical regression model.2 Vlll'labi BI In this example.The distribution and some properties of the OLS estimator 75 The K elements of this vector are the sample covariances between the explanatory variables (Xtk) and the di sturbance term (Ut) . 1: • • • the classical regression model.24.1 X' u) = {3 + Plim 71+00 1. 71 +00 n Otherwise some of the elements are finite constants unequal to zero: . P lim (. that are in accordance with the three types of models that were introduced in Section 4.  .X'u n = {3 + f! = {3.I · Plim n+ (1) .27) is valid.27) • Plirn (. so expression (4.!. (4.. 71+ 00 n (4. If the disturbance term and the explanatory variables are independently di stributed.X'u) = O..
I X'u of the OLS estimator for this specific equation . <7~) for all t. n E +E I: Yi. ~ . ~ I: Y' _I U.I n + I: X. property (4. x Plim n ~oo I: x.I Because of contemporaneous nonzero correlations between elements of the matrix (X'X ). I: Y. The expression jj = is written below: /3 + (X'X ).I I: Y.:'I . the probability limit of /3 is determined:  fll Plim n~ oo fl2 fl3   fll n fl2 fl3 + Plim n ~oo 1 n n I: X..I I: Y'~ I Look at the expectation of /3 to see whether the OLS estimator is unbiased or not. The OLS estimator is still consistent.Y. I: Y. I: xl I:X'Y'.I I: Y'.u. . In this case. only asymptotic properties of the estimation result can be obtained. For a first course it is sufficient to establish the conclusion that al l the results are asymptotic now. This implies biased results in small samples.I I:X.I I: Y. To demonstrate these properties of the OLS estimator we look at the following dynamic model Y. Description of the Reduced·Form Model • If lagged dependent variables are included as explanatory variables.Y. + fl3 Ytl + u" u. However.I I: X'Y' _I I: Y. To show the consistency property.I I: X. This has to be mentioned in a res~arch paper too. the OLS estimator is biased but it is still a consistent estimator. ..27) is still valid because Y' _I and u. = fl. ~ N I D (0.~.I with elements of the vector X' u (the matrix/vector elements contain all the observations because of the summations) the expectations operator cannot be moved into the last column. are independently dis· tributed.76 Chapter 4.I I: X.I ~ I:u. when the estimated model is not a classical regression model. + fl2X.
+ 1 u.) are assumed to be endogenous variables and investment (J. Plim i3n = {3 + Plim n~oo n~oo (. It is not difficult to compute the covariance between Y. f32 i33 The variables Y. we consider the model with endogenous explanatory variables. The reducedform equation for income is:  i3. and u..X'u) n n n~oo ={3+0x'x·q # {3.4 An encIogenoua structural roill. the use of another and consistent estimator is necessary.27).!.EYi) (u. is related to Y. nlOdel) variable Finally. . are independently distributed. 1 . = i3. Plim n>oo • i33 i32 i3. Consumption (G.Eu. In this example.The distribution and some properties of the OLS estimator then it follows that: 77 i3. The structural equations are: G. Reducedform equations are the model solutions for the endogenous variables (they are expressed in the exogenous and lagged dependent variables) as introduced in Section 1.) and income (Y.28) is valid instead of property (4. Example 4. .i32 I . The inconsistency of the OLS estimator f32 will be demonstrated in this example. Y.!. is easily observed in the reducedform equation for income. = 1 f32 + 1 . then the OLS estimator is inconsistent as property (4.. [n the consumption equation the income variable (Yi) is an endogenous explanatory variable. the most simple Keynesian model is used again.5. and u.i32 which shows that u. + u. [n this situation. are dependently distributed but Yi. as will be introduced in Chapter 9. + I. and u. so the OLS estimator is still a consistent estimator. ) an exogenous variable. . n f32 i33 +0xx' 1 0 0 0 i3. + f32Yi Yi = G.X'X)" PUm (. The relationship between Yi and u.)). At least one element of the vector q is not equal to zero.: E ((Yi .
oo The inconsistencies Cl and C2 are the result from the matrix multiplication In the Examples 4.EY. consistent and asymptoti call y efficient (in the class of linear consistent estimators). The reducedform model The linear model bas exogenous and lagged dependent explanatory variables: the OLS estimator is biased.: 1 (3. Plim n . . Next we observe that the OLS estimator is inconsistent. A summary of the res ults is given below. • • The classical regression model The linear model has exogenous explanatory variables. consistent and efficient in the class of linear unbiased estimators. The relevance of this knowledge for e mpirical work is evident. .24.78 Chapter 4. The properties of the OLS estimator are interpreted for large samples. and u. properties of the OLS esti mator have been established for the three different types of models..4.. The bias goes to zero for increasing sample size. Description of the Reduced·Form Model First find an expression for the first term: so: Y. which are considered as detenninistic variables in the equation: the OLS estimator is unbiased. = 1 (3 2 Substitute this tern] and evaluate the covariance betwee n 1'..
) =  n 2 u'u ln (2rr) .U /2. for example.2a 2 ' "2 u n The likelihood function has to be max. 'U In (L ) = In ((2rra~rnI2 e.8 Maximum likelihood estimation In following chapters. with u ~ N (O .( y .XP)' (y . In this section. As usual..XP)/2. to become familiar with ML estimation procedures. u . The multivariate density function for the observations of the endogenous variable is: 1 (y " /3' ( 2 ) = (27r(7~)n/2 e.i mised. the maximum likelihood (ML) estimator will occasionally be referred to./J . some specific probability distribution needs to be assumed for the disturbance term.u/ u/2q~ /2 ( 2rra~)n The likelihood function has the same form. ML is a general statistical principle to estimate the parameters of (oon)linear models.. a u = .~) ' J I 1 U This is a function ofy for given values of the parameters {3' and O'~ .2 ln (au) . See any statistical/econometrics textbook. but is interpreted as a functio n of the parameters for given observations: It is more convenient to work with the loglikelihood function. Differentiation with respect to {3 first gives the following firstorder condition: 8ln (£ ) a{3 8u'u a{3 1 2a 2 = O. For estimating the parameters with ML. Cramer ( 1989). y = X{3 + u . the assumption will be made that the disturbances are normally di stributed. Another estimator has to be used to obtain at least consistent estimation results. 4.Maximum likelihood estimation 79 • The structural model The linear model has exogenous. a sbort description concerning the ML principle is given for the general linear model. lagged dependent and endogenous explanatory variables: the OLS estimator is biased and inconsistent. The multivariate normal density in terms of u is: t>' 2) J (u .a~In) . __1::::=e .
we see that maximising the likelihood function with respect to the parameters f3 is identical to mini~sing the sum of squared residuals. For linear models it makes no difference whether the parameters are estimated with OLS or ML. . The value of the loglikelihood function will be used for hypotheses testing with the likelihood ratio test.9 Regression analysiS with EViews To perform a regression anaJysis in EViews one has to specify the equation in an 'Equation Specification' window. The ML estimator of a~. . is different from the OLS estimator a~ of O"~: n 1 "'2 .~ et n t= 1 n uU=n _ KL. = n ~U/U.2 1 '" 2 The ML estimator a~ is a biased estimator of a~. The hypothesis that will be investigated is a simple relationship where coffee prices influence tbe price of cocoa on the world market. u Solving this equation for "e gives the solution: u. The estimation procedure will be explained by using the workfile from Figure 3. but it can be biased in finite samples. ifUt ~ N ID (0.. The model that will be estimated is a reducedform model because lagged dependent variables are included in the equation. defined as a.4 later. We use Hendry's approach of modelling from general to specific.et· t= l .7. see Figure 4.. From statistical theory it is wellknown that the ML estimator is always consistent and efficient within the c1ass of consistent and asymptotically normally distributed estimators.80 Chapter 4. 4. We have established that minimising the sum of squared residuals yields ML estimates. This is a . "e. Description of the ReducedForm Model Tn this condition. The ML principle will be used in this way for nonlinear models that are found later in the text.. To obtain the ML estimator of In (L) is differentiated with respect to which yields the second firstorder condition: "e. which is one of the topics of the next chapter.2 au = .2 u n ~2 2uu + u'u 2U 4 = O. 8ln (L) a. This implies that the maximum likelihood estimator f3 of the parameter vector f3 is identical to the OLS estimator fj of f3.~). EViews regression output always gives the value of the maximum of the loglikelihood function .
lpcojfee and lptea. Next. This has been done in the window from Figure 4. A general specification for this relationship is: COC In (Pt ) ={3.3. The commodity prices are transformed into their logs and the new variables get the names lpcocoa. the regression equation is specified.l to ./ ) lpcocoa( . • Click the ' Quick button' in the EViews window and select 'Estimate Equation '. for example. The equation can be specified in the 'Equation Specification window' as follows: lpcocoa c lpcojfee lpcoffee( .7). It has been obtained in the following way./) lpcocoa( .co!) + L j= 1 L {hj In (Pt } ) co + L {33h In (p.3). its specification has been obtained in a couple of steps by reducing the general model (4.29). (In Chapter 7. This specification can also be written in a more convenient way: lpcocoa c Ipcoffee(O to . in this example the workfile 'cofcoctea' is opened (Figure 3.Regression analysis with EViews 81 recommended procedure to determine structural or reducedform models for timeseries data.) The relationship is analysed for the variables transformed into their logarithms.'e~) + Ut h= O L (4. The price of tea has no influence CO on the cocoa price and the inclusion of pt ! and I1o{ is sufficient to show the influence of the coffee price on the cocoa price. Notice that this equation is not a classical regression model but a reducedform model because of the lags of the cocoa price. Next click with the righthand mouse . just a few lags for the variables should be sufficient.2) Ipcocoo(3). On It is a specific model . The specification of the equation that will be estimated in this example is: C In (Pt OC ) = {31 + {32 1n (ptC f) + {331n (ptco f) O + {3. The sample concerns monthly data and because prices react quickly to each other./) lpcocoa( . Next the 'Equation Specification' window opens. this procedure will be considered in more detail. Estimation of this last regression model will be demonstrated here. + L i= O L {3li In (p. In (ptC + {35 1n (p. Another way to open an equation window is to select the variables involved. Open your workfile. three lags (L = 3) for each explanatory variable."~ ) + (36 1n (PtCO~) + Ut. The following steps are examples of what can be done: • • • Start EViews.29) with. The estimation result (not reproduced here) shows that this number of lags is sufficient to have nonautocorrelated residuals.
11Dni 0.D.102578 1. Error IStatistic Prob .040055 0.0131 3.602815 2. 0.044336 0.016100 0. of regression Sum squared resid 0.489336 0.020112 0.044301 0.772 O.9511255 2.041270 0.461665 0.4: Regression output .710014 2.3: Specification of a regression equation Dependent Variable: LPCOCOA Method: Least Squares Date: 1unQ2 Time: 15:25 Sample(adjuS1ed): 1960:04 "200=2:09 Included observations: 510 after adjusting endpoints Variable Coefficient S1d.485524 29.00::0 C LPCOFFEE LPCOFFEE(l) LPCOCOA(1 ) LPCOCOA(2) LPCOCOA(3) Rsquared Adjusted Rsquared S.660197 9506.97815 6. dependent var Akaike info criterion Schwarz criterion Fstatistic Prob(Fstatistic) Log likelihood DurbinWatson stat Figure 4.120061 0.800936 2.320071 0.0032 0.4235 0.0133 O. Description of the ReducedForm Model Figure 4.E.996166 0.070928 0.508912 2.OODJJ Mean dependent var S.OCOO 0.82 Chapter 4.
Error. The results can be copied into the text file via the clipboard as well. The standard errors of the estimates se as determined in Section 4. If it is relevant to print the results. are given in the second column: SId. This implies that the use of statistical tables will be hardly necessary. Looking at the statistics given at the bottom of the window you may not recognise them all at this stage of the course. The estimation result is presented in the 'Equation' window. see the other options in Figure 3. The indicated default method is OLS. With this option.4. other estimators will be encountered in Part ill. Many results in the output have stiU to be explained.16.3 is shown in Figure 4. An example has been given in Figure 4. When bypotheses are tested. EViews computes the value of the test statistic and generall y its 'probability' or 'pvalue' too.4.4. The maximum lag is 3 in this regression. When the tests for the disturbanceterm assumptions and restrictions on parameters have been introduced in the following chapters.  (13k) . click on 'Print' in the window in question. this example will be elaborated further to explain the other computed statistics in Figure 4. A number of those probabilities can be found in the output in Figure 4. the lags have to be filled in subsequently. but they will be explained in following chapters.7. The other columns will be explained later on. In the heading of the 'Equation' window you can observe that the sample has been adjusted and starts in 1960:04. Choose an estimation method and click OK. The output belonging to the specified relationship in Figure 4. It is also possible to change the estimation period in this window. Let us take a look at Figure 4. We have lost three observations by making the model dynamic with the specification of lags. Specify the regression equation in that window. .3. The (6 x 1) vector {3 with the point estimates is found under Coefficient. This change does not influence the sample in the workfile. only a static equation is specified.Regression analysis with EViews 83 • • button on the selection and click 'Open as Equation'.4. EViews automatically adjusts the sample when timelags are specified in the equation.
.
the gene ral linear model. could follow. Student 's ttests or Fisher F tests. 5. However. for example. and the OLS estimator for its parameters have been discussed. in this chapter. we know that only in thai case are the sta ndard errors ofthe OLS estimator correctly calculated. This means that standard errors can be computed for the OLS estimates. the logical follow up to Chapter 4 is a chapter where a number of tests are introduced for testing the disturbanceterm assumptions.3: Introduction In the previous chapter. From the previous chapter. we can correctl y perform statistical tests for hypotheses about the parameters. F . Based on these assumptions the distributi on of the OLS estimator has been derived .and LMtests are discussed. It is cl ear that when the stochastic assumptions have not been rejected. a chapter where tests can be djscussed. its stochastic assumptions. This is why we wiU first di scuss the tests for deterministic assumpti ons. it is possible to discuss the tests for the stochastic assumptions in a clearer way in Chapter 6. These tests are more convenientl y introduced in a chapter with statistical tests for the deterministic assumptions and can sometimes be used in a similar way when testing disturbanceterm assumptions. the arguments were given that it is necessary to get a model where the disturbanceterm assumptions are not rejected. In Chapter I. With the knowledge about the tests obtained in this chapter. but only under the strict assumption that the disturbanceterm assumptions are true. The t. . The refore. whi ch are needed for testing the specification of the systematic part of the model. knowing these fac ts. the tests will be discussed the other way around ! The reason is that some of the tests for testing the disturbanceterm assumptions are.Chapter 5 Testing the Deterministic Assumptions . After these 'tools' are known .
2 relationships among the variables of the linear model are discussed that are relevant to the topics that come afterwards.x'xfj = 0 (5. In this section. For example. The remaining sections of this chapter concern the abovementioned tests. The orthogonality property is also true for the predicted values of the dependent variable. it is mentioned once again that in thi s chapter it is assumed that the assumption holds! 10. Because R 2 is also used as a quality measure for the systematic part of the model it will he introduced in this chapter.1) and all these products are zero. All the explanatory variables are orthogonal to the OLS residuals. so all the explanatory variables are multiplied with e in equation (5. some important consequences are shown. This is why the formulation of Case 2 is found at the end of Chapter 6. .7.15) from Section 4.3. in Section 5.:.xfj) = 0 X 'e = O. we will consider a wellknown statistic regarding the quality of the linear fit to the data: the coefficient of determination R2. which will be explained in that section. In this section we look at these relationships. Therefore. in Section 5. Let there he no misunderstanding that in practice one has to test the disturbanceterm assumptions firs!. First. this orthogonality property will be used in the derivation of the Fstatistic in Section 5. 2 Relationships among variables of the linear m odel Some relationships among the variables of the linear model are relevant for the derivation of test statistics or have interesting computational consequences. we show that the residual s e are orthogonal to the explanatory variables that are in the columns of X.1 ) X' (y .86 Chapter 5. Some arithmetical problems.4 in the following way. In the matrix X ' the variables are in the rows. This statistic can only he interpreted correctly if the disturbanceterm assumptions are met. The presence or absence of a constant tenn in the model has arithmetical consequences for some relationships or statistics. This will he done in Case 2 when the first research paper will be written. Rewrite the normal equations (4. First. exist around this measure.::: 5. as this variable is a linear combination of the explanatory variables: y'e = = f3'X'e  fj' 0 .0. Testing the Deterministic Assumptions Besides a discussion about various types of tests concerning possible values of the parameters. X'y = X'Xf3 X 'y .
... the multiplication of the first row with the residual vector shows: . because e = 0: .K X 2K . First.. + f3 KX'K + e" . use this property to show that Y = Y when an intercept is included in the equation:  The means of the observed and predicted values of the dependent variable are identical.... + (3. = f3......... I n t =1 + f32 X'2 + f33 X'3 + . = f3. A model in deviation fann is a model for the variables in deviation from .. Then... Such a specification has an implicit constant term. The sum and the mean of the residuals are equal to zero if a constant term has been included in the model.. + f3Kn . X nK L e...X'2 t= l n 0 0 0 X . e2 en t= l n L e.X'K t=l Define a vector £ with all elements equal to I: £'= (1 1 • 1 )' ... .. this is not the case without a constant term in the model.. They concern arithmetic problems with practical implications when estimating a homogeneous relationship. With an intercept in the equation it is simple to demonstrate that the sample means of the variables are on the regression plane. Y.... 1 ..n ~ n L t= l 7l X' 2 + f33 n .....' " Y... Some of the effects around the specification or nonspecification of a constant term will be di scussed here. The orthogonality property can be schematically shown in the following way: n L et 1 X . I L t= 1 n X '3 + .... ...2 1 X 22 1 X n2 • • e..Relationships among variables of the linear model 87 If the model contains a constant term then the elements in the first row of X ' are all equal to I....'e = 0 ==> L t= l n '" = 0 ==> e = O.. In general..... I L t= l n X'K +n I L t= l n e" This property can be used in a model in 'deviation form' ..
t= l Y yt +e. In deviation form.3).Y) = L (y. Testing the Deterministic Assumptions their sample means. . for example. The estimate f31 of the constant term can always be computed from : 131 = Y . . The following expression is always true becau se of the orthogonality property ( f: Y. .= ~ n n n " ~ t= l n yt2 = Y. .Y) +L e. + 2 L (y.2 + . 2 ~ ~ ~ t= l n t= l n ~. is true with a constant term in the model: L (y.2) in 'deviation form ' .X) + (U.V) = (y.U).+e. the bivariate model : yt = f31+ f3 2X .732 X . + 2 "~ Y t= l n " ~ t= l [t = " 1'. ~ 2 t= l (5. .=Y (y.Y) + L e._V)2=L (Y . . (X. n n 2 n n 2 t t= l (yt  y)2 = .2 " 1'. .V)' t e.e. See. In fact. in deviation form .2 ~ ~ t= l + " e. .e. e.V) = {3. + L= l t= 1 t= l t= l t= l The last relationship has been obtained by substituting the following quantities " L y.e. .Y) e..Y) + e. The following relationship (5. ' t= l t= l t= l n n 2 n (5.= 1 t (y. because a model in deviation fonn has one normal equation less to solve.. + U. .88 Chapter 5.2) can happen during the derivation of a statistic that we can use an expression like (5 . to compute a regression in deviation form has historical interest only. L (yt . the model is: (yt . an advantage in the past. = 0).3) This can be established in a similar way. Y . . = 0 t= l . .
the (total) variance of Y. When R 2 is close to I it can be an indication that the estimated linear model gives a good description of the economy. . The coeffi cient of determination R 2 is a measure of the goodnessoffit of the estimated model to the data. This coefficient can be considered as an indicator concerning the quality of the estimated linear model. = Y L e.Y) 2 t= l . = O. ofte n just called Rsquared. 5. is frequently denoted as: TSS = ESS + R SS. which is only valid with a constant term in the model. Multiplied by l i n it is an expression of the variances of the vari ables. Y. R 2 can be used as such if the disturbanceterm assumpti ons have not been rej ected.'  n ( _ TSS t ="'.. The total sum of squares (TSS) equals the explained sum of squares (ESS) plus the residual sum of squares (RSS). t= l n n t= l Relationship (5.3 The coefficient of determination The considerati ons of the previ ous section influence the computation of the foll owing statistic.. and when R2 is cl ose to zero it can be an indication that this is not the case.3). R2 can be expressed in three ways. is split into the 'explained variance' and the 'unexplained variance' . .  )2 Y (y.The coefficient of determination and 89 L Y e.V { ESS 2 _ '0'"2 _  I: . therefore a subscript is added: n R~ = 1 _ RSS = 1 _ TSS t t= l '=1 ' I: e2 (Y. the coefficient of determination..
= Y" . If a constant tenn is in the equation. have been generated.. X. In EViews R~ is computed. However.m'n (Yi . + f3. R2 has no meaning and should not be included in a paper reporting the econometric results. = Y" . The definition of R~ is frequently used to compute R2 . then the three expressions are identical to each other: demonstrated by using the expression TSS = ESS + RSS. because R2 is always equal to the value r} x' the correlation coefficient between Yi and X. . . with only 14 observations. It is easy to see that the statistic R2 is not useful in the bivariate model. and ~ can become larger than I . .» 20 1 3 . which is sufficient for this purpose. in the bivariate model R2 is independent of the parameter estimates and cannot give information about the 'quality' of the model.mm (Yi.». The statistic is not useful in a model without a constant term. 1 Total variance of Y. Next three other 'dependent' variables have been computed by linear transformations of Yi to make it all much worse: R5 Y'2 = Ytl  Yi3 . In a model estimated without an intercept it is always possible to compute yourself (K > 2) and to report its value if you wish to publish an R2. is a linear transformation of X..2 (Y" . Testing the Deterministic Assumptions The following word expressions for R2 clarify the definitions: R2 = 1 _ Unexplained variance of Y.) The behaviour of the R'f can be demonstrated with the following simulated example. So. ~ = Explained variance of Yi ) Total variance of Yt R~ = ·The sq uared linear correlation coefficient between Yi and Y. and Y" which gives information about the fit of the model to the data. R~ can become negat. Then it can be shown that: • • • the three definitions of R 2 are not equal to each other. Yi. So. Then R 2 has also the property: 0 < R 2 < 1.90 Chapter 5. in a model without a constant term or with one explanatory variable.ve. as is the case in EViews. R~ is always between 0 and I because it is the squared correlation coefficient of Y. so  does not report this statistic in the above described situations. (Then mention in your paper the definition of R 2 you have used. This is conveniently demonstrated by evaluating the third expression R5: because Yi = {3. Two series Y" and X.5 (Yil . .
40 30 40 Y/3 •• .1 . the regression results are given for the four relationships estimated with and without a constant term. I.. .:• Y" 30 20 I• I.1) will show that it is very odd to estimate relationships between these dependent variables Yti and X t without using a constant term. • . X..Y) t= l n / (n.!< ) 2 L: (y. 30 •+~~~" 30 X. This is a bad property for an indicator about the quality of the estimation result of a model and is the reason that the adjusted determination coefficient (R2) is introduced.1) ' or R2 = 1 .K) ] .7 later. ~r' X.:. +. The results speak for themselves. 30 • I.~ • I.~~" 20 ~ • • I.1: Scatter diagrams of the simulated data Scatter diagrams (see Figure 5.The coefficient of determination ~ ... ~ . Figure 5. +~. . ~ .1) / (n . The nominator and denominator of R~ are corrected for degrees of freedom : R2 = 1 __.1~__________ L: n eU(n . . One more annoying aspect of the determination coefficient is its property to increase when arbitrary variables are added to the model. In Table 5. +~~~~ 30 I. • I. This is caused by the fact that L: t= 1 n e~ decreases when variables are added.R2) (n. . 91 4 .[ (1 . .. 30 • 4. Everything that has been discussed in the text can be found in Table 5. see Remark 5. 1. with standard errors in parentheses. 30 X..'..
1. In the regression showed in Figure 4. 13 ) X.84 1.84 1.and Fisher Fdistributions for testing restrictions on parameters. Such proofs concern a number of X2distributions that are needed to obtain the Student t.6.1: Ex:amples of the varying determination coefficient for model s with and without an intercept R2 increases when variables are added whose parameter estimates have a tvalue that is larger than one.05) Table 5.05 ) X .4 Survey of some useful distributions and test principles We wiU consider some necessary statistical topics as preparatory work before interval estimates can be computed and restrictions can be tested.84 0. 1).57 X . If x ~ N(O .65 0. • The definition of the Student tdistribution with r degrees of freedom is as follows.43 X. = .07) X .52) + (0.04) = 17. .03) (0.21 ) = 1.08 + (0 .03) I 2 3 4 5 6 7 8 Y" Y'2 Y" Yi3 Yi. because of the large sample.84 0. In small samples. see agai n Figure 4.Jv2 /r ~ t(.84 0.08 YB = 1.74 (3 .10 ( 1.84 0. the definitions of the Student t. v 2 ~ x2 (r ).·)..11. 13 ) X.84 . Y" R' 1 0.92 Chapter 5.jT ~ t(r).84 0. First.32 0.84 2.42 0.17 X. 1. = 15.53 R' 3 0.03) (0.43 = 1.63 ( 1. degrees of freedom : . 0.73 = 0 . The tstatistic will be derived in Section 5. (0.84 3. and x and v are independently di stributed. 5. then the following quotient has a Student's tdistribution with . Both statistics R 2 and R2 are printed in the EViews output.4 the difference between R2 and R2 is rather small.54 + (0.84 0. 0.1. A more detailed discussion and proofs regarding the issues in this section can be found in the statistical literature or in a statistical chapter in a more theoretical econometrics book.84 0.0. larger deviati ons will be observed. or X (5. 0.4) (x/v) .84 0.4 earlier. (0 .84 YB = 8.84 0.75 R' .53 0.84 . because these definitions will be applied to derive the test statistics. Testing the Deterministic Assumptions Regression (3.21 X .04) 0.23 0.and Fisher F distribution must be memorised.26 + (0 .
consider an (n x 1) vector X . in matrix notation : 2 XX = 0" 1 I X I (a 2In). • Thirdly. for deriving t. it is clear that we will need X2 di stributions. for example. 1= 1 • Secondly. which is normall y distributed with zero mean and an (n x n) covariance matrix 11: x (nx 1) ~ N (0.) • First. by definition. . In). Johnston and DiNardo ( 1997: pp. the sum of squares n L i= l n x.and Fdistributions. and the sum of squares L i= 1 n (Xii 0")2 has again a X2 (n) distribution: (5. in addition to the normally distributed OLS estimator. ~ x2(n). If X2 ~ X2(p) . Use the notation (4. is given below. y2 ~ X2( q).4) from Section 4. and X2 and y2 are independently di stributed.2: x (n x I ) ~ N (O.5) From these definitions. (5. a helpful summary of four useful X2 distributions.' x . consider a variable Xi with n independent observations from a standard normal di stribution.Survey of some useful distributions and test principles •  93 The definition of the Fisher Fdistribution with p and q degrees of freedom is as follows.7) Now write (x'x ) / a 2 in a differem way. has a x2(n) distri bution: (5. (5. the n the following quotient has a Fisher F distribution with p and q degrees of freedom : . containing n observations on the variable Xi. 11). Therefore.493495). which will be used to derive the test statistics.8) and see the benefit of this notation in the following situation. consider a variable Xi with n observations that is normally independently di stributed with mean 0 and variance 0"2 Then the (n x l ) vector x is di stributed as: Standardise the Xi. (See. Then.6) XiX = L x.
94 Chapter 5. • Finally. For example. In EViews an LRtest . We have a variable Xi with n observations that is normally independently distributed with mean 0 and variance .). If thi s null hypothesis has not been rejected the model can be reestimated without X tk . . The maximum of the loglikelihood function of the unrestricted model is L(O) .10) will be used in the following sections of this chapter. These types will be encountered in following chapters. The three test principles differ in which model is used to test the restriction. (See. Then the fol Further we have an (n x n) idempotent matrix A with rank r lowing X2di stribution can be derived: (5.. < n. which means that the restricted model is a special case of the unrestricted model. a frequently used test concerns the null hypothesis that an explanatory variable (X tk) has no influence on the dependent variable (Y. the restricted model has to be ' nested ' in the unrestricted model. we consider the following situation. is defined as L(OR) ! L(O). Section 4..2: (71 )( 1) x ~ N (0.9) and (5.8). for a compact discussion on this topic. • Likelihood ratio (LR)tests The LRtest can be app~ed after both the unrestricted and restricted models have been estimated.6 in Stewart and Gill (1998). ' Waldtests' and 'Lagrange multiplier tests'. 10) The two X2 distributions (5. which is formulated as Ho : 13k = O. and L(OR ) is the maximum of the log~kelihood function of the restricted model. The likelihood ratio>.2In) . and the test statistic and its distribution (for large samples) under the null hypothesis are:  The LRtest statistic can easily be computed because the maximum value of the likelihood function is always printed in the regression output. for example. Testing the Deterministic Assumptions It can be shown that: (5. A summary of the test principle is given below.9) See the analogy with the notation in (5. suppose that 9 linear restrictions have been imposed.) Forexarnple. Test principles Three types of test principles are disti nguished in the statistical literature: ' Likelihood ratio tests'. Then we have an estimated unrestricted model and an estimated restricted model. in other words Xtk does not belong to the model.
the Waldtest and not by the LMand/or the LR. the X2 distribution of the squared residuals e 'e (or a~ ) is established. Substitution of this expression in (5.~ In) . Greene (2000).~ in the various formulae. 12) . an LR. the estimator of the variance of the di sturbance term . the F .~ has the following X2distribution : (5. a degenerated normal distribution was derived for the residual vector e.. 11 ) • ]t is known that e = Mxu and that e'e = u'Mxu . but this test can only be used to test the hypothesis that variables can be excluded from the model. 10) we know that (u'Mxu) /.K . In this section. The X2distribution (5 . 11 ) gives: (5.) 5.10) from the previous section is appUed to derive the distribution of u~ . The matrix Mx is idempotent with rank n .and a Waldtype test is rejected by.. • The di sturbance term U t is a normally independently distributed variable with vari ance u~: u ~ N (0 . are applied after the parameters of an unrestricted model have been estimated. which will turn out to be a very useful di stribution. only the restricted model has to be estimated to test restrictions on parameters..4. • Lagrange multiplier (LM)tests For the application of an LMtest. in the following way.Distribution of the residual variance estimator 95 • is also found.. only the unrestricted model has to be estimated. An underlying reason for different test results is that in small samples differences are caused by using either the unbiased estimator (j~ or the biased but consistent ML estimator q~ for the disturbance variance . (See.test.~. This is often represented by the inequality: LM < LR < W. for example.and ttest that are discussed in this chapter. for example. Waldtests For the application of a Waldtype test for testing restrictions on parameters. In the linear model it is possible that a null hypothesis that has been tested by an LM. For example. or Stewart and Gill (1998) for more information on this property. Maddala (2001)..5 Distribution of the residual variance estimator In Section 4. Then according to (5.
5.7 is used by writing /3 and e as linear functions in the same random vector with disturbances u . 13) The X2di stribution (5. such as the normal distribution of the OLS estimator.13) will be used to eliminate the unknown parameter a~ from normal distributions that have as a parameter. The parameter can be eliminated by using the X2di stribution of a~ (5. .K). Also. With the following steps the tdistribution of the 13k is derived.96 Chapter 5. The notation as introduced in Section 4.or Fdistributions with a~ in their expressions instead of the parameter (T~ . These distributions are used in practice to test hypotheses. this distribution can not be used directly to compute interval estimates for the unknown parameters.13) implies that a~ is an unbiased estimator of a~ . ~ ~ ~ 0'. observe that (5. because u~ = e/ef (n . fj = /3 + (X'Xr' X'u e = M x u. This results in a Student's tdistribution for the standardised estimator fA of the parameter fh.5. Making use of the property that the expectation ofa x 2 distributed variable is equal to its number of degrees of freedom: a! or E (a~) = a~. 13). Testing the Deterministic Assumptions However. This is a more convenient way to prove its unbiasedness than the method given in Section 4. It will be shown that the elimination of a~ results in t. au ~2 (5.K ) the statistic is written as: 1 (n_ K ) a~ ~x2(n .6 Interval estimates and the ttest to test restrictions on parameters The normal di stribution of /3: ~ contains the unknown parameter a~: therefore. A first requirement is that the OLS estimator and the residuals are independently di stributed.
./x . x" is the k th diagonal ~ • The distribution (5.e .Interval estimates and the ttest to test restrictions on parameters Next.(3. Standardising 13. as done in (5. to yield the Student's tdistribution for the OLS estimator (3.f(). compute the covariance matrix Cov 97 (13. The di stribution of an individual parameter estimator (3.f( ) . a tdistribution is obtained according to its definition by dividing the standardised normally di stributed OLS estimator (for one parameter) by the square root of the X2 di stributed o~ (5. The expression oJ!.'. 13) of o~ is: (3. or simplified: ~ (3k ./x" ~ t (n .l X'· Var (u ) · = a~ (X'X r ' X ' M o. 13) to eliminate the unknown a~.(Notice!) 3) m _ S th Cov (ffi. tdistribution of (3.l X '· a~ In' M x (X'X ).e) = = M x = (X'X ). . (n. x .f( ) a~ ~x'(n. au ... 14) .e) .e )f The covariance matrix equal s a zero matrix as ~ Mx X = O. has been introduced in Section 4. the (5.f()~: ) ~ t (n.(3.f(). ./x" in the denominator is the standard error. ou.. it is established that {3 and e are independently distributed. So..: ~ ( J(n. 1) . is written as: se (13k ) .4).7 as: ~ where the earlier introduced notation has been used again: that gives: element of (X'X). ~ N (0. Notice that the assumpljons of nonautocorrelated and homoskedastic disturbances have been substituted to obtain this result. au ~2 Divide these two expressions.. of 13k' Next. On the basis of this result.
98 Chapter 5. ) and compare tbis value with the value of t (n ) K ) in the table of the tdistribution for any significance level.or tworailed leSI. It is important to specify the alternative hypothesis HI accurately.. . Tbere are three possibilities: In economic models. the researcher generally knows the signs of parameters of contemporaneous specified economic variables in timeseries models. The most common value for c is c = 0. the null hypothesis Ho : (3k = C. is absent. When <> is the chosen significance level.se (S.c /se (S.)· The tstatistics follow tbe Student's t distribution under the null hypothesis. can be tested: for example. ~ . 14) is the wellknown tdistribution for individual parameter estimators. and the alternative bypothesis is that the influence of X t k exists and can be either positive or negative._ (3. The test can be used as a one. a (100 . The null hypothesis is that the influence of X. The ttest is a Waldtype test because the restriction is tested by using the unrestrictedly estimated model. 14): Compute (S. or compute its matching pvalue. . on Y. with c being any constant. Tbe null hypothesis is that the variable X' k bas no influence on the dependent variable Yr . wbich depends on the formulation of the alternative hypothesis. All econometric software produces at least three columns in the output. a column with the estimates. the standard errors and the tstatistics.<» % confidence interval for (3. Hypotheses with respect to a particular value of (3.. The third column with tstatistics is the quotient of the first two columns to test the null hypothesis X .K ) can be found in a table of the t distribution. Two things can be done with this distribution. Ho : (3k = with ~ a t. The signs of the parameters for most of the variables in a model for crosssection data will be known too. is written as: The value t .../ 2 (n . The tdistribution is used under the null hypothesis by substituting the null hypothesis in the distribution (5. Testing the Deterministic Assumptions The result (5. Then a onetailed ttest is . to test for the significancy of the influence of on Y. Confidence intervals for the unknown parameters can be calculated and hypotheses about values of these parameters can be tested.
2 Underlying assumptions Be aware that all the assumptions of the classical regression model have been used! So. atstatistic tb1: (.and twotailed tests EViews gives a fourth column with pvalues for the tstatistics for the null hypothesis flo : 13k = 0 and the alternative HI : 13k '" O. orin EViews c(2) = 1. For the same reason this is true for the tstatistic that bas an asymptotic tdistribution_ With lagged dependent variables the standard errors and tvalues have to be interpreted asymptotically. Such a tstatistic can be computed directly in EViews by using the procedure with ·scalar' . For higber probabilities the difference can be more striking. the absence of autocorrelation and heteroskedasticity. That is 0. the probability of its tstatistic is 0. before interpreting tstatistics in the regression output! Remark 5. But as a positive inftuence from the coffee price is expected.Forexample. The sign of the parameter of pt'0{ is not known a priori. look at the estimation result in Figure 4. So it is not necessary to use the table of the tdistribution. This hypothesis can simply be tested by the researcher by computing the tstatistic se . Then it might also be appropriate to test the null hypothesis Ho : 13k = 1. Remark 5. for example.3 Probabilities.32%.16% in the rightband tail of the distribution. and decisive for rejecting the null hypothesis or not. These pvalues correspond to a twotailed test. . Then compute. Examples are given below. the pvalues have to be halved. once again: in practice.7. when a lagged dependent variable is included in the model. The value of the kth parameter estimate is also returned wi. on&.4 where the tstatistic of the estimated coffee price parameter has a probability of 0. Compare this value with an appropriate value in the table of the Student tdistribution.16% in the lefthand tail and 0.8k) scalar tbl = (@coefs(2) . In your research papers you have to mention asymptotic standard errors and asymptotic tvalues. Remark 5. the parameters represent constant elasticities.4 Lagged dependent variables As discussed in Section 4.1)/ @stderrs(2). Then a twotailed test is needed.th @coefs(k)andthevalueofitsstandarderrorwith @stderrs(k). Remark 5.take the null hypothesis Ho : (3t = 1.1) / (. If lagged variables occur in timeseries models then their signs are often less clear. first test the normality. The difference is small in this example. for large samples.33%. For example. and comparing the value with the appropriate tvalue in the table of the tdistribution. the estimation results can be judged asymptotically.Interval estimates and the ttest to test restrictions on parameters 99 the obvious test.8k .16%. so the ttest is a twosided test with a probability of 1. When in fact the test is onesided.1 Testing the elasticity When the variables have been transformed in logs.
02 .68 Example of lhe one· tailed . se({32).K = 40: t . including the constant term). 1 1. So Ho (no influence of X '2) is rejected in favour of H. se(f32 ). Testing the Deterministic Assumptions Example 5. so a onetailed test will be performed: Ho : {32 = 0.   Student f·distribution 5% o Figure 5. The null hypothesis will be tested.K) .02.68. (a positive influence of X'2) at the 5% significance level if: se ({32 _ ) > 1.025 (40) = 2.1 A onetalled Hast and a confidence Interval Suppose we have a model with four explanatory variables (K = 4. the interval in thi s example is:   f32 ± 2. is valid. that the variable X'2 has no influence in the explanation of the dependent variable Y. Under the null hypothesis the following di stribution of tff. The parameters have been estimated from a sample with 44 observations.68. with or = 5% and n .{37:2~ ff' = se(ih ) IjJ t (40). .Q25 (n .05(40) = 1. HI : {32 > O. A possible positive influence of X 2 is expected from the economic theory. In the table of the t distribution we find that to. As to. U you want to report a 95% confidence interval for {32 then you have to calculate: f32 ± to.100 Chapter 5 .2: ttest in Example 5 .2. This is shown in Figure 5.
Distributed lag models are extensively discussed in Chapter 14. we can judge whether the influence is positive or negative depending on the sign of P3 :  Ho : P3 = 0.02 (2 ~% of the left tail and 2 ~% of the right tail). On the other hand. Student tdistribution HI 21/2% H. If P3 > 0. + "t. If the null hypothesis is rejected.2 A twotalled ttest Suppose that in the model of Example 5. Suppose the influence of X. that X .3 has no influence. it is possible that to observe that P3 < 0.3 A different type of Consider the model : Y. on Y.02 Figure 5. See Figure 5.2.3 as an illustration of this test. X. + P3 Xt.02. se ( P3 or P3 < (~ ) se P3 . = Pt + p.3: Example of the twotailed t test in Example 5. we read again : to. The sign of P3 can be positive or negative. is positive for econOrrllC theoretical reasons: f3z > O.J + P4 Z. is tested with a twotailed test. has been completely effected in Yi. 1 the influence of X'3is unknown. Then Ho is rejected in favour of H I if: P3 ~ ) > 2. In the same table. 21 12% . Then the null hypothesis.2.2 Example 5.O' 5 (40) = 2. H t : P3 oF O. the model is called a distributed lag model: it takes more than one period before a change in the level of X.02.Interval estimates and the ttest to test restrictions on parameters 101 Example 5.02 o +2. which can be seen as a .
if it looks that fh '" f33. cov ~ ($2. the (restricted) model can be respecified with the first differences of X. o. or written as Ho : f32 HI + f33 : fh + f33 = 0 . Click on ' View' in the 'Equation' window. This printout is obtained in EViews as follows. Examples to test this null hypothesis in a more convenient way are given at the end of this chapter in Example 5. and estimated as: which gives one more degree of freedom and less multicollinearity (see Section 7. However. $3) . and select 'Covariance Matrix' . Here the use of the ttest is shown. after the equation has been estimated. and the distribution of (h + $3) becomes: The tstatistic under the null hypothesis is: To obtain the last component in the denominator. a printout of the co variance matrix of {3 is needed.ication that more or less than just the change of X.. For example.4).102 Chapter 5. ('\7 X" see Remark 5. f33. For example. this null hypothesis can be tested with a ttest. . after a regression. Testing the Deterministic Assumptions correction for an overreaction on the change of X t in period t . Write the null hypothesis as a homogeneous linear equation Ho : fh + f33 = 0.5 below).6. then it might be relevant to test twosided: ~ ~ Ho : fh = f33 HI : f32 . If the null hypothesis is not rejected. It is just an example because the ttest is not the most convenient method to test this restri ction. in period t is relevant for determining the level of Y. . it can also be an ind.. an Ftest and a Likelihoodratio test (LRtest).
. The example still concerns one restri ction.The Wald Ftest to test restrictions on parameters 103 Remark 5. (5. Two difference operators are distinguished. = X. 15) R is a (g x K ) matrix. . To illustrate thi s notation Example 5. which can directly be done in the EViews 'Equation' window.7 The Wald Ftest to test restrictions on parameters In Example 5. The null hypothesis is: Ho : (32 + (33 = O. However. This is only one restriction.X' _l. with the Ftest it is possible to test more than one restriction simultaneously. Ho : R i3 = r. The restriction is written in the general notation as: (0 1 1 0) = (0). and the delta: t.3. In general notation in thi s section. and The 'Voperator will mainly be used in economic models. the nabla: 'V. 9 Hnear restrictions on various parameters will be tested. starting in the previous Example 5. Simple examples are found in this chapter. a hypothesis was tested about more than one parameter by using attest. and more extensive examples are found in Pan IV.3 from the previous section is used. Such a hypothesis can be tested more conveniently with a Wald Ftest. so 9 = 1 and the assumption was that K = 4. as differences with respect to a previous period in the past are more usual . 5.3.5 Notation for a difference variable Modelling variables in first differences will regularly occur in econometrics. The null hypothesis concerning 9 linear restri ctions on K parameters is written in matrix notation in the following way. i3 is a (K x l ) vector and r is a (g x l ) vector.. These operators are defined as follows: 'V X.
. tbe F statisti c can be computed and the . Nex t. The Fstatistic (5. (n _ K ) (a~/(J~) 1(n _ K ) After substitution of the null hypothesis null hypothesis can be tested: R = r .8)) ~ F (g. The OLS estimator and a~ are independently distributed.K )distribution of a~ . n . write the distributi on of by using the properties and notati on discussed in Secti on 4.O'~R (X ). the quotient of both X2distributi ons is taken: .7: R .4. Testing the Deterministic Assumptions The teststatisti c and its distribution will be derived as follows. according to the definiti on of the F distribution.8 ~ R~ ~ N (R. It can be eliminated by using the x' (n .8)' (R(X )JR 'X ') which is rewritten as: .1 (R~ .R Ig .R ~ N (O.9) from Section 5. 17) is computed after only the unrestricted model bas been estimated. O'~R (X r' .K ).8 ~ (<k (R~ .8. The notati on of the X2distribution (5.8 This is illustrated in Figure 5. The distribution of the OLS estimator {3 is: ~ assuming that all the class ical assumpti ons have been satisfied .4 is used to determine the distribu The variance O'~ is an unknown parameter of this statistic. So.1 R') 'X (R~ . For that reason the F test is a Waldtest.8) 'X tion of the corresponding quadratic form : R') .104 Chapter 5.R.
'. After the equation has been estimated: click 'View'.4 Zero slopes test In the standard output of a least squares procedure.1 restrictions on K .l ) x K ) matrix.. For example.4: Ftest Example of the use of the Wald Example 5..' . Example 5.. there are K .l ) x l ) vector with zeros.4.. consider again a model in general notation : where the hypothesis will be tested: Ho : {32 = {33' . next specify the restriction(s) in the 'Wald Test' window.1 parameters: 9 = K . . X ' K' Using the standard formulation in (5. = . This Fstatistic is used to test the null hypothesis of zero slopes: Ho : fh = {33 = {3. 'Coefficient Tests' and 'Wald Coefficient Restrictions . see Figure 4. and the explanatory variables X'2 .distribution 5% o Figure 5. All the tests on parameters are found under ' View' in the 'Equation' window.5 Testing a restriction with EVIews How can we use EViews to compute the F test? In EViews.The Wald Ftest to test restrictions on parameters · 105 F..l.15). which implies that no linear relationship exists between Y. R is a (( K . = 13K = 0. . an Fstatistic is always found. restrictions on parameters are very simply tested with the Wald Ftest. and r is a (( K .
The Fstati stic has a pvalue of 7. The differences with the output of the unrestricted model in Figure 4.square Value df Probability I Null Hypothesis Summary: Normalized Restriction (= 0) Value Std.6). so the null hypothesis is not rejected at the 5% significance level.8 later. so the restriction does not disturb the original unrestricted fit to the data.6: EViews output concerning (he Wald F test . The output of the regression with the restriction is given in Figure 5.5).010372 Restrictions are linear in coefficients.018263 0.7. That is all! In the example below. The output is given in the equation window (see Figure 5. The Wald Test: Equation: EOOl Test Statistic Fstatistic Chj.5: EViews Application of the Wald F lest in The restriction Ih = /33 or Ih + /33 = 0 is written in EViews as: c (2) + c (3) = 0 in the 'Wald Test' window that has appeared (see Figure 5.85%. The x2 statistic is explained at the end of Section 5 . is tested for the contemporaneous and lagged coffee price in our relationship between cocoa and coffee prices. Err. although it is not a very convincing result.c°! and P. as formulated ahove.":'{ individually.106 Chapter 5. C(2) + C(3) 0. Figure 5.. Testing the Deterministic Assumptions Figure 5.4 are rather small. a null hypothesis.c°! can be specified instead of P. If Ho is not rejected then \l P.
745467 O .044025 1.044036 0.{33 Akaike and Schwartz information criteria are discussed in Section 14. the following identity: Fo. The first one is a repetition of the definition of the tdistribution (5.ocom 0. Error t·Statistic Prob.and Ftest In case of one restriction From the statistical literature it is known that no difference exists between the F test and the (twosided) ttest if the null hypothesis concerns only one restriction. Look at the following two formulae.707789 2. r) .071036 0.988733 Mean dependent var S.0063 3.D.666275 11833. n .622490 2.The Wald Ftest to test restrictions on parameters 107 I Oependenl Variable: LPCOCOA Method: Least Squares Dale: 02I20All Time: 09:31 Sample(adjusled): 1960:04 2OO'lm02:09 .7: Regression output with the restriction {32 = .062181 1.120869 Std.K) = t~.040786 0. Tbe second relation is tbe definition of the Fdistribution (5.o5 (1. This implies.o::ro R·squared Adjusted R· squared S. dependent var Akaike info criterion Schwarz criterion F·stalistic Prob(F.810081 30.31 O .9B9360 0. statistic) Figure 5. these can also be used to judge the estimation results in this example.558213 2.114613 1. . for example.7.337967 0.4).996166 0. 0.38338 6.E.989444 0.5) applied to the variables x 2 ~ X2(1) and v 2 ~ x2(r): X x 2 ~ F(l. Remark 5. Included observations: 510 after adjusting endpoints Variable Coefficient O .952551 695.00:0 0.025 (n .4863 1.018527 0.6 Relationship between the t.K ) . of regression Sum squared resid Log likelihood DurbinWatson stat 0. Jv 2 / r ~t(r).1053 0.0051 0.(OO)5() ' 0.602815 2.
17) is exactly equal to: F= (8 R 8) / g 8/ (n . In fact. This is done 'by using the Lagrange function : Differentiation of 8 • (jj R . Proof The result (5. The model with restrictions on the parameters is written as: y=Xi3+u with 9 linear side conditions: Ri3 = r.18) when software other than EViews is used. as will be seen in following chapters. Then estimate the restricted model and save the sum of squared residuals 8 R . Testing the Deterministic Assumptions 5. which can be convenient depending on the software used. where it is less easy to test linear restrictions. 18) is frequently encountered in econometrics. Because expression (5. compute the OLS estimator by minimising the sum of squared residuals subject to the 9 linear restrictions on the parameters.K ) = g(j~ .K ) (5. It is now possible to prove that the Fstatistic (5. estimate the unrestricted model and save the sum of squared residuals 8 . it is useful to prove (5. 18) This expression can be used to test very generally formulated restrictions. the only thing that has to he proved is that the numerator of (5.2X'y + 2X 'Xi3 R  ~ R 'A = 0 . A) with respect to jj R and A results in the firstorder conditi ons: = .8 Alternative expressions for the Wald Ftest and the LMtest It is possible to compute the statistic (5. 18) is achieved after the following steps. Next.1 7): because for the denominator you immediately find : g8/ (n . 18).108 Chapter 5. First. The statistic is simply computed by hand. It is also a convenient way to compute the Fstatistic (5.17) in a completely different way. 18) is equal to the numerator of (5 .
yields: 109 with jj = (X'X) .7 The coefficient of determination Equation (5.Alternative expressions for the Wald Ftest and the LMtest Multiplying the first condition with R (X'X ). (5. Write the residuals from the estimated restricted model as:  eR = Y .x jj + X (X'X ).20) or by writing SR = eneR and S = e'e: This is equal to the numerator of (5. It is not necessary to go on with determining the di stribution of {JR. 17). 19) shows the difference between the residuals of the restricted and unrestricted model. 18). The equation demonstrates that the residual sum of squares (RSS) decreases by adding variables! The equation . Next. The crossproduct vanishes because of the orthogonality property X 'e = 0 .1 X'y . gives: .  With this formula restricted OLS estimates can be computed.1 and solving for A. which was derived in equation (5. Substituting for A and solving for {JR . Remark 5. the restricted OLS estimator.X{JR =  Y . etc. which proves (5.1 ).20) has one more important consequence. compute the sum of squared residuals by multiplying with its transpose both sides of this equation. The result is: (5 .1R ' ( R (X'X r 1 R yl (Rjj  r) This is equal to the following expression.18) will be proved.19) Relation (5. because only (5. as the expression for the unrestricted OLS estimator.
This argues the statement made earlier that tbe coefficient of determination R2.S )/g .4: L(8) = 697.   LR = 2l n (A) = 2 (695. Next.6 was F = 3.S/ (n. This LRstatistic has a X2 (1) di stributi on.1.(SR .0536. Read in Figure 5. All the results concern the null bypothesis of one restricti on: Ho : f3.7: L (8 R ) = 695.4863 . 18).1 .940586) / 1 = 3.K )' F = (1.1") t= l n increases when any (correct or nonsense) variable is added to the model! Example 5._ _ .0536) = 3. • • The result from Figure 5. Nex t. In Figure 4 .value of 0. defined as: R2 = 1 _ L e~ . in our example with cocoa and coffee prices. compute (5. Testing the Deterministic Assumptions implies tbat e'e < eReR. . some computational results are shown concerning statistics discussed in this chapter.107280 with a p. compute the F statistic by using (5..0766. + (33 = O . 1346.4863.697.8 Computational In thjs example.18): F .:='::.7 we read SR = 1.940586/ 504 5 The smaU difference with the result from Figure 5. 2' L (Yi .952551 . The matching pvalue is 0.110 Chapter 5 .6 is due to rounding errors. The same hypothesis can also be tested with a like lihoodratio test (see Section 5.4 we read S = 1.107 .4): • Read in Figure 4.0785. 1.940586.952551. In Figure 5.
as introduced above.8 Other expressions for the statistics It is possible to obtain a large sample test for the Waldtest.6). In EViews. It concerns the R2 of the regression of the residuals from the restrictedly estimated model on the explanatory variables of the unrestricted model. Greene(2000). see. the LM test statistic is often computed as nR2. with 9 equal to the number of restrictions. The statistics can also be expressed in the coefficient of determination R2. by using the equation: TSS = ESS + RSS. which can be written as: and for the Lagrange multiplier (LM) test: with S R and S as the sums of squared residuals. • . for example. Once again. or Stewan and Gill (1998) for more information about these test statistics. but is not used in this book. For example. Both test statistics have under the null hypothesis a X2 (g) distribution. The expression • LM = nR2 for the LM test will regularly be encountered in Chapter 6.Alternative expressions for the Wald Ftest and the LMtest 111 Remark 5. the value of this X'test is also found in the output window of the F test (see Figure 5. MaddaIa (200 I). An expression of the F test in terms of the R2 does exist as well.
.
and homoskedasticity. If these assumptions are not rejected. Errors made in the specification of the equation are found in the disturbance term. It makes no sense to test hypotheses about parameters in a wrongly specified model. as formu lated in Chapter 4 . If that is the case. a number of statistical and econometric tests will be di scussed so that all the disturbanceterm assumptions. The disturbanceterm assumptions can be compactly written as: (6. any systematic process for the disturbance term should be absent. Therefore residual autocorrelation and/or heteroskedasticity are primarily considered as an indkation of specification errors in the systematic part of the model. If these symptoms are found in the residuals. 1). In Chapter 8. then the right way in general is to respecify the systematic part of the model to get wellbehaved residuals. can be tested.Chapter 6 Testing the Stochastic Assumptions and Model 6. When a proper economic model is specified. the validity of the stochastic assumptions about the disturbances has to be tested. the t. then the t. the beterosked asticity . Before interval estimates.1 Introduction In this chapter. can be used in practi ce. we shall also see that heteroskedastic behaviour of the disturbance term is not always caused by specification errors but can belong to the model. no autocorrelation.and Ftests are carried out assuming (6. as discussed in the previous chapter.1 ) For individual disturbances this implies normality.and Ftests.
9. Most of the items in 'Residual Tests' that are shown will be disc ussed in this chapter. [n this chapter. Residual'. The dotted lines around the residual grapb are ODe standard error of regression: 0 ± 1 . [n the following sections. a simple normality test and some tests for autocorrelation and for heteroskedasticity are introduced. outliers or other systematic behaviour can directly be observed in a picture of the residuals. it is time to check the validity of the last assumption: model stability. In addition to using these statistical and econometric tests.1: The options for residual tests in the 'Equation' window can be modelled too. 'Residual Tests' have been selected. as is demonstrated with the SUR model (see Section 8. Testing the Stochastic Assumptions and Model Stability LPCOFFEE{·l) LPCOCOA{l) LPCOCOI'{·2) LPCOCOA{·3) Mean dependem var Figure 6. How . The window shows the different views of the regression result. In the 'Equation' window in Figure 6.1 the options under ' View' are shown. it can be informative to look at a plot of the residuals.3). Fitted. This is the plot of the residuals from the regression in Section 4. ii•. At the end of the modelling process. Choose under ' View' . we will look at a variety of residual tests. and select 'Graph' to get a picture like Figure 6.114 Chapter 6. Some outliers are clearly observed and can be investigated. Autocorrelation in the disturbance term can arise by building a specific type of model. Cycles. 'Actual.2. The same plot is obtained by clicking on the button 'RESIDS' in the equation window. when you are satisfied with the results of the tests with respect to the stochastic and deterministic assumptions.
Fitted 5 .5..~f7="'r~UJ~~ ry6 Actual...Fitted I Figure 6..Bera (J B) test. The J Bstatistic is computed in EViews as: JB =(n.0 . In the last section of thi s chapter. Always the use of the tests with EViews is discussed immediately after their theoretical introduction.2 .. the first research project is formulated in Case 2. The J Btest statistic.~ 115 _ 1'1". but nowadays the test is always referenced as the Jarque.2) . (6. Case 2 is preceded by a section where some practical recommendation s are given about writing a research paper and presenting estimation results. See also the statistical literature for more detailed information.1 . I {j ill "" ''' ' Ir1 .2 Residual n_• . is defined in the statistical literature as : JB = n 52 (K _ 3)2 24 '6 + .. For that reason.r"'!rr.1"". This normality test goes back to its introduction by Fisher (1948).2: The actual and fitted observations of In (PtCOC) and the residuals to judge the model stability is discussed in Section 6. The test can be found in many statistical and econometric textbooks. This normality test is included in EViews.r7f'1 ~17"T"'""': . I II! 111111" rr.Bera test (Jarque and Bera (1980)). and is X 2 (2) distributed under the null hypothesis.2 A normality test A standard normality test is the Jarque. Actual .A normality test .1 . The null hypothesis is Ho: the residuals are normally distributed. 6. which can be used for models with a constant term.K ) 52 (K _ 3)2 24 '6+ .
The individual statistics Sand K of the test are printed too. The nortnality assumption is rejected at the 5% significance level if JB > 5. In Figure 6.9 again. the output with the histogram of the residuals and some statistics are shown. as mentioned in the introduction. but if the observation is right and a reason for the outlier can be found. Deviations from nortnality can be caused by outliers in the residuals. It is known that the thlrd moment of a sy mmetric distribution is zero. K=1'4 = 1'4 a4 The moments can be estimated from the OLS residuals: ~ 1" i J. so for a normally distributed variable we have 1'3 = O. The null hypothesis is clearly rejected . When an outlier is observed. you may eliminate that observation by specifying a dummy variable in the regression equation. the X2 (2) distribution has been drawn. probably on account of some heavy outliers: see the plot of the residuals in Figure 6.4. The test has a X2 (2) distribution under the null hypothesis of normality. 4.2.~ etl n t= l n ·th 1.99. then try to find the reason for the outlier. A good tip.3. In EViews the JBstatistic (6. With the help of the moment generating function for a variable from a normal distribution. Testing the Stochastic Assumptions and Model Stability with S being the measure of skewness and K the measure of kurtosis.· = WI 2) 3. (The use of dummy variables will be discussed in Section 7.2) is computed by selecting 'Histogram . is to look at a plot of the residuals. In this example. These values of Sand K imply that both components of the JBstatistic in (6. . 2) are zero under the null hypothesis of normality. making it possible to observe which component causes possible rejection of the normality assumption. The figure concerns the residuals obtained from the regression example in Section 4. If normality has been rejected.the residuals are not normally distributed.Li = .116 Chapter 6. The JBstatistic is expressed in terms of the thlrd and fourth moments of the disturbances: S = 1'3 = a3 1'3 (1'2 )3/ 2 tL~ .5. tbe test gives no clue about how to 'repair' this failure. This nortnality test is a nODconstructive test. In Figure 6.) The pvalue of the JBtest will increase in that way. you can derive that: 4 2 1'4 3a 31'2'  This implies the following values for S and K when the variable is standard normally distributed: S = OandK = 3.Normality Test' from the 'Residual Tests' . If it is a typing error the observation can be corrected. the nortnality assumption has clearly been rejected.
