This action might not be possible to undo. Are you sure you want to continue?
Long-Run Growth Forecasting
Long-Run Growth Forecasting
Stefan Bergheim Raimundstr. recitation. etc. The use of general descriptive names. Heidelberg. specifically the rights of translation. whether the whole or part of the material is concerned. reproduction on microfilm or in any other way. 121 60320 Frankfurt Germany stefan. and permissions for use must always be obtained from Springer-Verlag. trademarks. Violations are liable for prosecution under the German Copyright Law. and storage in data banks.bergheim@googlemail. in this publication does not imply. broadcasting. reuse of illustrations. 1965.com . that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Germany Printed on acid-free paper 987654321 springer. Cover design: WMXDesign GmbH. reprinting. even in the absence of a specific statement.com ISBN 978-3-540-77679-6 e-ISBN 978-3-540-77680-2 Library of Congress Control Number: 2008923365 © 2008 Springer-Verlag Berlin Heidelberg This work is subject to copyright. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9. registered names. in its current version. All rights are reserved.
To my mother .
Seminar participants at ETH Zurich. Chiara Osbat introduced me to the non-stationary panel literature. Special thanks go to Marco Neuhaus u for the long discussions in 2004 and 2005 that inspired large parts of this study. the Max Planck Institute of Economics. the Institut der deutschen Wirtschaft Cologne and the Macroeconomic Policy Institute (IMK) in D¨sseldorf also made helpful suggestions. Elga Bartsch and Sarah Rupprecht were so kind to comment on drafts of the book. and Magdalena Korb. Deutsche Bundesbank. Andrea Schneider. I would like to thank Michael Frenkel for supervising this project and J¨rgen Weigand for being co-supervisor. u .Acknowledgement This book is based on my dissertation at WHU Otto Beisheim School of Management.
. . . . .3 The concept of TFP is not helpful . . . . . . . . . . . . 2. . . . . . . . . 2. . . . . . . . . . . . . . . . . .1 Production function cannot be estimated . . . . . . . . . . . . . . . . .1 Application in cross-country analysis . . . . . . . . .1 The search for a dynamic model . . . . . . .1 Choosing the appropriate data source . . . . . . . . . . . . . . . . . . . . .2 Modeling barriers to riches (Parente & Prescott) . . .3 Open-system models .4. . . . . . . . . . . . . . . 2. . . . . . . . .4. . . . . . . . . 2. . . .4 Models with deeper insights . . . . . . . . . . .2 Strong demand . . .6. . . . . . . . . .2 Evolutionary models of growth . . . . . . . . .5. . . . . . . . . . . . . . 2. . . . . . . . . .3. . . . . . . . . . . . . . . . .5. . .7 The augmented Kaldor model . . . . . . . . . . . . . . . . . . 1 1 3 6 6 8 9 9 10 12 13 15 16 17 18 19 20 21 24 25 25 28 29 29 30 2 3 The dependent variable: GDP growth . . . . . . . . . 1. 2. . . . . 2. . . . . . . . . . . 2. . . . . . . . . .1 Models with scale eﬀects . . . . 2. . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6. . . . . . . . . . . . 2. . . 2. . . . . . . . . . . .2 Aggregate production function does not exist . . . .6 General critique of the standard approach . 1. . . . . . .3 Focus on convergence . . . . . . . . . .3. . 1. . . . . .2 Choosing the best econometric technique . . . . . . . . 1. . . .1 Including human capital (Lucas) . . . . . . . . . . . . . . .4 Beyond neoclassical economics . Assessment of growth theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . . .1 Tests for conditional convergence . . . . . . . . . . 35 3. . . . . . . .6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . .Contents 1 The importance of long-run growth analysis . . . . . . . . . . . . . . . . . . . . . . . 2. .1 Frequent forecast failures . . . . . 2. . . . 2. . . . . . . . . . .5 Opening the theories further . . . . . . . . .2 The basic neoclassical model . 2. . . . . . . . . . . . . . . .5. .1 Choosing a sensible theoretical model . .3. . . . . . . . . .3 Plan of work . . . . . . . . . . . . . . . . . .6. . . .but little supply . . . .2. . . . . . . . . . . . . . . . . . . . . . . . 36 . . . 1. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . 6. . 5. . .diﬀerent investment ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . 97 8. . . . . . . . . . . . . . . 5.2 Hours worked per capita are important . . . . . . . .1 Population growth is endogenous . . . . . . . . . . . . . . . . 7. . .and macroeconomic theory . . . . . . . . .1 Investment and changes in capital stocks . . . . . . 5. . . . . . . . . . . . . . . . . . . . 6. . . . . . . 7. . . . . . . . . .1 Micro. . . . . . . . 98 8. . . . . . .1. . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . Human capital . 7. . . . . . . . . . . . . . . .2 Investment ratios do not diﬀer much across countries . . . . . . . . . . .2 Main insights on capital accumulation . . . . . . . . . . . . . .2 Good macro polices and more competition .1 Spatial economics .3 Age structure of the population . . . . . . . . . . . . . . . . . . . . . .2.1. . . . . . .1 Measuring capital accumulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . .2 Relative location: rich neighbors . . . . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . 101 . . . 7. . . .1 Investment ratios are not constant . . . . . . . . . . . Physical capital . . . . . . . . . . . . . . . 97 8. . . . . 6. . . .2. . . . . . . . . . .1. . . . . . . .2 Measuring openness . . . . . .2. . . . . . . . . . . . . . . . . . . . . .2 Macroeconomic models with diﬀerent conclusions . . .1. .1 Extent of the market and specialization . . . . . .2. . .location matters . . . . . . . . . . . . . .3 Sum: Spatial linkages not much help . . . .1. . . . . .2. . . . . . . . . .3 Best measure: adjusted trade share . . . . . . . . . . . 7. . . . . . 7. . . . . . Openness .1 Microeconomic analysis: labor economics . . . . . . . .1. . . . . .X Contents 4 Labor input . . . . . . . . . . . . . . .1. . . . . 43 43 46 48 51 52 52 53 54 56 56 56 60 61 61 63 63 67 69 70 71 74 75 81 83 84 84 86 86 87 87 88 90 5 6 7 8 Spatial linkages . . 5. . .2 Constructing spatial GDP . . . . . . . . . 7. . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 8. . . . . . . . 5. . . . . . . . . . . 5. . .1. . . 7. . .1 Best measure: years of education . . . . . 5. . . . . .6 Capital accumulation is not exogenous . . . .2 The openness dummy . . . . . .4 Investment ratios are not proportional to levels of the capital stock . . . . . . . . . . . .1 Theory: higher eﬃciency . 4.1 Black market premium and tariﬀs . .2.3 Empirical debate: levels versus growth .5 Capital productivity does not correlate with income . . . . . . . . . . . . . . . . . . . . . . . . . .3 Additional inﬂuences of trade on income . . . . . . . . . . . . . . . . . . . . . . . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Diﬀerent databases . . . . . .3 Proper modeling of capital accumulation . . . . .1 Absolute location: latitude and climate . . . . . . . . . . . . . . . . . . .3 Capital stocks from perpetual inventory . . . . . . . . . . . . . . . .2 Measures and empirical analysis . . . . . . . . . . .2. . . . . . . . . . . . . . . . . 6. . . . . . . . . . . 5. . .2. . . . . . . . . . . . . . . . . . . . .2. . . . . . . . 97 8. . 4. . .1. . . .3 Investment ratios are not proportional to changes in the capital stock . . . . . . . . . . . . . .1. 5. . . . . . . . . .
Other determinants of GDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
10 The theory of forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 10.1 The beneﬁts of forecast experiments . . . . . . . . . . . . . . . . . . . . . . . 106 10.2 The characteristics of good forecasts . . . . . . . . . . . . . . . . . . . . . . . 106 10.3 Intercept correction and forecast combination . . . . . . . . . . . . . . . 109 11 The evolution of growth empirics . . . . . . . . . . . . . . . . . . . . . . . . . . 113 11.1 Still widely used: cross-section . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 11.2 Weaknesses of cross-section regressions . . . . . . . . . . . . . . . . . . . . . 116 11.2.1 Same production function assumed . . . . . . . . . . . . . . . . . . 117 11.2.2 Long-run growth path assumed to be constant and the same across countries . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 11.2.3 Same pace of conditional convergence assumed . . . . . . . . 118 11.2.4 Errors are assumed uncorrelated with the explanatory variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 11.2.5 Right-hand side variables assumed exogenous . . . . . . . . . 118 11.2.6 In sum: many assumptions are violated . . . . . . . . . . . . . . . 119 11.3 The climax of cross-section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 11.4 Advantages of panel techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 11.4.1 Initial technology can diﬀer across countries . . . . . . . . . . 123 11.4.2 Dealing with endogeneity bias . . . . . . . . . . . . . . . . . . . . . . . 123 11.4.3 Addressing lagged dependent bias . . . . . . . . . . . . . . . . . . . 124 11.4.4 Modeling heterogeneous technological progress . . . . . . . . 125 11.4.5 Summary of results from panel regressions . . . . . . . . . . . . 125 11.5 Non-stationary panel techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 11.5.1 Pooled mean group technique . . . . . . . . . . . . . . . . . . . . . . . 126 11.5.2 Testing unit roots and cointegration in panels . . . . . . . . . 128 11.5.3 Panel unit root tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 11.5.4 Panel cointegration tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 11.6 A two-stage estimation method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 12 Estimation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 12.1 Correlation analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 12.2 Panel unit root tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 12.3 Panel cointegration test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 12.4 The short-run forecasting models . . . . . . . . . . . . . . . . . . . . . . . . . . 146 13 Forecast competitions and 2006-2020 forecasts . . . . . . . . . . . . . 151 13.1 Forecast competition 2001-2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 13.2 Forecast combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 13.3 Forecast competition 1996-2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 13.4 Forecasts for 2006-2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 13.5 Other long-run forecasting models . . . . . . . . . . . . . . . . . . . . . . . . . 160
14 Conclusion and outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 List of ﬁgures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 List of tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
List of variables and coeﬃcients
Y total GDP y GDP per capita (Y /L) y change of GDP per capita dy/dt ˙ y percentage change of GDP per capita y/y ˆ ˙ K physical capital k physical capital per capita ˆ k growth rate of per capita capital stock a share of capital in national income α elasticity of output with respect to capital input r real interest rate I investment in physical capital κ growth rate of investment L total labor supply n growth rate of population Lt = L0 ent w real wage rate H human capital h human capital per capita ˆ h growth rate of human capital per capita S years of education R work experience o trade openness per capita A level of “technology” or of total factor productivity TFP g growth rate of technology At = A0 egt E country-speciﬁc eﬃciency x other drivers of per-capita income s savings rate sk saving to accumulate physical capital sh saving to accumulate human capital t time i index of countries u fraction of human capital reinvested
or regression coeﬃcient γ capital-output ratio K/Y δ depreciation rate regression error η coeﬃcient in Dickey-Fuller regression λ rate of convergence µ coeﬃcient on trade openness in technical progress function π coeﬃcient on cointegration errors ρ rate of time preference σ rate of risk aversion φ return to human capital ω coeﬃcient on human capital in technical progress function .XIV List of variables and coeﬃcients β elasticity of human capital.
List of abbreviations ECB European Central Bank ECO Economic Outlook database (by OECD) GDP Gross domestic product IMF International Monetary Fund MRW Mankiw. Romer and Weil (1992) OECD Organisation for Economic Co-operation and Development PMG Pooled mean group estimator PPP Purchasing power parity PWT Penn World Table R&D Research and development RMSE Root mean squared error TFP Total factor productivity UN United Nations WDI World Development Indicators (by World Bank) WEO World Economic Outlook (by IMF) .
the USA was seen by many as a sclerotic economy destined for anemic economic growth with high unemployment and to be overtaken by Japan within a few years. companies and individuals play. Germany is another case where trend GDP growth has been overestimated signiﬁcantly for the past 10 years. these predictions could not have been more wrong. Year after year. some investment . In the early 1990s. history is full of examples of poor predictions and therefore poor decisions. a decade of economic stagnation. As we know today. By contrast. knowledge about the drivers and linkages that determine the future will allow these players to actually shape the future themselves.7%. which would allow them to act appropriately. 1. To detect challenges and opportunities in a timely manner decision-makers require a good forecasting framework.1 Frequent forecast failures Unfortunately. Asset markets in the US surged as they became increasingly conﬁdent that the future would be much brighter than assumed in the early 1990s. Actual growth over the years 2001-05 was just 0. Had investors known already in the mid-1990s just how low Germany’s growth potential was.1 The importance of long-run growth analysis Forecasts are usually made to help and guide decision making. falling asset prices and banking sector problems followed and made many forecasters look incompetent.3% per year between 1992 and 2005. Given the role governments. From about 2% in 1995 the consensus forecast for trend growth was revised down to around 1% by 2005. informed decisions. Good forecasts are preconditions for good. Japan in the early 1990s was seen as a role model. decision-makers should be as well prepared as possible for the future. growth expectations of investors and companies had to be revised downwards. In the event. Growth of US gross domestic product (GDP) averaged 3. These decisions may vary from a ﬁnancial market bet on interest rate changes to the policy decision on how to structure a country’s pension system. Ideally.
First. The forecasts in the International Monetary Fund’s semi-annual World Economic Outlook (WEO) displayed a tendency to systematically overpredict real GDP growth as Timmerman (2006) shows.4%. Between 1991 and 2003. The frequent crises in emerging markets over the past two decades tended to be even more severe and surprising. Economic growth. This bias points to a signiﬁcant overestimation of trend growth. in some way.and wished they had had a framework to tell them to stay engaged in these countries.not a negligible amount. while growth over 2001-05 actually averaged just 1.” In trying to help reduce forecast errors in the future. on everything that goes on in a society. data and/or statistical tools to adequately model the developments of national economies. 9). For example.1 percentage points. Italy and Japan were on average a full percentage point too high. For the year 2003 the staﬀ forecast for GDP started out at 2. Even worse. 13) observed that “The growth rate of an entire economy is not an easy thing to move around. the next-year forecasts in the September WEOs for France.5% . cit.7%. These persistent and large forecast errors indicate that economists do not yet have the appropriate theories.way above the ﬁnal outcome of 0. Forecasts for 2002. Repeated small diﬀerences in growth rates can lead to large diﬀerences in outcomes many years down the road. the 1997 crisis in emerging Asia caught many investors by surprise . This anecdotal evidence is supported by more formal analysis.5% does not look all that diﬀerent from a growth rate of 2%. the task at hand is really enormous: Lucas (1988. This would entail a more thorough modeling of trend growth using elements of growth theory. necessarily depends. In the ﬁrst quarter of 2001 the ﬁve-year ahead GDP forecast was 2. But over 20 years. after retreating from Asia during the crisis.5%. A growth rate of per capita GDP of 1. being a summary measure of the activities of an entire society. many companies were surprised by the rapid rebound of countries like Korea and Malaysia . it provides an assessment of the main existing theories of economic growth and proposes an augmented Kaldor model as the most . The IMF is not alone in having made these systematic forecast errors. Unfortunately. Indeed.who wished they had been better able to anticipate the diﬃculties. They are also visible in the European Central Bank’s (ECB) staﬀ forecasts for the euro area and in consensus forecasts. Forecast errors are not conﬁned to developed markets. p. Germany. p. one of Timmerman’s recommendations on how to address these forecast errors is to have “more frequent reviews of estimates of potential output growth” (op. this translates into a 10% diﬀerence in income levels . 2004 and 2005 were also too high by between 0.4 and 2.2 1 The importance of long-run growth analysis plans would have turned out quite diﬀerently: production capacities would not have been expanded as much and investors would have avoided companies with a large exposure to German domestic demand. this study makes three contributions. The consensus according to the ECB’s Survey of Professional Forecasters did not fare much better.
In economics the term “growth” already refers to the long-run development of an economy. the theoretical and empirical growth literature has not yet produced a consensus on some of the most important questions: How important is the accumulation of physical capital for GDP growth? Is investment exogenous or endogenous? Should population growth be treated as exogenous? Does an increase in education lead to higher output? What is the best econometric technique to try to answer these questions? 1. it presents longrun growth forecasts for the years 2006-20.2 Strong demand . the datasets used and has to crosscheck the insights with real world experience. Just-in-time production may ease some of these diﬃculties.but little supply 3 reasonable synthesis. the path to these forecasts is at least as important. A forecast competition will show that forecasts based on the theories outlined here can outperform consensus forecasts and simple time-series models. Growth forecasts are used in many areas in business and ﬁnancial markets and by governments. this study applies modern non-stationary panel estimation techniques to test these hypotheses for 40 countries for the period from 1971 to 2003. The need for a neutral forecast is particularly strong here because individual business units have a genuine interest in presenting high forecasts. . losses for the whole company may ensue. A forecaster has to understand the assumptions made. If production and inventories are too high relative to actual demand in the future. While the ultimate goal of this study is to derive a set of forecasts. this study uses the term “long-run growth” to avoid any uncertainty regarding the time horizon. Businesses require forecasts for economic growth for their budgeting. Since many corporate investments. have investment horizons of 10 to 15 years they also require GDP forecasts over a similar horizon. such as a new chemical plant. but production capacities nevertheless have to be aligned with expected demand.but little supply Demand for substantiated long-term growth forecasts is high following the surprises and forecast errors made in the past. since the media and ﬁnancial markets frequently use “growth” when referring to changes in GDP over shorter periods of a year or even a quarter (which are the combined result of trend and cyclical factors). which may steer the allocation of resources to their unit. However.2 Strong demand . strategic plans and selected business cases are based on wrong assumptions. strategic planning and for the analysis of business cases. Unfortunately. The time horizon is the medium to long run of 5 to 20 years. Second. then prices may need to be set below initial plans to clear inventory.1. If budgets. It is the ﬁrst work to derive hypotheses of pair-wise cointegration among the key variables in growth models. the limits of theories. the evolution of its potential or trend output. And third.
Wrong estimates of revenue and expenditure may lead policy-makers to cut tax rates and expand welfare spending. If a central bank overestimates the trend growth rate. this may lead to a series of downward revisions of growth forecasts and upward revisions of inﬂation forecasts . g. A systematic analytical framework and a set of conditional forecasts for growth would make their tasks easier. this ”over-optimism may lead to complacency regarding the adequacy of growth-oriented structural reforms pursued by a country. Fund managers try to outperform their peers by comparing the growth forecast that the market is pricing in at the moment with their own. For example. possibly model-based. Business cycle analysis usually starts from trend growth and then adds or subtracts from it depending on the current state of policy variables and exogenous developments (e. many pricing models are based on the economy’s underlying growth trend: Government bond yields are often priced on the sum of expected real GDP growth and inﬂation. policy-makers are interested in speciﬁc advice on how to strengthen their countries’ growth performance . As Batista and Zalduendo (2004) emphasize.or how to prepare for geopolitical changes resulting from diverging economic outcomes.as seen in Europe and especially in Germany since 2001. national ﬁscal authorities require solid forecasts for trend GDP growth to estimate future tax revenue and pension liabilities. And these bond yields are themselves the benchmark against which other assets (equities. real estate) are priced.4 1 The importance of long-run growth analysis Financial markets make heavy use of long-run GDP forecasts in many ways. central banks need a good grasp of the growth rate of potential GDP over the medium term. This meant that budget deﬁcits turned out much higher than expected and led to major political upheaval inside the European Union because several countries did not comply with the Stability and Growth Pact.as seen in many European countries since 2001. Around the turn of the millennium. A stabilization program and the associated recommendations may look quite diﬀerent depending on the economy’s underlying growth potential. They would prefer to invest in markets that few others see as promising today but that will show their strength in the near future. exchange rate). The result would be unexpectedly high ﬁscal deﬁcits . It turns out that the IMF’s medium-term growth projections have a tendency to err on the high side. But most of these analyses use past trends as a starting point. Furthermore. the current growth rate of GDP is compared to the rate of potential GDP growth. many European governments used GDP forecasts that turned out to be too high because they were too optimistic both on the cyclical and on the trend development of GDP. Long-run growth forecasts are also important for international organizations like the World Bank or the International Monetary Fund. If trend growth is on a downward trajectory. it may supply too much liquidity and end up with unexpectedly high in- . When assessing the risk of overheating of an economy. oil.” In addition. forecast for long-run growth. Likewise.
Often. .but little supply 5 ﬂation. In the event. When trying to apply principles of economic theory to build a forecast model. The Handbook of Economic Growth published in 2005 does not include a chapter on forecasting in its two volumes with a total of 1998 pages.5% . the GCI or the IMD index does not seem to be a good predictor of economies’ future growth prospects. While demand is strong.just ahead of a decade of very weak growth.just ahead of a decade of very strong growth. The theoretical model outlined in chapter 2 and the discussion of the individual variables in chapters 4 to 9 beneﬁted strongly from having to be useful for forecasting. this natural rate is derived from the economy’s long-run growth potential.1. there is a slight negative correlation between the 2001 GCI ranking and actual GDP growth over 2001-05.2 Strong demand . (2003). 1 See Blanke et al. A stronger focus on forecasting in growth economics may also have positive eﬀects on the development of economic theory. This study is partly about what economists can learn when applying the ideas of growth theory in a real-world forecasting context. in 1999 and 2000 the European Central Bank came under pressure to revise its reference value for the expansion of M3 money supply upwards because strong current GDP growth had led many ﬁnancial market analysts to revise upward their forecasts of the euro area’s potential GDP growth. there is a scarcity of substantiated long-run growth forecasts. the WEF produced a joint ranking with the Institute for Management Development (IMD). The ”Growth Competitiveness Index” (GCI) developed by the World Economic Forum (WEF) claims to “evaluate the potential for the world’s economies to attain sustained economic growth over the medium and long term”1 and receives a lot of media attention every autumn partly because it is one of just a few models in the face of the strong demand.possibly because it has superior capabilities in modeling the economy’s potential GDP. For example. the usefulness of these principles is put to a test. Similarly. This forecasting weakness has some tradition: In the early 1990s. Some models used by central banks and international organizations just extrapolate the past trend of GDP growth or of labor productivity into the future using simple statistical tools like the Hodrick-Prescott ﬁlter. Finland and Korea ranked 25th and 28th in 1993 . Private institutions either make ad-hoc assumptions on future growth or generate models that may not deliver what they claim to do. The 1993 version saw Japan and Germany ranked at numbers 2 and 5 . the ECB correctly opted for maintaining its assumption of potential GDP growth at 2 to 2. Neither ﬁlls the gap between high demand and low supply of long-run growth models. Therefore. By contrast. Unfortunately. p. Academic research into economic growth shies away from exploring the forecasting performance of growth models and focuses mostly on explaining the past. central bank reaction functions such as the Taylor rule require a ‘normal’ or natural real interest rate as an important input. 3.
Evaluating annual observations either of growth rates or of GDP levels is likely to produce much larger absolute forecast errors than evaluating the averages. Wherever possible. By contrast. However. I will use data until 1995 to derive forecasts for average annual per capita GDP growth over the period 1996 to 2005 (and data until 2000 for forecasts over 2001 to 2005). there is still no consensus on the drivers of economic growth even today. discuss the individual drivers of economic growth and decide on the most appropriate econometric technique. However. Indeed. the model does not aim at explaining the business cycle. these models and their forecasts are not usually available to the public. to allow some out-of-sample testing of the model. In addition. The forecast horizon of 2020 is motivated by the average investment period of large projects initiated by companies or governments.3 Plan of work This study is about forecasting long-run GDP growth both per capita and in total. in most countries. the history of the past 200 years shows that economies can create tremendous riches and move far away from poverty. In principle. private-sector institutions such as banks and consultancies added their own models. 1. Chapter 13 will discuss these contributions as well. dooming mankind to unending poverty and hardship.3. It will derive forecasts for average annual GDP growth for the period 2006 to 2020 for 40 economies around the world based on models with annual frequencies using the most current data available. it would be possible to calculate and evaluate estimates for each year over the forecast horizon. Recently. 1. which economists continue to struggle to explain. the span from 2000 to 2005 seems to be close to a peak-to-peak period. . I will compare my models and insights with those from these forecasters. On the way to these forecasts I will evaluate the diﬀerent theoretical models. Ricardo and Marx drew similarly gloomy conclusions.1 Choosing a sensible theoretical model Economics carries the stamp of the ”dismal science” partly because of the gloomy predictions of Thomas Malthus in 1798 that population would grow faster than food supply. The reason is substantial technological progress. The model forecasts will be evaluated against a set of alternatives in chapter 13. A ﬁve-year horizon should be long enough to average out business cycle disturbances. While this shows that the models of Malthus and Ricardo were clearly misspeciﬁed.6 1 The importance of long-run growth analysis International organizations such as the IMF and the World Bank maintain a set of models to forecast long-run GDP growth for a large number of countries. Exceptions are working papers. for example those by Batista and Zalduendo (2004) and Ianchovichina and Kacker (2005).
My strategy is to combine the information from historical country experiences (e. this would imply a centuries-long divergence of income levels. a drop in energy prices or a change of government may all lead to a signiﬁcant acceleration in GDP growth for several years. When assessing the theoretical growth literature. g. short-term monetary and ﬁscal shocks). A change of government may have a short-run conﬁdence-boosting eﬀect which already anticipates measures that will have a visible impact on drivers of growth in the long run. My analysis focuses on these long-term eﬀects of policy changes. Judgments. The neoclassical model contributes the importance of diminishing returns to factors. there is less and less art in economics because modern applied policy models must be speciﬁed in a way that can be directly empirically tested. There is no single best way to conduct empirical analysis in the social sciences. Landes ) and careful econometric analysis to build a model that uses variables that are as exogenous and as causal for economic growth as possible.1. g. This relegates many important developments to the sidelines. i. settler mortality. which does not appear to be observed outside Africa.e. in the countries considered here. religion. This is in line with the view of Brock and Durlauf (2001). Furthermore. chapter 2 will gather the most useful elements from diﬀerent theories. it is crucial to understand the advantages and disadvantages of the diﬀerent approaches in order to ﬁnd the most suitable framework.3 Plan of work 7 A crucial challenge in model building is to distinguish between correlation and causality. In this study they will be excluded from the analysis because they are either highly unpredictable (e.g. being landlocked etc. The approach in this study is what Colander (2000. e. . i. However. However. For example.” Colander also quotes Keynes as deﬁning the task at hand: ”Economics is the science of thinking in terms of models joined by the art of choosing models which are relevant to the contemporary world. 137) calls modern economics or the economics of the model. but not all of them have a causal link. assumptions and compromises will have to be made. Since the focus is on growth rates rather than levels of GDP.2 This signiﬁcantly reduces the realm of possible theoretical models. ”the study of the economy and economic policies through empirically testable models. I will leave aside constant factors that aﬀect mainly the level of economic activity such as climate. a colonial past. an unusually expansionary monetary or ﬁscal policy. Hausmann.” However. Many variables are likely to be correlated with economic growth. p. Pritchett and Rodrik (2005) ﬁnd that events like these explain most of the accelerations of GDP growth over time. New growth models 2 Parts of the literature claim that these factors also permanently aﬀect the growth rates of GDP. who believe that historical and qualitative studies play a crucial role in the development of credible statistical analyses. exchange rate moves) or because they will eventually be followed by a reversal of policy (e. Indeed. the focus has to be on developments that are reasonably predictable. a depreciation of the exchange rate.
but initial attempts did not take into account the non-stationarity of the underlying data. On the other hand.” What holds for cross-country models applies equally to time-series and panel models: Forecast performance may be an important indication for a model’s validity. in line with Romer (1994c. We can thereby enshrine the economic orthodoxy and make it invulnerable to challenge. taking into account the information gained in the ﬁrst stage. 1. Chapter 12 will present the estimation results.. p. chapter 11 includes an evaluation of the diﬀerent empirical growth models and assess their strengths and weaknesses. but for periods beyond the usual samples. Panel models are more appropriate. Chapters 4 to 9 look at some of the most important drivers of growth in more detail. Each chapter focuses on the theoretical rationale and the best available data for measuring each driver.even though these models are still widely used today. 202) rightly points out that “the litmus test for the cross-country growth literature will come when we ﬁnd out how useful our current models are in predicting the variation in growth rates. In general. This implies that empirical analysis has to help with selecting an appropriate model. we will indeed end up with too little data. I will propose a two-stage approach. . Evolutionary models emphasize the importance of complementarities. which ﬁrst analyzes the long-run linkages between the levels of the key variables (panel cointegration).” Temple (2000.3.2 Choosing the best econometric technique The survey of growth theories will show that no consensus is available on how to best model long-term economic growth. Chapter 2 also explains why the production function should be used with considerable care in growth models and why many endogenous models with their scale eﬀects are not helpful for building a model of the real world. my approach will be rather pragmatic.8 1 The importance of long-run growth analysis add human capital and barriers to technology adoption. p. The second stage is the modeling of growth rates of GDP. the theory of forecasting sketched in chapter 10 shows that a sound theoretical basis is not a necessary condition for good forecasts. chapter 13 conducts two forecast competitions and presents forecasts for GDP growth over 2006 to 2020. Cross-country regressions will be dismissed as not ﬂexible enough for modeling the complex process of long-run economic growth . 2001 add-on) suggests that an ”alternative strategy might be to begin with unprejudiced empirical study of the determinants of the speed of technological innovation. 20): “If we set our standards for what constitutes relevant evidence too high and pose out tests too narrowly. My synthesis builds on Kaldor’s technical progress function and augments it with insights from the other models.. Finally. Solow (1987.”. With these diﬃculties in mind. not for existing data.
So far. This chapter cannot possibly summarize the vast body of theoretical growth models. Earlier he made clear that the ”art of successful theorizing is to make the inevitable simplifying assumptions in such a way that the ﬁnal results are not very sensitive” (Solow ). most importantly. Textbooks such as Obstfeld and Rogoﬀ (1996).” These are the standards against which I will evaluate the theoretical literature and build my own model.2 Assessment of growth theories Long-run economic growth is a highly complicated process that involves the decisions and the complex interaction of a large number of economic agents over a long time horizon. availability of data to test the theory and. The ﬁlters used in the chapter are usefulness for forecasting. real world validity of the theoretical conclusions drawn. This chapter only gives an overview of the diﬀerent approaches. p. At the end. 2. . Frenkel and Hemmer (1999). The detailed treatment of individual drivers of growth is left to the subsequent chapters. its causal arrows should rest on some sort of behavioral mechanism. this chapter presents a model that takes most of these lessons on board but is also tractable empirically. It also summarizes the important lessons to be learned from each model and the pitfalls to be avoided. Rather. 383) always thought “of growth theory as the search for a dynamic model that could explain the evolution of an economy over time” or equivalently “the theory of the evolution of potential output”(p. Therefore. 286).1 The search for a dynamic model Robert Solow (2001. this chapter cannot possibly present the ”true” model of growth. p. 383) also demands that “an economic model should have some internal structure. Aghion and Howitt (1998). Solow (2001. Jones (2002a) or Barro and Salai-Martin (2004) provide comprehensive overviews of neoclassical and endogenous models with all the mathematics behind them. it aims at highlighting the strengths and weaknesses of diﬀerent models. modeling this process has remained incomplete and a major challenge for economists.
p. The details of the model have been outlined competently and in detail in many places (e. I will use the same deﬁnition as Lucas (1988. a causal relationship between inequality and growth has no implications for whether there exists a causal relationship between trade policy and growth. two factors of production . p. This heterogeneity is helpful under the surface of macroeconomic models because it moves the economy forward: loss-making ﬁrms will shrink.10 2 Assessment of growth theories Also.but it remains hard to exactly deﬁne and explain changes in technology. Solow’s intention was rather narrow and straightforward: He wanted to explain the growth performance of a developed economy like the US and show that decreasing returns to capital imply that accumulation of physical capital alone cannot explain the long-run growth performance of an economy. g. Countries cannot get rich by simply accumulating more and more physical capital. any search for a useful growth theory has to start there. A severe problem in trying to explain the path of per capita GDP is that growth theories are “open-ended”. so a short summary should suﬃce. macroeconomic modeling is still an art. a constant returns to scale technology given exogenously. This chapter will also brieﬂy evaluate which elements of post-Keynesian models. However. Schumpeterian models. Keeping the model and the resulting empirical tests reasonably simple is also a goal. Brock and Durlauf (2001. Jones [2002a]).” Any macroeconomic model also has to abstract from the fact of reality that ﬁrms and households are highly heterogeneous: some ﬁrms make good proﬁts. ) remains the workhorse of growth empirics and the benchmark against which all other models are still compared. Therefore. it cannot be part of a long-run growth model for forecasting purposes. This implies that I will have to abstract from some potentially relevant aspects of the real world. 235) deﬁne openendedness as the validity of one causal theory of growth not implying the falsity of another: “So for example.” And I will try to avoid constructing a model that produces “logically possible outcomes that bear no resemblance to the outcomes produced by actual economic systems” (ibid). while others lose money. a constant and exogenous savings rate s. avoiding “kitchen-sink” models that have become so widespread in growth empirics. 5): A theory is “an explicit dynamic system. Decreasing and ultimately zero (or negative) returns will prevent this. One focus of this chapter will therefore be on modeling knowledge and the barriers to technology. evolutionary models and open-system models provide meaningful insights. while proﬁtable ﬁrms will grow. 2. The basic Solow-model is for a closed economy with only one good Y . something that can be put on a computer and run.2 The basic neoclassical model The neoclassical Solow model (Solow . To a large extent. All models agree that technology is a key factor shaping economic growth . no government involvement.