139687 0. Skewneu 0247529 ll .589014 3.4: The X2 (2)distribution . two exceptions exist where systematic behaviour arises in the di sturbance term. In general. However.959870 49. One reason is the explicit specification of 5% 2 5.99 Figure 6.3: The 18 nonnality test 6.88E·16 ll. no economic reason exists to make an assumption that the disturbance term in an economic model is autocorrelated. After that has been done.06835 O.rn std. Mode ls with autocorrelated disturbances because of theoretical reasons are discussed in Parts III and IV.COOOOO Figure 6. Oev.B) is unsystematic. When an economic model is specified from the economic theory you specify the systematic part of the model.oo~475 MilIint.3 Tests for residual autocorrelation It is obvious that autocorrelation is a phenomenon that can only occur in models for timeseries data. the assumption is made that the difference (u ) between the dependent variable (y) and the systematic part (X .Tests for residual autocorrelation 117 $er.es: Residuals Sample 1960:04 "200""'2:09 OI>servelions 51 0 K\I1osis Mean _an 3 .D61746 0.
where u.3) with u.l (6. (Some of these transformations are discussed in Part IV in more detail. When low frequency data is analysed. In both situations. The differences concern the order of the autocorrelation.2. A second theoretical reason that autocorrelated disturbances can arise is by a transformation of an original model. for example. for example. All tests have the same null hypothesis: the absence of autocorrelation in the disturbance term.) is written as: Vt = Ut . which can result in a movingaverage or MA(I) disturbance term. quarterly or monthly data is analysed. The tests have different alternative hypotheses.118 Chapter 6.5 for some explicit expressions of the matrix 0 in case of autocorrelated disturbances and estimation procedures for the parameters of this modeL ) In a standard situation. <7.) Then we have autocorrelated disturbances which belong to a correctly specified model. a model where an (adaptive) expectations variable has been eliminated (see Section 13. or a distributed lag specification for an explanatory variable has been eliminated by Koyck's transformation (see Section 13. When. we do not expect strong dynamics. it is assumed that the covariance matrix 0 is known. An MA(I) disturbance term (v. The Student's tdistribution of the tstatistic has been obtained on the basis of the N I Dassumptions of the disturbance term. An example of an MA( I) disturbance term is. theconsequences of autocorrelation for the statistical properties of the OLS estimator are discussed. In Section 7. you have knowledge about the form of the autocorrelation andlorheteroskedasticity. that high values for the tstatistics and R2 are observed in the output.) . . as happens in the SUR model in Chapter 8. Such a model is specified with differeD! assumptions. Various autocorrelation tests exist. the model had been specified with an insufficient number of lagged variables or not all the relevant explanatory variables have been specified in the model. The consequences of autocorrelation in the residuals are • • that standard errors are underestimated. more dynamics can be expected and testing for higber order autocorrelation is relevant.AUt. leading to seriously misleading results if the output is not correctly interpreted. (See Section 8. as the critical values are much higher than the values as given in the table of the tdistribution. For example. we know that residual autocorrelation means that the systematic part has been misspecified. We will see that a lot is wrong with the estimates and the test statistics when the model has been misspecified. is the di sturbance term of the original model. OJ. Testing the Stochastic Assumptions and Model Stability autocorrelation. Thus it is not necessary to test for the absence of autocorrelation as it is known that v. is autocorrelated.3). ~ N I D (D. If these assumptions have been violated.8 + u u ~ N (0 . so tvalues are overestimated. the evaluation statistics in the computer output may not be interpreted in the usual way.4). for example: y = X. So testi ng for first or second order autocorrelation can be sufficient. [n theory.
. The process is called an AR(P) process for Ut: Ut = 'PIUt. Because the 'omitted' variables are residuals and not independent variables. withVt ~NID(O .5). Autocorrelation of the pth order is represented by an autoregressive model of order p as specified in (6. The BreuschGodfrey LMtest • Breusch (1978) and Godfrey ( 1978) have developed an autocorrelation test that is introduced here with an example. + 'PpUt .4) is the alternative hypothesis. when using quarterly or monthly data it is useful to look at seasonal fr~uencies in the residual autocorrelation function.4) The alternative hypothesis. a. The statistic can be computed as nR2 . are zero. = 'Pp = O. They are discussed below. in words: the disturbance term has autocorrelation of the pth order.5) The null hypothesis is identical to the hypothesis that: 'PI = 'P2 = .Tests for residual autocorrelation We wiu look at a number of tests with the null hypothesis: 119 Ho: no autorreJatioD in the djsturbance term. So specification errors produce ARprocesses in the disturbance term.p + V t· (6. BG(p) has an asymptotic X2 (P) distribution under the null hypothesis of no autocorrelation: BG (p) = nR2 l:!J X2 (P). The idea behind the BreuschGodfrey (BG) test is as follows . (6. such as four or 12. the residuals have been computed as et. For example. Suppose that the following linear model has been estimated with OLS: Yi = {31 + {32X t + (33 Z . with the R2 of equation (6.1 + 'P2 Ut . The test statistic of the BGtest is an LMtest that is denoted as BG(p) when (6.. and the alternative hypothesis HI : Ut = 'PIUt.1 + 'P2Ut.4). + "t· Next.2 + . + 'PpUt.. . In EViews the BG(p)test is called the serial correlation LM test. the exact finite sample distribution of this F statistic is not . In EViews two tests for pth order autocorrelation are found when selecting ' Residual tests': the BreuschGodfrey test and the LjungBox test.2 + . ). The test is performed by estimating the following equation with OLS.p + Vt .. or a multiple of these orders. EViews also computes an Fstatistic to test that all the 'P.. to check the presence or absence of seasonality in the residuals.
120 Chapter 6.347999 0.Ol l 077 . Test Equation: Dependent Variable: REStO Method: Least Squares Date: 02118. for the same reason as the F statistic. Select ' View\ ' Residual Tests'.539467 0.536213 0.151461 .263162 ·0.260997 .351167 0.387986 0.7925 0.021336 0.(l3 Time.387045 0.Q. as shown in Figure 6.503877 1. versus H I: autocorrelati on of the 3rd order.5: BGtest Lag specification known.7942 0.598339 0.Q. I dQ Speclfl r: cthon EJ Figure 6. " then the BreuschGodfrey test will be computed after the order p has been given.Q.6.Q. You have to type a relevant value fo r p yourself. 0.1251 C LPCOFFEE LPCOFFEE(·1) LPCOCOA(1) LPCOCOA(·2) LPCOCOA(3) RESIO(I) RESIO(2) RESIO(·3) . The order is specified in the ' Lag Specification window'.7062 0.OO5569 .I67665 0.145996 O .042768 0. or at best onl y in a possibly indicative way.377200 1. It is quite possible that the probabilities of the two statistics n R2 and F differ considerabl y. 'Serial Correlati on LM test .213713 0. .6: Result of the BG(3)lest .523348 .390378 1.2426 0.Q. Test the null hypothesis Ho: no residual autocorrelation.6964 0.031767 .169!XE 1 .Q. The default value in EViews is p = 2. 09:35 Presample missing value lagged residuals set to zero. Error '·Statistic Prob. The regression output of equati on (4..9747 0. so individual !PiS and their tvalues are shown but.5.Q.Q.70c001 . The output appears in the 'Equation' window.OOI359 . 12) is also given. The example with the commodi ty prices is continued.1332 0. therefore do not consider the tvalues. Variable Coefficient Std.138142 .2254 Figure 6.042090 0.Q. the distribution of these tstati stics is unknown. A part of the output windows is given below. Testing the Stoc hasti c Assumptions and Model Stability ..Q. as shown in Figure 6.
It is interesting to look at a plot of the coefficients Pi. The Ljung.7.Pierce test will be shown as the Q' test in this book.4 for more information. has an asymptotic X2 distribution. in fact. In EViews.7: The X2 (3) distribution The BG(3)test has apvalue of 37. The Q' test does not 'look' at the individual autocorrelation coefficients but considers the sum of a number of squared autocorrelation coefficients.Pierce Q' test is defined as: p Q·=n L P~. The Box. The theoretical autocorrelation coefficients Pi of the disturbances are zero under the null hypothesis of nonautocorrelation. with X5. The BoxPierce test. koown as the autocorrelation function (ACF). These tests make use of autocorrelation coefficients of the residuals. i= l . or beller formulated.05 (3) = 7.Box test will be discussed in more detail in Chapter 14 with respect to its use in estimating univariate timeseries models.8 Figure 6. or its modified version the Ljung.Tests for residual autocorrelation 121 5% 3 7. we check whether the computed sample autocorrelation coefficients of the residuals (Pi) are not significantly different from zero. so the null of nonautocorrelation is not rejected against the alternative of thirdorder autocorrelation. The BoxPierce and the LjungBox tests The Q' test of Box and Pierce (1970) is a useful test. when testing forpthorder autocorrelation. The X2 (3) distribution is plotted in Figure 6. as it yields a useful picture of the correlation behaviour of the residuals. and therefore the Box. So.815. the LjungBox test is called the Qtest. The use of the ACF in a regression analysis is illustrated in this section.Box test. The estimated autocorrelation coefficients are defined as Pi: see Section 14. with p degrees of freedom under the null hypothesis of nonautocorrelation as well.03%.
. Notice also the large difference in the pvalue of the B G(3) and . an example is given showing the correlogram of the cocoa residuals. In Figure 6. The first column is the autocorrelati on function. 'CorrelogramQstatistics·.6) is found in the ' Equation' window under 'View' . Therefore. In Section 14. The difference between the Qtest and the BGtest is that for the BGtest you have explicitly to choose one specific order of the autoregressive process specified under the alternative hypothesis. it is more customary to use a significance level of 10% instead of 5% when using the Q' test. or the probabilities of the Q statistic give some information about possible residual autocorrelation.. In the standard linear regression model it makes no difference in which way the Qtest is computed. ).8. The Qstatistic is defined as: P i= l ::2 Q=n(n+2) " Pi . 3. it is convenient to keep (a part of) the residual correlogram in the comer of the screen because after each regression the correlogram is refreshed with the ACF of the residuals from the new regression. no significant autocorrelation coefficients are observed. n . Testing the Stochastic Assumptions and Model Stability and has a X2distribution under the null hypothesis: Q' ~ X2(P). the pattern of the computed autocorrelation coefficients Pi (i = 1. The third column gives the Qvalues and in the fourth column the probabilities of the cumulating Qvalues are found. One more practical problem: if you wish to use the statistic as a test. Look at the pattern of the pvalues.6) This test is widely used. The second column is the partial autocorrelation function (PACF): this function will be explained in Section 14. When modelling from general to specific you can see immediately when things go wrong.. However. However. (6.122 Chapter 6. whereas the test may have low power if too large a number of lags have been chosen. ACF. It is also possible to doubleclick on RESID and to choose 'Correlogram'. In practice. but the Qtest of Ljung and Box (1978) is used. The Q' test is not computed in EViews. . Q ~ X2(p). In EViews. the Qtest (6. 'Residual Tests '. Report in your paper the result of the BGtest for one or two values of p. in practice. you must choose the number of lags to use for the test. in the econometrics literature it is discussed that this test has low power in detecting autocorrelation. The LjungBox statistic is a modified BoxPierce test that has the same asy mptotic di stribution but is closer to the X2 (p )di stribution in finite samples. L. 2. but in the econometrics literature it is argued that the test is not an appropriate test in autoregressive models or models with lagged dependent variables. Then the Qtest can only correctly be computed via the 'Equation' window.5.7 we will see tbat a difference between these two procedures exists when the null hypothesis is tested for the residuals of a univariate timeseries model. By choosing too small a number of lags you may not detect higher order serial correlation.
Therefore. for example.047 0. . Its use is very limited because of these conditions. versus the HI: firstorder residual autocorrelation.176 10.714 0.044 0. the DW test has the same limitations as the Qstatistic in dynamic regression models.. and Ct ~ N ID (o. .. .l + Ct..048 0. the examples with EViews regression output windows. 0.515 0.698 0. 0..: 11 :17 Sampl.036 0. Durbin and Watson (1950) test the Ho: no autocorrelation..043 1.: 1960:04 2002:09 Included observations: 510 Autocorrelation Partial Correlation PM.104 10. .031 0. The DW statistic can be used if the explanatory variables are exogenOus and a constant term has been included in the model.6868 0.8: The Ljung. 4 0.Box QIesl the Q(3)test.894 3 0.036 0. .868 .5161 4.021 2.594 .. Higherorder autocorrelation is not detected.7) .246 0. . Figure 6.043 0..048 0. The DW test will be discussed here for that reason: it is always found in the output of a regression analysis. so you have to know the background of the test.: 02I1MJ3 Tim.00.. One more limitation is the alternative hypothesis of firstorder autocorrelation. .. you have to be sure that firstorder autocorrelation is the only possible form of autocorrelation that is relevant for your model. with Icpt! < I .7229 0.Tests for residual autocorrelation 123 Oat.7553 10.oD (6.012 0. .020 0.em 0.Watson (DW) test is the most wellknown autocorrelation test.031 0. . Therefore.897 2 0. .0168 0. but it is often not possible to apply it.431 0.5428 6. it is a strange phenomenon that the DW value is always printed in the regression output.797 0. However. OStat Prob 1 0. . . But why would you limit yourself to test only for firstorder autocorrelation as the previously di scussed general autocorrelation test(s) can simply be used? Firstorder autocorrelation in the disturbance term is represented by an AR(1) model: Ut = CP IUt. see.020 0.2242 0.012 0..793 5 6 7 8 9 10 11 12 0.D18 0.079 0.663 0.050 0. The DurbinWatson test Probably the Durbin.3352 3.048 o..5573 5.047 0.742 0. . which could have been caused by the ineptitude of the Qtest as a statistical test in the dynamic regression model.045 0. ..00.
9) Expression (6. If DW > 2 because of possible negative autocorrelation. Negative residual autocorrelation: 'PI < 0 ==> DW > 2. .8) where e. (6. There is one more problem with this statistic.2 E e. • • • No residual autocorrelation : 'P I = 0 ==> DW '" 2. However.124 Chapter 6. the DWstatistic is usually tabulated for ·values smaller than 2. therefore. to obtain upper (du ) and lower (dLJ bounds for the true critical values of the DW statistic. Testing the Stochastic Assumptions and Model Stability The DWstatistic has been defined as: (6. we see the following properties of the DWstatistic.8) is between the distributions of the lower and upper bound. . the same table can be used for the statistic (4 . The unknown distribution of (6. as (4 . Positive residual autocorrelation: 'PI > 0 ==> DW < 2. A simple relationship ellists between the parameter 'PI in equation (6.2'P1 = 2 (1 'P1) . If DW < dL: reject the Ho in favour of firstorder residual autocorrelation. These two distributions depend only on the number of observations (n ) and the number of explanatory variables (K ). Tbe following co nclusions are drawn (in case the DWtest is used). If d L < DW < d u : the test is inconclusive. it is possible to derive distributions of a lower and an upper bound for the DW statistic. you regularly find the recommendation to use du as the true critical value because the inconclusive interval can be rather wide.e'1 E er E er "' 1 + 1 . From this relationship. In the econometrics literature.7) and the DW statistic: DW '" 2 (1  'Pd . • • • If DW > du: do not reject Ho. are the OLS residuals.DW). The distribution of (6.d' E er + E erl .'PI ) if 'P I < O. This means that the DWstatistic cannot be tabulated.e.9) is found by rewriting the DWstatistic (6.8): DW = E (e.8) depends on the matrix X with observations on the explanatory variables.DW ) '" 2 (1 + 'P d which is identical to 2 (1 . One can prove that the distribution is symmetric around 2.
which has been introduced as a new test statistic.= 1 . it is more informative to publish the results of the BreuschGodfrey test in your paper. In Chapter 8.8) ljJ N (0. They have been discussed or mentioned in this section because you still find the m in many articles and software. but that is a clear paradoxical si tuation. In facI. all the ingredients of the following expression are known . beteroskedasticity occurs in the mode l. In thi s situation. When heteroskedasticity occurs in the model it is possibl y caused by a rela tionship between the disturbance variance and one or more variables. 6. The value of the DWstati stic can be very misleading in such a mode l. In this section. Stewart and Wallis (198 1)). Therefore. a second variant exists . n . the htest cannot be computed. are clearly inconsistently estimated by using the autocorre lated OLS residuals (see. Take good note of the facts from this secti on and mention the DW statistic in your paper only if all the conditi ons have been sati sfied. Analogous to the situation with autocorrelation.. when trends occur in the observations of timeseries data. . it is quite possible that. This test is not found in EViews.Tests for heteroskedasticity '125 As menti oned above. For this situation. However. the test can only be used if the explanatory variables are exogenous and a constant term is included in the model. Although heteroskedasticity can be an indication for specificati on e rrors 100. etc. like a DW. If (1 . for exampl e. beteroskedasticity in the residuals causes underestimated standard errors. Its definition is: _C.v aT (. 1) . Durbin 's htest has been computed in empirical articles. 0 0 not use the DW stati stic iflagged dependent variables are present as explanatory variables in the model. but it can easily be computed if you wish to know its value. heteroskedastic di sturbances can also be found in models for timeseries data. The test is consiste nt when 'PI = 0. [n such a situation. But it can also be caused by the data. elc. then they can induce trends in the variance of the variables. it can happen that a similar trend is found in the variance of the disturbance tenn as weU. overestimated tvalues. In Chapter 12 we revert 10 the DWstatistic. contrary 10 autocorrelation. you will freque ntly see that.the Durbin's alternative test . More variants around these tests can be found in the literature. models are di scussed with an explicit specification of heteroskedastic di sturbances.8 is the estimate of the parameter OfY' _ l in the model.4 Tests for heteroskedasticity The phenomenon of heteroskedastic ity in the disturbances occurs mainl y in models for crosssecti on data. test for quarterly data. It is not difficult to demonstrate that cp" and so the DWstatistic too. or their variances.8) ) is negative. for exampl e. we consider tests for the following three situations. For example.v aT ( . all these lests around the DW statisti c have to be seen in a historical context. Durbin's alternative test is similar to the BG(I )tesl. where . For both mentioned fonns of heteroskedastic ity an alternative hypothesis can be formulated to test the null hypothesis of homoskedastic di sturbances.
one has to estimate such a relationship. one of the following regressions is calculated with OLS. Because of the generally formulated alternative hypothesis. in words: • • Ho: the variance of the disturbance term is constant. The hypothesis that a~. The White test is actually an LMtest. Trends in the variance can occur in timeseries and crosssection data. 10) is the equation . This test is not a standard procedure in EViews.g. is also possible in models for timeseries data. in a crosssection the data will be organised according to the sequence of one of the variables (e. the White's heteroskedasticity test is found with two options: 'no cross terms' and with 'cross terms' (see Figure 6.. The White test In the di scussion of the tests the subscript i is used to indicate that these tests are more relevant for the analysis of crosssection data than timeseries data. For the test. and (the variance of) explanatory variables exists. the White test is discussed. caused by the levels. but it can be calculated simply in EViews. in 'Residual Tests'. especially in series across a long period of time. • • To conclude these introductory remarks: Ho implies the disturbances are homoskedastic and H I will be different with respect to the form of the assumed heteroskedasticity. For example. in a regression on linear and quadratic expressions of all the explanatory variables. The null and alternative hypotheses for the White test are. we look at possible monotonous trends in a~ . With the BreuschPagan test (Breusch and Pagan: 1979) the null hypothesis is tested against the alternative that a concrete relationship between a~. has just an increasing or decreasing trend in the sample period can be tested with the Goldfeld and Quandt test (Goldfeld and Quandt: 1965). so the nR2 is used again as a test statistic. followed by the two other tests. the test is called a nonconstructive test in the econometrics literature. This is done by using squared OLS residuals as the dependent variable. In EViews. In the third situation. The procedure is clarified with the next example. Also the Goldfeld and Quandt test statistic can be computed in a simple way by using regression results. squares and products of all the explanatory variables. It is the only heteroskedasticity test in EViews. Testing the Stochastic Assumptions and Model Stability • A general heteroskedasticity test is White's test (White: 1980). Suppose the specified model is: Yi = fh + fh X i + f33Zi + Ui · Next the parameters are estimated with OLS and the residuals ei are computed. The test is nonconstructive because it gives little or no indication about the type of the heteroskedasticity in cases where the null hypothesis has been rejected. A time trend in a~.126 Chapter 6.1). First. Equation (6. from low till high values). To make the test operational. HJ: the variance of the disturbance term is heteroskedastic of unknown form. It tests the homoskedasticity bypothesis against the alternative that heteroskedasticity is present.
0685 0.073977 0.03lJ76 I·St atistic Prob.(J.004316 0. C LPCOFFEE LPCDFFEE<2 LPCDFFEE(1) LPCOFFEE(·Iy.825824 1. nR2 ~ X2 (5) in equation (6.040357 0.Z.(J. We resume the example with the commodity prices.11) The null hypothesis is tested with the statistic nR2. 10).005919 0. e~ ="1'I+~~+~~+~~+~~+~~~+~.391273 0.10) (6.(J. 0.B0C6 0. Be aware that the number of degrees of freedom of regression (6.047850 0.274254 0.11).2252 0.040517 0. is not known and should not be used .329170 1.10) or (6. Test Equation: Dependent Variable: RESI()II2 Method: Least Squares Date: 02I11W3 Time: 11 :57 Sample: 1960:04 2002:09 Included observations: 510 Variable Coefficient Std.004336 0. is given in Figure 6.(J.001623 0.214265 0.660003 2.! LPCOCOA(·1) LPCOCOA~ 1)'2 LPCOCOA(·2) LPCOCOA(·2)"2 LPCOCOA(·3) LPCOCOA(·3)"2 Figure 6.007600 . which can reduce the power of the test. Erro.1844 0.009322 . the di stribution of the F statistic.092453 .003773 0. The statistic nR2 is asymptotically X2di stributed under the null hypothesis.0995 0. Similar to the BGtest.007157 . with n and R2 from the regression (6.001461 0.650343 ·1 .004581 0. 11 ) becomes very large when the model has many explanatory variables. which is computed by EViews again.020819 ·0.11) is with cross terms (in this example just one term).0313 0. applied on the cocoa residuals.017539 0.435094 .7840 0.29)936 ·2. (6.6637 0.0224 0.011592 0.Tests for heteroskedasticity 127 without cross terms and equation (6.040356 0. e~ ="1'1 + "1'2Xi + "1'3 X? + "1'.252693 .9.9: . 11): nR2 ~ X2 (4) in equation (6.5090 0.(J.003735 ·0.(J. Part of the window with White's heteroskedasticity test without cross terms.6958 Part of the output of White's heteroskedasticity test .(J.030362 0. + "1'sZ? + Vtl.159669 ·1.
the Breusch. Let us suppose that the variable X . Originally. + 0'. The variables for the testequation can also be selected by looki ng at scatter diagrams between squared OLS residuals and explanatory variables. The proposed procedure given above is simpler and asymptotically equivalent to the original procedure. for example. Next. the regression from Section 4. the White test is very general because after rejection of the null hypothesis. although the t statistics give some indication. regress the squares of these residuals e. So assume that the (lagged) coffee prices possibly generate heteroskedasticity. on X i and xl : e. so that the nR2 is used as a test statistic. More infonnation is obtained by using the next test. Breuscn and Pagan suggested regressi ng eU(j~ on some variables and computing ~ESS of this regression as a test statistic.128 Chapter 6. when it is used the first time. (See lohnston and DiNardo ( 1997) for more details and further references. The Breusch. nR2 has the following distribution : The number of degrees of freedom is eq ual to the number of parameters in the null hypothesis without the constant tenn.Pagan test.9 will be elaborated further. for example. As mentioned earlier.9 is unknown. or by using ' Procs'. the test gives no clear indication as to what has been the cause of the heteroskedasticity. + 'Y3 X . however. The BreuschPagan test If an idea exists as to which variable(s) may cause the heteroskedasticity in the disturbance term. for example: genr res = r esid. the following rel ationship can be postulated: The original OLS residuals have been saved with a different name. Testing the Stochastic Assumptions and Model Stability The null hypothesis of homoskedastic disturbances is not rejected with a pvalue of 28. although some problems with the disturbanceterm assumptions of 0 . in fact to test the hypothesis Ho : 'Y2 = 'Y3 = O. can be expected (see.Pagan test is an LMtesl too. Although the distribution of the tstatistics in Figure 6. the coffee prices have the hi ghest tvalues. which can be used as an indication of a possible source for heteroskedasticity. " which gives the suggested name residOl . then you can specify and estimate a testequation yourself and perform the test.) To illustrate the use of the BPtest. After clicking 'Quick' and 'Estimate Equation'.Pagan test (BPtest). . Greene (2000)). The regression of the residuals on the selected explanatory variables (or their squares) is the basis for the Breusch. But be careful with the tstatistics as their distribution is not known as well. seems to be the source of the heteroskedasticity. then. which requires.. As an asymptotic equivalent test the zeroslopes F test can be used.22%. a specific alternative hypothesis. = 'Yl + 'Y2X. In this example. We continue with the simple previous example. and selecting 'Make Residual Series . the following equation is .
009768 0.00l553 ·7.003IIJ5 0. with the pvalue in parentheses.1321).1I .01386l1 0. Next. The calculated BPtest.0397!1l 0. Error IStatistic 0.615 1.0514 0. is: nR2 = = 510 x 0.0358 0.006533 0.004243 0.021553 1844.0527 0. E. of regression Sum squared n:Jsid Log likelihood Durbin·Watson stat 0.Tests for heteroskedastieity 129 Figure 6.017344 0.On6 Mean dependent var S.006057 0. dependent var Akaike info criterion Schwarz criterion 0.077254 0. See the 'Equation Specification' window in Figure 6.004251 0.007502 0.10: Specification of tbe equation for the BPtest specified. compute the value of nR2 by using the results from the output window in Figure6.105ll93 1 . 10.2 Rsquared Adjusted Rsquared S.O.11: Regression output for computing the SPtest .0727 (0.CDm2 0.563211 2.214176 F·statistic Prob(F·stalisUc) Figure 6. Dependent Variable: RESIOO1"2 Method: Least Squares Date: 02/1BAJ3 Time: 12:09 Sample(.768204 Prob.039B60 0.942007 1.5735 0. without generating new variables beforehand: residOl7 c /pcojfee /pcojfee7 /pcojfee(.013868 7. e LPCOFFEE LPCOFFEE"2 LPCOFFEE(·1) LPCOFFEE(·1 ).djusted): 1960:04 2002:09 Included observations: 510 after adjusting endpoints Variable Coefficient Std.881553 0.95m:3 1 .I) Ipcojfee(1)7 .083910 O.
Testing the Stochastic Assumptions and Model Stability The 5% and 10% critical values of the X2 (4)distribution are: 9. has an increasing/decreasing monotonous trend. We conclude that the variance O'~. Microsoft Excel or with any other statistical computer program. For crosssection data. otherwise drop the middle observation). which are the conditions in the definition of the F distribution as summarised in Section 5. for example.. +u +· · ·+u!n 2 2 2 2 U ~n+l + u !n+2 + . Then the idea behind the test is to split tbe sample into two periods (timeseries data) or into two parts (crosssection data) and to compare the residual variances in the two subsamples (n is even. See the result of the zeroslopes F test in the window in Figure 6. Timeseries data have their fixed historical ordering. The GQtest is used to test the null hypothesis of bomoskedastic disturbances: Ho : a~i = u~ for all i. The following statistic has an F distribution because the di sturbances are normally. An economically relevant reason determines the choice of the ordering. first arrange the data in any logical order according to one of the variables. The Ho is not rejected with a probability of 13. althougb its distribution is not clear in small samples. the dependent or an explanatory variable.: HI : the variance a~. is homoskedastic.and Ftests are rather close because of the very large sample size.130 Chapter 6. The GoldfeldQuandt test Finally the GoldfeldQuandt test (the GQtest) is di scussed as an alternative beteroskedasticity test. . the quotient would be close to one. lfthe variances are constant.n 2 2 ..O'~ In) U = F 2 . independently di stributed. or: e~ = c(l) for all t.24%. the F test can be considered.11 to test the hypothesis that the disturbances are homoskedastic: Ho: c(2) = c(3) = c(4) = c(5) = O. The unobserved disturbances cannot be replaced by the residuals from the original regression as they are not independently .7794. The null hypothesis of homoskedasticity is not rejected again at the 5% or 10% level. Crosssection data are ordered according to one of the variables of the model.4877 and 7. + 2 2 "' Un (1 1) n. Alternatively. The pvalue of 0. 1: u ~ N (O . but now versus the alternative of a monotonous trend in O'~.1321 can be computed with. The probabilities of the X2 .
I n / 2) . So.1 4) and (6. according to the definition of the Fisher F distribution.12) (6. Write the two regressions. u (1 ) ~N (6.6 + u (i ).O}l)u. as: (0.a [2)u. 15) The quotient of the statistics in (6. According to (5. one has to do two additional regressions. I n/ 2) . Y(2) = X (2).Tests for heteroskedasticity distributed.5 we know that the two sums of squared residuals have the following independent X2 distributions: (6. estimate the parameters twice with OLS and compute the two residual vectors. e(1 ) and e (2). 13) The null hypothesis. Y(l ) = X (i ). the numerator and the denominator will not be independent: 131 For that reason.14) (6. you only have to divide these two values and determine the pvalue. the two residual variables have the same parameter O'~ in their distributions.6 + U(2). The sums of squared residuals are given in the output.12) in Section 5. the following F distribution: To apply the test in practice with EViews. The two subsamples can be separated more clearly by deleting more observations from the middle of the ranked . The model is estimated twice for the two subsamples. 15) has. To compute the F statistic of the GQtest. can be written as: for all observations. the sample is divided in two subsamples. This is written as follows: The two residual vectors e (l ) and e (2) are independently di stributed. Ho. U(2) ~ N (0. Under the null hypothesis. with subscripts (1) and (2) corresponding to the two subsamples.
but the parameter estimates are quite different in the two subsamples. For that reason.5 Checking the parameter st ability Finally. As an example. the null bypothesis of no heteroskedasticity is clearly rejected at the 5% level in favour of a decreasing variance of the disturbance terro. . The estimation results are not given here. This is not really a test. Period I bas 261 observations (n. This time. with the pvalue in parentbeses. The assumption has been made that the parameters of the explanatory variables are constant over time. " and mark 'Recursive Coefficients' in the obtained window. The null hypothesis. . but the pictures of the recursively estimated coefficients are informative regarding the stability of the estimates. F . and the matching 5 % critical vaJue are as foUows: e (l) e ( l ) e ( 2) e (2) = 1. no observations have been deleted from the 'middle' of the sample. If there are K coefficients to be estimated in the (3 vector. we return to the specification of the estimated model. 13). e ( l ) e (l)/ (n . This form of heteroskedasticity has not been found with the previously introduced heteroskedasti city tests. then first K + 1 observations are used to estimate (3. The date 1980(12) has been chosen to split the sample.9. That is the smallest number of observations that can be used to estimate the .K ) 1. The idea behind recursively estimating the parameters of the equation is repeatedly adding one observation and subsequently reestimating the parameters. . H o. 3773 (0.12.K ) = 0. and select 'Recursive Estimates (OLS only) . The number of observations that can be dropped from the middle is rather arbitrary and has to be done carefully because of the loss of degrees of freedom in the two regressions (6.824421 . The equation bas been reestimated for the sample period 1960( 1}. as shown in Figure 6. consider the unrestricted relati onship between the cocoa and coffee price from Section 4. In the 'Equation ' window click on ' View'. = 261) and the second period has 279 observation (n2 = 279).0047). This result is not unexpected for the reason given above and you can reproduce the results yourself by using the accompanying data set.824421 /(279 _ (2) Fo.05 (255.2245. AU the coefficients are shown in the list.060631. 6) 6) = 1. = 0. 'Stability Tests'.1980(l2) and the period 1981(1}2002(9). An impression about the stability of the parameter estimates in the sample period can be obtained by computing recursive coefficient estimates and looking at their plots.. e (2)/(n. Testing the Stochastic Assumptions and Model Stability data set. 6. is that the disturbances are homoskedastic and the alternative hypothesis. which makes the test more powerful. but you can also select just a few parameters. 273) = 1. H" is that the variance in the first period is larger than in the second period. because the period after 1980 was quieter for the coffee market than the preceding period with heavy price movements. The resulting F statistics.132 Chapter 6. 12) and (6.060631 /(261 = e .
These estimates are plotted in separate graphs. as shown in Figure 6. For example. only the coefficients C(2). It is normal that the first estimates behave in rather an unstable manner. The expectations of the CUSUM of Squares statistic run from zero at the first observation until the value . which will briefly be introduced here. Then the next observation is added to the data set and [{ + 2 observations are used to compute the second estimate of {3.12 shows that more output options concerning model stability can be chosen. for example. for example. C(3). C(4) and C(S) have been selected to present a clear Figure 6.it produces estimates from a number of subperiods. From these plots. the 'CUSUM Test' and the 'CUSUM of Squares Test' can be selected. You can rescale the figures to show the estimates for only the second half or twothirds of the sample period.[{ estimates of the vector (3. The CUSUM statisti c is based on cumulative sums of scaled recursive residuals and is plotted against time. then we have only one degree of freedom. because of the very low number of degrees of freedom at the beginning of this procedure. also produce informative plots about the stability of the structure. The drawn 'zero lines' are not default. but can be made default via the 'Options' button.12: Specification of the coefficients that will be estimated recursively parameters. 13. Johnston and DiNardo ( 1997) or Greene (2000) for more detailed theoretical infonnation about the statistics. The window in Figure 6. When the graph of the CUSUM statistics revolves around zero within its confidence bounds the null bypothesis of parameter constancy is not rejected. The last recursive estimate of {3 is equal to the OLSestimate obtained earlier for the whole sample. The CUSUM of Squares statistic is a cumulative sum of squared residuals. The background of both the statistics will not be discussed here because their interpretation is straightforward.Checking the parameter stability 133 Figure 6. you could conclude that the estimates are stable. The CUSUM statistic is plotted in EViews with 5% significance confidence bounds. These tests. in your paper. Together with the standard errors they are infonnative about the significance of the estimated parameters: for example. This process is repeated until all the n sample data have been used. note that C(2) and C(3) become significantly different from zero at the end of the sample period. as they hardly vary in the second half of the sample period. Notice that recursive least squares is not an estimation method to estimate variable parameters . (See. In this figure.) The expectations of the CUSUM statistics are zero under the null hypothesis of constant parameters. I 3. yielding n .
...~ 2 S. _ .13: Plots of recursively estimated parameters eo . 1. .... &tin • .E.. 51 $9Iifioa'lct: . ~ 0. . . _. _ ..> .r.  12r~~~' CUSLllllof SqurH .. D~ ·0.E ~: 1..• I........: .8 : " • • • ..• Figure 6. "/' ....£  Rt<:v'ItYe C(2) EstirnaH .."'..... . ...~ _ ...' . " .." ..:l 2 S. ...'" 1. . • .. D • 0. ~ .E.2~r .". v "" " 02 .  ReoIniv.....".... .. . 20 DO  CUSUM ... .....8 1.. .... ___ .._••••• _... '" .... '" .eo to ....r t·.. . .._.' ..... C(4) ·.. " ~ ...I...2 ... .. . ...' " .. "1 _ . C(3) ~IS ··· :t 2 S...20 ' . .. .14: Example of the CUSUM and CUSUM of Squares Test . I~ .:. . '" ' " ' =. .... n ~ I~ o 'V ..' . r.  . " " " " .. 7D 75808$OO~OO Figure 6..._ ~.. ...
29)... It is very irrational to suppose that an intervening period exists where the variable X . Next. X ' .... . . Suppose that the estimated parameters (32 and (3.Model estimation and presentation of estimation results 135 of one at the end of the sample period.. ) Multicollinearity might have been the reason for the abovesketched problematic estimation result. as the graph of the CUSUM of Squares statistics hits the 5% significance bound. Y. In Figure 6. That is an estimated model.. 6. where a part of an estimated model is considered: . these plots are shown for our example with the regression of cocoa prices on coffee prices. il is possible that a better result is obtained by deleting the 'significant' variable X '2 because the variable X' _1 can take over the role of X' _2. does not influence the dependent variable Yi... the 'estimation and test' stage will be fini shed and the results have to be written and di scussed in a research paper. High corre lation between the explanatory variables X" X' _ 1 and X' _2 causes multicollinearity. .. 14. The influence of X" X ' _1 and X' .. specified with a sufficient number of lags for the explanatory and dependent variables. 14.6 Model estimation and presentation of estimation results In this section. The CUSUM of Squares statistics are also plotted with 5% confidence bounds. = (31 + (32X.. Therefore. After the research problem has been introduced. all three variab les behave in a similar way. the data have been described and analysed as indicated in Case I. After a number of steps. First find a 'general model' like mode l (4.. look at the following situation... Of course. such that residual autocorrelation is absent..... but that is not correct because of economic reasons (see Remark 1.2 + .2 on Y. under the null hypothesis of constant coefficients and variance..1 ). (This subj ect will be discussed in the next chapter.. This is well demonstrated by    .. you next need to estimate an econometric model... In Figure 6. we see that the price explosion of the commodity prices at the e nd of the I 970s caused some instability.. It is clarifying for the reader to show some of the striking steps in the estimation procedure..1 + (3. you may consider removing the variable X' _ 1 from the mode l for stati stical reasons. In Section 4. some practical aspects of the estimati on procedure are discussed and suggestions provided for correctly publishing estimation results in an article.9.. + (33 X. then a good dec ision had been taken by deleting X' _2 from the model. cannot clearly be discriminated. try to find a specific model by imposing zero and other restrictions on parameters.... do not include all the steps that have been done. . as proposed in Case 2 in the next section. Next. only some of them will be interesting. U reestimation of the model without X' _2 still results in significant parameter estimates and absence of residual autocorrelation. are significantly different from zero whereas (33 is not. For example.. the idea of modelling from general to specific was introduced as a structura l method of modelling timeseries data.
2 (0. .39) 0.23(0.. ·. as described above.19(0.07 1) .016 0. Model 4 is the 's pecific model'.I .328 p eol 1.83(0.070 (0..23) 50.045) • •• (0 .020) (0. Model I is the 'general mode l'.14(0.99 R2 ·.3 (0.121 (0.1: Schematic example of a regression table inserting a regression table in the paper. and Model 2 (no results have been given in thi s example) and Model 3 represent two interesting models to show the reader. which is easier to survey than listing a number of estimated equations or regression outputs in the paper.04(0.75) 47. Notice the cbange in the significance of the parameters by deleting P" o£ from the model.37) 49.041 ) p eo l 1.00) 5.462 (0.044 ) .071 ) (0 .07(0. The asymptotic standard errors of the estimated parameters have been given in parentheses.103 (0.041 ) 0.044 ) (0 .00) 12. ·.069) • (0 .062 (0.Ql8 (0.328 Pf~ ~O.0.043) • •• •• (0.00) 14.328 ••• (0. An example is given in Table 6.044) 1. In the first column in Table 6. .114 • 0.I · (0 ..99 3.0..338 1.044) 1.0..467 0.121 p eo! I.28) 0.1..044) 1. which must have been caused by the multicollinearity problems.136 Chapter 6 .044) 0.36(0.99 Table 6.044) • •• "peol I p eoc I.0.019) 0. . Testing the Stochastic Assumptions and Model Stabi lity Variables Dependent variable: Pt'C Explanatory variables: Model l Model 2 Model 3 Model 4 Constant p eo! I 0.119 (0.115 1. 4. (0.1 aU the explanatory variab les used in tbe general model are incl uded.0. The prices are in logs.07 1) · .3 0.106 0.14 (0. ·.020) (0.069) • •• •• .04 1) 0.79(0. 73) 0.110 Sample period: 1960( I }2oo2(9) BG (3) JB White 1.466 0.030 (0.
00).99.))  X' 2+ !h ( . probabl y not everything will be correct as only the OLS estimator has so far been used .16) is a standard method of writing an estimated model. the JB statistic.e(i3'») Both equations are exact equati ons.044) ' 3 R2 : 0.370). = {3. The consequences of th at distinction for esti mating the parameters of the model will be considered in Case 6. (. as will be seen in Chapter 12..07 1) '.e(i3. The n the null hypothesis Ho : {3k = c.1 (0.))  X '2+ {33 (.28). In the e mpirical literature you can see that standard errors or tvalues are placed under the estimated coefficients in pare ntheses.2002(9).14(0. 6 . If possible. write the standard errors rather than the t val ues under the coefficients as tvalues onl y test Ho : (3k = o. 17) (.020) + 0. BG(3) : 3.(i3K ))  X'K + e. (6.. the equation has not been written correctl y.I (0. Generally.044) '. The notation of eq uation (6..328 p coc_ 0.07(0. give pvalues in parentheses behind the values of these stati sti cs. I 6) Y. instead of zero only. e(i3K ))  X'K (6. White: 12.) values. and possibl y (if allowed) the DW va lue. The presentation of equation (6. JB : 49. If a hat C) has not been placed above the dependent variable or the residual variable et has not been included...Case 2: an econometric model for timeseries data 137 The fi nal estimation result will be repeated in the text and discussed as your proposed econometric mode l. Not onl y has the estimati on stage to be done properly.e(i33))  X'3 + . 0. can be tested directl y by the reader of the paper for any value of c. In general.e(/3. 17) is sometimes useful to show the estimation result fo r a specific situation. ))  + fJ. From now on. (i33))  X'3 + .. + {3K (. are wrinen. The result has to be written in a mathematically and stati sticall y correct way and concerns a couple of simple matters which are useful to restate. a suffic ient number of tools have been in trod uced and di scussed to start your firs t complete paper.7 Case 2: an econometric model for timeseries data At thi s stage of the course. + {3K ( .016 (0.. '. e(i3..2 (0. Sampl e period : 1960( 1).041)' (0.  + f32 (. The correct way to present an estimation result is shown in the following two equations: ~ Yi = fh or (. where . some BG(. The fi nal model for the cocoa prices can be written as follows (with variab les in logs): pc = .04(0.121 p co! _ (0.041 ) 0. Under the estimated equation other statistics such as R2 or R2. No di stinction will be made between exogenous and endogenous explanatory variables at thi s stage of the course. models will be estimated and assumpti ons will be tested in a 'correct' way.462 p coc + 0.103 peo! + 1. standard errors are better placed under tbe estimated coefficients. the presentation of the estimati on results also has to be correct. etc. However.110 p c::c.
You can do both and decide later on which model describes the economy in the best way. Make a decision about modelling the original variables. In Case I. After Chapter 8 has been studied. At thi s stage. deflated variables or the logs of the variables (parameters are constant elasticities). Reconsider the results of that analysis and perform the estimation procedure in the correct order: model from general to specific according to the reference points given below. Preparations have been done in Case I with the analysis of the timeseries data. you can only report the presence of heteroskedasticity in case the null hypothesis of homoskedastic disturbances has been rejected. which gives you information about the stability of the elasticity in the sample period.Pagan). as far as possible. When not modelling the logs of the variables you can make a graph showing an interesting elasticity. A short introduction to the three subjects has been given in Chapter 2. Reconsider your text and decide which results have to be described in the paper and write. as was demonstrated in Section 3. but sometimes a null hypothesis concerning first differenced variables can be relevant. or a model for the money demand in the USA. Restrictions can be zero restrictions (excl usion of variables). The same questions can be answered if you use your own data. a complete paper. The first two sections of your paper have been written in Case I.and F tests to test restrictions on parameters. If you are content with the results. you are working in a structured way to find a good model. followed by some directions for possible models that can be estimated with the data. Impose the restrictions. • Open in a corner of your screen the correlogram of the residuals 'RESID' to have 'continuous' infonnation about the residual behaviour with respect to autocorrela· tion. Test the normality assumption and the null hypothesis of homo skedastic disturbances for your final model (White. GoldfeldQuandt or Breusch. Estimate a dynamic equation. Evaluate the estimation result by checking the stability of the parameter estimates by computing recursive parameter estimates and CUSUM statistics. In doing so. A good model is a parsimonious model that describes the economy according to the economic theory and that does not reject the stochastic assumptions. then also compute some interval estimates for parameters of interesting variables. it will be possi ble to tackle thi s problem in Case 4. Both the estimation results will be compared and evaluated in Case 6. Choose your final model and justify your choice.138 Chapter 6. Go on until a 'general' model without residual serial correlation has been found. Specify in the 'Equation Specification' window an equation with an equal number of lags for all the variables and estimate the parameters with OLS . then respecify the equation with more lags for all the variables. If autocorrelation is indicated.4. Testing the Stochastic Assumptions and Model Stability you will reestimate your model with the 2SLS estimator for obtaining consistent estimation results. • • • • • • • . reestimate the model and check the residual autocorrelation function . Look at the correlogram of 'Resid' and the plot of the residuals. Start reducing the model by using t. a model explaining the demand for vehicles in the USA or the retail gas price. you have chosen one of the three offered 'research projects' with timeseries data: one of the macroeconomic models.
4. In Section 7. This section also demonstrates why the procedure of modelling 'from general to specific' is recommended to obtain a correctly .. in Section 7.. A number of different subjects then follow. and are sometimes employed to eliminate outliers.5. 7.2.. attention wi ll be paid to prediction issues. It is a theoreticaloriented section. the consequences of specification errors are discussed. the use of qualitative explanatory variables is illustrated. effects of multicollinearity (correlation between explanatory variables) on estimation results are demonstrated.3.2 Specification errors In this section. the basic course is completed by di scussing a number of subjects related to the reducedform model.1 Introduction In this last chapter of Part n. but with clear practical consequences. We start with studying the consequences of specification errors in the linear model. It will become clear why it is a good idea to use Hendry's approach of 'general to specific modelling' to find a satisfactory econometric model... The chapter covers a variety of different subjects. This approach to obtain a model by working in a structured way is obviously better than some ' trial and error' or datamining procedure.Chapter 7 A Collection of Topics the Linear Model &. in Section 7. Qualitative explanatory variables can be used to specify a deterministic trend.. and in Section 7. 7. seasonality and different regimes.
The expressions are: (7. which are useful expressions for examining the propenies of the OLS estimator in both the situations. These expressions do not have any practical relevancy but are useful for the theoretical   analysis.3) It is possible to derive individual expressions for the subvectors. 1) (7. Two types of specification errors can be considered. for example. He ndry (1995) or Charemza and Deadman ( 1992) for discussions on thi s topic).81 and.2) has been estimated with OLS.140 Chapter 7. The situation that the first model is true and the second model has been estimated can be compared with the start of 'general to specific modelling'. To examine the effects of these two types of errors.5) will be introduced.2) is 'true' but model (7.3).2) bas been estimated. First.81 and +U (7. A Collection ofTopics Around the Linear Model specified parsimonious model (see also.2) in the partioned form: Then the OLS estimator for all the parameters is wrinen in partioned form as: (7. 1) has been estimated.1) is 'true' but model (7.4) and (7.2) Xl is an (n x K .5) Proof If model (7.6) . The consequences of these two types of specification errors are elaborated in this section. model (7. ) matrix and X 2 is an (n x K 2 ) matrix. then the estimated model is written as: (7.82 in equation (7.4) (7. The following two situations are analysed: • • model (7. Write model (7. the following two models are considered: y = X l . Before these two situations are di scussed the algebraic expressions (7. it is possible to have a model that is a number of variables short. and secondly it is possible that too many variab les have been specified in the model.
7) on the left: ~ X~ Mx.81 can be derived in a similar way: ~ . X2.8.4) for the subvector.Mx . y = X~ MX. Provided that (X . = 0 . 1) has been estimated.5) for the subvector . Using these (7. X X.82 + X~e X~Mx. not a recommended procedure First we look at the situation that model (7 .X.82 + e . : ~ the expression (7.Xl f l X . Modelling from small to large. y = X~ M X. Multiply both sides of the equation with X~ o. on the left: ~ 141 M x. In other words.3). (X .6) with M x . Y = M X.Specification errors Define M x .5) are the subvectors of expression (7. ~ Remember the property that M x .X 2. the expression (7. = ( '. = In .X2. the expressions (7.82' ~ where the property is used that X~ and e are orthogonal too.82 is obtained: ~ By defining M x. X I )1 .X. X I .2) is true but model (7.) e = Ine . (X. and observe that: Mx .4) and (7 .) . and multiply the estimated equation (7. e .l X. the model has been specified too small. X . M x X 2) is a nonsingular matrix. X\ e = properties gives the equation: M x. X.0 = e. y· So. X 2.Y = M X. + MX. e = ( In .82 + Mx. It is easy to demonstrate that the OLS .8. Because the residuals and explanatory variables are orthogonal. M X.
X .8) separately. 1) is true and model (7. ).2) has been estimated.6.6.X.142 Chapter 7.' X'. look at the subvector .4) and (7 . + (X.6. Therefore.) .) . Modelling from large to small. . ). The parameter vector {31 is estimated as: fj.X . The expectation of. it is not a recommended procedure.' X'IX2. (X . + (X.X. = + X 2.8) when model (7. but that test will not give specific information about the individual specified variables.X 2i32 + (X.62 + u ) . No conclusion can be drawn from the OLS estimation result. = + (X. f ' X.6. An LM test can be applied.62 X 'I .62 + (X '.X .).X2. The residuals will be highly autocorrelated as the variables from X 2 are included in the disturbance term.) .' . is (the explanatory variables are assumed to be exogenous):  Efj. Eu i .X .5) it is easy to prove that the parameter vector is unbiased and consistently estimated in a model which has been specified with too many variables.6.u.y = (X.I X '. :  This implies (with explanatory variables stiU assumed to be exogenous): . To show (7.' X. = . A Collection of Topics Around the Linear Model estimator for the parameter vector {31 is biased and inconsistent. It can be proved that: (7. a recommended procedure With the notational result from (7. This implies that the Wald t.6.8) we consider the two subvectors in (7. First. = (X.61 · This proves that the OLS estimator is biased and inconsistent as the bias does not reduce to zero when the sample size increases.' X.and F tests cannot be used to find a correct model.X .
for example. In that situation we have: It can be shown that the difference between the covariance matrix (]~ (Xi M x. The estimates of the parameters f31 in model (7. (Xlf3l + u) = (X~ MX .and F tests may be used to find a correct model. the terminology used in this section relates more to timeseries models.2) are less efficient when too many variables have been specified. There is no reason for residual autocorrelation. X2 ) . XI ) .I of the correct model is a positive definite matrix.).) The econometric model can be used for prediction purposes when the assumptions which have been made in the sample period remain unchanged in the prediction period. . Clemens and Hendry (1998) for a detailed analysis of this subjec!. lf lagged variables have been included in the model. This .1 X~M x. x l ) vector y f with predictions of the dependent variable.Prediction ~ 143 Next. so the Wald t.K 2) instead of (n . This result proves that in a model that has been specified with too many variables the OLS estimator remains unbiased and consistent.1 X~ Mx . 7.3 Prediction Assumptions One of the things that can be done with an estimated model is forecasting or policy simulation. This demonstrates that the 'general to specific modelling' strategy is a good one. U . X2) . This implies (again the explanatory variables are assumed to be exogenous): ~ Ef32 = O. J V = 0 = 0 + (X~ MX. the computation and properties of postsample predictions are considered by using the estimated linear model.1 X~ Mx. In general notation.X2) . The variance (]~ has been biasedly estimated because the wrong number of degrees of freedom has been used (n . Although it makes no difference in computing predictions with an estimated model for timeseries data or with a model for crosssection data.X2 ) . u .I and the actual covariance matrix (]~ (X'X ).y = (X~MX. + (X~ MX. (See.X2) .1 X~Mx.K . [n this section. the results obtained above are asymptotically valid. the model will be used to compute an (n . look at the second subvector f32: 32 = (X~MX .K. X I f3.1 X~Mx . One may expect that variables which do not belong to the model will have parameter estimates and tvalues that are close to zero.
The first aspect can be summarised by the following statement for timeseries data. • It is clear that the predictions of YI will be computed for given XI by using the OLS estimates of the parameters from the sample period. A Collection of Topics Around the Linear Model suggests that the (nl x K ) matrix X I with future observations of the explanatory variables is known. However.144 Chapter 7. There are two possibilities to obtain these observations: • If the model is used for a policy simulation. then the researcher provides the data on the explanatory variables. but a prediction of a stochastic variable. If the assumption is realistic that the economy. Therefore. The quality of the result is clearer when the predictions are published with so me confidence interva1. Then the predictor Y is defined as: I First. it is necessary to check which additional assumptions have to be made to know that I this is a correct procedure. a new terminology is introduced: prediction intervals.O"~q future data is: Both models are identical and the disturbances U and U I are independent. when reaHsations . Prediction intervals give information about the accuracy of the computed forecasts a priori. The model for the n sample data is: Y = Xj3 + u . for example. and so the structure of the model. Secondly. then forecasts of future values of the explanatory variables have to be computed too. u! Prediction intervals The presentation of only predicted values of the endogenous variable gives too little information. that is not possible. have to be valid for the nwdel in the prediction period too. the properties of the predictor Y have to be studied. and can be used a posteriori. and identical statistical properties are assumed in the sample period and in the future period. will not change in the near future it makes sense to use the estimated model for forecasting. You can estimate a parameter and compute. a 95% confidence interval around the estimate. The assumptions that have been made about the model for the sample period. A forecast is not an estimate of a parameter. The assumptions are mathematically simply formulated as fo llows. The same parameter vector {3 and disturbance variance have been specified. this can be done by using ARIMAmodels for these variables (see Chapter 14). with and the model for the n l U ~ N (O . If the model is used for computing postsample forecasts of the endogenous variable.
10) + a~In.' X ' .S ) +ilf) = Xff3 .Xf (X'X ). a~ (X' X).Uf) = .S)) = + Var (uf ) var( .X i) = X f . Because the two components are independent we get the following expression for the vanance: Var(f ) = Var (Xf (f3 .' X' u. which are normally distributed variables. Next determine the expectation and variance of f.Xff3 = +0 o. Xi = + Var (uf) . x nIl covariance matrix is zero: Cov (Xf (f3  S) . a~ ( In.X fS) + Var (uf) = . it is shown that the two components of f are independent: S = f3 + (X'X ). E( f ) = E(Xf (f3 . Cov (u .' . The prediction error f is the basis for the derivation of the prediction interval. (. First.l X i) .X f' Var (S) . Uf) = 0.• Prediction 145 of the forecasts are known.9) = X f (f3  S) +ilf S) ~ The prediction error consists oftwo components: the sampling error (f3 and the future di sturbances U f. The distribution of the prediction error will be used for the computation of prediction intervals.' X' u = X f (f3  S) = . asCov (u . + X j{X'X ). Define the (nl x 1) vector f with prediction errors as: Rewrite the error as follows: f = X f f3 + uf  ~ X ff3 (7. The following (n. Next we have to detennine the expectation and variance of f . Because f is a linear expression in f3 and uf. to test the assumption that the structure of the model has not changed in the prediction period. f is normally distributed too. uf) = 0.Xj{X'X). (7.
As before with the distribution of the OLS estimator. this will be done by using the definitions of the t. provided that the prediction error f and the residual vector e are independently distributed. 2 (7. A Collection of Topics Around the linear Model This implies the following normal distribution for the prediction error f : f ~ N (O.))..a~ (In. Var(u) · M:" +Cov(u f' u ) · M:" = . (7.. + Xf (X'X) . This property is shown by using equation (7. (7.I f 2 au ~x(n. Again. the unknown has to be eliminated to make it possible to compute a prediction interval. .13) This is possible.K) a~ ~ au 2 x2 (n  K ).Xf (X' X).).' X' u and The covariance between f and e can be computed as: Cov(f .IX'. is eliminated by using the X2distribution of the sum of squared (n .and F distributions. If an individual forecast has been computed.146 Chapter 7.I X'f ) .11) expressed in one prediction • error.11) it follows that the X2distribution for the quadratic form of the prediction error IS: f' (I nl + X f (X'X) .x tCX' X ). with laggeddependent variables the results are asymptotic in this section. The explanatory variables are assumed to be exogenous. a 95% prediction interval for the forecast by using the tdistribution.X f (X'X ). divide (7.13) to get.' X'M~ (X'X).I X. by (7.9): f = Xf (f3f3) + uf + uf = .' ·0 ' = 0. a. From (7. e) = .' X'· Var(u)· M :" = a~ Xf = a~Xf (X'X).11 ) Notice that all the assumptions of the linear model have been used.12) The disturbance variance residuals again: a. for example.
+1 that is distributed as: The (K x l ) vector x. Divide this distribution by the di stribution 7. 13). se (fn+!)  < Yn+! < Yn+.to.15) is not plotted by EViews.+! has been predicted by Y.1 Uu . The prediction interval is an interval for the variable Y" +.+! . A confidence interval is computed for a parameter.' Xn+l) I n+! In+1 ~N(O. l) . which is a stochastic variable.K ) .2 later).K ) ·se (fn+.2 .. to get bounds with the following F statistics: (f' (In..Q25 (n . The standard errors (7. (7.!J F (n" n . a prediction interval I for the vector YI can be determined by dividing both X2 distributions (7. 14) Then the 95% prediction interval for Yn+ 1 can be written as Yn+1 . 10 case an (n1 x l ) vector Y of predictions bas been computed..2 I. 12) and (7.K).) for notational convenience: (7.025 (n . we have a prediction error 1.   (7. + 2 · se (fn+!)  is plotted (see Figure 7. se(fn+I ). but instead the interval Y n+1 ...13 to get a Student tdistributed statistic: UuJ(l + x~+! (X'X r ~t(n .16) These F values can also be used in retrospect when the observations Yn +i have been realised and the prediction errors f can be computed.K ).+! is a vector with future values of the explanatory variables. + XI (X'X r ' XI) f) in. However the prediction interval (7.) < Yn+! < Yn+1 +to. Next standardise I n+1: UuJ (1 + X~+l (X'X).Prediction 147 If Y. Then they are used as test stati stics to test . 14) are computed in EViews.15) Notice once more that this interval is not a confidence interval. . 1 Xn+ 1) Define the standard error of In+! as se (fn+.
In EViews an alternative form of this test is computed that is called the Chow forecast test (see the end of this section for a discussion of this test). Measures of accuracy of forecasts Two things can be done when the observations of the prediction period have been realised. For the forecast Y I a forecasting interval has been determined and for E (y I ) a confidence interval can be computed. Various measures have heen proposed for assessing the predictive accuracy of forecasting models. the forecast quality can already be analysed when the model has been estimated for a subsample and has been used to forecast the last observations in the sample. the mean absolute error (MAE) and the Theil inequality . but also the forecast quality of the model can be analysed with measures of accuracy. A forecast test can be computed to test that the structure of the model has not been changed. when estimating the mean sales of any product in the next period with a 95% confidence interval. The distribution ofyI has a smaller variance than the distribution of the prediction error f. 17) = a~XI (X'Xr ' Xj .the root mean squared error (RMSE).Xj (7. The first two moments of YI are: E (Y/ ) = E (X/i3) = X Jl3 Var (Y/ ) = Var (X/i3) = X I . In practice. The predictor YI can be shown to be an unbiased estirnatorof E (y I )' We know already that it is an unbiased predictor (the new terminology) of YI because the expectation of the prediction error is zero. Compare this expression (7. which is a parameter. Estimation of the mean of the dependent variable in the prediction period Forecasting nl values of the stochastic variable Y with YI is equivalent to estimating the mean E (y I ). For example. Var (i3) .10)! So the distribution ofYI IS: This distribution can be used. We look at a number of wellknown inequality measures that are computed in EViews . for example.148 Chapter 7. The Fvalues are often computed in software packages. A Collection of Topics Around the Linear Model the null hypothesis that the model structure has not heen changed in the forecasting period. the forecasting interval is mainly used. The test is called a forecast test. 17) with the variance off in (7.