Capital is in any model accumulated according to: (2. we can derive the level of per capita GDP yt (lower case variables will in general denote per capita amounts: yt = Yt /Lt ) in the steady state as:2 (2.4) ˙ Kt = Kt − Kt−1 = It − δKt−1 where It is gross investment in period t and δ the rate of depreciation of the capital stock.3) ˆ ˆ Yt = g + αKt + (1 − α)n The steady state (or balanced growth path) is deﬁned as the situation in which output Y and the capital stock grow at the same rate.2. The neoclassical Cobb-Douglas production function with Hicksneutral technological progress is:1 (2. p. The labor force L grows at the exogenous rate n. Lt = L0 ent .e. and technology advances at the constant and exogenous rate g. Taking logs and then diﬀerentiating with respect to time leads to: (2. where the hat above a variable denotes percentage changes. α When using Harrod-type technological progress (Yt = Kt (At Lt )(1−α) ). .1) α Yt = At Kt Lt (1−α) where the subscript t denotes time and α the elasticity of output with respect to capital input. 40. and a level of technology (or total factor productivity TFP) A.2) ˆ ˆ ˆ ˆ Yt = At + αKt + (1 − α)Lt . so that the capitaloutput ratio K/Y remains constant. technological progress is often assumed to be of the Harrod-neutral.2 can be written as: (2. while a dot above a variable (used below) will denote absolute changes. and combining the above equations with the insight that capital per eﬀective worker does not change in the steady state. per unit of eﬀective workers At Lt ). but negatively on population growth. I will show later that this assumption (a Kaldor stylized fact) actually holds in reality and will use it in the theoretical model as well. 1 2 Below. expressing all variables in intensive form (i.2 The basic neoclassical model 11 capital K and labor L. laborα augmenting type Yt = Kt (At Lt )(1−α) See for example Jones (2002a).5) yt = At s n+g+δ α (1−α) The level of GDP depends positively on the level of technology At and the savings rate. At = A0 egt Therefore equation 2. the rate of technological progress and the depreciation rate.
the three variables y. This insight has to carry through into any growth model and I will make use of it as well. investment is the accommodating variable to guarantee an equilibrium. Countries with high savings rates and low rates of population growth are seen to have a higher steady state level of GDP per capita. the model leaves the main driver of economic growth. there should therefore be two cointegrating relationships as will be explained in detail below. He also pointed out that he never used his model for developing economies because the underlying mechanisms in those countries . However. In the Solow model. partly because of problems of omitted variables and reverse causality. i.even if he left it to future researchers to model that progress. Romer and Weil ). Capital reacts endogenously to the advance in technology and therefore to the change in capital’s productivity.6) ˆ ˆ yt = k t = At = g ˆ In econometric terminology.in sharp contrast to the knife-edged models of Harrod and Domar proposed in the 1930s and 40s. the growth rates of GDP per capita and of the capital stock per capita in the steady state are independent of the savings rate and depend only on the pace of technological progress g. p.2. One of these uses is cross-country applications (prominently in Mankiw. (2. 283) has “been skeptical from the beginning about the interpretation of crosscountry growth regressions”. Solow himself (2001. Solow made clear that sustained economic growth depends on technological progress . There is no role for entrepreneurs. The Solow model has received a wide following because it generates a stable equilibrium . p. 2.e. “all the action was not in capital accumulation.” Technological progress is needed for continuing accumulation of physical capital. see chapter 6) are all driven by a single shock g.12 2 Assessment of growth theories Importantly. g.1 Application in cross-country analysis Although intended for a narrow question. Solow’s model was also used for many purposes that it was not intended for. Or as Frankel (2003. 189) puts it. while technological progress prevents the marginal product of capital from falling too much and thus keeps capital accumulation going. Among the three integrated variables y. unexplained. Since real-world economies tend to move in a rather smooth way over longer periods of time (exception: the Great Depression). Unfortunately. k and h. a theoretical model of long-run growth should be able to lead to stable equilibria. this stability is possible because of the assumption that investment is always equal to savings. So any criticism linked to the Solow-model is mostly a criticism of the (mis-)use of the model and not of the contribution of the Nobel laureate himself. k and A (and h in an augmented model. but rather in the residual. Decreasing returns make the model mathematically tractable. animal spirits or expectations about the future to aﬀect investment and growth.
This approach is still the workhorse of empirical growth analyses today. Applying natural logarithms to equation 2.7) ln yi = ln B0 + α α ln si − ln (ni + 0. I will show in chapter 11 that most of the assumptions used to derive equation 2.05) + (1 − α) (1 − α) i The level of per capita income of country i depends only on that country’s savings rate si . but this approach will not be used in my empirical model. Diﬀerences in the level of technology or productivity are not explained but relegated to the random error. 2.crucially .2. however.but the US investment ratio is lower than that of Thailand and population growth is higher.that the initial level of technology A0 is equal to a common constant B0 in all countries plus a random. all economies would grow at the same rate in the longrun.3 Focus on convergence 13 are quite diﬀerent. enter as one of the contenders in the forecast competition in chapter 13. Diﬀerences in observed growth rates across countries would then stem from diﬀerent distances to their own steady states. Romer and Weil (1992) leads to the basic empirical equation: (2. This is the notion of conditional convergence where economies approach their own country-speciﬁc steady state path over 3 The assumption that all countries grow at the same rate will. For example. However. In addition.3 With the main driver of long-term growth outside the standard neoclassical model. . As a result. country-speciﬁc shock . Researchers in the crosssection literature make the assumptions that g is constant and identical across countries (i. Since growth rates diﬀer signiﬁcantly in reality. this is exactly what much of the empirical growth literature has done over the past 15 years. that the coeﬃcient α in the production function is identical across countries.5 and assuming that g + δ = 2% + 3% = 5% as in Mankiw. complete knowledge spillovers).7 do not hold in reality.” Therefore. and . its population growth rate ni and a random error. GDP per capita in the USA is more than 15 times the level in Thailand .3 Focus on convergence The practical use of the basic neoclassical model is severely limited by the fact that long-run growth g is assumed to be exogenous and identical across countries. saying that “the neoclassical model is not a theory of development. Prescott (1998) is very explicit. e. the empirical growth literature quickly focused on the gap between where an economy should be in the long run and where it is today. this may not be the ideal starting point for long-run growth forecasts. the cross-country diﬀerences in savings and population growth rates are not nearly large enough to explain the variation in income per capita. awareness of the cross-section models is crucial for any assessment of the literature. that the depreciation rate δ is constant and identical across countries. Nevertheless.
1 illustrates conditional convergence: the steady state levels of income advance at the same constant rate in both economies. the trajectory of actual GDP in economy 1 is much steeper than that of economy 2: Economy 1 will post stronger growth rates as it converges to its own steady state despite the common pace of technological progress.3 Steady state path of economy 1 2.7 Fig. leading to the parallel dotted lines. Below the steady state an economy has a low capital-labor ratio and a high marginal product of capital. 2.8) yt = ln yt − ln yt−1 = λ(ln y ∗ − ln yt−1 ) ˆ . The capital-labor ratio and output per capita both rise as they converge from below to their steady state paths.9 Time 1. However. Figure 2.5 2. the resulting growth rate of GDP for g = 0 is: (2. This leads to high capital accumulation. Assuming that 10% of a gap between actual and steady state GDP closes each period. Two economies converging to parallel steady-state paths The drivers behind this convergence process are diﬀerences in the marginal products of capital in a production function with diminishing marginal products.1 Convergence trajectory of economy 1 1. which then dampens the marginal product of capital on the way to equilibrium. while economy 2 is slightly above its steady state. If income per capita closes the gap between its steady state level y∗ and last year’s level at the rate of λ . economy 1 starts out well below its steady state path.14 2 Assessment of growth theories time.7 Steady state path of economy 2 Convergence trajectory of economy 2 2. 2.5 above.9 ln GDP 2. The level of that path is determined by savings rates and population growth rates as shown in equation 2.1.
However. this speciﬁcation is sometimes interpreted as cointegration between ln yt−1 . s and n. Furthermore. Romer and Weil (1992. As I will explain in detail in chapter 11. 2. neoclassical authors claim that a 2% convergence rate is like an “iron law. say 20% of the gap closes every year. then the idea of conditional convergence is of minor use for long-run growth analysis. On the other hand. but not for growth forecasts over a 10 to 15-year horizon.2. However.1 Tests for conditional convergence Given that convergence to a country’s steady state income level is the only thing that diﬀerentiates between countries’ actual growth experiences in the neoclassical framework. Wacziarg (2002. Therefore. A signiﬁcant coeﬃcient on yt−1 is interpreted as evidence for conditional convergence. more appropriate techniques ﬁnd convergence speeds of 10% (Caselli. 375) point out that the whole idea of conditional convergence has little economic meaning because this common growth rate does not exist (the parallel dotted lines in ﬁgure 2. Pesaran and Smith ). as Caselli et al. the empirical result of a 2% convergence rate stems from the problematic cross-section regressions.3 Focus on convergence 15 Replacing the steady state income level with its determinants βx leads to a simple convergence regression like (2. They also ﬁnd that the rate of technological progress has been higher in OECD countries over the period 1960 to 1989 with a smaller dispersion as compared to the world as a whole.” Mankiw. Esquivel and Lefort ) or even 30% (Lee. 910) thinks that “With thirty or so years of data. convergence studies also suﬀer from sample selection bias (only successful countries provide reliable data on GDP growth) and measurement error (poor countries have a larger informal economy). In general. but the literature hardly ever formally tests for either non-stationarity or cointegration. If it is very fast.” These high speeds of adjustment make convergence an important notion for business cycle analysis (the output gap closes). “economies spend most of their time in the neighborhood of the steady state. Pesaran and Smith (1997.3. p. . 423) even derive this 2% from the supposed parameters of the economy. In addition. Even more.9) yt = −λ ln yt−1 + λ βx ˆ This speciﬁcation is used for example by Mankiw.1). interpreting the coeﬃcients on the exogenous variables x may turn hazardous. as they are often erroneously interpreted as measuring the unconditional impact of the variables on the growth rate of GDP. conditional convergence assumes a common steady state growth rate. p. put it. it is understandable that a heated debate developed over how quickly that convergence proceeds. it is simply impossible to distinguish steady-state growth from transitional growth” given estimates of a half-life of transition of around 32 years. However. p. Lee. Romer and Weil (1992) or in forecasting by Ianchovichina and Kacker (2005).
This is clearly not the case. there may still be some value in looking at the distance to the leader. Still. any catching-up is likely to be associated with policy decisions and changes in the determinants of economic growth. the main task of growth theory and empirics should be to explain the long-run path of an economy and not a short-run convergence to that path. e. . the prediction following from absolute convergence is that all countries should achieve the same level of income per capita over time . convergence to a country-speciﬁc growth path was the focus of a large body of literature and subject to a heated debate that is still ongoing. A useful model has to go signiﬁcantly further. However. as argued above. For example. The error is being corrected. It is intimately related to the idea of cointegration. Likewise. Hausmann. He sees no sign of absolute convergence. Nevertheless. The real issue is not how quickly an economy approaches its long-run growth path. As this brief review illustrates. Likewise. not convergence.4 Models with deeper insights A forecasting model would be highly unsatisfactory if growth was either predetermined outside the model or only stemmed from the convergence to the steady state. but has to remain tractable mathematically. the simple prediction of absolute convergence is not supported empirically. Pritchett and Rodrick (2005) show that growth accelerations are more likely in poor countries than in rich countries.16 2 Assessment of growth theories A diﬀerent issue is whether countries’ GDP per capita converge to that of the world leader. However. a stable steady state with low levels of per capita income possibly resulting from inferior production technology.” Similarly. any deviation from a linear relationship between the two tends to get smaller over time. Quah (1993) ﬁnds a tendency towards a twocamp world. Parente and Prescott (2000) think that any useful theory of economic development has to be able to explain the catching-up of countries like Japan and South Korea in the second half of the 20th century. Rather. The theoretical reasoning uses the idea of a “poverty trap” i. Growth does not come about automatically but is a complicated process and requires hard work. The main focus of this study is on that long-run path. big time. divided between haves and have-nots. 2.possibly by now. However. is the big story. and the two variables converge to the common path. the idea of convergence will feature prominently in my empirical model. Easterly and Levine (2001) analyze economies over the period 1980 to 1992 and conclude that divergence. The leaders may set examples that followers may decide to apply at home. Therefore. If two variables tend to move together over time. but the trajectory of the longrun path itself. This so-called “absolute convergence” would be extremely helpful for growth forecasting: the distance to the leader ”distance to frontier”) that would explain growth is reasonably easy to calculate. Pritchett (1997) sees “divergence.
This model provides much richer insights than the Solow model because there are now several parameters that policy may be able to inﬂuence. but produces a more complicated formula for long-term growth. Lucas’ assumption of exactly constant returns to human capital in the production function for human capital is a highly restrictive one . And all key variables grow at the same pace per capita in the steady state: (2.4.11) yt = ˆ 1 (B − δ − ρ) σ This growth rate depends on the underlying characteristics of the economy: The rate of risk aversion σ. . this so-called “augmented Solow model” does not add any signiﬁcant further insights.2. 2 See for example Frenkel and Hemmer (1999). 209ﬀ. In Lucas’ model with intertemporally optimizing households but without spillovers of human capital. Still. thereby avoiding the reason why Solow had to use g. p.1 Including human capital (Lucas) A big step forward in growth theory came with Lucas (1988) adding a second sector to the economy that produces new human capital H by using a fraction u of the existing human capital in a constant returns technology. It eﬀectively assumes that human capital and knowledge are one and the same. However. this leads to a long-run growth rate of per capita GDP of:5 (2. the rate of depreciation of (human) capital δ and the rate of time preference ρ determine the share of human capital devoted to teaching u.12) 1 ˆ ˆ yt = kt = ht = (B − δ − ρ) ˆ σ This result stems from the assumption of a constant returns production function for human capital.small deviations will make the model look starkly diﬀerent. The model maintains perfect competition. moving from a variable that drops like manna from heaven to a variable that has more to do with economic policy is very important and I will use much of this thinking below. p. 4 5 See Aghion and Howitt (2005b).4 Models with deeper insights 17 A very simple way of expanding the Solow model adds a third factor of production. However.4 The law of motion for human capital in per capita terms is: (2. human capital (see also chapter 6 on human capital). The long-run growth rate of all per capita variables is still determined outside the model by g. 2. In the words of Solow (1994. the long-run growth rate is just as constant and exogenous as it is in the Solow model.10) ˙ ht = ht − ht−1 = B(1 − u)ht − δht−1 ˙ where ht denotes the absolute change of human capital per capita. the technology of teaching B.
bribes or severance packages to redundant workers. “this version of the endogenous-growth model is very un-robust. capture work rules that require ﬁrms to use a minimum number of workers per machine or per plant. it is impossible to combine free trade and barriers to technology at the same time: If rules prevent domestic ﬁrms from using the best available technology. The theory of relative eﬃciencies (or of diﬀerences in total factor productivity TFP) presented in Parente and Prescott (2000) decomposes a country’s technology into a pure knowledge component At common to all countries and an eﬃciency component Ei speciﬁc to each country i. . In principle. The Parente and Prescott model suggests that the potential for rapid growth is larger the further a country is behind the technological leader. The production function as outlined in Parente and Prescott (2005) therefore extends the standard neoclassical production function: (2.” Also.18 2 Assessment of growth theories p.13) α Yit = Ei At Kit Lit (1−α) This function gives the “maximum output that can be produced given not only the technology constraints but also the constraints on the use of technologies arising from policies” (Parente and Prescott . The diﬃculty with the Parente/Prescott approach is that eﬃciency is difﬁcult to measure. 51).4. 1994. Opening an economy to free trade is only sensible if barriers to technology are reduced at the same time or earlier. Or it could capture the cost of switching to a new technology in the form of regulations. for example. This Ei can also be interpreted as a measure for the quality of a country’s institutions. It is straightforward to allow Ei to vary over time as well. The eﬃciency component could. However. But then it will not be able to make use of all the technology that is incorporated in imports or conveyed in the exchange with foreigners. a country can reduce its domestic barriers without opening to free trade. They propose to model the barriers preventing a country from using the best globally available technology as a way to explaining diﬀerences in income levels as well as some growth miracles of the 20th century.2 Modeling barriers to riches (Parente & Prescott) Another useful model that has its roots in neoclassical thinking but tries to capture more aspects of reality is the ”barriers to riches” approach by Parente and Prescott (e. 82). then foreign ﬁrms would be able to capture the market if they were allowed to enter. 2000 and 2005). g. p. which measures the degree to which that country exploits the usable knowledge that is generally available. the Lucas model has no role for active individuals like entrepreneurs and does not include explicit research and development (R&D) activity. 2. The conclusion is that low barriers (good institutions) and free trade tend to go hand in hand.
as research and development: “Long-run growth is driven primarily by the accumulation of knowledge by forward-looking. 51) applauds the attempt to “model the endogenous component of technological change as an integral part of the theory of economic growth. proﬁt-maximizing agents. 3). It also implies that careful modeling of any changes in barriers is particularly important when trying to model GDP growth in emerging markets. e. 287) put it: “The good thing about the fox that Paul Romer started chasing more than 15 years 6 Romer (1986). Therefore. so that it does not depend on the exogenous rate of technological progress g. who tend not to engage much in R&D.5 Opening the theories further 19 i. that only a small subset of all possible goods will ever be introduced” (p. services and ideas. Since I want to model the growth process of a large number of countries with extremely diﬀerent income levels. 19). Raising the pace of knowledge acquisition is likely to raise the speed of GDP growth.” Another term often used is “new growth theory” to distinguish it from the “old” neoclassical models that treated technology as given. trying to avoid the Solow model’s feature that accumulation of capital alone cannot make a society rich. The other strand of the endogenous growth theory focuses on the creation of ideas and their spillovers. Societies with institutions that allow ideas to be generated or applied more quickly than elsewhere tend to grow faster or to be richer. He criticizes general equilibrium theory for operating within a ﬁxed set of goods: “It is an inevitable fact of life that economies will forever operate on the boundary of goods space. 2. Technological progress is modeled as a purposeful economic activity. It drops the neoclassical assumption that the same technology is available in each country and recognizes that “technological advance comes from things that people do” (ibid. p.2. p. p. 12). any useful model of economic growth will have to try to model the creation of new goods. This is an important insight that I will use when including human capital and trade openness in the forecasting model. the lower its Ei . Or as Solow (2001.5 Opening the theories further The Lucas model is one example of how to circumvent the inability of the Solow model to explain the evolution of long-run growth by relaxing the assumption of diminishing returns. this idea is very appealing. p.”6 This strand of the literature tries “to uncover the private and public sector choices that cause the rate of growth of the residual to vary across countries” (Romer [1994c]. According to Romer (1994a) new growth theory is mainly about the creation of new goods. 1003. . Solow (1994. This is consistent with the observation that growth miracles are more likely to happen among poor than among rich countries. This strand of endogenous growth theory tries to explain growth endogenously inside the model. p.
endogenous growth models oﬀer some valuable insights into the evolution of growth and knowledge in leading developed countries. They can grow quickly simply by adopting technologies available elsewhere. 2. This is how the Japanese in the 1970s. Some explicitly used ”creative destruction” in the title like Aghion and Howitt (1992).1 Models with scale eﬀects The 1990s saw a revival of Schumpeter’s ideas with many contributions to the endogenous growth theory.5. However. The modeling of purposeful R&D activity cannot help explain these success stories. 2005]. As Parente (2001) makes clear. any economy is full of abrupt changes as new ideas are implemented in the quest for monopolistic proﬁts: new ﬁrms drive out old ﬁrms and improve technology .” Large countries or countries with fast population growth do not necessarily show faster growth of GDP per capita than smaller countries. which are then forced to be innovative themselves. In his view the evolution of an economy is linked to individuals doing new things or doing things diﬀerently. Likewise. poor countries do not engage in R&D because they don’t have to. According to Schumpeter. Together with the paucity of reliable data. they do not help to explain the huge cross-country diﬀerences in incomes.the famous “creative destruction”. 2002b. Radical innovations may be introduced and then diﬀuse into the general economy with a series of incremental innovations.” Modeling the creation of new ideas and knowledge is in line with the thinking of Joseph Schumpeter. However. who saw capitalism as an engine of progressive change. the Koreans in the 1980s and the Chinese in the 1990s achieved much of their stellar growth rates. this is enough reason to not include R&D in my model. The growth rate of per capita GDP depends positively either on the size of the population (or of human capital. As a result. people lose their jobs and have to ﬁnd new things to do.20 2 Assessment of growth theories ago is that it leads us to focus on the analysis of the economic incentives to create new technology. Segerstrom . all models with spillovers lead to so-called “scale eﬀects” as Jones (1999) outlines in a detailed and structured way. Howitt  and Aghion and Howitt [2005a]). It is possible for innovations to be clustered in time partly because one innovation may disrupt the business model of other ﬁrms. This would be compatible with the conditions for a growth model outlined in chapter 1. Firms go bankrupt. which is supposed to capture the performance of nearly 40 highly heterogeneous countries. Grossman and Helpman . Bottazzi and Peri (2007) see ”a clear rejection of the existence of a strong scale eﬀect. see Romer . Rudd (2000) ﬁnds little . economic growth is not a balanced process where all sectors of an economy expand in the same way every year. Young  and Aghion and Howitt ) or on the growth rate of the population (see Jones [1995b. However. In general. Parente (2001) highlights that “the prediction known as scale eﬀect is not born out by the data”. Similarly.
The benchmark book by Nelson and Winter (1982) summarizes their earlier contributions that focus on capitalism as an engine of change. Fagerberg (2002) and Andersen (2004). the activities of entrepreneurs are very important in changing the economic and social landscape. The smarter these workers. However. some researchers classify the models just sketched as “new neoclassical” or as Schumpeterian. the higher the quality of their proposals and the higher is the level of productivity. 2002).5. managers and customers are. even before the contributions of Romer and others. Theses models explicitly think of growth as a process that takes place in historical time. But any analysis of the role of entrepreneurship in growth such as in Beaugrand (2004) quickly turns to the environment within which the entrepreneur operates: Will he be able to hire enough qualiﬁed workers? Will he be able to reap the rewards from his activities? 2.” Likewise. Fagerberg highlights the diﬃculties of being able to publish in mainstream .2 Evolutionary models of growth The growth models sketched in the previous section borrow heavily from Schumpeter. this feature makes it diﬃcult to create realistic models that do not show explosive growth paths or scale eﬀects. their qualitative transformation in historical time. Acemoglu and Angrist (2000) are unable to ﬁnd external returns to education that diﬀer signiﬁcantly from zero. p.” Classifying the diﬀerent models remains hazardous. 53) because there are many cases where the improvement of a product or the cost reduction stems from the input from workers.5 Opening the theories further 21 evidence of an external eﬀect of human capital and concludes that “human capital spillovers of the form postulated in the new growth literature are unlikely to matter much in practice. evolutionary economists (“neo-Schumpeterians”) started to model the evolution of economies. Furthermore.8 7 8 The empirical support that Jones (2005) derives from Alcal´ and Ciccone (2004) a is highly problematic (see chapter 7). However. R&D expenditures clearly are an “inadequate measure of the resources devoted to increasing productivity” (Solow 1994.7 In sum. Partly for good reasons. The focus on knowledge is in line with the observation that rich countries tend to be those with the highest level of knowledge.2. The feature that knowledge is nonrivalrous and leads to increasing returns is on the one hand important to drive the endogenous models. who tried to combine the dynamic thinking of Marx with the historical school and with neoclassical microeconomic foundations. Parente (2001) concludes that “endogenous growth theory is not a reasonable theory of economic development. the evidence is not strong enough to make models based on spillovers useful vehicles for long-run growth analysis. Outlines can be found in Verspagen (2000. management or customers. Certainly.
They adopt new routines and output rises back to the production possibility frontier possibly after temporary disruptions during the adjustment process may have depressed it further. Unfortunately. e.these are not the characteristics of the economic society in which we actually live. 31) highlight that their focus is on “how individual skills. these routines are the equivalent to genes in biology. However. Usually. i. and disequilibrium notions to model how economies evolve. 11) stresses that the essence of human cultural evolution is that “knowledge is accumulating in the heads of human beings. Nelson (2005. economic selection. Neoclassical models (including endogenous models) use representative agents with complete foresight who pick the best technology from a freely accessible set of technologies depending on prevailing prices and who always keep the economies in equilibrium. advanced technologies and modern institutions come into being” in a “trial-and-error cumulative learning partly by individuals. p. Romer or Alwyn Young. their continuous change from a lower or simpler state to a higher or more complex state. Output would stay below a maximum possible value until companies for some reason decide that things have to change. organizational routines. as a look at real-world companies and markets shows.9 This section attempts to sketch this link and to see to what extent diﬀerent theoretical models satisfy requirements from evolutionary theory. But . They are constrained by the limits of what they know and by old habits and routines.to borrow from Keynes .22 2 Assessment of growth theories Nelson and Winter (2002. p. this makes formal modeling rather diﬃcult. the pace of acquiring new routines probably depends on the pace at which human capital grows and on the pace at which a country opens to routines from abroad. partly by organizations. It is possible that companies hang on to old routines for a very long time despite new possibilities becoming available. Lucas. they develop and adopt new routines when the old ones no longer work. strateUS journals and concludes that “evolutionary modelling does not appear to have been accepted as a welcome addition to the discipline by hardcore mainstream economics” (p. Firms act under bounded rationality and use rules of thumb rather than fully rational proﬁt maximization. Evolutionary models try to move closer to the real world by using bounded rationality. 37).” Evolutionary models use biological notions such as natural selection and genetic mutation. In evolutionary economics.” Unfortunately. heterogeneity of agents. partly by society as a whole. The evolutionary literature usually does not link its insights to models like those of Lucas or Parente and Prescott. the insight that the process of growth should be analyzed at the company or at least the sectoral level is something I cannot use explicitly in my macroeconomic model. The counterpart to natural selection in biology is economic selection among the heterogeneous actors: ﬁrms with better products.” 9 . Beaugrand (2004) notes that the profession keeps cultivating its diﬀerences: “articles in the Journal of Evolutionary Economics rarely quote papers from Aghion and Howitt.
a link to the Parente and Prescott framework. Nelson (1981. However. page 42 in reprint) argues “it may be fruitful to consider the several sources of growth as being like the inputs into a cake. There are strong interactions between technology. The more pressure ﬁrms experience to improve their technology. this close linkage among the factors of growth leads to the question of what ultimately determines their evolution. culture. The evidence presented by Lentz and Mortensen (2005) strongly supports the view that worker reallocation from shrinking to growing ﬁrms is an important source of productivity growth . evolutionary biology is also mostly about maximization and equilibrium . any measure that improves on a ﬁrm’s old ways of doing things in order to avoid destruction. lack of foresight or discernment) behavior.an aspect that macroeconomic models cannot capture. For example. however. which also includes imitation and learning. innovation is crucial to generate new varieties.” In other words. Nelson also refers to the “general features of the economic environment and of political and social institutions that support all three sources and the growth they promote”. it makes little sense to try to divide up the credit for growth. while biology assumes myopic (i. science. formal models usually have to resort to an exogenous rate of innovation.2. i. treating the factors as if they were not complements. institutions and the economy that should be taken into account. But will they reshape their labor and product markets in order to take full advantage of these changes? In the words of Paul Romer (1994b). In practice.e. the more productive they will be . However.in the techniques we use. which makes them look similar to the simple neoclassical model outlined earlier.5 Opening the theories further 23 gies and routines will grow and survive. growth accounting would oﬀer few useful insights. politics. e.” This includes allowing entry . While economic selection reduces variety. which points to the same research agenda as the neoclassical growth theory has turned to in recent years: what ultimately explains the exogenous rate of technological progress that shapes all the proximate drivers of economic growth? One useful insight from evolutionary theory is that we cannot really distinguish between economic and non-economic factors when trying to explain economic growth. at the moment we see technological progress and the rapid integration of the global economy exerting signiﬁcant pressure on institutions in many European countries.and progress .” He goes on to conclude that “where complementarity is important. “a government must create an environment that fosters change . Selection improves the average ﬁtness of the ﬁrms but has a negative aspect as well in line with Schumpeter’s creative destruction: ﬁrms with old technologies will shrink or disappear.just as neoclassical economics . Krugman (1996) sees economics and evolutionary biology as sister ﬁelds because they are both about the interaction of self-interested individuals.and less about myopia and dynamics as it tries to come up with comprehensible models. The main diﬀerence is that economics tends to assume intelligent behavior. Evolutionary theory uses a very broad concept of innovation. All are needed. Evolutionary models lead to conclusions that diﬀer signiﬁcantly from neoclassical models.
often at the expense of external validity or realisticness. history. conventions.24 2 Assessment of growth theories and exit of ﬁrms. The economic system is “embedded in and connected with politics. An open system is a system that interacts with the outside world. values. In their deﬁnition. this captures only the last step in their models. Or at least it has to model some of the symptoms of sclerosis such as a decline in labor usage in economies that have developed highly protective institutions. By contrast. but sometimes in ways which lead the system to evolve. evolutionary economics made use of methods developed in biometrics from the 1930s to start an evolutionary econometrics or evometrics. So keeping the necessity . social systems and behavior. a system is a network. a structure with connections. When evaluating their models. Change tends to be disruptive and societies in many developed countries have institutions that try to shield their citizens from these disruptions. Over the past decades. where to close the system. evolutionary economics models only some of the issues crucial to explain long-run economic growth. evolutionary economists make heavy use of data on patents and R&D. new technologies to be introduced? So far. all the elements of social life” (Chick 2003. philosophy. 2. often equated with a long-run result as in the neoclassical model.3 Open-system models Evolutionary models point in the same direction as the “open-system” approach advocated by Chick and Dow (2005). Ideally. open systems try to take history and initial conditions into account in the form of changing networks and institutions. This is easier said than done in practice. exogenous variables are especially helpful. 3). In most models one has to decide which variables are determined outside the model and have an impact on the system under investigation. A deeper question would be what allows and causes these new patents and the new R&D spending? What allows new ﬁrms to develop. This is similar to the notions of routines and innovations just highlighted. The argument for closed systems rests on internal consistency being the only test of rigor available in economics . Parente and Prescott). Economists often abolish time. e. i.5. However. Closure and openness are a matter of degree and the researcher has to choose where and how to close his system. p. Andersen (2004) gives an overview of the development of evometrics and illustrates a microeconomic example using selection eﬀects and innovation eﬀects at the ﬁrm level. a model of long-run growth should therefore try to take into account the evolution of change-inhibiting institutions (cf. for example by modeling convergence to an asymptote. This becomes an especially big issue when dealing with the time dimension. within which agents act. which is the end-point of the analysis. In forecasting. mostly in ways which reproduce and reinforce the system.
this chapter has shown how diﬃcult it is to construct a formal dynamic theoretical model that is able to endogenously explain economic growth and lead to reasonable predictions about the real world. 9). Going even further. This is consistent with the calls from so-called broadband economists such as Fullbrook (2004). However. which models the complex interaction of independent agents with spontaneous self-organization. positive feedbacks and complexity. The next modeling step would be chaos theory. (f) expand the set of methods and (g) conduct an interdisciplinary dialogue. 2.2.6. However. into account. Kibritcioglu and Dibooglu (2001) also propagate an interdisciplinary approach to long-run growth analysis. who suggest economics (a) use a broader concept of human behavior. The following chapters on the diﬀerent elements in my empirical model try to show this awareness. 417) posits that the marginal product of human capital “β is between one third and one . most theories of economic growth assume or use an aggregate production function and make statements about its coeﬃcients. (e) ground studies empirically. (d) develop a new theory of knowledge. p. (b) recognize culture as embedding economic activities.6 General critique of the standard approach 25 of an open system approach in mind is important in trying to build as realistic models as possible. 2. The following sections will show that they can produce misleading conclusions and should therefore be avoided. taking the insights from other disciplines like demographics.6 General critique of the standard approach So far. the situation gets even more complicated because of the widespread use of some vehicles that are supposed to increase our understanding of the growth process: the production function and total factor productivity. For example.1 Production function cannot be estimated As shown above. One has to be conscious of any assumptions made to create a workable model and in deﬁning the boundaries of the model. “Any number of arbitrarily small perturbations along the way could have made the world as we know it turn out very diﬀerently” (Romer 1994a. (c) consider history as economic reality. p. the simple Solow model would imply a marginal product of capital of one third. Paul Romer highlighted back in 1994 the possible chaotic behavior of economies. The augmented Solow model in Mankiw. culture and politics which is largely inherited from the past. Romer and Weil (1992. history etc. this is clearly too ambitious for this study. The conclusion from this overview of diﬀerent theoretical approaches is that economic growth is a highly complex process: many heterogeneous agents interact in an evolving system of science.
p. at = a = rt Kt /Yt .16) ˆ ˆ ˆ b Yt ≡ ˆt + aKt + (1 − a)Lt which is equivalent to the standard equation used in Solow accounting exercises with ˆt called “the residual”.7 to test whether the coeﬃcients of the production function really have the postulated size.18) a (1 − a)wt + aˆt ˆ a a r ln rt + ln s− ln n+δ+ (1 − a) (1 − a) (1 − a) 1−a . 373). The fact that the coeﬃcient on labor equals labor’s share in income should not be taken as support for any theory. The income identity is: (2. the wage sum plus the income from capital. although it is simply a transformation of the income identity using the assumption of constant factor shares. which states that the value of output (value added) always has to equal the value of inputs.” Tests for endogenous growth models tend to focus on whether α+β > 1 in the augmented model. which can be diﬀerentiated to lead ˆ ˆ r to gt = Bt = (1 − a)wt + aˆt . which is a stylized fact in macroeconomics. Empirical papers estimate an equation like 2.16 and taking antilogarithms leads to (2. e. as was outlined in Felipe and McCombie (2005) on which the following sections are based.7: ln y = b0 +ln wt + (2. The general ﬁnding is a high R2 and coeﬃcients that match theoretical priors. If we use the same deﬁnition for capital accumulation as in the Solow model outlined ˆ ˆ ˆ above (Kt = sYt /Kt − δ) and the steady-state assumption that Kt = Yt . Bt is called the ”dual” measure of productivity.26 2 Assessment of growth theories half.17) a Yt ≡ B 0 r t w t (1−a) (1−a) a K t Lt (1−a) a ≡ Bt Kt Lt (1−a) . However.14) Yt ≡ wt Lt + rt Kt . for example by Herbert Simon in his 1978 Nobel Prize lecture. Equation 2.15) ˆ ˆ ˆ ˆ Y ≡ art + aKt + (1 − a)wt + (1 − a)Lt ˆ r ˆ now deﬁne ˆt = aˆt + (1 − a)wt to get b (2. a where Bt = B0 rt wt (see Felipe and McCombie . b Integrating equation 2. we arrive at the accounting identity in growth form: (2. The simple reason is that these tests basically only estimate the national income identity. This point was made earlier several times. i. Totally diﬀerentiating the identity and assuming that factor shares are constant over time i.17 looks like a standard CobbDouglas production function. where wt is the real wage and rt is the real interest rate. e. these tests are meaningless and the high R2 is not at all surprising. we can derive an equation similar to 2.
the labor share in income is 0.25 and of capital of 0. So it should not come as a surprise that the human capital-augmented Solow model improves the ﬁt. All that has been done is to use the income accounting identity plus the assumptions of constant factor shares and of a constant capital-output ratio. But both approaches do not change the basic fact that an identity is estimated and that no insights can be gained about the coeﬃcients of the production function or the degree of returns to scale.18 are therefore not helpful when trying to test hypotheses about the Solow model or any other model with a production function. on whether markets are competitive and on whether factors are being paid their marginal products. then the data must necessarily always give a near perfect ﬁt to the model. the regression of output on labor and capital inputs generates a near perfect ﬁt and a coeﬃcient on the aggregate capital stock of 0. a measure of human capital is correlated with the real wage wt across economies. estimates of production functions today do not and cannot use physical quantities of capital. However. Given the focus of macroeconomic models on monopolistically competitive economies (New Keynesian theory) it seems reasonable to assume that at least one of the neoclassical assumptions is violated: factors are not being paid their marginal products. The ﬁrms set prices by a markup on unit labor costs worth 33%. To illustrate this point. Instead they have to rely on values which aggregate diﬀerent types of capital using the prices in the base period. it follows that rt = 0: the real interest rate does not change over ˆ time.2. consistent with the assumption of perfect competition in the factor markets. When using value data for the capital stock. these prices stem from the rate of proﬁts.18 looks like a modiﬁed version of the Mankiw. while the capital share is 0. However. Equation 2. The idea of a production function refers to real quantities of factor inputs such as the number of machines and the number of hours worked. Romer and Weil model.identical to the factor share but very diﬀerent from the true capital elasticity of 0.25. which is in practice derived as the residual from the national accounts identity 2.75 with a random error to avoid multicollinearity. which is driven by the real wage rate. Because of the mark-up. If these two assumptions are correct (since Kaldor  both seem to be widely accepted).6 General critique of the standard approach 27 Y Since by assumption a = constant. although no assumption has been made here on the degree of returns to scale. The underlying problem is the nature of the data used. Equations such as 2. The reason is that factors are not being paid their marginal product in the simulated economy as a result of the mark-ups. K/Y = constant.75. Felipe and McCombie (2006) simulate an economy consisting of 10 ﬁrms with constant returns to scale Cobb-Douglas production functions and output elasticities of labor of 0. using ﬁxed eﬀects panel techniques also should improve the ﬁt.75. Most likely. and r = a K = constant. . Augmenting the standard Solow model leads to even better estimated ﬁts for the identity if the additional variables are good proxies for Bt .25 . Likewise.14. Wages and the return on capital are the same in all ﬁrms.