THEIL. The U coefficient is also called the standardised root mean squared forecasting error (SRMSE). in: _ Yi . The square of the numerator can be written as: Squaring the terms in brackets results.Y". Yi The value of the THEILcoefficient lies between zero and one. They can be used as relative measures to compare forecasts produced by different models for the same series.s y ) .= )2 +(SYSy) 2+2(1r)sy. Yi . after some algebra. whereas a value of zero implies a perfect forecast. . Y Y.Prediction coefficient (U): 149 RMSE = 1 n+n l MAE= L nl L=n+l U= These measures will reflect the model's forecast accuracy: large values indicate a poor forecasting performance. EViews computes two more statistics which are scale invariant: the mean absolute percentage error (MA PE) and another Theil inequality coefficient. )2 = ( ((_ . These measures are not bounded and the values of the RMSE and MAE depend on the scale of the dependent variable. .Y".
U a lagged dependent variable has been specified you have to choose between a dynamic or a static simulation. One option concerns the method of forecasting: indicate whether static or dynamic predictions have to be calculated. representing measures of unequal central tendencies.... Of course both procedures are identical when the model is a static model. . + fJ.... . · ... and imperfect covariation. Yn + J = /3.........:  static: etc.... a variance component and a covariance component...r ) By .. ....... Click on 'Forecast' and various options can be selected in the 'Forecast' window (see Figure 7. the static procedure computes oneperiod ahead forecasts. assume the following estimated model : . = /3..+ Y . Xn+2+/33Yn+'...' respectively.." ~ So the square of U or the square of the RMSE is decomposed into three components: a bias component... . where Yn is the last observation of Y in the sample period.... + /33Yt .." ~.. and r is the correlation coefficient between Yt and VI. + /32 X n+l + /331'. 2 (1 ... X n+2 + /33Yn+l. + fJ... A Collection ofTopics Around the Linear Model where By and By are the (biased) standard d!viations of Yt and VI ... First..150 Chapter 7.... unequal variation. Predictions can be calculated in tbe following two ways.......... Yn +2 can be predicted in two different ways... . . by using either Yn+l (if known) or Yn+. .: . or dynamic: Yn+2 =/3.. X. In EViews the three components are defined as: e<Y!Jariance : ::7::''''.....B y nE" (Yt _ 9/.....1 later). .. For example... .. Y. Further.. +fJ.. In a dynamic model you have to choose whether the forecasted or the observed (when they are known) values will be used for the computation of multiperiod ahead forecasts. predict Yn +.... n 2 = /3... . In the latter situation.)2 t=n+ l Forecasting with EViews In the 'Equation' window you will find the button 'Forecast'.
you can miss the cbange and are one period bebind from then on (see Figure 7. It is more interesting to plot the actual and predicted prices together with the ta/ 2 (n ..96 * sei . Indicate whether they should be computed and saved in the worlcfile as a new series.Prediction 151 A problem with simulations is the prediction of turning points of the variables.K ) x the standard errors. You can also mark the 'Forecast evaluation' and the prediction period can be defined.4 later).2.15). The relationship between cocoa and coffee prices has been reestimated for the period I 96O(01}2ool (05) Next p[oc has been forecasted for the period 200 I (06}2002(09). This can be done in the following manner. for example. together with ta / 2 (n .025 (488) = 1.96).1. Next. genr se. for example in the following way: • genr se_u = lpcocoai + 1. inequality coefficients for the observed and the predicted variables are computed. This is a picture obtained by plotting both the actual and forecasted variables on one graph. Give a name for the standard errors if you wish to see them in a plot of the predictions. A second option concerns the computation of standard errors of the predictions. you have to find informative leading indicators that can be included in the model (see. The calculated standard errors are the square roots of the diagonal elements of the matrix (7. Chapter 9 in Clemens and Hendry (1998». You have to indicate whether the predictions and standard errors are to be plotted in a graph but such a grapb is not very informative. This subject will not be discussed in this introductory course in econometrics. and a plot of the forecasts and ±2 x the standard errors is made.l = lpcocoai . This is specified in the 'Forecast' window shown in Figure 7. Many studies can be found on this topic in the econometrics literature. You can change that name to a more suitable one if you want. You generate two more series.14): With these standard errors it is possible to compute the prediction interval(s) (7. some examples of EYiews output are given.96 * sei . (to. The graph in Figure 7. Standard forecasts are stored in a variable that talces the name of the observed variable plus an ' f' added at the end. 1. A more informative plot is discussed later on in this section. This result is shown in Figure 7.10): as written in (7. To tackle this problem. Indicate whether the simulation is dynamic or static when a lagged dependent variable has been included in the model.K) x the standard errors in one figure.2 is not really very informative. When the model has no information about turning points. Predictions are computed. The standard errors are stored in a new series in the workfile (bere the name sei has been chosen). .
083go2 0... 3.8 3.st : LPC OCOAF Act.50<1822 0 .eG5<187 0. c.416280 g. I: LPCOCOA Forecast ._ _ _ _ _ _ _ _ _ _ _ _.138225 0. Proponion Vari3nce Proportion Covariance Proponion 0..152 Chapter 7.1: The 'Forecast' window 4.2.615718 0..ample : 2001 :08 2002:0g Included observ3lions: 1G 4 ..2<18057 2001 :10 2002:01 2002:04 2002:07 LPCOCOAf Figure 7. Percent Error The~ Inequalily C ffic~nI .4 Root Mil" Squared Error Mean Jlbsolute Error Mean JIbs ... A Collection of Topics Around the Linear Model Figure 7...2: Forecast output in the 'Equation' window . Bi. F.0 • .
. i"'"' 01 :07 01:1 0 02:01 02:04 02:07 Figure 7.. .. as outlined before.d L ... 1 i J_. as given in Figure 7..2 4..9 3.. I I .1"""1' IltJllIIIII "".. .3: Actual and forecasted cocoa prices with ±1..LPCOCOAF 4. seJI and seJ to plot them all together...4 4.4. This is not strange. ' 3.04 02:07 Figure 7. .. I ". .V . . v' / ". I I I 'l .96 x sef and make a group of /pcocoa.1 4.. .6 3.4 I ... !..Pred iction 4. /pcocoaj.. price iPredict~d coco~ price ' I i ! / j/ i . it concerns only a test for the hypothesi s that the coffee market has influence on the price formation of the cocoa market. . etc.1111111(11" ~I·':'·~~ . as the relationship does not concern a real economic model for the cocoa market..6 4.. I I 3. .. From Figure 7. 4. The realisations from 200 I (06}2002(0 I) are within the bounds.... 7+_~_~_r_~~~~ .".2 4.. "1ir:::1! 4.. • . . . The problem with forecasting turning points..B 3.8 . .1101."I I I . . I I . ..3 4. Change label names.. but only the four shortrun forecasts up to 2001(09) look acceptable... to get a nice plot of the actual (realised) and dynamic predictions.3 it is clear that the relationship between the cocoa and coffee prices is not very accurate in forecasting the cocoa prices... is clearly seen in this fi gure.0 3. r....3. 'T .. ! . .. / 01:01 01:04 01:07 01:10 02:01 02.4: In (PtO' ) Static (oneperiod ahead) forecasting of . The result of a static simulation (oneperiod ahead forecasting) is seen in Figure 7. .5 LPCOCOA .4 4.Cocoa .0 153 ~.1 1 . ! I I ..6 . rl' '' '. .96 x jE ..2 01 :04 1 _ I .l 1. . . \ 3 ..I .
10%.6. • (7. 'realised' observations in the prediction period. The number of degrees of freedom is d! = n . ) In.) observations and obtain the restricted sum of squared residuals Sr = e~er. it is not necessary to compute the regressions yourself. It is a sirrular test to the Ftest (7 . the breakpoint test is described in Section 7. The number of restrictions is 9 = n.e\e.. Both tests can only be used expost. observations for the prediction period. The null hypothesis implies that the n. The Chow forecast test yields the output partiaJJy reproduced in Figure 7. observations. ' Stability Tests'. A second Chow test is the Chow breakpoint test.eJ/ (n . 19) is found in the 'Equation' window under 'View' . A Collection ofTopics Around the Linear Model The Chow forecast test Besides the forecast test used to test the hypothesis of no structural change in the prediction period. We have n observations in the sample period and n. With the forecast test. The start of the forecasting period 2001(06) has been used to test the null hypothesis that the assumptions made in the sample period before June 2001 are still valid in the forecasting period after June 2001. Compute the F statistic (7 .154 Chapter 7. e'. So we have at our disposal a complete data set of n + n. as described above. which is divided into n observations to be used for estimation and nl observations to be used for testing.K ) (7. the null hypothesis is tested that the vector {3 is identical both outside and within the sample period. as in EViews the test (7. . 18) that follows an F distribution under the null hypothesis of parameter constancy: (e~er F= . you can use the Chow forecast test that has been programmed in EViews. See the EViews User Guide for an explanation of the 'Loglikelihood ratio test' that is also reported in the output.19) However. The idea behind the Chow forecast test is as follows: do a restricted and an unrestricted regression and compute the wellknown F statistic: F = (SnS)lg Si d! ' in the following way. with a pvalue of 9.5.16). forecasts have been generated by the same structure. Next estimate the parameters of the same model with all (n + n. the Chow test is considered as a test for parameter constancy but it is also a test of predictive accuracy. In fact. The null hypothesis is not rejected at a 5% significance level. after which the sampling distribution of the prediction errors is analysed under the null hypothesis of parameter constancy.K.18) • • Estimate the parameters of the model for the sample with n observations.5 for the opening window). and obtain the unrestricted sum of squared residuals S = e\ e. 'Chow Forecast Test' (see Figure 7. . which tests a null hypothesis about structural breaks in the sample period. that is computed by estimating the parameters for the sample period with n observations and predicting the n.
0000 0.1m66 6.041524 0. For this introductory course it is sufficient to demonstrate the computational problem by giving some examples of the consequences_The explanatory variables are called multicollinear .1Jl6572 1.462!1l8 0.6: Output from the Chow forecast test 7.0B4852 29. Multicollinearity is a data problem.003190 0.8751 0.307543 O.044938 0. Error IStatistic Prob.04«75 0. C LPCOFFEE LPCOFFEE(·l) LPCOCOA(l ) LPCOCOA(2) LPCOCOA(3) Rsquared Rsquared 0.041074 0. it has nothing to do with the violation of statistical assumptions.4 Multicollinearity One more problem that you encounter when estimating the parameters of the linear model is the multicollinearity problem.157226 2.0103 Mean dependent var S.03 Time: 12:12 Sample: 1960:04 2001 :05 Included observations: 494 Variable Coe1ficient Std.0000 0.123:173 O.071222 0.S: Starting the Chow forecast test T8s1 Equation: Dependent Variable: LPCOCOA Method: Least Squares Date: 02/25.0202e9 0.Mu lticollinearity 155 Figure 7.114484 0.0376 0.996407 2. Figure 7.574135 0. Multicollinearity influences the accuracy of the parameter estimates. Problems caused by multicollinearity are of a computational nature.0029 0.467972 2.D.
156
Chapter 7. A Collection of Topics Around the Linear Model
when they are highJy correlated. In fact, they are always correlated, but serious correlation might give problems. The following problems will be highlighted. • The parameter estimates become inaccurate. The covariance matrix of the OLS
estimator is:
Var
(3) = (1~ (X'X ) l .
If the correlation between some variables goes to I, then the matrix X 'X is nearly singu lar and the determinant, det(X ' X ), reduces to zero. The inverse matrix (X 'X r ' can be written as:
•
•
where the last component is the adjugate or adjoint matrix . So the elements of the inverse matrix will increase as the correlation between explanatory variables increases, implying large standard errors of the parameter estimates. Large standard errors result in small tvalues; so be careful with the exclusion of variables because of low tvalues only! Serious bias problems can be expected when the deleted variables belong to the model. When that is true, correlation exists between explanatory variables in the systematic part of the model and the disturbance term that contains the omitted variables. The sample bias can become so large that the estimated coefficients get the incorrect Sign.
In the second edition of Johnston ( 1972: pp. I60163), an example is found that very clearly clarifies the consequences of this problem. In the later editions (third and fourth) of Johnston this example is no longer found. The example starts with a model in deviation form (x" = X"  X i):
y, =
and suppose:
/3,X'2 + f33x'3 + (Ut  u),
and state for reasons of simplicity:
""" X t2  """ X t3 L 2 _ L 2 
1,
r23
L V,
= 0, and
L V,X'2
Xt3
= O.
:
This makes the correlation coefficient
between Xt2 and
equal to a
Multicollinearity Compute (X'X ) and its inverse matrix (X'X)l :
157
(X'X )l =
1
1 (1 a) . a a
2
1
Next compute the variancecovariance matrix of the OLS estimator:
The last expression sbows that the variances of the estimates go to infinity when a > 1. It can be proved that a~ is not affected by the multicollinearity. Compute the sample bias for the parameter estimates of this model :
fj =
f3
+ (X'X )l X'u,
!Ja
Continue for !h only:

1J2 = {32
=
=
+ 1 ~ a 2 (L X'2 U, 1
a
i32 +
fh +
fh +
~ a 2 (L X'2 U, 
L X'3U,) a L (ax'2 + v,)U,)
a2
02
1 1 _ 02
L
Xt2 U t

1_
L
a
Xt2Ut 
1_
(}'2
E
VtUt
=
L
X t2 U t 
1::0'2
L
VtUt·
Suppose that {32
> 0, L v,u, > 0, then:
0 "2 1 0
a
~
L. v, u, >
00
if 0 > 1.
So, it is quite possible that i32 < 0 for some 0 as the last term will dominate the other two terms. The same can be demonstrated for {33 ·

158
Chapter 7. A Collection of Topics Around the linear Model
If, for example,
f32 > 0, and ~ v,u, < 0 then:
7
1
" 2 = ,," L.. v , ti, + +00 if " 
'"
+ 1.
Then the estimate f32 gets higb pnsitive values, which is unrealistic from an economic pnint
~
of view.
Two questions remain: how is multicollinearity detected? and what can we do about it? To detect multicollinearity problems in practice you can look at the following symptoms. • • • • High correlation coefficients in the correlation matrix of the variables that is computed as part of the data analysis. Unexpectedly large standard errors and so low tvalues of the estimated parameters. Wrong signs or implausible magnitudes of parameter estimates for economic reasons. Very unstable parameter estimates. Compute recursive coefficients to observe this pnssible phenomenon. (See, for example, Greene (2000) p.257 for an empirical example.) One can use the included data file in Greene to compute recursive estimates for observing the effect of multicollinearity.
What can be done when the estimation results have been influenced by multicollinearity? Multicollinearity can be reduced by restricting parameters, for example, by giving a parameter of one of the multicollinear variables a value, based on knowledge from related publications published by other authors, or from the economic theory, and estimating the remaining parameters. In the case of lagged explanatory variables you can expect correlation between the lags, similar to our relationship between cocoa and coffee prices. Three lags of P,coc have been specified: for example, the first three autocorrelation coefficients of
the cocoa prices are:
r
(P,C pt""!) = OC, (P,CC Pt~) O,
0.990
r
= 0.975
r ( p,COC ,~) = 0.962 .
One remedy can be the respecification of the equation with, for example, \IX, and X'  2 instead of X" X'_l and X' _2 separately, if pnssible. For that reason, this hypnthesis was tested with a Wald or LRtest in Sections 5.7 and 5.8. However be cautious, as the firstdifference operator eliminates a linear trend in a variable (see Remark 7. 1 below). If tbe linear trends in all the explanatory variables are eliminated by taking first differences, and the dependent variable also has a linear trend, then you get specification problems when that variable has not been adj usted. For more discussion see Chapter 12 and Remark 12.1 in particular.
Artificial explanatory variables
159
7.5
Artificial explanatory variables
Artificial variables are variables which are not observed; their observations are detenninistically detennined values. The constant term, deterministic trends and dummy variables are examples of artificial explanatory variables (often called qualitative explanatory variables). It is also possible that the dependent variable is a dummy variable. That is the feature of logit and probit models. These models are introduced in Chapter I I. When an additive seasonal pattern of the dependent variable is not explained by one of the explanatory variables, dummy variables can be used to specify a seasonal pattern of the dependent variable. Dummy variables can be used to eliminate outliers. or to model different regimes, etc. In this section these variables are discussed in more detail.
The deterministic linear trend
If the dependent variable exhibits a linear trend and not one of the explanatory variables has such a pattern. then the estimation result will show a lot of residual autocorrelation. The model is not able to explain the trend. This problem is solved by including a detenninistic linear trend variable in the equation as an additional explanatory variable. This yields a good fit of the equation to the sample data and possibly good shortterm forecasts thereafter. The 'linear trend variable' t is defined as:
l' = (1
2 3
...
.. .
n)' ,
and a model specification as described above is, for example, specified as:
Y,
=
!It + /h X, + {Jot + !L,.
(7.20)
In this example, the deterministic trend explains the trend in Yt, whereas the economic
variable X, explains the variation of Y, around the trend. In EViews the linear trend is known as a variable with an @sign: @TREND(d). with d being the start date that gets an 'observation' equal to zero. (EViews starts with zero instead of one.)
1' = (0 1 2
n  l)',
this implies a different estimate for the constant tenn. You then have the following options to specify a linear trend variable in the equation. If the trend variable is generated as a new series, then it is included in the workfile, for example:
genr [series name] = @ TREND(d) genr t = @trend(1960:01)
and specify equation (7.20) in the 'Equation Specification' window:
YC Xt.
160
Chapter 7. A Collection of Topics Around the linear Model
But the trend variable can also be specified ' locally' in the 'Equation Specification' window. For example, equation (7.20) is specified as:
Y CX@ TREND (1960:01 ).
Remark 7.1
You bave to be alert wben specifying a detenninistic trend variable in a model with differenced variables. See what bappens in the following example:
Yi VYi
because
= /31
+ /32X, + /33t + u,
.j. take first differences
= /32 VX,
+ /33 + Vut
Vt = t  (t  1) = 1.
So in a model with the variables in first differences, when the constant tenn (/33) has been included explicitly, itis the coefficient of the linear trend tenn. The real constant term is present in the model in an implicit way and, as seen in Section 5.2, it can be computed as:
If an equation is estimated in first differences and a trend tenn does not belong to the model, no constant tenn should be specified in the 'Equation specification ' window.
Remark 7.2
The specification with a deterministic linear trend term makes little or no economic sense in the long run. Realise that Y, + 00 if t + 00, which is not really realistic behaviour for an economic variable. The specification of the trend yields a good fit of the model to the data in the sample period and can be used for sbortrun model forecasts only. This idea has an impact on the specification of the trend tenn. Other possibilities to model trend behaviour exist and are possibly more realistic from an economic point of view (see, for example. the discussion on deterministic and stochastic trends in Chapter 12).
What to do with outliers?
If the variable Yi has an outlier which is not explained by one of the explanatory variables, this outlier is located back in the residuals. A dummy variable can be specified to eliminate that outlier. Be careful this is only good practice if you know the reason for this outlier. If
Artificial explanatory variables
161
this is unknown, you have a probJem. as it is not common sense to eliminate observations which you do not like! If an outlier is eliminated by a dummy variable, the reason why that observation is an outlier should be explained in the text of the paper. In fact, the dummy variable should have a name. Suppose, for example, that the third observation of the dependent variable is an outlier, then specify the dummy variable D, as:
D' = (0
0
I
0
0)' .
In EViews, generate a series equal to zero, and change the third observation in the series
window. The equation is specified as:
In this way, it is clearly visible that an observation has been eliminated. This means that the constant term is equal to f31 for t = 1, 2,4, ... , n and is equal to (f3, + (33) for t = 3. But, once more, only use such dummy variables if they have an economic interpretation. It does not make sense to eliminate every disturbing observation!
Seasonality
If the dependent variable exhibits a seasonal pattern and the explanatory variables do not exhibit a seasonal pattern, then the seasonal behaviour cannot be explained by these variables. When nothing is done, the seasonal pattern is found in the OLS residuals. You can choose either to model the seasonal pattern or to adjust the data for the seasonal pattern. EViews has a number of possibilities for seasonal adjustment. Doubleclick on the variable and click on ' Procs', 'Seasonal Adjustment' (see Figure 3. 14). Next, an adjustment procedure has to be selected, such as additive or multiplicative adjustment under 'Moving Average Methods . . . ' . Multiplicative methods are appropriate if the series can be divided
into a product of the trend component and a seasonal component, whereas additive methods are appropriate if the series can be djvided into a sum of the trend component and a seasonal component. An additive seasonal pattern is rather restrictive, as it assumes that the seasonal
effects in, for example, each quarter have the same magnitude. The multiplicative adjustment is probably the recommended choice in most cases. Other procedures are available in EViews such as the procedures XII and XI2 of the Census Bureau in Washington DC. If you want to model an additive seasonal pattern, seasonal dummies can be specified. For example, in the case of quarterly data the model specification looks like:
The constant term f31 is corrected in the first three quarters. The constant term is equal to f31 + f33 in quarter I, to f3, + f3. in quarter 2, to f3, + f3s in quarter 3, and is just f3, 10 quarter 4.
162
Chapter 7. A Collection ofTopics Around the Linear Model
The matrix X looks like this:
1 1 1 1 1 1
Xl X2 X3 X. Xs X6
•
1
0 0 0
1
0
0 0 1 0 0 1 0 0 0 0 1 0 0
1
X =
•
•
0
•
• • •
• •
• •
•
1 Xn
0 0 0
Observe that the rank of the matrix X is equal to K  1 if four dummies and a constant tenn are included in the equation! Then X'X is a singular matrix and OLS estimates cannot be computed. In EViews, seasonal dummies can be generated to be included in the workfile, for example, by typing in the command line:
genr d1 = @seas (1) , genr d2 = @seas (2), etc.
The figured in @seas(d) represents the dth season. But,just as with the trend, these dummies can also directly be specified in the 'Equation Specification' window. For example:
y C X @seas( l ) @seas(2) @seas (3).
Then the workfile does Dot expand with more series.
Modelling different regimes
In the practice of econometric research, different regimes can be observed in a sample period. For example, when analysing the price formation on commodity markets that are ruled by an international commodity agreement (lCA), periods with an effective ICA and periods with a freemarket situation can be distinguished. If the price of that particular commodity is below a certain level the agreement can become effective. A bufferstock manager can buy quantities of that commodity to stockpile during a period when the price is too low, or producers have to restrict their exports to quotas that are settled in the ICA, etc. So two different regimes ruling the market are observed in the sample. In fact, two different models are relevant to model the price formation for the differeDt regimes. As an example, suppose that a simple model, in general Dotation, for the price of a commodity in
= D. = /1. For example: d' = (1 1 1 0 0 1 1 0 1) . . + u. that has the value 0 if the ICA is not effective and the value 1 if the ICA is effective. from our own research. + U'I. + (3. = f36 = 0 (7. = Both the constant term and the regression coefficient are adjusted: (3.21) Then both the parameters. + f3.Artificial explanatory variables 163 periods without an effective agreement is: y.D. = f3. + ((3. After that period no market interventions took place. • • Genr dummy = 0 (or = 1. X.22) and f3.7. As an example. depending on what occurs the most) for the sample period.X. or by specifying D • X as a variable in the 'Equation specification' window.21) can be created as follows. An international tin agreement existed in the period 1976(0 1)1985(06). because in both regimes the variance of the disturbance term O'~ is identical! If two separate regressions are performed then they will be different: 0'~1 and 0"~2. + f3. In EViews the dummy variable d ' (7. + f36D. a picture of the residuals from an equation for the price formation of tin is given in Figure 7. + U'2· The specifications of the equations are kept simple to illustrate clearly the problem and its solution. (7. have different values in the two regimes by specifying the model in the following way: Y. (7. + i36' This equation is estimated as: Y. The picture needs no comments! Modelling with the use of dummy variables is out of the question. is computed via a 'generate'. + (3. = f3 . Doubleclick on dummy. . the intercept and the slope coefficient. i33 = f3. and that a model for periods with an effective agreement is: y.X. D. With a Wald Ftest the null hypothesis: Ho : i3. = f33+ f3. + f36Z. The two regimes can be specified in one equation by using a dummy variable D. + u" where Z.)X.X. Realise that this is a restricted model.23) can be tested to verify that the data cannot be modelled with one model specification. + f3.
all the parameters are allowed to vary in each subsample. is equal to the unrestricted sum of squares S in the Chow Breakpoint Test: 2001 :CE Figure 7.6 earlier). With the Chow breakpoint test the null bypothesis.' .8 later that can be compared with Figure 7. A Collection of Topics Around the Linear Model 0. of no structural break(s) in the sample period. More dates can be specified for the breakpoint test. This would produce too many dummy variables for estimating the parameters of the equation. then l: i= l m S. Replace the zeros (or the ones) with ones (or zeros) for the dates that a different regime was effective. In fact it is an F test. otherwise the test breaks down.1 0.8: Output of the Chow breakpoint test for o ne date . In subsample i we have m. similar to the test in equation (7.. > K . suppose the sample has been divided into m subsamples. The Chow breakpoint test When different regimes are observed you may want to test whether they bave to be modelled with different specifications or not.1 76 78 80 82 84 66 88 90 92 94 96 98 Figure 7. observations. . but a different null hypothesis is tested with a different test statistic (see the first example in Figure 7.164 Chapter 7. For a break at one date the test resembles the Chow forecast test. . If parameter constancy is tested.. will be tested. Let SR represent the RSS for the entire sample (the restricted sum of squares) and S. This resuJts in m residual sums of squares.7: Residuals from one model for a sample period with two regimes • • Click in the ' Series window' the button 'EdiH/. m).23) Ho : /35 = /36 = O. So be sure that the sample size in every subset is suffi ciently large so that it is possible to estimate the parameters (m . the model is estimated for all m subsamples. In general. i = I . the RSS in subsample i. Instead of using dummies.
'Chow Breakpoint Test ..nmK ). with the following F statistic: (SR F= (LSi i=l Si) /( mI ) K mi') I!J F((m .. Notice the different pvalues in Figures 7.8. the F distribution is valid if the disturbances are independently and identically normally distributed. '. the null hypothesis of identical model structures in four subsamples of the entire sample is tested. Two examples of the use of the test are shown in Figures 7. ~ /( nmK ) As usual.mK as degrees of freedom. The null hypothesis of no structural break is clearly rejected.Artificial explanatory variables 165 Figure 7. Test the null hypothesis Ho: no structural changes in the sample period at the specified dates. In the first example..  _ " I 'I'Mtum 11101 "" ~ ·"1.8 to 7..9: Specifying four subsamples for the Chow breakpoint test wellknown Ftest: F = (SRS)/g Si d! ' with 9 = (m .I )K .10.1 Rr1£j . . the null of no structural break in June 200 I is tested (previ ously this date had been used as the start of the forecasting period)._ Chow Breakpoint Test: 1975:07 1985:01 2001 :01 Figure 7. 'Stability Tests'.I )K._  Workhh· t til ( ()( 'I A ~~~• • • • ." . and d! = n . This null hypothesis is clearly also rejected. The test is found in EViews in the equation window under ' View' . In the second example.6 and 7." " ...10: Output of the Chow breakpoint test for more dates ..
or if you do not find any outlier(s) with an interesting economic interpretation. A Collection of Topics Around the Linear Model 7. then just report your results with supporting commentary. If nonseasonally adjusted data have been used. Decide what to do: eliminating or modelling the seasonal variation are options.4. • It is quite possible that the estimation result from Case 2 has suffered from multicollinearity. Append the results to your paper. • • • • • . this point is not relevant. If such outliers are present then they can be eliminated by reestimating the model by specifying one or more dummy variables. otherwise adapt the specification.166 Chapter 7. Apply the Chow forecast test. which is the case with Data set 4. For seasonally adjusted data. compute static and dynamic forecasts for the omitted quaners and compare the predictions with the actual values according to the directions discussed in Section 7. Is the structure of the model constant in the sample period? Reconsider the results of the CUSUM tests and the recursivelyestimated coefficients. Look also at a graph of the residuals to search for any structural break at dates that can be explained by economic reasoning. Find out whether that is the case by applying the criteria mentioned in Section 7.3: graphically and by interpreting inequality coefficients. and use the Chow breakpoint test to test the null hypothesis of no breaks. then check whether the seasonal pattern can be considered as additive or not. If you cannot explain an outlier. Probably the same model specification will still satisfy. Investigate the residuals of the model from Case 2 to ascenain whether outliers are present that can be economically explained. An additive seasonal pattern can be modelled by specifying quanerly dummy variables. The following points may be of interest relating to the model that has been estimated in Case 2.6 Case 3: exercises with the estimated model Case 3 concerns the various subjects that have been discussed in this chapter and is an extension of Case 2. With this model. Reestimate 'your model' without using the last four or eight quaners.
~ Part III .'''.
In Parts I and II sufficient subjects are discussed for a basic course in econometrics. With the subjects discussed in Part m. the SUR (seentingly unrelated regression) model whose parameters can efficiently be estimated with GLS. In each chapter. but combining these equations in one system can improve the estimation results. the simultaneous equation model (SEM) is introduced.Introduction to Part III In Part m. which is a multiple structural equation model with endogenous and exogenous explanatory variables included in the equations. It is known that the OLS estimator is an inconsistent estimator for the parameters of such a model. the twostage least squares (2SLS) estimator will be introduced as a consistent estimator for this model. Not all the subjects bave to be discussed in a followup course. [n Chapter 9. A third variant around the subject of different djsturbancetenn assumptions are the models with autoregressive conditional heteroskedastic disturbances. the structural model is considered. . a linear model will be introduced with different disturbanceterm assumptions than bave been made in Part II for tbe reducedform model. The parameters of such a model can be estimated efficiently by the generalised least squares method (GLS). that basic course can be extended. Therefore. the ARCH and GARCH models. In Chapter 8. We also come across the first multiple equation model. Only if you choose to study the simultaneous equation model in Chapter 10 it is then also necessary to study the estimation problems in Chapter 9. The equations of a SUR model are treated as reducedform equations in this chapter. after which we will show in what way its parameters can be estimated and what can be done with that model in practice. In Chapter lO. a specific model will be introduced. a selection can be made. where endogenous explanatory variables are introduced in the linear model. several special structural econometric models will be di scussed.
All the introduced models can be handled with procedures in EViews. which will he shown. . models are discussed that have a dummy variable as the dependent variable. Empirical cases are included in the chapters.Introduction to Part III 169 In the last chapter of Part Ill. these models have qualitative dependent variables.
.
I has been obtained by making use of the assumption that . Remember that the expression VaT (.3. properties of the reducedfonn model have been established under the assumption that the disturbances are independently and identically distributed. it was also emphasised that the standard errors have been computed erroneously if the disturbances are autocorrelated andlor heteroskedastic.Chapter 8 Estimation with More Disturbance 8.1 Introduction In Part n. it was shown that the OLS estimator is biased and inconsistent in that case. There is one remedy to eliminate the violation of the disturbancetenn assumptions. A number of autocorrelation and heteroskedasticity tests have been discussed to test this assumption. In Section 7.3) = O'~ (X'X ). because the economic theory yields an economic model for the systematic part and not for the unsystematic part of the model. It is not opportune to model the disturbance tenn. which is to respecify the systematic part of the model to get random residuals. If autocorrelation andlor heteroskedasticity is detected. In Part n. then this often has to be interpreted as the result of specification errors in the systematic part of the model.
These reasons for systematic behaviour are of a theoretica1 or empirical nature. Estimation with More General DisturbanceTerm Assumptions as was shown in Section 4 . In the last section of this chapter. we will discuss models that have disturbances that are not random for obvious reasons.2 ::. The linear model is now written as y = X{3 + u . We will apply thi s method by modelling the disturbance term according to the indications from the theory or the empirical analysis. The meaning of these processes in the disturbance term for the specified model is discussed in Sections 8.. with u ~ N (0 . In Section 8. This substitution is not valid when autocorrelation or heteroskedasticity should belong to the data generating process (DGP).172 Chapter 8.7.. and it is certainly wrong when the model is misspecified. n ) ..2A). Sometimes it is convenient to write the distribution of the disturbance term as follows: N(O. Next. n ) = N (0. This is a multiple equation model where the matrix n is derived from the way thi s model has been specified. the GLS estimator is introduced as a theoretical solution to the problem of estimating the parameters again in an efficient way. the generalised least squares estimator (GLS) is an estimation method that can be used to obtain efficient parameter estimates. In that situation..6. this method also estimates parameters of a structure for the disturbance term. with u ~ N (0 .. Another result from Chapter 7 is that the OLS estimator is an unbiased andlor consistent estimator when autocorrelation or heteroskedasticity is part of the DGP. When the null hypothesis of no autocorrelation has been rejected. n ) . The (n x n) covariance matrix n may have unequal elements on the main diagonal (varying disturbance variances) and covariances unequal to zero. If thi s is a correct specification of the model then the parameters have to be estimated with the GLS estimator to get efficient estimates.2.: The generalised least squares (GLS) estimator in theory The GLS estimator will be derived to estimate the parameters of the following model: y = X{3 + u.4 with Case 4. The standard errors of the parameter estimates are wrong and the estimator is not efficient any more. Hypotheses about the form of the autocorrelation andlor the heteroskedasticity will be made to limit the number of parameters that has to be estimated. . a brief introduction to ARCH and GARCH models is glven .:8 . Then we return to the linear single equation model.3 for estimating the parameters of the Seemingly Unrelated Regression (SUR model).5 and 8.1) where n is assumed to be a known (n x n) positive definite variancecovariance matrix... In this chapter.. a direct application of the GLS estimator is given in Section 8. The SUR model is applied in Section 8. then we have to consider AR. .or MAdisturbances. (8.
The result will be a model with transformed disturbances C u that are nonautocorrelated and homoskedastic: Cy = CX{3 + Cu. (8.2) Derive the moments of the transformed disturbance term Cu to show the mentioned properties: E (Cu ) = CE (u ) = 0 = C · (12A·C' = (12CAC' = a2In. (8. is autocorrelated but homoskedastic) or as a proportionality factor in the case of heteroskedasticity.1y. This equation can be rewritten to obtain the following property: A = C . 1 (8.l X'C'Cy = (X'A. Var (Cu ) = C · Var (u )· C' So model (8.2) to get an expression for the GLS estimator. Both sides of the equation are multiplied by the matrix C. That means that a Choleski decomposition can be applied to A . Then its inverse matrix is positive definite also.'Xr1X'A.The generalised least squares (GLS) estimator in theory 173 where (12 can be interpreted as a variance (u. The matrix A is assumed to be a known positive definite covariance matrix.2) is an efficient estimator again.I = C'C. 139 " 1 = ((CX)' (CX)r (CX)' Cy = (X'C'CX).4) .2) satisfies the classical assumptions of no autocorrelation and no heteroskedasticity in the disturbance term.3) and analogous to OLS we get the following expression for the covariance matrix of (3: Var (13 9.IXr ' . Apply OLS to equation (8. Then it follows that the OLS estimator applied to model (8. 1 The matrix C will be used to transform the model.1 (CT CAC' = In.. ) = (12 ((CX)' (CX)r = (12 (X'A .L it is possible to find an (n x n) nonsingular lower triangular matrix C such that: A.
the GLS estimates can be computed by esti mating the transformed equation (8. For example. representing identical unsystematic influences like world market developments or similar influences of economic cyc1es. This means that GLS is the appropriate method to efficiently estimate the parameters of the model.7) Often A or C are unknown in practice. is computed from the GLS residuals: egis = y .2): Cy = CX(3 + Cu. (2 (X . Indirectly. with Cu ~ N (0.5).3) and (8.. 8.7).2 (X' A. These correlations do not influence the OLSestimation results by equation.6) . a number of coffee consumption equations can be estimated for various countries.. It consists of equations explaining identical variables. = n 2 1 I Kegl s egl s . you can test. or a number of production functions concerning the same product can be estimated for different companies. the more efficient the estimation results will be. X Var (. by using the tests for autocorrelation and heteroskedasticity from Chapter 6. The disturbance term has a covariance matrix n that originates in the manner in which the model is specified. (8.8gls)   = (. Directly by computing the GLS estimators (8. . To obtain an . A SUR model is a multiple equation model with seemingly unrelated regression equations. etc. in case it is the disturbance variance.5) where (12. The GLS estimator can be computed in two ways. The different equations are seemingly unrelated. (8. . but that correlation between the disturbance terms of the equations exists.A 1 )') . Estimation with More General DisturbanceTerm Assumptions The sampling distribution of the GLS estimator is (X is exogenous): ~ (3gls ~ N (3. but a wellknown statistical phenomenon is that the more information that is used. whether the transformed disturbances Cu are indeed nonsystematic.2In) . When the hypotheses of homoskedastic and nonautocorrelated disturbances are not rejected we consider the used matrix C as a feasible transformation..1Xr ' .3 The SUR model The SUR model is a proper example of a model with autocorrelated disturbances caused by the model specification process. but for different samples.8gls) can be estimated as: Var (. then it is recommended you estimate equation (8.X(3gls (8. In that situation. The idea behind this model is that these equations are independent.174 Chapter 8.
X. With t. . = Yl Y2 Ym X.The SUR model 175 estimator for all the parameters of all the equations. and u. This can compactly be written in matrix notation as: To combine all the equations in one model. if t =F S . ift =s. consistent and efficient estimator. /31 /32 = • U.) matrix and /3. it is usual to assume that the contemporaneous correlation is identical in each period. s = 1. = 0 Xm /3m Um . nonzero correlation between the different disturbance terms is expected. .U'l) etc. + U" U. The number of explanatory variables (K. the following vectors and matrices are defined. . applied to each individual equation. implying that OLS. In) In general notation the ith equation is written as: The number of observations is equal to n in all the equations. is an (unbiased).In) Y2 = X 2i32 + U2. U2 . u. This means that y. is a (K . ~ N (O . and not to make the model too intricate. the equations have to be written in a system as one model.'). This is reflected in the following common notation for the variances and covariances of the disturbances. To limit the number of parameters. To derive the specification of this model we consider m linear equations that are written as: y. and that the correlation exists only contemporaneously. y.a~.a~. . is an (n x K. which can be written as: with variancecovariance matrix As written above.n. U2 ~ N (O. a. with: 0'11 = O'~I I and 0'12 = aU t . X. X2 0 0 0 J = f3. Each disturbance term is distributed as N I D (0.. are (n x 1) vectors. x 1) vector. 0 0 . = Xl/3.) may differ in each equation.
= n .) = (' x. UI) Var (u. u rn) . according to (8 . 0'1m 1n . • First. ) . is an (nm x and {3. The following feasible GLS procedure can be used to compute the SUR estimates.n.ei e j = 12: etietj..) = Cov (u~. with t". t= l .4): Var ( /3. u . the matrix rl. The use of the contemporaneous covariance a ij in estimating the parameters of the model will result in smaller standard errors and more efficient estimates compared to OLS by equation. ..8) ) . In practice..) Cov (U2. This is not a scalar matrix so the parameters of model (8.. estimate the parameters of all the individual equations with OLS. Var (u . is unknown.~I f: K matrix i) We have to know what the covariance matrix of the disturbance term u .. 1n Var ( u~) 0'22 1n . so its parameters have to be estimated. . . This knowledge is necessary for correctly choosing an estimator for all the parameters in the vector /3•. are (nm x 1) vectors.' .. Estimate the unknown variances and covariances as follows: • ~ (Tij n 1 .. that yields m residual vectors ei l with i = 1. ) = 1J ' " n n .. 0'2m.m. U2) Var (u2) • COV (u. with covariance matrix according to (8. the matrix X . looks like. . an ( ~ i~ K i x 1 vector. I x .) Ull I n 0'21 In 0"12 I n Cov (u.. The theoretical GLS estimator. m. urn) Cov (U2. Estimation with More General DisturbanceTerm Assumptions The vectors y . ' = . All the m equations are written in one system: (8.8) can be efficiently estimated with GLS. and u .3).176 Chapter 8.. for this model is: ~ n1x n/3" = (X'" " " )' X'" " Iy .
then select 'Seemingly Unrelated Regression ' as the estimation method..1: Creating the object 'system' . In the EViews window click the buttons 'Objects'. the GLS estimator has been directly apptied to estimate efficiently the parameters.' x:n. In small samples. You have to define a system of equations (see the help function for all the information).1 . 'New Object' . Remark 8. and compute the feasible GLS estimator: (In  n. Type in the 'System' window the equations according to the explicit form.2. and next click 'Estimate'. which results in the window in Figure 8. An example is given in Figure 8. . and covariance matrix./n K. The SUR estimator can conveniently be computed in EViews. K j ) instead of Substitute these estimates into the matrix fl. that is when U ij = 0 for all i ¥ j.The SUR model 177 • This is often done without making adjustments for degrees of freedom. giving the matrix fl .1 Check that the SUR estimates will be identical to the eartier obtained OLS estimates if the equations are 'really unrelated' . possible adjustments can be made by dividing through (... = ( x:n. I CiiiCOI'" 1 Figure 8.5): In the SUR model. according to (8. 1y. x.) . /3. 'System'. Give the 'System' object a name and it is included in the workfile.).
Therefore keep the workfiles saved as you will need them again later. 8. Two equations are suitable for this exercise: macroeconomic consumption equations or investment equations. It is interesting to compare the singleequation results with the fullmodel results. For an example of a SUR model see. the OLS estimator is still the only estimator that has been discussed and will be used. Estimation with More General DisturbanceTerm Assumptions Figure 8. for example. .2: Specification and estimation of a SUR model Notice the other estimation procedures that are available in this window. where a SUR model is estimated that explains the attendance at various clubs in the English Football League.178 Chapter 8. Harvey ( 1990). in this exercise. The results from this case will be reconsidered in Case 6 after the model with endogenous explanatory variables has been discussed. Also. Choose one of these equations for this case.4 Case 4: a SUR model for the macroeconomic data The macroeconomic data of the five countries that accompany this book are used to estimate a SUR model in this exercise.
and the use of the GLS estimator is not appropriate. Use the specifications that have been found for the individual countries. So. Open the new object 'System' and save it with any appropriate name. 8. White's estimation option is not available in the system window. Under 'View'. Are the SUR results more efficient • • in this case? Write a clear and complete research paper. Theoretical reasons for an AR( I) model are not present. Under 'Procs' in the 'System' window the procedure ' Make Residuals' (for all the equations) is found. it is a correct procedure to determine the covariance matrix of the disturbances and to use GLS for getting efficient estimates of the parameters. AR(I ) residuals are often found after a misspecified model has been estimated with OLS. ' Residuals' the view option 'Correlation Matrix' (of all the residuals) is available.2. Then an AR( I) model for the disturbance term is added to the model: n However. That situation is utilised to show the use of GLS. Estimate the parameters of all the equations with GLS by marking the SUR option. neither are residual autocorrelation tests. .Autocorrelated disturbances The following checklist can be used to write a paper about this subject. the DWstatistic) has rejected the null of no autocorrelation. Consider and compare all the obtained results. Compute the correlation matrix of the five residual variables to get an indication of whether or not the SUR model might improve the earlierobtained singleequation results.6). • 179 Estimate dynamic models for each of the five countries for identical sample periods in the same way as has been done in Case 2. In case the autocorrelation exists because of a specific model specification. the residuals are not allowed to have significant autocorrelation. in many textbooks an AR( I) disturbance term is introduced. like the SUR model. Thus it is always possible to look at the correlograrn of the residuals. However. You can look at graphs of the residuals. This can be an alternative for the procedure mentioned after the first • • • • • buUet. or because of model transformations.g. after an autocorrelation test (e. Write the five consumption or investment equations in the explicit way in the 'Systern' window as shown in Figure 8. If the option OLS is selected then the single equation results are obtained again. but its use depends on the source of the autocorrelation. and save the five residual variables. Reestimate models with the White option if heteroskedasticity has been found in the residuals (see Section 8. as stated earlier: the systematic part of the model has to be respecified. But it is not expected that the SUR procedure generates autocorrelation in the residuals.5 Autocorrelated disturbances Sometimes the GLS estimator can also be appued when the disturbance term is autocorrelated. Of course.
Ut = 131+ f32 X'2 + u" + Vt· (8. For example. if you are interested. The objection against these procedures is that the model has been extended with a new parameter 'PI . etc. Est imation with More General DisturbanceTerm Assumptions However. 1 'PI 1 'P I n = a 2A = 2 'PI 'PI • 'P~ 'PI 1 n. these procedures will briefly be mentioned here. are wellknown. the economic theory does not indicate specifying an AR(I ) model for the disturbance term in some equati on from a macroeconomic model.l Various methods exist to compute GLS estimates for the 13k and the 'PI in this model. 'PI 'PI 1 (8. . then the following matrix expressions can be derived. A second objection is that an AR( I) di sturbance makes no sense with respect to the model specification from an economicaltheoretical point of view. the only interpretation being that the researcher has a misspecified model! Therefore. . When the firstrow of the matrix C (incorrectly) is deleted. In EViews the misspecified bivariate model (8. .Lu. then the following transformed model is obtained for the observations t = 2.n: • . The procedures of CochraneOrcutt. . Hildreth.'Pf c= 'PI 0 0 1 'PI 0 0 1 0 • 0 0 0 0 1 0 0 'PI For example. See the EViews User's G uide for a description of the nonlinear estimation procedure to obtain estimates of the parameters of such a model .I 'PI u 1 .I 'PI 'PI and 11'1 . once more: do not model the disturbance term when the systematic part of the model has to be modelled in a correct way.2 'PI n.9) n.2 n. 10) can be estimated by spec ifying an AR( I) term in the 'Specification Equation window': Y C X2 AR(I ).180 Chapter 8. because AR( I) procedures are found in the econometrics literature and software. Durbin. so that you will recognise them and know not to use them.'PI au 2 . 10) = IPI Ut . If it is assumed that the di sturbance term follows an AR( I) process. let us look at the bivariate linear model Y.
For example. Then the GLS estimator can be correctly used to obtain efficient estimates or asymptotic efficient estimates. it would be more sensible to derive 'MA(I )' procedures instead of 'AR(I )' procedures.2) u~ _>.'l'lu. this procedure is found in the (historical) literature. This procedure transforms the original static equation to a restricted dynamic equation. the matrix n is wrinen as: 2 _ >. forexarnple. + /32X'2 + v.uu n= 0 2 _ >. An MA(J ) process for the disturbance term can occur when this specification has arisen from a model transformation to solve other problems.>'u~ i > 2 : Cov (v" V'_i) = o.uu (1 + >..2) u~ .d = v.. 2.2) u~ i = 1 : Cov (v" v.AUt _ l.2) u~ .uu 0 ••• 0 (1+ >. For this reason.a~ 0 0 • • 0 • • • 2 (1 + >.. is easily derived: Compute Cov (v" V' _i) for i = 0. The covariance matrix n of v. . A low DW value is also found when the disturbance term follows. The transformation shows that the specification errors in the original model consist of a lack of dynamics. depending on the specification of the transformed model. but correct. consider a bivariate model with an MA( I) disturbance term: Yi = /3.). an MA( I) process. specification of a dynamic model. Sometimes. Examples of autocorrelated disturbances caused by model tran sformations are found in Chapter 13. implying the model should be respecified in a dynamic direction.. ..uu 0 • • • • • 0 2 _>. = (1+ >.d = . i = 0: Next. with Ut "" Vt = Ut . 1. The random di sturbance term u. belongs to the originally specified model.uu 2 _ >. N ID (0. u~) .Autocorrelated disturbances with 181 (u. Respecification of the model by the researcher to eliminate the autocorrelation will result in a different. : Cov (vt) = u.
6 Heteroskedastic disturbances Heteroskedastic disturbances are mainly a problem in modelling with crosssection data. First. it is also possible that an economic reasoning exists for the existence of heteroskedasticity in the disturbance term of the model. Heteroskedasticity can be part of the DGP for empirical and/or theoretical reasons. but it is not interesting for economic reasons to determine a particular relationship for the variance O'~ . GLS is an incorrectly used estimation method when no a priori reasons for autocorrelated disturbances exist. then GLS estimates can be computed by estimating the parameters of a transformed equation like (8.182 Chapter 8. but the problem can also exist in causal models for timeseries data.7). the source of the heteroskedasticity has to be found. However. Of course. .) If beteroskedasticity belongs to the model then efficiently estimated parameters can be obtained by applying the GLS estimator. the parameters of this model can be estimated by adding an MA(l) term to the equation in the 'Equation Specification' window: YCX2MA (1) See the EViews User's Guide for more information about the estimation procedure. or an explicit expression for n has been derived for some structural model. Summarising • GLS is an estimation method that cao be used to obtain a correctly specified aod • estimated econometric model. (See once more Section 6. as has been discussed in Section 8. autocorrelation for theoretical reasons only. because in that case specification errors are the cause of the systematic behaviour of the disturbances.2 concerning equation (8. The heteroskedasticity problem is di scussed in the next section. Then a second possibility exists to get asymptotically consistent estimates of the variancecovariance matrix of the parameter estimates. The heteroskedasticity problem can be tackled in two ways. Then the systematic part of the model has to be respecified.7) with OLS. The method is called weighted least squares.4 for some possible causes of heteroskedasticity and the tests to detect these causes. the same statements are true with respect to heteroskedasticity in the disturbance term. Estimates can be computed by using the correction a/White when estimating the parameters with OLS. Both methods are discussed below. The first method is the application of the GLS procedure. 8. Estimation with More General DisturbanceTerm Assumptions In EViews. . when theoretical reasons are known for some timeseries model for the disturbance term.
i for all o o o • • • • o o = (12 • • • 1 ~.2 .. .. 0 0 v0: 1 0 0 v).. This pattern can also be found for the disturbance variance. • • A. Then the question is: what does the covariance matrix Var (u ) = 11 look like in these situations? We are more interested in the matrix C .. 0 X. the transformation matrix C is derived. variables can be observed with strong trends. o o . in a model for crosssection data that explains personal consumption of some lUxury article by personal income and some other economic variables. . It is common to rewrite the matrix 11 in the following way by introducing a constant (72 such that (7~.' ~n o 0 1 0 0 1 ~n In this situation. In models for timeseries data with many observations. .1 and the transformation matrix 0 ). 0 0 (72 Un ..2 0 0 0 . The covariance matrix of u..Heteroskedastic disturbances 183 Weighted least squares The point of departure is that the heteroskedasticity belongs to the model. In the following steps...n =C= 0 0 0 0 . 0 0 11= • • 0 0 . . 1 • • • • • • 0 • 0 0 • • • • 0 • = (72 A .. when the di sturbances are heteroskedastic but not autocorrelated can be written as: (72 u.. 0 . as we wish to transform the model to a model with a homoskedastic disturbance term. In models for crosssection data. heteroskedasti ciry can be observed in tbe form of a relationship between (7~ .. .n . or has been caused by the sample data. and explanatory variables.. GLS can be applied to estimate the parameters of these models with heteroskedastic disturbances to obtain efficient estimates. . it is easy to derive the inverse matrix A C: ). = (72 / ). one can expect that the variance of the consumption is small for low income levels and large for higb income levels. For example. . 0 v).. Then it is also possible that a trend can be found in the residual variance.I = • 0 0 0 ).. 0 2 (7u..
It is not a new model but a transformed equation for computational reasons only. fl . Estimation with More General DisturbanceTerm Assumptions Look at the transfonnation of. + flK X iK + ". Then the transformed equation is: (8.. = fl. . This equation can be estimated with OLS to get efficient estimates of the fl•.. for example. can efficiently be calculated from the following regression: ". X 2 . for example the result of a Breuscb..>:. 0 o 0 o Cy = v'>:2 .184 Chapter 8. Suppose that the following form of the heteroskedasticity has been detected with a BPtest: Then the matc hing weigbts J>:i are found as follows: A i 1 2 = X i 2 ==> .:. and X i.."... . All the observations are multiplied with J>:i. = X .. • o • • o .. X i2 + fJ2 I X i3 + fl3 X i2 '" + . The result (8. = x i2 Y. are weighted with J>:i. to show the effect of this transfonnation manix C on the variables of the model: v'>:. . + fJ2 X i2 + . the dependent variable y. Suppose we have the following equation: Y.. + flK X ii K + X i2 .. 11 ) In other words: the variables Y.. The transformed disturbance term v'Xiui has a constant vanance: • Relevant weights have to be established. 'i2 1 The parameters fl. 1 I) of the transformation explains the name weighted least squares.Pagan test (BPtest) can give a good indication.. . This happens with the explanatory variables too.
. Ut .. But suppose.2 = .3: Application of weighted least squares in EViews X i2 is homoskedastic..Heteroskedastic disturbances 185 s ' ct R ¥lhn • Acorp. . Click on 'Options'. See Figure 8.. fh is the constant term in the regression but remains the coefficient of X i 2 in the economic model. only for the following schematic example.3 for an example concerning the relationship among cocoa and coffee prices. that we bad detected the following form of heteroskedasticity: . mark 'Weighted LSrrSLS'.2 (pca!)2 t.. The weighted least squares estimator is found in EViews as one of the options in the 'Equation Specification' window. Spa Ii 'hud to favllf:. The transformed distUIbance term Ui / I A!f 1 Figure 8. In Chapter 6... and type the name of the weight. we did not detect heteroskedasticity in the residuals of the equation for the cocoa price in the form of a relationship with one of the explanatory variables for the entire sample.. Only a decreasing trend in the variance was found with the test of Goldfeld and Quandt.
has been rejected with a Breusch. only as instructive example. as that is tbe model.. Generate the new variables. Next.Pagan test. of this relationship. estimate.3) is calculated. This procedure has one di sadvantage: it means that the GLS estimator (8.).) + (w. the equation: 2 + 0'2 X i· e2 i =  0'1 Compute the predicted values e 2 . 12) the homoskedasticity of the disturbance term ". X i ) + f33 (WiZ. and estimate with OLS the equation: (Wi Yi) = f31W. However. because it belongs in the GLS residuals.. A White test will find heteroskedasticity in the GLS residuals. So you do not have the possibility of checking whether the covariance matrix n has been sufficiently accurately approximated to reduce the heteroskedasticity.186 Chapter 8. in favour of the alternative hypothesis that: Next. Estimation with More General DisturbanceTerm Assumptions then the matching weights are: This weight is shown.3. or specify the transformed variables in the 'Equation Specification' window. + f32 (w. unknown parameters are part of that relationship. Then a beteroskedasticity test can be applied to the OLS residuals. The theoretical procedure to obtain GLS estimates of the parameters can be summarised as follows.. .6). Therefore. 12). in the 'Estimation Options' window in Figure 8. In that situation. Of course the GLS residuals are heteroskedastic. we have studied the GLS procedure with a strai ghtforward proportionality of the di sturbance variance with an explanatory variable. from equation (8. and we only get the GLS residuals (8. These parameters have to be estimated before the weighted least squares procedure can be applied. the weights can be computed as: Wi= 1 ~. it is possible that the variance of the disturbance term can be expressed in a linear relationship of explanatory variables. 11 ) with OLS. good advice is do the procedure yourself by estimating an equation like equation (8. So far. by using the OLS residuals e. Suppose that in the model : (8.
.. the decision bas been taken to use the square roots of the predicted values. The estimation results are not reproduced here. which can be used as a simple formula to compute these values: GEN R RES2F = (RESID02 "2) . are hornoskedastic. The model is the estimated regression equation. They have been stored as LPCOCOAYIT.03.. as weights. For example.et.. Choose in the 'Model ' window: 'Solve'.. the regression of the squared residuals RESID02 has been perfonned on LPCOCOAYIT (see Figure 8.. If this null is not rejected then efficient estimates have been obtained and the econometric model is written in the original specification by using these estimates:    Yo = f31 + fh X i . Finally. .1 lllual8S We noticed in Part n that the relationship between cocoa and coffee prices was not very stable... This can be done in various ways. 1. the predicted values are computed.RES! D. 4.. . Next the predicted values are calculated.1 an application of this procedure is given.. 3.. Next. In Example 8..fh (Wi X i) .. Therefore. The null hypothesis of homoskedastic disturbances is rejected at a significance level of 3%. So the relationship has been estimated once more but now for the sample period 1990(0 1)2002(09). Example 8. The regression output of the test suggested that almost all the variables contribute to the heteroskedasticity. which is a linear combination of all the explanatory variables.Heteroskedastic disturbances 187 Test whether the newly obtained residuals ih from this regression Vi = (WiYo) . but the result of the White . + f33 Zi' . remember the definition yt =  Yi .f33 (WiZi). In the original equation. 5.4).76 with a pvalue of 0.they got the name RESID02.. 2.. the OLS residuals have been saved . ..test is nR2 = 19. In the original equation the option 'Make Model ' has been chosen under 'Procs'.fhwi .
as is shown in Figure 8.0010 0.04.iduals of the last regression from step 4.049617 0.(83926 16. This is more convenient here than the procedure in steps 2 and 3.004222 0.002874 0. Error tStatistic Prob.4346 1.rJ22452 153.044502 6. Error I·Statistic Prob.054670 0.123573 0.0394 0 ..997706 0.3695 0. It is not necessary to Dependenl Vanable: LPCOCOAI(RES2f'O.0024 0 .421 ·2.DBl131 0.4: Regression of squared residual son theexplanalory variables The variable RES! D contains the res. 0. the weighted regression equation is specified and estimated by dividing all the variables by J RES2F..009660 Rsquared Adjusted Rsquared S.5) 0.153164 LPCOCOA(·l)1(RES . ~4 0. Estimation with More General DisturbanceTerm Assumptions Dependent Variable: RESIOO2A2 Method: Least Squares Dale: rnfZ7100 TIme: 16:05 Sample(adjusled): 1990:01 2!l2!lO022:09 Included observations: 153 after adjusting endpoints Variable Coefficient Sid. dependent var Akaike info criterion Schwarz criterion Fstalistic Prob(F.002691 620. 1. 0.lDl14 LPCOFFEE(·l )I(RE .001694 3.D.E.005007 0.732522 0.99479 2.048409 0.188 Chapter 8.E.368734 ·2.634400 3. 0..994515 0.5.076136 0.256S84 LPCOCOA(3)1(RES .119065 0. of regression Sum squared resid Log likelihood 0.023204 0.D. of regression Sum squared resid Log likelihood 0. 0. ·0.0036 0.4344 Mean dependent var S.rJ26790 Figure 8. Subsequently.079005 LPCOFFEEI(RES2F .920711 3. Now the predicted values of the regression from Figure 8.253655 LPCOCOA(2)l(RES .049352 0.(00) 0..076799 0.006428 0.0093 0.955C64 DurbinWatson stat Mean dependent var S.6751 ·217..004328 8.67009 20..997628 1.5) Method: Least Squares Dale: 02127100 TIme: 15:55 Sample: 1990:01 = Variable :09 Coefficient Std. Included observations: 153 1/(RES2F'O.statistic) Figure 8..DCl68a8 0..£0J230 2.4 are in tbe seriesRES2F.5: The weighted least squares regression .003628 C LPCDCOA FIT Rsquared Adjusted Rsquared S. dependent var Akaike info criterion Schwarz criterion DurbinWatson stat 60..084115 8.039552 2.