Had this been a test for the null hypothesis of increasing returns (α + β > 1). The second problem lies in aggregating the coeﬃcients of production functions of diﬀerent ﬁrms. 2. L. Felipe and Fisher also emphasize that the usual aggregation cannot even be interpreted as an approximation of the true relationships. it does not have a sound theoretical foundation as was shown by Fisher (1969). there would still be the second aggregation problem.28 2 Assessment of growth theories In a second exercise Felipe and McCombie simulate a model with increasing returns to scale for all ﬁrms and a mark-up of 33%. Even if capital was physically homogeneous. Theories of growth cannot be tested in this way because they cannot be refuted. The production function cannot even be used as a parable as was suggested by the .” The problem is not as severe in aggregating labor input (common physical unit: hours of work) or human capital (common physical unit: years of education). The aggregation of ﬁrm production functions has been discussed since the 1940s. the hypothesis would have been erroneously rejected.g. ”the value of human capital in Spain is EUR 40. but that interest rate is usually determined by using the amount of capital in relation to output. p.6. Both approaches require an interest rate (discount rate). The estimated production function now has a higher intercept. all micro production functions have to be identical except for the coeﬃcient of capital eﬃciency. extreme care has to be taken if one wants to express these in value terms as well (e. At least since Wicksell it is well known that capital goods cannot be measured and aggregated in physical units because of their heterogeneity: how does one add up an airplane and a printing machine? Therefore valuation measures must be used. The circularity is clear. The production function takes on diﬀerent forms and purposes during the course of the studies. The survey by Felipe and Fisher (2003) lists several extremely restrictive conditions for aggregation to be valid: for example. The conclusion must be that it is impossible to estimate the coeﬃcients of an aggregate production function using standard macroeconomic data. labor L and other inputs X. I will try not use the notion of output elasticities below and will not try to estimate or interpret them. In any case. Kaldor (1975. retraced in Cohen and Harcourt (2003) and surveyed by Felipe and Fisher (2003). As a consequence. The value of a capital good can be the cost of its production or the value of the output that it will produce in the future. the quality of a forecasting model does not depend on estimated elasticities being equal to the factor shares. but the same coeﬃcients of 0. 348) noted that the diﬃculty of isolating or measuring the change in the quantity of capital ”makes it impossible to attribute to capital a marginal productivity of its own. However. However.75 on labor.000 billion”).2 Aggregate production function does not exist All students of economics use an aggregate production function from their ﬁrst term onwards. capital K. X) is the use of aggregates for output Y .25 on capital and 0. The ﬁrst problem of writing Y = f (K.
evaluating input changes at the factor shares leads to an estimated rate of TFP growth of only 1. 2. However. Factor shares depend on the relative bargaining power of labor and capital . it seems surprising that the aggregate production function and growth accounting exercises are still widely used in growth .5%.19) At = Yt − αKt − (1 − α)Lt Many studies use the ”Solow residual” or total factor productivity (TFP.3 The concept of TFP is not helpful Simple time diﬀerentials of the production function are still widely used as so-called Solow accounting exercises.4 Beyond neoclassical economics Given all these diﬃculties.6.2. growth accounting would attribute a third (equal to the share of capital in national income) of output growth to the changes in the capital stock even if those are induced by the change in A.15 to 2. 2. In other words. Their simulation uses a rate of technical progress at the ﬁrm level of 5% per annum for each ﬁrm. ˆ With constant population Lt = 0 (which is a relevant approximation for many ˆ ˆ ˆ OECD countries).they need not have any link to technology. At is by deﬁnition nothing but a weighted average of the wage rate and the interest rate.6 General critique of the standard approach 29 Massachusetts side of the Cambridge capital controversy. with a constant population all of the output gain should be attributed to TFP. it follows that Yt = Kt = At . ideas of growth accounting will not be used in this study and literature based on it will not receive a prominent role below. In levels this is: (2. Growth accounting assumes that factors are paid their marginal product. However. Attributing parts of it to capital does not improve our understanding of the growth process. long-run growth of per capita GDP in the neoclassical model depends only on the progress of technology. Indeed. using the reasoning from equations 2.6. At ) as a measure of technology and try to explain it. The capital stock only adjusts to keep the capital-output ratio and the real return on capital constant. The conclusion is that growth accounting exercises may lead to misleading estimates of actual technological change. However. Barro (1999) also highlights this fact. Together with the assumption of a production function Felipe and McCombie (2006) illustrate that the rate of technological progress as measured by TFP need not have any relation with the true underlying rate of technological progress. because that would require the same time pattern and capital-labor ratios for all goods: machines would have to be produced in the same way as consumption goods. Therefore. the construction of TFP is tautological: countries with high income per capita will have high TFP (and high levels of physical capital per capita). According to Felipe and McCombie (2007) what is usually labeled TFP growth is nothing but a ”measure of distributional changes”.17. As was shown above. There is not much to be learned here.
which they do. This holds in particular for tests of Ramsey’s AK model. It will also make use of the empirical insights outlined in later chapters. However. p.” Similarly. in our opinion. Kaldor (1957. Romer (1994c.their similarity across countries. some formal link between inputs and output is necessary to help structure the analysis. which is nothing but a test for whether output and capital grow at the same rates . p. 2001 add-on). help answer the central question of why some countries are richer than others. If the aggregate production function does not exist. 591) argued that the “purpose of a theory of economic growth is to show the nature of the non-economic variables which ultimately . the unrealistic predictions of some endogenous models with their scale eﬀects make them of limited use for forecasting real world economies.30 2 Assessment of growth theories analysis. This ﬁt provides no information about the existence let alone the shape of an aggregate production function. Although coming from a diﬀerent angle. With the available data it is impossible to assess the coeﬃcients of a production function. This takes us to Cambridge (England) economists such as Joan Robinson or Nicholas Kaldor. the kind of data [used] cannot be used to make you recant. The question rather is whether it exists at all. as pointed out above. Romer and Weil (1992) or Barro and Sala-i-Martin (2004) do. who have always criticized the use of aggregate production functions. As Felipe and McCombie (2005. The model presented in the next section will abide by the requirements derived above.in cross-country analysis . This is very diﬀerent from what the bulk of the empirical growth literature and prominent examples such as Mankiw. 2. The simple reason seems to be their good ﬁt to the data. thereby following Solow (1987. p. No dataset can refute the null hypothesis that the elasticities equal the factor shares and that there are constant returns to scale.7 The augmented Kaldor model From the discussion above it should be clear that a theoretical growth model should not make use of an aggregate production function of a speciﬁc shape or use marginal productivities in the reasoning. In sum. the neoclassical model is not the appropriate starting point for an analysis of long-run growth. then there is no point in speaking of a marginal product of aggregate capital or of aggregate labor. The crucial discussion about an aggregate production function is not whether it exhibits constant or increasing returns to scale. who suggests (as mentioned earlier) that an ”alternative strategy might be to begin with unprejudiced empirical study of the speed of technological innovation” and to take insights from other branches of economics and social sciences into account. 10) is also aware that “if you are committed to the neoclassical mode.” However. I will therefore not try to estimate coeﬃcients or make statements about the degree of returns to scale. this good regression ﬁt stems from the constancy of the factor shares in income over time and . 24) put it “this neoclassical framework does not.
And it links inputs to outputs. In other words. g. i. My augmented technical progress function is: 10 When moving to cross-country comparisons. The simple technical progress function of equation 2. In Kaldor’s model.20 is a compromise solution to the challenges outlined in the previous sections. one has to carefully analyze which coeﬃcients are assumed to be identical across the economies (e. 599). Kaldor thought it was impossible to distinguish between a rise in output induced by new technology and one induced by additional capital because a higher capital stock per worker must inevitably have been preceded by the introduction of a superior technology. Kaldor modeled a single economy. macroeconomic modeling would be impossible. combined with the willingness to invest capital (p. e.” Robert Solow would probably agree with this statement. . p. a constant capital-output ratio and a constant rate of proﬁt (real interest rate) over time. He used the stylized “Kaldor facts” of a constant labor share in income. 286). Both assume full use of the economy’s labor and capital input.2. but the disadvantage of assuming that aggregate capital exists. economic growth depends on the readiness of an economy to absorb technical change. labor shares.20) ˆ yt = c + akt ˆ Equation 2. So it is mostly concerned with the supply side of the macroeconomy” (2001. 593). capital-output ratios or proﬁt rates).20 can easily be augmented to include human capital h and a variable that measures openness o (capturing the Ei in the Parente/Prescott model from equation 2. Likewise. p. I will not make use of models that focus on the demand side of the economy or on balance of payment constraints such as in Thirlwall (2002) because these constraints are unlikely to be binding in the long run for the countries in the sample of this study. Models deviating from the full employment assumption have been developed by Post-Keynesians such as Robinson and Kalecki and recently reviewed by Stockhammer (1999). Just like Solow. but in a less restrictive way than a Cobb-Douglas production function because there is no link between the coeﬃcients and a marginal factor product. “Solow accounting” exercises would make no sense to Kaldor. not by eﬀective demand” (1957. output “is limited by available resources. who thinks of ”growth theory as precisely the theory of the evolution of potential output. Kaldor treated long-run GDP growth as exogenous.13).7 The augmented Kaldor model 31 determine the rate at which the general level of production of an economy is growing. Without a link between output and input and without assuming that a capital stock exists. It has the advantage of allowing some structured modeling. I will not follow them except for acknowledging the development of labor usage over time. This is also in line with Solow. Using per-capita variables his ”Technical Progress Function” is: (2.10 It is a model of the long run.
whereas openness partly measures the diﬀusion of existing knowledge across countries. This is in line with Prescott (1998.which are diﬃcult to measure anyway.21 does not use R&D separately.21) ˆ ˙ yt = akt + bht + (1 − a − b)ot ˆ ˙ Output grows at the weighted average rate of the percentage change of physical capital and the absolute changes of human capital and trade open˙ ness.just as in the Solow model . Physical capital plays a passive role in the Kaldor model. For emerging markets. one can now .23) y y ˙ yt = aˆt + (b/ω)ˆt + (1 − a − b)ot ˆ ˆ ˙ ˙ yt = kt = ω ht = µot .12 and in the evolutionary economics literature. This close link between the growth (percentage or absolute) of the main variables in this model mirrors the results from other models summarized in equations 2.22) or (2.21 includes the most important drivers of growth in line with the models sketched earlier in this chapter: Human capital (discussed in more detail in chapter 6) captures knowledge in a broad sense. p. Input this into 2. It adjusts to the intensity of innovation driven by openness and human capital. where At is the stock ˆ of total available scientiﬁc and technological ideas. ˆ (1−a−b) where µ = (1−a−b/ω) .32 2 Assessment of growth theories (2.21 to get ˆ (2. Trade openness (discussed in detail in chapter 7) and human capital are closely linked to the quality of institutions. though without any of the labels used here. Equation 2. A similar equation appears in Bernanke and G¨rkaynak (2001) and in u Bottazzi and Peri (2007). The ˆ ˆ version in Bottazzi and Peri (2007) is yt = akt + bAt . so equation 2. 525) arguing that “the reason that capital per worker is high in rich countries is that total factor productivities are high in rich countries.invoke the stylized Kaldor-fact that physical ˆ ˆ capital and output grow at the same rate in the long run (k = y ) to ensure a constant capital-output ratio. One aspect of human capital is that it measures the development of new knowledge in one country. the absolute change of human capital times a constant is equal to the growth rate of GDP per capita as I will show ˙ in the empirical part in chapter 12: ω h = y . Having established this augmented technical progress function.6 and 2. so no additional series is included for institutions . Similarly. The capital stock per capita obeys the following rule: . which inﬂuences the proﬁtability of new investments. Openness is assumed to be exogenous and ht may either be exogenous or it may be determined as in the simple Lucas model in a human capital producing sector. openness is also a measure of how intensely an economy is making use of knowledge used or generated abroad.” A stronger development of openness and human capital raises the productivity of physical capital (the proﬁt rate) and therefore the rate of capital accumulation.
physical capital. 360) highlighted that this is common practice even if the formal layout of papers is the other way around: “An author starts from some doctrine which he wishes to defend [. Each chapter on the determinants includes an overview of the theoretical and empirical literature and concludes with an assessment of how to best model this variable.” This chapter was about ﬁnding that “least unplausible” model . In addition. However. Joan Robinson (1961. The following seven chapters will prepare the ground for the empirical analysis thereafter by presenting datasets on GDP and its main determinants. p. e.they are pair-wise cointegrated.24 is more mixed. Equilibrium in the Kaldor model is ensured by assuming that sP − sL > β(Y /K). human capital and openness indeed follows equation 2. These results will be described in detail in chapter 12. the levels of the grand ratios such as the capital-output ratio turn out to be quite diﬀerent across countries. In principle.25) it = kt+1 − kt = (yt − yt−1 ) α + β pt−1 pt−1 pt yt +β − kt−1 kt kt−1 If GDP and proﬁts both do not change then there will not be any addition to the capital stock.2.and one that matches the empirical results. i.7 The augmented Kaldor model 33 (2.24) kt = αyt−1 + β pt−1 yt−1 kt−1 It is equal to a coeﬃcient α times output in the previous period plus a coeﬃcient β times proﬁts p in relation to the capital stock in the previous period. This possible endogeneity of human capital was highlighted by Bils and Klenow (2000). my empirical analysis in chapter 12 shows that GDP reacts to slow-moving human capital. I will show in chapter 12 that the evolution of output. which is the same as saying that there are three cointegrating relationships in the paths of these four variables . Reformulation of this equation leads to the rule for investment per capita which depends on the change in GDP and the change in the proﬁt ratio: (2. As is the case in most economic papers. The evidence for physical capital reacting to output and to the deviation of the capital-output ratio as in 2. This is in line with the fact that building human capital is a lengthy process that is determined by decisions of society and policy-makers several years earlier.23.] and sets out ﬁnding the least unplausible-looking assumptions that lead to the conclusions that he requires.. the savings rate out of proﬁt income has to exceed the savings rate out of labor income by an amount larger than the sensitivity of the capital stock to the proﬁt rate times the output-capital ratio. the signiﬁcant link between GDP and human capital and openness. my theoretical model is to a large extent motivated by the empirical results such as the absence of scale eﬀects.. . and the enogeneity of physical capital. this model could be extended to allow human capital to react to higher GDP just as physical capital does.
It is well known that GDP is not a measure of wellbeing. A more meaningful measure of labor productivity is GDP per worker or per hour worked. GDP only measures market production. a hurricane is not accounted for. By contrast. high taxes on labor contribute to a lot of home production while non-market production might be particularly important in emerging markets. The growth rates of total GDP are in focus when international organizations or the media discuss an economy’s performance. which also takes into account that working hours diﬀer signiﬁcantly across countries. It is often labeled (labor-)productivity and will be the main focus of this study.3 The dependent variable: GDP growth The purpose of this study is to analyze and forecast the growth rate of gross domestic product (GDP). e. . of the value of all ﬁnal goods and services produced in an economy.2 on page 40 shows that most studies in the area of growth empirics focus on GDP per capita. strong growth rates in many Asian economies may stem partly from non-market activities becoming market activities and therefore being included in oﬃcial GDP estimates . then economic fundamentals such as human capital may only explain parts of measured GDP growth. it is not clear why people automatically become unable to work on their 65th birthday. Some studies focus on GDP per capita of the working age population. Table 3. If the share of these activities varies over time. i. many countries have a signiﬁcant amount of home and non-market production. While I will not investigate GDP per hour worked. However. For example. usually referring to people aged 15 to 64. Likewise. In continental European countries. while the value destroyed by. say. GDP tends to rise after natural disasters because the rebuilding activities enter GDP. labor input as measured by overall hours will be an important variable to help explain GDP per capita. However. This is the most widely used measure of an economy’s success.which also may help explain why some countries’ incomes have converged to the world leader. for example. GDP per capita is more useful when comparing the economic situation of individuals.
chain-weighted real GDP in Spain jumped by 10. For example.3% in 1 2 In cooperation with the database agent Global Insight. Thirdly. The PWT uses domestic currency expenditures for consumption. The IMF’s IFS provide history for GDP only from 1980 as index levels and their series were full of errors when I ﬁrst retrieved the data. I asked the IMF to upload corrected data for Belgium.2 It provides. Firstly.edu . The most widely used database for academic work (see table 3. Data are available on pwt. The main disadvantages for my purpose are that all series end in the year 2000 and that the annual growth rates do not equal those reported by international organizations or the media.1 For these reasons. Real GDP per capita is available with 1996 as the base year (rgdpl) and as a chainweighted series (rgdpch). This database was ﬁrst made available in the early 1980s and became very popular in the early 1990s with the release of version 5 (Summers and Heston ).upenn. it has to provide reasonably current data. And fourthly.econ. GDP data have to be available for around 40 rich and poor countries around the world.” In addition. since I want to compare my forecasts with others. The OECD’s ECO database is not too helpful for the purpose of this study because it only covers OECD countries.2 was released in late 2006. the OECD’s Economic Outlook database (ECO). investment and government spending and transforms each component into ”international prices” using purchasing power parity exchange rates. Secondly. Four datasets satisfy at least some of the conditions outlined above: the IMF’s International Financial Statistics (IFS). Its major advantage is the easy and free accessibility.1) in early 2006. the growth rates have to be equal to those published and forecast by the IMF or the OECD.36 3 The dependent variable: GDP growth 3. it has to provide a history at least back to 1970 to allow panel estimation. among others. the PWT showed some bugs. These purchasing power parity (PPP) components are then added up to total GDP. IFS data will not be used here.1 (PWT 6. and Greece. p. nominal and real GDP data per capita for 168 countries from 1950 to 2000. But some large diﬀerences between the IFS database and the IMF’s own semi-annual World Economic Outlook remained. Heston and Summers (1996. Version 6. Choosing a dataset is not a task of minor importance: Hanousek et al. (2002). See Heston et al.2 at the end of this chapter) is the Penn World Table. (2004) show that the results of studies on growth determinants are sensitive to the choice of database. 24) point out that many researchers used the PWT growth rates “unaware that the rates they obtained are not the same as the rates implicit in the countries’ own national accounts. the Penn World Table (PWT) and the World Bank’s World Development Indicators (WDI). with a lag of at most two years.1 Choosing the appropriate data source The dataset for GDP has to satisfy several requirements besides being calculated according to a common standard. Finland. available as release 6.
39 0.50 0.58 0. Hungary Peru Venezuela Morocco UAE Kazakhstan 2.00 2.31 0.21 0.70 China India Brazil Mexico Korea Indonesia Taiwan South Africa Turkey Thailand Argentina Philippines Colombia Egypt Malaysia Chile Nigeria Israel Singapore 15.31 0.73 1. The weights in world GDP of the 40 countries considered 21 rich countries 19 emerging markets United States Japan Germany UK France Italy Canada Spain Australia Netherlands Belgium Austria Sweden Greece Switzerland Portugal Norway Denmark Ireland Finland New Zealand Sum 21: 20.89 0.48 0. All rely on information from the United Nations’ International Comparison Project (ICP).81 1.78 1.95 2. converting local currency values using purchasing power parity exchange rates (in the case of the PWT and now also the WDI called “international dollars”).55 0. Source: IMF WEO April 2006 .60 1.6 Table shows PPP weights in world GDP in 2005 in %.00 3.1.20 Not included Russia Iran Poland Pakistan Saudi Arabia Ukraine Bangladesh Vietnam Algeria Hong Kong Romania Czech Rep.03 0.50 0.8% the next year. The WDI.27 0.82 0.10 6.27 0.41 0.17 48.53 0.1 Choosing the appropriate data source 37 1995 only to fall back by 4.58 1.3.44 0.22 0.33 0. Table 3.21 Sum 19: Sum 40: 36.28 0.91 0.32 0.45 0.87 0. The WDI provides long histories on GDP at constant prices in local currency units and in international dollars for a large number of countries. The database is updated once a year in the spring.58 0.41 5.39 0.93 0.63 1.27 0.81 0.40 4. The only small downside is that the most recent data available stem from two years prior to the release year. The main focus of this study is on growth rates of GDP.9 85.26 0.28 0.38 0. The WDI is the database used in the empirical analysis in this study.03 0.32 0.13 3. However. The most reliable and comprehensive database appears to be the World Bank’s World Development Indicators (WDI).76 1.41 0.68 0. OECD ECO and PWT databases all provide GDP in PPP US dollars.28 0.93 0. for policy analysis cross-country comparisons of levels of GDP in a common currency are necessary as well.55 0. None of the other databases showed these changes.31 0.66 0.
Figures 3.6% of world GDP in purchasing power parities in 2005.1 lists the countries included with their respective shares in world GDP according to the IMF’s April 2006 World Economic Outlook using PPP exchange rates. My country sample covers 85. And countries such as Pakistan or Bangladesh lack some of the data needed for a fundamental model (e. The levels of overall GDP are then divided by population levels to get GDP per capita. 40 30 20 10 0 -10 -20 -30 -40 -50 Percentage differences to the level in the USA in 2000 PPP GDP per capita PPP GDP per hour worked Norway France Germany Austria Ireland Belgium Netherlands Italy UK Denmark Sweden Finland Greece Japan Switzerland Australia Fig. This procedure maintains the “headline” growth rates of GDP but makes the levels roughly comparable across countries. 3. The biggest diﬀerence appears for Norway New Zealand Portugal Canada Spain USA . Earlier and later years are calculated using the growth rates from the series in local currency units. Oil exporters like Iran and Saudia Arabia are just as impossible to model based on growth theory.g. years of education). On average. Table 3. Ukraine and Vietnam) because not enough time-series history is available and because they have been through a major structural break. Poland.2 and 3.38 3 The dependent variable: GDP growth The GDP data used here are in 2000 international dollars.1.3 at the end of this chapter plot the trajectories for GDP per capita. g. the PWT income levels are higher than the WDI levels in the poor countries and lower in the rich countries. Russia. Large diﬀerences in distance to frontier The diﬀerent databases provide slightly diﬀerent pictures of how rich the countries are relative to the USA as measured by GDP per capita. The starting point is the level of GDP from the WDI database for the year 2000 for each country. The most important countries not included in my sample are transition countries (e.
The literature on the distance to frontier models economic growth depending on how far a country’s productivity is away from the technology leader usually the USA. these can make a signiﬁcant diﬀerence as ﬁgure 3. However. By contrast. who calculate total factor productivity as output per adult minus capital per adult multiplied by capital’s share in income. . Similar conclusions hold for other continental European economies with their low working hours per capita. However. these studies tend to use GDP per capita or total factor productivity per capita.1 Choosing the appropriate data source 39 which the PWT sees 19% below the USA in 2000. Chile’s GDP per capita was 72. Dowrick (2002) argues that the ”true” ratio may lie somewhere in between the two. While French GDP per capita was 26% below the US level in 2000. while it was 70. French GDP per hour worked was almost 10% above the US level according to the WDI data. (2006). Since data have to be as up to date as possible for forecasting purposes.3.2% in the PWT. while the WDI sees it 3% above the US level.9% of the US level in 2000 according to WDI. but one has to keep in mind that comparing income levels is a hazardous task. No account is taken of diﬀerences in working hours. the WDI database is the preferred choice.1 shows. A good example is Vandenbussche et al.
6 PWT 5 Nehru & Dhare.1 PWT 6.0 PWT 5 PWT 6.40 Table 3.6 PWT 5 PWT 5. PWT 5 PWT PWT 5.6 PWT 6.1 & WDI PWT 6.0 PWT 6. Leblebicioglu & Schiantarelli Doppelhofer. Ley & Steel Krueger & Lindahl Sarno Dowrick & Rogers Bosworth & Collins Alcala & Ciccone Barro & Sala-i-Martin Batista & Zalduendo Bond. Esquivel & Lefort Klenow & Rodriguez-Clare Lee. . Subramanian & Trebbi Hausmann.6 PWT 5 PWT 5. Romer & Weil Loayza Islam Caselli.6 PWT 5 WDI ECO PWT 5.1 & WDI WDI n Sample length Source Year Author(s) 3 The dependent variable: GDP growth 1991 1991 1992 1992 1994 1995 1996 1997 1997 1997 1998 1998 1999 1999 1999 2001 2001 2001 2001 2001 2001 2002 2003 2004 2004 2004 2004 2004 2004 2004 2005 2005 Barro DeLong & Summers Levine & Renelt Mankiw.6 WDI & OECD PWT 5. Scarpetta & Hemmings Easterly & Levine Fernandez.1 IMF WEO PWT 6. Some papers may use several measures and sample sizes. Pritchett & Rodrik Ianchovichina & Kacker This study Note: Table is not comprehensive. Pesaran & Smith Sala-i-Martin Edwards Evans Frankel & Romer Hall & Jones Judson & Owen Acemoglu. Miller & Sala-i-Martin Hauk & Wacziarg Rodrik. 1960-85 worker 98 1960-85 capita 96 1960-85 capita 97 1960-85 worker 98 1960-85 capita 102 1960-89 capita n/a 1960-92 93 1960-90 capita 54 1960-89 capita 98 1985 worker 127 1960-88 capita 64 1950-88 capita 69 1995 capita 15-64 21 1971-98 capita 73 1960-95 capita 72 1960-92 capita 110 1960-90 capita 7 1950-92 worker 51 1970-90 worker 84 1960-00 worker 138 1985 capita 86 1965-95 capita 89 1961-00 worker 98 1960-98 capita 98 1960-92 capita 69 1960-00 capita 140 1995 capita n/a 1950-99 capita 70 1966-99 capita 40 1971-03 PWT 4 PWT 5 PWT 5 PWT 5 PWT 5 PWT 5 PWT 5 PWT 5 PWT 5. Johnson & Robinson Bassanini. Databases and sample sizes Speciﬁcation Measure Growth rate of Growth rate of Growth rate of Level & growth of Level of Level of Growth rate of Level and growth Level of Growth rate of Growth rate of Level of Level of Level of Growth rate of Level of Growth rate of Growth rate of Growth rate of Growth rate of Level of Growth rate of Growth rate of Level of Growth rate of Growth rate of Level & growth of Growth rate of Growth rate of Level of Growth rate of Growth rate of Level & growth of per per per per per per per per per per per per per per per per per per per per per per GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP TFP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP GDP per per per per per per per per per per capita 98 1960-85 worker 25 1960-85 capita 98 1960-89 capita 15-64 98 1985.2.
GDP per capita in 21 rich countries 25.000 in PPP USD of 2000 35.000 USA 25.000 15.000 Taiwan Argentina 10.000 Switzerland 20.000 0 1971 1976 1981 1986 1991 1996 2001 Fig.000 Korea Israel 15.000 in PPP USD of 2000 Singapore 20. GDP per capita in 19 emerging markets .000 5. 3.000 Norway Ireland 30.000 Portugal 5. 3.1 Choosing the appropriate data source 41 126.96.36.199.000 10.000 1971 1976 1981 1986 1991 1996 2001 Fig.
In their time in the late 18th century this might have been an appropriate description of economic reality. 4. the causality runs primarily from income to population growth as this chapter will argue. Furthermore. If there is more food. But with a ﬁxed supply of land and slow technological progress. Conversely. this is not the case in most countries. Therefore. The usual assumption in the growth literature is that utilization rates of labor supply are constant. many empirical growth models use the growth rate of the population to explain the level of GDP per capita. However. 1957). while there is a negative correlation between these two variables. animals and humans will have more oﬀsprings. the larger population will then depress the average consumption possibilities and eventually birth rates. population growth is assumed to be an exogenous variable with a constant growth rate of . In the neoclassical model pioneered by Solow (1956. the rate of population growth (or the fertility rate) will not appear in my model of per capita GDP. but it is no longer relevant in most economies in the 20th and 21st centuries. the age structure of the population may be important for the evolution of GDP: a large share of experienced workers in the overall population is likely to be positive for average income as outlined in the work of Malmberg and Lindh (2004a and 2004b). a variable that is usually not included in empirical growth models receives a signiﬁcant role in my framework: hours worked. However.1 Population growth is endogenous The presumed negative impact of population growth on income levels that is still used in many growth models has deep roots in classical and neoclassical analysis. However. Classical economists like Malthus and Ricardo relied on biology to model population growth.4 Labor input As mentioned in chapter 2. population growth is crucial for calculating forecasts for overall GDP growth from the forecasts for per capita growth.
these scale eﬀects do not reﬂect macroeconomic reality.” In addition. the growth in per capita income during the past 150 years has little to do with population” and with Easterly (2001. The message that many empirical economists took from this reasoning is that countries with faster population growth will be poorer than countries with slower population growth. Temple (1999.5%. 95) calls ”development. Endogenous growth models with spillovers kept the assumption that population levels or growth rates were exogenous. there is unlikely to be a positive impact on per capita income. p. while it fell from 2. 2002] or Segerstrom ). as outlined in chapter 2 this also does not reﬂect reality.” Likewise. Richer societies tend to have lower birth rates. Furthermore. However. who argue that technological progress raises the return on human capital and induces a substitution from the quantity of children to the quality. the opportunity cost of having children goes up (foregone earnings and costs of educating children) as income rises. the best contraceptive. 146) who conclude that “in the modern view.44 4 Labor input n. p. population growth does not aﬀect per capita income growth . steady state labor productivity is lower. p. Partly responsible are the better health systems in rich societies which lead to lower child mortality rates. in these models a higher level of population tends to lead to higher growth rates because knowledge spillovers are more powerful if there are more people nearby. Keynesian models also tend to see a positive impact of population growth on GDP as more people bring a stimulus to demand. However. Higher population growth reduces the steady state level of GDP because the additional workers have to be equipped with physical capital which reduces the average capital available per worker. p. As outlined above.” This is in line with the model of Galor and Weil (2000). large social security systems in rich . Fertility rates and population growth depend on the level of income.g. 91)“that there is no evidence one way or the other that population growth aﬀects per capita growth.” In general. Easterly (2002. Therefore. p. High immigration into Germany around the time of reuniﬁcation appears to be an example for this link.1% in the 19 emerging markets. 422) already pointed out that ”people shift from saving in the form of children to saving in the form of physical and human capital. Again. Jones [1995b.nor is it ﬁxed or exogenous. 142) concludes that “the popular belief that population growth is economically harmful is not yet well supported by statistical evidence. Why are the one billion people of India so much poorer than the 5 million people of Finland? In the second generation of endogenous models the rate of population growth has a positive impact on the rate of per-capita GDP growth (e.5% to 1. average population growth in my sample of 21 rich countries has fallen from 1% per annum to 0. investment and income growth. I side with Becker. These theoretical considerations led to development policies in the 1960s and 1970s focusing on birth control as a path to higher income. Over the past 40 years. Glaeser and Murphy (1999. Barro (1991.
1 Population growth is endogenous 45 countries imply less dependence on own children for old-age provision.ggdc. And it explains why cross-section regressions ﬁnd a signiﬁcant relationship between birth rates and income. This introduces a link from high income to low birth rates. Because fewer children die and those surviving are better educated. admitting that ”if s and n are endogenous and inﬂuenced by the level of income. I use the data from the Groningen Growth and Development Centre.] are potentially inconsistent”. which may lower birth rates as well. Countries with higher levels and growth rates of income tend to attract immigration. 411) relegate the issue of endogenous population growth to a footnote. then estimates [.net/dseries/totecon. as the UN forecasts are based on sometimes outdated assumptions. However. Net migration rates are becoming an ever more important source of variation in population growth rates. but also by the relative attractiveness of countries according to gravity models. For total mid-year population levels until 2005. This serves as a warning that the UN forecasts . Admittedly. where the government instituted a one-child policy in the late 1970s. This was not part of neoclassical models. fewer have to be born to support their parents in old age.. When trying to forecast GDP growth. 1 http://www. this is probably the exception rather than the rule.html .. Population data are still necessary to calculate per capita levels of the other variables and to calculate forecasts for total GDP from those for per capita GDP.just as any other forecasts . This circular reference should be avoided. The decision to migrate is partly driven by migration policies. p.4. In some cases this leads to jumps in growth rates between 2004 and 2005 (Switzerland. My conclusion is that population growth should not be treated as an exogenous variable despite the great regression ﬁt both in cross-section and in panel models. which reduced population growth far below the level indicated by income and education.may not always materialize. They recommend the use of appropriate instruments. there are cases where population growth is exogenous. Singapore and Israel). But population forecasts use the development of GDP as an explanatory variable. which focus mainly on closed economies. a regression of GDP on population growth will probably have high explanatory power. but acknowledge that this is a formidable task. South Africa. I interpolated the levels published for ﬁve-year intervals using a spline and appended the resulting annual changes to the Groningen data to calculate population levels. Romer and Weil (1992.1 The forecasts for annual percentage changes until 2020 stem from the UN Population Division’s 2004 revision. Mankiw. The prime example is China.
the high levels of GDP per capita in the USA and in Japan owe a lot to the high level of annual hours worked per capita there.one of the most frequently used variables in growth empirics . Prescott (2004) sees tax rates as explaining why Americans work 50% more in the market economy than many Europeans. from diﬀerent unemployment rates (i. working hours will have to be an important element in long-run forecasts of GDP over time. being determined by taxes. the slow growth rate of GDP in Germany over the past decades stems partly from the shrinking labor input per capita.2 Hours worked per capita are important While population growth .) and preferences for low and falling hours worked in Europe. there is more household production in many European economies. a point made in detail by Ramey and Francis (2006). labor market institutions and preferences. The long-run trend of hours worked per year is largely exogenous to per capita GDP. In the debate about diﬀerences in income and growth between Europe and the US. which does not aﬀect GDP but should aﬀect overall welfare. This variable captures the signiﬁcant variation over time and across countries in the overall usage of the available labor resources in market activities. Olivier Blanchard (2006) and Robert Gordon (2004) highlight the role of both institutions (regulation. or are not tired if the working day ends after six hours. so that there need not be an increase in leisure. The variation may stem from diﬀerences in the hours worked per worker. another less frequently used variable should play an important role: hours worked per year and per capita of the total population. Measured hourly productivity of those remaining employed will be higher. but a negative link to GDP per hour worked: The least productive people are usually the ﬁrst to exit employment. Ramey and Francis (2006) argue that the increase in hours spent in school can explain a large part of the decline in hours worked. Likewise. Similarly. Some countries will decide to use the additional income and productivity 2 On the other hand. starting from a 50-hour workweek for example. There should be a positive link between hours worked per capita and GDP per capita.e.should therefore play no role in models for per capita GDP growth. where hours worked are still high today.2 Figure 4. . Furthermore.1 below shows the trajectories for the 21 rich countries. taxes etc. In emerging markets. the ratio between workers and the labor force) and/or from diﬀerent labor force participation rates: (4. The use of per capita data (instead of per worker) is the natural choice for a study that focuses on per capita GDP (rather than per employee). fewer hours worked per employee could mean that productivity goes up because workers are well-rested after a long weekend.1) hours worker laborf orce hours = ∗ ∗ population worker laborf orce population As was shown in the previous chapter.46 4 Labor input 4.