It is possible that 'missing values' are encountered in this regression as elements of the series RES2F might be negative. then asymptotic consistent estimates of the standard errors of the estimated parameters can be obtained in the following way.95250 0.09%.5517Bl OJXl4435 0.522400 Figure 8. See the White test in the window in Figure 8. They are specified in the 'Equation ' window.237141 0.B42879 0.2741B 1 .712253 0.57B6 11 .250521 0.B2879 Probability Probabil~y Test Equation: Dependent Variable: RESI{)A2 Method: Least Squares Date: 10113. Error I· Statistic Prob.154167 0. .7). It is possible to estimate the parameters and standard errors consistently and asymptotic efficiently with OLS by using the EViews option 'Heteroskedasticity Consistent Coefficient Covariance' in the 'Equation Specification' window (White (1980)). owing to theoretical or empirical reasons.BD3B67 0.BI29 O.178434 0.Heteroskedastic disturbances 189 White Heteroskedasticity Test: Fstatislic Obs~·squared 1. the weighted least sq uares estimates. The estimates in Figure 8.BD2B 0.Bm 0. in the disturbances in a particular model. or more specifically for this situation. Indeed.6022 C 25. but that will be indicated in the heading of the window. This option is found under the 'Estimations Options' (see Figure 8. Heteroskedasticity (W hite sta ndard errors) It is not always necessary to detect a particular relationship between the disturbance variance and explanatory variables in cases of heteroskedasticity. the null hypothesis of no heteroskedasticity is not rejected at a significance level of 25.1474 0.6. use the White test again to check whether our hypothesis concerning the source of the heteroskedasticity was a reasonable assumption. O .252CS8 14.\)3 Time: 14:45 Sample: 1990:01 2002:09 Included observations: 153 Variable Coefficient Std. of an unknown form.5 are the GLS estimates. Lastly.0:12317 HE.6: estimation Result of the White lest after the weighted least squares generate the transformed variables and to include them in the workfile. When it is known that we have heteroskedasticity.
7: Specification of the White option Then the matrix n is estimated with the heteroskedasticity consistent covariance matrix estimator of White.l o White replaces the al by el and the covariance matrix is consistently estimated by : . The option is used when the form or the source of the heteroskedasticity is unknown. Another option is the NeweyWest option.l aT o o 0 a~ 0 X (X'X).190 Chapter 8.lx'nx(x'x). With this option. The 'White covariance matrix' can be chosen when the disturbances are only heteroskedastic. But as in general residual autocorrelation is an indication of specification errors. The NeweyWest estimator of the disturbance covariance matrix is more general and is consistent in the presence of both heteroskedasticity and autocorrelation of an unknown form. The covariance matrix of the OLS estimator is computed as follows: Var(i3) = (X'x). The point . Var({3) = (X'xt 1 (X 'x t l . the covariance matrix is estimated in a different but consistent way by making allowance for the heteroskedasticity. or is irrelevant. it is advised only to use the White option. Estimation with More General DisturbanceTerm Assumptions Figure 8.:. and the systematic part of the model has been correctly specified.
0092 0.150053 U666.081748 0.1291 0.695335 2.0084 0. [n Figure 8.6115 C LPCOFFEE 0.0931 0.8: correction Comparison of the OLS output with and without the White estimates of the parameters are not affected by the use of the White option.268618 0.670717 0.690245 2.338447 0.128175 0.526045 2.6_ Find a correct model that explains the percentage of products that are new to the industry (new.045309 0. The variables in the data set are described in Section 2. In the econometrics literature sometimes it is suggested that heteroskedasticconsistent standard errors should always be computed because some residual heteroskedasticity wi ll always be present.091377 0.835120 15. Error I·Statistic Prob. this is only good advice when you are sure that the heteroskedasticity is not the result of specification errors in the systematic part of the model.49419 2. 1900:01 2002:09 Variable Coefficient Std.536900 0. by usi ng all or a selection of the other variables and by taking into account .054127 0.045309 0.920032 15.05503B 0. the OLS output is given.0079 0.5921 Figure 8. with and without the use of the White option.084378 1. The differences in the calculated standard errors are only small in this example.0B9004 1.795268 2.101209 0.148346 0.509039 0.148346 0.126725 0.150053 1.053070 0.0B0B50 0.640500 0.0052 0 .Case 5: heteroskedastic disturbances 191 Dependent Variable: LPCOCOA Method: Least Squares Date: 02I'2B.154449 0.COl) 0. However.055748 0.ml>l. 8_ 7 Case 5: heteroskedastic disturbances Crosssection data In this case.sjndustry ).(00) 0. See the EViews User's Guide for detailed information and the formulae.154449 0.0059 0. only the standard errors are estimated in a different way.66622 2.8.338447 0.0041 O.8 0. C LPCOFFEE LPCOFFEE(l) LPCOCOA(l ) LPCOCOA(2) LPCOCOA(3) 0. you have to tackle the heteroskedasticity problem in a model for the crosssection data from Data set 5.01J3 TIme: 10:21 i.
Pagan. Reestimate the model with OLS by making corrections for heteroskedasticity and other problems. • Estimate the model with OLS and test the assumption of homos ked as tic di sturbances with the White test. Transform the model according to the outcomes of the test and esti mate the parameters of that equation with OLS to obtai n the GLS estimates. Test the homoskedasticity assumption of the di sturbances of the tran sformed equation with the White test. Follow the procedure that has been di scussed in Example 8. Timeseries data At the end of the first ' research paper' written in Case 2 the result of one or more heteroskedasticity tests had been reported. the ARCH and GARCH models are discussed at an introductory level. etc. The model will be estimated twice: ( I) by using the option 'Heteroskedasticity Consistent Coefficient Covariances'. Write a clear paper about the accomplishment of your estimated model. so apply the GoldfeldQuandt test. Reestimate the mode l twice: with GLS to get efficient estimates of the parameters. S. Compare the results and add the evaluation to your paper.S Introduction to ARCH and GARCH models Heteroskedasti city is usuaUy associated with models for crosssection data (with correlations between disturbances assumed to be zero). A second possibility is the use of the GLS estimator. A number of hints are given below. This can be done by using the White option for 'Heteroskedasticity Consistent Coefficient Covariances'. 1 and compare the results with the previous results. Also investigate the residuals for the presence of any particular regular behaviour or outliers by looking at a residual graph.192 Chapter 8. which is the weighted least squares estimator in thi s specific siruation. whereas models for timeseries data are tested for autocorrelation usually with the assumption of homoskedastic disturbances. along si milar lines as those in the first paper concerning the timeseries data. In thi s section. Testing and . and by specifying dummy variable(s) for possible outliers. • • • • • Estimate the parameters with OLS. and with OLS and the White option for the standard errors. If you concluded that the residual s are heteroskedasti c then try to find an explanation for the heteroskedasticity: could thi s be due to trends in the variance? Possibly. and (2) by using the GLS estimator. • That gives the first possibility for correctl y estimating the sta ndard errors of the estimated parameters of thi s model. Test the homoskedastic ity assumption with the test of Breusch. Estimation with More General DisturbanceTerm Assumptions any possible heteroskedastic behaviour of the disrurbances. A form of heteroskedasticity that can be encountered in timeseries models is the (generalised) autoregressive conditional heteroskedasticity (ARCH) or (GARCH).
1.l " " J Ut . The conditional disturbance variance is the variance of Ut I conditional on information available up to time t .. This is written as follows: a. The process described above is called an ARCH(p) process. t where et is a standard nOI11)ally distributed variable. Engle postulated the following relationship: (8. to study the properties of an ARCH(I) model. Notice that two distinct model specifications are considered: one for the conditional mean (of the dependent variable) and one for the conditional disturbance variance. for reasons of simplicity. I Ut. 13) Volatility from the previous period is measured by the lags of the squared residuals. + {3K X tK + Ut. The unit variance is not a restriction as any other variance can be rescaled to unity by suitable adjustment of the other parameters.Introduction to ARCH and GARCH models 193 estimation procedures are available in EViews and are simple to apply. + (lp 'U 2_ p ) 1/ 2 . Therefore. . The unconditional mean of the disturbance term ut is zero and from the assumption that £ t is standard normal it follows for the conditional mean that: E(Ut I ut. the conditional variance. This co ncept may successfully be applied in models for volatile markets. Engle formulated the notion that information from the recent past might influence the conditional disturbance variance.p) = E(u.13) can arise from a disturbance term defined as: Ut = et ( Ct'o + 01 Ut2_ l + . Ut_p). In speculative markets like exchange rates and stock markets you can observe that large and small errors tend to occur in clusters. consider the unconditional and conditional moments of the disturbance term . as will be seen later on. The concept of ARCH disturbances was introduced by Engle (1982).d = O. ..· . The variance as modelled in equation (8. = ct/ao +alu~_l· Next..t = Var (Ut I Ut.. This model is written as: Yt = {3I Ut + fh X t2 + . of a variable will be modelled.I . Recent disturbances influence the variance of the current disturbance and thus the variance of the dependent variable. or the volatility. We go on.. It looks something like 'autocorrelation in the heteroskedasticity'.
194 Chapter 8.) = Var (E(u. O.d = E (u~ 1 Ut _.8 I'): Var (Ut) = E( Var(ut 1 ut. (Var (y I x».fi". a theorem about the decomposition of the variance of a variable from a joint distribution is used (see. Bollerslev (1986) has introduced the generalised autoregressive conditional heteroskedasticity (GARCH) models. I u. (E (y I x)) Thus here: + E. Estimation with More General DisturbanceTerm Assumptions The conditional variance of Itt is: Var (Ut 1 Ut . I U'_I)). for example.Il) +E(Var (". 1 ClOCl l ) andE (utut _i)=Ol. Altogether the unconditional distribution of Ut has been derived : Ut ~NID(O. AJ so.the disturbances are not serially correlated. Greene (2000) p. that a GARCH model with a smaller number of terms can perform as well as or better than an ARCH model with many parameters .797). I "'I» Var(u. The implication of these properties of Ut is that OLS is still an efficient and consistent (linear) estimator. The disturbance term u. The unconditional variance of Ut can also be derived.. q) model is defined as: Bollerslev has shown. In the first step of the derivation. for example. is heteroskedastic conditional on Ut l. • If x and y are bivariately di stributed then : Var(y) = Var.) = E = 00 (cD (ClO + ClIU~_ [) + alU~_ l l which is similar [0 equation (8. which estimates all the parameters of the complete ARCH model (see. A GARCH(p. But a more efficient (nonlinear) GLS estimator exists.1 3). so Ut is homoskedastic. .d) = 00 + alE (U~_ l) + Cl [ Var (ut _d Cl[ = ClO ClO 1  This variance only exists if ClO > a and ICll l < 1. We see that the unconditional variance of Ut is constant. it can be shown that the autocovariances of Ut are zero . = 0+ E(Var(". with an example. Greene (2000) p.
As before. 2 Be aware of the fact that if specification errors in the systematic part of the model are the reason for ARCH residuals. More variations of the ARCH model exist. for example. which is in agreement with the assumption that an ARCH(p) process is absent in the disturbances. a TARCH(I. + O:p e t _ p ' Test the null hypothesis Ho : = = . If it is. then it makes no sense to estimate the model with ARCH disturbances. 'Residual Tests' . from infonnation about the volatility observed in the previous period (the ARCH term). because of economical reasons. These models are not part of a basic course in econometrics. 1) model is often used in quantitative financial research.9 where the null has been tested of no ARCH process against the alternative hypothesis that the disturbance term follows an ARCH(2) process. This is described as you would have to do it yourself ifit were nota program in the software package.. ~ ~ ~ Perform WI·th OLS the regressIOn e t = 0 0 + Ctl et_ l + . relevant to assume that downward movements in the market are less volatile than upward movements of the same magnitude (a kind of irreversibility) a TARCH model can be specified.. the examples given in Chapter 18 in Greene (2000). = = 0. This model is also consistent with the volatility clustering often seen in financial returns data. it concerns the relationship between cocoa and coffee prices. when discussing autocorrelation and heteroskedasticity problems.Introduction to ARCH and GARCH models 195 The GARCH(l. .. • • • Regress y on the variables in X and save the residuals e . A straightforward method used to test for ARCH in the disturbance tenn is an LM test perfonned in the following way. The 'T' stands for threshold .l . Just as before. and the forecasted variance from the last period (the GARCH tenn). 2 al a2 ~ 2 a. For example. . Testing for ARCH effects and estimation with an ARCH disturbance in EViews An example will be given that shows the use of an ARCH test and an estimation method to estimate all the parameters of an ARCH model. . In EViews. Use the LM test statistic nR2. the ARCH test is found in the 'Equation' window under 'View' .a TARCH model is a GARCH model with a dummy variable which is equal to 0 or 1 depending on the sign of Ut . where a trader predicts this period's variance by forming a weighted average of a longtenn average (the constant tenn). See the example in Figure 8. the model has to be respecified if the disturbancetenn problems have been caused by misspecification. where large changes in returns are likely to be followed by further large cbanges.I) model is specified as: See the econometrics literature for a more extensive description of this type of model and for more ARCHtype models.
003)95 Std.022979 0. dependent var Akaike info criterion Schwarz criterion F·gletistie Prob(Fslalislic) 0.006500 0.006563 ·7. of regression Sum squared resid Log likelihood DurbinWatson slat 0.E.048953 0. O.021338 1838.394770 1.009D2 0.D.019110 0.196 Chapter 8.044090 0. 11K I Figure 8. Estimation with More General DisturbanceTerm Assumptions Test Equation: Dependent Variable: RESI[)1l2 Method.10: Specification of a GARCH(I.93:) 2.140761 0..044094 8.9: The ARCH lesl AfICIfM r.OCOJ e RESID"2(·1) RESID"2(·2) Rsquared Adjusted Rsquared S.838736 0. Error O .002823 Figure 8. Least Squares Date: 03118104 TIme: 13:48 Sampla(adjuslad): 1960:00 "200"""2:09 Included observations: 508 after adjusting endpoints Variable Coefficient 0 .I) model .COO~ t·Stal istic Prob.203088 5.228072 ·7.2674 0..110311 3.0015 0. HeN SOlD.003816 Mean dependent var S.192309 0.
5%) significance level. C LPCOFFEE LPCOFFEE(1) LPCOCOA(1) LPCOCOA(2) LPCOCOA(3) 0.036798 0. this means that the variance is influenced by the volatility of the two preceding months.1DXl Mean dependent var S. dependent var Akaike info criterion Schwarz criterion f statistic Prob(Fstatistic) 3.0097 0.629185 0.691B3B 19.019881 0.319134 2.!XXXl 0.002299 1.602815 2. of regression Sum squared resid Log likelihood DurbinWatson stat 0.040574 0.025714 0.0956 1. see the EViews User's Guide for details about thi s procedure. because it is a popular alternative hypothesis in quantitative research in volatile markets.)5 ARCH(I) GARCH(I) Rsquared Adjusted Rsquared S.450B50 0.670748 5894.098436 2.().607117 27JJJII9 5.00J116 0.D.989319 0.094932 0.E.745473 2. The LMtest statistic nR2 is asymptotically distributed as X2(P) under the null hypothesis of no ARCH(p) process.04/162 0. A GARCH(l. I) model Both test statistics are reported: an Ftest for the parameters of the lagged residuals (see the same outcome as the zeroslopes test) and the nR 2 statistic. The output is given in Figure 8.OOB.Introduction to ARCH and GARCH models 197 Dependent Variable: LPCOCOA Method: ML .016339 0.00lDJ Figure S. Economically.4.944456 709. .29aJ3 0. The value I is the default value for the order of the ARCH and GARCH model in EViews.00J2 0. Ju st as before with other residual tests.U: The output after estimating a GARCH( I. Error 0. The n you can choose ARCH as the estimation method in the 'Equation Specification' window.048627 zStatistic Prob.COl) 0. Check your own estimation results.0139 Variance Equation C 0. when it can be expected that volatile situations did occur in the sample period. It is not a bad idea to use the ARCH test as one of the standard tests after estimating a regression equation.120212 . Compare thi s result with the original OLS output shown in Figure 4.105700 1.434009 0.2763 0.I) model is estimated. The null hypothesis of no ARCH is clearly rejected at the 5% (even 0.061594 0. the distribution of this Fstatistic is not known under the null hypothesis.400 0. 11.0019 O.119664 Ul64003 3. It is a nonlinear estimation procedure.ARCH (Marquardt) Date: 0311Ml4 Time: 13:59 Sample(adjusted): 1960:04 "2002m>:09 Included observations: 510 after adjusting endpoints ConV1!rgence achieved after 24 iterations Variance backcast: ON Coefficient Std.989487 0.996166 0.622657 3.04161 2 2.OO9t O. and the window shown in Figure 8. The significance of the contemporaneous and onemonth lagged coffee prices has clearly increased .912032 0.63131 0.10 appears.
.
3 and 4.4. on the basis of your economic knowledge of the subject. [n these sections. Knowledge about asymptotic properties is necessary when the sampling distribution of an estimator cannot be derived. As it is not the intention to discuss asymptotic properties of estimators . then it is not necessary to specify a complete monetary block as part of a macroeconontic model.8. endogenous or exogenous. When is an economic variable an endogenous variable? [n Section 1. With endogenous explanatory variables in the model the consistency property of the OLS . a discussion and examples concerning this topic were given. This happens with the OLS estimator when.1 Introduction Brief introductory remarks about asymptotic behaviour of the OLS estimator have been given in Sections 4. It is not necessary to specify a complete SEM if you are only interested in analysing how one economic phenomenon can be explained. for example. To summarise that discussion: we consider an economic variable as endogenous if it can be explained in another equation which belongs to a complete simultaneous equation model (SEM) that can be specified for the research topic concerned. if you are interested in explaining money demand in a country.Chapter 9 Models with Endogenous Explanatory Variables 9. Consider the explanatory variables in the moneydemand equation and deterntine their nature. only some general aspects of asymptotic behaviour will be introduced and discussed to make results comprehensible and acceptable. the classical assumptions of the classical regression model are relaxed. It is important that the consequences for econometric practice are clear.in detail. we have seen that the OLS estimator is no longer a consistent estimator when an endogenous explanatory variable has been specified. For example.
then tbe following two expressions are valid: Plim n+oo (~X'X) l n = 0 1 xx P lim n+oo (~X'u) n = O. The estimator can be derived directly in a very simple way by using GLS : this will be shown at the end of Section 9. Secondly it was assumed that the explanatory variables and the disturbance term are independently distributed. It was assumed that the matrix X'X with sample moments of the explanatory variables converges to a nonsingular matrix O x x with population moments or with finite constants: ! Plim ~X'X = n+oo n O xx.7. another but consistent estimator is introduced in this chapter: it is the twostage least squares estimator (2SLS) or instrumental variable estimator (IV). it is convenient to repeat the derivation of the probability limit for /3 here:  P n+oo lim jj = /3 + P +oo ((X'X). This retrospect is a convenient introduction to the problems that are discussed in this chapter. so nXIX exists. /3 + Plim n +oo Plim n +oo (~X'u) n (9.I X'u) lim n = /3 + ~E~ = (>.200 Chapter 9. Therefore. Considering the issues in the following sections.1. The last expression is true if X contains exogenous and possibly lagged endogenous variables. The matrix O x x is a nonsingular matrix. Models with Endogenous Explanatory Variables estimator is lost. As usual.3. we look at the consistency property of the OLS estimator.1 ) = /3 + 0 x'x' O = /3. . Before discussing the estimators. as discussed in Section 4.xf ·(>'u) (~x'x) n . but no contemporaneous endogenous variables. the property is analysed by using the expression of the estimator in its sampling error: jj = /3 + (X'Xr ' X'u Remember the assumptions that have been made. The names of both estimators are of an historical nature so the estimators are introduced in an historical context to better understand how their names derived and the original procedures.
4 in Section 4. In the (classical) simple form. Notice that all the J( parameter estimates n MO n have an inconsistency if only one of the explanatory variables and the disturbance term are dependently distributed. 9. t.{3) can be derived: r.7: P nH:O lim (~X'u) oF O. Point this out in the text of the papers you write: standard errors. For the IV estimator.::A vn ({3 . as was shown in Example 4.2 The instrumentalvariable (IV) estimator If endogenous explanatory variables are included in the model. a matrix Z with instrumental variables is chosen of the same order as the matrix X with explanatory variables.1 ) because Plim ('!'X'u) oF O.1 for the sample period. This is easily seen in formula (9.{3) ~ 2 ') N ( O. This result implies that for a model with lagged endogenous variables OLS is used in a similar way as before. In that case the (K x 1) vector P lim (lX'u) is a vector with zeros only. the disturbance term and these endogenous variables are dependently distributed. The instruments are chosen in such a way that the instrumental variables and the disturbance teon are independent. its general form will be discussed which is called the 2SLS estimator. First. but that the estimation results have to be interpreted as asymptotic results.The instrumentalvariable (IV) estimator 201 The independence of the explanatory variables (X ) and the disturbance term (u ) is crucial for the consistency property of the OLS estimator ({3). v'n(fj . the limiting distribution of the 'rescaled' fj. n The OLS estimator is no longer a consistent estimator of the parameters. This estimator can always be used in situations when the OLS estimator is not consistent. then in the following section. In this section the IV estimator is introduced as a consistent estimator. The IV and 2SLS estimator are si milar estimation methods.and Fvalues are asymptotic. n+oo n As the limiting distribution of {3 collapses in {3. A different estimator is necessary to obtain consistent estimation results. The matrix Z is an (n x K ) matrix with instrumental variables with the following properties. The instrumental variables and . the IV estimator is introduced in the classical histori cal way.aJ!xx '  then the asymptotic distribution of {3 is given by:  The covariance matrix a~~ n xlx is approximated by i1~(X'X) . instrumental variables are necessary.
2) the explanatory variables (X ) and the dependent variable (y ) are weighted with the matrix Z. Thus the IV estimator {3IV of {3 is indeed a consisten t estimator. Rewrite this estimator in the expression with {3 and its sampling error. O. . expressed in y : i3IV = (Z'X r 1 Z'y.!. P lim i3IV = {3 n+oo + P lim n+oo + ~!~ + Plim n+oo ( (Z' X r 1 Z'u) = {3 (>'Xr (>'u) (. Determine the probability limit for increasing n .zlu) = O.ZIX) .l Z' (X{3 + u ) = {3 + (Z'X r Z' u .2) In expression (9.ZIX) = fl z x . Next consider the following linear estimator.l Z'y 1 = (Z'X ). Plim ('!'Z'u) n n+oo n = {3 = = ~ {3 + fl z~ . Finally.l . Models with Endogenous Explanatory Variables the disturbance term are independently distributed. So: P lim (. It is not difficult to see that {3 I V is a consistent estimator of {3. {3.!.ZIZ) = fi zz.!. n +oo n The sample moments of the instruments have to converge to finite constants for increasing n.!. ~ (9. n +oo n The (K x K) matrix fi zz is a nonsingular matrix . Plim (. Plim (.202 Chapter 9 . This implies that the instruments are exogenous or predetermined variables. the instrumental variables and the explanatory variables are correlated. i3IV = (Z'X ). n+oo n The (K x K ) matrix fl zx is also a nonsingular matrix .
. If the correlation is high. is: Y. However. Example 9. The two parameters /3) and /32 of the consumption function are inconsistently estimated with OLS .7 as: Cov (Y" U. are dependently distributed.. A possible choice is to use. The consumption (C. An example is given below to illustrate the use of the IV estimator. u. then a high sampling bias can still be expected because of the resemblance of the variables. The covariance between Yi and Ut was computed in Section 4. As an instrument variable we need a variable that correlates with income but is not endogenous.=C.+ U. = 7'~+ /3) 1132 1 I 1/32 . and". ) .) . .The instrumentalvariable (IV) estimator The asymptotic covariance matrix of f3IV can be derived as 203  In practice. However. and investments (I. A simple slmultaneOUl equation /lwelel The simplest Keynesian macro model that has been used frequently as an illustrative example in this book is: c. It as an instrumental variable for Y. in spite of this problem. Y.+ft.) are considered as endogenous variables.1. A difficult problem is finding variables to play the role of instrumental variables. for example.EY. The reducedform equation for Y. The matrix Z is a (n x 2) matrix. the estimator can often successfully be applied. This implies a high price for consistency. . That bias goes to zero when n increases because the estimator is consistent. When correlation between the instruments and the explanatory variables decreases.) = E (( Y. In the first column we maintain the constant term. 1 132 This equation shows that Y. this covariance matrix is estimated by . + u.) = 1 0_u 2 i32 . it is simple to see that the standard errors of the estimates increase. but they should also be highly correlated with the explanatory variables.) (Z'Z) (X'Z). = /3) + /32Y.Var (fiIV) = i1~ (Z'X ). the parameters of the consumption equation can be consistently estimated by using the instrumental variables method.) and income variables (Y. ) is considered as an exogenous variable in this model. You have to be sure that they are independent in the limit with the disturbances.
3 The twostage least squares (2SlS) estimator The 2SLS estimator is discussed in two ways: first a historical but long description of the estimator is given.."" n z~· Plim n . This problem can be solved. I. 9. which explains the name twostage least squares. and secondJy a simple direct derivation of the estimator is given. we could also use. In Example 9 .. 12 h . Historical introduction When more variables are available that can serve as instruments. to some extent. Yi .. for example. Y2 Y3 . Models with Endogenous Explanatory Variables Then the matrices X and Z are: 1 1 1 Y. A problem of using the IV method is that another computational result is obtained with every different choice of instrumental variables. by using the twostage least squares estimator..l as an instrument. X= Z= and then it follows for the IV estimator: Plim n . then each choice of instruments yields different (consistent) parameter estimates.204 Chapter 9."" So this choice of an instrument shows once again the consistency of the estimator. In such a situation.. The optimal linear combination is obtained by a regression of the endogenous explanatory variables on the possible instrumental variables. . If you are only interested in the mathematical expression and the consistency property of the estimator then you can skip the 'Historical introduction' and go immediately to the direct derivation.. 1 1 1 I. we can take a linear combination of tbese variables.
we shall see that we can forget the two steps.~). are orthogonal (see Section 5... The variable Yi2 is an exact linear expression in the exogenous    variables. and the estimator will be derived in a direct way. as has been done in econometric software packages. (9. + /h Y'2 + 'Y. one formula is derived to compute 2SLS estimates. which clarifies the name. X n + 'Y2 X'2 + ut..3) u. are variables that can be used as instrumental variables. whereas only one instrumental variable is needed for Yt2.) and the explanatory variables are independent. + /he. and Y'2 . First. This means that consistent estimates of the parameters have been obtained in two stages.X" + 'Y2 X'2 + (u. Assume that X'3 and X.3): Y" = (3.The twostage least squares (2SLS) estimator 205 that is a reducedform regression.4) into equation (9. Historically. the estimator will be derived according to the two steps that have been described above. substitute equation (9. and the predicted values Yi2 have nothing to do with the disturbances Ut. Consider the following model.. This twostep procedure is presented first in this section. However. (9. but the idea behind the 2SLS procedure and its name will become clearer by describing the two historical steps of the procedure. and exogenous variables X" and X'2: Y" = fh + /h Y'2 + 'Y. The variable Yi2 and the residuals e.4) In the second stage. and all the instruments have been used by taking the optimal linear combination of them via the reducedform equation. these estimates can also be computed with one formula that can be programmed. Secondly. ~ NID (O.). Later on. Consistent estimates were computed with a twostep procedure. This will be shown in two ways. The 2SLS estimator will be introduced by using the following example. 2SLS was not one procedure for computing the estimates. with endogenous variables Yi.  (9.Then in the first stage a reducedform equation for the variable Yi2 is estimated with OLS: Yi2 = ii't + 1T2X" + 1T3X'2 + 1T. 5) Next equation (9. . In the notation {35 are used as parameters of endogenous variables and "IS as parameters of exogenous variables. X'3 + 1T5 X'4 . + /he. OLS was applied twice.2).5) can consistently be estimated with OLS as the disturbance term (u. or written in the notation with the residual variable as:  Y'2 = Yi2 + e.
. is an endogenous explanatory variable the equation is a real reducedform equation: ~ X ti = Oli + a2iZ2t + a3iZ3t + . .' ii 2. endogenous andlor lagged variables...' Z'x " (Z' Zr ' Z'X2... expressed in the originally observed variables . .. . .. (9. (Z' Z). . .....XK) = (Zii" Zii 2. + 1Z" + . If the X . Define the matrix X as the (n x J() matrix whose columns are the vectors Xi and rewrite that matrix in the following way: ~ X = (X"X2 .' Z'X ..lZ'(X" X2". Models with Endogenous Explanatory Variables The first stage The model is: y=XI3 + u u ~ N (O . .. predicted values are computed. . . . The matrix Xcontains the predicted values from the estimated reducedform equations. with K < p < n. . the 'reducedform' equation is written as: ~ X " = 0 + OZ'2 + OZ'3 + . K ): ~ The vector X i is an (n x 1) vector with observations of X ". + (lpiZpt ) and if X" is an exogenous or lagged explanatory variable. + O with Z'P Zti = X ti .206 Chapter 9. .. with the (n x K ) matrix X with explanatory variables consisting of exogenous. and according to the equations above they are wriJten in matrix notation as(i= l ... The first step is to estimate K reducedform equations for all the explanatory variables expressed in the p instrumental variables. Afte r the explanatory variables have been regressed on the instrumental variables. a~I). . XK) = Z (Z' Z). Further we have an (n x p ) matrix Z with instrumental variables. ii K) z'x .6) The regression coefficients iii from the regressions of X li on the instrumental variables Ztj are in the columns of the matrix (Z' Z). Zii K = Z (ii ) " = Z ( (Z' Z).' Z'XK) = Z (Z' Z).
This has not changed the specification of the model. ~ y = X{3 + (E{3 + u ) .8) This expression (9.l Z'y . the expression (9. (9. the reduced form residuals. Multiplying this model with the (p x n) matrix Z' with instruments at both sides of the equation yields the 'weighted' model: Z'y = Z'X{3 + Z' u.6): jjlV = (X'Z (Z' Z).The twostage least squares (2SLS) estimator 207 in X and Z. jjlV = (X'x) X'y. X = X + E. The original model is: y = X{3 + u.8) of the 2SLS estimator can simply be computed in one step. the regression of y on the predicted variables in X is performed. is included in the disturbance term by the substitution. (9. in the next stage. again with OLS. as the difference between the observed and predicted values of the variables.7) or written in the originally observed variables by substituting expression (9. ~ where E is the (n x K ) matrix with OLS residuals in the columns from the K regressions in the first stage. (1~In) . Substitution of X in the model gives the following equation. The direct derivation of the 2SLS estimator Alternatively. The second stage ~ In the second stage.9) . ~ Then OLS yields consistent estimates because X is orthogonal with E and is independent of u . with U ~ N (0.I (9. .I X' Z (Z' Z).8) can conveniently be derived as an application of GLS . That makes it possible to write the 2SLS estimator in one formula that contains observed variables only. just as shown before in the simple example: y = X{3 + u .l z'x) . as is shown below.
1 If the number of instrumental variables is equal to the number of explanatory variables (p = K ) then Z is an (n x K ) matrix.I In this way.k!1zx = {3 r l !lzx !lz~ . 0 = {3. Then the estimato r can be derived as above with the name IV estimator. Remark 9. the original twostep procedure is no longer recogni sable. Models with Endogenous Explanatory Variables with the foll owing covariance matrix of the transformed di sturbance term Z' u : Var {Z'u ) = Z' Var {u )· Z = Z' '<7~ I. with rank K. for simplicity. and all the matrices in (9.2): JiIV = {Z (Z (X'Z)I (X'Z){ 'Xr 'Z) Z'Z)I Z 'y = { Z'Xr ' Z 'y.8) are square nonsingular matrices. So the transformed disturbance term Z is distributed as: 'u Z ~ N (O . l . <7~ (Z 'u 'Z)).8) equals the rv estimator (9. The parameters of equation (9...I Plim {3IV = {3 n+oo + Plim (>'Z) G z'zr n +oo G z'x ) x (>'z) G z'zr (>'u) + (!lzx !l. .208 Chapter 9. and the 2SLS estimator (9. Perhaps.8): JiIV = (x'Z(Z z'x 'Zr l r X'Z(Z'Z)IZ'y. l . it would be better to forget the hi story of thi s estimati on method. It is simple to see that {3IV is a consistent estimator of {3: ~ JiIV = {3 + (~ x'z) (~ Z'Z) I (~ z'x) x ~ (>'Z) G z'zr G z'u) . Z = <7~ (Z'Z) .9) may be estimated with the GLS e stimator which also results in expression (9.
1 0). as di scussed in Section 5.' are obtained from the regression in the 'second round'.t. 11 ) The problems are caused by the fact that we have to deal with two residual variables.jii (. Look at the definiti ons of the residuals: e2 = y . problems exist when you want to use the Wald Ftest. to test g restrictions on parameters.fJ) can be derived and looks like: This gives the asy mptotic di stribution of fJIV : For the sample period.S) / g S/ (n.7. so they are orthogonal to the explanatory variables.. in its 'alternative fonn': F = (SR .. t= l and with the TV residuals denoted by el v. and e2. are the correct residuals as defined in equation (9..6/V .The twostage least squares (2SLS) estimator 209 Asymptotic standard errors The asymptotic distribution for . These residuals have not been obtained after a regression and so they are not orthogonal to the explanatory variables.XfJIV. computed as: (9.n_ K _ 1 "' 2 n ~ e / v. denoted by elv. The residuals e2.. .K ) (9.' . The residuals el v. 10) Testing restrictions on parameters After the parameters have heen estimated with 2SLS restrictions on them they can he tested in an identical way as to when the parameters were esti mated with OLS. However. the variancecovariance matrix is approximated and computed as: u~ (X' Z (Z' Z).l Z'X ).l with: .
The name for 2SLS is TSLS in EViews.R = Y . : log per capita real expenditure on gasoline and oil. an estimation method is selected after the model has been specified in the ' Equation Specification' window.4 IV and 2SLS estimation in EViews In EViews. IfTSLS has been selected. thi s is not valid with the IV residuals e[V. Models with Endogenous Explanatory Variables which imply that: X ~. we cannot compute (9.R of the restricted equation: ~ e[V. e2 = O. 2SLS estimates are computed if the list of instrumental variables is longer than the number of explanatory variables.X /hv. 11 ). where the orthogonality property has been used .19). See the EViews User's Guide for more details. 11 ). The consequence is that we get different expressions for sums of squared residuals in the application of the Ftest in its alternative form. See Section 5. X'2 : log of real price of gasoline and oil.8 for the derivation of the Fstatistic.R have to be used. However. As an example a data set is used from Johnston and DiNardo ( 1997). so it is not possible to proceed with the equation above.• 210 Chapter 9. but that is beyond the scope of this book. X'4 : log of miles per gallon. for the period 1959(O J ). . . X" : log per capita real di sposable income. 9.. If we repeat that derivation here. the list of instrumental variables has to be specified (the matrix Z). after which the test is performed.1990(04): Y. In more advanced textbooks one finds alternative expressions of quadratic forms of the residuals that can be used in a similar way as (9. " and specify the restriction(s) in the 'Wald Test' window.19 it was possible to square the equation without getting the cross product on the righthand side of the equation because the orthogonality condition X ' e = 0 is valid. It concerns the following quarterly data. Therefore. then the residual s e[V. the relationship between the residuals of the restricted and unrestricted model is written as: In equation 5. 'Coefficient Tests' and 'Wald Coefficient Restriction s . Similar to equation (5. In EViews we test restrictions on parameters in the same way as before: select in the 'Equation ' window: 'View'.
985789 0.00m Tim.1.1265S0 0.032494 0.mxl Mean dependent var S.326367 5.tistic) 7.985202 0.118535 5.173:176 0.034149 0.D.'". Next the equation is reestimated with 2SLS. In the output window the IV residuals are used to compute the sum of squared residuals S S R and the standard error of regression <Tu: X t . the influence of the lagged price has disappeared.0001 0.: 13:26 Sample(adjusted): 1959:21990:4 Included observations: 127 after adjusting endpoints Variable Coefficient Sld. So the endogeneity of the price variable has apparently influenced the OLS estimates.IV and 2SLS estimation in EViews 211  Fl'r:Tr.n4 O. The model can be refined further as the lagged price is not significantly different from zero. The price elasticity has clearly reduced. )(2 )(2(1) Xl X4 Y(l) Rsquared Adjusted Rsquared S.T ' I'C"rr=' · Fr=~ .063660 0. dependent var Akaike info criterion Schwarz criterion F·statistic Prob(Fst.: 03.061427 3. Error I·Statistic Prob. 1"4 as an instrument for X t 4  SSR and = e7v.32 have been used as instruments for X t 2 .1: OLS estimates An acceptable (concerning absence of autocorrelation) model estimated with OLS is given in Figure 9.COD O .00::0 OJXOJ O. The variables X t ..031209 0. and X t .2.4936/4 10.99'7258 7. the price variable is an endogenous variable.55913 0.227326 0.025159 361 .2 and The output of this 2SLS estimation is shown in Figure 9.' Dependent Variable: Y Method: Least Squares Oat.E.063183 0.505971 0.t L t= l n 1 n"o"O _ K L t= l n e7v.014419 0.594380 ·5. of regression Sum squared resid Log likelihood DurbinWatson stat e 0.t· • ..1'1'ldl". The window to specify the model and the matrix Z are given in Figure 9. rprr'"' ~~ .2431 2.co:mJ Figure 9.2 .0000 0..n IUIiI .761041 0. whereas the other parameter estimates have hardly changed.· fdlill. whereas some doubts exist concerning 'logs of miles per gallon' .648618 0.187604 0.3.346005 0.~tol.2Hm2 5.450009 1678. However.489836 5.
oo:xJ O.OOll Mean dependent var S.007902 0.1054 0.4 after adjusting endpoints Instrument list: e X2(1 TO 4) X3 X4(I) Y(·l) Va.212 Chapter 9.D.078565 0.11D!4 0.1B3460 0.649616 0. e X2 X2(I) X3 X4 Y(·1) Rsquared Adjusted R·squared S.015163 1336.2: Specification of the model and the matrix Z with instruments Sample(adjusted): 1960:1 1990:4 Included observations: 12.999709 5.943203 0.994912 9.003199 1.631622 0. dependent var Sum squared resid DurbinWatson stat ·7.OCOOlJ 0.oo:xJ O.128188 0. of regression Fs1atistic Prob(F·statistic) 0.3195 O.047388 4.962020 0.3: 2SLS output .542325 0.138944 0.E.able Coefficient Std.On211 0.810 O .036729 0.065333 3.754900 0.027132 2.962751 0.342730 O. Error I·Statistic Prob.077189 0.056331 Figure 9.00)2 0. Models with Endogenous Explanatory Variables ~I 1992:1 Figure 9.
• • Compare the OLS and 2SLS estimation results in an econontical and statistical way. a model has been established with OLS as an estimator for the parameters. then lagged (predetermi ned) variables can be used as instruments.? In Case 5. • • Determine the nature of the explanatory variables of your econometric model. etc. Are they all predetermined? So was it correct to use the OLS estimator? [n case you find that endogenous explanatory variables are part of the model .Case 6: consistent estimation of structural models 213 9. the parameters of a SUR model have been estimated with OLS . estimate the parameters of your model in a consistent way by using the 2SLS estimator and check whether the specification of the model can be maintained. . Do you consider the new (consistent) estimated elasticities are more realistic. the results are reconsidered. [n this exercise.5 Case 6: consistent estimation of structural models In Case 2. Reestimate these parameters with 2SLS also and compare the results with the early obtai ned results. Use the following reference points to adjust your paper once more. • Write a comment on your previous paper using the results from this case. If no new economic variables are avai lable to serve as instrumental variables.
.
So in a SEM we have to deal with endogenous explanatory variables that are explicitly specified in a structural equation.5 and applied in Case 8 in Section 10. But do we know what we are estimating? Is the specification of each equation unique in the system? This is an important aspect of the SEM: the identification of its equations. and fullsystem (fullinformation) methods. is the subject of this chapter. other but consistent estimation methods will be considered when estimating the parameters of the structural form of the model. which results in inconsistent OLS estimates. An equation is identified if no linear combination of other equations exists that produces an identical equation. Each equation has to be unique in the model.1 Introduction A second multiple equation model. It is a multiple equation model. These concern singleequation methods. The identification problem is discussed in Section 10. We know that endogenous explanatory variables are correlated with the disturbance terms in all the structural equations of the SEM. Using all the information from the entire model results in more efficient estimates.Chapter 10 Simultaneous Equation 10. With an estimation method such as IV. Homogeneous linear restrictions on the parameters will be necessary for the identifiability of the equations.3. such as 2SLS. we are able to consistently estimate the parameters equation by equation. Therefore. called the simultaneous equation model (SEM). methods will be discussed to check the identifiability of the equations in a straightforward way. sometimes together with restrictions on the covariance matrix of the . otherwise you do not know what kind of equation you are trying to estimate. where explanatory variables from one equation can be dependent variables in other equations. A SEM is a structural model formulated from the economic theory. After the problem has been illustrated.6. The estimation methods are described in Section 10.
The structural form has K exogenous explanatory variables X tk . + {J. The structural form contains the behavioural equations stemming from the economic theory.. with parameters I'ik . A SEM has aSlrucluralform. or the other way round: that in each equation some variables have to be excluded. Then the model is said to be complete. with t = 1. . .2 Mathematical notation of the SEM A clear mathematical notation of the model is helpful to define criteria for determining the identifiability of its equations.l Yt c . because that notation will be used to establish criteria as to whether or not an equation is identified. These identification criteria are applied in Case 7 in Section LO. . • • The structural model bas G endogenous variables. and every linear combination of equations results in an identical equation.. + 'YI KXtK + UtI + {J2 t Y tI + . The standard notation of the structural form of a SEM is as follows. G. + {JCC..7 where the recursive models are discussed. j = 1. . n.3. The notation is presented in thi s section.216 Chapter 10. This implies thai for identifiabitity of an equation not all the variables can occur in the equation. .2. 1) Y tG = 'YCt + {JCt Ytt + . So both i and j run from I till C. . we need G structural equations. . This last aspect is shown in Section 10.. In matrix notation. n and k = 1..C Y tG + 'Y12Xt2 + . A matrix notation for the SEM is convenient. The subscript i corresponds with the equation number and the subscri pt j refers to the jth variable. . then the identification of the equations is discussed in the next section.. eliminating . both dependent and explanatory variables.. Simultaneous Equation Models disturbances. which is introduced in Section 10. We will need a number of exclusion (or zero) restrictions on the parameters. The criteria for identification are conveniently applied when the SEM is written in matrix notation. specified with parameters {Jij. observed for t = 1.t + 'YC2 X t2 + . Because the reduced form is the solution of the structural form for the endogenous variables Ytj. .4. a reducedform and afillalform. . denoted as: Yt j. Criteria for the number of these restrictions are derived in Section 10.. + 'YC K X tK + UtC· It is clear that the equations are not identified as every equation looks like the otherequations. For a description of the final form . see later in this section ... + 'Y2 KXt K + Ut2 (10. .c Y tG + 'Y22 X t2 + . K. 10. The structural model without any restrictions on the parameters can be written like: Y tI = 'YII Yt2 = 'Y21  + {J12 Y t2 + . • . ..... + {J...
or has a zero when an equation is an identity..of Gi endogenous variables and [<i exogenous variables. .2) The disturbance tenns are assumed to be identically contemporaneously correlated in each period. Each disturbance term has the usual stochastic assumptions: This notation will not be used for the model as it is more convenjent to use a different vector notation. The vectors are defined across the variables instead of across time. which is identical to the assumptions made for the SUR model. Ut = (ex I) (Kxl ) (ex l ) Ute The vector Yt is the vector (now eliminating the variable subscript) with all the endogenous variables. Xt is the vector with all the exogenous variables at time t. .Mathematical notation of the SEM the time subscript. Xt = . n ( 10. Define the vectors Yb X t and U t as follows: 1 X t2 Yt = • UtI U. 0 ).. t = 1. and U t the vector with all the di sturbance tenns of the model.3) . observed at time t. model ( 10. I) is written as: 217 Y l = Z l<>l Y2 = + UI Z 2 <> 2 + U2 + ue· Ye = Z e <>e In equation i the matrix with explanatory variables Zi consists .2 . Bring all the variables to the lefthand side of the equation. Then the model is written as one equation: ( 10. The contemporaneous correlation between the disturbance terms is visible in the notation: ~ Ut N I D (O.
which gives the reducedform model: or in general matrix notation..'Ut : Yt = IIXt + V. . and Vt = B . [f B is not a singular matrix. 7r2 K X tK + Vt 2 7ru .. . . The (G x G) matrix Band (G x K ) matrix r with the parameters are: [n the ith row of each matrix are the parameters of the i th equation.lr. then the model can be solved for the endogenous variables. This is a similar notation for the individual reducedform equations as that introduced in Section 1.5: Yil = Yi 2 = + 7l"1 2 X t2 + 7T1 3 Xt3 + . Simultaneous Equation Models with: all a 21 a' 2 a' G " an .218 Chapter 10. a 2G n (Gx G ) • • • a GI a G2 " . a GG The diagonal elements have the same notation as introduced for the SUR model in Section 8. with II = _ B . 7r lKXt K + Vt l 7r2 J + 7T2 2 X t2 + 7l"23 X t3 + .3: a" = a~..
The reducedfonn equations are written in matrix notation (identical to the general linear model) as: Yl = X7rl + VI • Y2 = X7r2 + V2 YG = X7rG + VG. If the model is dynamic then lagged endogenous variables are among the variables Xli.4) The parameters of the reducedform model can consistently be estimated with OLS..1 1 0 . and " a vector with elements equal to I (the constant term).I Ut2 Gt a .l +Ut2 Yi =Ct+lt+Gt · The endogenous variables Ct. Bodkin and Hsiao ( 1996).jIIe 10.. The notation .B21 Yi +'Y22 Yi.Bll Yi + It Uti = 'Y21 +. as introduced in this section.1. Consider a simple macro model with the following three structural equations: C t = 'Y ll +. and }It represent consumption. Solving these difference equations results in the final form of the model. and the reducedform disturbances are function s of all the structural disturbances.j) are nonlinear functions of all the structural parameters.I. a . There is one exogenous variable G t for government spending. and one predetermined variable.B21 1 It Yi a a a . [n matrix notation the structural form of the model is written as follows.. with X = (" X 2 X3 XK ). This example is based on an example in [ntrilligator. the lagged income variable Yi. implying that some of the reducedform equations are difference equations. An example of this matrix notation applied to the equations of a simple macroeconomic model is given in Example 10. is a standard notation in most of the existing econometrics Uterature. which sbows the dependency of all the structural disturbances and all the endogenous variables. Ex. as always.wmlc mode. investment and national income respectively.1 A m. It. or more usually the entire reducedform model as introduced above is: Yt = IIXt+ v t ( 10.Bll 1 1 Ct a + 'Y22 .1 1 Uti Yi.croeco.Mathematical notation of the SEM 219 The reducedform parameters (rr.
IS: = 1T32 = 1 1 . = + 'lT12G t + 7l'"1 3 Yt./321 c.6) (10. (10./J21 In these equations./311 . In general notation./311 .5) ( 10. whereas equation (10.t + V'3' (10.I I .!hI + (I  /321) Un + /311 U'2 I . both shortrun and longrun .2 . + "33 Y..7) has to be solved to know the influence of the exogenous variable G. yields Yot = 7T31 2 3 '(1 + 71'"33 + 7r33 + 71'"33 + . the reducedform equations are more simply written as Ct = 11'"11 I.1 .7) Notice that the reducedform equation ofY..3 + . + 'Y22 Y.!hI + /311 G. + 1 . = 'Y ll + 'Y21 1 .. aG./311 . .. as solutions of the structural model are: C.!hI J . The equations ( 10. t This equation is the finalform equation for income../311 ) + !hI G ./321 ../311 . + "33Gt! + "~3G'2 + "~3G'3 + .:.l 1 .3 + 7l'"33Vt ... Thus the impact multiplier. + 7r33 I) + "~3 Yo + "32 (G.5) and (10.220 Chapter 10..I + V'2 "31 + "32G. From this equation all the multipliers for income.11 ./311 .. . + "23 Y./32 1 '  I + /321 UtI + (I . G./311) Yo 1 . we clearly see that all the three endogenous variables are influenced by all the disturbance terms of the structural form./32 1 1 . is obtained from equation (10. + 7r33 I VI 3 ) ../311 . Simultaneous Equation Models The reducedform equations..!.3 + 1T33Vt./311 . Y. through time.l + VtI "21 + "22 G.. = 'Y ll (1 !hIl + /311 'Y21 I .:::.'Y21 (1 . Recursively substituting back to t = 0./311 1 !h. + "h ' G I) 2 3 '(10.8) + ( Vt3 + 7T33Vt./311 . giving the effect on current income of a change in current government expeoditure.=~_.3 ./311 !h.!hI + /311 'Y22 Y.8) as: ay.10) .9) The effect on current income of a change in government spending in the previous period (10.::. ./321 ' + 'Y22 (1 .l. = Y. 1 .. is a difference equation.!hI + U'~ + U'~_ 1 ./311 .6) are finalform equations already./311 ) U'2 1 ./311 . illustrating the dependence of the disturbance terms and the endogenous variables. can be calculated.
lC.9) and ( 10. the structural equations of the model are identified if the reduced form has exactly one structural form. 10. The result is the twoperiod cumulative multiplier: Similarly the threeperiod cumulative multiplier is: etc. Finally the longrun multiplier is ("33 < 1): "32 1 .lc.B . To demonstrate this concept mUltiply the structural form : (10. unique equations have to be specified in a simultaneous equation model. However. Then the reducedform model is the solution of only one structural form model. as is shown below by solving model (10. It is necessary to check whether each equation in the model is identified."33 In Chapter 12. again in other words.lCrXt + B . in other words.' CrXt ' CUt = . Or. This reduced form belongs to an infinite number of structural forms as every choice ofa nonsingular matrix C yields another structural form with the sarne reduced form. otherwise you do not know what you are working on.B .3 Identification From the economic theory.12) has exactly the sarne reduced form as the original model (10.12): • Yt = .(CB).lUt.1CUt = .11). A structural equation is identified if the equation is unique in the model or. . 10) gives the effect of a change in government spending over both the current and the previous period. longrun and shortrun effects are discussed in more detail.lrXt + (CBr + B . 11) with a (G x G) nonsingular transformation matrix C_ The transformed structural model (10.Identification 221 Adding ( 10. an equation is identified if no linear combination of structural equations exists that is identical to that particular equation.
1 . we look at the consumption equation. 12).1 do not occur in the consumption equation.C13) yt Identical restrictions with respect to normalisation and exclusion of variables have to be imposed on this transformed equation. G t and Yt . If these restrictions result in the situation that the only allowed transformation matrix C is the identity matrix l a. 1 + CI3 C'3 = 0 = 0 = 0 C12 ')'22 (10. (C11 ')'ll . e ll C12 CI3 :::.1 That now results in the following transformed consumption equation: (C11  C13) Ct = + C12'Y21 ) + (CI2 + C13) I t + (Cll~1 1 + C12~1 + CI2')'22 Yt 1 + Cl3Gt + (CllUtt + C12Ut 2) . Next. As an example. so the structural model is unique and therefore identified. their coefficients are zero. to ensure that the first equation is a consumption function as originally specified.13) = ( Cll C'2 CI3)= (1 0 0) . the investment equation can be checked to see whether or not it is identified. then the consumption equation is identified.222 Chapter 10. The transformed parameters of the consumption equation are: ( Cll C12 Ct3) o . The coefficient of Ct is equal to I.1 1 0 1 . First. This is the first row of the identity matrix 13 • Notice that the constant. These conditions are required for the transformed parameters too. Thus the reduced form belongs to just one structural form.11) (exclusion of variables and normalisation) must also be imposed on the parameters of the transformed model (10. o . the identificatlon of the structural equations of the macro model of the previous section is checked. The transformed coefficients are computed from the multiplication given below. The variables It. the parameter of yt and the disturbance term are identical to the original ones. Simultaneous Equation Models the restrictions on the parameters of the original structural model (10. In a si milar way.1 1 0 1 1 ')'11 o ')'22 and (Cll CI 2 ')'21 o o o o . If it is possible to show that the first row of the matrix C is the first row of the identity matrix. then the model is identified. the elements of the first row of the matrix C can be determined.1 ~ll ~21 o and (C21 ')'22 1 o o o .
when two equations should have identical restrictions. Then the equations are not identified. The number of equations for the conditions (10.811 + c22. which is exactly the necessary number of equations to solve for C2j .1 + cnG t + (Cll Un + C12Ut2) • (C21/11 .14) = ( C2 1 C22 = (0 1 0). Three variables have been excluded to get an overidentified consumption equation.B. The number of equations for the conditions (10. For that reason.C23) Y.14) that specify the identifiability conditions? The number of exclusion restrictions are sufficient to establish that both behavioural equations are identified. This result can be generalised and formalised in two conditions which are called: the order condition and the rank condition. A necessary and sufficient condition is the rank condition.cn) It = + C22'Y2. This order condition is a necessary condition but is not sufficient because identical restrictions can occur in different equations.C23=1 C21 + C23 C23) = 0 cn = 0 (10. we call the consumption equation overidentified and the investment equation exactly identified. This is the second row of the identity matrix 13 . The order condition From the above example it is clear that a necessary condition for an equation from a Gequation SEM to be identified is that it has at least G . also contributes to making an equation unique in a model.83. derive the conditions so that the equation is again an investment equation as originally specified. A restriction like . C22. for example. 2.1 4) is three for the three unknown coefficients C2j (j = 1. What do we observe when we look at equations (10.Identification 223 The transformed investment equation is: (C22 . in this example with three endogenous variables and three equations we need at least two excluding restrictions for identifiability of each of the equations. We observe that two variables have been excluded to obtain an exactly identified investment equation. The reason is that restrictions in more than one equation can be 'dependent'. whereas the number of restrictions in each equation is sufficient for identification. So. Next. .82 + .1 exclusion restrictions.3) . = . This condition can be generaHsed by extending the excluding restrictions to linear homogeneous restrictions in general : a necessary condition for an equation from a Gequation SEM to be identified is that it has at least G . An exclusion restriction is a linear homogeneous restriction.821 + C22'Y22 Yt . or written as .13) is fourforthe three unknown coefficients C'j (j = 1.1 linear homogeneous restrictions. 3).13) and (10. which is one equation too much.83 = 0 . 2.) + (C21 + C23) Ct + (C21. The third equation is an identity and is always identified because no parameters have to be estimated..
224
Chapter 10. Simultaneous Equation Models
The rank condition
A necessary and sufficient condition for an equation from a Gequation SEM to be identified is that the rank of the matrix with parameters formed by the restrictions in the equation under study has rank G  1. U this condition is fulfilled, then dependencies, as mentioned above,
do not occur.
To summarise the conditions for identifiability we see that three situations are possible.
Denote the number of linear homogeneous restrictions in equation i with ~ , and lile matrix with columns that represent these restrictions with A i. then the following proposition has been made plausible.
Ri < G  1 =* R.; = G  1, and " (A i ) = G  1 =* R.; > G  1, and r( A ;) = G  1 =*
equation '; is not identified, equation '; is exactly identified, eq uati on i is overidentified.
The order condition can be rewritten in an interesting way. The order condition says that eq uation i is identified if:
R.; > G  1.
The number R.; is the number of excluded variables, so as (K  K ,) and (G  G ;) are the number of excluded exogenous and excluded endogenous variabJes in equation i respectively, the inequality becomes:
R.; = (K  K,)
+ (G 
G,) > G  1,
( 10.15)
=* K  K , > G,  1.
On the lefthand side of (10.15) the number of possible instrumental variables that is needed to estimate equation i is found, whereas on the righthand side of the ineqUality the number of necessary instrumental variables appears because (G,  1) is the number of endogenous explanatory variables in equation i . So if the (i n)eq uality holds, the parameters can be estimated by lV or 2SLS. The distinction between exact and over identification is, in fact, historical too. In the past, we stated that the parameters of an exactly identified equation had to be estimated with lV, and the parameters of an overidentified equation with 2SLS. Those were two different methods, viewed computationally. Nowadays, this distinction is not really relevant, as computational limitations no longer exist.
10.2
condition
These conditions will be applied to the macro model again. For reasons of convenience the model in matrix notation is shown again:
o
 1
1
0
1  1
Ct
It Yt
o
+
 1'22
o
o o
 1
1
Utl
rtl Gt
Ut2
o
Identification
225
Count the restrictions: R, = 3, R2 = 2, whereas G = 3. So according to the order condition it is possible that the first equation is overidentified (R, = 3 > 2 = G  I) and the second equation is exactly identified (R2 = 2 = G  I ). So check the rank
condition for the consumption and investment equation:
o
1  1
0
"122
0 0
 1
1
, A2 =
0
o
 1
0
 1
0
Here it is not difficult to recognise that r (A, ) = 2, and that ,. (A 2 ) = 2, so the first equation is overidentified and the second equation is exactly identified. Both structural equations can be consistently estimated by 2SLS (or IV). Lf an equation is not identified it can become identified by imposing more restrictions on its parameters or by adding variables in other equations of the model. Although the investment equation is identified, this equation will be used to present an example of how a restriction can be added to its parameters. Suppose that the following null hypothesis is not rejected: Ho : Ih.l + 1'22 = 0 1 after the parameters of the investment equation have been estimated. Then the investment equation has one more linear homogeneous restriction and can be written as:
1,
= "121
+ ,B\7Y, + U'2.
This equation only makes sense if It does not show a linear trend, because the transformation \7Y, eliminates a possible linear trend in Y, . A model that has a dependent variable with a linear trend and explanatory variables without trends has not been correctly speci fied because the trend in the dependent variable is not explained. These aspects concerning trend characteristics of an economic variable will be elaborated and formalised in Chapter 12. The concept of order of integration of a variable will be introduced, which is useful to check the trend behaviour of all the specified variables. It is
important to mention this point here. The restriction is a linear homogeneous restriction, so R2 = 3 > 2 = G  I , and the matrix A 2 becomes a (3 x 3) matrix, because a third column is added for this extra linear homogeneous restriction:
C,
excluded
G,
excluded
"
1
•
o
A2
=
o
 1
o
 1
,Bll 0
1
The rank of A 2 is still equal to 2, r (A 2 ) = 2, but in this situation the equation is called overidentified. In the original specification the equation was exactly identified.
226
Chapter 10. Simultaneous Equation Models
10.4 Case 7: identification of the equations of a SEM
This case is an exercise to practise the conditions for identifiability on the equations of a simple macroeconomic model. Consider the following dynamic macro model with a
behavioura1 equation for consumption and investments, and an identity defining income:
Gt
= 1'n
It = 1'21 Yt =
+ ,anY, + 1'12Yt l + Ut! + ,a21Yt + 1'22Y,  1 + 1'24It l + Ut2
It
C, +
+ Ct.
The variables
G" It, yt are considered as endogenous variables, and the variable C, for
government spending is considered as an exogenous variable. Then consider the following:
• • Check the identifiability of the equations of this macroeconomic model. Suppose that the government spending was constant in the sample period:
C t = C.
Check again the identifiability of the equations of the macroeconomic model with
this assumption.
• Next assume that the government spending C t is also an endogenous variable, so one more behavioural equation is added to the model, for example:
What can be said about the identification of the three behavioural equations of the new model?
•
The new model of the previous question is maintained. After the parameters have
consistently been estimated, the null hypothesis
Ho : ,8" =  1'22
was tested and was not rejected. What is the impact of this result on the identifiability of the equations?
10.5
Estim ation methods
Tho types of estimation methods are distinguished for a SEM: singleequation meth· ods and fullmodel (or fullinformation) methods. Singleequation methods do not use the information that contemporaneous correlation exists between the clisturbance terms of the complete model. They are not asymptotically efficient, but only consistent. Some
Estimation methods
227
wellknown, singleequation estimation methods are IV, 2SLS and LIML (limitedinformation maximum likelihood). The first two methods have been discussed before. LIML is the maximum likelihood estimator applied to one equation, and is called ' limited' for that
reason.
Complete system methods are, for example, FIML (fullinformation maximum likelihood) and 3SLS (threestage least squares), which are consistent and asymptotically efficient estimation methods. These two methods are asymptotically equivalent (see, for example, Greene (2000) Section 16.6 and included references). The principle of 3SLS is identical to that of the SUR model. One mathematical expression for the 3SLS estimator can be derived, but historically the three stages were considered separately. For didactical
reasons, to understand in a clear way what is done, these three steps are illustrated here in
an example. The method is applied to a twoequati on system:
Yi1 = I'll Y'2 =
+ i112 Y'2 + 1'12X'2 + Uti 1'21 + 1121 YtI + 1'23 X '3 + U'2'
Remember the disturbanceterm assumptions (l0.3) for all the disturbance terms of the model :
u,
~
N I D(O, n ), t = 1, ... ,n.
This can be written in this example as:
The first two stages consist of the computation of the parameter estimates of the individual equations with 2SLS. Similar to the SUR estimator, the variances and covariances of all the (2SLS) residual veclors of the model are computed in the third stage, which gives an estimated covariance matrix n, after which GLS estimates can be computed. Both equations are exactly identified as each equation has one exclusion restriction. A summary of the 'three stages' is given in the following scheme.

•
•
In the first stage, OLS is applied to each of the reducedform equations and the predicted values Yi1 and YiA are calculated. _ Secondly, substitute Yi1 = Yi1 + e", and Yi2 = Yi2 + e'2 in the structural equations and estimate the structural parameters with OLS to obtain consistent 2SLS estimates of the parameters:


Y tI = I'll Y'2 =
+ i112Yi2 + 1'12 X '2 + (Uti + i112e.d 1'21 + i121Yi1 + I'23X,3 + (U'2 + i121e,,).

Save the 2SLS residuals eIV," and eIV,'2·
228
Chapter 10. Simultaneous Equation Models
• Next. in the third stage, estimate the variances and covariances of the disturbance terms:
This gives an estimate of the covariance matrix 0 :
The two equations are combined in one model in an identical way as in the SUR model, with identical notation : X i being the matrix with the explanatory variables and "'i the vector with parameters of the ith equation:
y. = X . o:.
ll. ~
+ u.) with
N (0, 0 ) .
For this model, the feasible GLS estimator ii. is calculated:
The elements of the vector ii. are the 3SLS estimates of all the parameters of the SEM. The procedures in EViews to compute 3SLS and FIML estimates are similar to the procedure to estimate the parameters of the SUR model. The following must be done: select ' New Object', ' System', and type the structural equations in the 'System window'. However, be aware that identities are not allowed in the system window! A list with instrumental variables has to be included in the system. These [V variables can be specified for the entire system or by equation; see the examples below. Possibly, identities should be substituted in the other equations and these equations can be specified in the system. For example the structural equations from Case 7 are:
Ct
=
I, =
+ /1n Yi + "1'12Yt  l + Uti "1'21 + /121 Y, + "1'22 Yi  l + "1'2. [ tl + U'2·
"1'11
Case 8: simultaneous estimation of a macroeconomic model
229
When. for example, the variables G t and Yt  1 , . .. 1 Yt 4 and It  i , ... 1 ft  4 are chosen as instrumental variab les for the entire system, the 'System' window looks as follows:
CONS = C( l ) + C(2 ) INV =
* Y + C(3 ) * Y( l ) C(4) + C(S) * Y + C(6 ) * Y( l ) + C(7 ) * INV( l )
INST GOV Y(  1 to  4 ) INV( l to  4)
II is possible 10 slay closer 10 each equation by specifying IV variables by equation, by separating each equation and its IV variables with the @ sign, for example in the following way:
CONS = C( l )
+ C(2) * Y + C(3) * Y( 1)
<lI CONS(l to  2) INV (  1 to  2) GOV Y( l ) INV = C(4 ) + C(S)
* Y + C(6) * Y( 1) + C(7) * INV( 1 )
<lIY( lto  3 ) INV( l )
Next, selecl the estimation method 3SLS or FIML under the 'Estimate' button (see Figure 8.2 for an example of that window). More information and advanced information about the use of a system is found in the EViews User's Guide .
•
10.6
Case 8: simultaneous estimation of a macroeconomic model
With the knowledge obtained in this chapter, we are able to estimate the pararoeters of the structural equations of one of the macroeconomic models simultaneously. Some suggestions are given as follows :
• •
• •
•
Choose one of the countries from Data set 2 with macroeconomic data. Specify a dynamic consumption and investment equation and check the identifiability of the two structural equations. Estimate the parameters of the two equations with 2SLS. If Data set 2 had been used in Case 2, one of the equations has already been estimated. Open a new object 'System' and specify both equations in the system window, in the first instance according to the previous individual results. Specify the instrumental variables according to one of the two mentioned possibilities in the previous section. Alternatively, estimate the pararoeters twice in both manners to see the effects of this choice on the results. Click the button 'Estimate' and select 3SLS as the estimation method.
230
Chapter 10. Simultaneous Equation Models
• Check the resulls in an economical and statistical way,just as was done with the SUR model in Case 5. Are the estimation results realistic and do the residuals obey the underlying assumptions? Compare the fullsystem resulls with the singleequation results. Finally, write a paper about the estimation results and all other relevant findings.
•
10.7
Two specific multipleequation models
In this section, some attention is given to recursive models. Two forms of recursivity in a SEM can be distinguished: simplerecursive models and blockrecursive models. These are models with specific restrictions on parameters and covariances of disturbance terms. The simplerecursive model is more restrictive than the blockrecursive model. These two models also playa role in tbe history of econometrics. The simplerecursive model is, in fact, a historical model. In the past, wben the computational possibilities were limited by lack of computers, the simplerecursive model was an interesting structure for a SEM. Everyone was happy wben OLS could be used by equation, instead of full SEM methods. Nowadays, these computational restrictions no longer exist, and many have doubts about the reliabi lity of the assumptions underlying this model and its practical usefulness . The blockrecursive model is still useful, as in large SEMs the number of predetermined variables can become too large, larger than the number of observations. Then computational problems arise when estimating the parameters of reducedform equations in the first step of the 2SLS procedure. Or, in other words, when only talking about the mathematical expression (9.8) fo r the 2SLS estimator, si ngular matrices appear in that expression. This problem is overcome by splitting the model into a number of simultaneous blocks. The two models will be described brieHy in this section. As these models can still be encountered in the literature it is good to know their features.
The simplerecursive model
The SEM is called a simple· recursive model if in the simultaneousequation model:
BYt + rXt = Ut Ut ~ NID(O , n ), t = 1, .. . ,n,
the matrices B and n can be restricted in the following way:
1 (h,
(G x G )
0 1
...
•
.. .
.. .
0 0
Ull
0
U22
0
0 0
B
,
(G x G )
n
• • • • •
{3G1
{3G2
...
0 1
0
0
.. . .. .
0
UGG
R. according to the order and rank condition. . This can be made clear with the following example based on an example in Johnston (1984). with Uti ~ N I D (O.CI2. Consider the model: Yn Yt2  'YllXtl = Un !hlYtl 1'Z lXU = Ut2 . it is possible to show that the equations are identified. which is caused by the extra restrictions on the covariance matrix n of the disturbance terms. The disturbance terms of the SEM are assumed to be independent.. Wben checking the identifiability of the equations we see that only the first equation is identified according to the order and rank condition: RI = G .C12.Two specific multipleequation models 231 The matrix B is triangular and n is a diagonal. The transformed equations become: (C I1 . The model is written in matrix notation as: 0) ( ~21 1 (Y'I) _ ("111) (Xtl) = Y'2 "121 (Uti) . Then the matrix C looks like: . Using the matrix: C = ( CII C21 C 12) C22 as a transformation matrix and using the additional restriction on the covariance matrix 0 .1 for i = 2.62J)Yn + CI2Y'2 + C2 2Y'2  (CI1"111 (C21"111 + C12"121)Xn + C22"12tlxn = CI1 Un = + CI2Un.62J)YtI (C21 .621 = 1 CI2 = 0 C22 = 1 which implies: CII = 1 and CI2 = 0 and C22 = 1.. C21U'I + C22 U '2 · The transformed coefficients have to obey the same restrictions as those of the original equations. .1. so this gives the following conditions on the transformed parameters: CI1 .C22. If the model is transformed by using a transformation matrix C as before. it is possible to show that the second equation is also identified. U'2 with The first equation of this model is identifiable and the second is not. < G . G. "il) and "1 2 = O.
. and cov(ut2. This implies that: C2 1 = a. ) · C' = COC'. The parameters of the equations can be consistently estimated by OLS.But there are more 0 _  (0"1 1 a 0"22 a) . the restriction of the zero covariance must also be valid in the transformed model. Estimation of the parameters with a simultaneous estimation method is more obvious.232 Chapter 10.utI) = If you really want to use the properties of this simple recursive model you should be convinced that the strong restrictions on the parameters afthis model are realistic. that is: or 0"22 a) ( C2 1) 1 = a. Of course.XtI) =a. or: C21Ul1 = O. because cav (Ut2. This matrix must obey the restriction that the nondiagonal elements are zero. The covariance matrix of the transformed di sturbances is Var (CU t). (rllXtt + Uti) ) = a. in the second equation. The first equation is a reducedform equation and. then: Var (CU t) = C · ~lr ( u . Thus the only admissible transformation matrix is the identity matrix: which means that both the equations are identified. a. as 0"1 1 i' a. Simultaneous Equation Models The second row is restrictions: DOl the second row from the identity matrix 12 . Identification has been achieved by restrictions on the structural parameters and the covariance matrix of the disturbances of the model. the di sturbance term Ut2 and the explanatory endogenous variable Y'1 are independently distributed: cav (Ut2' YtI) = cov (Ut2.
Two specific multipleequation models
233
The blockrecursive model
The simplerecursive model can be modified to a lessrestrictive model, the blockrecursive model, when not all the elements in the rightupper part of B are zero. The model consists of independent blocks of simultaneous equations. This model produces 'blocks' around the main diagonal. The order of the blocks around the diagonal of n is assumed to be identical to those of the blocks around the diagonal of B. Let us look at an example to clarify this model. Suppose that we have a SEM with five endogenous variables and equations, and the following restrictions on its parameters are given:
1
!321  !312
1
!332 !342 !352
0 0
1
(5x5)
B
!331 !341
0
!353
0
0 0 0 1 0
0
0 0
 !345
O'll
0'21
" 12
" 20
0
0
"33
0
0
(5 x 5)
n
=
1
0 0 0
0
0 0
0 0
0'44
0
0
0'45
0
0
"54
" 55
The model has been divided into three blocks. Equations I and 2 are a simultaneous block, equation 3 is statistically independent of the system, and equations 4 and 5 form a simultaneous block too. If these restrictions are valid, then the identification of the equations is checked by block and the parameters of the equations are estimated by block. Identification is not checked in this example as no matrix r is considered. It is just an example to illustrate the properties of this model. Suppose that all the equations are (over)identified, then the parameters of the first two equations are estimated with, for example, 2SLS, the parameters of the 3rd equation are estimated with OLS, and the parameters of equations 4 and 5 are estimated again with 2SLS . The advantage of this mode l is that when a SEM is very large, it can be convenient to handle it as a blockrecursive model, as the number of predetermined variables of each block will be smaller than the number of predetermined variables in the entire model. This overcomes possible estimation problems when the number of observations is smaller than the number of predetermined variables in the entire model. In practice, it is not impossible that the restri ctions of the blockrecursive model are acceptable and can be applied in very large models.
Chapter 11
Qualitative Dependent
11.1 Introduction
The last specific structural model to be discussed in Part III is the model with a qualita· tive dependent variable. The dependent variable is a dummy variable. This subject can be discussed very extensively, for example, Maddala (1983) provides a very complete text. Maddala (2001) also considers this subject in his Introduction to Econometrics, 'This is a complete area by itself, so we give only an elementary introduction,' and this is done, in an even more elementary fashion, in this chapter. The dependent variable is often a binary variable. However, that depends on the formulation of the research problem. The dummy variable can take more than two values, but that is not a topic in this book. Dummy variables which have more than two values compticate this subject more than those taking only two values. In this introductory course, only binary dummies that take zero or one as values are considered. Such a dependent variable is called a dichotomous or binary variable. The necessary data for estimating the parameters of such a relationship clearly concern a crosssection. For example, look at a dichotomous variable that a person possesses a personal computer or not (PCi ), which is related to a number of relevant variables Uke income (Yi ), education (Ei ), age (Ai), etc.: PCi = PCi =
fJl + fJ, Yi + fJ3 E i + fJ.Ai + u.; .
1 if person i has a PC,
(ll.l )
oif person i has no PC.
236
Chapter 11. Qualitative Dependent Variables
The (crosssection) data necessary to estimate the parameters of the model are collected by asking a group of people whether or not they have a PC (PCi = 0 or I) and what their age,
income and education is. Education will be a qualitative variable too, for example, ranked from 0 (no education) to 10 (highest level of education). If such a relationship has been estimated, tben it will be possible to predicI whether a person owns a PC given his or her
income, education and age. It will be clear thaI the variable PCi does nOI have a normal distribution, 10 mention juS! one of the problems. Model ( 11.1 ) is also known as a discrete regressioll model. More Iypes
of these models exist, even for timeseries data. In cases where the parameters are estimated
with OLS the model is called the tillear probability model. The /illear discrimillant full ctioll is an example of another model which is related to the linear probability model. These models will be explained to a certain extent in Section 11 .2. One more option is to say thaI there is a ' Ialent variable' Y,' underlying the process that is not observed. Only a dummy variable Y; is observed in the following way:
Yi = { 0,
I , ifY,' > 0,
otherwise.
This idea results in what is known in the econometrics literature as the probit and logit models. These models are discu ssed in Section 11.3. Sufficient introductory knowledge about a number of these models will be obtained from thi s chapter, to make it possible to apply them at an elementary level. Applications are presented in Section 11 .4, where the use of EViews is described to estimate the parameters of a discrete regression model .
11.2
The linear probability model
The linear probability model, mentioned in the previous section, model (ll.l), denotes a model with a di chotomous dependent variable (P C;) which takes the values I or 0 (yes or no):
with E (Ui) = O. This implies that:
For notational convenience we consider one explanatory variable X i 2 in this chapter to explain the models:
E (PCi ) = {31 + /h X i2 .
The linear probability model
237
The expected val ue of a qualitative variable like E (P Ci ) has an in teresting interpretation wi th respect to probabi lities. Define the probability Pi as the chaoce that person i has a PC:
Pi = Prob (PCi = 1) .
But realise that PCi can only take the value 0 or 1 which means thal: Prob (P Ci
= 0) = 1 
Pi .
Next, compute the expected value of the binary variable:
E (P Ci )
= 1 . Prob (PCi = 1) + o· Prob (P Ci = 0)
= I · Pi + O · (I  Pi)
=
Pi.
The expectation of the binary variable E (PC i ) is simpl y the probability that person i has a PC given the values of the exogenous variables:
P,
=
(3,
+ f3, X i2 ·
This model is called the linear probability model. The equation can be estimated by using OLS . With the econometric model , predicti ons of the probabilities can be computed given particular values of the ex planatory variables. However, this model has a number of problems. It is clear that the value of a probabijity is in the interval (0 , 1). That can go wrong with the predicted values PCi . PC, is the estirnated probability that the event of the ownership of a PC occurs. But it is clear that these predicted values cao lie outside the interval (0 , 1). Further, we have a problem with the distribution of the disturbaoce term. The disturbances of this model caonot be normally distributed. They must have a binomial distribution. Because P Oi is equal to 0 or 1, the disturbance term U i can also take only two values: 1  {3,  {32X'2 and {3,  f3, X i2 , with probabilities Pi = {3, + f3, X i2 and (1  P,) = 1  {3,  (32 X '2. Recall the statistical expression for the variaoce of a random variable:
 
var(Ui) =
=
E(u, "j
IL)2)
L (u,  IL) 2 f (u;) = L (U;)2 f (u.),
" j
as IL = O Then compute the variance of u, from the linear probability model: .
CT~ • = (1  ({3,
=
({3, = E (PC,) (1  E (PCi )).
+ {32Xi2))2 ({3, + (32Xi2) + ( ({3, + f3, X i2 ))' (1 + (32X i2) (1  ({3, + f3, X i2))
({3,
+ f3, X i2))
•
238
Chapter 11. Qualitative Dependent Variables
This means that the disturbance tenn of the linear probability model is beteroskedastic and that OLS estimation produces nonefficient estimates. Efficient estimates can be computed by using a weighted least squares estimator (GLS) as discussed in Section 8.6, because the form of the heteroskedasticity is known in this model. This is done by first estimating the equation with OLS, after which the predicted values P C i are computed. With these predictions the weights W i are calculated:

Wi
=
1::=::===::::==
1
";Pci (1  Pci ) ,
and the weighted least squares estimates can be computed. To keep it simple, it is also
possible to use White's 'beteroskedastic consistent covariance matrix estimator' to obtain consistent estimates of the standard errors.
More problems can be mentioned such as standard test statistics that cannot be used in a straight manner as the disturbance tenn is not nonnally distributed: and, more importantly, as mentioned before, the computed probabilities can lie outside the interval (0,1) in many cases. Figure 11.1 shows an illustration of this problem. The variable X i2 is any explanatory variable to serve as an example for the scatter diagram. Since PCi can take the values 0 or I only, a scatter diagram will consist of just two horizontal rows of points. OLS will fit a line through these points similar to the one drawn in Figure 11.1. From this figure it is also seen that the coefficient of determination R' is likely to be low, as no straight line could fit the scatter. R' is not a good measure for
pc;
1
Fitted line
o
Figure 11.1: A scatter diagram of data for the linear probabiJjty model
Probit and logit models
239
the quality of the perfonnance of the model when the dependent variable is a qualitative dummy variable. This problem is discussed in more detail in Section I 1.3. For a more detailed discussion on the models in this chapter see, for example, the above mentioned books of Maddala or Chapter 19 in Greene (2000).
Remark 11.1 The linear discriminant function As mentioned in the introduction, the 'linear discriminant function' is another model whicb is related to the linear probability model. This function will not be discussed in this book, only the idea bebind this function is mentioned here. For example, n ob
servations on an economic variable bave been collected. 1be economic phenomenon is explained by K explanatory variables. Next, suppose that the n observations can be split into two groups, such that n, of them belong to one group and n2 belong to a second group, with n = n, + 02. If a linear function of the K variables can be found that separates the two groups in an optimal way, then this function is caIled the linear discriminant function. The discriminant function makes it possible to predict whether new observations belong to group one or group two. The linear discriminant function is the best discriminator between the two groups according welldefined criteria. The statistical analysis used to find this function is caIled discriminant analysis, but is beyond the· scope of this book.
11.3
Probit and logit models
Because of the problems that we bave established with the linear probability model different models will be used. One of the possible approaches that has been developed in the econometrics literature is to model a latent variable Y,' and to specify the model as follows. The bivariate model is used again in general notation. The model is:
Y,' = {3\
where
+ {J, X i2 + Ui,
is an unobservable variable. For example Yi* represents the amount of money that person i has spent on buying a PC. What we observe is a dummy variable Y, defined
Yi.
by:
I , ifY,' > 0, Y,= { 0, otherwise.
Denote by Pi the probability that the event Y,'
> 0 occurs:
Pi = Prob (Y, = 1)
= Prob (Y; = Prob ({3\ =
> 0)
•
+ {J, X i2 + Ui > 0) Prob (ui > ({3\ +{J, X i2 ».
. This means for the probability P. = 1) = Prob (Ui < {3.Y. +/3. then we get a model that is known as theprobil model. + /h X i2) . + {32 X '2). 1 1 + exp ({3j .P 1 1~~= Fitted line o Figure 11. With the logistic function.: P.. Thus we have the following distribution function: F({3. in that case it appears that the abovederived probability is rather complicated to work out and therefore researchers prefer to use the logistic distribution. The cumulative logistic distribution is used for u.2. > .) 1 ( I 1. = Prob (Y.{32 X i2) .2: The logi stic function . + {32 X '2) = ef3t +. Qualitative Dependent Variables if we use a symmetric distribution of the disturbance term Ui that is symmetrical around its zero mean then the following equation applies: Prob (u. 1\vo distributions are considered in this context: the normal distribution and the logistic distribution. The resulting model is called the logit model.240 Chapter 11. If the Ui are assumed to be normally distributed.2) Y. The probability Pi depends on tbe distribution of the disturbance term Ui . we have the following logit model for Pi: Pi = "~77'" 1 1 + exp ('.({3. < {3.82 X j2 1 + e/3. The logistic function can be drawn as an Sshaped curve that is bounded between 0 and 1. + /h X'2 )) = Prob (u. However. as shown in Figure I 1. A convenient notation is introduced in the last expression.X.
(PI + PZ X iZ + u.25 and further adding 0. _ .from which it follows that: (P. + P2X iZ + Ui)) 1 + exp (. Therefore.Y. then it follows that Pi 4 1. To compare the parameter estimates of the linear probability model and the logit model.Yil ) n 1 _ P. = 1 + exp (.2 and is also easily established by looking at the following limits: IT X i2 4 if X i2 4 00  with fh > 0 . then it follows that P.5 to the constant term. The coefficient of determination R2 is an inappropriate measure of goodness of fit.3) In( This result is easily verified: Pi ) = PI +fh X iZ+ Ui.Pi: 1 _ P.' ) • = Y/ = PI + PZXiZ + Ui · The logit model (11. This null hypothesis is tested with a likelihood ratio test. The resuIts of the logit and probit models will not differ very much.2) for 1 . However. Let Lu be the value of the loglikelihood function of the unrestrictedly estimated model and let L n be the value of the loglikelihood function of the . a similar property when comparing the normal and the Student's tdistribution. The standard OLS zeroslopes Ftest is not used due to its relationship to RZ.')  .') x exp (.6 = 0.4) is a nonlinear model .Probit and logit models 241 This equation ensures that Pi lies between the values 0 and 1.In 1) ( exp (.551 ) to be comparable to the estimates obtained from the probit model. and 00 with fh > 0. Amemiya (1981) proposes to multiply the logit estimates by 1/ 1.625 as this transformation produces a closer approximation between the logistic and the standard normal distribution. Amemiya suggests multiplying all the parameters of the logit model with 0.Y. its parameters have to be estimated by a nonlinear estimation method.))' exp ( ( 11. 1 .Y. as was the case with the linear probability model.Pi (11. Rewrite the expression (11. To test the specification of the model an LRtest can be used to test the null hypothesis that the parameters are zero. The cumulative nonnal and the cumulative logistic curve are very close to each other except at the tails. the esti mates of the parameters are not directly comparable. 4 O. This property is clear from Figure 11. the estimates of the parameters obtained from the logit model have to be multiplied by V3/" ("" 0. The variance of the logistic distribution is " z / 3 and the variance of the standard normal di stribution is of course equal to I.4) I ( Pi ) in ( 1 1 + exp (.
person i has a PC.242 Chapter 11.000 < income of person i: < $40.000 $40. In the EViews User's Guide.000 < income of person i: < $30. ranking in the interval 015. = 2 3 4 5 income of person i: < $10. Below an example is given. where artificial data have been used. For estimating a linear probability model just specify the equation and estimate with least squares.000 $30. 11. The variables that have been simulated are: PGi = n person i has no PC.3. In the EViews procedure that estimates logit and probit models. so it has nothing to do with the coefficient of determination as di scussed in Section 5. an extensive description of the procedures can be found. This LRtest statistic is always printed in the EViews output when a logit or probit model has been estimated with a constant term. o 1 lNG. .In (Lu )). The McFadden pseudoR2 is defined as: PseudoR = 1 . More information about these subjects can be found in the earlier mentioned references and other econometrics literature.000 $50. $20.R2 is computed and printed in the output.000 AG Ei : age of person i. The sample contains 60 observations. so it is possible to reproduce the exercise in this section.L 2 Lu R ' It is a pseudoR2 because this coefficient does not have the interpretation of a ratio between unexplained and total variation of the dependent variable Y. Then the LRtest statistic is the wellknown statistic: LR = . but they are not often used. the McFadden pseudo.000. that has an asymptotic X2 (K . PseudoR 2 s are found in the literature. .xls.2 (In (LR) .4 Estimation of binary dependent variable models with EViews Estimation of the parameters of models with qualitative dependent variables with EViews is just as simple as all the discussed estimation methods from previous chapters. Qualitative Dependent Variables restrictedly estimated equation that has only the intercept as regressor.000 $10. These artificial data are available at the website in the file pcdata.000 < income of person i: < income of person i: > $50. ranking in the interval 188 1.1) distribution under the null hypothesis of zeroslopes coefficients. EDUGi : number of years of schooling after primary school. but remember all the mentioned problems with respect to the estimation result.000 < income of person i: < $20.
stic Prob(F9tatistic) Figure 11.3.1.023636 0. ~.! '11).. actual and fitted values have been plotted.576157 0.54.054552 3.4: Predicted ' probabilities' of the estimated linear proba bility model . .43 I : 1 59 .:~~1l Dependent Variable: PC Method: Least Squares Date: 07131.Fitted Figure 11.224211 25. but not aU results are relevant.4.58895 1. with two extreme negative values of . In Figure 11..737575 3...· 243 nFr:.183E192 0.!:00261 0.249640 0.320461 1.46222 O.5 1. The option 'Heteroskedasticity Consistent Coefficient Covariance' has been used to work with the heteroskedastic disturbance term. C AGE EOUC INC Rsquared Adjusted Rsquared S.~III.OO))l) • Mean dependent var S.. Variable Coefficient Std.1IDl 0.'.D.. t '.1"".0. Ii/ 1'1 1.1XD4 0.003015 1.687:JJ9 0. although the source of the heteroskedasticity is not unknown here.0 Residual Actual . where the residuals.0016 0.5lIl64O 1.I1AII\ r~t"""" ·· r:"'7r=~ 'r'""~r"""~ 3.1550 0.020056 0.00 Time: 12:.441843 4. II... ~~~~I'i"'i~ _1" .525424 0. dependent var Akaike info criterion Schwarz criterion Fstat.440981 0.. I).11[::3'.143865 14.410489 0..1".•. you can observe that several predicted probabilities are outside the interval (0. 1'1 W".3: An estimated linear probability model In Figure 11.932844 0.016393 0.Estimation of binary dependent variable models with EViews .. Error IStatistic Prob.386693 8.!~ld.lltolA . of regression Sum squared resid Log likelihood DurbinWa190n slat 0..· l'.E. the least squares output is shown.
250  .P. v. See again the EViews User's Guide for a description of the possibilities.021 AGE. 1 .620 .25 and subsequently the intercept term is increased with 0. It is important to realise that the output has to be read as follows: ~ In P.184) (0. Lagl """"""=== E.024 1.250 INC" (0 . 016) (0.182 0.139) (0.plli.0..478) According to the arithmetical rules of Amemiya the parameters are multiplied with 0.868) (0 . + 0. + 4. The specification of a binary model may never be entered as an 'explicit' equation. See Figures 11 .0.246 EDUC. Various estimation options are available under the button 'Options' .728 IN C.. + 0.5 to make them comparable with the estimated parameters of the linear probability model: Logil: OLS: 2..0. The third possible method (extreme value) is not discussed here.687 .024 E DUC.i ct the chance that a person has a PC given the not exceptional values of the explanatory variables age. ! IKnathr melhDd: ('" ftotJi r.061 0. Select 'BINARY' as the estimation method and choose the method: logit or probit. the standard options have been used.453 AGE. EViews maximises the likelihood function with an iterative method.0. Qualitative Dependent Variables Br. Figure 11. The model cannot predict correct probabilities for extreme values of the explanatory variables.405 0.6 as examples of estimating a logit model for the PCequation.021 0.244 Chapter 11. To estimate a logit or probit model the equation is specified again in the 'Equation' window. = 7. 10 this example.5: Example of the selection of the logit estimation procedure The estimated model is written as:  P C. + 0 . (2. = 0 .5 and 11.113 .244) (1. education and income. .055) The model can be used to pred.006) (0.687 .
3). the McFadden pseudoR 2 and the LRtest are found in the output. Copy the 'Substituted Coefficients' .00) .503640 0. this has been done twice.:niS114 5. log likelihood McFadden Rsquared Obs with Oep=O Obs with Oep=1 Total obs Figure U . 1) (see Figure 11. Finally. of regression Sum squared resid 7.8.83854 2.763480 0.Estimation of binary dependent variable models with EViews 245 Dependent Variable: PC Method: ML · Binary Logit (Quadralic hill climbing) Dale: 03J13.525424 O .198719 S.252704 1.83 (0. The value of the LRtest is 47.6: The estimated logit model All the logit estimates are clearly larger than the OLS estimates.3148 0. Change 'PC' .0011 0.477964 2. 0.90011 40.005194 3. as marked in the figure.30E·l0 26 31 2. This is illustrated in Figure 11.849328 0.245739 4. then the 'Equation' window looks like the window shown in Figure I 1. Error zStatistic Prob. ln the figure.3). Avg.727590 0. the residuals. The formula is based on equation ( 11. As an example.7[6478 0.0014 0.585978 59 C AGE EOUC INC Mean dependent var S.620367 ·0. See the EViews User's Guide for an explanation of other printed statistics. actual and fitted values are given again.286443 0.139391 0. the probabi lity (p) has been computed that a person aged 35 with little education (edue = 1).867752 0. Next probabilities Pi can be computed for given values of the explanatory variables. but high income (ine = 5) possesses a PC. where now all the predicted values are within the range (0.179793 ·16. The following steps have been done after the logit model bas been estimated as given in Figure 11 .0079 0. log likelihood LR statistic (3 dt) Probabilily(lR stat) Schwarz criterion HannanOuinn criter.81938 47.657261 3. dependent var Akaike info criterion Log likelihood Reslr.D. in the command window. both to show the formula and to show the formul a with the substituted val ues of the explanatory variables in the figure.7). education and income it is convenient to compute a probability by using expression ( 11 . so the null hypothesis ofzero slopes is clearly rejected.244489 1. As explained in the previous section.8 and explained below.ed): 1 59 Included observations: 59 after adjusting endpoint s Convergence achieved after 7 iterat ions Covariance matrix computed using second derivatives Variable Coefficient Std. For given values of age.()3 Time: 11 :03 Sample(adjus.453397 0.6 earlier: ~ • • • Under the button 'View' in the ' Equation' window select ' Representations'.E.
PC = 1@LOGJT(·(C(1) + C(2)"AGE + C(3)"'EDUC + C(4)' NC)) vbl Coefficients: Figure U.Fitted Figure 11.. Residual Actual ..~tll.246 Chapter 11. Qualitative Dependent Variables nRf"? ' Fr""'r'""" ' Fr'"~ . 1 '1" .7: The residuals of the logit model · 0.t". I II.4533967565"35 + 0..S: Computation of a probabil ity with an estimated logit model .I)f. II) 1'1 w(.· I '.2457385435"1 + PC CAGE EOUC INC Equation: Yb l~~ ..~r:'""r""'"'"t'?~ .
A number of reference points are given below: • Look at the description of the available variables in Data set 5. U. • • • . educ and in c. Determine which of these variables are candidates that influence the need for external technological knowledge by a finn. Let us look at a linear probability model and a logit model for this qualitative variable.23.5 Case 9: modelling a qualitative dependent variable Data set 5 with the crosssection data on many Dutch companies is used to estimate a model for a binary dependent variable.Case 9: modelling a qualitative dependent variable 247 in 'scalar p' and substitute the values for age. we can consider the variable extech. For the logit model.99. Then the probability can be computed that a finn wishes to have external help given values of the other variables. After doubleclicking on p the probability is shown at the bottom of the EViews window: p = 0. When the probability is computed with the linear probability model. The linear probability model can be estimated with weighted least squares or with White's correction for heteroskedasticity to compute the standard errors. which is a dummy variable that indicates whether a finn needs to acquire external technological knowledge. and determine the signs of their parameters. by using plausible and extreme values of the explanatory variables. Compare both the estimation results by using the arithmetical rules of Amemiya. Analyse the data. For example. the outcome is p = 1. Look at the residual behaviour. In EViews use the procedure 'scalar' to compute these probabilities in a neat way. whi ch is a useless result. Estimate probabilities that a firm needs external technological knowledge with both models. where the minimum. mean and maximum value of the variables are shown. Estimate a linear probability model and a logit model. This can conveniently be done by looking at the descriptive statistics of the involved series. use the LRtest as a goodnessoffit test. as an interesting variable that is dependent on other variables.
.
Part IV .
The question . then a possible solution for the improper longrun behaviour of the model is restricting the longrun parameters. But if a longrun relationship is expected to exist for theoretical economic reasons. The central topi c in Chapter 12 is an integrated discussion of modelling the short and the long run of the economy.is discussed in more detail. The restricted longrun model is imposed as a solution of a matching dynamic shortrun model. explicit attention will be paid to the specification of a lagged dependent variable. [n Chapter 13 we study a model with many lags of one or more explanatory variables. distributedlag models and univariate timeseries models. At the end of that chapter.Introduction to Part IV In the last part of this first course in econometrics. In Chapters 13 and 14. Errorcorrection models. Before discussing these models. Such a specification results in estimation problems. unitroot tests and the concept of cointegration are introduced and extensively discussed. such as a very few degrees of freedom. as in fact the model is only valid for the sample period that has been used to establish it. some general aspects of causal dynamic economic models are discussed in more detail in tbe first three sections of Chapter 12. dynamic shortrun model? It is possible that these properties are unrealistic from an economic point of view. three types of dynamic models are discussed in its three chapters. This is not necessarily a bad property of the model. An interesting question is: what are the longrun properties of an unrestrictedly estimated. and inaccurate estimates because of severe multicollinearity. like improbable values of longrun parameters or elasticities.What is the economic interpretation of a lagged dependent variable and what is the interpretation of the magnitude of its coefficient? . That restricted shortrun model is called an errorcorrection model. The problem is solved by using a distribltled lag specification for those explanatory variables. different dynamic models are discussed. These are errorcorrection models. .
Jenkins methods. the ARIMA models. the univariate timeseries models. Univariate ARIMA models are statistical models for one variable. are introduced. in Chapter 14. The variable is explained by its behaviour in the past and by lagged disturbances. These models bave the ability to produce accurate shortrun forecasts. which are extensively discussed with many examples.Introduction to Part IV 251 Finally. . The procedures to determine an ARIMA model and to compute forecasts are called the Box.
.
this is only done if there are econontic reasons for studying a possible longrun equilibrium model. Examples of this type of quantitative economic analysis are given in Section 12. Unit Roots 12. That concerns the introduction of the lag operator. mathematically seen.2. the lag operator will be applied to analyse the longrun properties of a dynantic shortrun model. a new notation for representing dynamic models is introduced in Section 12. or the researcher tries to estimate a static longrun model. It is possible that an acceptable equilibrium relationship is found. To overcome the problem of an unrealistic longrun model you can examine whether restricting the parameters of the longrun equation yields econontically realistic results.1 Introduction Before the dynamic models are discussed.3. The solution of this difference equation is a static equation that perhaps can be interpreted as the unrestricted longrun model. The use of the lag operator will tum out to be rather useful in the analysis of dynamic timeseries models. but it can also happen that such a solution results in a nonsense model from an economic point of view. When that is not the case. An unrestricted shonrun model is.3. • .Chapter 12 Dynamic Models. Restricting the longrun parameters can be done in two ways: • the researcher sets a longrun parameter to a value that is plausible from the theory based on his or her economic knowledge. a difference equation. Of course. In Section 12. the project ends with the estimated shortrun model that can be used for several other targets.
as di scussed in Section 12. That is true because the static model is a misspecified model for the short run.4 and 12.6. The lag operator is useful for notational and algebraic convenience only. When the economic theory is not clear about the exact specification of an eqUilibrium relationship.3. The errorcorrection model will be introduced and detailed in Section 12. applied on the variables and on the residuals of the static regression. serve as examples with respect to this notation. two situations will be examined in this chapter: • model and estimate the shortrun and check the longrun properties of the model.5. it is perhaps possible to estimate the equilibrium relationship.6. and • estimate or restrict a static equation. notation and examples This section is a preparatory section concerned with the use of a different notation: the lag operator L is introduced. longrun multipliers can easily be computed by using lag operators in the model specification. The procedures to model longrun and shortrun models are explained in Section 12. This notation is convenient in studying theoretical aspects of dynamic models. Define the operator L as follows: LX. The solution of an errorcorrection model is known a priori as it has been imposed: the model converges to the restricted or to the estimated longrun model. That is the second mentioned method: estimation of an equilibrium model. representing the long run. in addition. The unitroot tests are discussed in Section 12. = X' _I. Unit Roots and COintegration When a longrun model has been imposed as a given solution of the shortrun model by the researcher. An errorcorrection model is a dynamic model for the short run that has a special explanatory variable . 12. The tools for this analysis consist of unitroot tests. For more detailed information see e.254 Chapter 12. Such an estimated longrun model is a static equation. Restricting a longrun parame ter requires economic knowledge about the long run. then it causes restrictions on the parameters of the dynamic shortrun model.4. in Section 12. which is an indication for model misspecification. Some specific models. If the variables have comparable trends then it is possible that the static model is a longrun model. then the static equation is a misspecified model and the regression is called a spurious regression. books that extensively deal with these topics. In Part I we saw that estimated static equations usually have autocorrelated residuals.6 we investigate conditions on which an estimated static equation can be considered to be a longrun relationship among the economic variables. But. Hendry (1995) and Maddala and Kim ( 1998). which is di scussed in detail in Sections 12. the restricted dynamic shortrun model converges to the imposed longrun model.the errorcorrection term . Dynamic Models. which will be studied in the next two chapters. If those conditions are not satisfied.2 Lag operators. and estimate a matching corresponding dynamic shortrun model. By means of tills term. So summarising.which represents the equilibrium equation.g. The conditions for a static equation to represent a longrun model are based on the trend behaviour of the involved variables. . However. Hamilton (1994). This restricted shortrun model is called an errorcorrection model.
like Stewart (1991). it is sometimes convenient to use it as a variable to perform some simple algebraic operations. Mathematically seen. .1 ) is written in a simple notation as: (12. However. Next define the lag polynomials {3 (L) of order p and "I (L) of order s: {3 (L) = 1 .. L .. . + u. To check the stability of this model.. That is written as: Y. . it looks fishy using operators like variables. Bring the lags of the endogenous variable to the lefthand side of the equation: ( 12. is a difference equation. = {30 + (yo + "IlL + .I + 'Y2 X. A necessary and sufficient condition for stability is that all the roots of {3 (L) lie outside the unit circle in C.. = L (LX.. Y. In some books and journals the B (backwardshift operator) is used instead of L. Now the general dynamic model ( 12.2) Model (12. + u.2).. .{3. + "II LX.Y.. the result of a computation with the operator L is defined as a new operator. . L' X.p = {30 + "loX.. .2 + . X. .. Then apply the lag operator for the lagged variables and rewrite the equation: Y.1) Y. and after the desired computations have been done the /1. (1 . notation and examples Then the following use of the lag operator will be clear: 255 L2 X...f3.Lag operators. . where they have p and s complex roots respectively. + 'Y. = {30 + {3. 2 + .2 + . + 'Y.{3pLPY..) = LX.s + u. + u. + .X'.{3... + {3pY. = X' _2.. this is illustrated with examples in this and foUowing sections .f3. X. L 2 . = {30 + "loX.. the lag polynomials {3 (L) and "I (L ) are considered as polynomials in the 'complex plane' C.. + 'Y. + 'Y._.. YtI + {32 Y. We start by looking at the specification of a dynamic model of two variables. the result is identical to the result when just the L had been used for the computations. . with p lags of the dependent variable and s lags of the explanatory exogenous variable. L') X.{3. Some authors.{3pY.{3IL ......p +'YoX. ..2 . ..{3pLP) Y. Although L is an operator... define the lag polynomials in a variable /1.. + 'Y2 X. or written as model (12.is replaced by the operator L.LY.. In fact. + 'Y..1 . I). .{3pLP. X' . . + 'Y.
. 1 1 _ LX" . are given in the next section. = f30 In this case !3 (L) is a firstorder lag polynomial: + u.!3IL ) Y. 1). the polynomial !3 (L ) can lie factored as: !3 (L ) = (1 . The only reason to inlroduce these algebraic operations is that the result is convenient in analysing the properties of dynamic models. two sum operators will now be introduced.. Special expressions with the lag operator will be defined as new operators. . : yt = or written as: f30 + !31Y'. that sums the observations of a variable from ' the Big Bang' up till now is defined as the operator 1 1. + . Unit Roots and Cointegration This will be made clear by looking at a simple AR( I) model (AutoRegressive model of order I) for Y. !3(L) = 1 . Applications of the mathematical results on econometric models obtained in this section.. + L2 X..!3.L 00 1 L i=O X' _i = X. then the roots I l bi must be such that ILl> I .I + u" (1 .. .blL) (1 . po These stability conditions yield restrictions on the parameters !3i of the original model ( 12. This condition is the wellknown condition for the AR(l) model to be stable. For example.!3IL ) = 0 L=. Solve this polynomial for L: !3( L ) = 0 (1 . implying that Ibi l < 1 for i = 1.. . = (I + L + L 2 + . + LX.b2 L) .256 Chapter 12. In general. !31 The condition ILl> 1 implies that 1!311 < 1. (1 . all the p roots have to lie outside the unit circle.. . + X'_I + X' _2 + .L. = X..bpL) . Dynamic Models. The sum operator... )X.
One more feature of the new notation bas become clear: when the lag operator appears in the denominator of an operator then we have an infinite (weighted) sum of past values of the involved variable. declines geometrically in time.2 + >.j (1 . It is a weighted infinite sum where the' influence of the past fades away. in the sense that their values are bounded and not all zero. If the observations of the series {X.c X.j (1 . = 1 _ L = l . notation and examples 257 so: X. 1 1 _ >'L X. The assumption of a decreasing explanatory influence in time is not unrealistic from an economic point of view.LX. These are important features of tbe two sum operators. 1 . _.. The assumption is written as: . as seen below: 00 "" ~ i=O A X. . + AX' _J + A2 X' .Lag operators.3) Another useful sum operator is the operator with parameter A: 1 1 AL' with 0 < A < 1... )X.3 + .j (1 .'. The essential difference between the two operators is that one operator has a parameter that is less than one.2 L 2 + >. More details about these sum operators. originally introduced by Koyck (1954). = (1 + AL +>. A solution to the problem of the infinite number of parameters is to assume that the influence of the variable X.L ) is finite. + >. = X.L I: X. + >..3L3X. = X.3L 3 + . and the variance of X .2 L2 X. i=O 00 (12. + >. + .AL) will frequently be encountered in the econometric timeseries literature. whereas the other operator has a parameter that is equal to one.} are 'nonnal' economic observations. The distributed lag model is discussed in the next chapter in more detail. and related aspects are discussed in Section 12. _. An example where this notation is used is the following (distributed lag) model with an infinite number of lags of one explanatory variable: The parameters of this model cannot be estimated without making further assumptions. This operator will be used for the sum of the geometrically declining influence of lags of a variable..5.>. 3 X' ... Expressions like X . It is useful to remember this property from now on'. we can expect that the variance of X .L) is infinite.
. Unit Roots and Cointegration Substitute this in the equation and use the lag operator: y. is called deha. .... + "(0). +u. + "(0). resulting in a twicely differenced variable.5: Vt = t . = (:Jo +"10 _ )"L X. = = f30 + "loX. 1 in Section 7. only three parameters have to be estimated. This implies the following relationship between the operators V and L: V = 1 . This is the 'discrete analogy' of dif ferentiating and integrating of continuous functions.3L 3 + .' L' + ).258 Chapter 12.. .. Dynamic Models.6 but once more the names are mentioned: V is called nabla and fj.L + ). Second differences are wriuen as V' X . .1) = 1. + "Io )"LX. The definitions were given in Section 5.6. Of course. (:Jo +"10 (1 + ). l .X'_I = (1 . + u.' L' X. is: Both operators are simply related to each other: In econontic research.L ) X. the 'nabla' is mainly used.L.L ~ We see that summing is the inverse of differencing. we can difference a differenced variable once more... + . the firstdifference operator V is defined as: V X.: The second difference operator eliminates a trend in the growth rate of a variable. As introduced in Section 5. + u. as it is more useful to look at changes with respect to the previous period in the pas!. Y. 1 Y. with a sum operator.3L3X. . The first difference operator eliminates a linear trend in a variable. = X. The last example in this section of the use of lag operators concerns difference operators. 1 In the last non linear equation. . ) X.(t . The definition of fj. Then the inverse of the operator V : V I has the following expressions: VI = 1 = "' . as was shown in Remark 7.
0) for the second difference: V' 2 y.Y (. the difference operators can be used. then we are interested in knowing the . 4) for the first difference of a seasonal difference: V'V' 4Y' . 12.Y' . In EYiews. Yi . n . is defined as: V'. d (Y.Longrun effects of shortrun dynamic models 259 Sometimes. • The transformation 'V 'V s is not unusual in the empirical econometrics literature. (See also Chapter 14 for the use of this notation in ARIMA models. after the 'Forecast' button has been selected. An example of the use of a seasonal difference operator in a sample with quarterly data is the operator V' 4 that is defined as: V'4Yi = = (1 . in the si tuation that the model is specified in the differences of the variables. . s) is identically defined as d (Y. d (Y. The operator dlog (Y .3 longrun effects of shortrun dynamic models The main topic in this chapter is a discussion about estimating and solvi ng dynamic shonrun models for economic variables to obtain longrun models.) The use of difference operators in EViews is very convenient for forecasting the level of a variable. The workfile grows with every new 'generate' .4.£4) Y. See Figure 12. 1.O O) for the first difference: V'Yi . it is not necessary to generate difference variables in EViews.I ). Instead of generating a new variable like dY: genrdY = Y . the operator dlog is also defined. Figure 12.= IU . n . but it returns differences of the logs of the variable. The difference operator is defined and used in the following way: • • • d ey ) or deY. d (Y.1 for an example. 1. .1 relates to a relationship between the commodity prices (in logs) that have been estimated in first differences. 2. For example. Mostly you will be interested in predicting the level and not the change of the variable. Because we would like to predict future cocoa prices. The seasonal difference operator 'V 8) with period 8. 0) or d (Y. s). The same is true for other transformations with the difference operator. Therefore. seasonal differences of a series can be taken as one more method to adjust for seasonal movements in a variable. The difference operators are also defined in EViews. 0.1. it is possible to choose whether forecasts for the level or for the first differences of the variable have to be computed. when it is relevant. 4) for the seasonal difference in case of quarterly data: V' 4Yi. if a dynamic shortrun model is used for simulation experiments. the difference operator was used as shown in Figure 12. In that situation. 2) or d (Y.
( 12. Look at the following example with two different dynamic models. intermediate.and longrun responses .see Table 12.~ SjaIic Figure 12.5) In model (12._.5) has a lagged dependent variable. The longrun multiplier "10/ (1 .5) the stability condition: 1f3t1 < 1 is required. A shortrun dynamic model can explode or converge in a straight or in a cyclical way.1: Short. 1 . Model (12. Yi Yi = f30 = + "lo X .4) Model (12. Next.5) t= 1 'Yo ')'0 t=2 t=3 'Yo + 'Y. as an illustration of dynamic shortrun models and matching longrun models. + 'Yom = 'Yo (I + (3. + u. Dynamic Models. The solution of a dynamic model can result in a sati sfactory longrun model.4) has a lagged exogenous variable and the other model ( 12. X ' .260 Cha pter 1 2 . and longrun response to Yi of a unit change in X t in consecutive periods. which are the properties of difference equations. ')'0 +/ 1 + 1'1 'Yo + 'Yo(3. but there is also a possibility that the values of the longrun parameters are economically unrealistic.) .1 Table 12. f30 + "loX. + "I.4) ( 12. = 'Yo (I + (3. 'Yo / (1  'Yo + /31) /3.depends on the Period Model (12. Unit Roots and Cointegration fOrC(d'>t £J : .f3tl in model (12. because: 1 < 1 /3.1: Forecasting P{·oc in a model that explai ns V pt~ (in logs) properties of the model with respect to its longrun behaviour.I + u . Y.) 'Yo + 'Yo(3. look in Table 12.5) . + f3.1 al the shortrun .
'Yo/ (1 . Yi. {Jo (1 .5) with the lag operator and solve the equation for Yi. Next a shortrun model can be estimated with thi s restri ction on the long run .{JIL) Yi.6) Remember that we had established that the longrun response of Y. if 90/( 1 . it is possible to restrict this parameter to an acceptable value that is known from the theory or from other research.{J.4) and ( 12. by substituting the restriction in the shortrun model.) = c. + u. is: ( 12. In Section 12. + 'Y.: Y.6): Then the static equation represents the longrun equilibrium equation between Y and X.L) + u. Observe that this value of the longrun response is found by substituting L = 1 in the expression (12.. : The 'parameter' of the variable X.L) (12.{J. to a unit change in X. However.{J. + u.L) X.4) and rewrite the model with the lag operator: Yi. + u. Notice that: = + 'Yo X.. Restrictions on parameters will not be imposed in this way but by specifying an errorcorrection model for the short run. (1 .) has  an economic unrealistic value. is 'Yo + 'Y. Consider model ( 12. The restri cted shortrun model can be maintained when the restriction does not generate residual autocorrelation.{J. The errorcorrection model converges to the longrun model with the imposed parameter value(s). = = /30 + 'YoX. /30 + (yo + 'Y. The shortrun model has been respecified in such a way that the longrun restriction is clearly visible in the specification.) Yi. Next. (1 . first we look at the corresponding longrun models of models (12. = + {JILYi. = /30 + 'Yo X.4 the errorcorrection model is introduced. rewrite model (12.7) /30 (1 .LX. for example.Longrun effects of shortrun dynamic models 261 values of the sample estimates of 'Yo and {J.{J.5) by using the lag operator. + u. (1 . {Jo + 'Yo X. But the other way around.{JIL) (1 {JI)' .
then compute the longrun response by substituting £ = 1: "(( I) "(0+"(1+ "('+ . Unit Roots and Cointegration as a constant term cannot be lagged. /3. As a side effect.5)./3. /30 Y = {3(£ ) . The longrun response to Yi of a unit change in X. ./3. Now look at the general dynamic model (12. y = /30 + "(0 X 1 . 1 .£) of X. In equation ( 12... {3p · By substituting £ = 1 all the parameters are summed. + {3(£ )' "( (£ ) u. First rewrite the model : /3 (£) Yi = {30 + "( (£) X . is computed in the following way. .. is "(0/ (1 .262 Chapter 12.2)... equation ( 12. Dynamic Models. The longrun equilibrium equation 's . +"(. {3(1) 1 {3. In model ( 12.7) shows why an estimated misspecified static equation suffers from serious residual autocorrelation: the disturbance term is specified with a distributed lag. Substitute £ = 1 and the longrun multiplier is: "(0 1The static equation /3./3..5) we bad established that the longrun response of Yi to a unit change in X. With these examples it has become clear that the use of the lag operator is convenient in computing longrun multipliers. represents the longrun equilibrium equation between Y and X as a result of the shortrun model ( 12.7) we observe again that thi s longrun response is found by substituting £ = 1 in the operator "(0 (1 ./31) . + u" + {3 (£ ) X .
the value of f(x) at x = c + £ is: . \710g (X t ) = log (X.J . Rewrite the equation as before: Y t = 7r21 + 7r22Gt + 7r23Yt. This can be obtained more conveniently by using the lag operator in the reducedform equation (10. inflation is approximated by the first differences of the logs of the consumer price index: \7ln (C P 1.X t _1 '" '='''X t.8) 7r 22 7r23£ Yi v = 7r21 + 7r22Gt + Vt2 7r23 ' t= 7r2 J 1 +1 G t+c. Example 12.7r23£) = 7r21 + 7r22G t + 7r23£Yt + Vt2 (12. This can be seen by means of the foUowing algebraic exercise. Then we see the approximation shows up: log (1 + £) '" log (1) + d (log (x)) dx x=1 . x=c The approximation for 10g(1 + x) is obtained when computing a firstorder Taylor series for f(x) = log(x) with x = c + £ around c = 1..:::t2==7r23£ (12.7r23). This result can be verified graphically.1 A model In the macro model from Example 10. . the longrun multiplier for income with respect to government expenditures was recursively computed as 7r22 ! (1 . In empirical econometrics we frequently observe that the transforonation \710g (X..£ =0+ 1 ·£ < c.2.1 + Vt2 y.8). In general. which is of course equal to 7r22! (1 .£ 2 + .J Xt . the fact has been used that for x close to zero.9) is identical to equation ( 10.J . so notice the notational simplicity achieved by using the lag operator! The longrun multiplier is obtained for £ = I . Such a transformation can be used as a proxy for the relative change of the variable up to a maximum of about 15 %.1 in Section 10. but it can also be shown by using a Taylor serieS.7r23) as before.1 __ V.log (Xt.log(l + x) '" x.9) Equation ( 12.7)..tl = log (:J = I og ( 1 + X tX t .• Longrun effects of shortrun dynamic models 263 Example 12. (1 . For instance in the next example.Xt _ l) I In the last step.J is applied as an approximation for the relative changes of X t .1.
0. Because of thi s property.74 (0.014) 4a: C. of these equations are given in Table 12. = 0.72 0.93Yi . ) is frequently used as a proxy for the inflation in applied econometric work.) .264 Chapter 12. .O. is reported to give an indication about the quality of the estimated result: absence of firstorder autocorrelation. the following estimation results have been obtained . Explanatory variables are real personal di sposable income (Yi) and inflation (VCPI. By using annual data for the sample period 1960.2. _ 0.O. but In(l.549) The BreuschGodfrey test shows tbat only model 4 is acceptable.O.36 (0.24 + I .88Y. . with pval ue in parentheses.68 .I .0.24L ' Be (I ) = 5.06Yi _l or 2b: C.0. + 0.16 0.71 y. . .  .15 VCPI 1 .033) 2a : C.OOY .71Y.14VCPI. use your calculator and see that In (l. based on an example in Stewart ( 199 1: p.0. concerning a model for U K consumption (C.1984.196).68Yi + 0. = 0.24L .25 + 0.140. The estimated longrun elasticity of model 4 is as expected from economic theory. The variables are in logs.53 (0.24C' _1 or 0. so that the approximation has already deviated by 1 percentage point. .01 1 . + 0.0.25 + 0. = . Vln (C PI. = 0. thi s result looks even too good! The computed longrun equation of model (4) is obtai ned by the substituti on of L = 1 in equation (4b): C = .16 + 0. Dynamic Models. I : C.15VCPI. 12.002) 3a : C.40Ytt + 0.2 A macroe_nlc consumption function The longrun implications of dynamic shortrun models are illustrated with this example. by substituting L = 1. Unit Roots and Cointegration For example. = _ 0.18 V CPI 3b ' C.98 (0.  Be (I ) = 4.0.25 + (0.0. = 0.15) = 0.0.1 _ 0. 0.0. approximated by the first differences of the logarithm of the consumer price index.54VCPI .88 + 0.18VCPI .0.40Ly.05 ) = 0.   Be (I ) = 9.16VCPI. as is the case in the next example. wbere only the BreuschGodfrey Be (I )stati stic.72L ' 1 .049.72Ctt or is .0.01 4b: + 0.).04 + l.06L ) Yi ..16VCPI . The estimated contemporaneous and longrun elasticities.72L ' Be (I) = 0. .
Some historical remarks It is interesting to see how this modelling concept has evolved in the econometrics literature over time. the description of an econometric analysis is based on an unrestrictedly estimated mndel for the short run with a solution for the long run.88 + 0. together with low values of the DWstatistic when a static regression of two independent random walks is computed.71 / (1 .40) / (1. . It is a new way of modelling a shortrun dynamic model with a given longrun relationship.0. Srba and Veo ( 1978) the errorcorrection model was introduced.0. high R 2 s are found.4 and 12. In the wellknown article of Davidson. In Sections 12.00. These restrictions work through in the accompanying shortrun model.93 0. So the other way around.24) = 0.2: Shortrun and longrun income elasticities So far. This knowledge is used in the study of the nature of an estimated static equation. Their approach will be shown in this section. The general idea of ECMs is intrnduced in this section. The longrun equilibri um relationship has parameters that have been fixed at values known from some economic source. The plan for the rest of this chapter is as follows . which will be discussed in Section 12.0. a hi gh R2 and a low DWvalue.94 0.Errorcorrection models Model Model Model Model Model I: 2: 3: 4: 265 Short run 0. starting with restricting the longrun equation. Then. However.6 matching econometric models for the short and the long run are discussed. we start with a longrun mndel that is a given model. in Section 12. Granger and Newbold (1974) show that in most cases high tval ues. in this section.93 0.88 0.6.4 Errorcorrection models When we are convinced that a longrun mndel exists. Hendry.06 = 0. This restricted accompanying shortrun model is called an errorcorrection model (ECM). An estimated static equation can be either a spurious regression or an estimated equilibrium equation. • 12.68 . An estimated longrun mndel has also a matching ECM.72) = 1.71 0. and the unrestrictedly determined longrun model is an unrealistic model. we may try to impose restrictions on the parameters of the longrun model.94 (0.5 the concept of unit roots and of unitroot tests is introduced as tools to analyse trends in the data. Table 12. The ECM converges to this relationship. when your (static) result shows high tvalues. Such a regression result is called a spurious regression .68 Long run 0. based on the economic theory or known from empirical observations. then be aware of spurious results.
in Section 12. have a root equal to one. This makes the distinction between a spurious relationship or not. L'. A sufficient discussion on these topics along the lines as outlined above.. we look at the Engle and Granger ( 1987) formal twostep procedure to test for. If the polynomial has a unit root then the variable has a linear trend. 'Y (L) = 'Yo + 'Y. To obtain the longrun relationship. The trends will be analysed by looking at the 'order of integration' of the variables. such as the tests of Phillips (1987). Restricting the long run The concept of errorcorrection models will now be introduced and discussed in general.11) . Then finally. and estimate.. This is tested with a unitroot test . and in Section 12. is given in this and the following sections.L + 'Y2L2 + .5. 10) has been rewritten as: ( 12. These tests (and some more) can be applied with EViews.O'~). to make this methodology operational. and Phillips and Perron (1988). The static relationship will not be a spurious regression if the variables have common trends and the residuals are stationary. The concept of cointegration will be introduced and discussed in Section 12. longrun and matching shortrun error correction models. The unitroot tests were introduced by Dickey and Fuller (1979) and many unitroot tests were later established.10) iJ(L) = l iJJL . ~ NID (O.. equation ( 12.L2.iJpLP .2 as: iJ(L)yt = iJo +'Y(L)X.2. The static regression is either spurious or conditions can be established that the static equation can be interpreted as a longrun equilibrium relationship. bas been specified in Section 12.6. which will be explained in Section 12. .fJ. which have been introduced in Section 12.5 the unitroot tests are discussed. with the lag polynomials iJ (L ) and 'Y (L) : ( 12. + 'Y.6 the tests around the cointegration analysis are described. An unrestricted dynamic model between two variables yt and X. . after which we will look at some examples. Important in this context is analysing whether the timeseries have common trends or if they behave according to their own individual trends. The order of integration of economic variables will be determined by testing whether the lag polynomials. the errorcorrection model is introduced .266 Chapter 12. but the discussion in this book will be limited to the DickeyFuller tests.+u" u. Then the variables are integrated in a similar way and for that reason they are called cointegrated. In this section. Unit Roots and Cointegration Granger (1981) combines the concept of the errorcorrection model with his previous work by showing that a static estimation result is not necessarily a spurious regression.. Dynamic Models.