For smaller samples. lack of data prevents similar calculations for the emerging markets sample. hours worked per year per capita of the overall population for the 21 rich countries are derived. However. Limited availability of data is the standard explanation why hours worked are not considered in large cross-country studies. In the rich country group. Unfortunately. a speciﬁcation in GDP per person employed is likely to yield diﬀerent results from that in GDP per capita. These country-speciﬁc diﬀerences should be useful in explaining past GDP growth . The resulting GDP growth rates are likely to diﬀer signiﬁcantly.with ups and downs driven by the business cycle (see ﬁgure 4. so we use the number of workers instead to measure labor input. There are no series for hours worked in the Penn World Table or in the World Bank’s WDI.” Bassanini.1). As mentioned above. the OECD publishes data in its Economic Outlook database.. hours worked in Spain have trended upward since the mid-1980s and are now above the rich-country average. the number has fallen to around 750 in the early 1980s and has stayed there since .and for making better informed forecasts. Hours worked per capita kept falling in Germany throughout the sample period and are now almost 20% below the rich-country average. the empirical growth literature tends to omit this important element in explaining diﬀerences in levels and growth rates of GDP across countries.4. diﬀerent countries show vastly diﬀerent trajectories. but one that does not depend simply on the level of income but on society’s decisions. In short. p..2 Hours worked per capita are important 47 to extend their leisure time. although this trend seems to have slowed or even stopped since the 1980s. Scarpetta and Hemmings (2001. By contrast. Hall and Jones (1999.if only in the data appendix: “However. From the data on hours worked per employee combined with the employment and population data. any forecasting model has to include a variable for labor input. p.] employment rates have changed signiﬁcantly over time in most Member Countries and in particular in Continental Europe where signiﬁcant declines were recorded in the 1980s and 1990s. The Groningen Growth and Development Centre publishes hours actually worked per person employed per year for 40 countries.” Bassanini. 8) recognize the omission: ””We do not have data on hours worked for most countries. From 860 hours worked per year and per capita in the early 1960s on average in the 21 countries. [. hours worked per year have tended downward since the 1960s. . These data combine a variety of sources and are quite up to date and therefore useful for forecasting purposes. 53) also are aware of the omission . others will decide to continue to work as long as they are doing today. Scarpetta and Visco (2000) comprehensively consider hours worked. Under these conditions.
Demographers speak of the ”demographic transition” when mortality rates fall ﬁrst and birth rates decline with some delay. savings rates might be higher when youth dependency ratios are low. is what matters for the level of per capita income. The ”demographic window” is the period when the birth rate has already fallen .1.taking the youth dependency ratio down .3 Age structure of the population As outlined above. which should be positive for total hours worked and for per capita income levels (but not necessarily for output per hour). the growth rate and the level of the population probably do not have a signiﬁcant impact on the level of per capita GDP or on per capita GDP growth. In addition. history and forecasts . This leads to a high ratio of workers in the total population (see equation 4. The big advantage of considering the age structure for a forecasting model is that its shape in the year 2020 is relatively easy to forecast in 2006: most of the population of 2020 has already been born and fertility and mortality rates do not vary that much over a decade. p.1). 4.48 1000 4 Labor input Japan CH 900 800 700 600 France Spain 500 1971 1976 1981 1986 1991 1996 2001 Fig.” This would address the issue that simply using overall population data equates a 5-year old to a 40-year old and to a 70-year old even though their respective productivities are likely to diﬀer signiﬁcantly.and the baby boomers are in their prime earnings phase. Annual hours worked per capita in 21 rich countries 4. and not population size. Malmberg and Lindh (2004b. In addition. 2) argue that a consensus has emerged ”that population age structure. Demographic research suggests another variable that may have a signiﬁcant inﬂuence on per capita income levels: the age structure of the population.
health. .e. and therefore also the age structure.2 shows. Birth rates. lifestyle and work-environment of.but is set to fall from there as ﬁgure 4.30 Japan Germany 0. the share of ”prime-age” people. Their out-of-sample forecasts produce smaller mean errors.2. those aged between 30 and 49. Malmberg and Lindh (2004a. say. 2004b) use the relationship between the age structure and GDP to derive forecasts for GDP growth into 2050 for 111 countries using data from PWT 6. My empirical analysis in chapter 12 ﬁnds only weak support for these conclusions.28 0. 4.3 Age structure of the population 49 are easily accessible on the website of the UN Population Division. From their forecasts until 2050 they conclude ıve that because of recent decreases in fertility rates. 0. has risen from 25% in 1975 to almost 30% in 2005 using the unweighted average. but larger mean absolute errors than the na¨ alternative models. Share of 30 to 49-year age group In the long run some of the endogeneity problems outlined above persist.32 USA 0. They regress the natural logarithm of GDP per capita on diﬀerent age shares and a linear as well as a quadratic time trend and ﬁnd the largest positive eﬀect on GDP coming from the 30 to 49-year age share. i. A more serious issue is the fact that the characteristics. the shares of “prime age adults” and of 50 to 64year olds seem to be stationary variables in my sample of 21 rich countries. First of all.34 China 0.1 and from the UN Population Division.22 0.26 0. A 50-year old in an advanced economy today is likely to be healthier and more productive than a 50-year old three decades ago. ultimately depend on income levels.4.24 India 0. the share even reached 33% in 2001 . today’s poor countries will start to catch up with developed economies in which the growth process will stagnate because of the growth of the elderly population. In my sample of 21 rich countries. In China. a 50-year old keep changing over time.20 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 2025 Fig.
they cannot by themselves explain the path of non-stationary GDP per capita. only 13 of the 40 short-run models outlined in chapter 12 ﬁnd a signiﬁcant role for the age structure in explaining the growth rate of GDP . Second.50 4 Labor input Therefore. As a result. but focus attention on other variables such as human capital and openness as will be seen in the following chapters. . I will keep the age structure in the overall model. The panel estimations do not ﬁnd a signiﬁcant relationship.Sweden is not one of them in contrast to Lindh (2004).
This view reconciles the opposing views in the empirical literature mentioned above. Output growth is unlikely to happen without accumulation of physical capital. p. human capital. 1125) ﬁnd ”an enormous eﬀect” of the growth rate of capital per worker on GDP growth in their cross-section estimations but they suspect some endogeneity bias. Krueger and Lindahl (2001. For example. empirical studies should ﬁnd that rapid output growth goes hand in hand with rapid growth of the physical capital stock. this chapter will show that many of the results of earlier empirical studies which ﬁnd a signiﬁcant link between the level of the investment ratio and the level of income are sensitive to the use of investment valued at international prices instead of domestic prices. This chapter will build on the augmented Kaldor model outlined in chapter 2 and it will argue that the accumulation of physical capital reacts to variables like labor input. Easterly and Levine (2001) argue that factor accumulation is not the main driver of output growth. but accumulation is not the main or ultimate reason for output growth.5 Physical capital One of the most obvious and most widely investigated determinants of economic growth is the accumulation of physical capital. trade openness and institutional settings. Output and the capital stock are cointegrated as Kaldor (1957) found (although the term ”cointegration” was not coined back then). Nevertheless. In addition. . The conclusion is that there is neither a theoretical nor an empirical reason why investment ratios should correlate with output levels across countries . Their results are challenged by Bond et al. (2004) who argue that investment has a positive long-run eﬀect on both the level and the growth rate of output. Despite this simple and intuitive reasoning.in contrast to the Solow model as usually applied in cross-country analyses. The general idea is that consumption today is postponed to allow even higher consumption in the future: investment today leads to a higher capital stock and therefore more output tomorrow. a heated debate on the link between capital accumulation and economic growth is ongoing in the literature.
2. this is no longer feasible today.1) ˙ Kit = Kit − Ki. we get: (5. Capital intensities do not vary systematically with income levels. 5. As outlined in section 2. that Kit /Yit = Ki /Yi = γi is constant over time leads to the conclusion that the percentage change in the capital stock is proportional to the investment ratio: . Over long time spans.t−1 Ki.t−1 = Iit − δKi.1 Investment and changes in capital stocks The basics of the accumulation of physical capital were already highlighted in chapter 2. which is itself derived from the national accounts income identity.t−1 Yit Invoking the assumption that the Kaldor ratio holds. The investment ratio (gross investment as a share of GDP) in any given country is not constant. if we expand equation 2. Investment ratios are not proportional to the level of the capital stock either. This chapter will also show that there are signiﬁcant diﬀerences in investment ratios depending on databases and on whether real or nominal ratios are used.52 5 Physical capital Cross-checking the neoclassical assumptions outlined in chapter 2 with reality reveals that none of them hold. where Iit is gross investment in country i in period t and a common depreciation rate δ is assumed.1.6. but can vary over time. For example. 5.t−1 .4 to model diﬀerent cross-sections i. Investment ratios are not proportional to the change in the capital stock. 5. This leads to a circularity problem.2 above. investment ratios barely diﬀer across countries and therefore cannot explain the large variation in income levels. Kaldor (1957) avoided this issue by using the weight of steel embodied in capital equipment to aggregate the diﬀerent types of machines with the same physical unit. However. The data reveal ﬁve important features: 1. the simple relationship between investment and physical capital presents some pitfalls that are worth discussing in more detail.e. value measures rely on the rate of interest. Dividing both sides by Ki. value measures of physical capital are unavoidable. 4. In order to make progress towards a long-run forecasting model. Unfortunately. i.2) ˙ Yit Iit Kit = −δ Ki.t−1 and expanding by total GDP Yit gives: (5. 3.1 Measuring capital accumulation Physical capital is diﬃcult to measure.
but they do not test it. these exceptions include Bosworth and Collins (2003) and 1 2 See page 36 above for a general description of the PWT and table 5. Bosworth and Collins (2003. investment goods are expensive relative to domestic services in poor countries. and decide which route is most promising for the purpose of forecasting. p. 1434) prefers using domestic prices since these “characterize the trade-oﬀs faced by the decision-making agents. . Most empirical growth analyses have used and continue to use the investment ratios supplied by the Penn World Table. which evaluates all output components at international prices. but these databases are hardly used in growth empirics.is to also assume that γi is the same for all countries: γi = γj = γ.that is unfortunately usually not talked about .1 on page 65 for an overview of the measures and databases used in growth studies Dowrick (2002) documents a negative correlation between the relative price of investment to consumption and the level of real GDP per capita.1 The relative price between investment and consumption goods is exactly the same in every country in a given year and equals the weighted world average relative price. several databases for investment and capital stocks are used in this study.diﬀerent investment ratios To improve the understanding of the link between physical capital and output.2 Summers and Heston (1991.5.3) 1 Iit ˆ Kit = −δ γi Yit The big step in cross-country analysis . p. but few researchers seemed to take note. As table 5. To do so requires databases for investment ratios and capital stock.2 Diﬀerent databases . 12) also think that ”capital input should be valued in the prices of the country in which it is used”.1 on page 65 shows. the PWT shows much lower investment ratios in low-income countries than the investment ratios based on local prices. However.” Knowles (2001) also argues that it is preferable to use investment ratios (and government consumption shares) at local relative prices. 5. 337339) clearly made that point when releasing the ﬁrst versions of the PWT. countries with relatively high investment ratios would see proportionally higher growth rates of physical capital. Nuxoll (1994. In that case. This allows three main goals of the study to be achieved: cross-check theory and reality.1. evaluate ﬁndings of previous empirical research.1 Measuring capital accumulation 53 (5. Most empirical crosssection growth papers simply assume that this relationship holds. I will show later in this chapter that capital-output ratios are not identical across countries. The World Bank’s WDI and the OECD’s ECO database use local relative prices. which are the subject of the next sections. pp. the international ratios of investment and consumption goods prices are not the relative prices that steer resource allocation inside each individual country. Since in reality.
but if it is not yet scrapped it may produce just as much as when it was ﬁrst set up. the time paths of nominal and real investment ratios diﬀer signiﬁcantly. some are gross capital stocks. As a result. the simple ratio between real investment and real GDP may have no meaning. As a result. while others are net of scrapping. Moreover. The recent spread of chain-weighted GDP statistics has added a further complicating factor: the components of real GDP no longer add up to overall GDP. Therefore. However. therefore I will only analyze the latter here.1 above. With the help of real and nominal GDP data. The OECD’s ECO database only includes data for the OECD economies. even in countries that do not use chain-weighted GDP. nominal investment ratios ranged from below 15% of GDP in Colombia and Argentina to 29% in South Korea and even 40% in China. so the levels are not comparable. Scarpetta and Hemmings (2001).54 5 Physical capital the OECD’s own growth project as in Bassanini.1 shows. Furthermore. Capital stock series in the national accounts are calculated using the perpetual inventory method in equation 5. while the two ratios were equal in the base year 2000 as ﬁgure 5. dating back to 1970 or even further. Real ratios maintain the relative prices of the base year throughout the sample period. As outlined above. 5. The WDI provide nominal and real ﬁxed investment spending in local currency units for a large number of countries. the data released in spring 2005 run until 2003. the time path of the nominal ratio shows a slight decline while the real ratio has been roughly constant since the late 1970s.5 percentage points higher using nominal values than using real values. . prices of investment goods tend to rise at a signiﬁcantly lower rate than the overall GDP deﬂator. The rate of depreciation should be interpreted as a scrapping rate: a machine may have been fully depreciated. The OECD’s ECO database provides data for the business sectors in OECD countries. which makes it less useful for the purpose of the present broad-based investigation. Empirical studies might come to diﬀerent conclusions depending on the measure used. the business sector may not capture the evolution of the overall economy and the country sample is too small for my purpose. the unweighted rich-country average investment ratio in 1980 was 3. getting data on capital stocks is not a straightforward task.3 Capital stocks from perpetual inventory Unfortunately. I assume a scrapping rate equal to the depreciation rate. In 2003.1. the use of nominal investment ratios is indispensable when investment is added on the right-hand side of growth regressions. However. For the available countries the ECO investment ratios are basically identical to the WDI ratios. nominal and real investment ratios for all 40 countries were constructed. As a result.
the same rate as in Hall and Jones (1999) or in Vandenbussche et al. Yao and Lyhagen  or Chen and Dahlman ) or even 3% (Mankiw. For the growth rate of investment κ. while 3% clearly appears too low. An experiment with 5% did not show a major diﬀerence in the paths of the capital stock. Romer and Weil . (2006).4% for the rich countries and 5% for the emerging markets. Their use of long-run averages gives less weight to the initial investment ratio than the 3 An alternative calculation using country-speciﬁc growth rates for the ﬁrst 10 years leads to implausible levels and paths of the capital stock. Klenow and Rodriguez-Clare  and Sarno ). the growth rate of GDP g and the depreciation rate δ: γ = [I/Y ]/[g +δ]. The capital stock in the base year is set equal to real gross investment in the base year divided by the sum of the average growth rate of investment and the depreciation rate. . Other authors use slightly higher rates of 7% (Easterly and Levine ) or lower rates of 5% (Prescott . So K1971 = I1971 /(κ + δ).1. Nominal and real investment shares in 21 rich countries A rough estimate for the capital stock in the 40 countries in my sample can be constructed by perpetual inventory calculations following Hall and Jones (1999). They then multiply this γ by the average GDP level in the ﬁrst three years of the sample.5. For the initial capital stock they calculate country-speciﬁc steady-state values of the capitaloutput ratio γ from the steady-state investment ratio I/Y .1 Measuring capital accumulation 55 29 % of GDP 27 WDI nominal WDI real (base=2000) 25 23 21 19 17 15 1961 1966 1971 1976 1981 1986 1991 1996 2001 Fig. 5.3 The depreciation rate δ is set at 6 percent for all countries in all years. Easterly and Levine (2001) use a slightly diﬀerent approach. I use the grand mean of 2.
000 in Nigeria and USD 4. However. the nominal ratio in Finland went from almost 30% in the early 1960s to 25% in the late 60s.5% per annum in Finland and Argentina to 6. In per capita terms. the investment ratio in the country that originated most of the growth models tended to be stable over time. But with several of theses shifts happening. this assumption does not hold in reality. However. 5. but any guess of the initial capital stock ceases to be important after roughly a decade anyway.1 Investment ratios are not constant As outlined above. I will not explore this route.000 in Japan as indicated in ﬁgures 5. a crucial conclusion from neoclassical models is that countries with high investment ratios have high income levels per capita because . this would render cross-section models . the simple neoclassical model has a rather narrow focus. growth rates of the per capita capital stock ranged from 0. 5. For example.with their reliance on long-run averages . The capital stock data based on real investment from the WDI are expressed in international dollars of 2000 just as the GDP series are. Over the 10 years between 1993 and 2003. it is possible to test the explicit and implicit assumptions used for capital accumulation in the literature. up to more than 30% again in the mid-70s and down to below 20% in the mid-1990s using the WDI data. Mainland China stands out with a growth rate of 10.2.5% in Korea and Taiwan. These large diﬀerences are of the same order of magnitude as the diﬀerences in per capita GDP. It analyzes an economy with a constant savings or investment ratio.9%. In the neoclassical framework this could be interpreted as reﬂecting a shift in the steady state of the economy including the adjustment of the capital stock to that new steady state.2. Investment ratios tend to vary signiﬁcantly over time even in rich countries. capital stocks in 2003 ranged from USD 3. Bond et al. 5.56 5 Physical capital approach by Hall and Jones.useless.2.2 Investment ratios do not diﬀer much across countries As outlined above.8 and 5. see ﬁgure 5.000 in Norway and 81. The backward sum will just rise more quickly.9. (2004) suggest using the backward sum of the investment shares as a measure of the capital stock and ﬁnd that output and the backward sum of investment cointegrate in a panel when allowing for country-speciﬁc deterministic trends. their calculation is equivalent to the perpetual inventory method without depreciation and with a lower initial level.700 in India to 78.2 Main insights on capital accumulation With the datasets for investment ratios and capital stocks at hand. Incidentally.
the USA and the UK had rates of below 20%.2 Main insights on capital accumulation % of GDP 57 40 South Korea 35 30 Finland 25 20 USA 15 Source: WDI 10 1960 1965 1970 1975 1980 1985 1990 1995 2000 Fig. the long-run averages of the nominal investment ratios from the WDI vary only slightly across countries. per capita GDP levels in 2000 in international prices were 90% or more below the US level in many countries. As shown above. However.5%. The average investment ratio over the period 1970-2002 was 23% in my sample of 40 countries with a standard deviation of 2. This is in contrast to DeLong and Summers (1991). Bond et al. On average. 5.2. while Argentina. who ar- . Nevertheless.3. South Korea and Thailand had investment ratios of more than 30%. The correlation coeﬃcient for the 40 countries is -0.” Similarly.all else equal. This already indicates that diﬀerences in investment ratios cannot possibly account for the large diﬀerences observed in income levels. many papers draw conclusions such as ”a 1 percentage point increase in the investment share brings about an increase in steady-state GDP per capita of about 1.5. Nominal investment shares over time they have a high capital stock per capita . Malaysia.2. while the upper bound in practice is around 40%. as illustrated in ﬁgure 5. The theoretical upper bound is 100%. (2004) make the point in their pre-testing section arguing that therefore the investment share ”should have neither a deterministic nor a stochastic trend in the long run. the investment share in a given country cannot grow without bounds and therefore cannot explain long-term income growth. During 1981-2000.7 percentage points. countries with low per capita income levels had higher investment ratios over that period.” The conclusion is that investment ratios by themselves cannot be a driver of income or growth diﬀerences: “Diﬀerences in the fraction of product invested do not account for diﬀerences in international per capita incomes” (Prescott ).
425) alluded to the role of “variations across countries in the ratio of the investment deﬂator to the GDP deﬂator”. The correlation coeﬃcient with per capita GDP jumps from -0. Barro (1991. Why then is the investment ratio still one of the most robust correlates of income in cross-section studies? The reason is the wide-spread use of data from the PWT which value investment goods at international prices as table 5. It seems that the widespread use of international prices drives the results in cross-section estimates .58 5 Physical capital gue that there is a robust. .c. p. Investment ratios and the level of GDP p.1 on page 65 shows.48 according to the PWT data. but he did not elaborate on this point. Knowles (2001) shows that the results 40 USA Per capita GDP in '000 international $ in 2003 (WDI) 35 30 Japan 25 Korea 20 15 10 China 5 Egypt 15 20 Nominal investment in % of GDP (WDI).2 according to the WDI data to +0.” And that the “conversion to international prices introduces a strong positive correlation between the investment rate and the level of income per capita” (p. Bosworth and Collins (2003.3.a fact that is usually ignored when the results are interpreted. avg 1994-2003 25 30 35 40 0 Fig. which does not proxy for some other determinants of growth.6% to 35.4 illustrates how the gap between the World Development Indicators (WDI) and Penn World Table (PWT) investment ratios varies systematically with the level of GDP per capita. 11) point out that “investment shares for low income countries are much smaller when measured in international prices than when measured in national prices. causal link from equipment investment to economic growth. 5. Figure 5. Poor countries with low prices of non-tradables are attributed low investment ratios by those constructing the data . 12). Nominal ratios from this database show a much wider variation with a range from 9.4% over the 1981-2000 period. p.and empirical estimates recover this relationship.
But why do time-series or panel estimates (like the pooled mean group technique) also tend to ﬁnd a link between the development of GDP and the development of the investment ratio over time? It seems that these techniques pick up co-movements of investment and output at the business cycle frequency. As the other correlates of income that usually enter on the righthand side tend to move only sluggishly over time. My overall conclusion is that investment ratios should not be used to model income levels. Fig.5. Predictable diﬀerences in investment shares . three conclusions arise from the analysis of investment ratios: (1) The cross-country variation in nominal investment ratios from the WDI database is not nearly high enough to explain the large variation in income levels across countries. (3) Time-series estimates seem to pick up business cycle co-movements of GDP and investment.4. a cyclical acceleration of GDP is attributed mainly to the coincident and also cyclical acceleration of investment. 5.2 Main insights on capital accumulation 59 of Barro and Lee (1994) are sensitive to the use of international or local price levels. In sum. (2) The use of PWT investment data at international prices seems to be behind the ﬁnding of a robust empirical relationship in cross-section estimates. This could be examined by plotting the (probably stationary) time-series residuals of a regression without the investment ratio against this investment ratio.
In reality.5. However. For example. 5. India and Chile all had nominal investment ratios according to the WDI numbers near 23% during the decade from 1994 to 2003 (see ﬁgure 5. there can be strong capital accumulation at low gross investment ratios (if the starting capital stock is low) and weak capital accumulation at high investment ratios (if the starting capital stock is high). p.9 in Switzerland but just 2. Austria. countries with high investment ratios do not necessarily show large changes in the capital stock. the average annual changes in their capital stocks over the same period ranged from 1. One reason is that countries with a high capital stock need higher absolute investment to replace depreciated capital.2. The main reason for these diﬀerences is that the capital-output ratio over that period was 2.0 in India according to the perpetual inventory data introduced above. They see very little correlation between the change in the capital stock and the mean investment ratio in their sample for two reasons: the capitaloutput ratios are neither constant over time nor the same across countries. Bosworth and Collins (2003.60 5 Physical capital 5.3 Investment ratios are not proportional to changes in the capital stock As highlighted above. .5). Switzerland. Fig.8% in Switzerland and 2. Investment ratios and changes in the capital stock Therefore. 10) are among the few to question the proportional relationship between changes in the capital stock and investment ratios. standard cross-section analyses assume that capitaloutput ratios γ are identical across countries so that there is a linear relationship between investment ratios and changes in the capital stock.6% in Chile.5% in Austria to 6.5% in India and 7.
Therefore.2. Capital-output ratios are systematically larger in richer countries”.5 Capital productivity does not correlate with income Easterly and Levine (2001. This is not the case: the correlation between nominal investment ratios and the per-capita capital stocks across the 40 countries is slightly negative using averages over 1994-2003 as ﬁgure 5. not a high investment ratio per se. it supports the view of signiﬁcant diﬀerences across countries in the capital-output ratios. footnote 9) claim that the capital-output ratio ”systematically varies with income per capita. focusing on the investment ratio alone does not provide a complete picture. Klenow and Rodriguez-Clare . 5. driven mostly by China and the Asian Tigers (see ﬁgure 5.5.4 Investment ratios are not proportional to levels of the capital stock Given that the investment data are for gross investment.2. No link between investment ratios and capital stocks 5. Similarly. 5. Fig. Changes in the capital stock are what really matters for output growth.5). but the correlation for the 1994-2003 period is 0.2 Main insights on capital accumulation 61 While my analysis disagree with the ﬁrst reason. one might suspect that countries with a large per capita capital stock tend to have high investment ratios. In my sample there is no ﬁxed relationship between changes in the capital stock and the investment ratio.8.6.6 shows.
who already noted in 1975 (p. p.7 2. Malaysia and Switzerland had some of the highest capital-output ratios in 2003.7). while the USA and the UK had among the lowest.7.9 3. Klenow and Rodriguez-Clare’s results are mostly driven by the use of investment data at international prices from the PWT. The data used here do not support these claims. while the capital-output ratio is around 1:1. The result of no correlation does not change when GDP per hour worked is used instead of GDP per capita. 356) that the dramatic increase of the capital-labor ratio in the course of progress happened “without corresponding changes in the capital-output ratio”. so they also push down any derived capital stock proxies. Japan.1 3. ﬁgure 5.62 5 Physical capital (1997.5 Fig.7 also shows that the capital-output ratio is not roughly identical across countries as Kaldor claims and Felipe and McCombie (2007) use.5 2.1.” However. Using the capital stocks calculated from WDI investment data there is no correlation between income and capital productivity (see ﬁgure 5.4. where the ratio of “the capitallabor ratio is of the order of 30:1.2 and illustrated in ﬁgure 5. As described in section 5.3 2.1 2. This indicates that returns on capital may not be equal across countries in line with the standard conclusion from the international ﬁnance literature. 40000 GDP per capita in 2003 in international $ of 2000 35000 30000 CH JP 25000 20000 15000 10000 5000 Capital-output ratio in 2003 0 1. His example compares the United States with India.7 1. PWT data systematically push down the investment ratio in poor countries. No link between GDP per capita and capital-output ratios This is further support for Kaldor.9 2. 5. .3 3. 89) conclude that richer countries tend to have higher K/Y and higher H/Y than poorer countries.
Even more. Blomstr¨m.2. . physical capital reacts to any deviation from the long-run relationship between GDP and the capital stock.6 Capital accumulation is not exogenous With the investment ratio neither constant nor a good candidate to explain changes in the capital stock or the capital productivity of a country.3 Proper modeling of capital accumulation Accumulation of physical capital is an important driver of (labor) productivity and GDP.in spite of the aggregation issues highlighted earlier. The new panel methods and their use of cointegration analysis usually run into a key problem with investment ratios: they are stationary. but that investment does not Granger-cause growth. the possibly very long lags between investment and GDP growth as well as the variability of capacity utilization over the business cycle may make it diﬃcult to disentangle the individual eﬀects. However. the standard ways of modeling this in the empirical growth literature do not appear to be appropriate for reasons outlined in this chapter. Therefore. growth econometrics should move towards using the capital stock instead of investment ratios to explain income levels. However. Investment picks up if proﬁts earned on physical capital rise above their long-term average because of a productivity shock. there was no need to use capital stock data in cross-section analyses.5. 5. which in turn depends on the development of technology and on the input of other factors. The focus should shift to the development of the capital stock . Because of the (erroneously) assumed correspondence between investment ratios and changes in the capital stock. this does not rule out that in the short run and even in the long run investment has a positive impact on GDP (not least because it is a part of GDP). In the Kaldor model the development of the capital stock depends on the productivity of capital. Results of Granger causality tests are in line with this linkage between capital and GDP. Capital accumulation is driven by the growth rates of GDP and by the returns on capital as outlined by Kaldor (1957) and here on page 31 and supported by the empirical results in chapter 12. this leaves us with the question of what really drives capital accumulation. According to the Kaldor model.3 Proper modeling of capital accumulation 63 5. Lipsey and Zejan (1996) ﬁnd that growth Grangero causes investment.
000 Portugal 20.000 60. Capital stock per capita in 19 emerging markets .000 in PPP USD of 2000 80. Capital stock per capita in 21 rich countries 70. 5.000 Korea 30.64 5 Physical capital 90.000 CH 50. 5.000 0 1971 1976 1981 1986 1991 1996 2001 Fig.000 10.000 Japan NO 70.000 in PPP USD of 2000 Singapore 60.9.000 50.000 Argentina 20.000 Israel 40.000 USA 40.000 Greece 30.000 1971 1976 1981 1986 1991 1996 2001 Fig.8.
0 PWT 6. Hauck & Wacziarg Hendry & Krolzig Hoover & Perez Batista & Zalduendo Ianchovichina & Kacker This study 5. 65 .3 Proper modeling of capital accumulation Note: Table is not comprehensive. with growth rate Cross-section Positive in cointegration vector Cointegration Signiﬁcant positive relation Panel GMM No signiﬁcant impact on growth Cross-section Cross-section Signiﬁcant positive relation Cross-section Large.1 PWT 1 PWT 1 IMF WEO WDI & own Real electrical and non-el. GMM & LSDV Signiﬁcant positive relation Cross-section PCGets Signiﬁcant positive relation Cross-section No eﬀect Cross-section & time Cross-section Signiﬁcant cointegration 2-step panel Year Authors 1991 1991 1992 1992 1994 1996 1997 1997 1999 2001 2001 2001 2001 2001 2002 2003 2004 2004 2004 2004 2004 2004 2004 2005 DeLong & Summers Barro Levine & Renelt MRW Loayza Caselli.6 PWT 5. Some papers may use several measures and sample sizes.1.Table 5. Ley & Steel Bassanini et al. inv (% GDP) Real non-resid inv. (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Nominal investment (% of nominal GDP) no investment included Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) no investment included Capital stock per capita Robust and causal Negative cor.6 PWT 5. inv (% GDP) Physical capital stock (PPI) Physical capital stock (PPI) no investment included Real electrical and non-el. Overview of the empirical literature on physical capital Variable used PWT 5 PWT 5 World Bank PWT 5 PWT 5 PWT 5 PWT 5 PWT 5 PWT 5. inv (% GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real investment (% real GDP) Real electrical and non-el. Barro & Sala-i-Martin Bond et al. Esquivel & Lefort Sala-i-Martin Klenow & Rodriguez-Clare Hall & Jones Easterly & Levine Fernandez.6 OECD & WDI PWT 6.1 PWT 6. Bernanke & Guerkaynak Sarno Dowrick & Rogers Bosworth & Collins Doppelhofer et al. with initial GDP Positive and robust correlation Positive on level Signiﬁcant positive relation Positive on level Signiﬁcant eﬀect Positive correlation Depends of institutions Source Impact Method Cross-section Cross-section Cross-section Cross-section Panel Pi Panel GMM Cross-section Cross-section Cross-section Panel GMM Robust relationship Cross-section Signiﬁcant positive PMG Positive cor. signiﬁcant eﬀect Pool Signiﬁcant positive relation CS.6 & own PWT 5 OECD PWT 5.
Adam Smith (1776) saw two genuine roles for the government beyond defense: to support institutions ”facilitating the commerce of the society” and institutions providing education. While potentially frustrating for policy-makers. Individuals tend to ﬁrst go to school and then earn higher income. He argued that the state should provide access to general education for the broad population and that attendance should possibly be mandatory: ”For a small expence the public can facilitate. and can even impose upon almost the whole body of the people. This was also Smith’s answer to the basic problem in economics: the social question. Support for the view that education is exogenous comes from Stevens and Weale (2004). this slow and gradual movement of human capital can be very informative for a long-run forecasting model. Preferences and educational policies determine how many children complete a certain level of education. who note that the ”spread of formal school seems to have preceded the beginning of modern economic growth. Any resulting rise in GDP will be spread out over many years.in contrast to mainstream neoclassical models . the necessity of acquiring those most essential parts of education” (p. an increase in a country’s level of human capital occurs mostly because young people entering the labor force are better educated than old people leaving the labor force.” Today’s decision to raise a country’s level of human capital will have its biggest impact in 15 years or even later when the better-educated young enter the labor force. As a result.population growth and the accumulation of physical capital are largely endogenous to economic growth. History shows that these decisions do not depend much on current income levels. . Investment in education usually precedes the pay-out period by many years.6 Human capital The previous two chapters have argued that . This chapter turns to an important driver of income that is exogenous at least over a 10 to 15-year horizon: human capital. Equal educational opportunities are a more sustainable vehicle towards social peace than transfer payments from rich to poor. 843). can encourage.
68 6 Human capital In a similar spirit. 70. less crime etc. For example. there are several indirect eﬀects both on income and on measures of a nation’s wellbeing. p. there might be additional eﬀects on co-workers in the same ﬁrm and even in other ﬁrms. decode and understand information more quickly and can also apply more new ideas and be more innovative: “educated people make good innovators. human capital has some similarities to physical capital. higher human capital also leads to additional positive consequences such as improved health. The role of experience was highlighted most prominently by Arrow (1962. banning the import of foreign books and evicting Jews. Another similarity to physical capital is that most of the beneﬁts from a person’s higher human capital accrue to that person and to the ﬁrm where he works. p. He sees part of the economic success of protestant societies stemming from their emphasis on instruction and literacy partly by expecting people to read the Bible themselves. From a growth perspective. social and economic wellbeing. for example. human capital correlates strongly with other factors that are often seen as explaining the level of GDP. The ideas developed by one person (because of his high education level) can help somebody else create more ideas as well 1 Nelson and Phelps (1966).). On top of the productivity-enhancing eﬀects. This human capital can be acquired in many diﬀerent ways at home. p. These beneﬁts are probably quite large. skills and competencies embodied in an individual that facilitate the creation of personal. who stated that “technical change in general can be ascribed to experience. . Landes (1999. Countries with a high level of education or human capital tend to also score highly in economic freedom indices such as the Fraser Institute’s Economic Freedom Index or the Heritage Foundation’s Index of Economic Freedom.” Higher human capital allows an individual to perform higher value-added tasks more eﬃciently and more quickly. In addition to the direct eﬀects of human capital on income. higher life expectancy. Expenditures on research and development also tend to be higher in countries with a high level of human capital. He or she can receive. Human capital can be deﬁned as an individual’s general level of skills or more broadly as the knowledge. in school or at work. Both types of capital have to be produced by sacriﬁcing some of today’s output of ﬁnal goods. further education etc.”1 In short. social peace. 180) traces the decline of the Spanish empire in the 16th century to the catholic church stiﬂing education by. It can rise ahead of formal school education (parents) and continue to increase afterwards (experience. However. 156). Both are investments in future output. There can be no research without human capital and not much innovation without research. but they are extremely diﬃcult to quantify. Similar to additional physical capital (machines). Somebody enrolled full-time in school today cannot contribute to today’s GDP. additional human capital raises the productivity of raw labor. who had been so important for the intellectual life. higher human capital leads to more output per hour worked.
In this model. It is not surprising that the survey by Sianesi and Reenen (2003) concludes that the empirical literature is still largely divided on whether education aﬀects the long-run level or the growth rate of the economy. As early as 1890. Alfred Marshall noted that “the most valuable of all capital is that invested in human beings. The usefulness of the wheel is a simple example. while a high level of human capital explains a high level of income. The education sector produces new human capital with the help of existing human capital (teachers). These spillovers and feedback eﬀects of non-rivalrous ideas make human capital so special and can lead to market failure with too little investment in human capital because the individual can only reap part of the returns on his own investment.and macroeconomic theory The link between human capital and income should be obvious. The interpretation of these regressions is not always clear.2. As is the case particularly in the openness literature. especially during the last decades. In addition. Therefore. Robert Lucas modeled the link between human capital and economic activity by splitting the economy into two sectors.and macroeconomic theory 69 or another person can apply the same idea in his ﬁrm. As outlined in chapter 2.1 Micro.1 Micro.1 on page 79 reviewing the empirical literature often avoids the attribute of level or growth. Economic policy that raises the rate of growth of human capital will lead to higher rates of GDP growth.” And Benjamin Franklin was aware that “investment in education pays the best interest. Convergence regressions with initial income and the level of education on the right-hand side may not uncover the unconditional eﬀect of the level of education on GDP growth.” The theoretical models of the impact of human capital on economic activity have become increasingly sophisticated over the past centuries. Rich countries today are rich because they have accumulated know-how for many generations and passed it on to the next. transmission and development of knowledge. table 6. . 6. while the ﬁnal goods sector uses both human and physical capital as inputs. Section 6. As Dowrick (2004) puts it. the empirical human capital literature does not always clearly distinguish between eﬀects of the level of human capital and of its growth rate. the measurement error in the Barro-Lee data makes results derived from that dataset highly questionable especially when using ﬁrst diﬀerences.6. a rise in human capital leads to a rise in national income. If we were to forget this and other advances.” Gary Becker reﬁned and deepened these insights and coined the term “human capital” with the title of a book published in 1964. all “economic activities depend on institutions that encourage the preservation. income levels would plunge.1 will show that proper handling of the data makes this confusion disappear. Education is an investment.
In this view. Microeconomic studies tend to ﬁnd that an additional year of education raises an individual’s wage by between 5% and 10% depending on country or study subject . his work experience Ri plus a constant ci and a random error i .70 6 Human capital An alternative view is that a high level of human capital allows for a high growth rate of GDP. the same should be true for whole economies.1 and move all variables from individual levels to countrywide measures. (6. 6. post one of the highest rates of income growth given its high level of human capital. Germany (unless it gets a lot wrong in other areas) should.80 for all 40 countries. for example.1. .a useful benchmark for macroeconomic studies. we quickly arrive at a macroeconomic function. In these micro studies the economy-wide state of technology and the capital stock are identical for all individuals. The classic contribution by Mincer (1974) assumes that the only opportunity cost of attending school is foregone wages and that the increase in wages generated by this extra schooling is constant over time.1) ln Yi = ci + φSi + βRi + i The coeﬃcient on Si is interpreted as the rate of return on investment in education. This is particularly the case in the analysis of human capital and income. If macroeconomic returns exceed microeconomic returns this may be indicative of positive externalities. If we replace wages with national income in equation 6. The survey by Psacharopoulos and Patrinos (2004) reports that an additional year of education leads to a 10% higher wage on average in a large sample of countries. However. This point of view is highly controversial on theoretical and empirical grounds . there is no clear-cut empirical evidence for his thesis. Some researchers claim that individuals only go to school to signal their potential future employers how motivated and capable they are. Paul Romer stirred up a lot of attention in the late 1980s with his model of knowledge spillovers. they do not really learn much in school that would increase their productivity and innovative capacity. as labor economists have long analyzed the impact of education on an individual’s earnings.64 for the 21 rich countries analyzed in this study and of 0. In this case. If individuals with education earn more than those without it. the natural logarithm of an individual’s wage or income Yi is proportional to the level of schooling years he completed Si . A simple least squares regression of the natural logarithm of per capita GDP on the level of the years of education leads to an R2 of 0.1 Microeconomic analysis: labor economics Using insights from other branches of economics can be very helpful for growth economists. where the stock of knowledge determines the growth rate of GDP.and I will not use it below.
8% depending on the level of education. The other camp sees the level of human capital aﬀecting the growth rate of GDP per capita. My empirical results below support the ﬁrst camp.and macroeconomic theory 71 The lesson that growth economics has taken from these micro studies is that there is a signiﬁcant positive link between education and income .2 Macroeconomic models with diﬀerent conclusions As indicated above. Hall and Jones (1999) use the Mincerian equation to derive the level of human capital H = eφS L where φ is between 13. the standard physical capital K plus human capital H. For example. The augmented technical progress function in the Kaldor model in equation 2. In the . exogenous fraction of output in each of them with the same technology.and the level explaining the level.4% and 6. but schooling in levels as in equation 6.6. 7) use H = (1. Bosworth and Collins (2003.5%. the production function 2. With constant returns to scale and laboraugmenting technological progress. while Klenow and Rodriguez-Clare (2005) use 8. The experience dimension tends to get neglected because on a macroeconomic level the average work experience of the overall labor force does not change much (except through an increase in life expectancy that increases the working life as well). p. One camp sees the growth rate of human capital aﬀecting the growth rate of GDP per capita .and that the most appropriate measure for human capital should be the years of schooling. In addition. macroeconomic models developed over the past 20 years come to two diﬀerent conclusions on the eﬀect of human capital on income.1 Micro. the theoretical insights are the same as in the standard Solow model.07)S as measure of human capital.1.2) β α Yt = Kt Ht (At Lt )(1−α−β) This so-called augmented Solow-model was introduced by Mankiw. where income enters in log form.1. Romer and Weil (1992) and used for example by Bassanini and Scarpetta (2001) and Cohen and Soto (2001). I will take account of this generally accepted speciﬁcation by using the log-linear form in my empirical model. There are two types of capital.21 already made use of this log-linear relationship. Both are accumulated by investing a constant. The easiest way to introduce human capital into a model of long-run economic growth is to simply add it as another input factor into a production function. assuming a 7% return on each year of education. 6. a natural starting point for macroeconomic analysis should be a log-linear relationship. The ﬁrst camp consists primarily of two models. Similarly. While this model allows a richer empirical modeling.1 becomes: (6.
which raises the question what a fraction of inﬁnity is.72 6 Human capital long run. p. the Lucas-Uzawa model was a major step forward and I will use its conclusions below. While life expectancy in advanced countries increases by around 8 years every 20 years. As shown in chapter 2. this model is one of endogenous growth even if the actual growth rate is again determined by a set of (ﬁxed) parameters. In addition. Therefore. There is no need for a manna-from-heaven rate of technological progress g. while the remainder is devoted to producing ﬁnal goods by combining physical and human capital. Another assumption is that individuals have inﬁnite lives. 206ﬀ) or in chapter 10 of Aghion and Howitt ) is that GDP. the Lucas-Uzawa model provides clearer policy recommendations and allows more ﬂexibility. human capital and physical capital are all driven by the same process. in percentage terms the rise in education was roughly twice as strong as the rise in life expectancy. In contrast to the exogenous rate of the augmented Solow-model. the depreciation rate. . in reality we observe signiﬁcant rises in life expectancy. Ignoring the possibility of externalities in Lucas’ production function. the productivity of the education sector and the relative risk aversion. the testable hypotheses for empirical work are the same as in the augmented Solow model: GDP. for example by changing the rate of time 2 Taking this broadening to the extreme leads to the AK model. human capital and physical capital all grow at the same constant rate. The measure of capital is just deﬁned more broadly. where a broad measure of capital is the only input into production and where there are no diminishing returns to capital anymore. This rate is determined by the rate of time preference. education has increased by “just” 2 years over the past 20 years on the OECD average.2 A much richer theoretical model is the Lucas-Uzawa model (Lucas ) brieﬂy alluded to in chapter 2. the result (derived in detail in Frenkel and Hemmer (1999. this setup would preclude a rise in a measure of human capital such as the average years of education over time. so they should be pair-wise cointegrated. one of the diﬃculties is the assumption that a constant fraction of time is spent in education. which increase the incentive to invest in education (see Cohen and Soto ). where a deviation of GDP below the level indicated by human capital would point to stronger GDP growth in the future. When taking the Lucas-Uzawa model literally. This ﬁts well into my empirical framework outlined in chapter 12. A constant fraction of the labor force’s time is spent accumulating human capital. which uses a more complicated production function for human capital in a two-sector economy. GDP and each type of capital grow at the same exogenous rate g. They argue that a high ratio of human capital relative to income per head is a prerequisite for strong GDP growth. Still. Nevertheless. The model of Azariadiz and Drazen (1990) complements this analysis and points to the use of error correction models. Under constant life expectancy.
This interpretation seems to be most in line with observed developments. The production of knowledge can be captured in a function such as: (6. faster GDP growth could be a result of a faster increase in human capital or in the share of people doing research. However. GDP can continue to grow strongly even if human capital does not grow anymore. The underlying assumption is that there is an inﬁnite universe of ideas waiting to be discovered. This camp draws on Nelson and Phelps (1966) and was advocated most prominently by Romer (1986. 1998). Jones (1995b) argues that researchers may also be stepping on each other’s toes or may just reinvent ideas (“ﬁshing out”). but that the pace of progress need not accelerate over time. One can reconcile these two views by arguing that standing on the shoulder of giants allows researchers today to discover ever more complicated ideas and products. p. My estimates in chapter 12 will show a highly signiﬁcant cointegration between the level of GDP per capita and the level of human capital per capita. rejecting models of the second camp. S99) who concludes that ”an economy with a larger total stock of human capital will experience faster growth. given the stationarity of GDP growth. 1990) and Aghion and Howitt (1992. this cannot be the case.3) ˙ At = θLA. The larger the set of known ideas. The second camp of theoretical models on human capital argues that the stock of human capital determines the growth rate of GDP. . The second camp would expect the level of human capital and the growth rate of GDP to be cointegrated. In this case. Here. Models focusing on spillovers and externalities were repeatedly criticized for generating models with unrealistic scale eﬀects (see chapter 2) in the sense of Romer (1990. the existing stock of an economy’s disembodied knowledge (or human capital) determines the pace at which new ideas or knowledge are generated.” Both camps of models generate testable hypotheses. the easier it becomes to make additional discoveries (“standing on giants’ shoulders”). That is. According to the ﬁrst camp. there are also good reasons to think that it gets harder and harder to make new discoveries the more there have been made already. However.1 Micro. According to the second camp they should not be cointegrated: a deviation of GDP above any scaled level of human capital will not be corrected by GDP coming down in the future. the more neighboring ideas there are.6. time series for the level of GDP and human capital should be cointegrated.t Aν t The change in technology depends on the number of workers in the knowledge-producing sector LA and on the level of existing knowledge A.and macroeconomic theory 73 preference. The fraction of time devoted to schooling (and other parameters assumed exogenous in the Lucas-Uzawa model) may change over time. Lucas (1988) also allows for the possibility of human capital externalities.
the statistical basis for analyzing that link quantitatively is rather weak. in a country with a high stock of human capital. say 12 years of education.2 Measures and empirical analysis While the importance of human capital for a nation’s income level is theoretically well established. Barro (1991. say 5 years of education. Romer and Weil  and Levine and Renelt ) and in the ﬁrst applications with panel data (e.74 6 Human capital 6. Miller and Sala-i-Martin (2004) and Hauck and Wacziarg (2004). Similar calculations can be made for primary or tertiary education. A better measure is attainment rates. For example. which measure what share of the population has completed a certain level of education. but the same is not (yet) the case for human capital.g. However. rather than for the initial level. which compares the number of people going to secondary school in a given year with the number of people in the typical age group. 420) already noted that: ”One problem with the previous result is that the 1960 school-enrollment rates could be proxying for the ﬂow of investment in human capital. where the enrollment ratios are just high enough to keep human capital constant. secondary and tertiary) for each country. enrollment rates are not useful measures of human capital.g. Either we need the information about the initial stock and combine the two measures to get a sense for the future path of human capital.” However. Loayza  and Caselli. Secondary enrollment is the measure used in the early empirical growth literature in the cross-section world (e. p. which . It is standard practice for national statistical agencies to publish an estimate of their country’s stock of physical capital. there will be at least three ratios (primary. In sum. Sometimes enrollment is used to measure the stock of human capital. For the forecasting model I am interested in exactly these diﬀerences in the paths of human capital across countries.1 at the end of this chapter). this assumption has nothing to do with reality. the same 70% secondary enrollment rate may not even be enough to maintain the initial level of human capital. Esquivel and Lefort  see table 6. However. On the other hand. It was still used heavily even in the early 2000s for example by Doppelhofer. Or we have to make the extremely strong assumption that all countries are on the same steady-state path. enrollment rates are not a useful measure of either the stock or the ﬂow of human capital. a 70% secondary enrollment rate will lead to a signiﬁcant rise in the workforce’s average education level over time. Barro . where the paths of human capital are very diﬀerent across countries. in a country with a low stock. In other words. sometimes to measure the ﬂow. The same 70% secondary enrollment rate in two countries may go hand in hand with diﬀerent paths of the stock of human capital. A measure frequently used in the past is enrollment in secondary education. Mankiw. enrollment data alone tell us nothing about either the level or the growth rate of human capital.
especially in early datasets. De la Fuente and Dom´nech (2000) incorporate a greater e amount of national information to avoid implausible breaks in the data. Bassanini and Scarpetta (2001) .1 Best measure: years of education Over the past decade a more useful measure of human capital has gained an increasing following: the average years of education of the working age population as a proxy for the average educational attainment level. They argue that the ﬁnding of insigniﬁcant or sometimes even negative coeﬃcients in Benhabib and Spiegel (1994) or in Pritchett (2001) stems mostly from the use of Barro-Lee data which include signiﬁcant measurement error. These data are often referred to as ”average years of schooling”. Therefore. A combined measure is the average attainment level. De la Fuente and Dom´nech (2000) show that their dataset leads to signife icant positive coeﬃcients on human capital both in a levels speciﬁcation and in a ﬁrst-diﬀerence speciﬁcation. Barro and Lee (1996) extend this dataset to 1995 and later in Barro and Lee (2001) to 2000 and to measures for the age group 15 and over as well as 25 and over. Portela et al. The remaining 60% of the observations are derived using the perpetual inventory method by using enrollment data.2 Measures and empirical analysis 75 makes each individual measure less useful for empirical analysis. (2005) also emphasize the measurement error in the Barro-Lee data that stems from the calculation of non-census observations using enrollment data.2. These datasets have been used in the majority of empirical studies of human capital over the past 13 years as table 6.or census-based data as reported by UNESCO on the type of education that individuals have completed. In de la Fuente and Dom´nech (2002) they argue that their dataset e has the highest signal-to-noise ratio of all available sets. For 40% of the country-year combinations they used survey.” One advantage of this measure is that it is the same as the one used in the microeconomic studies outlined above. De la Fuente and Dom´nech e (2000) also add information on education levels that were attended but not completed. It covers 129 countries at ﬁve-year intervals between 1960 and 1985 for the population aged 25 years and above. de la Fuente and Domen´ch (2000) detect signiﬁcant measuree ment error in the Barro-Lee data for 21 developed countries. the measurement error in macroeconomic data appears to be large. but that could be misleading if it suggests that university level education is not included. while the Barro-Lee dataset leads to lower or (in ﬁrst diﬀerences) insigniﬁcant coeﬃcients.6.1 on page 79 shows. I will use the term ”education” throughout this study even if the original source uses ”schooling. Their time series for the population aged 25 years and older appear much smoother and more plausible than the Barro-Lee series. 6. They suggest a systematic way of correcting these errors using the time elapsed since the last census. The path-breaking work by Barro and Lee published in 1993 made a comprehensive dataset of average years of education available for the ﬁrst time. However. Unfortunately.
While this may be a valid concern over the very long run. the years of education in 2003 varied between 4 in Nigeria and 12. Over the 10 years to 2003 human capital rose by 2. Intermediate data were interpolated by spline. witness the fact that the average years of education jumped by 1. Over the 10 years to 2003 human capital rose by between 0. The return on human capital is estimated at 8% for each additional year of education. in the future the data are likely to include these aspects of education as well. with a mean of 12. Cohen and Soto (2001) ﬁnd a signiﬁcant positive eﬀect of human capital accumulation on growth. there are no signs of a leveling oﬀ in the data yet (with the exceptions of Germany and Switzerland) and there is a lot of . Since especially the second element is becoming increasingly important over time in advanced economies.1 years in Switzerland.76 6 Human capital use the de la Fuente and Dom´nech dataset and the pooled mean group panel e technique to ﬁnd that an additional year of schooling raises output by around 6% in the long run. they conclude that there are no externalities to human capital. For Taiwan. In the rich countries.2 in Portugal and 13.2 years in Israel and 2. The average 10-year change was one year. Since this is similar to microeconomic return measures.6 years in Singapore. Another interesting question is whether there is a ceiling or optimum for the number of years spent in formal education.2. This was built on the work of de la Fuente and Dom´nech (2000). For the 19 emerging markets I use the dataset from Cohen and Soto (2001) with observations every 10 years between 1960 and 2010. Current measures of years of education do not include learning-on-the-job or lifelong learning. I extend the dataset through 2003 using the OECD’s Education at a Glance of 2005.2 at the end of this chapter.1 and 6. The two datasets diﬀer in the age groups covered (25-64 vs. 15-64) and in the frequency of observations.8 years in Ireland and Spain but just 0.8 in Norway and the USA. For the 21 rich countries I use the dataset from 1971 to 1998 published in Bassanini and Scarpetta (2001) in the annex table.7 in South Korea with a mean of 8. The jump was even larger in New Zealand with two years. the years of education in 2003 varied between 8. For my purpose it is comforting to note that lifelong learning is highly correlated with the highest formal degree attained.2. I will use two diﬀerent datasets on the average years of education of the population aged 25 to 64. For the emerging markets Cohen and Soto (2001) conduct similar adjustments for the population aged 15 to 64 at the beginning of each decade from 1960 to 2000 plus a projection to 2010 for 95 countries. The datasets are plotted in ﬁgures 6. In the emerging markets. The average 10-year change was 1.2 years in the USA between 2002 and 2003 according to the OECD’s Education at a Glance issues of 2004 and 2005. I have to use the Barro-Lee data. who in turn e used the Barro-Lee data as a starting point.2 years. There may still be considerable measurement error. which I extrapolate.
since all economies consist of highly heterogeneous actors. So far this combination has not been used in the literature. Another problem is that the data on average years of education treat a school year in Japan just as one in Egypt and they treat a year of primary school just as one in university. for a given point in time there is a high correlation between the level and the quality of education suggesting that omitting the quality dimension may not be too detrimental. Their cross-section estimates show a signiﬁcant explanatory power beyond schooling quantity.2 Measures and empirical analysis 77 room in advanced economies before.1). the best available measures of human capital . there should be room for skilled and unskilled workers in every economy. one could also measure the value of an economy’s human capital using cost-based or income-based approaches. And the most appropriate estimate uses non-stationary panel techniques. The Centre for the Study of Living Standards publishes time series on cost-based values of human capital. these data are not available in long enough time series. Furthermore. What really matters for average income is the average level of education.ca . 50% of the working population have a university degree and most others a high school degree. Therefore. the quality of education may diﬀer signiﬁcantly across countries as the OECD’s PISA tests repeatedly show. Vandenbussche et al. However. However. the income-based measure requires as an input something that growth empirics tries to identify: the rate of return on human capital. 3 www. The log-linear speciﬁcation seems to have gained a wider following in recent years as the fourth column of table 6. The most widely used data are those from Barro and Lee (see table 6. Therefore. In sum. I do not consider a ceiling as binding over my forecast horizon until 2020. Gibson and Oxley (2003) survey these methods and highlight that on all measures the stock of human capital greatly exceeds that of physical capital. and Cohen and Soto (2002) for the emerging markets. countries with high salaries of teachers need not necessarily produce higher quality workers. the average human capital data do not distinguish between diﬀerent types of education.are those from de la Fuente and Dom´nech (2002) upe dated with the OECD’s annual reports Education at a Glance for the OECD countries. Hanushek and Kimko (2000) present a dataset on labor-force quality that combines achievement tests for diﬀerent cohorts.6.1 shows.with all the remaining deﬁciencies . In addition. Years of education are a rather crude and simple measure of the stock of education.csls. Le. The most appropriate empirical speciﬁcation is a log-linear relationship between the logarithm of the level of GDP and the level of years of education. In particular. In addition. say. the focus of this study has to be on a stock-based measure as outlined above.3 However. Alternatively. (2006) illustrate why skilled human capital is more important the closer an economy is to the technological frontier. The same is true for the International Adult Literacy Survey (see Coulombe et al ).
6.1. 6. Average years of education in 21 rich countries 14 Korea 12 10 8 6 India 4 Nigeria 2 0 1971 1976 1981 1986 1991 1996 2001 Fig.2. Average years of education in 19 emerging markets .78 14 6 Human capital years Germany 12 10 Spain 8 Portugal 6 4 1971 1976 1981 1986 1991 1996 2001 Fig.
panel 6. Primary enrollment 1960 Years of education Human capital stock School quality Years of education Primary enrollment 1960 Years of education Secondary enrollment Years of education Years of education Years of education Years of education Years of education 15-64 Secondary enrollment Years of education Years of secondary schooling Years of education Higher education enr. This study Secondary enrollment 1960 Secondary enrollment Secondary enrollment Secondary enrollment Years of education Secondary enrollment Primary and sec.Table 6. enrol. males Secondary enrollment Years of education Years of education Years of education United Nations Barro (1991) UNESCO MRW (1992) Barro-Lee (1993) Barro-Lee (1993) Barro-Lee (1993) Barro-Lee (1993) Barro-Lee (1993) Own data Own data Own data Barro-Lee (1993) DD (2000) WDI DD (2000) Barro-Lee (1996) Barro-Lee (1993) Barro-Lee (1993) Own data UNESCO CS (2001) Barro-Lee (1996) BL (2001) & CS (2001) Barro-Lee (1993) Barro-Lee (1993) Barro-Lee (2000) BL (2001) & CS (2001) Barro-Lee (2000) Barro-Lee (2001) WDI Barro-Lee (1993) corrected BL (2001) DD. Ley & Steel Bassanini & Scarpetta Bernanke & Gurkaynak Bassanini et al. CS & EaG Positively related to growth Fragile Strong positive Not shown. Barro-Lee (2000) is WP version of Barro-Lee (2001). Source Impact Method Year Author(s) ln ln ln ln ln ln linear linear linear ln linear ln linear ln linear linear linear ln ln linear linear 1991 1992 1992 1994 1995 1996 1997 1997 1999 2000 2000 2000 2001 2001 2001 2001 2001 2001 2001 2001 2001 2002 2002 2003 2004 2004 2004 2004 2004 2004 2005 2005 2005 Barro Levine & Renelt Mankiw. pos. GMM & LSDV Cross-section Cross-section & t Cross-section Cross-section Cross-section & panel Cross-section & panel 2-stage nonst. in ﬁxed eﬀects No signiﬁcant link Negative relation to GDP Small positive Positively related to growth Small explanatory power GDP causes human capital Sign. positive on top of quantity Signiﬁcant positive Weak positive Signiﬁcant positive Signiﬁcant positive Signiﬁcant positive Signiﬁcant positive eﬀect on growth Pos. related to growth Signiﬁcant positive impact Positive impact of level on growth Signiﬁcant positive impact Signiﬁcant relation with growth Small positive Positive impact of level on growth Growth and level are signiﬁcant Signiﬁcant positive Cross-section Cross-section Cross-section Panel Pi Panel LSDV Panel GMM Cross-section Cross-section Cross-section Cross section Cross-section Cross-section Cross-section PMG Cross-section PMG Panel GMM Cross-section Cross-section Cross-section Cointegration System GMM Panel GMM Cross-section Cross-section Cross-section CS. Romer & Weil Loayza Islam Caselli. 1961 Primary enrollment 1960 Secondary enrollment Years of education Years of education Years of education. Doppelhofer et al.2 Measures and empirical analysis 79 ”Spec. DD (2000) is Domenech and de la Fuente (2000) and CS (2001) is Cohen and Soto (2001).1. MRW (1992) refers to Mankiw. Hauck & Wacziarg Chen & Dahlman Batista & Zalduendo Barro & Sala-i-Martin Ianchovichina & Kacker Benhabib & Spiegel Portela et al. Easterly & Levine Krger & Lindahl Pritchett Cohen & Soto Sarno Soto Dowrick & Rogers Bosworth & Collins Doppelhofer et al. Overview of the empirical literature on human capital Measure ln ln ln ln Spec. impact of change on growth No signiﬁcant link Signiﬁcant positive Positive in cointegration vector Signiﬁcant positive Negative relation to growth Limited positive impact No robust relation with growth Robustly. Romer and Weil (1992). Esquivel & Lefort Klenow & Rodriguez-Clare Sala-i-Martin Hall & Jones Bils & Klenow Hanushek & Kimko de la Fuente & Domenech Fernandez.” indicates the transformation of the education measure. . Some studies may use more than one measure and/or more than one method.
39) concluded that “the challenge of identifying the connections between trade policy and economic growth is one that still remains before us. Some authors argue for adjusting trade shares for gravity variables and price diﬀerences. or Lee. 3. Trying to ﬁnd a link between the level of openness and the growth rate of GDP is likely to lead to inconclusive results as I will show later.2 I will argue that the level of trade openness is a good explanatory variable for the level of income per capita . Trade policy is highly correlated with other growth-relevant policy measures such as the rule of law. 1 2 For example. and cross-section convergence regressions make interpretation of some of the results rather hazardous. They seem to try to uncover a link between openness levels and GDP growth rates. Measures such as tariﬀs. Therefore I am not concerned by ﬁndings like those of Vamvakidis (2002) that “the positive correlation between free trade and growth after 1970 is an exception”. There is no consensus on how to best measure openness and trade policy.and that the change in openness helps explain the change in income. I will avoid these diﬃculties by interpreting trade openness as an overall policy variable that captures a broad array of growth-relevant policies. The literature is not always explicit about whether it is looking for an eﬀect of openness on the level or on the growth rate of GDP. . Ricci and Rigobon (2004) that the eﬀect of openness on growth is small. black market exchange rate premia and trade shares have been proposed. p.1 Three aspects appear to be behind this challenge: 1. 2. Lee. Disentangling the eﬀect of trade policy is therefore a diﬃcult empirical undertaking. Ricci and Rigobon (2004) and Dollar and Kraay (2003).” In the early 2000s. Theory gives conﬂicting priors. while others use the raw shares.7 Openness Not long ago. Rodriguez and Rodrik (2001. papers on growth and openness included terms like “once again” or “revisiting” in their titles.
By contrast. Likewise.4 Today. The experiment ended in bankruptcy and foreigners forced their way back in. the pro-openness consensus is overwhelming among economists . Trade openness (or short: openness) measures how exposed the average ﬁrm or individual of a country is to the exchange of goods and services with foreigners. a decline in transportation costs (container shipping.even if they are aware that some groups will be hurt by free trade. Case studies are important for providing background for the analysis of the empirical link between openness and income. For example. Two of the most striking case studies on how detrimental the closing of an economy can be come from Asia. Quoted from Baldwin (2003). the Secretary General of the United Nations Economic Commission for Latin America and later the founding secretary general of UNCTAD. Korea was called the ”Hermit kingdom” in the 19th century after it closed itself to foreigners. For example. . Unfortunately. advocated infant industry arguments for all of manufacturing in developing countries. This could include cuts in tariﬀs. services. 3 4 But chapter 8 will show that growth does not simply spill over across borders.3 While Rodriguez and Rodrik (2001) focus on policies. the trade share measures the actual exchange of goods and services with foreigners as the average of the shares of exports and imports in GDP. but also governments promoting the construction of new harbors. Until about 1250 China was one of the richest countries on earth. Hausmann. The failure of import substitution strategies in many former colonies in the 1960s and the success of the Asian openers in the 1970s add further anecdotal evidence. This raises the returns on physical (and human capital). economists have to accept some of the blame for countries following the wrong strategies. The result was that income levels fell back to those of the 10th century and stayed there for centuries. Pritchett and Rodrik (2005) ﬁnd that growth accelerations tend to coincide with a signiﬁcant rise in the trade share. promoting more accumulation in the second round. Both the trade share and trade openness may be aﬀected by policy as well as non-policy changes. technology and ideas will have a positive impact on incomes because the existing amount of resources is used more eﬃciently.82 7 Openness Trade policy can be deﬁned as any measure that policy takes to promote the exchange of goods and services with foreigners. the focus here is on outcomes: anything that promotes the cross-border exchange of goods. A country’s geographic proximity to other large and rich economies would by itself also imply a higher level of openness. But the centralization of power during the Ming period was accompanied by massive regulation and in the mid-16th century the government even banned ships with more than two masts. Ben-David (1993) analyzes episodes of trade liberalization and concludes that more liberal trade systems tend to go hand in hand with higher income levels. Raul Prebisch. massively restricting foreign trade.) could lead to a rise in trade without any policy changes. ﬁber optic wires etc.
A presumed positive link between openness and GDP is derived from Ricardian trade theory. which is assumed to be exogenously given anyway. GDP is not a measure of welfare.as is now common in Europe . which emphasize the eﬀect on levels. while using the Fisher ideal index pushes it close to zero. these are important theoretical and statistical reasons why the empirical literature had such diﬃculties uncovering a signiﬁcant positive relationship between openness and GDP. Each country will produce more of the good that is becoming relatively more expensive.1 Theory: higher eﬃciency 83 7.reduces the bias. there may be an eﬀect of the level of openness on the steady-state level of GDP. opening may have a negative eﬀect on the terms of trade or on the dynamic comparative advantage. Therefore. where the movement from autarky to free trade (or any smaller movement in between) leads to an increase in welfare in spite of constant production possibilities. The use of chain-linked indices . however. In a dynamic interpretation. The reason for these complications is that usually Laspeyres quantity indices are used to calculate time series of real GDP.1 Theory: higher eﬃciency From theory there is no clear conclusion on whether changes of openness have any eﬀect on GDP at all and whether the eﬀect will be on the level or on the growth rate.5 Other models see a positive eﬀect either on the growth rate or the level of GDP. as Neuhaus (2006) points out. However. Opening would be negative for growth according to infant industry models and models of optimal protection. . Macroeconomic data do not make it possible to identify which of the many potential transmission channels is the most important. Valuing this additional production at the lower prices of the base period (and the reduced production of the other good at the high prices of the base period) leads to a decline in reported real GDP. in the neoclassical trade model an increase in specialization driven by diﬀerences in relative prices will lead to a decline in GDP using Laspeyres indices. What is necessary. all these models advocate protection at most as a short-term strategy. if changes in openness make the steady-state GDP level move over time. is some sound theoretical basis for the model. welfare gains of trade cannot be uncovered by regressing GDP on trade.7. using a wide range of arguments. 60) summarizes the main theories and references. this eﬀect is spread out over many years as an economy opens gradually. as Neuhaus (2006) highlights. But this is not necessary for a model of long-run GDP growth which is interested in the overall eﬀect. Still. However. In the simple neoclassical growth model openness has no eﬀect on long-run growth. Likewise. then there would also be an eﬀect of changes in openness on changes in GDP. It only captures the real value of ﬁnal production in an economy. 5 Vamvakidis (2002. This disagreement partly stems from the later development of endogenous growth theories. In practice. However. However. p. Under free trade each country produces more of the product that was relatively cheaper (using real exchange rates) in that country before trade.
1. p. The “size” of an economy rises with trade and so does eﬃciency and income.6 A larger market implies a higher payoﬀ of innovations. tailors and merchants in the city of Nuremberg outsourced spinning and then weaving to nearby F¨rth. 7.1 Extent of the market and specialization In the Ricardian example sketched above. other factors must be behind any positive correlation between openness and income. 3) highlighted. a larger variety of capital goods. p.. However.” And the division of labor .). which contributed to England becoming the country of the Industrial Revolution. a larger accessible stock of knowledge. endogenous growth models with their focus on scale eﬀects see beneﬁcial eﬀects of openness both on growth and on the level of GDP. more R&D eﬀort. the production possibility frontier remains unchanged. This makes it empirically diﬃcult to uncover partial eﬀects. this reasoning does not necessarily imply that the growth rate of GDP is permanently higher when there is a higher level of openness. In the u 17th and 18th century. ﬁberoptic cables etc. 7. Clearly. helped by a signiﬁcant fall in transportation costs (railroad. “The greatest improvement in the productive powers of labour [. airplanes. An open trade regime coincides with other growth-positive policies such as prudent monetary and ﬁscal policy or a solid infrastructure. this extent of the market has kept increasing over the last 500 years. In their Communist Manifesto they noted that ”The 6 Very explicit in Rivera-Batiz and Romer (1991). Specialization continues today as China and India become more integrated into the world economy. outsourcing and specialization was least restricted in England.1. they promote openness and vice versa. Strong institutions go hand in hand with openness.84 7 Openness Since we cannot expect to uncover welfare or growth eﬀects stemming from specialization due to comparative advantage when using standard GDP data. 19). Endogenous growth models of the Grossman-Helpman type can be seen as modeling specialization through the use of varieties of intermediate inputs.] seem to have been the eﬀects of the division of labour. In addition. trade can help push that frontier outwards as specialization increases an individual’s productivity.. As Adam Smith (1776. promotes learning by doing and raises the return on physical capital. 43). . where fewer restrictions applied (Landes . However. In the middle-ages. more technological spillovers etc.2 Good macro polices and more competition The literature on openness and that on institutions are closely connected. This appears to be an area where models with scale eﬀects have a point.or the ability to specialize .is limited by the extent of the market (ibid p. Marx and Engels were among the ﬁrst analysts of the links between openness and insitutions.
287) who thinks that ”the nontechnological sources of diﬀerences in TFP may be more important than the technological ones. who sees technology as ”something that is common to all countries. it compels them to introduce what it calls civilization into their midst. This includes openness to trade and to ideas from abroad. Indeed. Crucially. the Parente-Prescott model of barriers to technology adoptions has a close link to openness. A one percentage point rise in the trade share raises income per person by at least . As indicated in chapter 2 on growth theory. Today we can interpret ”becoming bourgeois” as adopting institutions that allow the eﬃcient use of the factors of production. Frankel and Romer (1999) try to disentangle the impact of trade openness from the impact of domestic institutions.. by the immensely facilitated means of communication. although existing evidence on their individual eﬀects suggests that both are important. sound domestic political and ﬁnancial institutions and so on. 15). barriers to technology adoption are indeed important for understanding international diﬀerences in income levels. Thurow (2003) emphasizes that the biggest danger of globalization is to be left behind.1 Theory: higher eﬃciency 85 bourgeoisie.. on pain of extinction. Bassanini. both domestic technology barriers and foreign trade barriers require policy decisions. Frankel and Romer (1999. p. in this case great prudence is required when interpreting any regression coeﬃcients or when deriving policy conclusions: a change in openness likely requires a change in institutions for the full eﬀect on GDP to materialize.7. even the most barbarian. Scarpetta and Hemmings (2001) explicitly use trade openness as a policy variable.” This does not diﬀer much from Solow (2001. I will focus only on one of the two: openness. They ﬁnd that countries that are large. far away from other countries and landlocked trade relatively little with other countries.. draw all. Given the lack of long time series on measures of institutional quality.. Parente and Prescott use a similar deﬁnition of technology as Lucas (1988. [. Similarly. Dollar and Kraay (2003) address this point with a variety of econometric techniques to conclude that “the cross-sectional evidence is not very informative about the relative importance of trade and institutions in the long run. i. 394) conclude that trade indeed raises income. by the rapid improvement of all instruments of production. Both may be determined by third factors such as geography or history. to become bourgeois themselves” (quoted from Sachs and Warner . This should not come as a surprise. especially in developing countries. Barriers to domestic production can only prevail if the more productive competition from abroad is kept out.” The conclusion from panel analysis is likely to be the same. 5). to adopt the bourgeois mode of production. It is next to impossible to distinguish what came ﬁrst: good institutions or openness. p.” In sum.] something whose determinants are outside the bound of our current inquiry.] It compels all nations. However.e. p. they may control the technological ones. [. nations into civilization. But the fact that these geographic characteristics should impact income only through the openness channel can be used to identify the partial eﬀect of trade on income. p.