Errorcorrection models Rewrite the constant term and define f3
267
o, l' (L) and u;:
f30
1
' f30 f3o f3 (L) ,
,
E
i= l
p
,
f3;
1 (L)
1
1 (L)= f3(L) '
u, = f3 (L) Ut ,
and substitute them in equation (12.11):
Yi =f3o+1'(L ) X t+u;.
The longrun response of Yi to a unit change in X, is l' (1):
The longrun relationship between Y, and X , is:
Yi =f3o+ 1'(I)X,+u;.
The parameter l' (1) is the unrestricted longrun response of Yi to a unit change in X,. Suppose that a longrun equilibrium between the variables Yt and X, is imposed according to some economic theory; in particular assume that the theory states that Y and X are in equilibrium in the long run:
Y = X.
( 12. 12)
Jmposing ( 12.12) as the equilibrium relationship ( 12. I 2) implies restrictions on the parameters of the shortrun model (12.10):
1'(1)= 1.
This restriction will be found again in tbe shortrun model. Because of this restriction we have assured ourselves that the dynamic shortrun model ( 12. I 0) converges to the longrun model ( 12. I 2). This concept is elaborated in this section for two models.
An errorcorrection model for two variables with firstorder lag polynomials
Let us first look at a dynamic model for two variables that are represented by AR(l ) processes: botb the lag polynomials f3 (L) and 1 (L) are of first order. This is the most
268
Chapter 12. Dynamic Models , Unit Roots and Cointegration
simple example to start with:
f3 (L ) = 1 
1 M,
+ I'l L .
I'( L ) = 1'0
Then, with these polynomials the unrestricted model ( 12. 10),
f3 (L ) Y, = f30
is written as,
+ I' (L) X, + tt"
y, = f30 + I'o X , + I'lX'  l + f31Y'  l + U"
Ut
( 12.13)
~ NID (O,O"~) ;
1f311 < 1.
Rewrite the model using the lag operator and the notation introduced ahove.
Y, =
flo
1  f31
+ 1'0 + I'lL X, +
1  f31L
tt,
.
1  f31 L '
(12. 14)
fJO 
'" _
f30 1  f3 l
1" (L) = 1'0 + I'lL 1  f3 ,L
Then equation ( 12.14) is written as:
Y, = f30+ 1" (L ) X,
The unrestricted longrun response is 1" (1):
+ u;.
1" (1) = 1'0 + 1'1 1  f31
( 12.15)
If the parameters of equation (12.13) bave been estimated and a longrun relationship exists,
you have to cbeck whether the estimated model converges to an economically realistic longrun equation. Calculate and evaluate (12.15): ('Yo + 'Yl) / If thi s value of the longrun parameter is not in accordance with what is expected from an economic point of view, then the parameter can be restricted . The restriction can conveniently be imposed when the model has been rewritten in the following way. Write model ( 12. 13) in first differences and correct for the newly introduced lags (the equation remains
(1  ill) .
Errorcorrection models unchanged):
Y, = fJo + "loX, + "II X'  I + fJI Y' _I +
U"
269
\7Y, + Y' _I = fJo + "10 \7 X, + "IO X'  I + "II X tI +fJ,Y, 1 +u" \7Y, = fJo+"Io \7X, + "Io X'  1 + "II X'  I +fJ1 Y' I Y' _I +u" \7Y, = fJo + "10 \7 X, + bo + "Itl X' _l + (fJI  1) Y, I + u"
\7Y, = fJo + "10 \7 X, + (fJI  1) (Y' l + ;1
~ ~l X' _I )
+ u"
or ( 12. 16) The parameter in parentheses for X' _I is exactly equal to the longrun parameter (12.15) and can easily be restricted. As a simple example the following longrun relationship is imposed:
y = x.
This gives a restriction on the parameter of the longrun response "I' (1):
"I' (1) =
"10 + "II = l. 1  fJI
Substitute this value in the shortrun model ( 12.16), which results in the restricted shortrun model:
\7Y, = fJo + "10\7 X, + (fJI  1) (Y' I  X, tl + u ,.
(12. 17)
Model ( 12.17) is the errorcorrection model. The last term (Y, I  X' _I) is called the errorcorrection term, which is exactly the longrun relationship written with subscript t  1. An errorcorrection model is an equation specified with variables in first differences and with an errorcorrection term. The model shows the longrun relationship and the shortrun deviations from the equilibrium relationship. Notice that the disturbanceterm
Ut
has remained unaltered after rewriting the equation, so the original disturbanceterm
assumptions are unaltered! The parameters can consistently be estimated with OLS. When residual autocorrelation is absent, the restriction and the specification can be maintained. The coefficient of the errorcorrection term is negative as IfJd < 1. II can be shown that a value close to one implies a fast convergence to the equilibrium relationship and a value close to zero a slow convergence. This will be explained in Section 13.4. II is also possible and interesting to compute the parameters "I: from the infinite polynomial "I' (L ):
270
Chapter 12. Dynamic Models. Unit Roots and COintegration
The coefficients relationship:
'Yi
show the pattern and the speed of convergence to the equilibrium
"YOI
• ('"Yo + 1'1 ) (. + "Yl• + 12 , ... , . .) 10 .)
In this example. the last coefficient is equal to one:
The series of coefficients is called the stepresponse function. These steps show the con· vergence rate to the equilibrium relation: Y = X . For example. these coefficients can be computed in the following way by evaluating 'Y ' (L) X, .
• (L) X = 'Yo + 'YI L X 'Y ' l f3I L '
1 _ f3I L
1
bo + 'YIL) X,
= (1+f3IL+ f3iL2+f3~L3+···)bo+'YI L ) X,.
Next compute. term by term:
'Y' (L)
X,
=
'Yo X , + b l
+ f3no) X'_l + f31 bl + f3no) X ' _2 +f3T bl + f3no) X'  3 + ....
Then we have found the polynomial 'Y' (L):
'Y' (L ) = 'Yo + b l
+ f31'Yo) L + f31 b , + f3 ,'Yo) L2
( 12.18)
+f3T bl + f3no) L3 + .. . .
The coefficients in the polynomial (12.18) form the stepresponse function. When the parameters of the errorcorrection model ( 12. 17) have been estimated. then it is easy to compute some coefficients of the lag polynomial 'Y' (L ). See the elaborated example at the end of this section for an application with a relationship between coffee prices on the world market and in importing countries.
An errorcorrection model for two variables with second<>rder lag polynomials
The exercise will be done once more for a model where the two variab les are AR(2)
processes. [t is an identical exercise as before, but the example illustrates how things look
for higherorder lag polynomials. Both the lag polynomials are of second order. Just as
Errorcorrection models before, the assumed equilibrium relationship is:
271
Y = X.
The secondorder lag polynomials /3 (L) and 7 (L) are:
/3{ L ) = 1  /3,L  fh L 2, 7{L ) = 70 +7,L + 72L2
The unrestricted shortrun model (12. 10) is:
Y, =
/30 + "lo X, + 7, X,  , + 72 X ' 2 + /3, Y, , + fh Y'2 + "t ,
u, ~ N ID (O,u~).
Substitution of the lag operator and solving the equation for Y, gives:
~ ~ L~ fh L2~ =/30 +~~+ " L~+~O~+u,
(1  /3, L  fh L2) ~
_ Yot 1
=
/30 + (yo + 7 ,L + 72 L2) X, + u,
u, 2" 1  /3, L  /32 L
/30 + 70+7,L+ 72 L2 X + /3,  /32 1  /3, L  /32 L 2 t
This corresponds to our general notation of the unrestricted shortrun model:
Y, = /30+ "I' (L) X, +
u;.
The unrestricted longrun model is obtained by substituting L = 1:
~ =
/30+7' (1)X, +u;.
The longrun response , ' (I) is:
"I' (I ) =
70 +7, + 72. 1  /3,  fh
(12. 19)
The given equilibrium relationship Y = X implies restrictions on the parameters by restricting the longrun response "I' (I ), as seen before:
,'(1)= 1,
implying the restri ction for the parameters of the shortrun model too:
70 + "I, + 72 1  /3,  /32
= 1.
(12.20)
272
•
Chapter 12. Dynamic Models, Unit Roots and Cointegration
Again, rewrite the model in first differences and correct for the newly introduced lags .
•
\7Y, + Yi  1 =
f30 + "Yo \7 X, + "YO X'  1 + "Y1X' 1 "Y2 \7 X' _1 + "Y2X' 1
+ /hYi 1 + /h Y' 2 + u"
\7Yi = (30 + "Yo \7X, "Y2\7X' 1 + "YO X '  1 + "Y1X '  1 + "Y2X '  1 + (3, Y' I + (32 Y '  2  Yi l + U" \7Y, =
f30 + "Yo \7 X,  "Y2 \7 X' _l + (yo +"Yl + "Y2) X t _ 1
+ ((3, + (32  1) Yil + u, .
This results in the following unrestricted equation for the short run:
( 12.2 1) Just as in the first example, the coefficient of X' _l in parentheses in model (12.21) is equal to the longrun response (12. 19), and can easily be restricted. Substitution of the longrun restriction (12.20) in the shortrun model gives the errorcorrection model:
\7Y, = (30 + 10 \7 X, "Y2 \7 X'_l  (32 \7Yi _1 + ((31 + (32  1) (Yi  1  X, tl + u, .
(12.22) Because the variables Yi and X, are AR(2) processes, we observe that \7 X ' _1 and 'VYi 1 have entered the unrestricted model ( 12.21) and the errorcorrection model (12.22). In practice, the orders of the processes are unknown , so by modelling the ECM from 'general to specific' the orders will be empirically determined, as shown later on. For this model, the stepresponse function can also be computed. This can be done in the following way. Rewrite "y' (L) X,:
"Y (L ) X, = (3( L)X' "Yo + Il L + "Y2 L2 X 1  (3,L  (32£2 '
,
"Y (L)
= 1 (3I;/hL2(~+~L +~~)~
The coefficients of "y' (L) can be computed in the following way:
"Y' (L ) = (3 1 (L )"Y (L )
Errorcorrection models
or:
273
(3 (Lh' (L) = "I (L )
(1 
(31L  /h L2) ("10
+ "Ii L + "I;L 2 + "I3 L3 + .. .) 
("10
+ "IlL + "I2L2)
=
o.
Then, collect the parameters for the identical powers of L and compute the coefficients recursively:
"10"10= 0 L : "Ii  (3no  "11 = 0 • • " . ,..2"10 L 2 . "12  ,..1 "11 " .  "12 = 0
1:
=> => =>
or:
L 3 . '  ,..1"12  ,..2"11 0 . "13 " . " ' . .. ...
=>
"10 = "Ii = "I. = "I. = "Ij =
.. .
"10 (31"10 + "11 (31 ((31 "10 + "Ill + (32"10 (3ni + (32"10 + "12 (3n. + (32"1i
+ "12
The coefficients "Ii of the polynomial "I' (L ) form the step response function as before and can be computed after the parameters of the errorcorrection model (12.22) have been estimated. It is possible that the first values of the "Ii become larger than the restricted equilibrium value, and after that converge to the equilibrium value, which is an overshooting effect (see the elaborated example at the end of this section).
Errorcorrection models in practice and irreversible behaviour of explanatory variables
A practical aspect when estimating an errorcorrection model, is that no a priori knowledge is available abou t the order of the lag polynomials. Therefore, Hendry's principle of 'modelling from general to specific' will be used again. Start with a specification that has suffi cient laggedendogenous, laggedexogenous variables (both in first differences), and an errorcorrection term as explanatory variables. For example. when we have a bivariate model
and the longrun equilibri um relationship is assumed to be Y = X, we can start with
the ass umpti on of p lags for Yt and X t as a general dynamic model withou t residual
autocorrelation .
V'Yt = (30 + (31V'Yt  1 + ...
Ut
+ (3p l V'y" p+ l + "10 V' X t + ... + "Ip l V' X, _p+ l
+"1 (Y' I  X,  ,)+ut, ~ NID (O,(J~).
Test whether the specification can be narrowed after the parameters have been estimated with OLS . This happens in the usual way by checlcing the residuals for absence of serial correlation and by testing the significancy of the parameters. When changes of variables are modelled, an interesting question arises: has a positive unit change of an explanatory variable an identical influe nce on the dependent variable as a
In fact.. generate two dummy variables that have. the names DX P and DX N. < O an d .274 Chapter 12 . The variables 'V' + X. and 'V'. DXP and DX N are dummy variables with values 0 and I that have been assigned as follows: DXP = g if'V'X. .. < 0 if'V'X. > O if'V'X. the estimation results can be improved by specifying increases ('\7 + X. Generate the variable DX as the first differences of the variable X: DX =XX(.X . > O' The variables DX P and DX N obtain the real values of the positive and negative changes by multiplying the two dummies with the first differences in DX: DXP = DX . These dumnties are created as follows.X . Next. ) and decreases ('V'. Write the 'generate' in EViews exactly as written below (see also Figure 12. Dynamic Models. Unit Roots and Cointegration negative unit change of that explanatory variable? The relationship as given above is called a reversible model because this specification shows an identical change of the dependent variable for a positive or a negative unit change of the explanatory variable. >0' For example. ) of the explanatory variable as separate variables.l) . Define the new variables as follows: . if'V'X.2): DXP = DX > O DXN = DX <= 0. < 0 if'V'X. can easily be generated in EViews by using dummy variables. a simple bivariate reversible ECM is: ( 12.23) which is respecified as an irreversible model: ( 12. A model specified with such variables is called an irreversible model.> O if'V'X. DXP DXN = DX . that is a restriction! If this restriction is not valid.24) The practice of estimating an irreversible model with EViews is not difficult. < O and DXN = g if'V'X. . for example.X... DXN.{'V'X' _ 0 v if'V'X.
PIcot in the EViews workfile is shown. different time paths to the equilibrium level can be computed for responses to positive and negative cbanges of the explanatory variables.X.Errorcorrection models 275 genr dlpcoilee = DOpcoffee) genr dlpcofP = dlpcoffee > 0 gen.1. and V . For example. dlpcolN = dlpcoffee • dlpcolN Figure 12.10 · 10  IT • ~+ _~ '0 .: V + X. = DXN.24) the null hypothesis: pt' . The hypothesis that the influence of a differenced variable is reversible can be tested in various ways. dlpcolN = dlpcoilee <= 0 gen. With an irreversible model. These variables will the creation of V + be used as explanatory variables in an elaborated example of an ECM for coffee prices at the end of this section. and the result shown in a group The resulting irreversible variables DX P and DX N are equal to the abovedefined variables V + X. dlpcofP = dlpcoffee • dlpcofP gen. and V . = DXP V. In Figure 12.p~of (in logs). test in the estimated equation (12.X.2.2: Creation of '\7+ p tCo / and \1 .
C)nt + Ut.24) have been estimated.25) In this way. If is homoskedastic then it implies that the original disturbance term is heteroskedastic.23) and the unrestricted model ( 12. Vogelvang (1988) has shown that the response of retail prices of coffee in importing countries to coffee price changes on the world coffee market is an irreversible matter for many countries. Information from the coffee branch suggested u. Unit Roots and Cointegration with the Wald Ftest. It is assumed that the costs of roasting coffee are proportional to the consumer price index cpit: k.orrection model In an econometric study about the price formation of coffee. As no observations exist on kt some assumption has to be made about this variable. = "'(I C)nt + 7)) + (1 + 7)h (L) w Pt. C)nt ( 12.(I +7)) 'Y" (L) = (1 + 7)h (L) The longrun coefficient is 'Y" (1) = 1 + 7). . the prices have become real prices. For an example using the Wald Ftest. Dynamic Models. Substitution of this assumption gives: Divide the two sides of the equation by cpit: r Pt. = '" . or the likelihoodratio test (LRtest) when both the restricted model (12. . This results in the following economic model : p r where 1] represents the 'markup'." = ". and on the costs of roasting coffee kt in quarter t. with ".276 Chapter 12. We could expect that this property will be valjd for more commodities than coffee only. + U"t .25) is written in real prices as: r Pt = '"" W + 'Y"(L) p. The quarterly model explairling tbe coffee retail prices has been derived in the following manner. An elaborated example of an irreversible error{. and equation ( 12. cpi. Assume that the retail price in an importing country depends on the costs to purchase coffee on the world market Pl" in quarter t and preceding quarters. see the example at the end of Section 12.6.
(2.42 0.I  1. the irreversible results for France and the Netherlands are given successively (with asymptotic tvalues in parentheses): f/I'Prt .57 (0.01 0.0." 1) ' _ VI'N't = .4.19 + 2 1.42) ' (PN' t.0. = 5." _1."+ 0.07 '\7.1'.52 I 1.24) (I'Pr ' t.82 + 0.0.4: Stepresponse function for the coffee retail price in the Netherlands . The model has been estimated for vari ous countri es for the sample period 1976(0 1) 198 1(03).24 4 1. For France.1'. as shown in Table 12. 10) 1.63 2 1.17 1.12 1. 10 the Netherlands.19 '\7+1'.60 '\7+1'.23 1. the stepresponse function clearly exhibits asynarnetric behaviour in response to changing world market prices."+ 0.006 '\7+1'. Lag Change ofp~ o 0. Lag Changeofpr o 0 . 96) + 0.52 (3."_1.20 1. 75) (4 .82) 0.32 1.Errorcorrection models a markup of 30%.97) + 0."_1 (0 . The stepresponse function sbows that the regulation really worked out.25 Table 12. Tbe errorcorrection terms are clearly significantly different from zero (it is a onesided ttest).47 '\7.3). 14 '\7.1'. 19 0.70) (2. As an example. and rusa an overshooting effect for price increases is observed (see Table 12. a strong governmental price regulation was effective during the estimation period."+ 0.07 + I 1.38) '\7.1'." _1)· Contemporaneous price increases on the world market have a more significant influence on the retail price than worldmarket price decreases.1 7 4 1.31'."_1 (3.14 .05) (3.26 (2.22 3 1.68) (0.3: Stepresponse function for the coffee retail price in France The estimation results for the Netherlands are quite different.50) ( 1."+ (2.01 3 1.35 1. The accumulated step responses have been computed to increases and decreases of the world market price.I  1.25 Table 12. 79 '\7+1'.31 1.31'. so the restri cted equjjjbrium equation is: 277 Thi s longrun restricti on was necessary because no realisti c values for the markup had been obtained by estimatin g and solving unrestri ctedly shortrun dynamic equations.
In thi s chapter. Often. These random variables are not independently distributed. makes the process stationary. A timeseries is. then the assumption is made that the growth rate of the series is a stationary stochastic process. If the graph of the variable shows heteroskedastic behaviour of the variable. 12. such as taking the first differences of the observations. . . A timeseries will be considered as a stochastic process and some aspects of a stochastic process will be investigated for the involved variables.. attention will be given to the analysis of the trend behaviour of an economic variable. then the process is probably not stationary with respect to the variance. If this is true. because the normal distribution is completely determined when the first two moments of the distribution are known. is it deterministic or stochastic? These are the topics addressed in this section. This implies that weak stationarity of the second order is a sufficient condition for strong stationarity. . . More discussion on thi s subject can be found in the statistical literature. stochastic processes that are normally distributed are considered and entitled Gaussian processes. We will look at timeseries that have a joint distribution that is independent of time. Thus if the logs of a variable are specified and if the variable has a trend. A stochastic process for n observations is a timeordered sequence of n random variables with their own probability distribution. Unit Roots and Cointegration Hardly any differences exist between the response function for price increases and for price decreases. The \110g transformation is a proxy for the relative change of the variable. there is no place for a theoretical introduction into the subject of stochastic processes. In this section. The means. It does not matter where the time slice has been made. Because of the choice for an em piricallyoriented 'first course in econometrics' as presented in this book. . Y2 . If the distributions are not equal with respect to all the moments of the distributions. A timeseries can be seen as a realisation (or a draw) from ajoint density function p(Y1 . Dynamic Models.. only the first two moments are independent of time. so the distribution of Yb Y2 ) . for exannple. they have an n variate distribution. variances and covariances will be examined whether or not they are independent of time. Economic timeseries often behave with a trend and in such a case the process is not stationary with respect to the mean. in fact. as shown in Section 12.. Are trends present or absent in the variables? If a trend is present then what type of trend is it. where s can be either positive or negative. in that case. then the process is called a strongstationary process. Y2 + s .278 Chapter 12. If the trend behaves linearly in time then a simple transformation. . Only introductory remarks are given here concerning some elementary aspects of a stochastic process. Yn). A combination of transformations such as the 'Vlog transformation is frequently encountered in the literature.. after that relationships between nonstationary variables will be analysed in more detail in the next section. it is important to have knowledge about the trend behaviour of the economic variables that are modelled. • . a 'time slice' .3 .5 Trends and unitroot tests To obtain a satisfactory econometric model with respect to economica1 and statistical assumptions. Yn +s . then the process is called a weakstationary process. Yn is identical to the distribution ofY1+s . the log transformation is useful in obtaining more stationary behaviour.
The variance of a random walk process does not exist. (12...27) with the lag operator conveniently showing these two characteristics.d + E (Ut. E (Y. increasing or decreasing. Write ( 12.C7~)..) = E(u. Such a variable can be written as: Yt Ut = Qo+Q1t+Ut.26) The process ( 12.26) is called a trendstationary process (TSP process). ~ NID (0.2) + .. = au 00 . Yi = f30 + Y t. Then the process is called a random walk with drift.l + Ut 2 + Ut . = 0.. The drift parameter (constant) pushes the process in one direction. Yt= l _L Ut = The mean and the variance are: Ut 1 + Ut . many economic timeseries will behave like a trended variable.1 Ut + Ut ~ N ID (0. The random walk process can be extended with a constant term.Trends and unitroot tests 279 Trends Let us look at various possible trends in an economic variable.I + Ut· . C7~) ..  2 2 2 + au + U u + . This process can be made stationary by correcting Yi for the trend: (Yt  00  Olt) = Ut· ( 12.. Var (Y. This process will appear to be a process that can frequently be used to describe a stochastic trend in a nonstationary economic variable: . a random walk can be considered as a process that moves in all directions: Y t = Y t.5 .3 + .) + E (ut. Usually. as shown below..) = Vm'(u. Next. The process {Yi} is clearly a nonstationary process. which is the reason why the variables are nonstationary processes with respect to the mean .. =0+0+0 + .27) This process is stationary with respect to the mean but not with respect to the variance.) + Var (ut _ d + Var(Ut _2) + . A linear deterministic trend in a tim~series variable Vi has been introduced earlier in Section 7.. . .
i ) I with i = 1. it is important to di stingui sh between a stationary and a 000 stationary process. For economic reasons. which proves that the process is strong stationary.'.. . E (V.Yi.L U. An important economic difference between the nonstationary random walk and the stationary AR(l ) process is that a shock in 11. it is also called a unitroot process. 2. ~ NID (O . has a specific lag polynomial: {3 (L) = 1 .28) with the stationary AR( l) process: Yi.280 Chapter 12.L.. ( 12. .29) the influence of the shock on Yt fades away over time. This process of Y...29) ~ NID (O . or in other words the polynomial has a unit roar. has a petrnanent influence on Y.. Yi. Y. is a differencestationary process (DSP process). That polynomial has a root equal to one.28). O'~ ) ..... Dynamic Models.} is Gaussian. whereas in model ( 12. (1 . 11. = {31Yi. Use the lag operator to rewrite model ( 12.J + 11.2 + {3~11. Unit Roots and Cointegration Taking first differences of the observations of Yi.L) Y. so independent of time. = 11.. possibly with a drift parameter: (Y... O'~). 3. compare the non stationary random walk model : + Ut . in model ( 12. which explains ilS name. as the process wrinen as {3 (L) Y. = {3..d = 11. .29) and to compute the mean and variance to see that the process is stationary : v. makes the randomwalk processes stationary.. )= O+ O+ O. and the first and second moments are constant.. The process {Yi.'3 + .1 + {3~ U' .. [n fact this has been done before in Section 8. = 11.5 when determining the elements of the covariance matrix n of an AR(l ) di sturbance tetrn in (8. = O lo a si milar way. Yt = Yt  1 ( 12. For that purpose.9). = I _{3 .I + 11. are constant. one can verify that the covariances cov (Yt l Yt ." with 1{3t1 < 1 11. 1 Yi. + {3J11..
This is written with the following notation: y.l = (1 .} is nonstationary and the process {\7Yi} is stationary then the lag polynomial f3 (L ) of Y. but has been repeated here because a test for non stationarity will be formalised in terms of the roots ofthe polynomial f3 (L ). for model ( 12. in notation: An I (2) variable will be stationary after transforming the variable in second differences: (1 . If the polynomial f3 (L) has two unit rooLS (see later on in this section) then the variable is integrated of second order. (1 . and for theAR( I) model the polynomial f3 (L) has one solution. This concept is in accordance with the analogy around 'summing'l'integrating' and 'differencing'l'differentiating' that was introduced in Section 12.\7Yi. In the random walk model. In other words. The. = u. Tfthe process {Y.29). < f3t < 1. L = 1/ f3t. For economic reasons. = = \7 2 Yi \72y. the polynomial f3 (L ) has one solution. the polynomial of the nonstationary process has a unit root and the polynomial of the stationary process has a root larger than one.2. • . which means that we will test the null hypothesis that Yt is nonstationary against the alternative that Yt is stationary. and . ~I ( I ).L)2Yi \7Yi .L)\7Y. null and alternative hypotheses with respect to equation ( 12. The null hypothesis will be tested that f3 (L) has a unit root against the alternative that f3 (L ) has a root larger than one.28). An I (1)variable is stationary after transforming the variable into first differences: (1 . L = 1. This property had been established earlier. it will generally be more relevant to test the alternative H. : 0 Notation A new notation and terminology will be introduced that is convenient in the context of unit roots.L ) Y = Itt . has a unit root and the variable is known to be integrated of first order. ~ 1 (0) .Trends and unitroot tests 281 Compare the two models by using a lag polynomial to see the difference more clearly: (1 .30) 1f311< 1.L)Yi = u" or \7Y.29) are: Ho : f31 = 1 Ht : ( 12.f3tL ) Yi = u" for model ( 12. \7Yi ~ 1 (0) .
the 1970s. V 2 Yt = u. and has at most one or two roots equaJ to one. The lag polynomial (3 (L) is a polynomial of degree p. and V 2 Yt is stationary. stationary. which imply exponential trends in the growth rates. it is not possible to use a standard Student's ttestto test the null hypothesis: Ho : (3. The economic interpretation is that Y. < 1. whereas the remaining roots are larger than one. trend in the growth rate. ~ J (2): Yt ~ I (2) ~ VYt ~ I (1) ~ V 2 Yt ~ [ (0). VYt has a linear trend. Many economic variables are trended and will behave like an I (1) process. linear trend in the data. . that the polynomial is a specific seconddegree poly nomial: (3( L ) = (1 .282 Chapter 12. Which behaviour can be expected of economic variables? Economic variables will be: [ (0). = 1 H. Dynamic Models. However. but do exist sometimes. DickeyFuller unitroot tests To test the null hypothesis of a unit root you have to estimate the (3. [ (1).L )' (3( L)Yt = u. Examples of [ (2) variables are prices and wages in some inflationary periods. (1 .29) with OLS. See also the following notational sequence ifY. for example. . Unit Roots and Cointegration This implies for the model that if (3 (L) is of the second degree.L )2 Yt = u. in equation ( 12. Greene (2000) mentions money stocks and price levels in hyperinflationary economies such as those of interwar Germany or Hungary after World War II as examples of [ (3) series. Timeseries that are [ (3) (exponential trend in the growth rate) or higher are very unusual from an economic point of view. has a trend in the growth rate. we will look at general lag polynomials (3 (L ) other than only firstor secondorder polynomials. : 0 in the equation: < (3. or I (2) . Later in thi s section.
. an expression for the asymptotic distribution of the OLS estimator in the K variate model was given: In the model (12.(3m with l(3d < 1.) = It follows that: .fii.(3. for 1(3. . under H. This distribution can be determined as follows . in model (12. 1< 1. (12 t.l ' var ((3.31). = 1 because of a zero variance. L.. use these expressions for the OLS estimator (3.J t .  (3. l3. The tstatistic tii.Y.(l3.31) the variance of yt is equal to: Then.. so: . This distribution is valid under H./8e (13.2. ( l . does The distribution makes no sense for (3.te (3 ( t(n . The asymptotic distributi on of the OLS estimator (3. of (3. We do not have a distribution of (31 under Ho : (31 = 1.l ). The same is valid for the Student's tdistribution of the tstatistic tii..) ~ N (0.Trends and unitroot tests with the tstatistic tii = 283 not follow a Student's t distribution as is demonstrated below.e (31 1_ ) I. _ L: Y. "y'2 L. = s.'2 l .:  tii. looks like: . of (3.). In Section 9..
• A fourth situation can be distinguished. 1) To apply the test. 1) vn 1) .Fuller ttest (D F ttest). Hereafter. Therefore. This case is not considered here. (12. and statistical tables are available for these three different situations._t + 'Ut . Hamilton ( 1994) for delai ls.1) Yi. Fuller ( 1976)). The situations are described below· .284 Chapter 12. Because the di stribution of n ~13t (f31the OLS estimator of f31 is called super consistent. Three different situations are distinguished with three di stributions for the test stati stic. the test is called a Dickey. = (f31 . = IhY'. = BY.mple. : Y. In the description of the three situations the most simple assumption is made that the variable Yi follows a firstorder autoregressive process. Critical values and probabilities are computed by EViews when performing a unitroot test and printed in the output.1). Unit Roots and Cointegration The tstatistic tii.. a more general AR(P) process will be used for the application of the test. or HI : 0 It is convenient to subtract parameter in the equation: < f31 < 1.. Tables with critical values for this nonstandard distributi on can be found in Hamilton ( 1994). ~ 1 (0). Dickey ana Fuller. (f31 .I Write the model with B = + u. for ex. Yi.32) Ho : Y.32) and to estimate the \7Y. computed critical values for this di stribution by simulation (see. for example. does not follow the Student's tdistribution as was shown by Dickey and Fuller ( 1979). Thus we cannot test the null hypothesis that f31 = 1. In the 1970s. see. \7Y. . Unitroot tests are included in EViews. for example. In that situation the SlUdent's tdistribution can be used. H. with references to exists instead of the wT'k of Dickey and Fuller.1 + u. but that di stribution has no analytical expression. Situation 1: no constant term and no linear trend in the model The equation that will be estimated for performing the test is: Yi We want to test the null hypothesis. or Ho : f31 = 1. Dickey and Fuller have found that a nondegenerated distribution of n ( 131 exists. called a unitroot rest. against ~ J (1). Dynamic Models. it is not necessary to specify the regression equation yourself in most of the welJknown econometric software.I from both sides of eq uation (12. and others afterwards.
~ I (1) vers us HI : yt ~ 1 (0): 'V'yt = f30 + (f31  1) Y'. then the critical value is . Situation 2: a constant term but not a linear trend in the model The omission of the constant term in the equation (12.1. In the output.l + 1£t estimate e with OLS and compute the tstatistic tii: tii = e (~) .3. 5% 1 .l + Ut.95 at the 5% significance level.3: Testing for a unit rOOl in Yt = (31 Yt . go to the 'Series' window and find the option 'Unitroot test' under ·View'. Then the null and alternative hypotheses are identical to: Ho : HI e= 0 : e < O. For example.95 Figure 12. When using EViews. With a constant tenn the equation that will be estimated is: Again this eq uation is rewritten to make it more suitable to test directly the null hypothesis Ho : Y. EViews gives the critical values and probabilities belonging to the computed tstatistic tii' This t statistic follows a D F tdistribution under the null hypothesis of a unit root.32) can be rather restrictive. The testing of the null hypothesis has been iUustrated in Figure 12.e e This tstatistic is directly given in the computer output of the unitroot test.Trends and unitroot tests 285 DickeyFuller Itest H. suppose we have a series with 100 observations. and with e = f31  1 the equation is: . s.
= (8) This tstatistic follows a DFtdistribution under the null hypothesis of a unit root.4: Testing for a unit rOOl in Yi = f30 + f3. = ~ + fJ.89.89 Figure 12.3.05 (100) = 1.1) Yi. Just as before. Situation 3: a constant term and a deterministic linear trend in the model The last situation that we consider is a situation that a linear determinjstic trend has also to be included in the equation.66. = fJo + (fJI . Using the same example as in the first situation of a timeseries with 100 observations. .45. . Unit Roots and Cointegration DickeyFuller I·test H.l + "It + IL. For example. Yi. Yi. the 5% critical value for n = 100 is .286 Chapter 12. Then the equation is: Y. by using a DFtdistribution ..I + "It + IL" + OYi_1 + "It + IL. . Dynamic Models.O to. the critical values for the test are computed.4. t.l + U. estimate the (J with OLS and test Ho : Y. with the tstatistic to. the 5% critical value is 2. Estimate the parameters and compute the tstatistic (J S. : Yi ~ 1 (0). The Dull hypothesis of a uoit root will be tested in the rewritten equation: VYi = fJo VY.89 clearly deviates from the 5% critical value of the Student's tdistribution: to. ~ I ( 1) versus H. The critical value of . In EViews. The picture that illustrates the test is given in Figure 12.2. 5% 2.
8(L)Yi Yt =.Perron tests. The general model is : .I is nonstationary but the first differenced lagged variables 'VYii are stationary. When the null hypothesis has not been rejected. Of course. choose p in such a way that the residuals are not autocorrelated. In general..8p Yt . we do not know the order of the process and so the order has to be determined empirically. For an economic variable. do not use the DWstatistic to test for residual autocorrelation. the Augmented DickeyFuller test (AD F test). = iiil se (iii) is Student's tdistributed and t. which can be done by testing Ho : 'VYt ~ 1 (1) . This implies that the number of lags 'VYt . Other wellknown unitroot tests are the Phillips and Phillips. When a specific specification has been accepted.. In practice. That is done with equation (I 2.i can be determined by using the Student's tdistribution by testing that <>i is significantly different from zero.80 =. A linear trend term can also be included in equation (12.i + Ut· i= l practice.= O (0) is DFtdistributed.p + Ut· eli The equation.I <>i'VYi. .2 + .8. but look at the correlogram (Qtest) of the residuals.80 + OYt _ 1+ L p.1) Yi. situation 2 will be the most appropriate because a constant term will be necessary and an additional trend term is generally superfluous. + . the test will also be applied on regression residuals where the first situation can often be used as the residuals have a zero mean. Rejection of the null hypothesis implies stationarity of the process. In the description of the DFttest.Yi.1 + IhYi.Trends and unitroot tests 287 The use of the DFtest in practice In the econometric practice. Later on. in rewritten form (using as notation for the transformed Pi ) is: 'VYi = or . an AR(p)process will be assumed and the orderp will be determined by ' modelling from general to specific'.I + L p.33) i=l 'VYi = .33).8.Bo+Ut + . (12. a firstorder autoregressive process for the variable had been assumed. This means [n lse that to. as it is possible that a second unit root exists. These tests assume any ARMA process for Yi. test the null hypothesis Ho : 0 = O. you have to choose one of the DFttests. Under the null hypothesis of a unit root we have the following properties of the 'explanatory variab les': Yi. the conclusion is drawn that at least one unit root exists in the process.Bo + (. we have to test fora second unit root by testing the null hypothesis Ho : Yi ~ 1 (2) . This 'variant' of the unitroot test has got its own name. If the unitroot hypothesis has not been rejected.33) specified for'VYi : .I l>i 'VYt _ i + Ut .
~ 1 (1) Ho not rejected ==> Ho : Y./3. 1. also the disturbance term is different..£ . = /3" (£) \7 2 y. If /3 (£) is a lag polynomial of degree p and the polynomial has one unit root..£ /3. . = Ut U" where \7 2 y. = u. = u" where \7y.5. the relevant variable is selected by doubleclicking on its name: the 'Unit Root Test' is found under 'View' in the 'Series' window.··· /3. To perform a unitroot test in EViews. A procedure for testing for the order of integration of an economic variable is shown in Table 12. ~ 1 (1) H . = u.288 Chapter 12. : Yi ~ 1 (0) Ho rejected: Y. is a stationary process and /3' (£ ) is a lag polynomial of degree p /3(£ ) Y. you have to specify whether a constant term (situation 2).£)2y. : Y. then the process is not stationary and the polynomial can be factored as follows: /3 (£) = 1 ./3p P /3(£ ) = (1 . ~ 1 (1) H. so that: /3(£ ) Y. ~ 1 (2) fl o rejected: Yi ~ I (2 ) Ho not rejected: problems Table 12.£) (1 /3. /3' (£) \7y. To apply the (A)DFtest. The same is valid for two unit roots: /3" (£) (1 .£ )/3' (£).33)./32£2 . _1P..5: A sc heme for testing for (he order of integration of an economic variable The notation with the tilde above the parameters has been used to di scern the parameters from the parameters in equation ( 12.:y.2. ~ 1 (0) H 0 not rejected ==> Ho : Y. Unit Roots and Cointegration Ho : Y. /3'(£)(1£)Y. Besides the (A)DFtest you find other unitroot tests that are not di scussed in this book. is a stationary process and /3" (£) is a lag polynomial of degree p . ~ I (1) H . ~ 1 (2) or Ho : 'VY.£2 . = u. an intercept and a . ~ I ( I ) Ho rejected: Y. Dynamic Mode ls. ~ 1 (3) or fl o : 'V 2 y.1) /3 (£) = (1 .
l. first or second differences. the automatic selection yielded an acceptable result... but five other criteria. 'Trend and intercept' is Situation 3..1).. of the variable to perform the test.. When a correct specification has been found. and 'None' is Situation I. To test for unit roots in economic variables situation 2 will mainly be used. for regression residuals).7. In the given example.6... Without significant autocorrelation it is possible to use the Hest to test for the degree of the lag polynomial. f"}jane Figure 12. . Then you have to specify the degree of the polynomial (P) by indicating the number of lagged differences (p . I0Il . such as absence of residual autocorre.:. For specifying (or searching) for the order p of the polynomial you can choose whether the computer or yourself selects the lag length.Trends and unitroot tests Ilml Hoot T p~' 289 n tt:3 I t: Atla. first doubleclick on RESID to look at the Q test and the sample residual autocorrelations. In Figure 12..5 you can see the start of the test. EViews uses an information criterion.lation and the significance of the parameters. These two criteria are discussed in Section 14. The Schwartz Information Criterion (SIC) appears by default in the window. bas to be included. some EViews examples are given of the windows and output concerning the DickeyFuller test to test the null hypothesis that the cocoa price is firstorder integrated (in logs). . the lowest value of the SIC has been cbosen for a regression with a maximum of 18 lags.. like the Akaike Information Criterion (A I C). 10 Figures 12. .5 and 12. . I' "we . but we consider the ADFtest only. After this procedure has been done. Six tests can be found under 'Test type '. For the automatic selection. or nothing (situation I).. You also have to select the level.= • )oIaI" »end and i . Level and Situation 2 have been selected as shown.. A priori it is not obvious what is the best choice. are available.5: Selection and specification of a unitroot test linear trend term (situation 3).1'. This selection can be done by yourself but also with the help of any of the available information cri teria. EViews makes it easy to perform these unit root tests and deciding the order of integration of an economic variable.. If the automatic selection has been chosen one has still to check the results.. The three discussed 'situations' can be recognised: ' 1otercept' is Situation 2.. whereas situation I can be the appropriate situation for a differenced variable (and later on.... the null hypothesis of a unit root can be tested by usi ng the tstatistic of the estimated parameter of Y'..
8.017639 ·2. After a first regression it was obvious that the constant term could be dropped from the regression (a zero coefficient and tvaluel. LPCOCOA(·l) D(LPCOCOA(l )) D(LPCOCOA(·2» C Rsquared . Augmented Dickey·Fulier Test Equation Dependent Variable: D(LPCOCOA) Method: Least Squares Date: 03I2Dm Time: 14:27 Sample(adjusted): 1960:04 ""2002""':09 Included observations: 510 after adjusting endpoints Variable Coefficient Std.· 0.290 Chapter 12.112138 0.699357 0. MAXlAG=l8) IStatistic Augmented Dickey·Fulier test statistic Test critical values: 1% level 5% leval 10% le. Next proceed with testing for a second unit root. Unit Roots and Cointegration Null Hypothesis: LPCOCOA has a unit root Exogenous: Constant Lag Length: 2 (Automatic based on SIC.600590 8.00743J 0.4774 ·1 .002585 Figure 12. Dynamic Models.044100 0. An AR(3) process has heen assumed for the cocoa price ((p .7 and 12. by testing for a unit root in the first differences of the cocoa price V In (PtC ): DC Ho : In (PtCOC ) ~ 1(2) Ho : V In (PtCDC ) ~ 1(1).018644 ~ 1 .1) = 2).""C) is not rejected.442945 ·2.1003 0.353641 .600590 ·3..044213 0.536291 1. so tbe cocoa price is integrated of first order: In (PtCDC ) ~ 1 (1) .004619 0.Q.031683 0. Error IStatistic Prob.oogg Mean dependent var n /iAnFioriAnt vaf n 0. .115854 0. so the final result for situation I is shown in Figures 12. The null hypothesis of a unit root in V (In (PtCDC )) is clearly rejected.569733 "MacKinnon (1996) onesided pvalues.6: Output from the ADF ttest applied on In (PtCOC ) The null hypothesis of a unit root in In (P.Q.1 Prob.0115 O.1lXXl 0.
O~ ) Selection and specification of a unit~root test for . Error IStatistic 14. Augmented DickeyFuller Test Equation Dependent Variable: D(l. MAXLAG=IB) I·Statistic Prob.2) .Trends and unitroot tests 291 Figure 12.116599 0.PCOCOA. 0.641933 Prob. Null Hypothesis: D(lPCOCOA) has a un~ root Exogenous: None Lag Length: 1 (Automatic based on SIC.000290 Figure 12.941445 MacKinnon (1996) onesided pvalu8S.0.763663 0.78657 2.00:0 O .ooes D(lPCOCOA(·1 )) D(lPCOCOA(1 ).044134 Mean dependent var 0.8: Output from the ADF ttest for V' In (Pf OO ) .· I 5% level 10% level 1 .2) Method: Least Squares Date: Time: 14:57 Sample(adjusted): 1960:04 "200"ro2:09 Included observations: 510 after adjusting endpoints Variable = Coefficient Std.7: V' In ( P.051659 0.
it is not uncommon to see this type of misspecification in published researcb. we considered estimated static equations. Theo a lot of autocorrelation will be found in the residuals. This also means that unitroot tests belong to the data analysis preceding the model estimation. However. + u" with the following orders of integration (varying between 02) of the explanatory variables: The model bas been correctly specified with respect to the orders of integration if the order of integration of the dependent variable (Yt ) is equal to the maximum order of integration that occurs at the righthand side of the equation: Otherwise the model is misspecified. it is important to know the order of integration of the dependent and all the explanatory variables. under particular conditions a static equation can be interpreted as a longrun equmbrium equation. The trend is found in the residuals. instead of the assumed static relationships that were used in the previous section. Suppose the model is: Yt = (31 + (32X. as the trend in the dependent variable is not explained by any of tbe specified variables. . The obtained knowledge about trends and the unitroot tests from the previous section will be applied to study the properties of longrun and shortrun relationships between nonstationary variables. 12. as spurious regressions. will be discussed in detail. Granger (1981) writes that. + (33 Z. or the other way around. Consider the following example. Up to now.292 Chapter 12. with a lot of residual autocorrelation.1 Unltroot tests and causal models For obtaining a correct specification of a causal economic model. Unit Roots and Cointegration Remark 12.6 Relationships between nonstationary variables A different approach for modelling an ECM is the use of an estimated longrun equation as the imposed longrun solution of a dynamic shortrun model. Dynamic Models. Look at the following situation: the dependent variable Yt has a trend (Yt ~ I (I )) and none of the explanatory variables has a trend (they are I (O)). the properties of stati c equations and the conditions that the equation is spurious or a longrun model. in practice. This is done using a clear and consistent way of testing and modelling. In this section.
they can have similar trends that cancel in a linear combination of the variables: the static regression.1) static relationships exist between J( variables. ~ 1(0). .34). In that case l'I'd < 1 and u. and further that the DWstatistic reduces to zero. ~ = '1'1 = 1.(l2 X" is stationary. They drive apart when they have their own specific trend. it can be shown that the t statistic does not have an (asym ptotic) tdistribution and that tvalues rise for n 700. In other words.4.'1'1). This difference in the behaviour of the disturbance tenn is decisive in the conclusion about the nature of the static relationship of nonstationary variables: a spurious equation or a longrun equation. is stationary: u. whereas the linear combination of the two variables et. to keep the di scussion simple. But if the nonstationary variables are DSP. The variables may not drive apart from each other when they form part of an equilibrium relationship. = Y. When this occurs. Both variables have a similar stochastic trend. In the beginning of this chapter. + U. it is possible that a maximum number of (J( . written as ~ ~ e. together with low values of the DWstatistic. DW = 0 U. but the DWstatistic is not equal to zero. Yt and X t . (12. .Relationships between nonstationary variables 293 We start by looking at a spurious regression of two variables.(l. So the variables are nonstationary. These are properties of a spurious regression. perhaps something different is going on. the study of Granger and Newbold (1974) was mentioned. So the nature of the relationship depends on the properties of the processes in the variables and in the residuals from the static regression. If the variables are TSP then the regression will be spurious as the variables follow their own deterministic trends. The discussion of the concept of cointegration in this 'first course in econometrics' will be limited to the investigation of the existence of one longrun relationship among a number of economic variables.'l't L )Ut = c" DW "" 2 (1 . It is possible that a low value of the DWstatistic is found. the di stinction was introduced between variables that are DSP and those that are TSP.34) will frequently have high tval ues and hi gh R Zs. but a stationary linear combination of the variables exists. the integrated variables are called cointegrated. For the static model (l2. However. In Section 12. the residuals have a unit root: (1 . Which property the variables has is of importance in judging the static relationship of the variables. Mathematically seen. They demonstrated that static regressions between two independent unitroot processes.ilibrium relationship when the variables are integrated of first order and the OLS residuals are stationary. as in the following equation: yt = (ll + (l2 X. one can test for the number of cointegrating relationships by using the maximum likelihood . This means that a static regression between nonstationary variables is a longrun equ. In a multivariate model. representing specification errors with respect to the dynamics of the model. 1 (1 ) .
35) A condition for the existence of cointegration of the variables Yt l X t2 1 . The twostep procedure will be discussed together with its application in EViews. then the next step is to test the hypothesis that the differenced variables have a unit root: Ho . starting with the articles of Jobansen (1988) and Jobansen and Juselius (1990).. Y. If Ho is not rejected. We will limit ourselves to the Engle Granger twostep procedure (EG procedure) as mentioned earlier. Cuthbertson. Haldrup's contribution in McAleer and OxJey (1999) for a survey of the literarure on 1(2) processes.35) concerns a cointegrating relationship and can be considered as a longrun equilibrium relationship. Cboose the number of lags of the dependent variable or choose the 'Automatic selection'. .36) has to be specified in an 'Equation Specification' window and estimated with OLS to perform the Ftest.35) are stationary. . the null hypothesis that the variables follow a difference stationary process is tested.. Various applications of this methodology can be found in the econometics literature. For I See. Then equation (12. A second condition is that the residuals of equation (12.. First. ~ 1 (1). 1 (2) series are not considered here' After the unitroot bypotheses for the variables have been tested and not rejected. Extensive econometrics literature exists about this subject. X'2 ~ 1 (0) •. V'Y. Johansen and Juselius' approach is beyond the scope of this introductory course. Dynamic Models. ~ 1 (1). the null hypothesis that the variables are 1 (1) is tested: Ho .• X'K ~ 1 (1) Hj . X tK is that these variables have similar stochastic trends. we consider the static relationship: (12. Because only 1 (1) variables are considered as candidates for a possible cointegrating relationship. V'yt ~ 1 (0).. will be tested.• X'K ~ 1 (0) • by using the ADF test. V' X'2 ~ 1 (0) •..• V' X'K ~ 1 (1) Hj . . to test the null hypothesis that the process is a DSP (or a unitroot process). for example. Y. in such a way that residual autocorrelation is absent. • V' X'K ~ 1 (0) . V' X'2 ~ 1 (1) •. If such a cointegrating relationship exists then a corresponding errorcorrection model exists for the short run and will be estimated in step 2. For the formal discussion of the cointegration concept. Hall and Taylor ( 1992) or Verbeek (2000). X'2 ~ 1 (1) •. In step I the existence of a cointegrating relationship between two or more variables..294 Chapter 12. but the interested reader can find good descriptions in many textbooks such as Hamilton (1994). In EViews an equation like equation (12. see Vogelvang (1992) where a coi ntegration analysis of spot prices of various types of coffee is presented.. as specified from the economic theory. Unit Roots and Cointegration methodology of Johansen... for example. given below. The procedures of Johansen are included in EViews witb the option 'Cointegration Test' under 'View' in the 'Group' window (see the EViews User's Guide for a description). ~ 1 (0) .
where the bypothesis is tested at the 5% significance level. First. After the parameters have been estimated. example.34 6. as this Fstatistic is not Fisher F di stributed. = flo + (fl .16 Table 12. the test of the null hypothesis that the variable follows a DSP process is illustrated in Figure 12.6: Critical values for the DF Ftest Source: Part of Table B.6 for the Dickey. again for a series with about 100 observations. In the following example.6 1 5.91 5.10. as shown in Figure 12. has not been rejected.49 6. _1 + I'H The variable Yt is DSP if the null hypothesis L: <>. fll = 1 and l' = 0. . instead use Table 12.7 in Hamilton ( 1994).24 6.05 0. DickeyFuller Ftest 5% 6.I t. + Ut· (12 .34 7.49 Figure 12. 'Coefficient Tests' . the null bypothesis is tested that the cocoa price follows a unitroot process.65 7.36) Ho .10 0.025 25 50 100 250 500 00 5.81 7. But do not use the standard Fdi stribution.44 7.value in the EViews output! For example.Fuller F statistic. for the variable Yt the following equation is specified: VY.1)Y. based on the source Dickey and Fuller ( 198 1).36 5. VY i= l p.25 8.20 7. ' Wald Coefficient Restrictions ' to test the null hypothesis: C (2) = C (3) = 0. the Ftest is computed via ' View' .Relationships between nonstationary variables Critical values Sample size n 295 0.73 6.39 5. So do 1101 look at the printed p.9: Example of a DFF distribution to test the hypothesis that the process is DSP .25 7.47 5.30 6.9. specify the equation in the 'Equation Specification' window..
4 Std.74 (0. The regression concerns 510 observations.Fuller F~ lest The number of two lagged dependent variables was suffic ient to have a specification with nonautocorrelated residuals.."""~W"""""~[""'j~_ WIld Test: Equation: EQ_DSP_CDC Test Statistic Value df (2 .84E05 0.1o /01111111" Pirr: F~t:=""rr.K ) (12.II.505) Probability 2 Null HYFWlhesis Summary: Normalized Restriction (= 0) Value 0. Figure 12.1'1..4%.6. The null hypothesis that the coffee price follows a difference stationary process is clearly not rejected at the 5% (or 10%) significance level. which shows a 5% critical value of 6.37)  . .10: Specification for the Dickey.20E. Dyn amic Models.30. Err. as was checked beforehand.72). III 1"1' 1111 w . LB(36) = 30.009762 1.0. The Fisher Ftest indicates a probability of 19. Unit Roots and Cointegration Figure 12.Q5 Restrictions alll linear in coefficients. but this value comes from a FisherF distribution: we are concerned in this case with a DickeyFuller Fdi stribution. The OLS output is not given.:71f.19" 0.S ) / g S/ (n. therefore we use the row enlTy for n = 500 in Table 12.11."10".193.U: Result of the F test .0053$ 2. the Ftest can also be computed as: F= (SR . C(2) C(3) . The test statistic has been co mputed with the result shown in Figure 12. Of course.296 Chapter 12.
07 .4.44 3.02 . .45 2.52 3. . but not equal to zero. because the residuals have been obtained conditional on the estimated model.K· The last condition is checked by testing the null hypothesis that the OLS residuals have a unit root..16 4. which implies that they are not stationary.67  5% 2.37 .46 500 500 500 500 500 1 2 3 4 5 Table 12.4. Then the linear combination of K nonstationary vari ables is stationary.45 4.68 . . This procedure yields an estimated longrun relationship.76 3. This equation is a longrun equilibrium relationship if the OLS residuals e.4.43 2.14 2.99 3.42 3.7: DF t~ distribution s for unitroot tests for the residuals of the static equation Source: Part of Table B.f3KX.9 in Hamilton ( 1994).83 .99 4.f32X '2 .77 4. The critical values of the standard DF t test in EViews are not vatid for the OLS residuals. . Use Table 12.38) and the residual sums of squares (S) of the unrestricted equation (12.4. [f this null hypothesis has not been rejected. based on Fuller ( 1976). H. The DW statistic wiH be low.. : e. in contrast with the Situation 1 n Situation 2 Situation 3 K 2.77 5. ~ 1 (1) ~ I (O ) .36).64 .4.13 4.7 when the ADFtest will be used to test Ho.80 4..81 4. After we have not rejected that all the variables are unitroot processes the static rela tionship is estimated: Y.05 3.5 % 3.11 4.3. = Yi .4. = 13.4.13.37 3. with et: e. and that the variables are not cointegrated. the static relationship is a spurious regression. ~ ~ ~ Ho : e.71 10 % .49 4.45 .84 4. .5% .07 .3.02  5% 3.98  5% 3.55 3.37) with the residual sums of squares (SR) of the restricted (12.74 10% 3. Then of course use the table of the D F Fdistribution.3.3.16 4.38) 297 V'Y. are stationary. .27 3.20 4. The same procedure is apptied on the explanatory variables X t2 . Other critical values than before need to be used.' U" (12. . In case the hypothesis has been rejected.1 3 3.Relationships between nonstationary variables Estimate ( 12.5 % 3.XtK .71 . + f32X'2 + f33X'3 + . with OLS.74 4.38) and compute (12.38 4.4. + 13KX'K + U..39 . the variables are assumed to be coi ntegrated and the static relationship is an estimated longrun equilibrium relationship. = (30+ L "'i V'Yi i + i= l p.40 10 % 2.36) and equation ( 12..
as usual. However. which is the errorcorrection term in this model. Short run: \7Yf. equation ( 12.6. When the conditions for cointegration are not satisfied. In equation (12. a specification in first differences (possibly in irreversible form) and an errorcorrection term (e. and X.d. + e. Unit Roots and Cointegration given longrun relationships in Section 12.34) is a spurious regression.6j\7Yf. . A sideeffect of this property is that the bias and inconsistency of the OLS estimator. are cancelled out.39).1 pl (12. are not related. or tbe other way around.298 Chapter 12. Yf. the lagged residual variable e. according to step 2 in Engle and Granger (1987). are assumed to be of order p and s.6.X. e. i=O j= 1 s. which can also be considered as a test for cointegration. R 2 reduces to zero. DW goes up. \7Y. and X. the variables are TSP..35). (k = 1. The parameters .61 + . If the model were to be estimated in first differences (variables are stationary in that case) nothing is left of the 'good statistics'. = ..2 Procedure It is possible that one of the nuU hypotheses in the scheme above is not rejected whereas it is actually true (or the other way around) because of the ADFtest's low power. Dynamic Models. Qj \7Yf. + L 'Yi \7 Xti + L Long run: ._I.60 + 'Yo\7 X. = . and t. they have a unit root and so the DW statistic is not significantly different from zero. It is not necessary to use the 2SLS estimator.60+ I: l'i \7Xti+ I: . ~ I (1). a bivariate model for the variables yt and X t is used here as example.4. For notational convenience only.39) The lag polynomials of Yf. . By using the residuals (e.. = . Variables that are not included in the longrun equation can be included in the shortrun model. stems from the longrun equation (12. still go on and estimate the ECM and test whether e'_ 1 has a clearly significant influence (tstatistic with a low pvalue).35) a shortrun error correction model is estimated.. The longrun equation is written with a residual term and the shortrun model with a ' hat' above the dependent variable. caused by endogenous explanatory variables. because the rate of convergence is n instead of Vn (see the previous section). K) ofthe static equation have been super consistently estimated with OLS. Remark 12.j +l'e'. A correct way to publish the estimation results in a correct notation is: Yf.1 +u.39).and Fvalues are no longer 'significant'.3 Specification The shortrun model and the longrun model can be different with respect to the number of specified variables. Remark 12. .) from the longrun equation (12.j + 'Yet! . Then the residual s are not stationary. Draw conclusions only when the 'complete picture' of aU the results can be considered. The ECM is of the form as equation (12..
In this chapter.Relationships between nonstationary variables 299 '23 This chapter is concluded with an example concerning the relationship with the two commodity prices P[OC and p tCo ! in logs. store the residuals (.12.ted. .3 that R2 = r~<o. 14. The prices clearly move in a similar way in time. p . Observe the low value of the DW statistic. we have already established that Ptcoc is a unitroot process. The same can be found for pr ! (show that yourself by using the available coffee data). and apply the unitroot test for the null hypothesis that the prices are not cointegrated: . ' Make Residual Series .o. the static regression will be compared with the solution of the estimated shortrun model shown in Figure 4. Therefore. Next the analysis proceeds with investigating the properties of the static relationship between the two prices. these results will be evaluated. Figure 12.<o! are cointeg. With respect to the interpretation of the value of the determination coefficient. The result of the OLS estimation of the static equation is given in Figure 12. it is interesting to investigate the hypothesis that PtCOC and p. remember that for a bivariate model we have established in Section 5. " The output of the ADF !test is shown in Figure 12.4.+. 5+j+'h i 4++_. ').' Next.12: The logs of the commodity prices . We will investigate whether a longrun rela tionship exists between the two worldmarket prices along the lines as discussed above. . A graph of the logs of the prices is shown in Figure 12. which has been done under the name RESI D_ COI N. Procs'. At the e nd of the example..13. U not rejected. For notational convenience we write Plcoc and p tCo ! instead of In (PtCOC ) and In (PtCO! ) in this example.
00).E.352 \1 P.462 p coc+ 0.00) + 0.40) is a longrun equilibrium relationship between the logs of the both prices.76 from Table 12.14 (0.1 + l.136019 0.c.788209 Mean dependent var 0. of regression Sum squared resid Log likelihood DurbinWatson stat 0.328 p coc_ (0.0. Error IStatistic Prob.121 (0.016 + 0.00). If we want to test the null hypothesis at a significance level of 5%.277915 Akaike info criterion 39..047) \1 p. We observe that all the estimated parameters.002 (0. Unit Roots and Cointegration Dependent Variable: LPCOCOA Method: Leasl Squares Date: 04AJ8. BG(3) : 3.2.99. except the constant term but including the errorcorrection coefficient. Dynamic Models.897914 0.5% level. It is interesting to compare these results with the originally obtained unrestricted results from Section 4.1 (0.603299 o 2fOl86 0.l . JB : 49.043) (0.00). Thai result was: FoC= .D.992180 0. So we conclude that Ptcoc and pta! are coinlegrated and that the equation <12.07 (0.300 Chapter 12.370).28). (0.523628 0.coc = 0.OllDl LPCOFFEE C R·squared Adjusted Rsquared S.041 ) p co f _ 0..044) '. .40) is written as: ~ p. 0. White: 30. are significantly different from zero.13.020) 0.044) '.110 \1 P.93 (0.089273 43. dependent var 0.04726 F·sta1iS1ic 0.110 p coc.1281 3.023 e.9.95 (0.090501 Prob(Fstatistic) Figure 12.297417 1901. JB : 48.050) (0.011 ) R2 : 0.122 (0. 15. The option 'White HeteroskedasticityConsistent Standard Errors' has been used because of the heteroskedasti c disturbances.103 p co! ' (0 .0000 0.020590 0.0.40). ' 2 (0 .3 R2 : 0.13: The static regression Do not use the critical values and the pvalue that are shown in the output in Figure 12.CO! + 0.7.'1.c°f.04 (0. 14. we compare the value 3. even at the 2. BG (3) : 2.071) 0.041 ) '. So the correspond ing shortrun ECM to (12.1l3 Time: 13:34 Sample: 1960:01 2002:09 Included observations: 513 Variable Coefficient SId. White : 12.752 O.787794 S.85 (0.30 of the DFstatistic with the value .60908 1. Next <as step 2 in the EO procedure) a shortrun errorcorrection model is estimated by modelling from general to specific. The final result of that procedure is shown in Figure 12.46789 Schwarz criterion ·70. (0. The result is that the null hypothesis of no cointegrati on is rejected.
0010 O .667 + 0. the longrun implication of this model is found by the substitution of £ = 1: cof ?pc = 0.29 = 0.75. of regression Sum squared resid Log likelihood Akaike info criterion Schwarz criterion Durbi~Watson stat Figure 12. RESID_CDIN(I) D(RESID_COIN(·I» D(RESID_COIN(·2» Rsquared Adjusted Rsquared O. D.616282 Augmented DickeyFuller Test Equation Dependent Variable: D(RESID _COIN) Method: Least Squares Date: 0403.601650 0.044137 O.40) seems reasonably close (p .04aE ·3.\ll Time: 14:24 Sample(adjusted): 1960:04 "200m?2:09 Included observations: 510 after adjusting endpoints Variable Coefficient Std. dependent var S. As the quotient of the sample means of In (PtCOC ) and In t !) is 3.1099 0. the estimated longrun equation (12.630018 ·1 .121 . Next.0. Error IStatistic Prob.016 1 .E.328 + 0.cgo12 0.Fuller totes! applied on the cocoa residuals Rewrite this model by making use of the lag operator: Foc _ t  0.0444 Mean dependent var S. 1 .462£2 . The estimated longrun parameter is 0.110 + 0.293029 0.462 .0e:s5 ·2. whereas the solution of the unrestrictedly estimated shortrun model gives a lon~ CO run response of 0.208017 ·2.00200B 0.0B6623 3.103L p co! 1 .ocm 0.0.93.328£ + 0.941445 ·1.04461l8 0.110£3 t .1.90.Q1352B 0.99/ 4.071923 O.75Pt This resuJt looks like an acceptable relationship too.14: Dickey.0.1. MAXLAG=1B) t·Statistic Prob.'" 5% level 10% level '"MacKinnon (1996) onesided pvalues.995452 0.183109 1.Relationships between nonstationary variables 301 Null Hypothesis: RESID_COIN has a unit root Exogenous: None Lag Length: 2 (Automatic based on Ate .243821 566.297434 6.
981753 0.023308 0. dependent var Akaike info criterion Schwarz criterion Fstatistic Prob(F·statistic) Figure 12.000000 Log likelihood DurbinWatson stat Mean dependent var S.12.697. of regression Sum squared resid 0. the irreversibility of the ECM is investigated: has an in crease of the coffee price an identical influence on the cocoa price as a price decrease? The variables \7 + pta! and \7.134270 0.4.) Therefore. (See also the graph of the logs of Pt'°c and pta! in Figure 12.8971 .83%.309 has a probability of 57.011443 IStatistic Prob. .520133 ·2. Dynamic Models.713322 ·2.122385 0.D.oom Tim.16 the estimation result of the irreversible ECM is shown.: O4. The same result is obtained with the LRtest: 2 In (oX) = 2 (696. .066382 ·2.351634 ·0.684230 7. The result of the Ftest is given in Figure 12.15: The estimated errorcorrection model for In (P{OC ) to what can be expected.228711 2.4941 0.036847 0.0516) = 0.(l.110470 0.p ta! have the same influence on PtCOC is clearly not rejected with a probability of 57.002585 0.2009 1. The null hypothesis that V + p tCo / and V . In Figure 12.p ta! have already been computed in Section 12.13%.049567 0.127412 0. The null hypothesis is not rejected. Ptcoc To conclude this example.: 10:17 Sample(adjusted): 1960:04 2002:09 Included observations: 510 after adjusting endpoints White Heteroskedasticity·Consistent Standard Errors & Covariance Variable Coefficient Sid.841808 ·2. The point estimates show that an increase of pta! has more influence than a decrease. Error 0.046759 0. C D(LPCOCOA(·I» D(LPCOCOA(·2» D(LPCOFFEE) RESID_ COIN(· I) Rsquared Adjusted Rsquared S.0047 0.941778 696.043066 0.1 0. The LRstatistic has a X 2 (1) distribution.0422 0. the value 0. implying that the model can be estimated in reversible (symmetrical) form for the entire sample period. E.0000 0.0263 0.001881 0. we conclude that the cointegration analysis of the commodity prices has resulted in acceptable models for the investigated relationship between and pta!.671808 19. Unit Roots and Cointegration Dependent Variable: D(LPCOCOA) Method: Least Squares Oat.309.8971 1. but is that difference statistically significant? That question is quickly answered by computing a Wald F test or an LRtest. 17.58(l.002749 0.302 Chapter 12.
C(4) .0272 0.2556 0. of regression Sum squared resid Log likelihood Ourbi~Watson 0.981m 0.002052 1. we take up agai n the papers written relating to Cases 2 and 6.C(5) 0.71([(6 2.7 Case 10: cointegration analysis in this case.16: The estimated irreversible errorcorrection model Wald Test: Equation: EQ_SR Test Statistic Value df Probability Null Hypothesis Summary: Normalized Restriction ( O) Value Std.PCOCOA(1 )) O(l. the parameters of the original dynamic model from Case 2 were reestimated in a consistent way with 2SLS .011448 0. Err.17: Testing the irreversibility property 12.PCOCOA(2» OLPCOFP OLPCOFN RESIO_ COIN(I) Rsquared Adjusted Rsquared S.1383X1 2.351556 0..495245 2.109084 Restrictions are linear in coefficients.PCOCOA) Method: least Squares Date: 07mm Time: 12:42 Sample(adjusled): 1960:04 "'2002""':09 Included observations: 510 after adjusting endpoints 'iNhite HeteroskedasticityConsistent Standard Errors & Covariance Variable Coefficient Std. Figure 12.Q180 0.046r04 0. Your model is a dynamic model for the short run that will be analysed further. 0 .660190 15..215079 2.CIl6923 O.(J52666 0..109523 0.(J51801 0. dependent Yir Akaike info criterion Sct'l'Narz criterion Fstatistic Prob(Fstat istic} stat Figure 12.8 0.0516 1.076362 0. E.0447 0002585 0.00ll 0.003619 0. In Case 6.126211 0.D.148724 O .OOlDJ C O(l.373235 1. Error I·Statistic Prob.134794 0.1J56382 ·2.70405 O .012448 Mean dependent var S.00J465 0.02:lJ37 0.94IJ502 697.Case 10: cointegration analysis 303 Dependent Variable: D(l.128471 7. .049444 0..
All the variables have 10 be specified : D (X ) D(X (.2.2). X' .3) or similar. Practical remarks • Read once more Remarks 12. • • • What is your overall conclusion with respect to the presence or absence of cointegration? Write a clear paper about your analysis of the shortrun and longrun behav iour of the model. Next. by 10 be • • • estimating the static equation and testing the residuals for the presence of a unit root. and remember that an 1(0 ) variable does not form part of a longrun model but can be included in the shortrun model.3) can of course be used . If the existence of a longrelationship has not been rejected. the variable DX has been generated as the first difference of X" the n DX (Oto .304 Chapter 12. At the end of the exercise you can compare these results with the unrestrictedly computed longrun parameters. Dynamic Models.3. Then overall conclusions can be drawn. \l X t .2. Investigate whether a longrun relationship exists between the J (l ) variables. a cointegration analysis will be done along the lines as discussed in this chapter. • . Draw conclusions about the sense of reality of the different longrun results.3)). compare these longrun parameters with the computed longrun parameters of the unrestrictedly estimated model. The variables \l X" \l X'_lo \l X' . estimate a shortrun ECM.3 can be written as X (O10 .3 cannol be specified as D(X )(O 10 . Next. Pay attention to the following points. Unit Roots and Cointegration • First. Attention! In the 'Equation Specification' window in EViews il is not possible 10 specify lags for differenced variables in the same convenient way as can be done for variables in levels: X" X'_I. Decide which estimator can be used to obtain consistent results. Depending on the conclusions from the interim results it is possible that a decision can be taken to stop the analysis. for example. Analyse the data to determine which variables are integrated of first order and follow a unitroot process.2)) D(X (. Compare this ECM with the unrestrictedly estimated shortrun model. When.l )) D(X ( .3). because common trends do not occur. It is possible that not all the variables are 1(1). for example.2 and 12. in this case the complete exercise has done (see Remark 12. Summarise the results from the unitroot tests in a compacllable. compute the longrun multipliers and/or elasticities from that unrestrictedly estimated dynamic model. However. X '. Conclude whether these longrun parameters have realistic values from an economic point of view.
isturbance term has Ut ~ NID (O .. caused by the specification of a distributed lag.IX' _I + . Two types of distributed lag models and some related models will be discussed.. and the d. (13 . + . As an introduction to the concept of distributed lags. like model (13. are multicoWnearity. At tbe .3.2). ( 13.1 A distributed lag model is another dynamic causal model. the two kinds of distributed lag models for two variables are distinguished.(7~).~ 13. = f30 the usual assumptions: + 'o X . One concerns a model with many but a finite number of lags of an explanatory variable. Tbe parameters cannot. the variable X t is assumed to be exogenous. 2X. and too few or no degrees of freedom. The name distributed lag model will become clear in Section 13. + u. These problems and some pnssible solutions will be discussed in the following sections... 1) or Y. like model (13. and the other with an infinite number of lags of an explanatory variable.Chapter 13 Distributed Lag Models . The estimation problems.2 + .2) In both models. 1). be estimated. or cannot accurately.
306 Chapter 13. . Substitution of the polynomial in the model will replace the q parameters "Ii by the r parameters a i . with r < q. based on the economic theory. The procedure will reduce the number of parameters by a transformation of the model and also the multicollinearity problem. for the unknown parameters Ii: (13. 13..1. The method assumes that the form oftbe graph can be approximated by a polynomial in i of a suitable degree. If sufficient degrees of freedom are available. for two periods after which the influence fades away till it is zero after nine lags. These models are related to each other owing to the specification of a lagged dependent variable because of economic model building. 2. Distributed Lag Models end of this chapter. A wellknown solution to these problems is the method of Almon (1962). The idea behind the method is that you have knowledge about the development of the influence of the explanatory variable in time. as a Yj o 1 2 3 4 5 6 789 lag length i Figure 13. 1).1: cients A hypothetical path of lag coeffi .3) The approximation of the lag structure with this polynomial of degree r implies restrictions on the parameters "Ii . The Almon method suggests you approximate the 'graph' of the parameters "Ii against the lag length i by a continuous function. you have an idea about the form of the graph of the parameters "Ii for i = 0.· .2 A finite distributed lag model [n this section. An example of an arbitrarily chosen graph that suggests an increasing influence of X.1). is shown in Figure 13. 1. Consider the following polynomial of degree r . In other words. then the OLS estimator can be used to estimate the parameters of model (13. consistent and efficient result. with an unbiased. Probably the degree r will be larger than 2. an adequate estimation method will be discussed for the parameters of model (13. we will look at some specific dynamic models that are encountered in the econometrics literature. q. but inaccurate estimates can be caused by the multicollinearity problem.
. as the original model (13. + a 2 + (3) X ' _I + (aD+ 2a l + 4a2 + 8(3) X' _2 + (aD + 3a .A finite distributed lag model 307 seconddegree polynomial is a parabola. + q2 X' _q) +a3 (X'_l + 8X '_ 2 + 27 X' _3 + .1). + X .... 10'3 from model (13. + X' _I + X ' _2 + ..4) with OLS and compute the corresponding values for the "ii and their standard errors. (X' _I + 2X'_2 + 3X' _3 + . ( 13. _q.. . = f30 + a D (X .i. Next redefine the explanatory variables: ZtI = Z'2 = Z'3 = Z'4 = which gives the equation : X. The procedure will be explained by using an example where the degree of the polynomial r = 3. fade away over time. . + q2 X' _q X' _I + 8X' _2 + 27 X' . + 9a 2 + 27(3 ) X' . so the disturbanceterm assumptions have not changed by the respecification of the model. ._q X' _I + 2X'_2 + 3X'_3 + . + X ' _I + X ' ._q) +a . + q3X' _q) + u.... The new variables ZtI. = f30 + a oX.4) has the same disturbance term u.. + q3X. which is rather restrictive as it bas no inflexion point and so no ' tail' that makes it possible to let the influence of X. Just estimate the parameters 0'0._q) +a2 (X' _' + 4X'_2 + 9X'_3 + ... instead of the original q + 2 parameters! Model (13. + (aD + qal + la2 + l (3) X' _q + u. .. . Then rewrite this equation by collecting the variables for identical ai : Y. + qX'_q X ' _l + 4X'_2 + 9X' _3 + ... 1): + q2 a2 + q3 a3· y..2 + . + X . Then the original q parameters "Ii are expressed in the parameters ai in the following way: aD "II = aD + al + a2 + a3 "12 = a D+ 2a l + 4a2 + 8a3 "13 = a D+ 3a l + 9a2 + 27a3 "10 = • • "Iq = aD + qa l Substitute these expressions in model (13.. Z'4 will be less correlated than the original lagged variables X ' . + (aD+ a. .4) This model has five unknown parameters to estimate... . .3 + .3 + . + qX. implying fewer or no multicollinearity problems..
K ) q r.z + Ztl) + az (Z'3 .4): (13.az + a3) Ztl + alZ. the validity of the Almon procedure can be tested with a standard Ftest as model (13. [n case of residual autocorrelation being absent. Substituting the restrictions in (13.4). which gives the restricted (S n) and unrestricted (S) sum of residual squares. az.n .I = 0 and/or "1. If both endpoint .5) Y.z + a z Z'3 + a3Z'4 + u. These restrictions are substituted in model ( 13. we do not know the degree of the polynomial. a3 in the restriction ( 13. Also. for example.. then substitute aD in equation ( 13. This exercise results in fewer regressors in equation (13. f30 + al (Z.3) gives: and/or aD + al (q + 1) + az (q + l )z + a3 (q + 1)3 = O. for example.1 ). The Almon procedure does not always directly give a satisfactory result. This yields the restricted regressio n equation : with the new variables W ti defined as: Wtl = Z.Ztl W'3 = Z'4 + Ztl· The parameters f3o. if only a nearend restriction is imposed.z = Z'3 . aI.z + Ztl W.Ztl) + a3 (Z'4 + Ztl) + u.4) is a restricted model with regard to model (13.308 Chapter 13.5). but the Student's Hest can be used to test the significance of the last parameter a3 in equation ( 13.+1 = O. For example.4). Estimate both mode ls. the "Ii start at a level that is too high from an economic point of view o r the graph of the "Ii decreases too slowly to zero. this tstatistic gives information about the suitability of the degree of the used polynomial. after which aD is computed by substi tution of the aI. Distributed Lag Models In practice. = f30 = + (al . It is possible that the estimation result is unsati sfactory if. az and a3 are estimated. and compute the F statistic: F = (SnS)/(qr) I!JF( K) S/(n. This is done by setting: "I.4) and have to be elaborated. with OLS. Then thi s behaviour can be ' repaired' by restricting the graph (and so the parameters) by the imposition of a 'nearend' and/or 'farend' restriction on one or two parameters that do not occur in the model.
all the Q statistics bave probabilities equal to zero. Both the polynomials are of degree 4 and have been specified for 12 lags. with the following explanation: m = 0. For example. the residuals are hi ghly autocorrelated. the specification in the window is done as follows (but with concrete numbers for q and r in parenthesis): Y C PDL (X. The parameters of the di stributed lags are calculated together with their standard errors and tvalues. In Figure 13a. q.r). To end this section.PDL08 for P("f. q.4). The Almon procedure is applied to the relationship between the commodity prices. especially as it makes no difference whether the example is successful or not in showing the output of the Almon procedure that is given by EViews. This example will show a failed attempt to find a good fit to the data by using the Almon procedure. the regression output is shown. m = 2: farend restriction. The 'variables' PDLOl . r.1 = 0). The results. The specification in the 'Equation Specification' window is shown in Figure 13. The procedure is used by specifying a polynomial distributed lag (PDL) in the ' Equation' window. Only ' far end ' restricti ons have been specified. (79+ 1 = 0). showing the lag patterns. are given in Figures 13b and 13c. [ ) is estimated unrestrictedly with Almon's method. !!f clearly have more C influence.co! and P(~ (in logs) have an influence on the price formation of cocoa by specifying polynomial di stributed lags for both the variables. The specification P D L (X . However. 'Nearend' and 'farend' restrictions can be imposed by specifying one more parameter m: Y C PDL (X . q. [n empirical work. From the estimation results we see that only two coffee prices have a significant . In empirical econometric research we are often confronted with failures so it is good to give some unsuccessful examples in a textbook. We will investigate whether more lags of p. This implies that the parameters of this specification have been inconsistently estimated.2 and the output has been split up in Figures 13a to 13c. m = 3: both ends are restricted. PDLOl . m = 1: nearend restriction. (7. the application of the Almon procedure is very simple in EViews.PDL04 for Pto! and PDL05. r) means that a polynomial di stributed lag is applied to X that has been specified with q lags in the origi nal model and that will be estimated by using a polynomial of degree r for the parameters. where also the estimated original parameters have been plotted. but the derivation is done in a similar way.A finite distributed lag model 309 restrictions are imposed one more regressor disappears. an example is given of the use and the output of the Almon procedure in EViews. m) . or no m specified: no restrictions. if model ( 13.PDL08 are the transformed variables Z" from equation (13. Lags of P. which are asymptotic because of the lagged dependent variables.
004600 0. The only conclusion is that the estimation results.024815 0.2: Specification of a polynomial distributed lag for ~of and PtC~{ (in logs) influence.(Jj(Jj18 Prob.0.000189 2.(X8127 .0. Dependenl Veriable: LPCOCOA Method: Least Squares Dale: 04114m Time: 13:20 Sample(adjusledj: 1961 :02 ')I2002m:09 Included observations: 500 after adjusting endpoints Variable Coefficient 0.544 O .OOW8 0. indicate that the PDL specification is not a good method for determining the relationship between p tCO ! and P.0. .OO PDlD7 PDLD8 Rsquared Adjusted Rsquared S.698761 .003496 0.IllEIli .(ll9687 ·{J.310 Chapter 13.c°c.004288 ·1.OCOOXI Mean dependent var S.004225 0. just as before.011796 0.001534 0.3a: The regression output . Distributed Lag Models Figure 13. Error tStatistic 1.072722 0.000174 0.0D8829 0.599282 0.45517 OJXm O .310642 4177. the use of EViews' PDL procedure has been clarified by this example. 0.CXl5297 0. because of the residual serial correlation.61603 5.5097 0.000185 0.053430 O.4850 C PDLDI PDlD2 PDLD3 PDlD4 PDLD5 PDl. of regression Sum squared resid 0.386505 2.CO)) 0. However.caw3 Std. dependent var Akaike info criterion Schwarz criterion Log likelihood DurbinWatson stat Fstatistic Prob(F·stat istic) Figure 13.002952 . But we cannot draw any conclusion from this result.s.0000 4.E.015758 11 .010957 0.D.1104 0.5995OD 2.985285 0.802482 18.
0 00252 .01549 . . ~ NID (O .o.00758 .04281 0.91202 ·9.00164 0.01240 0.2 + .65962 . .00021 .00771 0.02157 o00523 Sum of Lags Figure 13. + 'Y IX'.06937 0.3c: The estimated polynomial distributed lag of In (ptOC ) 13.0. Error l·Statistic 0.5'm5 5.00121 .02764 .0CB33 0.44513 6. .0. .00796 0.00581 0.. .l u. .41235 ug Distribution of LPeOFFEE i Coefficient .93192 2.00928 0.00942 .00848 0.06790 8.0.00207 .6) It is clear that the infinite number of parameters ri cannot be estimated without making further assumptions about the lag pattern.00849 0. .00803 0.01792 .04803 .3550 37.78727 ·7.3 Infinite distributed lags Model (13.6!BJ 1.04553 .00854 0 .00548 0.1Dl91 0.00679 0.17965 .80291 5 0.0.00014 0.02407 0. C7~) ..00057 0.0..oca:l5 .0.72727 64.0. Such assumptions and accompanying procedures.0.41467 2. .66017 ·1.07434 0.01780 0.05935 0.02216 0.0.00748 0. 0 1 2 3 4 5 6 7 8 9 10 11 12 0.0.05090 .00711 0.01868 . i Coefficient Std. + u" ( 13.06600 0.01534 1.01149 0 .0.00571 0.00530 ..00879 0.3b: The estimated polynomial distributed lag of In (Ptea!) Lag Distribution of LPCOCO .0.00826 0. 0 1 2 3 4 5 6 7 8 9 10 11 12 0. + 'Y2X' .96737 Sum of Lage Figure 13.08842 .00768 0. : Y.08629 .0.28235 .6) is a model that is specified with an infinite distributed lag pattern for the explanatory variable X.04052 0.18326 7.01145 0.0.0.0.0.96748 ·5.7820 .(0026 . = f30 + "loX. .01576 6.0. . \ .0. Error l·Statistic 0.20578 1.0.14693 .Infinite distributed lags 311 Sid.01248 0.0.01493 ' 93.00978 0.11046 1 .
but we will limit ourselves to the geometric distribution. With this distribution a declining influence of the exogenous variable can be specified.6) once more in general notation: = {Jo+ L i= O 00 "IiX'~i + u" (13. In that way. Various probability distributions can be used to solve thi s distributed lag problem. which is discussed as an example of such a procedure. which can be an acceptable and realistic assumption. A wellknown and oftenused ctistribution is the geometric distribution. This explains the name 'ctistributed lag'. declines geometrically in the course of time. the number of parameters is reduced. geo metrically declines over time. These probabilities are linked to the parameters in a simple way: there is a constant "( such that: The parameters "Ii are replaced by Pi"! . For example.6). the parameters must be positive and sum to one. Koyck (1954) assumed that the influence of the exogenous variable X . economic theory idea about the way the variable X. I and i = 1. Next the formula of the probability ctistribution for the Pi can be substituted. Write the specification of the ctistributed lag model ( 13. ~ NID (O . The form of the lag structure is in accordance with the . on Y.7) u .ies for the parameters Ii that a parameter>. Histori cally.. influences the dependent variable Y.. will briefly be discussed in this section.. Geometrically distributed lags The geometrically distributed lag model assumes that the influence of a change in X. 3. . we can assume that the form of the lag pattern is comparable with the form of some probability distribution. The form of the chosen ctistribution is used as an approximation for the theoretical lag structure. with Pi > 0 for all i . so they have to be rescaled. Distributed Lag Models which make it possible to estimate the parameters. This assumption implies that the influence of the explanatory variable geometrically fades away over time. exists such that: "Ii = A"Ii~ l' with IA < I . In general. Denote the probabilities with Pi . When the assumption has been made that the parameters stem from a probability distribution. by substituting the parameter(s) of that probability ctistribution in the model (13. . O'~). That impl. 2. The number of parameters "Ii can be reduced. the parameters will not sum to one. over time. then L Pi = I .312 Chapter 13. A suitable probability distribution is a distribution with a form that corresponds to the expected shape for the course of the parameters. A solution to this problem is to make assumptions about the form of the lag structure.
as introduced in Chapter 12: 00 l _\LX. = L i=O A' X t<.8) in the model results in the respecified model (13.9) is a nonlinear model in the parameters that cannot be estimated with OLS. I The infinite number of parameters of the original model has been reduced to three. IAI < 1. = i30 + 'Yo L i=O A' X' _i + u. = f30 + 'Yo 1 _ AL X. y. 1. = i30 + 'Yo X. First. it is more correct to use explicitly the formula of the geometric distribution. =i30+'Yo (1 +AL + A2L 2 + . 'Y.. + A'Yo LX. + u" = i30 + 'Y (1 .9) defines the sum of a geometrically declining timeseries.. rescale the probabilities to link them to the parameters: 'Yi = Pi"Y.. The notation of equation ( 13.10) in the model : 00 (13. + u. . Because of the solution mentioned in the introduction. + . .Infinite distributed lags Recursive substitution of thi s relationship gives: 313 (13. with IA < 1. . + A2'YoL2 X.9) This is the geometrically distributed lag model in its standard notation with the lag operator. Pi = (1 . + u. model ( 13.AP' X.8) This suggests that the infinite number of parameters 'Yi can be replaced by two parameters 'Yo and A. 10) Y. Y.+u.9): 00 Y. . y. _. = i30+ L 1=0 A''YO X' i 00 + u. 2. = 'Y( 1 . Substitution of (13. ( 13. However. )X.A) L i=O Ai Xt< . . y. + u. for i = 0. Then everything follows in similar way. = i30 + L i=O 'Y (1 .AP' Next substi tute (13. using a probability distribution.Api .
How can the parameters of model (13. where the GLS estimator was also introduced.9). so this model has an autocorrelated di sturban ce term that belongs 10 the model. The longrun response and the mean lag The longrun response of Yt to a unit change of X. The disturbance term Vt is an MA( I) process. The mean lag can also be used to check the . stati stic that can be used as an evaluation tool with dynamic models is the mean lag. + (1 . and a restriction exists that the MA( I) parameter of u.d (13.9). 11 ) has more problems than just autocorrelation.>" One more convenient. However equation ( 13.>.. = "Yo + 1 _ >'L X.5. the parameters were estimated by transforming model ( 13. They concern methods. like search and iterative methods.L ) + "YoX. In Section 8. and the AR( I) parameter of Yt are identical. 11 ) However. When this model is estimated with OLS in the form : Y. inconsistent and inefficient.0 Y.>'L) u. + >'Ytl + (u. Solutions for all the estimation problems can be found in.1 + v.314 Chapter 13. + u..u. This model was introduced in Section 8.>'L) : Yt (1 .5.9) be estimated? Historically.>'L) Yt = = . = 0'1 + 0'2 X. is found by substituting L = 1 in model ( 13. to find the maximum of the likelihood function of the non. However.. I> >. So the OLS or GLS estimator cannot be used in a correct way. AUt _ I . this can be demonstrated by 'multiplying' both sides of equation ( 13. but can be applied by writing a program of EViews procedures.transformed model (13. Dhrymes ( 1971).9) with (1 .>. These procedures are not standard in EViews. tbe statement was made that no a priori reasons exist to assume an AR( I) disturbance term.) are dependently distributed.>. two more aspects of this model will be introduced. which can be used for the evaluation of the obtained estimated model. Finally.l ) and the disturbance term (v. f30 (1 . this is not a simple equation.) + "Yo X. Distributed Lag Models with the parameter 1'0 written as: "Yo = "y( I .9) and is equal to the following expression: "Yo 1 . ( 13.) the result is identical to model (13. and simply to compute. The lagged dependent variable (Yt. f30 (1 . + 0'3Y'. With the use of instrumental variables (IV) only consistent estimates can be obtained.6). 12) Vt = Ut  then the OLS estimates are biased. but that an MA( I) disturbance is possible after a transformation of the model. for example. these methods are beyond the scope of this book and are not discussed further.
and calculate separately the numerator (Snum) and the denominator (Sdenom) .13). the mean lag can be computed and evaluated when formula (13..:. Here the mean lag will be computed for the geometrically distributed lags.13) In practice.A). Sdenom · The denominator Sdenom is computed straightforwardly: The numerator Snum is computed in the following way: Snum = LiAi = A+2A2 +3A 3 + 4A4 + .see.. for example.. Substitute the formula of the geometric distribution in ( 13.15) Snum . We have an idea about the mean number of periods that the adjustment to a shock from an exogenous variable on the dependent variable will last. .14) (13.36): m eanJag = "f 1 .ASnum = Snum (1 .. the formula for the mean lag is: m eanJag = (13. IA Also it is true that: Snum .A) = (1 _ A) A Snum = (1 _ A)2' . In general notation.ASnum = 2 A + A + A3 + A4 A .. Combine the two equations to determine Snum: A Snum (1 .14)(13. Next compute the difference (13. ASnum = A2 + 2A 3 + 3A 4 + .13) has been made concrete and after the parameters have been estimated. Stewart and Wallis (1981: p.15): (13..  Snum .Infinite distributed lags 315 economic reliability conceming the assumption about the distribution of the lags.
9) in a less restrictive way by assuming that the geometric decline does not start immediately but starts after several periods.l for i = 3.4. .16).5.l .1 in the equation : (13. Special attention will be paid to their economic interpretation and to the disturbance term.16) Five hypotheses are discussed in this section.1 It is possible to estimate model (13. 13. and "'12 an: estimated unrestrictedly.316 Chapter 13. The disturbance term Vt is either just randomly specified or it .4 Models with a lagged dependent variable In the previous section. which results in the following model: Yo = "" + "'loX.lt in a specification with a lagged dependent variable.6).A) In practice. The question that is rai sed in this section is: what is the economic interpretation of Yt .. Distributed Lag Models Next compute the mean lag: A/ (IA )2 meanJag = 1/ (1 . because of a geometricaUy distributed lag.. Remark 13.. 12) by transforming the original mode l (13.12) have been estimated._ +1_ "'12 AL X t + tit· The parameters "'10. . Theoretical reasons exist for the 'popular' specification of a lagged dependellt variable. which are all examples of models that result in a specification such as (13. Popular because a lagged dependent variable often eliminates firstorder autocorrelation when we have started with the specification of a static model (which is not always the best idea to start with). we look at some more economic models that resu. "'11. we saw that a lagged dependent variable appeared in model ( 13. For example: with 1'i = A1'i .l divided by I minus the value of that coefficient. the mean lag can easily be computed when the three parameters of model ( 13. + "'I1 X t. In this section. The mean lag is equal to the coefficient of Yt .
Then we make the usual assumptions for the disturbance term Ut.. V t . after a change in X" and specify a partial adjustment mechanism . does not completely adj ust to a change in X. 16) is a transfonned model that has arisen by transfonning a causal model with a geometric di stributed lag (with parameter .. 0 ). Vt = Ut . The disturbance tenn v. (1~). is not a transfonned disturbance tenn. The original model is one with a partial adjustment mechanism. !f model (13. Partial adjustment The model that will be discussed as the third model with specification (1 3. A geometrical distributed lag A second model with the specification ( 13. it makes the economic assumption that a change in X . !fthe dependent variable Y.16) is correct with respect to one of these hypotheses and the estimated parameters have economically relevant values. model ( 13. It is a model with a clear economic interpretation.. u. Just a dynamic model has been specified.x) for an explanatory variable.' as the desired level of Y.Models with a lagged dependent variable 317 is a transfonned disturbance tenn. probably with more dynamics for the other involved variables also. with Ut as the disturbance term of the original model: u ~ N (0. 16) can directly be specified by the researcherto show habit formation in the behaviour of economic subjects. in the same period t. It shows the stochastic trend in the dependent variable. and OLS can be used to obtain consistent estimates of the parameters. in period t is fully reflected in yt in the same period t. OLS results in inconsistent estimates of the parameters. v ~ N (0..AUt _ I. then the mean lag is: fit 1 fil ' and the longrun response is: Habit formation First. There we saw that the disturbance term V t is an MA(l) process.In) . N I D (0 .1 6) is also the result of a tran sformation. for example.. The model has been transformed according to Koyck's method as was shown in the previous section . then we specify Y.
 Y. + 1 is a .ll is specified as part of the desired adjusUTIem (Y.1l = A(Y. because the transformation does not affect the disturbance term u.l is not the adjusUTIent parameter.Y'.: (Y. The adaptive expectations mechanism is formulated as: (13. are independently distributed. and there is slow adjusUTIent with a value of A close to zero. Distributed Lag Models for the real change \7Y.18) yields a specification similar to ( 13.X~I'_ I) thai was made in the previous period. for period t + 1 that is formed in period t.18) Ut ~ NID (O . . By substituting this assumption in (13.l and the disturbance term u.'  V. . + (1  A) Y.19) the variable X:+ l it can be eliminated. A value close to zero of the estimated parameter OfY.20) means that the change in the expectation from period t to t fraction of the expectations error (X. . : (13.' is determined by the static equation without a disturbance term: Y. and a random disturbance term u .' = i3. an historical model that has not disappeared from econometrics literature will be mentioned. Notice that the parameter (1 . In this equation.l to Y. The desired level Y. No observations are available for this variable X~+ l l " So an assumption about the way the expectation is formed has to be made. When A is close toone.. is a fraction of the desired change. Substitution of the desired level (13.. The real change from Y.l + u.. and so a fast adjustment process!  Expectation variables With this fourth model.1l +u. the adjusUTIent is fast. but still with the original disturbance term: v.l ) .318 Chapter 13.17) and the actual adjusUTIent (Y.' .. + {h X " (13. = Ai3. (13. In the past. a~ ) . 0 < A < 1. the lagged endogenous variable Y'. lts parameters can consistently be estimated with OLS.A) of Y'.19) where X~+ l l' is the expectation of X . _1 implies a value close to one for the estimated adjustment parameter A.16).. it was not unusual to assume an adaptive expectations mechanism.Y.20) Equation (13. + A{h X .. 17) in the adjusUTIem equation (13. The original model is one which has an explanatory variable that represents an expected future value of an exogenous variable X .
48».A) L ) u" Y.(1 .d .A)U' _I.X .(1 .21 ) Again the specification of model ( 13.+I I' in a variable that is observed: (X. AX. Error<:orrection models The last example of a model with a lagged endogenous variable in the specification is an errorcorrection model (ECM). X.A) L ) Y. So. identical estimation problems arise to those with the geometric distributed lag model after using Koyck's transformation. AX.+II' = AX. Equation ( 13.. X.I' _I = AX. + (1 . = and rearrange the variables: /31 (1 .A) u. A more general approach is to identify and estimate a suitable ARlMA model for the variable X.(1 .(1 . The procedure of using adaptive expectations is rather oldfashioned and restrictive. ( 13.I' _' (X.+II' . X.(1 . .A) LX. 16) has been obtained as the result of a transformation.A) L)' Substitute this expression in the original equation: Multiply both sides of the equation with (1 .(1 .+II' = AX.I' _I) + AX.+II' . = /31(1 .(1  A)) + {3. = u.Models with a lagged dependent variable 319 Rewrite equation (13 .(1.(1. It can be shown that the result of the adaptive hypothesis is comparable wilh the specification of a specific timeseries model for X " an lMA( I.+II' .A) L): (1 . v.A) L ) X. .+II' .A)V' _I (see.A)) + {3.2 1) has a transformed disturbance term that has an MA(I) process. The ARlMA models are discussed in Chapter 14.X.A) X.I + (u.I' _') = A X..X . for example Stewart and WaUis ( 198 1: p. = v. + (1  A) Y. An errorcorrection model can have a specification with a lagged dependent variable when the degree of the lag polynomial of the dependent . AX. and to compute forecasts with that model that wiu serve as 'observations' for the expectations.(1 .20) to get an expression for X.(1 . (1 . '+II' . .  (1 . I) process: \lX.(1 .1tI = AX.
q~) . That result will yield relevant information about the interpretation of the parameter of the errorcorrection term. . Ut ~ NID (O.4.J + Ut . We have seen that the disturbance term Ut of the ECM has not been transformed. A low value of {3J implies a fast adjustment to the long. Its value indicates the speed of convergence to the equilibrium relationship. Now we know the interpretation of {3J .O"~) .O"~) .run equilibrium. = X t and the restricted errorcorrection model for this relationship was: 'VY.J + {3J Yi. Knowledge about the interpretation of the value of the parameter {3J that has been obtained when discussi ng the partial adjustment model can be used here to derive a property of the parameter of the errorcorrection term. = {30 + 1'0Xt + I'J X t .4: y. Distributed Lag Models variable is larger than or equal to one.320 Chapter 13. The longrun equilibrium in the example was Y. .Xt. and so it bas the usual assumptions: Ut ' " N I D (0.d + Ut Ut ~ N ID (O. Once again: a low value of {3J is identical to a coefficient close to one for the errorcorrection term ({3. This will be demonstrated with a si mple example. Remember the unrestricted bivariate model with firstorder lag polynomials. That is useful knowledge for empirical econometricians. = {30 + 1'0'V X t + ({3J . as has been shown in Section 12.J . So fast adjustment to the longrun equilibrium occurs when the parameter of the errorcorrection term is close to one. as discussed in Section 12.1) (Yt .1).
AR(P) and MA( I) processes for the disturbance term of a causal model. The variable Yt is considered as a stochastic process and the timeseries with observations on Yt are realisations of that process. a type of model that is clearly different from the causal models that have been discussed so far is considered. we became familiar with the concepts of AR(I ). for example. Besides an ARIMA model. One more advantage is that only data for that variable have to be collected.. It concerns the autoregressive integrated moving average models (ARIMA models). 1975)). An accessible comprehensive introduction to many aspects of timeseries analysis is. In previous chapters. We will not consider multivariate timeseries models in this book..1 In thi s chapter. Why have these models been developed? The idea behind them is that it is sometimes possible to get more accurate predictions for a variable with a TS model than with a causal model.Chapter 14 Univariate TimeSeries 14. Zellner and Palm (1974. In case the objective of the research is to produce predictions for an economic variable you can consider estimating a TS model for that variable. do not think that two completely different models exist that explain and forecast the variable Y. A univariate timeseries (TS) model explains one variable (e. in general notation) with regard to its own past and the history of the random disturbance term u. In thi s chapter. Y. which are stati stical models. for example. for example. However. That is. The . instead limiting ourselves to the discussion of univariate ARlMA models.g. a model with a deterministic or exponential trend as explanatory variables only. but also for a vector of variables in a multivariate context. a deterministic TS model can be distinguished. These models can be considered for one variable. Chatfield ( 1996). the stochastic TS models are discussed. There is a link between a causal model and TS model for an economic variable (see.
has been filtered to a white noise process. So when the observed Y. The input Ut is not observed. (T~) for all observations. .1(L ) has to exist. = + 1/I2L2 + 1/I3L3 + ". filter U" " with the condition that the inverse 1/1.. . As an introduction to the models.. .. a. ut + 1/11 _ I + 1j!. .322 Chapter 14. Y. The output series Yt is observed as a realisation of a stationary stochastic process that is assumed to be a Gaussian process: y ~ N (0. For example. as the inverse of the fi Iter. as the normal distribution is completely determined when . The linguistic terminology in the literature of timeseries analysis is somewhat different from what has been used so far around causal models. This process can be written in an equation. 11). and Jenkins (1976».2) ut ~ NID(O . or with a constant in the model: Without loss of generality. The necessary conditions of being invertible and stationary of the process imply restrictions on the parameters of the model. 11/1 1 (L ) 1 .) into a model (1/1 (L )) that results in an output series Y. U.. To check the stationarity of a Gaussian process. =1/I (L)u" with 1/1 (L) written as an infinite lag polynomial: ' 1/1 (L ) = Y... it is suffi cient to check that the process is stationary of the second order. Their methods will be explained in this chapter and the use of these procedures in EViews will be clarified.. possibly with a constant tenn.1): Y. 11/I (L) m odel I . the model has been found.. To delermine the model 1/1 (L ) consider the opposite direction of the process (14. .. the process ut is called a 'white noise process' because of the assumption that it is N I D (0. such as: Y.u'_2 + . Univariate TimeSeries Models TS models have originally been developed by Box and Jenkins in the 1960s (see Box .. (14. it is assumed that the mean is zero in the foUowing discussions.. is 'filtered' to become a whitenoise process: Y. ..1 ) There is a random input (u... the idea of the linear jilter model (or linear stati stical model) is briefly di scussed. U' 1 + 1/IIL ( 14.) forallt. That model can be represented schematically in the following way.
= B (L) u. as was mentioned in Chapter 12. .. as seen in Section 12.q) model without a constant term... by taking the logs (In).q) model for 'Vyt..p~ + .CP1L . In case of one unit root.cpiL .pi +. cp (L) = (1 .2) = E (un + ._IU.L ) (1 .. Nonstationarity of yt nnight be caused by a unit root in cP (L). For example. When d = 1 the variable yt has a linear trend (constant growth) and wben d = 2 there is a trend in the first differences. ) + . will also converge to a finite The variance exists as the series 1 constant...5. = = 0.. +Bq LQ. CP.Introduction 323 the first two moments of the distribution are known. An ARMA(p.. Mostly this is done by using transformatio ns that we considered in Chapter 12.p~ + 1/>~ + ... tbe polynonnial cp (L) can be written as: . The process is stationary by imposing restrictions on the parameters 1/Ji in such a way thatthe variance of Y.. (14. lfthe variable yt is not stationary (and most econonnic timeseries are no t).piE (U~_ I) + .q) process for yt: cpO (L) 'VY.CP2L2 . L2 .. converges to a finite constant: E (1'. for a stationary variable yt . • . like differencing ('V d with d < 2). orby doing both ('Vdln).. . = a~ (1 +..l) = (1 .L )cp' (L) = 'Vcp' (L).cppLP B(L) = I +B 1 L+B2 L 2 + .pi + .3) The model bas an autoregressive polynonnial cp (L) of degree p and a movingaverage polynomial B (L) of degree q that are written as usual: cp (L) = 1 .. ) = E(u.p~ E (U. so a linear trend in tbe growth rate.. the variable has to be transformed to a stationary variable.2): E(Y..p~ + ....1+1/>2U'2 + ._2) + ..p3+"') if the series 1 + 1/>1 + . ) 0 ·(1 + 1/>1 + 1/>2 +. with the condition d < 2 for econonnic reasons. wbich is defined as an ARIMA(p. I .p2 + . consider the mean and variance of model ( 14..3) and we bave an ARMA(pl...p3 + . is constant. +1/>1"'...cp.. Substitute this in model (14.. is usually defined as: cp(L)yt=B(L)u.
2 Timeseries models and some properties In this section.1 (L) converge. without a constant term. Univariate TimeSeries Models The process is integrated of order one.a pure AR and MA model. because 'P (L ) has one unit root. In the next section . is written as: or 'P (L ) yt = u" with ( 14.d. Y.4) are such that the coefficients of the infinite polynomial 'P.324 Chapter 14. . The range for the parameters where the AR(P) model is stationary can be determined a priori. an autoregressive integrated moving average process: 'P( L ) Y. the process is always invertible.2 for tbe AR(I) model and is shown here once more to complete the overview of this section.1 + u. = I1(L)u. we look at the properties of some TS models . and the model can be written as an MA (oo) process: The ARprocess is stationary if the parameters 'Pi of (14. a process with d (d < 2) unit roots in the autoregressive polynomial is called an ARIMA(p.q) process. 14.4) We know already that the process is stationary if all the p solutions of 'P (L ) = 0 are outside the unit circle. MA and ARMA models will be considered first. 'P' (L ) Vdyt = 11 (L ) u" ==} with 'P' (L) a polynomial of order p . = 'P I Y' . some examples of AR.d. In general terms. and the mixed ARMA model. which is identical with the previous mentioned condition for the roots of 'P (L ) = 0 lying outside the unit circle. Because this polynomial is a finite polynomial. The conditions for stationarity and invertibility of the polynomials are discussed in a more concrete manner. Autoregressive models The AR(P) model. This has been done earlier in Section 12.
is written as: or with the MA polynomial 8 (L) : Y t = 8 (L )Ut.5) The interested reader can find detailed descriptions concerning the computation of these inequalities in econometric books such as Anderson (1976).'P IL ) Y t = "t so solve L from : 'P (L ) = 1 . The same can be done for the AR(2) model : Y t = 'P I Yt.2 + "t. The MA(q) model.'P 1L = 0. Chatfield (1996).'P1 L 1 .Timeseries models and some properties or written with the polynomial 'P (L) : 325 (1 .I + 'P2 Yi. L= giving the restriction for 'PI : 2'P I ILl> 1 = l'Pd < 1. Moving average models Similar examples can be given for moving average models. without an intercept. Box and Jenkins ( 1976).1 < 'P2 'P I + 'P2 < 1 'P I + 'P2 < 1. Greene (2000) and Hamilton ( 1994). ( 14. or again written with the lag polynomial 'P (L): It is possible to derive the following region for the parameters where the AR(2) model is stationary: . with .
are finite constants. : 1 . the rool of 8d the polynomial 8 (L ) has to lie outside the unit circle: 8 (L ) = 1 + 81 L 1 + 81 L = 0 L = .81 . restrictions for the parameters can be derived in a similar way: or with the lag polynomial: The conditions for invertibility of the MA polynomial resemble the stationarity conditions of the AR(2) process (14.82 < 1 8.82 < l.l Yt = (1 +8. + 8~ + . This means that the moments of the distribution of Y.8 This gives the restriction for 8. + 8. and this is identical to the condition that the q solutions of 8 (L) = 0 are all outside the unit circle. 1 For the MA(2) model. look at the MA(l ) process: Y. the other way around..1 < 82 . .1 (L ) converge.). 1 U. L )u.It =U u 2 + 82 2 + 82 2 82 2 lU 2Uu+"'+ qa u u = (1~ (1 + 8. The MA (q) process is invertibleif the coefficients of the infinite polynomial 8.5) except for the signs of the parameters: .. Univariate TimeSeries Models The process is always stationary because the polynomial 8 (L) is a finite polynomial. . 1 +8 L Y. the mean and variance: E(Yt) = 0 • E( v2) . as can be checked quickly by looking at. = 1 • The infinite sum (1 + 81L) .1 converges if 1 < lor. As a first example. for example.326 Chapter 14. = or with a lag polynomial : Ut + 81Ut. .
4 and the PACF will be discussed in Section 14. Matching functions makes it possible to identify some ARIMA model. The Box. we are sure that they comply with the conditions.1 (L ) <p (L) Yi. Both functions are necessary for the identification of the TS model. Estimation of the parameters of the polynomials cp( L) and O( L) After an initial identification.Jenkins procedures provide us with tool s that are useful in identifying the order of the i. If the variable is not stationary it has to be transformed into first clifferences or first differences of the logs of the observations. the parameters of the selected TS model will be estimated. we look at the ARMA(p. In thi s section. Theoretical forms of the ACF and PACF of many TS models have been determined and are known in the econometrics literature. As invertibility of the polynomial 0 (L ) has been required. because the variable has been transformed to become stationary.5. <P (L) Yi = 0 (L ) Ut· It will be clear that the conditions for stationarity and invertibility are a combination of the conditions fo r the AR and MA models. The pictures of the sample functions are visually compared with figures showing forms of known theoretical ACFs and PACFs. which is a mix of the previously introduced AR and MA model. The roots of both the polynomials <P (L) and 0 (L) have to be outs ide the unit circle.q) model is written as ( 14. In practice. the Box.3 BoxJenkins procedures In practice. The ARMA(p. we do not know the orders p and q of the lag polynomials <p (L ) and 0 (L). Examples are given in Section 14. A useful property of the estimation procedure in EViews is that the (inverse) roots of the polynomials are printed in the output. model (14.3). The tools to determine p and q are the autocorrelation fu nction (ACF) and the partial autocorre lation function (PACF). which is the final goal of this exercise. polynomials. With regard to the roots of the AR polynomial. PACF and the theoretical TS model is unique.q) model. Identification of the TS model. resulting in restrictions on the parameters. First the stationarity of the variable has to be checked. Each TS model has a unique theoretical ACF and PACF. 14. The ACF is introduced in Section 14.6.BoxJenkins procedures 327 Autoregressive moving average models Finally. by applying an ADF unitroot test. the sample ACF and the sample PACF of the timeseries are computed. Then we will estimate the parameters <Pi and O and check the residuals of the TS model as to whether or not they are white noise. Then we will compute predictions for the concerned variable. Next. the orders of the polynomials p and q have to be determined.3) can be rewritten as: Ut = 0. . The relationship between ACF.Jenkins procedures are discussed in the following four stages. making it possible to check the conditions. for example.
Then you can check whether these inverted roots are within the unit circle. This expression is used for forecasting Y. The concept of stationarity was introduced in Chapter 12 and is considered here in more detail.3) can be written as: Y. the ACF and the PACF.4 The autocorrelation function (ACF) The autocorrelation function had been introduced in Section 6. 14. so model (14.Jenkins procedures is discussed and clarified with examples.1 (L ) 8 (L ) u" or with the estimated parameters as: Y. If residual autocorrelation is absent.1 (L )8 (L )u. But remember that data from outside the estimation period cannot be used to test the quality of the estimated model for the sample. EViews estimates the parameters with a nonlinear optimisation procedure and computes the roots of the polynomials which are printed as ' inverred roots'. Therefore. = '1'. the significance of the parameter estimates can be considered. A stochastic process is said to be covariance stationary if the means. shorrrun postsample forecasts can be   computed with their prediction errors. The presence of autocorrelation will have been caused by model misspecification. to test the null hypothesis of an independently distributed disturbance term. are introduced and illustrated with examples. The earlier introduced Qtest of LjungBox will be used to test the hypothesis of random residuals. All these procedures are implemented in EViews.328 Chapter 14. = '1'. Forecasting The polynomial 'I' (L ) is invertible.3 as a tool for the LjungBox Qtest. When the results are satisfactory. In the TSanalysis literature this step is called 'diagnostic checking'. after which the econometric practice of the Box. Diagnostic checking After the TS model has been estimated the absence of residual autocorrelation has to be checked. the identification tool s. In the following section. . The i maximum of the likelihood function is found by minimising I: u~ with respect to these parameters. Forecasting the last observations of the known sample belongs more or less to the diagnostic checking stage also. Univariate TimeSeries Models The disturbance tenn U t is a nonlinear function of the unknown parameters <Pi and B . . the following conditions . the variances and the covariances of the process are constant through time.
for all s.k.The autocorrelation function (ACF) for the moments of the distribution of Y. is written as: Pk = "Ik. for all s and t.6) The lheoretical ACF can be computed for all TS models of economic variables. "10 ( 14. . This ACF will be estimated by computing the 'sample autocorrelation coefficients' from the data on Yi. it is common practice in the econometrics literature to define the autocovariance between Y. The autocovariances "Yk for k = 0. 2.) = vaT (Y._. are required: 329 E (Ytl = E (Y.) . and Y'. I.). for all s and t. _" Yi.k as "Ik': If the moments are not time dependent because of the stationarity requirement. for all t: "10 = vaT (Y. For notational convenience. But the use of only lhe ACF is not suffi cient to . _form the discrete autocovariance junction of Yt · The 'autocorrelation coefficients' of the process Y. with k = 0. 2. t and k. then the oDly relevant parameter is the lag or the lead k: "Ik = "Ik' · The autocovariance with k lags is identical to the autocovariance with k leads: "Ik = "I.). cov (y" Y'k) = cov (Y" Yi+k) = cov (Y..k· Notice that "10 denotes the variance of Y.._.) .k· Then the theoretical discreet autocorrelation function Pk of Yt..) = cov (Y'+8> Y. are defined as: Because of the stationarity property of the process it follows that: Pk' = Pk = P.+k+. . vaT (Y. Comparing the sample ACF with lheoretical ACF can yield indications aboul the kind of process that may have generated the data. 1.
The residuals of these regressions (in notation: and ef) are equal to the variables yt and X t ... the partial correlation coefficient between Y. written as Pk between Y and Y. for a stationary process is defined as: Pk = Pkt! with k = 1.1 periods Yi.. 14.. X.l' . but it is useful to discuss it briefly here as an introduction to the partial autocorrelation coefficient. . . To understand the concept of a partial autocorrelation coefficient.po . It is the simple correlation between Y. . which is shown with the following formulae: er The partial autocorrelation coefficient Pkt between yt and Yt . What is. How can the partial correlation coefficient between Y. . and X.. The concept of a 'partial correlation coefficient' may be wellknown from statistics. v Pkt = PeY.k. The theoretical PACF pi. . we consider a Unear relationship among three variables: Y. corrected for the influence of the intervening k . and X .. as the correlation coefficient between the two residual s ei and ef. and X. For example.k .. The residuals e[ are now defined as: et = y v~t  ao  ~ ~ v QIIt. (notation: Pv.IIt . and Z..5 The partial autocorrelation function (PACF) The PACF is necessary as a second tool in the process to identify a TS model together with the ACF. Next. compute the partial correlation coefficient Pv.. an example of a partial correlation coefficient of two variables is given. .x ofY..k .k+l· ~ v The partial autocorrelation coefficient Pkt can be computed as the correlation coefficient of ei and Y.. and Y. .ll' .. Univariate TimeSeries Models identify all characteristics of the process. 1 Yt .k . for example..  Qk. Yt . is computed in a similar way. .l  . J t .k : ..) has been eliminated from these two variables.k+ 1. ..330 Chapter 14..k+ l . but 'corrected ' for the influence of Z.x) be computed? This can be done by first regressing both variables on Z. . within thi s group of variables? A partial correlation coefficient measures the correlation between two variables after the influence of other involved variables (in thi s example only Z. ... That means that we can only correct yt for the influence of its past: Yt..we also need the PACF of the variable. The PACF is introduced in the next section. Both functions give information about the type and order of the process.
.. > 2. we look at an AR(P) process. The partial autocorrelation function (PACF) 331 As an example. p} ..8) Define the partial autocorrelation coefficient of order k for any AR(P) process as 'Pkk. This is an important property for determining the order of the process.1 + 'P2'"Yi . Solve 'Pkk (with k = 1.2 + . . .. . As the maximum number of lags is equal to p... (j = 1. + I. .7) Obtain the same relationship for the autocorrelations by dividing both sides of the equation by the variance '"Yo: Pi = 'PtPi1 + 'P2Pi. ... and 'Pkk = 0 for k > p. In the econometrics literature. (14. it is clear that: Pk = 0 if k > p. .p + Ut· Multiply both sides of the equation by Yt i: . . but only the solution for 'P22 is relevant: P1 = 'P2 1Po P2 = 'P21P1 'PH = 0 for k + 'PnPI + 'P22PO. + 'PpYi . two parameters ('Pll and 'P22)are unequal to zero and are solved for k = 1 and k = 2.1 + 'P2 Yt. 'P21 and 'P22 as i = 1. For an AR(2) process. > 3...k. + 'Pp'"Yip' (14.{JklPi . with i = 1. .. + 'PpPi. Proceed as follows with the AR(P) model: Yt = 'PI Y t .9) Then the PACF is defined by the series {'Pkk : k = 1.2 + . .l + lPk2Pi .. which implies that one equation is solved for one parameter 'Pll (with Po = 1): P1 ='PllPO = 'Pkk = 0 for k 'P 11 =P1. k. The derivation of these eq uations is briefly summarised. p) from the set of k linear equations in 'Pk. . For k = 1 the result for 'P11 is identical to that above. When we have an AR( I) process. .k.2 + . k) that are given as: Pi = I. For k = 2 there are two linear eq uations that can be solved for two parameters. ... . then k = 1. We will look at some examples. the sample PACF is computed by solving a number of equations that are called the YuleWalker equations.2 + . Take the expectation at both sides of the equation giving a similar equation..v ( 14...{Jk kPi . but now for the autocovariances: '"Yi = 'P I'"Yi.
'P22 from a system of two equations.2 Pk .I PI IP>! 1 PI PI 1 Pk . as will be shown in the nex t section. and finally solving 'Ppp from a system of p equations.I Pk Pk ..2 'PH 'Pk2 .2 Pk. whereas 'PH = 0. But we will only calculate 'Pkk .I PI 1 'Pkk These equations are the earlier mentioned YuleWalker equations in the k unknown parameters 'Pkj. the PACF for an AR(P) process consists of p coefficients unequal to zero: {'PH: k = 1.I 1 Then the PACF.I Pk 2 PI PI 1 'Pkk = Pk . .I Pk .3 IPk l· Pk. 'P33 from a system of three equations.p. earlier defined by the sequence 'Pkk. The series 'fJkk is derived from the solutions from the system (14. for k = 1. these coefficients will be estimated by the sample autocorrelation coefficients Pk... Both functions are available in EViews and computed by doubleclicking on the variab le and selecting 'View' and 'Correlograrn' .3 PI P2 Pk. . p} .10) The theoretical ACF Pk and theoretical PACF 'Pkk can analytically be detenni ned for every ARMA model. written as: PI 1 PI PI P2 1 Pk . 'Pkk = 0 for k . • The p coefficients are computed by solving 'Pll from one equation. This can compactly be written in matrix notation . . p. is written as: 'PH = IP'I IP:I' for k = 1.. In practice. Both functions are expressed in the autocorrelation coefficients Pk. for k > p. Un ivariate TimeSeries Models So. . when modelling a variable.332 Chapter 14. in general.. . This is si mply done by using the Cramer's rule: 1 PI Pk.9) of k linear equations in the k unknown parameters 'Pkj. (14.. PI Pk Pk . > p. The order of the process is unknown and will be estimated by detennining the lag where the PACF becomes not significantly different from zero.