(2) the series have to measure the theoretical concept (here: actual openness). However. imports may introduce new machines that cannot be produced at home. importing ideas and technology may also boost GDP at home. Poor countries do not need to create new ideas to improve their standard of living. the same criteria as for the other variables have to be satisﬁed as much as possible: (1) time series at an annual frequency have to be available. the adoption of technologies developed elsewhere still requires some physical and human capital in the adopting country. 7. . However. (3) the measure should be reasonably easy to forecast and (4) there has to be an empirical link to per capita GDP. 5). For example.86 7 Openness one-half percent. (1994) trade restrictions raise the price of imported intermediate goods. If cross-country comparisons of levels are needed. it is worth brieﬂy discussing other measures used in recent empirical analyses.2 Measuring openness Several measures of openness have been used in the empirical growth literature. However. In a very diﬀerent framework. Before presenting the most appropriate measure in detail.3 Additional inﬂuences of trade on income In addition to these two important links between trade and income. They only have to apply existing ideas to the production of goods and services (Parente and Prescott . 7. Societies that generate and tolerate new ideas are likely to have higher income than countries stuck in their own past. Irwin and Tervi¨ (2002) show that these results o are not robust to inclusion of the distance from the equator. There is an even larger disagreement here than on how to best measure human capital. In the model of Easterly et al.e.1. This theory may have been relevant in the 1960s and 1970s when many countries could not aﬀord to purchase imports because of lack of foreign currency. these constraints have become less binding on average over the past decades. They also discuss the fact that high incomes in partner countries lead to higher bilateral trade. the average share of exports and imports in GDP. For my purpose. This section reviews these measures and critically evaluates them. These two are probably highly correlated: countries with the right institutions and incentives to innovate at home are also those open to learning from abroad. the best available measure is the time series of the trade share. For inclusion in my forecasting model. The tolerance and openness towards new ideas has both a domestic and an international dimension. i. p. it is best to adjust this measure for diﬀerences in relative population size. which depresses the rate of GDP growth. post-Keynesian models of balance of payments constraints in the growth process lead to Thirlwall’s Law that the rate of a country’s GDP growth is proportional to the ratio of its rate of export growth and to its income elasticity of demand for imports.
1 on page 94 with the overview of empirical studies) is that of Sachs and Warner (1995). one of the most widely used approaches to measuring openness (see table 7. I focus on outcomes. the exchange rate in the black market is 20 percent or more below the oﬃcial exchange rate. very few countries nowadays restrict their foreign exchange markets. a country may have zero tariﬀ rates. average tariﬀ rates of 40 percent or more. Rodriguez and Rodrik (2001) criticize the Sachs-Warner dummy for depending mostly on the state monopoly of exports and the black market premium. which are the result of a large number of factors. Several cross-section studies used the average tariﬀ rates on imported manufactures as reported by UNCTAD. openness is always a matter of degree. Trade can be . Wacziarg and Horn Welch (2003) present updated values of the SachsWarner dummy for 2000. However. Moreover. The . but the time series only start in 1995 which is not long enough for my purpose. 7. if tariﬀ rates remain unchanged but falling transport costs allow more trade. As a result. they conclude for example that China remains a closed economy. attention turned to non-tariﬀ barriers. 12). On the other hand. This black market premium is supposed to be a measure of trade distortions. For example. there would be hardly any eﬃciencyenhancing trade despite the country’s low tariﬀ rate.restricted in many other ways than through tariﬀs. They constructed a binary variable by assessing ﬁve characteristics in 113 countries for 1970 to 1989. No country is completely open.and has been .2 Measuring openness 87 7. As a result.2 The openness dummy To date.2. A country is deﬁned as open only if none of the ﬁve characteristics applies: non-tariﬀ barriers on more than 40 percent of trade. They argue that sample selection bias makes the state monopoly variable indistinguishable from a sub-Saharan-Africa dummy. then tariﬀs would also be an inappropriate measure for openness. This is calculated by dividing total import duties by the volume of imports. A problem with this variable is that it may be a poor measure of actual trade barriers for a number of reasons.1 Black market premium and tariﬀs Initial empirical studies often used data on the diﬀerence between the oﬃcial exchange rate of a country’s currency and the exchange rate in the black market. so black market premia are mostly zero.2. However. The Heritage Foundation’s Index of Economic Freedom includes a sub-component measuring the distortions in international trade. as the debate about the European services directive illustrates. Wacziarg and Horn Welch see “a distinct possibility that the classiﬁcation of countries as ‘open’ or ‘closed’ is too crude to provide much information for growth” (p. which appears to have little to do with what has been happening there over the past 20 years. but all potential trading partners have extremely high tariﬀ rates. a socialist economic system and a state monopoly on major exports. This appears to be a useful index.7. Therefore.
the ratio between exports and imports and GDP has increasingly become the variable of choice in empirical growth analysis (see table 7. even the number of years open cannot explain diﬀerent growth experiences across the countries seen open throughout my sample period . In many cases. However. its actual trade with others might still by relatively small. Trade shares are the sum of nominal exports and imports divided by two times GDP. Over the 20 years from 1983.88 7 Openness Sachs-Warner dummy does not allow for diﬀerent degrees of openness.which includes all advanced countries. see also table 7. Therefore. the un-weighted average trade share rose by 3. Companies in small countries have few compatriots to trade with.7 But even using the number of years open is not helpful in my context. In fact. Ley and Steel (2001). True openness can therefore only be measured using actual trade data and not with barriers. International trade stemming from this reason does not diﬀer much in quality from domestic trade.3 Best measure: adjusted trade share In principle. the literature has used a compromise approach by assuming that a country is more open today the further back it was ﬁrst assessed as open: the number or the fraction of years open is used (which is equivalent to using the ﬁrst year of openness). Thailand.1. 7. so they trade more across borders. Countries opening most quickly during that period include Malaysia. A mixture of the two is used in the Trade Openness Index presented by Gwartney et al. . as is the case for investment shares (see page 55) the relative price of traded and non-traded goods is likely to change over time: the development of the real trade share will diﬀer from that of the nominal trade share.2. (2001). a country can have done all the right things to promote trade but still not trade at all.1 on page 94) and I will use it as well.1. New Zealand is an example of a country relatively far away from potential trading partners. The literature often uses ratios of real variables to construct trade shares. Since my empirical analysis 7 8 The number is used for example in Fernandez. In 2003 these shares ranged from 11% in Japan and 12% in the USA to 104% in Malaysia and 200% in Singapore. actual trade is what matters when a country tries to reap the beneﬁts from free trade outlined in section 7. So even if New Zealand abolishes all tariﬀ and non-tariﬀ barriers. This might be the case if the country is inﬁnitely far away from potential trading partner. On earth. Ireland and Austria. real trade shares rose much faster than nominal shares in the past as prices of traded goods rose less than prices of non-traded goods. Therefore.1 points in the 19 emerging markets.8 This also holds for changes in openness.4 points in the 21 rich countries and by 14. The world is an example of an eﬀectively closed economy because of a lack of trading partners. However. the fraction is used in Hall and Jones (1999). because the annual change in openness would be the same for all countries once they are deemed open.
7. but do not consider that collinearity in their regression for productivity.6. They themselves report a correlation coeﬃcient between lnrealopen and lnpopulation of -0. Another important question for the empirical estimates is whether absolute or relative changes of trade openness aﬀect GDP.2 and 7. which in their reasoning would lower productivity. Ireland and Thailand. Also. However. However. Belgium +34.g. as the case of Germany shows.show a lower . Therefore. Alcal´ and Ciccone (2004) explain the a level of productivity using both the trade share and the population size on the right-hand side. my empirical analysis and country examples indicate that the log-linear speciﬁcation is more appropriate. to allow cross-country comparisons of openness levels and to derive policy conclusions. The R2 is 0. this ﬁnding stems from the use of both the trade share and the country size in the same regression. Figures 7. As a corollary they also ﬁnd that larger countries grow faster. a country such as Japan with a high domestic price level would . Prices for exports and imports tend to be similar across all countries. Japan -16%). Should we use a log-log or a log-linear speciﬁcation? Most often. To avoid these traps. I will ﬁrst correct for size . 614). g. for example. even large countries are able to raise their trade share by more than ten percentage points within just 10 years. My size adjustment stems from a simple regression of the average nominal trade share in 1995-2000 on the natural logarithm of the population level over the same period for 39 countries (Singapore excluded) as shown in table ??. This speciﬁcation implies that a 10 percentage point rise in the trade share has a smaller impact on GDP if it starts at 70% than if it starts at 10%. However. it is helpful to adjust the trade shares for the size of the country. the log-log form is chosen (e. Bassanini.g. It is not easy to correctly deal with the nexus of country size and magnitude of the trade share. Scarpetta and Hemmings ). For example.and then proceed with estimating.25. Importantly.call the result ”trade openness” . a ﬁnding quoted by Jones (2005) in support for models with scale effects.all else being equal .3 at the end of this chapter show the trajectories of the trade openness measures in the two samples. while countries above the line get positive numbers (e.2 Measuring openness 89 will use panel methods to explore the impact of changes in trade on changes in GDP. there is a case for another adjustment to the trade share in order to account for diﬀerences in relative price levels. They illustrate the upward tendency of openness and the particularly strong development in. The trade openness levels are then simply the deviations from this regression line. The example of Ireland shows that a rise in the trade share may be very important even if the starting level is already high. However. the raw trade shares would be suﬃcient as input. they ignore that this country is highly likely to also move down the size scale. the ﬁxed eﬀects would pick up the eﬀect of country sizes.2%). They conclude that “an increase in real openness taking a country from the thirtieth percentile to the median value raises productivity by 80 percent” (p. the absolute changes of openness over time are not aﬀected by these simple intercept shifts. Countries below the line get negative values (e. Finally.
90 120 7 Openness Nominal trade share (1995-00) Malaysia 100 Ireland Belgium 80 60 40 NZ Argentina 8 9 10 11 China 20 0 ln population level (avg. Small countries with large trade shares trade share than a country with a low domestic price level. using the PPP conversion factors for the year 2000 from the World Development Indicators and dividing the trade share by these conversion factors to generate a PPP trade share (or in the Alcal´ and Ciccone a terminology a ”real” trade share) signiﬁcantly reduces the trade share in poor countries such as India and China. 615ﬀ). Subramanian and Trebbi (2004) criticize their approach for introducing a spurious correlation between openness and income. However. The unavoidable difﬁculties include the Laspeyres-measurement of GDP mentioned above and the problem of ﬁnding an appropriate and exogenous measure of openness. because a rise in productivity in the tradables sector can produce a rise in real openness (and income) regardless of the origin of the productivity increase. Now the same 10 percentage point rise in the raw nominal trade share no longer can have the same eﬀect on GDP growth. Some are unavoidable.3 Empirical debate: levels versus growth The literature on the linkage between openness and income is marred by many diﬃculties. Alcal´ and Ciccone a (2004) were the ﬁrst to promote an adjustment for this eﬀect and they provide the theoretical background (p. some are home-made. Therefore. Rodrik. . I do not adjust for diﬀerences in price levels. 7. The home-made diﬃculties include a sometimes misleading handling of levels and diﬀerences of openness and income and the accompanying interpretation of the empirical results. 7. In addition.1. 1995-2005) 12 13 14 Fig.
Given my reasoning. they claim to analyze ”the eﬀect of changes in trade on changes in growth” (p. Therefore. capital. Because of these confusions and misconceptions an extensive literature review appears inappropriate in this chapter. the more appropriate interpretation in a convergence regression like the one used by Edwards would be that the level of openness has an impact on the level of GDP. who go straight to regressing GDP growth on the level of openness. Similar confusions and conclusions can be found in Lee. Similarly.initial GDP on the right hand side. their italics) without mentioning again that they added initial GDP on the right hand side. of capital and of openness. Rodriguez and Rodrik’s (2001. Edwards’ (1998) ﬁnding that ”more open countries will tend to experience faster productivity growth than more protectionist countries” should not be interpreted as the level of openness having a direct link to the growth rate of GDP. they do not check whether there is more than one cointegrating relationship among these three variables. all that has been done is to show that openness helps explain the level of steady state GDP. When they move to explaining GDP growth. Yao and Lyhagen (2001) use the pooled mean group estimator to ﬁnd a positive long-run relationship among per capita GDP. Ricci and Rigobon (2004). However. growth is higher for the same level of initial GDP. Now. when deriving their claims that openness does or does not inﬂuence GDP growth. p. a positive coeﬃcient on openness is interpreted as showing a link between the level of openness and the growth rate of GDP. 155. Some papers are clear and helpful. 1) main question is stated as ”do countries with lower barriers to international trade experience faster growth?” and this is what they seem to look for.together with the gap to initial GDP . it is not surprising that they ﬁnd no robust link and ”view the search for such a relationship as futile” (p. Rather.7. Subtracting initial income from both sides produces the convergence regression with the growth rate of GDP on the left hand side and the levels of labor. The starting point is a production function.explains the growth rate of GDP. however. which . . or in Chang. Dollar and Kraay (2003) conclude that ”countries that trade more grow faster” although they start from a regression of the level of GDP on the level of the trade share. openness and . Fendel and Frenkel (1999) regress the percentage changes of GDP on the percentage changes in trade shares and ﬁnd a signiﬁcant relationship.crucially . Even more misleading is the fact that some papers eﬀectively use models of levels on levels. 39).3 Empirical debate: levels versus growth 91 A considerable part the literature is either not explicit about what link it is examining or is concerned with the link between the level of openness and the growth rate of GDP. Much of the confusion stems from the use of cross-section convergence regressions described in detail in chapter 2. For example. where the level of GDP is explained by the level of labor input. For example. Kaltani and Loayza (2005) who regress changes in growth on changes in openness. per capita capital and the trade share in 28 Chinese provinces in the period 1978 to 1997 with a signiﬁcant positive coeﬃcient on the trade share. However.
Chapter 12 will show that there is a cointegration relationship between trade openness and per capita GDP in both the 21 rich countries and the 19 emerging markets. This is in line with my empirical approach although Brunner does not use the powerful panel techniques.92 7 Openness Another helpful contribution is Brunner (2003). . He rejects non-stationarity of GDP growth and concludes that the level of trade openness cannot possibly impact the growth rate of GDP. who separately tests the eﬀect of trade on income levels and on income growth for 125 economies over 1960 to 1992 using time series methods. this nevertheless indicates that trade openness includes a signiﬁcant amount of information that can be exploited for long-run growth forecasts. He starts from the premise that for the level of trade to have permanent eﬀects on GDP growth. it is necessary for both GDP growth and trade to be integrated of at least order one. While not as statistically signiﬁcant and economically important as in the case of human capital. In 10 of the 40 countries I ﬁnd that GDP adjusts to deviations of openness and GDP from the cointegration relationship.
2.3. Trade openness in 21 rich countries 50 Deviation from regression line 40 Thailand 30 Singapore 20 Nigeria 10 0 -10 -20 Argentina -30 1971 1976 1981 1986 1991 1996 2001 Fig.7. Trade openness in 19 emerging markets . 7.3 Empirical debate: levels versus growth 93 50 Deviation from regression line 40 Belgium 30 Netherlands 20 Ireland 10 0 −10 −20 Australia −30 1971 1976 1981 1986 1991 1996 2001 Fig. 7.
6 Sachs & Warner (1995) Sachs & Warner (1995) Sachs & Warner (1995) Own calculations PWT 6. Hendry & Krolzig Hoover & Perez Batista & Zalduendo Rodrik et al. Overview of the empirical literature on openness Variable Years open (number) Black market premium Distortions in intern. positive eﬀect on growth Signiﬁcant positive eﬀect Modest role Signiﬁcant and large eﬀect Signiﬁcant positive eﬀect Sign. Some papers may use several measures and sample sizes.1. Ianchovichina & Kacker This study Note: Table is not comprehensive. panel 7 Openness Year Author(s) 1997 1998 1998 1998 1998 1998 1999 1999 2001 2001 2001 2003 2003 2003 2003 2004 2004 2004 2004 2004 2004 2004 2004 2004 2005 Sala-i-Martin Edwards Edwards Edwards Edwards Edwards Frankel & Romer Hall & Jones Easterly & Levine Bassanini et al. positive eﬀect on growth No additional eﬀect No signiﬁcant relationship Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Sign. trade Import tariﬀ Openness dummy Outward orientation Trade as % of GDP Years open (fraction) Trade as % of GDP Trade exposure adjusted (ln) Years open (number) Trade as % of GDP Trade as % of GDP at PPP Trade instrument Years open Level of trade as % of GDP Openness dummy Trade as % of GDP (real) Trade as % of GDP at PPP Years open (number) Years open (number) Years open (number) Trade share adjusted Trade as % of GDP (nominal) Trade share adjusted Trade share adjusted Sachs & Warner (1995) Barro & Lee (1994) Heritage Foundation UNCTAD Sachs & Warner (1995) World Development Report IMF Direction of Trade Sachs & Warner (1995) As in Beck et al. and robust positive eﬀect Strongly and robustly related Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect No eﬀect on top of institutions Signiﬁcant positive eﬀect Signiﬁcant positive link Source Impact Method Cross-section Cross-section Cross-section Cross-section Cross-section Cross-section Cross-section Cross-section Panel GMM PMG Cross-section Panel Panel GMM Cross section Cross section Cross-section Cross-section PMG Cross-section Cross-section Cross-section PCGets Cross-section Cross-section plus time Cross-section Cross-section 2-stage nonst. Fernandez.94 Table 7. OECD & own calculations Sachs & Warner (1995) IMF Direction of Trade PWT 5.6 Frankel & Rose (2002) Sachs & Warner (1995) n/a Sachs & Warner (1995) China Statistics Yearbook PWT 5.0 Own calculations WDI Robust positive relationship Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Signiﬁcant positive eﬀect Sig. . Ley & Steel Brunner Dollar & Kraay Bosworth & Collins Bosworth & Collins Barro & Sala-i-Martin Chen & Dahlman Yao & Lyhagen Alcala & Ciccone Doppelhofer et al.
As outlined in chapter 3 on GDP. who shows that the beneﬁts of spillovers decline with distance in 14 OECD countries and that the amount of spillovers is halved after about 1200 kilometers. If a country has reached a high income level relative to its neighbors. are you more likely also to experience strong growth? The lack of attention to distance in growth empirics is surprising. the more it is supposed to beneﬁt from spillovers. Similarly.8 Spatial linkages “Everything is related to everything. The aim of this chapter is to explore the importance of geographic distance for the growth of national economies. Most growth models abstract from geographic distance by eﬀectively treating countries as randomly distributed in space or as isolated islands. An exception is Keller (2002). if your neighbors grow quickly. If you have a rich neighbor. the further a country is below the level of the world leader or the world average. are you more likely also to be rich? Similarly. theoretical growth models have made use of spillovers to explain economic growth:1 A country’s GDP growth rate depends on the growth and income levels of other countries. . this literature is usually not concerned with the position of the leader in geographic space and with the distance that the spillover eﬀects have to travel. thereby driving up income there. However. Recently. but near things are more related to each other. the literature on the ”distance to frontier” focuses on technological distance to the world leader and the diﬀusion of technology from the leader to the followers. this could potentially induce it to outsource activities to low-cost neighbors. since distance features prominently in the gravity models used in the analysis of trade and migration.” This ﬁrst law of geography formulated by Waldo Tobler (1970) is frequently ignored in empirical growth analyses at the country level. 1 Klenow and Rodriguez-Clare (2005) provide a helpful taxonomy of diﬀerent models.
(2004) ﬁnd for 17 Spanish regions that “a 10 percent increase in the level of total factor productivity of the neighbors raises the level of technology in one region by almost 3 percent. (2000) ﬁnd for 142 European regions that if any particular region is experiencing a high long-term growth rate it is more likely that the regions closer to it will also experience a high long-run growth rate. the ways of constructing spatial GDP and to the diﬀerent estimation techniques used. and show that this correlation reﬂects more than the existence of common shocks.96 8 Spatial linkages Spatial econometrics is a subﬁeld of econometrics that deals with the treatment of spatial interactions and spatial structure. for example across US states (e. spatial aspects have not been explored in a panel framework in the growth literature. The focus in this chapter will be on spatial dependence. although slope coeﬃcients may also diﬀer. Lim ) or European regions. my analysis in section 8.4 I attribute the diﬀerences in conclusion to the diﬀerences in country samples. which relates deviations from the mean of own values to the values of neighbors. . Ramirez and Loboguerrero (2005) use a pooled cross-section approach for 98 countries to conclude that ”spatial relationships across countries are quite relevant. Good economic performance is mostly home-made and does not simply spill over from across the border.” Several studies look at regional economic growth.g.” They argue that a country’s economic growth is aﬀected by the performance of its neighbors and that ignoring spatial interrelations can result in model misspeciﬁcation. Moreno and Trehan (1997) ﬁnd that ”a country’s growth rate is closely related to that of nearby countries. See Lim (2003) for a brief overview.2 Spatial dependence refers to observations in one country depending on the observations in other countries. active foreign intervention by foreigners through colonialism or war would aﬀect activity. 2 3 4 Anselin (2001) provides a comprehensive overview.2 in this chapter and more comprehensively in chapter 12 does not support these conclusions. To my knowledge. and tend to conclude that there are signiﬁcant cross-regional externalities.3 Moran scatterplots are the graphical counterpart that will be used below. For example. More generally. The simplest example is a diﬀerent constant (ﬁxed eﬀect) that is correlated with spatial aspects. However. The most widely used measure is Moran’s I statistic (introduced in Moran ). Certainly.” Cannon et al. Vaya et al. Similarly. Another issue is spatial heterogeneity which refers to parameters diﬀering depending on the spatial characteristics of a country. spatial autocorrelation simply examines the coincidence of value similarity with locational similarity.
Continent dummies were used for example in Barro (1991). hot regions. Landes (1999) argues that the favorable climate was one of the reasons why the Industrial Revolution happened in Europe and not elsewhere. It will be captured in the ﬁxed eﬀects in the panel analysis.1 Spatial economics .2 Relative location: rich neighbors A potentially more important inﬂuence on a country’s economic performance may be its relative location to others and the performance of these neighbors. Acemoglou. climate zone or continent. This would imply that strong growth in neighboring countries would be followed by stronger growth at home once these growth-positive policies are implemented. Authors using absolute latitude or the distance from the equator include Sachs and Warner (1995).1 Spatial economics . .location matters The literature on economic geography distinguishes between absolute and relative location.1. For example.1 Absolute location: latitude and climate The most obvious reason to focus on absolute location is the fact that agricultural productivity is lower in warmer climates. if there is political instability or even a civil war in a large neighboring country. Relative location refers to the location nearer or further away from other countries. However.location matters 97 8. Trade may slow. exchange of ideas may fade.1.5 Absolute location refers to the location at a particular point in space. Lee. Plenty of rain and forests allowed independent. then negative spillovers may depress activity also in the country of interest. military spending may rise.8. Technological progress was higher than in dry. In addition. But since absolute location is ﬁxed. successful neighbors may provide a good example to emulate. Likewise. it will not be of much help in an attempt to explain and forecast GDP growth rates. All this may be helpful in explaining diﬀerences in income levels across countries. Pesaran and Smith (1997). 8. A country far away from the instable country will suﬀer far less than an immediate neighbor. severe climate may lead to lower labor input (hours) because recovery periods have to be longer. the empirical diﬃculty is that the other drivers of growth such as human capital and trade openness would also accelerate and 5 Abreu. small economic units to prosper and compete. de Groot and Florax (2004) provide a useful survey and show that the literature on spatial econometrics and growth only took oﬀ around the year 2002. because winter frost is essential for rich soils and for keeping diseases in check. for example in a certain latitude. Miller and Sala-iMartin (2004). 8. Hall and Jones (1999) and Doppelhofer. Johnson and Robinson (2001) argue that heavy incidence of malaria kept colonizers from building strong institutions in some colonies with lasting eﬀects on income levels today.
htm . but could not ﬁnd signs of this spatial correlation. 8.98 8 Spatial linkages help to explain the ﬁnal outcome of stronger GDP growth. In the case illustrated above. For example. but it would be diﬃcult to disentangle this empirically. these distances have to be inverted ﬁrst. They expect to ﬁnd that the product of the residuals from two countries would tend to be high when the countries were close together. The resulting matrix shows that Germany has a weight of 4. This is the approach that will be followed here. where the inter-city distances are weighted by the share of the city in the country’s overall population. a country’s neighbors in my sample have to be weighted: important countries have to receive a higher weight than less important countries. However. this would give a very small role to more distant countries. while Argentina has a weight of 0.627 kilometers. having a common border is far too narrow a measure of spatial relationships. They provide a matrix of bilateral distances in kilometers for 225 countries. I construct time series of ”spatial GDP per capita”.080 kilometers. the openness variable already captures trade linkages. I use their measure distw. these inverted distances have to be standardized (”row normalized”). in line with Moreno and Trehan (1997).cepii. And since the weight of all neighbors has to sum to one for each country. An alternative speciﬁcation would be to use a quadratic instead of a linear function. It provides the distance between two countries based on bilateral distances between the biggest cities of those two countries. DeLong and Summers (1991) investigate whether the distance between capital cities has any explanatory power for the residuals from their crosssection estimates. the distance between Spain and Germany is given as 1. Since nearby countries have to get a higher weight than more distant countries. Also. Germany would have 38 times the weight of 6 http://www. or whether regions or countries have a common border (as in Lim  and Ramirez and Loboguerrero ). In this case. Another possibility is to use trade weights (as in Arora and Vamvakidis ). The ratio between these two shares equals the inverse of the ratio of the two distances mentioned earlier. I use distance data from the Centre d’Etudes Prospectives et d’Informations Internationales CEPII as described in Mayer and Zignago (2006). while the distance between Spain and Argentina is then 10.6 For my sample of 40 countries covering 86% of world GDP. The most often used and easy to implement weighting matrix uses the physical distances between the countries or their main cities.6% in Spain’s spatial GDP.74%.2 Constructing spatial GDP In order to capture the economic characteristics of a country’s neighbors. successful neighbors might be the ultimate reason for growth-positive reforms. To this end. However.fr/anglaisgraph/bdd/distances.
4 at the end of this chapter plot the series. there are some . almost 0. Egypt.0% in Japan’s spatial GDP) posted particularly strong growth over that period. This stems from the fact that Korea and China (with weights of 16.2 Constructing spatial GDP 99 Argentina in Spain’s spatial GDP if a quadratic function was used . . 7 Since I do not use all 225 countries in the weighting. The so-called Moran scatterplot in ﬁgure 8. this is probably not indicative of a causal relationship.8. 8.only much smoother as country-speciﬁc business cycles tend to get averaged out. g.2. Moran plot for the levels of GDP per capita in 2000-2003 Turning from levels to growth rates of spatial GDP per capita. Rather.1.6 percentage points above the average. Nigeria and Egypt whose actual neighbors are not in my sample will receive a relatively high level of spatial GDP.7 Not surprisingly. For example. Singapore and Japan show that it is possible to have high levels of per capita GDP despite having relatively poor neighbors.1 shows that there is a small positive correlation between the levels of spatial GDP and own GDP in my sample of 40 countries. Fig. Japan’s spatial GDP rose relatively strongly at 2. the resulting trajectories of spatial GDPs per capita look similar to the trajectories of the underlying series for GDP per capita . it is driven by the rich European countries being surrounded by other rich (European) countries. poor countries.diﬀerences across countries over longer horizons. over 1994 to 2003.6% and 8. Turkey and Mexico are examples of countries with relatively rich neighbors (at least in my 40 country sample) that nevertheless do not have a high level of GDP per capita themselves. Australia. Figures 8.92% per annum. e.3 and 8. However.albeit small .in the linear case the ratio was 6. The USA.
8. At the low end of the rankings of spatial GDP growth.c. Brazil.0% and 7. in the gravity models of migration and trade. 8. In my framework.0 0. as did Ireland and Chile in the top left quadrant. In addition. it is possible to take this into consideration in spatial analysis: Moreno and Trehan (1997) scale their initial weighting matrix with the .0 Chile 0.100 8 Spatial linkages Despite this strong growth of neighboring economies. (deviation from avg. when the home country is itself very large. Chile had to face just 1. Moran plot for the growth rates of GDP p. The absolute size of a neighboring economy does not play a role in my measure of spatial GDP. Here. In principle.9% weighted growth in its neighboring countries.0 Korea 2.0 Japan −2. these size eﬀects are already captured by the trade openness variable. Japan’s own GDP per capita rose by just 1. For Germany’s trade share it clearly makes a diﬀerence whether all neighbors are the size of Luxembourg or the size of France. However. Italy and Switzerland all faced a combination of weak growth at home and nearby (bottom left quadrant). the weak growth performance of Argentina and Brazil (with weights of 20. whatever happens in neighboring countries is unlikely to matter that much.2 Indonesia Spatial GDP growth p. A large neighbor should be more important than a small neighbor.6 India Taiwan China Fig.2.2 showing the deviations of growth rates from averages illustrates that there was no visible link between own country growth rates in 1994 to 2003 and neighbors’ growth: Japan in the bottom right quadrant had below-average own growth despite strong performance of its neighbors. It is probably more important for the German economy if French GDP grows by 3% than if Luxembourg’s GDP grows by 3%. (deviation from average) 0.0 Switzerland −4. China in the top right quadrant clearly outperformed its neighbors.0 −0.1% per annum.4 −0.6 Argentina −0. the weight of the attracting unit is an important element.2 0. 1994-2003 The Moran scatterplot in ﬁgure 8. in percentage points) 6.0 Ireland 4.2% in Chile’s spatial GDP) left its mark.c.4 0. c.0 Own GDP growth p.
as is the eﬀect of absolute distance. rich and close neighbors. Therefore my results do not support those derived from cross-section regressions for large and heterogeneous samples. The analysis of the Moran scatterplots in this chapter will be supported by the panel analysis in chapter 12.which may themselves be positively correlated across space as the Asian Tigers show. the errors from this relationship enter signiﬁcantly only in 9 out of 40 models for short-run GDP growth. While there is .3 Sum: Spatial linkages not much help 101 corresponding ratios of foreign to domestic output to also capture eﬀects of country size. poor neighbors.a long-run panel cointegration relationship between own GDP and spatial GDP.not surprisingly given the way the data are constructed . Simple correlation as measured by Moran’s I does not necessarily indicate causality. Again. if a country has large. Likewise. There is no correlation between the levels of spatial GDP and the levels of trade openness in the 40 countries in 2000-03. its trade openness measure will be higher than that of a remote country with small. for example that New Zealand is much further away in absolute terms from other countries than is Germany. I conclude that growth is mostly driven by policies and decisions taken at home . lagged spatial GDP growth enters only 3 of 40 short-run models. . I ﬁnd little evidence of a major impact of neighbors’ GDP growth on home country growth. it requires the right decisions at home. this eﬀect is already mostly captured in my openness variable . All else equal.3 Sum: Spatial linkages not much help While theoretically appealing at ﬁrst sight. Growth does not simply spill over from next door.8. 8.
000 USA 14. Spatial GDP p.000 6.000 Taiwan 12. 8.3.000 in PPP USD of 2000 Turkey 18.000 15. 8. of 19 emerging markets . Spatial GDP p.000 Singapore 9.000 Japan 10.000 1971 1976 1981 1986 1991 1996 2001 Fig.000 18. of 21 rich countries 21.4.000 1971 1976 1981 1986 1991 1996 2001 Fig.102 8 Spatial linkages 26.c.000 Australia 6.c.000 in PPP USD of 2000 22.
And second. While this lack of correlation does not imply that there is no partial eﬀect of R&D. They ﬁnd a cointegrating relationship between the domestic stock of knowledge. countries with high R&D shares tend to have high levels of income. Therefore. First. Furthermore. while the focus here is on 40 countries. constructing stocks of R&D is not an easy undertaking. forecasting both R&D resources and the stock of global knowledge is a tremendous task on its own but would be necessary for a forecasting model. so they may not be too helpful in explaining long-term GDP growth. Given my focus on growth rates of GDP. R&D shares barely changed over the past decades. Unfortunately. the correlation between the stock of R&D and the stock of human capital is very high. However. However. this would suggest the use of changes in R&D shares to explain growth rates of GDP.9 Other determinants of GDP Long-run economic growth is a highly complex phenomenon. a measure of the R&D stock could be the focus of future research. However. the ratio of R&D spending to GDP shows no correlation with the growth rates of GDP: countries with high rates of R&D spending do not grow systematically faster. domestic R&D resources and the global stock of knowledge. On the other hand. Still. Two diﬃculties prevent the inclusion of this relationship in the present study. expanding the list of variables beyond those presented generates a variety of problems. One very important candidate for help with modeling the progress of technology is spending on research and development (R&D). Using stock measures appears more promising and would be in line with the other variables suggested in the previous chapters. so that it is not clear whether there will be suﬃcient additional information given the high cost of constructing the data. their knowledge data are available for only 15 countries. it suggests the need for very detailed analysis. Bottazzi and Peri (2007) construct domestic and global stocks of knowledge using patent applications from 15 rich countries. . Either data are not available or the theoretical link is not clear or the empirical link is not strong enough. a point made clearly by Klenow and Rodriguez-Clare (2005). the measures discussed so far cannot provide a comprehensive picture of all the relevant factors.
and for my purpose the changes in their quality . A ﬁnancial system that channels resources to the most promising activities and allows risk sharing. ) and that better developed ﬁnancial markets ease ﬁrms’ external ﬁnancing constraints (Levine ). but these simply are not available. measuring these institutions . Therefore. equity market capitalization may not be a promising route as the high level of equity markets in Japan in the early 1990s illustrates. However. Rodrik. The empirical literature indicates that episodes of ﬁnancial market liberalization are followed by stronger GDP growth (Bekaert et al. would clearly be beneﬁcial for GDP. In addition. but also on GDP. In addition. In addition. Overall. Sometimes it is argued that democratic governments may not only have a positive eﬀect on non-material well-being. Another candidate is the political constitution of countries. However. I close my system at the borders marked in the preceding chapters but acknowledge that there are other elements that also deserve attention. Most empirical papers from the cross-section area use just one observation on quality for each country. these questions are not explored here although they might be the subject of future research. as the empirical growth literature has shown repeatedly. but time series would be necessary for a forecasting model. Johnson and Robinson  and North ) argues that high quality institutions are crucial for explaining the levels of GDP per capita. Subramanian and Trebbi . However. Acemoglu. In sum. new questions and variables keep coming up all the time. the assessments of institutions usually stem from investors or observers and may be subject to endogeneity bias. for example. Using.104 9 Other determinants of GDP A second candidate to consider would be the structure of a country’s ﬁnancial system. the literature struggles with how to best measure the quality of a ﬁnancial system. Tavares and Wacziarg (2001) even ﬁnd a moderately negative overall eﬀect of democracy on growth. a large body of literature (e. Rich countries may get a high rating partly because they are rich. A compromise would be to assume that there is no major eﬀect of the political setup on economic activity either way. g. the evidence is far from convincing as the example of poor but democratic India illustrates. . forecasting ﬁnancial system quality is even more challenging. it would be extremely helpful to have time series of objective measures of institutions.is a far from trivial task. with the accompanying decline in ﬁnancing costs.
the future is largely unpredictable. The basics of forecasting are outlined in Diebold (2004).S. There are things we know we know. The following sections will draw heavily on these two sources. judgments on how to deal with a host of issues have to be made. there are known knowns. The ones we don’t know we don’t know. on the basis of existing knowledge. Secretary of Defense Donald Rumsfeld at a Department of Defense news brieﬁng Feb. We also know there are known unknowns.”2 Before building a forecasting model it is necessary to review some of the basics of forecasting theory. This is particularly the case in macroeconomics where endogeneity problems abound. Singer is a biochemist. while Clements and Hendry (1999) is a more advanced exposition (summarized in Clements and Hendry ).10 The theory of forecasting As we know. Secretary of Defense was not warmly received by all listeners back then. . However. Singer (1997). i.S. I will not follow all the recommendations in that literature. That is to say we know there are some things we do not know. Academics would put it slightly diﬀerently: “because of the things we don’t know we don’t know. Forecasting often is a mixture of science and art. the general conclusion is that forecasting is extremely diﬃcult and that conﬁdence bands tend to be very wide. but it summarizes the diﬃculties of forecasting quite well. But some developments can be anticipated. 1 2 U. But there are also unknown unknowns. since forecasting is only one of the purposes of my analysis.1 This assessment by the U. 12. Unfortunately. This should help focus attention on the most relevant issues and avoid some of the traps in forecasting. e. or at least imagined. 2002.