6 Examples of theoretical ACFs and PACFs for some TS models In this section. together with their plots.l· 2 . a~).l· Compute the autocorrelation coefficients: Po = 1 p. Forthe AR(I) model the relationship for the "Ii is: Ii = CPl '"Yil  Recursive substitution of this relation gives: The ACF is computed as: Pi =  "Ii "10  .1 The AR(1) model The AR( I) model is: Use equation ( 14. ~ NID (O . Also ( 14. Exllmple 14. which is a geometrically declining function. for example.plPi.P.7) as a basis for the computation of the ACF. In all the examples we assume that: u. Various examples of theoretical ACFs and PACFs can he found in the econometrics literature (see.8) can he used to get a similar relationship for Pi: Pi = 'P IPi . a few examples are given of the computation of theoretical ACFs and PACFs of some simple TS models. Anderson (1976) for many examples and a compact description of theoretical correlograms and sample correlograms of matching simulated processes).Examples of theoretical ACFs and PACFs for some TS models 333 14. = 'PI Pi = <. = 'P lPo = 'PI P2 = 'P .<P I' .
.p? 1 .70 0.10): 'I'll = Although it is clear that 'l' kk = 0 for k computing '1'22 from (14. _I + u.80 0 .8Y'_1 + u" is given in Figures 14.'1']  1 . lb. An example' of the ACF and PACF of the AR( I) model.BY.00 1 2 3 4 5 6 7 8 9 10 Figure 14.60 0 . ('I'l.20 0.'I'? = 0. For theAR(l ) model: p = l ==>k = i = l. • The plots of lhe theoretical ACFs and PACFs in this section have been made by using Microsoft Excel.30 0. which already means that 'I'll = Pl.p 1) . 10). This result is also obtained by solving equation (1 4. la and 14. This implies that no intervening lags are present.50 0.1a: The ACF of the AR( I) process y. Univariate TimeSeries Models ACF 1. yt = O . = O.10 0.10): IpI! II I = PI > 2.90 0. The PACF is computed by using equation (1 4.334 Chapter 14.00 0.40 0. this can also be shown arithmetically by 1 PI 1 PI PI '1'22 = P2 PI 1 P2  .p? 2 2 '1'] .
8Y. The ACF is computed by using equation (14.30 0 .50 0.1 + 'P2Pi.1b: The PACF of the AR(I ) process 1'.70 0.40 0.00 1 2 3 4 • 5 6 7 • 9 10 Figure 14. Example 14.I = 'PI + 'P2 PI .1 'PI'P2 .2 The AR(2) model The AR(2) model is: with the stationarity restrictions on the parameters as given in the inequalities (14. Although the form of the two functions and the number of coefficients are identical for every AR( I) model.20 0.00 0.80 0. the figures are different for every value of 'PI .5).2' i = 0: Po = 1 i = 1: PI = 'PIPO + 'P2P.80 0. • i = 2: P2 = 'P IPI + 'P2Po 'P~ 1 ._1 + tt.8): Pi = 'PIPi.Examples of theoretical ACFs and PACFs for some TS models 335 PACF 1.'P2 + 'P2 · .10 0. = O.90 0.
2b. _2 + Ut has heen plotted in the Figures 14.Pl 1 . the Pi can he computed recursivel y by the relationship: Pi = 'PIPi. for k > 3. k = 1: = 2. First.'P2 Secondly. k = 2: 'P22 = 1 PI 1 PI P2 1 PI P2 PI 1  p? p? 1 'PIP I + 'P2  p? pi Pl Pl (1  'P2) 1 + '1'2 . An example of the ACF and PACF of the AR(2) model: Yt = 0. 2.40Ytt + O.1 + 'P2Pi2 · This function will decay to zero as i + 00. we bave k = 1. As p i = 1. The ACF of the AR(2) model is an infinite decreasing function and the PACF has only two nonzero values.p? 'P2 = 'P2· (1 1 Pi) pf The higherorder partial autocorrelation coefficients are zero: 'Pkk = 0.336 Chapter 14.2a and 14.Pl p? .25Y. .2 and 'P11 = Ipl l III 'P I 1 . Univariate TimeSeries Models iQrtlher.10). Next the PACF is computed according (14.'P2 Pl + 'P2 .
!l(U'i + O.2Ot.25Y'_2 + u... it can be shown that the AR(P) model has an ACF that decays to zero.20 3 4 5 6 7 8 9 10 0.25Y'_2 + u.20 0.. 0..00 2 0.O..Examples of theoretical ACFs and PACFs for some TS models ACF 0.' Figure 14... PACF 0..) = EI(u.3 The MA(1) model The MA( I) model is: The computation of the ACF is done as follows....40 0.60 Figure 14.2a: TheACF oftheAR(2) process Y.2b: The PACFoftheAR(2) process Y.60 0. + OtU. = . Conclusion On the basis of the preceding examples we conclude the following abou t the ACF and PACF of the AR(P) model. ..0.60 L _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _. Example 14.00 10 0. and an PACF that cuts off after p lags.40Yll + O.40Y.4 0 t ..40 0. = ...40 337 0..... Start with the autocovariance function : "Yi = cov( Y" Y.0....Uti!lI· .60 .. Generalising the result of the AR( J) and AR(2) processes....0.20 0._1 + O.
The PACF is computed by: ! This is an infinite function that decays to zero. The PACF is less easy to derive. = u.3a). is given in Figures 14.60 0.g. + 0.3b. or more generally: i >2 ~ 'Yi = O.40 0.70U' _I .00 1 2 3 4 5 6 7 B 9 10 Figure 14.3a and 14. Univariate TimeSeries Models Then it is easy to see that the covariances are not equal to zero for contemporaneous products: i = 0 ~ 'Yo = (1+ Or)<7~ i = 1 ~ 'YI = 0J<7~ i = 2 ~ 'Y2 = 0.338 Cha pter 14. it is important 10 notice that Ipti < because lOl l < 1(see also Figure 14.20 0. So the ACF of an MA( I) process cuts off after one lag. This implies for the autocorrelation coefficients Pi = 'Yiho : Po = 1 PI=l+O~ 01 Pi = 0 for i > 2.30 0. Anderson ( 1976)) for its derivation. + 0.50 0. Yi. An example of the ACF and PACF of the MA(l ) model. = u.30: The ACF of the MA( t) process Y. 10 ! 0. Further. ACF 0. See the econometrics literature (e. This is always valid for an MA(l) model.70ut! .
70"<1 Example 14.60 1 339 ~r: "6" 7 tit 8 9 10 Figure 14. + OIU' _1 + 02U'_2)(Utt + 0IU'_i_1 + 02U' _i_2) ). Notice that it is possible that p. Next compute the ACF again from Pi = "1'/"10: Po = 1 01 (1 + O ) 2 PI = 1 +0r+0~ P2 = HOr +O~ P.3b: The PACFofan MA( I ) process Yt = + 0.Examples of theoretical ACFs and PACFs for some TS models PACF 0. The autocovariance function 'Yi is: i = i a =? "10 = i = 1 =? (1 + Or + 0~ ) C7~ "11 = (0 .20 0.60 0. = a for i > 3.20 0.40 0. + 02 0dC7~ = 2 =? "12 = 02C7~ i > 3 =? "Ii = O .00 0. Yi i) = E[(u. The PACF can be shown to have no cutoff.40 {l. the function decays to zero. In the same way as for the MA(l ) model. > ~ for an MA(2) model. the expectations of contemporaneous products are Dot equal to zero.4 The MA(2) model The MA(2) mode l is: For the determination of the ACF the autocovariances are computed again: "Ii = cav(Yi . The ACF of an MA(2) process cuts off after two lags. O 2 .
The ARMA(p. . They do not have clear cutoff points.5. Therefore the identification and estimation of an ARJMA model for the log of the cocoa price will be di scussed. the relationship between P{OCand p tCo! was used to illustrate the computation of predictions for the cocoa price with a causal model.Jenkins procedures are di scussed on the basis of an example of a TS model for the cocoa price. You will observe waves in the pattern of Pk because of this. then it can be proved that: n. making it difficult to determine the order p and q of the ARIMA model.l k= ] L 'h = 0. The ACF and PACF are a mixture of damped exponentials or sine waves. An empirical example is given in the next section. Good advice is to use only a small number of the autocorrelations for the following reason. In practice. 14. In Section 7. Univariate TimeSeries Models Conclusion On the basis of the preceding examples the following conclusions concerning the ACF and PACF of an MA(q) model are drawn. Predictions for In (P{OC) will be computed with a TS model and compared with the previous results. and bas a PACF that decays to zero. Generalising the result of the MA(J) and MA(2) process it can be shown that the MA(q) model has an ACF that has a cut off after q lags.3.7 The BoxJenkins approach in practice In this section the Box. we will compute a series of sample autocorrelations {Pk} and try to recognise a matching known theoretical pattern to identify a TS model. adequate representations for many stationary variables can be achieved by models that do not have too higb values for p and q.q) model: cp (£) Yt = e(£)!Lt .340 Chapter 14. Remark 14.1 A property of the sample autocorrelatlons In practice.q) model The ARMA(p. So the behaviour of the ACF and the PACF of MA models is 'contrary' to that of AR models. If you have a sample of n observations. has a combination of AR and MA properties resulting in the following property. This property works as a 'restriction' when many Pk are computed and disturbs the theoretical pattern.
for i > k. doubleclick on the variable name. These are the standard errors that are plotted in the EViews window. more examples of identifying TS models can also be found in the literature (see. In Section 12. for all k. i= l ' 1 Pi = 0 .992. the sample ACF and PACF are compared with pictures of theoretical ACFs and PACFs as derived in the previous section. In fact it shows the unit root in the process: P = 0. we know already that In (P.) = under the null hypothesis that n 1+2I:Pl' .OC) To identify a univariate timeseries model for In (P.c°C) ~ I (1). . But in case we do not have that knowledge a priori.The BoxJenkins approach in practice 341 The identification procedure has been discussed rather extensively in the previous section but here the computational aspects of estimation and forecasting procedures will only briefly be discussed. Granger and Newbold (1977) for some clear empirical examples). therefore we will try to identify a TS model for \lIn (P. In Figure 14.5. click on ' View' and select 'Correlogram' .c°C Look ). we concluded that In (P. (See the more advanced textbooks for more detailed information about these aspects. Next the sample ACF and PACF of the variable are computed. It I is clear that the ACF cannot decline in a fast way. then the standard errors are computed under Ho : Pi = a for i > 1: se(p. then the 000stationarity is immediately observed in the correlograrn that is given in Figure 14.5.). For example.) = If. for the sample data of the stationary transformed variable. The standard errors se (P.c°C) is not stationary. Next. The figure shows the typical form of a correlograrn of a nonstationary variable. for values of P' in the ACF and PACF that are significantly different from zero by using the interval 0±2 x se(p. as had been tested previously with an ADF test. if you want to test that the process is a white noise process (such as the residuals of an estimated model). From section 12. A window opens where the order of integration of the variable has to be indicated.) of the estimated autocorrelation coefficients are approximately computed as: 1 se (p. for example.C the ACF and PACF are estimated OC). Identification of a TS model for In (P.) Of course. To identify an ARIMA model by using EViews.OOC) .4 the window is shown that appears after the option correlogram has been selected for In (P. so we can directly select '1st difference' .5.
72 .000 0.867 0.OSI 1400.027 2888.056 5026.2 . 0.7 0.936 0. Univariate TimeSeries Models Figure 14.3 0.004 5423.5: The correlogram of In (Pt ''' ) .342 Chapter 14.0.025 3334.946 0.0 COI"it1 "'I 0 .0.000 0.914 0.6 .992 fJJl.958 0.23J 1004.000 0.000 n ace n fin ~ Figure 14.{)1.880 0.000 0.u3 TIme: 11 :23 Sample: 1960:01 2002:09 Included observations: 513 Autocorrelation Partial Correlation 1 2 3 4 5 6 7 8 9 10 11 12 1? PAC Q. .000 0.5 O.0.D16 1966.000 0.893 0.5 0..969 0.1 0.0. .033 3772.900 0.000 0.053 4618.0.000 0.4 .Slal Prob .000 0.000 0.2 .4: Defining which transformation has the correlogram is computed 1 0 be applied before Dale: 05.026 4200.0 .o.992 0.024 2432.925 0.000 0.004 0.
Next in Figure 14. .n rY'f> h??1 F> n Figure 14.000 0.000 .082 51 .553 62.000 0.076 0.7. we notice that all the probabilities of the Qtest are zero.2) model.31 4 0.075 Prob 0.IXl) 0.042 0..117 0. The correlogram in Figure 14. the first difference of In (PtC C has been selected and the correlograrn of \lIn (PfOC) is given in O) Figure 14.935 50.000 0.052 0. The ACF and PACF sugges~ for example.().7 is used for the identification of a TS model for \lin (PfOC).D38 0. 0.. .ooo 0.439 61 .014 . indicating that the process clearly differs from a white noise process.000 0. The dotted lines around the correlation coefficients are twice the standard errors (±2 x ) l (n). estimating an MA( I) or maybe an ARMA( I . .032 0. This button can now be used to get the 'Correlogram Specification' window at once.976 51 .7: The corre)ogram of \ltn (PtCOC ) .D38 '(). In the first place.000 0.958 50.000 0. .().().029 .004 0.580 52. .n nH' .105 52. Date: !J5.The BoxJenkins approach in practice 343 Figure 14.547 51 .108 0.000 0.000 0.021 .074 0.058 0.6.314 .000 0.6: Selection of the transformation before the correlogram is com puted '.01 2 .023 0.().000 0.5 that a button 'Ident' has appeared.062 0.008 0. .01O 50.OO7 '().{)tm TIme: 11:44 Sample: 1960:01 2002:09 Included observations: 512 Autocorrelation Partial Correlation AC 1 2 3 4 5 6 7 8 9 10 11 12 1':t PAC QStat 0.003 62.000 0.'. Notice in Figure 14. .382 55.
MA( I) or MA(2). When anARMA(2.2) model for V'yt has to be estimated. Click on 'Quick' and 'Estimate' as usual. The minimum is found by using an iterative procedure (see the EViews User's Guide for the details as to how this is done in EViews). It has a similar interpretation but it gives a larger 'penalty' for the number of parameters. However. therefore a 'penalty' is added for the number of parameters in the model. theACF and PACF gave some indication of the order of an ARIMA specification. n S I C = In(iT~) +K . we have seen that the residual variance is reduced by increasing the number of regressors. For two models based on the same series you would choose the specification with the lowest value of the AlC or the SIC. n The SIC is an alternative to the AIC. That makes it sometimes difficult to make a choice between different model specifications. which can be an advantage. Because series of different length can be used. I ) and an ARIMA(1 .2) are estimated. The example of estimating a TS model for the cocoa price is continued. AR(2). etc. and specify in the 'Equation Specification' window: d(Y) c ar(l) ar(2) mall) ma(2) . In EViews. which is identical to maximising the likelihood function. These criteria cannot be used to compare models that have different levels of differencing. especially when specifications are very close to each other and estimation results hardly differ. I . forexarople. ARIMA models are statistical models. you always find the 'Akaike information criterion' (AlC) and the 'Schwarz information criterion ' (SIC). For that reason. In EViews. the SIC will indicate a simpler model than the AlC. I. AnARIMA(O. The values of the criteria can only be interpreted relatively. AR and MA terms can be specified in the 'Equation Specification' window by specifying AR(1).344 Chapter 14.1 (L ) 'P (L ) yt. A helpful tool in making a choice is the use of an information criterion that is often part of the computer regression output. the criteria are normalised by dividing them by the number of observations that have been used to estimate the model. The AIC and the SIC are defined as: MC= In(iT~) + 2K . Both criteria are based on the residual variance. Univariate TimeSeries Models Estimation of the parameters of the TS model The parameters are estimated by minimising the sum of squared disturbances. The (squared) disturbances are a nonlinear function of the parameters: Ut = 0. Initially. In Figure 14. In(n) . You want to have a model that has the smallest residual variance. they are related to economic models but they are not economic models themselves. the following must be done.8 the 'Equation Specification' window is shown. Next the number of p and q have to be chosen in such a way that absence of residual .
617571 . crr~ 0 . Error I·Statistic Prob.057950 0. I .[.002484 0.statistic) 0.97 .97 0.003372 0. Notice that the roots of the lag polynomials have been computed and have been printed as the inverted roots. Dependent Variable: D(lPCOCOA) Method: least Squares Date: 05AJUJ3 Time: 14:33 Sample(adju'ted): 1960:03 "'200""'2:09 Included observations: 511 after adjusting endpoints Convergence achieved after 21 iterations Baekeas!: 1960:01 1960:02 Variable C M(l) Coefficient Std. dependent var Akalke info criterion Schwarz criterion F·statistic Prob(F.3465!B 0.940414 ·7. ) .67548 8.E.3311 O. and the estimated parameters are significantly different from zero.111 .4033 1.113498 0.003280 0.OODJJ ·.000076 0.36 Figure 14.I .694338 ·2.939449 0.The BoxJenkins approach in practice 1'I!Jot".900726 692.9: An estimated ARIMA( I.9 and 14.10 the two output windows are shown.1(8252 0.661176 21 .0.D.(. In Figures 14. 345 It3 Figure 14. ofregression Sum squared resid Log likelihood DurbinWatson stat Inverted AA Roots Inverted MA Roots 0.CE2662 1.636118 O . Autocorrelation in the residuals will be checked later.OCOJ O. <'1'1.972740 16.8: Specification of an ARIMA(I. I) model has a lower value both for the Ale and the SIC.2) for In (P.066356 ·2. The estimated ARIMA(O. The inverted roots should be within the unit circle. MA(1) MA(2) Rsquared Adjusted Rsquared S.984577 .(1))) Mean dependent var S.Itro[. I.o.2) model for In (17"' ) autocorrelation is also found.966348 . .o.043644 0.
044 ) = (0.041550 8.0.D.618 L 0.(0.OO:X:OO 0.1 .(62548 1..967519 ·. (0.347 L2) u.0. Similar problems exist with respect to the AR polynomial: are the parameters initially written at the left· or right·hand side of the equation? The estimation results of the two models.00328) (1.352651 0.003) 0.346 Chapter 14.002398 Fstatistic Prob(Fstatist ic) stat Inverted MA Roots Figure 14.618 (0 .044) 0.5117 0.2) model: V In (P.00328 + V.00011 + (1 .701880 ·2..E. The interpretation of the printed signs is often not clear in software output.995221 693.110532 0.966 ( 1 . whereas others write them with negative signs.0.347 £2) u" 0.044 ) .0. The ARIMA(1 .112273 0. 069) (0. obtained with EViews.347 L2) u.003738 0.487292 0.0.10: An estimated ARrMA(O I. Error I·Statistic Prob.0000 S. I ) for In (Pt'~ ) . (0 . (0.50096 O .clllmo ·2.003 ) = (0.966 L) V In (P. = (1 .069) L.6813 1.'2OO"Il?2:09 Included observations: 512 after adjusting endpoints Convergence achieved after 6 iterations Backcast: 1960:01 Variable Coefficient Std.058) L) (v In (P.'OC) 0.COC) . dependent var Akaike info criterion Schwarz criterion o. Univariate TimeSeries Models Dependent Variable: D(lPCOCOA) Method: Least Squares Date: 05J1)1. Some authors write the MA parameters with positive signs <as in this book). have the foUowing interpretation.0.058) That is equivalent to: L) V. C MA(1) Rsquared Adjusted Rsquared 0.618 L.058) ( 1 (0.002454 0.35 0.(0. ofregression Sum squared resid Log likelihood Durb i~Watson Mean dependent \laT S.069) (0.685324 64.(l3 Time: 14:36 Sample(adjusted): t960:02 :."~) = 0.966 ( 1 . because the convention of writing the lag polynomial of the MA·part is not standard in econometrics literature.
I): (1 + 1I1L)(1 .618 U'_1 . in expression (6. For example.9! Diagnostic checking (testing the residuals to be white noise) Obviously. The model specification is unique for a panicular ACF and PACF.The BoxJenkins approach in practice 347 This can also be rewritten as: \lIn (P.6).002 (0. With the Ljung.002 +Vt (0 . earlier introduced in Section 6. Try to find the most parsimonious model. = (1 + 1I1L) "t for any III I 1 < 1.11 ) . the residuals have to be tested for the absence of autocorrelation.353 Remark 14. the following two models have the same correlogram because of a common factor in the AR and MA part of the model: AR(I): (1 .353 So this is equivalent to the model: \lin (PtCOC ) = 0.004 ) + (1 + (0 . Diagnostic checking concerns the evaluation of the estimation result. In timeseries analysis.069) 0.044 ) The ARIMA(O. L) Ut.c.042) U' _ l · 0."f) (0 .1.c°C) = 0.347 Ut2· (0 .042) 0.966 \lIn (p. See the COmmon factor in Figure 14. (14. with respect to the number of parameters.0.1) model: \l (In P.3 as the Qstatistic.'P1 L ) Y. except in the case of common factors. In time· series terminology this is called 'diagnostic checking' .004) Vt or = (1 + 0.002 (0. that is free of residual autocorrelation and bas significant parameter estimates. 11 ).042) L) Ut. \lIn (P.Box test.2 Common factors You must be aware of 'common factors' in the model specification.353 (0. it is usual to compute the LjungBox (Q) test.058) + Ut  (0.000111 + 0.004 ) + Ut + (0. = Ut ARMA(2.'P1 L ) Y.c°C ) = 0. we test whether the residuals behave like a white noise process.c°C) = 0. The formula is given once more in expression (14. So be cautious or economical with the specification of a TS model.
Stat Prob Partial Correlation I I I I I I AC I I I .746 0.044 5 .032 0.348 Chapter 14.072 0.374 0.079 11 0.547 Figure 14.11: for In Correlograms of the residuals of the estimated ARJMA model s (PfO C ) .OOJ 0. where 'Correlogram .1.045 .2935 1. In the ARIMA( 1.9658 2.254 0.011 .344 0.q) model. Date: ()5. Univariate TimeSeries Models where is now the squared residual autocorrelation coefficient of lag length k .6n9 5.705 0.2) comes first.003 . and for the ARIMA(O. I I Prob PAC Q.452 0. I I I I I I I I 1 0.0085 0.024 4 0.7050 4.062 0. .6610 8.ng the information criteria and evaluating its forecasting power.034 9 0.259 0. As an example. because the probabilities are given starting at lag 4.439 0.n5 0. although the pvalues of the MAO ) model are higher.063 6 0.002 12 O.Pi.038 7 .0.004 ·0.statistic probabilities adjusted for 3 ARMA term(s) Autocorrelation I I I . the Qtest has a X2 (K .0. [t can be shown that very high R 2 s can belong to very poor TS models. p + q = 3.003 3 .Qstatistics' is selected with all the values of the Qstatistic and the corresponding probabilities. 1.q) distribution: Pi You should choose an appropriate value for K.0.322 0.605 0.439 0. As usual. which is not too high or too low. 1) are visible in the background . I I .039 .n9 0.037 6 0.366 0. it is informative to look at the whole range of probabilities of the Qtest.032 10 0.03 Time: 12:18 Sample: 1960:03 2001 :05 Included observations. because it always starts at lag 1! The determination coefficient R2 is not very useful in TS models.0. the (1. for a stationary AR( I) process with zero mean it can be shown that R2 "'" The best thing to do is to judge a model by looking at the residual correiogram.0. it is a ntistake to doubleclick on RESID and to use that correlogram.2477 5.) For example.449 0.024 0. the pvalues of the (0.1.661 0.036 0.1)2.2) model.9833 8.004 2 ·0. I I I I I I I I . the correlograms of the residuals of both the ARIMA models for the cocoa price are given in Figure 14. by usi.0. (See Section 13.7753 8.0124 0.038 0.I) model they are given with lag 2. The null hypothesis of white noise residuals is not rejected for both models.538 0. 495 Q. I . Under the null hypothesis of white Doise residuals of an estimated ARMA(p. 11 .9833 0.0.2766 1.745 0 .007 0. In EViews the ' Residual Tests' have to be used again.P .7 in Maddala (2001) for a discussion of this topic.
In EViews. 12). The forecast sample and the choice between dynamic or stalic simulation have to be indicated.l ).3 has been repeated. a forecast window appears after clicking on the button 'Forecast' in the 'Equation ' window (see Figure 14..1.coc and to save the standard errors of the forecasts. just as earlier discussed.12: I The 'Forecast' window . or its first difference 'Vy. See the chosen options in Figure 14. The same prediction period has been chosen as in Section 7. In Table 14. I "11K Figure 14.3 concerning the specification of a differenced variable. you want to predict the level of the variable. the forecast evaluations have been summarised .3. With the specification ( 14. in the last column the result from Section 7. 1) model can be estimated as: dey ) c ar(l ) ma(1 ) and can also be estimated as: ( 14. with the variable ( 14. 1. 12) the choice has to be made whether forecasts will be computed for the level of y. Remember the remark that was made in Section 7.13) only predictions for 'VYt can be computed. With specification ( 14.13) dy generated in advance as: genr dy = y . 12 with the option to predict the level of lu P. In most cases. For example.12) dy c ar(l ) mal l ). an ARIMA ( 1.The BoxJenkins approach in practice 349 Forecasting with a TS model Forecasting future values of an economic variable with an ARIMA model follows a similar route to that of the causal model discussed in Chapter 7 (see also Clemens and Hendry ( 1998».y(. The models have been estimated for the reduced sample period and the predictions are computed for the chosen prediction period.
RIMA( I.~. i : . Although you cannot see it in Figure 14.. ! l.. a property of predictions from an ARlMA model is that they vary in the short run after which they become constant.058) (pt'.272837 6..~f..'"':.2· I 4.I .044) 0. The ARIMA(l . whereas the ARIMA model yields increasing trends.. I i .. i : Y : .~_i.350 Chapter 14. Univariate TimeSeries Models Forecast evaluations of various models for In ptcoe Forecast sample: 2001:06 2002:09 Inequality coefficient A.l  (0..2 +~_+. I .063902 Table 14. .284123 6...040965 AcIl~(O . 4 f i ! .665497 0. in Figure 14.042970 0. Perc. I . t ~ : I I .504822 0. I . Although we ought to realise that the causal model is not really a model for the price formation of cocoa on the cocoa market.1: Forecast evaluations o f the two ARIMA model s and the causal model The inequality coefficients of the ARIMA(l.069) 0...2 4.416260 9.604259 0...6 3... Predict~d coco~ price ! 4. but they are clearly smaller than the coefficients belonging to the predictions obtained from the causal relationship between the commodity prices.JIJlI1A. 1.. I . 13. ! I " ..+. This is the reason for the smaller values of the inequality coefficients as PtCOC is also increasing.(I.3...2) 0.347 U t .0 3. In Figure 7. 1Cocoa price i l !_ . That can be demonstrated as follows.330472 0.. Error Theil inequality Coefficient 0. ! ... we observed a decreasing trend in the forecasts.I ) model.13: Forecasts from the A..."n + Ut  (0 .~f__..2) model that has been estimated for the sample period is: \lin (PtCOC) = 0. I .6 .....618 Ut. ! 3.2) model b ...I) Causal model Root Mean Squared Error Mean Absolute Error Mean Abs.f+lr I ..347836 0. __ _ _____ _+ _____ _ ..  .  ~  01 :01 01 :04 01 :07 01 :10 02:01 02:04 02:07 Figure 14.. i .00011 + 0.1 i iii : i i : i L _ _ "i '.I.3..13 the results for the ARlMA(l ..2) model are plotted in the same way as was done in Section 7.8 3.966 \lin (0.4 t2xse ! / 1 4. ! ! I .2) model are marginally smaller than those of the ARlMA(O..+_~_+.346024 0...
618 · 0 . However. the conditional expectation OfUn+ l is zero. for period n 351 + 1.347 .966VI11 (P~+l)  0 .966Vln.00011 = 0.l .).0 . Un· The forecast is computed by using the last residual en: (P:.000 11 + 0.00328 as shown below in 'general' terms: I et = 0. .~'2) the model is: 0.~'2) = 0.+l) = 0.I : .0.o.347e n _1.o.(P:. In practice.966Vln (P~+l) + 0 .~'l ) . the model is: The conditional expectation of Vln (P:.0. That implies for the computation of the forecast V In (P.618un . is Vln (P:.i+.966VIn   The last expression shows that the computation of forecasts has become recursive. Then the ARIMA forecasts from this model will converge to a constant value. converge to the constant term of the model . is Un and U n . for t > n + 1.966Vln (P~') . Hence V In (P~+cl) is forecasted as follows: Vln (P'.c ) : 2 Vin (P~+2) = 0.347 · en· Vln It will be clear that Vln (P~ ) is forecasted by: +'3 .) .to given the past.966Vln ( P~O') + 0  0.347 · Un + 0.The BoxJenkins approach in practice Then.~'3 ) = 0.0.00011 + 0.00011 + 0.347un_l.+l) = 0.~c2)' Vln (P:.00011 + 0 .0. 0.(P:.00011 + 0. In this example.618e" Vln To forecast Vln (P:. and the condi tional expectation of Un and Un . given its past.(P:. the forecast is computed according to the estimated model and using the last residuals en and en . the forecasts of 'V in (P~+i)' with i > 3 .
where the forecasts show the growth of 0. showing the graphs of the ACF and PACE Include a figure with the newly calculated predictions.13 above. Compare these forecasts with the real values of the observations and. predictions for an endogenous variable have been computed by using a causal econometric model. • • I \ . Univariate TimeSeries Models so \J In (P~~ ) + 0. the predictions from Case 3 and the actual values.328% per month in the log of P. Compute forecasts for the omitted periods at lhe end of the sample period. the forecasts have become nearly constant. Do this for a sample without the last couple of periods. the variable is in first differences. But each timeseries from a data set can be selected for this exercise. This result is illustrated in Figure 14. which makes it possible to compare the quality of both the predictions. Write an illustrated paper about your analysis. After a number of periods.352 Chapter 14. In Case 3. It is interesting to estimate an ARIMA model for the same endogenous variable.Jenkins procedures will be applied to a series from one of the timeseries data sets.8 Case 11: BoxJenkins procedures In this case the Box. This implies that the trend in In (PtCOC) has become constant. 14. pay attention to stationarity (also seasonality). These (longrun) forecasts are not very interesting. However."oc . Check the residual behaviour for the absence of autocorrelation (diagnostic checking). which underpins the statement that an ARMA model can be useful for the computation of shortterm predictions. if relevant.00328. which means that the forecast of the change of In (PtCOC ) has become constant. compare the forecasts with the earlier obtained predictions made with the causal model. • Identify and estimate a TS model for one of the variables from a data set. Longrun I model simulations are better carried out using a causal econometric model.
pp. ( 1992) Applied Econometric Techniques. Hemel Hempstead: Philip Allan . ( 1976) Time Series Analysis and Forecasting.1287. Joumal o/the American Statistical Association 65.P. 5th ed..S. (198 1) ' Qualitative Response Model: A survey'. Breusch. Baltagi. Chichester: John Wiley & Sons. a . Anderson. ( J979) 'A simple test for heteroscedasticity and random coefficient variation'. Breusch. J. (1976) Time Series Analysis. Hants: Edward Elgar. London: Butterworths. T. Cramer. H. 2nd ed. pp. (1996) The Analysis o/Time Series: An introduction . (1978) ' Economenic modelling oflhe aggregate timeseries relationship between consumers' expenditure and income in the United Kingdom'. Clemens. $. (200 I) Econometric Analysis of Pallel Data . London : Chapman and Hall.E.1 5091526. Box. ( J986) 'Gereralized Autoregressive Conditional Heteroskedasticity'. 1488.334355. San Francisco: Holden Day. Cuthbertson. and Hendry. ( 1989) Econometric Applications 0/ Maximum Likelihood Methods.p.. ppA07423. ( 1970) ' Distribution of residual autocorrelations in autoregressiveintegrated moving average time series models '. Box. Australian Economic Papers 17. Econometrica 30... Cambridge: Cambridge University Press. and Jenkins. and Deadman. (1962) 'The distributed lag between capital appropriations and expenditures' . Forecasting and Comro/ revised ed.References • Almon. A. Econometrica 47. B. ( 1992) New directions in econometric practice. M. (1978) 'Testing for autocorrelation in dynamic linear models'.0 . pp. Hendry. 0 . W.307327. G.P. Cambridge: Cambridge University Press. Bo ll erslev. M. Davidson. Srba. and Pagan. D.1294. G. The Box·Jenkins approach. Chatfield.S.G.E (1998) Forecasting economic time series. and Taylor.. K. pp. The Economic Journal 88. J. Journal of Econometrics 31.R. S. T.E. Amemiya. Journal of Economic UteralUre . T.M. t . Charemza.. S.E. D. p. C. D. Hall.F.P.661692.F. . and Yeo.A. D.H. and Pierce. pp. F. T.S.
pp. Economics Letters 6.E. pp. and Box. and Granger. (1982) 'Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation' . J.J .G.S. New York: McGrawHilI.. Engle.M.S. and Applications. Granger.A. (2002) Introduction to Econometrics. D.C. J.M . P. and Bera.42743 1. IntriHigalor.409428. R. (2(x)() Econometric Analysis. Journal a/the American Statistical Association 74. and Fuller. Cambridge: Cambridge University Press. N. Switzerland: Ghisler & Co. Harvey. Econometrica 55. Dickey.B.A. A. Econometrica 50. Koyck.C. L.1302. M. W. and Quandt. Cambridge: Cambridge University Press. Econometrica 49. ( 19 80) 'Efficient tests for nonnality.303. (198 1) 'Likelihood ratio statistics for autoregressive time series with a unit root'. tn Moo (1998).H. C. 2nd ed. Journal of Econometrics 16. ( 1965) 'Some tests for homoscedasticity'. G.25 1276. (1979) 'Distribution of the Estimators for Autoregressive Time Series with a unit root' . ( 1997) Econometric Methods 4th ed. and Newbold. R. R.S. Cambridge: Cambridge University Press.23 1254.539..F. (1976) Introduction to Statistical Time Series..297. Upper Saddle River: Prentice Hall. pp. Fuller. 1293. and Hsiao. P. F.D.W. and Watson. P. pp. W.M.1. and NeWbold. Johnston. New York: McGrawHili. homo skeda~ ti c ity and serial independence o f regression residuaJs'. C.N. pp.179. C.) Econometric Methods. (2002) A Concise Introduction to Econometrics. M. with AppHcations for the Demand for Money' . (1987) 'Cointegration and error correction: representation. Dougherty. MaddaJa.547. Godfrey. Haldrup. G. Gujarati. Granger. (1983) LimitedDependent and Qualitative Variables in Econometrics. Franses. (1948) Statistical methods for research workers.J.354 References Dickey.J. pp. Bodkin.W. Techniques. pp. Phillips. Amsterdam : NorthHolland. ( 1974) ' Spurious regressions in econometrics' . Maddala. Unit roots.2 1O. K. S. 10th ed. Ghisler. A.. l oumal of the American Statistical Association 60. (1990) The Econometric Analysis of Time Series. Hamilton. and Fuller. pp. G.E. Dhrymes. L. 1. Princeton : Princeton University Press. (1981) 'Some properties of time series data and their use in econometric model specification'. C . (2003) Basic Econometrics. Oxford: Oxford University Press. pp. and DiNardo. Journal of Econometrics 2. III120. L. Oxley.M. pp. W. pp.K. G. San Francisco: HoldenDay. (1971) Distributed Jags: Problems of Estimation and Formulation .G .F. ( 1954) Distributed Lags and Investment Analysis.0.9871008. Econometrica 55. and McAleer. 169.... pp.). C. coilltegration and structural change..259.A. (1950) 'Testing for serial correlation in least squares regression '. C . Goldfeld. Biometrika 65.255. . estimation and testing ' . D.H.J. (1988) 'Statistical Analysi s of Co integration Vectors' .A.234. Ljung. Chichester: John Wiley & Sons. Edinburgh: Oliver and Boyd. R. Fisher. (1999) 'An Econometric Analysis of 1(2) Variables' in Practical Issues in Coinregratioll Analysis. D. G.S. (2001) Introduction to Econometrics. (1994) Time Series Analysis. D. New York: John Wiley & Sons. p. 4th ed. S.P. Johansen. ( 1987) 'Timeseries regression with a unit root '. 2nd ed.. pp. London: Academic Press. Maddala. W. ( 1996) Econometric Models. Hendry.30 I.W. and Kim. (1972 2nd ed. Granger. (1990) 'Maximum Likelihood Estimation and Inferen ce on Cointegration. C. Upper Saddle River: Prentice Hall .277.J. Engle. (1978) 'On a measure o~ Jack of fit in lime series models' . 1984 3rd ed. P. J. Johnston. New York: McGrawHill. (Eds. C. l 063. and Juselius. ( 1995) Dynamic econometrics. (2003) Toral Commander.130. Biometrika 37. P. ( 1977) Forecasting EC DrlOmic Time Series. S. Oxford : Oxford University Press.A. Econometrica 46. Jarque. 3rd ed. Johansen. pp.A. Journal of EcorlOmic Dynamics and Control 12. G. Greene. Hemel Hempstead: Philip Allan . Durbin.121 . Oxford: Blackwell Publishers. Oxford Bulletin of Economics and Statistics 52. ( 1978) 'Testing against general autoregressive and moving average error models when the regressors include lagged dependent variables' . R. W.
2.C. P. White.. (1997) Modem Econometrics.F. M.106. Groot. 2nd ed.e.L.. ( 198 1) Introductory Econometrics. P. Biometrica 75. pp.81 . Chichester: John Wiley & Sons. (2000) A Guide to Modem Econometrics.S. (200 1) Using Econometrics: a Practical Guide. and Walli s. Zellner. ( 1975) 'Time Series and structuraJ analysis of monetary models of the U. Ridder.F. Oxford: Blackwell Publishers. Journal of Applied Econometrics 7. USA: SouthWestern College Publishing. economy '. Hemel Hempstead : Philip Allan . and Watson. E. K. E. and Palm . Thomas. pp. 2nd ed. and Palm. ( 1988) A quarterly econometric model of the world coffee economy. H. 1. Stewart. Vogelvang.H. A . Z.M .. TI 941 13. Wooldridge. ( 1980) 'A heteroskedasticityconsi stent covariance matrix estimator and a direct test for heteroskedasticily '. 191 . ( 1998) Econometrics. .201 . ( 1994) 'Analysis of Longrun and Shortrun Price Behaviour o f Related Agricultural Commodities for Modelling and Forecasting'. and Perron.S. Van Montfort. Stock. ( 1988) 'Testing for a unit rool in timeseries regression ' . Studenmund. Boston: Addison Wesley Longman. Heme! Hempstead: Prentice Hall Europe. ( 1991 ) Econometrics. 4th ed. Series C. (2002). ( 1999) illtroductory Ecollometrics: A Modem Approach.C. an introduction. Journal of Econometrics 2.B.. pp. Econometrica 48. E. . Boston: Pearso n Education. F. 1754. New York: Springer Verl ag. L. ( 1974) 'Time Series analysis and simultaneous equation models' . Irvine: Quantitative Micro Software. 'The innovation decis ion and fixed costs' in The emergence of the knowledge economy.L. and Nijkamp. ( 1992) ' Hypotheses Testing Concerning Relationships between Spot Prices of Vario us Types of Coffee'. Tinbergen Institute Discussion Papers.335346. Verbeek. Pt. A. (2003) Introduction to Econometrics. Quantitative Micro Software (2000) EViews 4. H. F.). pp.1 User's Guide. 12. Stewart. A.56. Stewart. K. M.81 7838. Harlow: Addison Wesley Longman . J. and Kleinknecht. Sankhyii: The Indian Journal of Statistics 37. Acs.1. R. Vogelvang.References 355 Phillips. Vogelvang. Amsterdam: Free University 'Press. pp.H. (Eds. Zellner. A. P.W. J. and Gill . M. G. J .
.
34 1.3 ADF (A ugmented DF) tests 287. 209 Augmented DF (ADF) tests 287. W.82 autocorrelation 53. 34452 behavioural equations 12 binary variables 235. C. 264 DurbinWatson test 123. 0. 199.3 autocorrelatio n function (ACF) 121 .294. 327. ARMA models applied quantitative economic research 15.2. 321 . T.297 autocorrelated disturbances 179.Index .7 bivariate regression models 569 blockrecursive models 233 BLUE (best Linear unbi ased estimator) 72 Bollerslev. 33 1. 344 Almon procedure 3069 Amemiya.294. 33040.2 Akaike Information Criterion (A1C) 289.5 cau sal models 292 C haremza. 3478 BoxPierce test 1213 BreuschGodfrey test 11 921. 242. 340 artificial variables 15965 assumptions of models 1434 asymptotic theory 66. 114. 241.18 ARCH models 192. 99. 264 BreuschPagan test 12830 ARMA (autoregressive moving average) models 323.2. 11 3.3 breakpoint test 1645 BreuschGodfrey test 11 9. 74.25.297 adjusted detennination coeffic ient 91 . accuracy of forecasts 14850 ACF (autocorrelation fu nction) 121. 194 BoxJenkins procedures 322. 3234. 32830. T. 341. 11 7. 244 Anderson. 34(}"'52 BoxPierce test 12 1.33340. 3337 see also ARlMA models.0.33340.21. 325 Choleski decomposition 173 Chow breakpoint test 1645 .7 ARIMA (autoregressive integrated moving average) models 32 1. 327. 328. 6 C hatfield. 3413 aUlocovariance function 329 autoregressive (AR) models 3245. 3278. 327. 325 LjungBox test 1223 partial autocorrelation function (PACF) 327. 34 1. 32830.
245 and tim e~se ri es model s 348 coefficient of determination 86.11 likelihood ratio (LR) lests 945 relation ships among variables 869 residual variance estimator 956 coffee consumptio n model 910. 2767 command window 31 commodity price data 20. 192 data mining 6 Davidson. 1901 CUSUM statistic 1334 CUlhbertson.2956 interval esti mates 96. 3 17 GLS esti mates 3 14 habil fonnation 3 17 infinite distributed lag 311.767 errorcorrection models 3 1920 expectation variables 31819 finite distributed lag 306.11 geometrically distribuled lags 3 12. 24. 28 macroeconomic 212. 18990 correlation matrix 41 (Iests 85. D.12 lagged dependenl variable models 3 1620 long run responses 3 1416 mean lag 3 1416 multicollinearity 306 OLS estimales 314 partial adjustment mechanism 31718 di stributions and determini stic assumptions 924 geometric distribution 312 disturbance terms 14. D. 9. 3478 dichotomous variables see binary variables Dickey. 299304 coefficienl of delermination 86. 1901 economic 8 di screte regression models 236 di stribuled lag models 257. 11 . 97. 148 consistency of estimators 657. 2034. 3 14 diagnostic checking 328. 92. K.F.53 ARCH and GARCH models 1927 autocorrelation 17982 . 6 dependenl variables 89.14. 176 White covariance matrix 189. 85. J.1 complete model s 9 confidence intervals 98. 266.358 Index Chow forecasl lest 1545 classic regression model 51.10.478 commodity prices 201 crosssection 8. 200 consumption equation 67. 107 lest principles 945 Wald lests 95. 2934. 293 different regimes 1624 seriaJ correlation LM test 119. 265 Deadman. 78 determination coefficient see coefficient of determination deterministic assumptio ns 15. 1789 mi ss ing 30 money market 23 oil marketrelated 223 panelS sample 8 limeseries 8. 8992. 24 12 adjusted determination coefficient 9 12 pseudo coefficient of determination 242. 103.10311 . 749. 210 discrepancy variables 22 covariance stationary processes 3289 crosssection data 8.l . 100. 525. 2854 difference operators 2589 differencestationary process (OSP) 280.11 deterministic linear trend 15960 deviation form models 878 Dhrymes.E. 109. 1378.8. 75.H.16 lag pattern assumptions 3 11. 11 . 2645 correction of While 182. 294 data 1924 analysis 17.2 di stributions 924 FIeslS 85.16. J. 305. 190 DiNardo. 96103. 109.10.A.103 Lagrange mulliplier (LM) lests 95.20 see also autocorrelation covariance matri x 7 1. P.108.241 . 24. 93. 93. 147.111 classic regression model 523 deterministic assumptions 523 matrix notation 545 stochastic assumptions 53 coinlegrati on analysis 266.20 Almon method 3069 degrees of freedom 306 frequency of 3 12 generalion process (DGP) 172 importing 26. 8992. 222. 27.
265 nonstationary variab les 292303 notation 253.F. 20410.12.3 14 instrumentalvariance (IV) 20 14. 227 in EViews 2 1012 unbiasedness 65.9.12 restricted dynamic l SI structural I I .294. 22 I systematic part 14. 28 192 see also distributed lag model s econometric models 67 economic data 8 economic model s 67 efficient estimators 65.213 instrumentalvariance (IV) estimator 20 14. 200 desi rable statistical properties 647 efficiency 65. 210. 192. 180. 26577. 9. 266. 749. 222. 26 I longrun models 2534. 723 GLS (general least sq uares) 53. 51. 2 1012 endogenous variables 8.266. 42 importing data 26. 74.I 2.7 ca1culations 417 command window 3 1 correlation matrix 4 1 data ana1ysis 47. 261. 298 equations behavioural 12 consumption equation 67. 272 errorcorrection tenn 269 estimates 12 interval estimates 96.14 computations 45. 28 Newey· West estimator 190 and residuals relationship 612 simultaneous equation models (SEM) 21718 dlog 259 dummy variables 159. 2034. 2645 nonsystematic part 15 reducedform 11.7 testing 99 endogenous explanatory variables 199.103 estimators 12 asymptotic theory 66.2 15 and the OLS estimator 199200 twostage least squares estimator (2SLS) 200.8 twostage least sq uares (2SLS) 200. 235 DurbinWatson test 1235 dynamic models 10. II. 2812 shortrun model s 253. 778 Engle. 723 elasticiti es 12. 20410. 182. 199 consistency 657.I 5 errorcorrection models 254. 26577 • lag operators 253. 2549. 2549. 25965 sum operators 2567 trends 2788 I unit roots 254. 31920 firstorder lag polynomials 26770 hi story of 2656 .8 dlog 259 elasticity computations 457 forecasting 1503 frequency of data 3 12 Graph option 326 Multiple graphs 394 I group statistics 4 1 groups of objects 379 hi story list 26. 99. R. 1724. 2 I 5 in EViews 21012 maximum likelihood (ML) 7980 NeweyWest 190 residual variance 956 sample mean 67 simultaneous equation model s (SEM) 22630 and stochastic assumptions 1347 threestage least squares (3SLS) 227. 702 see also OLS (ordinary least squares) estimator evaluation of results 17 EViews 2548 Almon procedure 309 Arch test 1947 binary dependent variable models 242. 253304 cointegration analysis 299304 difference operat'Ocs 2589 errorcorrection models 254. 26 1.I Index estimation 63~ heteroskedastic 18292 359 irreversible behaviour of explanatory variables 2736 irreversible models 2747 restricting the longrun 2667 reversible models 274 secondorder Jag polynomials 2703 stepresponse function 270.
345 Chow forecast lest J545 confidence intervals 98. 93 .30 OoldfeldQuandttest 1302 workfiles 2632 exchanging variables 31 exogenous variables 8. 3 12. 294. W. 266. 220 finite distributed lag models 306. 325 Hendry. 28 final fonn simultaneous equation models (SEM) 219. 317 GLS (general least squares) estimator 53. 93.52 frequency of data 3 12 fullinformation estimation 2278 Fuller. 183. 182 and distributed lag models 314 . 75 correction of White 182. 2 15 in EViews 2 1012 interval estimates 96103 inverted roots 328. 2703 lagged dependent explanatory variables 767 J agged dependent variable models 31620 GaussMarkov theore m 72 Gaussian processes 278. 293 . 294 Hamilton. 147.O. 100. 1724. 227. S.A.3 geometric di stribution 3 12 geometrically di stributed lags 3 1214. 26770.16 instrumentalvariance OV) estimator 20 14. 379 Options button 3 1 polynomial distributed lag (POL) 309 preview window 28 Procs bulton 354 1 Range buno n 30 regression models 803 restriction testing 1057 Sample button 30 saving files 3 1 SUR estimator 1778 View button 324 1 OoldfeldQuandt test 1302 Oranger.W. 2824 irrevers ible mode ls 2747 JarqueBera (JB) test 115. 180.7 Keynsian macro model 9 Kim. C. 75 and stochastic assumptions 125.M . 11 3 irreversible behav iour 2736 Ftests 85.2956 hypothesis formulation 5. 294 Johnston . 6 hypothesis testing 6 identification of equations 22 16 identities 12 importing data 26. J . 265 heteroskedasti city 53. 254. 254.I2 lag polynomial s 2556.9 White test 1268 expectation variables 3 18. 257. 265. 322. 298 Oraph option 326 Multiple graphs 394 1 Oreene.360 Index EViews (Continued) instrumentalvariance (IV) estimator 21012 inverted roots 328. 266. K. 54. 294. 26 1 lag pattern assumptions 311 .11 fi rstorder lag polynomial s 26770 forecasting 14355 accuracy of forecasts 14850 assumptions of the model 1434 infinite distributed lag models 3 1 1. 10311 . 6. 294 OARCH models 192. 113. S. L. D. J. 18990 history list 26. 210 Juselius. II .32 in timeseries models 192 weighted least squares method 182. 18292 and binary dependent variable models 243 BreuschPagan test 128. 42 homoskedastici ty 53.H. 2549. 51 . 325 group statistics 41 groups of objects 379 habit formation 317 Hall . W. 148 in Eviews 1503 mean estimation 148 prediction error 1456 prediction iOlervals 1448 with timeseries models 328. 349.9. 3 19 lag operators 253.19 explanatory variables 89.17 Johansen.D.J. 345 log transfonnations 424 missing data 30 objects 26. In Moo 254 Koyck.F.
2245 order of integration 225. 244 longrun models 2534. F. 229. 1712 and distributed lag models 314 distribution and properties of 6779. 2812 matrix notation 545 si multaneous equation models (SEM) 2 1621 objects 26. regression models linear transfonnations 6770 linear trend 159{i() LjungBox test 1223 LM (Lagrange multiplier) tests 95. 263 Maddal a.8.16 missing data 30 model stability 133 money market data 23 moving average (MA) models 3256. 266 orthogonality properties 86 outliers 1601 Palm.C. 27()"'3 prediction error 1456 intervaJs 1448 see also forecasting preview window 28 probability limits 66 probit models 236.Index lagged variables S9. 767. 254 market regimes 1624 matrix notalion 545 maximum likelihood (ML) estimators 7980 mean absolute error (MAE) 148. 424. 148 mean lag 3 14.341 . 12. 245 Qtest 12 1. 293 NeweyWest estimator 190 nonstationary variables 292303 nonnality test 11517 notation 253. 33040. 289 qualitative dependen t variables 23547 binary dependent variable models 2427 linear probability models 2369 logit models 23942 probit models 23942 random walk processes 27981 Range button 30 rank condition 2245 recursive coefficient estimates 132 macroeconomic model 21921. 2549. 178. 266. 209. 23942.3 Perron. 287 polynomial distributed lag (POL) 309 polynomials 2556.52. 200 and endogenous explanatory variables 199.2. 266. 556 1. P.10 see also estimators partiaJ adjustment mechanism 3 17. 265 LR (likelihood ratio) tests 945 MA (mov ing average) models 3256. 33740 McFadden.30.3. 306 Multiple graphs 3941 multivariate regression model 5961 NeWbold. 0. 321 panel data 8 parameters 67. P. 51. 1617 estimation in time·series models 3278. 287 Phillips.200 Options button 31 order condition 223. 235. 37. lOSII latent variables 239 likelihood ratio (LR) tests 945 linear discriminant function 236. 266. 265.B.9 361 OLS (ordinary least squares) estimator 50.50 mean absolute percentage error (MAPE) 149 mean estimation 67. 23942.14. 242 see also reducedfonn model.5. P.9 oil marketrelated data 223 . 239 linear models 67. 11. 244 procedures 3541 Procs button 354 1 pseudo coefficient of detennination 242.18 partial autocorrelation function (PACF) 327. 26770. 10811 log transformations 13. 344 7 stability 1324 testing restrictions on 96107. 242 macroeconomic data 21 .99 Lagrange multiplier (LM) tests 95.278 logit models 236. 33740 multicollinearity 155. 10.C. 869 bivariate models 15 probability models 2369.
25 classic regression model 53 estimation procedure 1347 heteroskedasticil'Y tests 12532 normalilY lest 11517 parameter stability 1324 stochastic model s 14. 255 stochastic assumptions 16. 79.199.P. 51 .2. 259{j5 simplerec ursive models 2302 simuhaneous equation models (SEM) 9. 107 sum operators 2567 SUR models 1749 (Iesl 85. 778. 265 standard errors 71 .259 secondorder lag polynomials 2703 seri al correlation test 119.12. 96. 52. 117.12. 96103. 6 12. 2034. 2245 rank condition 2245 recursivity in 2303 reducedform 218.17.2.79 variance estimation 634 see also regression models reducedform simultaneous equation models (SEM) 21819. 265. 52. 75. 294 technical relationships 12 test principles 945 Theil inequality coeffic ient 1489 threestage least squares estimator (3SLS) 2278 timeseries dala 8. 192 . 11 338 autocorrelation tests 114. 265. M. 171 asymptotic 209 static models 10 stepresponse function 270.18 estimation methods 22630 final form 219. 525.10 correlogram of 138 variance estimator 956 see also autocorrelation restricted dynamic equations 181 restriction testing 1057 reversible models 274 rOOI mean squared error (RMSE) 14850 Sample bullon 30 sample data 8 sample mean 67 sampling error 65. J. 3289 stochastic vectors 6770 strongstationary process 278 structural equations 11 . 344 scientific quantitative economic research 56 seasonal adjustments 37.2. 220 singleequation estimation 2267 slfUclural form 2 161 8.15 stochastic processes 278.5 183 di sturbance term estimation 634 disturbances and residuals relationship 6 12 lagged dependent explanatory variables 767 maximum likelihood (ML) estimalors 7980 OLS (ordinary leasl squares) estimalor 50. 55{j.362 Index recursive models 2303 blockrecursive 233 simplerecursive 2302 reducedform equations 1112 reducedform model 17. 209.18. F.220 slfUclural models 11 . 17 1 spurious regression 254. 220 fullinformation estimation 2278 identifi cation of equations 2216 mathematical notation 21621 order condition 223. 92. 107 TARCH model 194 Taylor. 78 discrete 236 in Eviews 803 multivariate 5961 spurious regression 254. 1378. 70 saving files 3 1 scalars 41 .20 shortrun models 253. 92.103. 55{j1. 272 Slewart. 213 Sludenl's (leSI85. 67. 266 relationships among variables 869 research applied quantitative economic 1518 scientific quantitative economic 56 residuals 54. 22 1 structural form simultaneous equation models 2 161 8. 266 Srba. 151 . 161 . 220 regimes ruling the market 1624 regression models bi variate 569 classic 5 1. 27. 2 1533 di sturbance terms 217. 220 specification errors 13943. 44 Schwartz Information Crilerion (S IC) 289.19.
28192 Augmented DF (ADF) tests 287. 3478 forecasting with 328. 340 variables 8. 276.213 exchanging 3 1 exogenous 89. 33040. 190 white noi se 322. 321 zeroslopes test 105 . 332 see also dynamic models transformations linear 6770 log 13. 424. 327.2 too many 140. 702 unit roots 254. 265 and causal models 292 constant term and linear trend 286 constant term and no linear trend 2856 no constant term and no linear trend 2845 univariate timeseries models see timeseries models YuleWalker equations 331. 34 1.19 explanatory 89.3413 autocorrelation tests 328. 3234.3 1620 latent 239 nonstationary 292. 333. 294 View button 3241 Vogelvang. 1423 volatility of 192 variance estimation 634 Verbeek. 2736 lagged 8. 34952 heteroskedasticity 192 moving average (MA) models 3256. 3478 autoregressive (A R) model s 3245. E.10.3278. 332 Zellner. 33740 parameter estimation 3278.7 BoxJenkins procedures 322. 347.8. M. 141 . 235 endogenous 89. 11.10. 1839 White covariance matrix 189. 34 I. 278 trendstationary process (TSP) 279. 3447 partiaJ autocorrelation function 327.5 1.3 s[Qchastic process 278 YuleWalker equations 33 I. 204.2. 33340.303 order of integration 225.~I~"""""""""""""""""""""""""""""""""""""" Index timeseries models 321 .52 ARlMA models 32 I. 10311 asymptotic standard errors 209 direct derivation of 2078 in EViews 21012 first stage 2067 weakstationary process 278 weighted least squares method 182.77.11 . 75. 297 exchanging variables 3 1 Yeo. 18990 workfi les 2632 historical introduction 2045 second stage 207 unbiasedness of estimators 65. 294 volatility of variables 192 Wald tests 95.767. A. S. 32830. 11 .767 lagged 3 1620 qualitative 23547 363 autocorrelation function 327. 33 1.34052 coefficient of determination 348 diagnostic checking 328. 2427 dependent 89. 34452 ARMA models 323. 199213. 266 orthogonality properties 86 qualitative dependent 23547 relationships among 869 too few 140. 75 expectation 318.99. 11 .9. 266. 199. 3241 artificial 15965 binary 235. 294.8 White test 1268 correction of White 182.5 1.227 discrepancy 22 dummy 1~9 . 293 trends 27881 twostage least squares estimator (2SLS) 200. 11 .
Eoo..1OIIlIsIs 818 regularly confronIed resuIIs of qu&lltilatill8 economics research. showing how to conduct your own econometric research project and write a complete research paper. • The author has carefully selected only the relevant statistical and mathematical theory to discuss the application of tests. but avoiding superfluous manual computations. their underlying assumptions and how estimates of parameters or 818 computed. mathematical approach to applied econometric research. KEY FEATURES • An integrated discussion of econometric theory and applications with EVlews . Econometrics: Theory and AppI/cations with EVIews provides a broad Illboduction to quantitative economic how arise. econometrics in the Department of Econometrics. • A quantitati118 economic rasesrch project is simulated In the case studies throughout various chapters using real economic data. Free University of I SBN 0. economic The auIhor combines practice by demonstrating its with the software package EViews. • Includes examples. The is on how to select the right method of analysis for a gillen situation.illustrating the benefits of a statistical..pearsoned co uklvoplvang . • Numerous screenshots of EVlews windows have been included to clarify its use. and how to actually apply the "led !Ods in the right WBIf. and the consequences of those tests. or for those requiring knowledge on the practical application of EViews to economic data.273683748 FT Prentice Hall FINANC IAL TIMES 9 AddH:ional student support at An imprint of www. with datasets availabie to download from the book's website at LIS on troductory econometrics courses taken es a pert of an economics degree or quaJIftcaIIon. data and exercises.