1 The beneﬁts of forecast experiments The importance of long-run growth forecasts for many diﬀerent constituencies was emphasized in the introduction in chapter 1.3 The growth literature tends to avoid the topic of forecasting altogether. i. Successful models are those that adapt quickly to shifts in the economy . As argued in the introduction. Logical coherence requires that forecasts satisfy economic identities. Communication is easier if the underlying model is coherent. The textbooks of Jones (2002a).g. forecasting is not a key element of economic curricula and of econometric textbooks. Barro and Sala-i-Martin (2004) and Hemmer and Lorenz (2004) do not mention it.” 3 A basic econometrics textbook such as Greene (1993) has only six pages dealing with forecasting out of more than 700 in total.e.not necessarily models that are based on sound economic theory. Coherence is important for my model as well. the requirement for coherence may conﬂict with other requirements. It is not clear why a whole strand of economics avoids the forecasting issue even though the demand for growth forecasts is so strong. e. Despite its practical importance. A business cycle model where unemployment does not fall when GDP grows above trend is not coherent in this sense. growth theory may beneﬁt from forecast exercises because more focused questions may be asked and the practical relevance of the diﬀerent growth theories has to be shown. that the budget deﬁcit equals expenditures minus revenues. The panel textbooks by Arellano (2003). as the most coherent model need not necessarily be the most precise one.106 10 The theory of forecasting 10. the beneﬁcial eﬀects do not necessarily go both ways: Diebold (1998) concludes that there is no support for the belief that a greater reliance on economic theory will help forecasting. The 1998-page Handbook of Economic Growth published in 2005 does not include a chapter on forecasting.2 The characteristics of good forecasts The link between theory and forecasting is also important when communicating forecasts. which should also be useful for policy analysis. Economic coherence requires that forecasts make economic sense by obeying economic rules and regularities in the data. Few economics departments oﬀer courses in forecasting. logically integrated. consistent and intelligible. Unfortunately. . p. 10. At times. At least the benchmark book on time series econometrics by Hamilton (1994) has a full chapter on forecasting. 11) that we “should put logical and economic coherence above satisfying some statistical criteria. I side with Don (2000. This study tries to ﬁll the gap. Baltagi (2001) and Hsiao (2003) all do not mention forecasting.
reality is likely to be diﬀerent. e. 295).4 This could imply that forecasts must not change signiﬁcantly between one forecasting exercise and the next . simple. I will mostly focus on the second.6 New. More details can be found in Clements and Hendry (2003). a model that may have captured past events quite well may not be able to capture the events that will drive the future. it is important to keep the loss function of the users of these forecasts in mind. For example. i. The future rarely is like the past. great in-sample ﬁt is no guarantee for good forecasts. The general rule in forecasting is called KISS: Keep it so4 5 6 See Don (2000). Unfortunately. The loss may rise linearly with the diﬀerence between the forecast and the actual outcome. 32ﬀ. An optimal forecast should be unbiased. Or it may rise with the square of this diﬀerence: larger losses carry a more-than-proportinal weight.2 The characteristics of good forecasts 107 Another standard requirement for forecasts is stability. p. . the average error should be zero. Usually. Therefore.5 In the ﬁrst case. All econometric models are mis-speciﬁed and all economies experience signiﬁcant. unpredictable events will happen in line with Rumsfeld’s categorization quoted above. Forecasts tend to assume that the model is correctly speciﬁed and that it remains valid over the forecast horizon. In the case of a bias. the bias test will have to be along the cross-country dimension. Deviations in either direction may be equally relevant: overestimating growth may be just as costly as underestimating it. assuming a loss function where larger absolute errors matter more. p. p. As is now well known. Therefore. Since I will only produce one forecast per country. However. unanticipated shifts. When evaluating the quality of forecasts.otherwise the change requires a detailed explanation. a model forecasting 2% annual GDP growth in China in the next 5 years has next to no credibility because the forecast is so far away from the consensus following several years of GDP growth of around 9%.a clear violation of the unforecastability principle. a future loss (or gain) depends on a decision today and on the state of the world in the future. parsimonious models tend to have the best out-of-sample forecasting performance. the “mean absolute error” is the criterion to evaluate diﬀerent forecasting models. In principle. More details on loss functions can be found in Diebold (2004). regressing the forecast on a constant would help improve the forecast . In the second case the “mean squared error” would be used. The requirement of stability may also mean that forecasts must not completely contradict the current consensus view.10. I will compare the forecasts for 200620 derived at the end of this study with the countries’ past growth experience and with the forecasts in Bergheim (2005). truncated or asymmetric loss functions are also possible. 10. This implies for example that additional information should not be able to improve the forecast: the “unforecastability principle” (Diebold . On the contrary. the reverse also holds: the best forecasting models often do not capture the past development of the variable very well.
So the model has to be reasonably comprehensive. my set of forecasts assumes that the patterns characterizing the data in the past will continue to hold in the future. 29]). Therefore. shifts in the equilibrium of a cointegration relationship may lead to serious forecast errors: forecasts may be for a decline in the variable of interest exactly at a time when it rises strongly. business cycle forecasts may assume unchanged interest rates. 45). forecasts must be made conditional on a certain policy decision. p. the regression coeﬃcients are assumed to stay stable. The Lucas critique has made clear that this may not necessarily be the case in practice. “if we can never use past experience to predict consequences of some innovation in policy. The forecasts presented in this study are conditional on the forecasts for the exogenous variables materializing. Some simple models such as exponentially-weighted moving averages are very ﬂexible in capturing such shifts. For example. If these turn out diﬀerent.e. Forecasts may also be made conditional on many other factors. Therefore. as Frankel (2003) points out. while other simple models such as a linear determinist trend are not ﬂexible. A further severe problem for forecasts is that policy can inﬂuence them. which may be particularly large at the most recent observation? The trade-oﬀ for my long-run GDP model is clear. Decisions taken today may aﬀect the path of the economy and therefore the outcome of the variable to be forecast. Do they stem from the cointegration errors. Furthermore. So it is certainly necessary to analyze the sources of a model’s forecasting diﬃculties. On the one hand. it is important to clarify the assumptions made and to present alternative scenarios using diﬀerent assumptions on important but uncertain variables. On the other hand. In fact. I want it to incorporate economic theory to be able to help policy decisions. causally-relevant variables cannot be relied upon to produce the best forecasts when the model is mis-speciﬁed for a non-constant mechanism (Clements and Hendry [2001. then the forecast will be oﬀ even if the forecasting model itself is excellent. forecast failure can be an indicator of important structural changes in the economy.108 10 The theory of forecasting phisticatedly simple! The “parsimony principle” applies: simple models are preferable to complex models (Diebold . Even more. the general advantage of simple models stems from their higher ﬂexibility in adjusting to breaks. the model’s out-of-sample forecasts will be evaluated against very simple time-series models in chapter 13.” . p. Forecasts may be wrong because exogenous variables turned out diﬀerent from what was expected. Likewise. To get a sense for this trade-oﬀ. As Clements and Hendry (1999) point out. i. then we might as well give up and go home. Weak forecasting performance of a model tells us nothing about the policy implications of the model. However. Clements and Hendry identify shifts in the coeﬃcients of deterministic terms as the most important source of systematic forecast errors. poor forecasting performance relative to very simple models may damage its credibility. The policy links may still be valid. exchange rates and oil prices.
GDP forecast from the trend line is now attached to the last observation.0% growth per year.10.3 ln GDP Last actual Intercept correction 10. an intercept correction is neither necessary nor possible here.6% without intercept correction.0 Actual 9. However.02 *t 10. Clements and Hendry recommend to always use intercept correction when deterministic shifts are suspected. the error-correction components will capture the adjustment to any deviations from long-run equilibrium. I will also employ corrected forecasts in the forecast competitions. Because of the importance attributed to intercept correction in the forecasting literature. the average annual growth forecast with intercept correction would be 2.7 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 Fig. actual GDP would grow by just 1.1 illustrates the choices. In my fundamental model.1). These corrections set the error at the last observation equal to zero.8 Year 9. My fundamental models produce annual forecasts for GDP growth (rather than levels. The simple forecast model is a deterministic trend with a slope equivalent to 2. 10.3 Intercept correction and forecast combination 109 10.a. 10. The chart illustrates this upward shift in the trend line.6% p. Assuming that the error between actual GDP and the ﬁtted level disappears over the forecast horizon.1. the last actual observation for the year 2000 lies well above the ﬁtted line.2 1. as the trend model depicted in ﬁgure 10.3 Intercept correction and forecast combination An important practical issue is how to link the forecast and the actual data series. judgment enters the construction of forecasts. . Therefore. This can be interpreted as a closing of the output gap by 2005.1 lnY = a + 0. Therefore.6% per annum over the ﬁve-year horizon.0% instead of the 1.9 Forecasts 9. Intercept correction for a simple trend model Clements and Hendry (2003) show that intercept correction is a widely used and successful strategy to improve a model’s forecasting performance. a. Again. 10. Figure 10. The 2% p.
which may hurt the forecast’s public credibility. but will keep in mind that these are surrounded by considerable uncertainty. interval forecasts or density forecasts (Diebold . Usually. For the forecast competitions and to derive the forecasts over 2006-20. Another variable may be very important for the model. Another salient insight from the forecasting literature is that combining (or pooling) forecasts from diﬀerent models often leads to a forecast with lower errors and variances. One of the reasons to do out-of-sample forecasting exercises is to ﬁnd out which models perform well and which models . say. it is possible to present point forecasts. thereby introducing considerable uncertainty into the forecast of the variable of interest. However. In the context of forecasting GDP growth. Interval forecasts give a range in which the future outcome is expected to lie with.110 10 The theory of forecasting In general. In this study I will mostly use point forecasts. density forecasts. 38ﬀ). I need to generate forecasts for the exogenous variables. One exception is the Bank of England with its ‘fan charts’ on inﬂation which show a widening probability density as the forecast horizon expands. while another model might explain the long run. combining the sales forecasts from an econometric model with the judgmental forecast from the sales force stands a good chance of doing better than each of the two forecasts separately. The lesson to be learned for my question is that one should try to base the forecasts on exogenous variables that are relatively easy to predict. but it may not matter all that much for the model. one could weight the numbers from a business-cycle model and those from a long-run growth model to derive annual GDP forecasts. This trade-oﬀ already inﬂuenced the selection of the variables presented in the previous chapters. For example. In some cases it might make sense to let the weights vary with the forecast horizon: one model might be better at explaining shorter horizons. similar to combining diﬀerent assets in a ﬁnancial portfolio. I will do this by using either forecasts from other institutions (UN for population) or simple rules such as “average growth will equal the average of the past 10 years” or “average growth will equal the weighted average of past growth rates in that country and past growth in other countries”. regression analysis is used to determine whether there is any value in combining forecasts and what the weights should be. This is particularly often the case when individual forecasts are based on diﬀerent information sets. It turns out that these densities are also quite wide. it is not straightforward to produce these conﬁdence intervals. p. but extremely diﬃcult to predict. The third option. The trade-oﬀs are clear: a slow-moving variable may be easy to forecast. a 60% probability. Many examples in macroeconomic forecasting show actual outcomes well outside the calculated conﬁdence bands. is even more demanding and rarely used in practice. The uncertainty stems to a large extent from uncertainty about the path of the exogenous variables. Merely using the standard errors of the underlying equation is usually not enough because uncertainty about the forecasts of the explanatory variables will add to the overall uncertainty.
10.or if diﬀerences in behavior are properly taken care of. Therefore. but also strong views on what variables are reasonably exogenous. The advantage of the use of panel data is that we can learn about a country’s behavior by observing the behavior of other countries in addition to that of the country in question.3 Intercept correction and forecast combination 111 may be combined to produce the best overall result. I have too few observations to apply this technique. This will be the subject of chapter 13. Rather. . I will use panel data. However. Of course. Current best practice in macroeconomic forecasting appears to be multivariate vector error-correction models. I do not use a vector autogressive model (VAR). which are not discussed in the forecasting book of Clements and Hendry (1999). this is only helpful if the countries behave similarly .
1 Growth empirics took oﬀ with Barro (1991) and Manikw. Ley and Steel (2001). with Fernandez. Lee. Doppelhofer. This model has to be built with two main components: variables (appropriately measured and transformed) and econometric techniques. A whole industry of cross-section empirics developed in the 1990s and is still alive today. Capital adjusts to productivity gains. No consensus has yet emerged on the best technique. Miller and Sala-i-Martin (2004). This chapter therefore assesses the strengths and weaknesses of diﬀerent econometric methods for long-run growth analysis. but provides no independent contribution. In very important contributions Islam (1995) and Caselli. Finally. The initial reaction came from Levine and Renelt (1992). p. 459) combined with constant capital-output ratios implies that all growth in per capita income stems from changes in TFP in the simple models. and Hauk and Wacziarg (2004) recent examples. The preceding chapters outlined the diﬀerent theories. Pesaran and Smith (1997) showed that heterogeneity also applies to long-run growth rates 1 Chapter 2 showed that the also still popular growth accounting technique adds little value. who challenged the robustness of Barro’s results. This is not a straightforward task because a wide range of methods is used in growth empirics and new methods are becoming available almost every month. the most appropriate available econometric technique has to be selected. The argument of Barro and Sala-i-Martin (2004. The chapter on forecasting sketched some of the theoretical requirements for a successful forecasting model.11 The evolution of growth empirics The ultimate goal of this study is to create a sensible and robust forecasting model for long-run GDP growth. the candidate variables and how best to measure them. Esquivel and Lefort (1996) were among the ﬁrst to use panel methods to deal with heterogeneity issues. It also explains why I use a two-stage panel method that combines cointegration analysis and a stationary model. which used long-run averages of GDP growth and its determinants in a cross-country framework. . Romer and Weil (1992). The range of empirical techniques extends from the still popular crosssection methods to the recently developed panel cointegration techniques.
The newly developed panel techniques also imply higher data requirements: time series have to be available.7 shown in the chapter on growth theory: α β α+β ln sk. Labor input and technology in country i are assumed to grow at exogenously determined rates of ni and gi as outlined earlier on page 10 in the chapter on growth theory. Fortunately. MRW make a host of strong assumptions. i is the regression error and the other coeﬃcients are as described in chapter 2.i − ln(ni +g+δ)+ 1−α−β 1−α−β 1−α−β ln yi = A0 +gt+ i (11. The equation to be estimated is generally derived from a simple neoclassical production function. this approach has many weaknesses .1) where the index i runs across countries. the focus on these weaknesses has helped foster the development of more appropriate techniques over the past 15 years or so. A constant fraction sk of output is saved to accumulate physical capital and a constant fraction sh is saved to accumulate human capital. The pooled mean group estimator of Pesaran. My setup combines their work with stationary panel techniques and time-series models to allow enough parameter heterogeneity.and I will not use it to build my forecasting model. it is useful to sketch the basic estimation framework in some detail. Using the steady-state characteristic that physical and human capital per eﬃciency unit of labor do not change anymore one can derive an equation in levels as in Manikw. often augmented with human capital. given the scarcity of time series back then. Breitung (2000 and 2005) and others on panel unit roots and panel cointegration provides a toolbox to model economies’ long-run dynamic properties and exploit panel information. but does not yet test for cointegration. However.114 11 The evolution of growth empirics and convergence speeds. It was the best available technique when growth empirics took oﬀ in the early 1990s. Romer and Weil (1992) and similar to equation 2. The work of Pedroni (2000 and 2001).outlined below . t is a time trend. Shin and Smith (1999) makes it possible to deal with these heterogeneities. Techniques and datasets developed in parallel.1 Still widely used: cross-section The overviews of the empirical literature in the chapters on the individual drivers of economic growth highlighted the still widespread use of cross-section (since the section usually is a country: cross-country) estimation techniques over the past 15 years. In order to clarify assumptions as well as strengths and weaknesses of crosssection methods and to track the evolution of empirical methods. To estimate this in the cross-country dimension. Table 11. 11. For .1 lists some of the most relevant contributions in chronological order and shows that even in 2004/05 many authors used cross-section techniques.i + ln sh.
Rodrik.11.1 Still widely used: cross-section Table 11. Romer & Weil Loayza Islam Caselli.1. Leblebicioglu & Schiantarelli Chen & Dahlman Doppelhofer. Subramanian & Trebbi Yao & Lyhagen Benhabib & Spiegel Ianchovichina & Kacker This study Method used Cross-section Cross-section Cross-section Cross section Panel π Panel LSDV Panel GMM Cross-section Panel Cross-section Cross-section Cross-section Cross-section Cross section Cross-section Cross-section Pooled Mean Group Cross-section Cross-section Panel GMM Cross-section Cross-section Cross-section Cointegration analysis Panel GMM System GMM Cross-section Panel Panel GMM Cross-section Cross-section Cross-section plus time Pool Cross-section Cross-section CS. Ley & Steel Kr¨ger & Lindahl u Pritchett Sarno Dowrick & Rogers Soto Bosworth & Collins Brunner Dollar & Kraay Alcal´ & Ciccone a Barro & Sala-i-Martin Batista & Zalduendo Bond. Esquivel & Lefort Klenow & Rodriguez-Clare Lee. Empirical techniques used in the literature Year Author(s) 1991 1991 1992 1992 1994 1995 1996 1997 1997 1997 1998 1999 1999 2000 2000 2000 2001 2001 2001 2001 2001 2001 2001 2001 2002 2002 2003 2003 2003 2004 2004 2004 2004 2004 2004 2004 2004 2004 2004 2004 2004 2005 2005 Barro DeLong & Summers Levine & Renelt Mankiw. Pesaran & Smith Sala-i-Martin Edwards Frankel & Romer Hall & Jones Bils & Klenow de la Fuente & Dom´nech e Hanushek & Kimko Bassanini. GMM & LSDV Cross-section PCGets Cross-section Cross-section & panel Cross-section Pooled Mean Group Cross-section & panel Cross-section 2-stage method 115 . Miller & Sala-i-Martin Hauck & Wacziarg Hendry & Krolzig Hoover & Perez Portela et al. Scarpetta & Hemmings Bernanke & G¨rkaynak u Cohen & Soto Easterly & Levine Fernandez.
a positive coeﬃcient on a right-hand side variable should not be interpreted as the level of this variable having an unconditional positive impact on the rate of long-run GDP growth. it is hard to see how an inherently dynamic process such as economic growth can be modeled by abstracting from the dynamics through the use of long-run time averages and by focusing only on cross-sections. This may have been necessary in the early days of growth empirics because of the 2 Chapter 5 showed that this 3% is at the low end of the range of depreciation rates used in the literature. Therefore.116 11 The evolution of growth empirics example. The number of observations is equal to the number of countries. the elasticities α and β in the production function are assumed to be identical across countries despite the extremely heterogeneous sample of 98 countries. life expectancy in 1960 and the fertility rate).g. Over the past 15 years.2 Weaknesses of cross-section regressions In principle. A classic cross-section convergence contribution comes from Barro (1991): In a sample of 98 countries he regressed the growth rate of real GDP per capita over the period 1960 to 1985 on 1960 school enrollment and the 1960 level of GDP per capita. The move to convergence regressions in the empirical literature was already highlighted in chapter 2. In these models. . this setup works least well in the sub-sample where these assumptions are most likely to hold: the OECD sample. as ever more regressors were added to the model either to introduce a new theory of economic growth and/or to improve the ﬁt of the regressions. Adding more countries increases the degrees of freedom but also the heterogeneity of the sample.2 Population growth and the investment ratios enter in their country-speciﬁc averages over the years 1960 to 1985. the time dimension enters through the comparison of averages of growth drivers with the starting level of income or productivity. Unfortunately. Here. the literature has mushroomed. GDP growth depends on the diﬀerence between initial and steady-state GDP levels. As outlined earlier. one has to be careful when interpreting the results from convergence regressions. the control variables capture the level of steady-state GDP. Additional so-called “state variables” were introduced to better model A0 (e. 11. With the growth rate of GDP (or TFP) on the left-hand side and initial GDP on the right-hand side. As outlined earlier in this study one of the reasons for these diﬃculties may be that capitaloutput ratios are not the same across the OECD countries. MRW also assume that the growth rate of exogenous technological progress g is the same 2% across countries and that the depreciation rate δ is identical at 3% across countries. while so-called “control variables” were to capture other determinants of steady-state income such as average trade openness over 1960-85.
All ﬁve assumptions are likely to be false as the following sections will explain in detail. Pesaran and Smith (1997) ﬁnd that steady state growth rates diﬀer signiﬁcantly across countries and that ignoring these diﬀerences signiﬁcantly biases the estimates of the speed of convergence downward. but at least allow the initial level of technology A0 to diﬀer across countries.2. A slightly less restrictive assumption would see the elasticities of substitution between the production factors identical. Technology is a public good to which all countries have access.2 Weaknesses of cross-section regressions 117 scarcity of time series.1. 11. Caselli. Solow (2001) thinks that this is not enough heterogeneity: ”It is certainly unwise to assume that all economies are equally eﬃcient at reallocating inputs across industries” so that capital intensity may adjust to changing relative factor prices. Esquivel and Lefort (1996) show that the speed of convergence will be biased downward when individual eﬀects are ignored. recent research based on panel techniques shows that this assumption does not hold. The term g is absorbed in the constant term.1 Same production function assumed As outlined above.2. . However. However. the basic cross-country growth estimates assume identical production functions in all economies: in the long run the same amount of input will generate the same amount of output in all economies. The standard cross-section estimator will not be consistent. However. this requires a ﬁxed eﬀects panel model. In addition. the rate of technological progress and the rate of convergence. Leaving aside the general problems with the aggregate production function as outlined in chapter 2. then cross-section regressions will suﬀer from omitted variable bias if the explanatory variables are correlated with A0 in equation 11. this abstracts from a potentially very interesting aspect of the growth process: technologies may develop diﬀerently across countries over time.2 Long-run growth path assumed to be constant and the same across countries Another weakness of the cross-section method is that technological progress is assumed to be constant and identical across countries (gi = g) and is thereby pushed to the sidelines.and cannot be tested . He indeed ﬁnds signiﬁcant diﬀerences in A0 across countries.11. 11. but it does not seem appropriate anymore. If these diﬀerences in technology exist. Technically. Lee. Islam (1995) was the ﬁrst to apply this approach to growth studies.with cross-section methods. exogeneity of the explanatory variables is assumed and the errors are assumed to be uncorrelated with the regressors. The assumption that technology improves at the same rate in all countries in the sample is not tested . this appears to be a highly restrictive assumption. The other weakness of cross-section methods is that they assume all parameters to be homogeneous across countries: the production function.
In reality. countries are heterogeneous and their ﬂexibility to react to shocks diﬀers signiﬁcantly. p. Ordinary least squares estimation is only valid under the assumption of independence. Fixedeﬀects panel estimates can tackle this problem. 11. 1134) argues that “it is not entirely convincing that savings and fertility behavior will not be aﬀected by all that is included in A(0)”. However.2. Evans (1998) uses cointegration tests on the natural logarithm of GDP per capita for country pairs and ﬁnds widely diﬀering trend growth rates for a sample of 54 countries over 1950 to 1990. most panel models ﬁnd heterogeneous adjustment speeds across countries (for example: Bassanini.2. Quah (1993) argues that long-run growth rates are neither stable in one country nor the same across countries. e. as proposed by Islam (1995).3 Same pace of conditional convergence assumed Cross-section models assume that all countries converge to their steady state at the same pace: λi = λ for all i. From the discussion in chapter 4 on labor input it should have become clear that population growth is endogenous: higher per capita income leads to lower birth rates. The model used below supports this view of heterogeneity. Islam (1995. Also. Scarpetta and Hemmings ). This is in line with Caselli. Indeed. The results from the sub-sample of 27 relatively poorly educated populations lead him to conclude that technical knowledge may not be ultimately accessible to countries with relatively small stocks of per-capita human capital. so it includes the left hand side) which depends on many other factors such as the input of human capital. . 11.118 11 The evolution of growth empirics Likewise.4 Errors are assumed uncorrelated with the explanatory variables The identifying assumption in Manikw.5 Right-hand side variables assumed exogenous Among the most robust variables in cross-section studies are the population growth rate and the investment ratio. This assumption is not (and cannot be) tested with cross-section methods. Romer and Weil (1992) is that the error term in their regression is not correlated with the regressors such as the investment ratio and population growth. 11. In a similar spirit. in the country-speciﬁc features of an economy which MRW absorb in the error term. i. the country’s trade openness and its institutions.2. the discussion of physical capital in chapter 5 indicated that investment seems to react to the returns on physical capital (deﬁned as GDP per unit of capital. Esquivel and Lefort (1996) who argue that the rate of investment in physical capital is determined simultaneously with the rate of growth. not the other way around.
but it has not prevented the cross-section industry from ﬂourishing.3 The climax of cross-section 119 11. even if it does not mention the weaknesses of cross-section methods or the alternative methods that are becoming available. there is nothing that would raise questions about the initial assumption of identical technologies in all countries. Fernandez. 11. In spite of all these contributions. p. Sala-i-Martin (1997) uses millions of cross-section regressions to test the robustness of the variables used in empirical growth studies and to show that Levine and Renelt (1992) were too pessimistic in their conclusions that hardly any variable robustly explains growth. the cross-section industry reached a climax around the turn of the millennium.11.” (ibid) . with technologically-sophisticated developed countries. 51) and Romer (2001. which had GDP per worker around 65% below the US level in 1980. Combined with the assumptions outlined above. the invalidity of assumptions made and the problems with data especially for investment and human capital. this is problematic.6 In sum: many assumptions are violated In sum.” They nevertheless include Peru and Costa Rica. p. in their cross-section estimation. Also. 226) referring to the cross-section approach as a “dead end”3 . It is nevertheless important to be aware of the developments in this area and of the conclusions drawn in order to understand the current state of the empirical growth literature and to appreciate the diﬀerences from the panel work promoted in this study. These studies ignore the weaknesses of cross-section techniques outlined above.2. Sala-i-Martin reﬁned the analysis with coauthors using a Bayesian averaging of classical estimates (BACE) approach. This view was emphasized many years ago. DeLong and Summers (1991) were ”skeptical of what can be learned by combining in one regression very poor countries. several studies show that the assumptions used in the cross-section literature are violated. which appear to have productivity levels less than Britain before the industrial revolution. Ley and Steel (2001) use a Bayesian model averaging strategy on 2 million regressions and come to similar conclusions to Sala-i-Martin: there are robust correlates with GDP growth in cross-section regressions. cross-section methods have to use highly heterogenous countries to get enough degrees of freedom.3 The climax of cross-section Despite these fundamental shortcomings as well as Robert Solow’s 1994 revelation that he does “not ﬁnd this a conﬁdence-inspiring project” (Solow . which combines an averaging of estimates across models (a Bayesian concept) 3 He also says that looking “only at the cross-country regression evidence. For example. Zambia is unlikely to have the same preferences or technology as the United States. Barro and Sala-i-Martin (2004) is probably by far the most inﬂuential textbook on growth theory and empirics.
de Groot and Heijungs (2002) use a ”response surface metaanalysis” to show that a limited number of variables are robustly related to growth. In addition. 287). 282) repeatedly said that he has ”been skeptical from the beginning about the interpretation of cross-country growth regressions.120 11 The evolution of growth empirics with the classical approach of OLS estimation. We are back at the original Solow 4 Batista and Zalduendo (2004). Forecasting exercises at the IMF and at the World Bank published in 2004 and 2005 also used cross-section methods.. Cross-section analysis too narrowly focuses on the issue of convergence although the real focus should be the evolution of output over time.4 These ever more sophisticated cross-section exercises did not address the really important methodological issues outlined earlier in this chapter. They advocate the generalto-speciﬁc approach which is shown to have size and power near their nominal levels.. Brock and Durlauf (2001) criticize the robustness investigations of Levine and Renelt (1992) and Sala-i-Martin (1997) as being vulnerable to the likely collinearity of regressors: any coeﬃcient is highly unstable when alternative collinear regressors are added to the model. but large size). Florax. Hendry and Krolzig (2004) propagate a quick and easy way to reach the same conclusions as Fernandez. something that is common in the cross-section models. Doppelhofer. p. while the Sala-i-Martin approach is not stringent enough and accepts a false relationship too often (high power. . Hoover and Perez (2004) use a cross-section approach for some 100 highly heterogeneous countries. 326) thinks that the work of Lee. Based on Monte Carlo experiments they show that the Levine and Renelt approach is too stringent and rejects a true relationship too often (small size. It is time paths that need to be modeled and studied” (p. Hoover and Perez (2004) introduce another layer to the analysis by comparing the robustness tests used by Levine and Renelt (1992) and those used by Sala-i-Martin (1997) with a general-to-speciﬁc approach.” ”It is probably not a good idea to set cross-country regressions the task of explaining the rate of growth . but low power). Ianchovichina and Kacker (2005). Islam (1998. Miller and Salai-Martin (2004) ﬁnd 11 variables robustly partially correlated with long-term growth. For example. as far as the cross-country dimension of this concept is concerned”. p. Ley and Steel (2001) by also using a general-to-speciﬁc procedure. Solow (2001. Their conclusions are based on approximately 21 million randomly drawn cross-section regressions on a dataset of 32 variables for 98 countries over the period 1960 to 1992. he is especially wary of applying his own theoretical framework to developing economies. While these certainly are further methodological contributions. No reference is made to the weaknesses and disadvantages of cross-section regressions. they still do not address the crucial issues in growth empirics. Pesaran and Smith (1997) ”robs the concept of convergence of any real meaning. They contrast their approach with the millions of regressions that Sala-i-Martin and his coauthors require.
Johnson and Temple (2005) refer to the panel work of Islam (1995) as a pool analysis. p. However. a pool is a dataset with observations on diﬀerent units at diﬀerent points in time drawn randomly from a large population. The estimate does not 5 6 Often ”to pool” refers to adding additional units to a panel. studies such as that of Barro and Sala-i-Martin (2004). who split samples into smaller periods and treat the observations on the USA in 1970-80 just like those on Japan in 197080 and like those on the USA in 1980-90. data for GDP and investment for the same country would be observed over several consecutive points in time. In my context.4 Advantages of panel techniques 121 model that explicitly focuses on the development within one country. ﬁlters like the Hodrick-Prescott version suﬀer from the well-known end-point problems. This may explain why Durlauf. By contrast. which contradicts Hsiao’s deﬁnition. Capturing diﬀerences in countries’ growth paths should be the prime task of growth empirics. 16) who refers to a pool as a panel where all coeﬃcients (constants and slopes) are the same across the units (because one cannot determine whether the same unit has been observed several times).6 In growth empirics. 11. . 1 and Wooldridge . My goal is to avoid that vulnerability by focusing on growth determinants.5 This is consistent with Hsiao (2003. such as prices of houses in diﬀerent cities in diﬀerent years. Given the large number of potentially relevant regressors. However. Only panel models can capture the evolution of economies over time and still provide enough degrees of freedom to include a reasonable number of regressors.4 Advantages of panel techniques Estimating cross-section regressions does not allow the analysis of the complex dynamic adjustment processes or of the heterogeneity of growth rates across countries. An alternative would be to analyze countries individually with time series methods. 408): there is a cross-sectional dimension N and a time-series dimension T . p. p. are pools. However. Trying to deal with the short sample problem by using the mean group method. EViews for its own historical reasons distinguishes between a pool that has few cross-sections and a panel which has many cross-sections. In general. Another approach would be to abstract from any regressors and use simple ﬁlters for GDP to estimate a rate of potential growth. where equations are estimated country by country and then averages of the coeﬃcients are used. some of the relevant data are only available at annual frequencies starting in the 1960s or 1970s. this leads to a low number of degrees of freedom.11. a panel or longitudinal dataset follows several crosssections over time (see for example Hsiao . EViews allows unit-speciﬁc slope coeﬃcients only in “pools”. In contrast to the textbook deﬁnitions. is unlikely to lead to meaningful country estimates. Hodrick-Prescott ﬁlters will be used as one of the benchmarks for comparison with forecasts from fundamental models.
11.8 10.4 10. as better estimation techniques become available. especially as longer spans of data become available. Therefore.it + a2 ln sh.1. which is to model .in a slope of zero. By contrast. but it does not capture the time dimension of the data. Diﬀerence between cross-section and panel The second advantage is that the larger number of data points and more degrees of freedom in panels than in cross-section or time-series datasets lead to more eﬃcient estimates. panel techniques are ideal for my purpose.6 10. The large dots are averages of the annual observations (small dots) for each country and the thin regression line through these large (cross-section) dots has a positive slope.0 9.2 10. Figure 11.2)ln yit = a0i + gi t + a1 ln sk. 11. The ﬁrst and most important advantage of panel approaches over crosssection methods is that they make it possible to control for unobserved (omitted) individual-speciﬁc characteristics that may be correlated with explanatory variables but constant over time.1 illustrates how cross-section estimates using long-run country averages may ﬁnd a positive slope while ﬁxed eﬀects estimates would ﬁnd no relationship between X and Y. The panel equivalent of the MRW equation is: (11.122 11 The evolution of growth empirics take account of the fact that one country is observed several times. Temple (1999) already argued that ”panel data studies will increasingly be the best way forward for many questions. Pooling increases the degrees of freedom.in this stylized example .it − a3 ln(nit + g + δ) + it where the index i runs across countries and the a0i are the ﬁxed eﬀects. using all observations for each country and a countryspeciﬁc constant (ﬁxed eﬀect) in a panel model will result .0 10.6 11.2 11.8 17 18 19 20 21 22 23 24 25 26 Observations country 2 X 27 Fixed effects estimation: slope = 0 Y Observations country 1 Cross-section estimation using time averages: positive slope Fig.4 11. one might add. More interesting hypotheses can be tested.” And.
Islam ﬁnds a wide range of ﬁxed eﬀects and thus a wide range of eﬃciency with which countries are transforming their capital and labor resources into output: technology does diﬀer across countries.4. For example. panel models which allow enough ﬂexibility and permit some of the coeﬃcients to be heterogeneous across countries appear preferable. 11. the panel literature has relaxed the assumptions of identical production functions (Islam . The slope parameters are homogeneous and identiﬁed by using only the within-country variation in the data. Therefore. However. As I will outline below. Endogeneity issues still need to be addressed by a careful selection of the regressors. Esquivel and Lefort (1996.2 Dealing with endogeneity bias Unlike Islam. for example 5 years. using a least squares dummy variable (LSDV) ”within” estimator and the minimum distance (MD) estimator on ﬁve data points with variables in ﬁveyear averages over 1960-85 for 96 countries.4. In this case. The LSDV estimate uses a full set of country-speciﬁc intercepts (or dummies) to capture the heterogeneity in technology. 11. p.11. Pesaran and Smith ) and identical pace of technological progress. estimates of the relevant parameters may turn out inconsistent or meaningless (Hsiao . This leaves at most eight time observations given that datasets usually do not start before 1960 or extend beyond 2000. some of the slope coeﬃcients can be allowed to be country-speciﬁc. many studies used medium-term averages. 8). He estimated a dynamic panel model which makes it possible to capture the country-speciﬁc levels of technology. Evans ).4 Advantages of panel techniques 123 the development of the trend in (steady-state) GDP plus any convergence to that steady state. below: CEL) does not focus on diﬀerences in technology levels (individual eﬀects) so they eliminate the country-speciﬁc eﬀect by ﬁrst-diﬀerencing the data. one still has to avoid combining too heterogeneous data. Loayza (1994) uses a similar approach and draws similar conclusions to Islam.1 Initial technology can diﬀer across countries The seminal paper by Islam (1995) was one of the ﬁrst to advocate the use of panel methods in growth empirics. Ignoring individual-speciﬁc characteristics that exist among the diﬀerent cross sections can lead to parameter heterogeneity and biased coeﬃcients. With this approach. identical convergence speeds (Lee. In order to abstract from business cycle variation and to deal with the fact that many variables are measured less frequently than annually. omitted variables that are constant over time will not bias the estimates. the work of Caselli. They also . This implies that countries with very diﬀerent characteristics should only be included in the panel if the ﬁxed eﬀects can model that heterogeneity.
Because the errors are correlated with the lagged diﬀerence of the dependent variable. the between estimator) performs better than alternative estimation techniques. such as the ﬁxed-eﬀects estimator or the Arellano-Bond estimator. lagged levels are used as instruments. The corrected LSDV was suggested by Kiviet (1995) and calculates the bias with the use of a consistent estimator such as Anderson-Hsiao and then subtracts this bias from the LSDV estimate. And they ﬁnd strong evidence of endogeneity. Hoeﬄer and Temple (2001) because lagged levels are weak instruments for subsequent ﬁrst diﬀerences. this argues for caution when combining the data from very heterogenous countries in one panel. their focus is on the endogeneity bias that is present both in Manikw. Their focus is on the well-known bias of the LSDV method when estimating the coeﬃcient of a lagged dependent variable. However.124 11 The evolution of growth empirics eliminate time eﬀects by calculating deviations from period means. e. using lagged levels as instruments for the changes.3 Addressing lagged dependent bias Judson and Owen (1999) also use panel methods to identify country-speciﬁc eﬀects. if there is measurement error in the variables. However. Judson and Owen also evaluate the GMM estimator suggested by Arellano and Bond (1991). Judson and Owen also show that there is hardly any bias of the coeﬃcients on the exogenous regressors. LSDV does very satisfactorily in his Monte Carlo simulations. By contrast. Indeed. Hauck and Warcziarg (2004) run Monte Carlo simulations and conclude that OLS estimation of a single cross-section of variables averaged over time (i. they do not ﬁnd any improvement over the corrected LSDV. They indeed ﬁnd that treating the country-speciﬁc eﬀects properly raises the estimated speed of convergence. CEL propose a General Method of Moments (GMM) estimator following Arellano and Bond (1991). which uses information from all lagged values of the dependent variable and from lagged values of the exogenous regressors. Their discussion is therefore mostly relevant for research on the speed of convergence. GMM dynamic panel estimators are also used by Easterly and Levine (2001) and by Dowrick and Rogers (2002). . However. The Anderson-Hsiao estimator removes the ﬁxed eﬀect by ﬁrst diﬀerencing the equation to be estimated. They suggest the use of a corrected LSDV estimator for panels with a small time dimension (t <= 10) and of the Anderson-Hsiao estimator for panels with larger time dimensions. Instead. this technique is rather problematic because the relationship between the variables and their instruments is likely to be weak. Romer and Weil (1992) and in Islam (1995) and on the correlation of lagged income with the error term which was also not addressed by Islam. Again. the GMM approach was criticized by Bond.4. 11. Islam (2000) shows that the Anderson-Hsiao estimator performs poorly when lagged levels are used.
Barro (1997). Caselli. Pesaran and Smith (1997. 11. First. These movements of the steady state are the focus of the investigation in the present study. CEL (1996) argue that because countries are near their steady states. Felipe and McCombie (2005.4. the results from the panel literature outlined in the previous sections are quite striking. The methods just outlined improve the ﬁt to the accounting identity. The method used by Lee. and the theoretical speed of convergence using the identity is inﬁnity. Esquivel and Lefort  and Lee. e. the MRW equation is no longer a valid regression because the error term is correlated with the regressors. Furthermore. They promote the use of time-series models in heterogenous panels because only this approach allows researchers to ”distinguish between the convergence of countries’ outputs to their steady state paths and the cross-sectional evolution of the paths themselves. p. Pesaran and Smith ). conditional convergence rates are considerably higher than the 2% to 3% found in cross-section studies (Islam . Temple (1999). And they show that ﬁxed eﬀects panel estimation as in Islam (1995) yields inconsistent estimates of all coeﬃcients if technology grows at diﬀerent rates in diﬀerent countries. Evans ). parameters of the production function come closer to commonly accepted empirical values (Islam . probably because of a combination of excludability (patents). p. Certainly. all growth modeling should take good care to .1 in the theory chapter above. Pesaran and Smith (1997) allows countries to diﬀer both in the initial level of technology A0 .4 Advantages of panel techniques 125 11. However. Durlauf and Quah (1999). the rate of technical progress g and the speed of convergence λ. 1128). as outlined in section 2.4 Modeling heterogeneous technological progress Panels can identify technological progress in a simple Solow model because the coeﬃcient on the time trend can be heterogeneous when moving out of the cross-section world. the pace of technological progress is diﬀerent across countries. the observed diﬀerences in growth rates must stem from shifts in the steady state that occur at diﬀerent strengths. Pritchett (2001) and Wacziarg (2002) all argue that the use of ﬁxed eﬀects panel models has to be approached with care because the information on the variation between the countries is lost. 22) show that the better ﬁt and the increase in the estimated speed of convergence is to be expected. technology is not a global public good.6. i.” This is in line with Bernanke and G¨rkaynak (2001) who argue that if the u growth rate is stochastic. Third. learning by doing or some form of social or organizational capital. Second. p.5 Summary of results from panel regressions Overall. production functions are diﬀerent (CEL .11. Lee.4. 366) show that standard cross-section regressions yield estimates of the convergence speed that are biased downward even if technology is assumed to improve at the same rate across countries.
915) criticism that panel methods are “reducing the focus from the long-run to higher frequency data. heterogeneous convergence speeds and country-speciﬁc technological progress is the pooled mean group (PMG) technique presented by Pesaran. As the focus shifted from the between-country variation to the within-country variation of the data.7 11. The power of many tests was increased. The pooled mean group technique developed in the 1990s was a major step forward.126 11 The evolution of growth empirics capture both the within and the between variation.” In addition. which was used for modeling long-run growth by Bassanini. As is well known in the time-series literature. And the use of non-stationary panel techniques takes care of Wacziarg’s (2000. However. Shin and Smith (1999). the characteristics of the underlying time series required closer attention. The following sections will present the advantages and weaknesses of these non-stationary panel techniques. Recently developed panel unit root and panel cointegration tests help answer those questions. . A careful treatment is therefore necessary. one has to be particularly careful with time series that include interpolated values such as the human capital variable which I will use. non-stationarity can lead to spurious regressions. The insight of Lee. Short-run coeﬃcients are likely to be of little value. Shin and Smith is the observation that panels with large T and large N were usually estimated in two 7 See Durlauf. p. as they may capture measurement error. Scarpetta and Hemmings (2001) and others at the OECD. The non-stationary panel techniques outlined below deal explicitly with the non-stationarities that are typically present in datasets on long-run growth.5. The same holds for variables that show little year to year variation. However.1 Pooled mean group technique One method that allows ﬁxed eﬀects. Islam (1995) and CEL (1996) were signiﬁcant steps forward as they addressed the omitted variable bias. Johnson and Temple (2005). Using a richer dataset by combining several countries in a panel was extremely helpful. Pesaran and Smith (1997) that most of the parameters in growth regressions are heterogeneous made clear that more sophisticated estimation techniques are necessary. The starting point of Pesaran. panel methods allow enough ﬂexibility to do this. the panel techniques initially applied to growth and convergence issues did not account for the non-stationarity of the regressors.5 Non-stationary panel techniques The panel techniques proposed in the mid-1990s by Loyaza (1994). but left some questions open. leading to a conclusion on the most appropriate technique to be used for my forecasting model. 11.
g. Shin and Smith ): p−1 q−1 (11. and short-run dynamics with GDP growth in the previous periods and changes in the drivers in the current and previous periods. The long-run coeﬃcients (or a subset of them) are presumed identical across countries while constants. short-run parameters and error variances are heterogeneous. Their pooled mean group estimator thus involves both pooling and averaging. which constrains some coeﬃcients to be identical across countries and allows others to diﬀer. In addition.2. Pesaran. but my own work with the PMG estimator shows a signiﬁcant heterogeneity in both ﬁxed eﬀects and convergence speeds (Bergheim ). they recognize the non-stationarity of the underlying data and propose an error-correction model. in Pesaran. A time trend to capture heterogeneous technological progress is not signiﬁcant in Bassanini. In almost all cases. The country index i runs from 1 to N while time t runs from 1 to T . the average convergence speed λ is higher than in cross-section regressions. or alternatively by pooling the data and restricting all slope coeﬃcients and error variances to be the same across countries (e. But they also include a large number of additional variables that are meant to model the evolution of technology over time.3)yit = λi yi. ﬁxed eﬀects).t−1 + βxit + gi t + ˆ j=1 αij yi. Country-speciﬁc ﬁxed eﬀects µi and a random error it complete the equation.t−j + µi + ˆ it The growth rate of per capita GDP in period t depends on a countryspeciﬁc convergence coeﬃcient λi times the level of GDP in the previous period. Finally. They stick closely to the neoclassical model and focus on the investment ratio and population growth as explanatory variables. Short-run regressors are included to control for country-speciﬁc business cycle dynamics. convergence speeds. the product of the vectors of common long-run coeﬃcients β and the drivers of GDP x. The restriction that the slope coeﬃcients in the long-run relationship are identical can be tested. The coeﬃcient gi captures the country-speciﬁc rate of technological progress. The application of Bassanini. As expected.t−j + ˆ j=0 θij xi.11. In the notation used above the PMG approach estimates (see equation 3. Shin and Smith propose an intermediate solution. And . Hausman tests accepted the hypothesis of common long-run coeﬃcients β. p and q are the lag orders for the dependent and independent variables.5 Non-stationary panel techniques 127 ways: either by estimating N separate regressions thereby allowing all coeﬃcients to be country-speciﬁc and then calculating the means of the coeﬃcients (mean group technique). Scarpetta and Hemmings do not discuss heterogeneity in detail. Bassanini. Scarpetta and Hemmings (2001) because technology is modeled comprehensively by other variables. Scarpetta and Hemmings (2001) for a sample of 21 relatively homogeneous OECD countries over 1971-98 makes full use of the ﬂexibility of the PMG approach.
acknowledging that capital accumulation depends on the return on capital. “the demand for empirical studies [is] exceeding the supply of econometric theory developed” in the area of nonstationary panels. but allow trade to have region-speciﬁc eﬀects. there are three disadvantages of PMG in general and in the growth context. Thirdly. Fortunately. theoretical growth models imply several cointegrating relationships among the main variables. They impose a common longrun coeﬃcient only on capital.it is simply assumed. some recent research like Brunner (2003) carefully analyzes the non-stationarity properties of the underlying data. The next sections present the techniques to address these issues.2 Testing unit roots and cointegration in panels Most applications of the PMG technique do not test whether the variables are indeed I(1) or whether there is a cointegrating relationship among the I(1) variables. Then they explain how these techniques can be combined with a stationary ﬁxed-eﬀects panel to estimate the short-run dynamics.) and negative coeﬃcients for all variables that trend downward (e. a potentially serious problem that I will deal with explicitly in my approach. the PMG approach may ﬁnd signiﬁcant positive coeﬃcients for all other variables that also trend upwards (human capital. PMG does not allow to test cointegration .128 11 The evolution of growth empirics they model investment separately. They ﬁrst present the current state of panel unit root tests and cointegration tests.5. There is a risk of spurious regressions. Surveys of the literature on non-stationary panel techniques include Baltagi and Kao (2000) and Breitung and Pesaran (2006). but long-run growth processes are much more involved. which itself depends on the path of technology. In addition. the positive coeﬃcient found for the investment ratio may simply capture business cycle co-movements of GDP and investment. This relationship is simply assumed and the attention immediately turns to the sign and magnitude of the coeﬃcients. per capita capital and per capita trade. 11. Breitung . This may be appropriate when testing purchasing power parity or the link between consumption and income.g. PMG can only take one cointegrating relationship into account. As I showed above. Firstly. Secondly. As Baltagi and Kao (2000) note. but relies on single-equation time series techniques. One of the contributions of this study is to apply these techniques to growth issues in ways that have not been explored before. Results for the growth model are presented in the next chapter. with GDP moving higher over time. the techniques for panel unit root tests and panel cointegration tests have been reﬁned over the last few years and are now available in several software packages. By contrast. This calls for the separate use of panel unit root and cointegration tests. population growth). Despite its strengths. Yao and Lyhagen (2001) apply the PMG technique to Chinese provinces over 1978-97 and ﬁnd a positive long-run relationship between per capita GDP. trade etc.
t−1 + it The errors are assumed to follow the usual assumptions (independently and normally distributed with zero means and ﬁnite variances. 2. the basic idea behind all panel analysis is that there is some commonality across the groups in a panel. but possible serial correlation). Not surprisingly. the focus has to be on techniques that perform well in samples with T around 30 and N around 25.5 Non-stationary panel techniques 129 and Pesaran (2006) conclude that ”the analysis of cointegration in panels is still at an early stage of its development. which aims to ﬁnd out whether a1 = 1 in yt = a0 + a1 yt−1 + t by testing whether η = a1 − 1 = 0 in the following regression:8 (11.” At the moment.5) ∆yit = µi + ηi yi. The new. in particular that the error i is uncorrelated with the error j. All tests allow group-speciﬁc deterministic elements (intercepts and time trends). The assumption of the so-called ”ﬁrst generation” tests is that the groups are independent.3 Panel unit root tests The importance of unit roots has been well known since Granger and Newbold (1974) and the survey by Campell and Perron (1991).5. nontechnical overview of the literature. unit root tests have become possible in the panel dimension. The goal is to select the most appropriate techniques for my growth model.. N groups each with a mean of µi so the model becomes (11. It also addresses the possible heterogeneity of the model parameters across the diﬀerent cross-sections. Therefore. 8 For a detailed explanation see. The starting point for panel unit root tests is the familiar augmented Dickey-Fuller test. 181ﬀ. for example. one standard result from the literature in nonstationary panels is that an increase in the time dimension T is more informative than an increase in the group dimension N . Recently. Several Monte Carlo simulations assessed the size and power of the diﬀerent tests under diﬀerent conditions.4) ∆yt = a0 + ηyt−1 + t Moving to the panel dimension. the literature is mostly concerned with addressing the issue of cross-section dependence either in unit root tests or in cointegration tests. p..11. 11. we have i = 1. A larger number of groups allows analysts to better detect this commonality. . ”second generation” tests relax this assumption. Since the literature on panel unit roots and panel cointegration is very active and continues to evolve quickly. Enders (2004). the following sections give a brief. As mentioned above. but I will not focus on these since they are not yet available in standard software packages. An example for unit roots is Hlouskova and Wagner (2006). .
Their 1997 working paper version also includes a Lagrange multiplier statistic (IPS-LM). but smaller than zero (homogeneous alternative) so that all series are stationary processes. based on the working paper by Levin and Lin of 1992) use the homogeneous alternative of identical ﬁrst-order autoregressive coeﬃcients. Breitung orthogonalizes the residuals from the ADF regressions and the auxiliary regressions. compared to performing a separate unit root test for each individual time series” (p. Importantly for applied work.130 11 The evolution of growth empirics Usually. . T between 25 and 250). then estimates the ratio of long-run and short-run standard deviations for each group. They conclude that the ”power of panel-based unit root tests is dramatically higher. below: LLC. Their test is appropriate for panels of moderate size (N between 10 and 250. He ﬁnds that the LLC and IPS tests (see below) suﬀer from a dramatic loss of power when a trend is included among the deterministic terms. who uses the null hypothesis of stationarity. IPS show that their test performs better than the LLC test. The null hypothesis is that ηi is equal to zero for all groups. the null hypothesis is that a time series has a unit root in all the groups. Pesaran and Smith (2003. Lin and Chu (2002. A pooled regression of those residuals produces a t-statistic (UB-test) that has a standard normal limiting distribution. Im. Their t-bar statistic converges to a standard normal distribution as N approaches inﬁnity. The generalization in Breitung and Das (2005) accounts for correlations of errors across groups. Since Hlouskova and Wagner (2006) document the weak power of the Hadri tests I will not use them. They calculate ADF t-statistics separately for each individual group and then average these across the groups. 1). Breitung (2000) also uses the homogeneous alternative hypothesis.unlike in the standard ADF tests . In deviating from LLC. below: IPS) use a heterogeneous alternative which allows the ﬁrst-order autoregressive coeﬃcients to diﬀer across groups. Levin. This implies that . The exceptions are the tests by Hadri. a rejection of the null hypothesis does not imply that the unit root is rejected for all groups but only for a fraction of the groups (unfortunately they do not provide information on the size of this fraction). or that the ηi are diﬀerent across groups and a fraction of them are smaller than zero (heterogeneous alternative) so that some but not all series are stationary. The tests diﬀer in the speciﬁcation of the alternative hypothesis. so just right for growth datasets. Their three-step procedure ﬁrst calculates separate augmented Dickey-Fuller (ADF) regressions for each group plus two auxiliary regressions. The alternative hypothesis is either that the autoregressive parameter ηi is identical in all units.their t-bar test does not require a table of critical values derived from Monte Carlo runs. and ﬁnally computes a pooled t-statistic which has a limiting normal distribution. For the usual case of serially correlated errors they recommend that the lag order of the underlying ADF regressions should not be too low.
There may be more than one β and the number of linearly independent cointegrating vectors is called the rank. However. a vector of n time series xt is said to be cointegration if each series is I(1). In sum. Perron and Phillips (2005) show that the IPS and Breitung tests have low power if country-speciﬁc trends are present. p. LLC and Breitung tend to have the highest power (one minus the probability of accepting the null when it is wrong). Pesaran (2007) suggests a simple estimator that allows for cross-sectional dependence. single-group case were developed in the 1980s with major contributions from Engle and Granger (1987) and Johansen (1995). a signiﬁcant amount of information can be derived from a cointegration relationship between variables. The basics of cointegration in the simple time series. If GDP converges to the path traced by the fundamental drivers of growth we have a cointegration relation that can also be exploited for forecasting. Furthermore.11. i. I will show results from four diﬀerent test statistics in the next chapter: LLC. 2001) and Breitung (2005) appear to be widely accepted and quite robust.and it is evolving rapidly. Breitung and the two IPS tests.4 Panel cointegration tests The literature on panel cointegration tests is even younger than the literature on panel unit root tests .5. The general idea is that two non-stationary variables may not stray far from a linear combination of the two. where t is stationary. This is an attractive concept for modeling long-run growth. Over recent years the cointegration literature has expanded into the panel dimension for the same reasons as outlined above: richer datasets. 11. at least one must Grangercause the other. Hlouskova and Wagner concede that the IPS tests have been disadvantaged in their simulation study. The cointegration vector β leads to β ∗ xt = t . Indeed. there is a formal equivalence between cointegration and error-correction as established by Engle and Granger (1987). The tests by Pedroni (2000.5 Non-stationary panel techniques 131 The simulation study by Hlouskova and Wagner (2006) compares several diﬀerent unit root tests including the ones outlined here. not stationary. which uses homogeneous panels under both the null and the alternative hypothesis. e. Also. 189) point out that if two variables are cointegrated. which is only available in Matlab. They show that the sizes (probability of rejecting the null when it is true. They favor the test of Ploberger and Phillips (2005). usually set at 5%) of the LLC and Breitung (2000) tests are closest to their nominal size. while some linear combination of the series is stationary or I(0). so my main focus will be on them. Maddala and Kim (1999. Since there is no consensus yet on which estimator to apply in projects like mine. so that there is some long-run equilibrium relation between the variables. Cointegration is intimately connected to the idea of error-correction: the current deviation of a system from equilibrium conveys information regarding its likely future course and is therefore useful for forecasting. Moon. More formally. more pow- .
The issue of possible cross-section cointegration is discussed intensively in the panel cointegration literature. p. In all cointegration tests. the focus is on residual based approaches. but system approaches based on vector-autoregressions are being developed as well especially for cases with more than one within-group cointegration relation. The null hypothesis is for a common cointegrating vector βi = β larger than zero.” Pedroni and Vogelsang (2005) present tests for heterogeneous panels that are robust to cross-sectional dependence. Breitung and Pesaran (2005. However. this is the appropriate assumption for the analysis of long-run growth since general economic relationships may be assumed to be quite similar in all countries. Reimers and Roﬃa ) or demographics (Hondroyiannis and Papapetrou ). these tests are not feasible in this study given the large number of variables and the relatively small number of N and T . but faces the obvious diﬃculty that allowing all parameters to diﬀer across groups reduces the appeal of using a panel study in the ﬁrst place. The econometric literature tries to address the heterogeneity of parameters. Examples include Larsson. but that the short-run dynamics such as the adjustment speed diﬀer across the groups (see below). Pedroni’s group mean panel fully modiﬁed OLS (gFMOLS) The panel cointegration tests developed by Peter Pedroni are by now widely used on issues such as equilibrium exchange rates (Maeso-Fernandez. i. It turns out that some countries indeed diﬀer signiﬁcantly. Importantly. The idea is to estimate a homogeneous cointegration vector but to allow unit-speciﬁc ﬁxed eﬀects and short-run dynamics.e. the variables . the null hypothesis is for no cointegration. money demand (Dreger.132 11 The evolution of growth empirics erful tests etc. Breitung (2005) suggests a systems method that I will apply. multivariate tests are only admissible if the variables on the right-hand side are not cointegrated among themselves. Most tests assume that the cointegration vectors are identical for all groups. e. Marcellino and Osbat (2004). the residuals are non-stationary or I(1). Pedroni is explicit in highlighting the strengths of this estimator relative to the pooled (”within dimension”) panel fully modiﬁed OLS. Lyhagen and Lothgren (2001) and Groen and Kleibergen (2003). the often-quoted working paper is from 1996). The assumption of identical long-run relationships can be and should be tested. even in a panel setting the time dimension remains more informative on the long-run coeﬃcients than the cross-section dimension. However. The alternative hypothesis may be that all units have the same non-zero cointegration vector β or that they have unit-speciﬁc vectors. 29) conclude that “to date a general approach that is capable of addressing all the various issues does not exist if N is relatively large. So far. One of the most used tests is his group mean (“between dimension”) panel fully modiﬁed OLS (gFMOLS) method as outlined in Pedroni (2000. especially following Banerjee. Osbat and Schnatz ). which provides important information for policy analysis. i. Theoretically.
It is then used to correct for the bias induced by the endogeneity of the regressors. And . Breitung’s 2-step estimator Breitung (2005)9 proposes a computationally convenient two-step estimator. The resulting t-statistic is distributed normally. an oil price shock would aﬀect all countries at roughly the same time. The two-step estimator has been applied to health care expenditure in the OECD (Dreger and Reimers ) and to real exchange rate estimation (Dufrenot and Yehoue ). on unit-by-unit estimation that allows diﬀerent constants and slope coeﬃcients. So the null hypothesis of no cointegration (β = 0) is tested against the alternative of unit-speciﬁc cointegration. it ) is stationary with asymptotic covariance Ωi .e. The alternative hypothesis allows unit-speciﬁc (heterogeneous) cointegration vectors. Pedroni’s Monte Carlo simulations show that the biases are small in samples near the size of mine (T =30. The matrix αi allows group-speciﬁc ﬁxed eﬀects and β is the cointegration vector. Common time dummies would help address this issue.power is quite high. The group mean FMOLS test requires the time dimension to be at least as large as the crosssection dimension N . .11. for example. 9 The working paper version often quoted is from 2002.6) yit = αi + βxit + νit xit = xit−1 + it where the vector error process ξit = (νit . The same holds for size distortions.t−1 + γi it where it is a white noise error with positive deﬁnite covariance matrix Σi = E( it jt ). The covariance matrix of the error processes is estimated using a kernel estimator. i. Breitung starts from a transformed Vector Error Correction Model (VECM): (11. The estimator is based on the “between dimension” of the panel. It also requires errors to be uncorrelated across units. The mean value of the resulting slope coeﬃcients can then be evaluated using a t-statistic constructed as the simple mean of the individual t-statistics.7) γi ∆yit = γi αi β yi. which may be a problem in my context given that.5 Non-stationary panel techniques 133 cointegrate for each member of the panel with the same cointegrating vector β. This is similar to the IPS test for unit roots outlined above.not surprising given the panel framework . N =20). which is the case in my datasets. where the short-run and the long-run parameters are estimated in separate steps. Estimation is based on the following set of equations: (11.
This 2-stage estimation is similar to the Engle and Granger approach outlined in Engle and Yoo (1987. This second-step regression can be treated as ordinary least squares and the non-stationarity of the regressors can be ignored. The estimated short-run coeﬃcients αi and Σi ˆ are then used to transform ∆yit to zit and it to νit . This is one of the many compromises one has to make in growth empirics. Because of the small number of degrees of freedom. panel methods are superior although good care has to be taken to still capture the cross-section information.6 A two-stage estimation method As outlined above. Shin and Smith (1999) and with economic reasoning: the long-run linkage between variables is assumed to conform to underlying theoretical principles and is therefore common (homogeneous) across countries. I will use both the Pedroni and the Breitung tests in order to explore the robustness of the results. partly because their small-sample properties have been questioned by Breitung (2005). It is important to test whether long-run coeﬃcients are indeed homogeneous and to examine the heterogeneity of the short-run adjustment. On the other hand.t−1 + νit . my experiments show that the estimated cointegration vectors are quite similar regardless of the method used. The non-stationarity of the variables and the complexity of the growth process require panel cointegration techniques. ˆ ˆ The second step estimates a common cointegration vector β by OLS using ˆ the pooled regression zit = β yi. Among the tests not presented and used here are the DOLS tests proposed by Kao and Chiang (2000) and those by Mark and Sul (2003). 11. Some information in the covariance matrices may therefore not be exploited. it is not feasible to estimate the whole model in one stage. Breitung shows that the long-run ˆ parameters are asymptotically normally distributed. On the other hand. Therefore. simple time series methods do not have enough power and allow too much heterogeneity. However. I proceed to the next stage with the coeﬃcients estimated using the Breitung test.134 11 The evolution of growth empirics In the ﬁrst step. 149). . I use a two-stage approach where I exploit the information in the long-run linkages between the drivers of income and the level of income as well as the information in the short-run dynamics. However. cross-section methods are not useful for constructing forecasting models because they do not take account of the dynamic characteristics of an inherently dynamic process and because they have to impose too many assumptions that do not hold in reality. My estimation approach is in line with Pesaran. p. the speed and the channels of the short-run adjustment to any long-run equilibrium may diﬀer across countries. the unit-speciﬁc short-run parameters αi and Σi are estimated using separate models for all N cross-sections allowing for unitˆ speciﬁc cointegration vectors.
while the gFMOLS Pedroni test makes use of the modiﬁed NPT kindly made available by Hlouskova and Wagner.11. This is in line with the call of Brock and Durlauf (2001) to combine only countries that are exchangeable. This allows me to recover the error terms from these cointegrating relationships which are on average equal to zero and stationary.ect.just one cointegration relationship among ﬁve variables used in the augmented Solow model and that he does not use panel cointegration techniques. he also estimates short-run models for each country. at least one of the adjustment coeﬃcients must be non-zero if two variables are cointegrated. Few authors have approached long-run growth issues with cointegration techniques. Therefore. The two main diﬀerences to my approach are that he looks for . According to the Granger representation theorem. The time series of the error terms and the ﬁrst diﬀerences of the explanatory variables enter into short-run models for each country and into stationary panels.html . These errors will be used in the second stage Stage 2: Stationary models. For these tests the non-stationary variables included on the right-hand side must not be cointegrated among themselves. Since the main focus of my analysis is on GDP. Both stages are done separately for the two sets of countries. I also check for cointegration among these drivers of GDP growth.uni-bonn. The panel cointegration tests use country-speciﬁc constants.syr.htm http://www. Otherwise. the LLC and IPS tests are from the NPT page by Chiang and Kao (2002)11 . I pay most attention to variables that cointegrate with GDP. An exception is Sarno (2001). I estimate the cointegrating relations among the candidate variables using the panel techniques of Pedroni and Breitung outlined above. Dowrick (2004) also uses the idea of analyzing growth in an error-correction framework. The program codes for the Breitung tests are from Breitung’s homepage10 . too much heterogeneity could damage the estimation. In the country-speciﬁc models. In the ﬁrst stage. However.6 A two-stage estimation method 135 Stage 1: Cointegrating relationships. but does not apply non-stationary panel techniques.de/mitarbeiter/breitung/breitung. All panel analyses of unit roots and cointegration were done in Gauss 6.edu/maxpages/faculty/cdkao/working/npt. who estimates country-speciﬁc cointegration vectors for the G7 economies for 1950 to 1992. the coeﬃcients reﬂect the countries’ own business cycle dynamics and adjustment speeds. Bottazzi and Peri (2007) use an approach similar to mine.1. The adjustment coeﬃcients on the cointegration errors show how strongly the variable on the left-hand side adjusts to any deviations from the cointegrating relationship.and ﬁnds . although their second stage is a stationary VAR instead of my single-equation models.maxwell. 10 11 http://www.
I now have all ingredients for the two-stage estimation approach.1 shows the simple correlation coeﬃcients between the natural logarithm of the level of GDP per capita and the other variables using averages over the most recent 10-year period available. Also. 1994-2003.12 Estimation results After assessing the diﬀerent growth theories in chapter 2 and the diﬀerent candidate variables in chapters 4 through 8 and after evaluating the empirical techniques.71 and 0.1 Correlation analysis Before starting with the cointegration analysis and building the forecasting models.72. with correlation coeﬃcients of 0. In the 21 rich countries there is a very high correlation between GDP per capita and the two per capita capital stocks for physical and human capital. The emphasis in this chapter is on the result of pair-wise cointegration that was derived from diﬀerent theories in chapter 2.77. The other variables do not show a high correlation with GDP per capita. 12. it is helpful to look at the correlations among the variables chosen. there is a high correlation between the 30-49 age share and the level of education and .a strong negative correlation between the education level and population growth. However.in the full sample only . there is also a high correlation among the two capital stock variables. Not surprisingly. This seems to support the idea that countries with a high level of education have a lower birth rate and therefore a higher share of prime-age adults. Again. being surrounded by high-income countries . The next highest correlation coeﬃcient for the income level in the ﬁrst data column is with the share of the population aged 30 to 49. the coeﬃcient is even higher in the full sample at 0. The ﬁrst column in table 12. This correlation is even higher for the full sample with all 40 countries: high-income countries tend to have high stocks of both physical and human capital. The fact that we do not see this in the 21 rich countries chosen could be related to immigration as an important driver of population growth in some countries.
40 -0. Correlations of the levels of key variables (1994-2003 avg.26 -0.67 -0.138 12 Estimation results Table 12.66 -0.2. for example the negative correlation between trade openness and hours worked in the rich country sample. Open ln Ysp 30-49 Pop.03 -0.46 1.89 0.00 0.46 -0.c.67 0.07 1.00 0.00 correlates negatively with population growth.00 -0. ln K p.2 keeps the levels of the three stationary series age share.49 0. This supports my . Education Openness ln GDP space Age share 30-49 Pop.c.28 1.) 21 rich countries ln Y p.08 1. Educ.22 0.01 0.c.08 1.00 0.06 1.28 1.00 0. population growth and investment share.15 1.46 -0. GDP growth shows a high correlation with the changes in the two capital stocks and with changes in hours worked in the rich countries.63 -0.56 1. Inv/Y ln GDP p.34 0.89 -0.09 0.33 0. All 40 countries ln Y p. Open ln Ysp 30-49 Pop.43 -0. Educ.11 -0.72 -0.46 1.51 0.37 0. Since the primary purpose of this study is to explain changes rather than levels of GDP. growth Invest/GDP Hours p.08 0.00 0.13 1.35 -0.00 0.51 -0.00 0. This is in line with the theoretical considerations outlined in this study.c. Table 12.77 0.00 -0.98 0. it is helpful to also analyze the correlation of the changes in these variables over the period 1993 to 2003.27 0.c. growth Invest/GDP 1.c.c.09 0.00 0.02 0.01 -0.13 0.00 -0.13 -0.01 0.23 0.11 0. ln Capital p. ln Capital p.00 -0. Inv/Y ln GDP p.01 1.68 0.57 0.36 0.29 1. The slight negative correlations between GDP growth and both spatial GDP and the 30-49 year age share in the rich countries already raises some doubts about the explanatory power of these two variables.41 -0.11 -0.45 0.31 1.16 1.00 0.74 -0.36 0. Education Openness ln GDP space Age share 30-49 Pop. changes in the capital stock per capita correlate positively with population growth in the rich country group and in the overall sample.29 0.71 0. There are some correlations between some variables which are not easy to explain.00 0.45 0. ln K p.00 -0.c.13 0. More openness possibly goes hand in hand with higher incomes and more leisure. Interestingly.c.1. which appears to be the result of the clustering of rich countries in Europe referred to in section 8.17 0.
2. the diﬀerent types of unit root tests allow insights into which series are trend-stationary and whether that trend is common or country-speciﬁc.75 1. physical capital.54 0.31 0.26 -0.14 0. 12.00 -0.64 1.39 0.12 -0.05 0.27 0.75 1.49 1.26 1. The panel unit root tests ﬁnd six series that appear to have a unit root in the 21 rich countries and 19 emerging markets for the period 1971 to 2003: GDP.04 1.00 -0.07 -0.31 0.41 0.48 1. p.06 0.38 0.00 0.11 1.08 -0.15 -0.2 Panel unit root tests Much can also be learned from careful pretesting of the time series.2 Panel unit root tests Table 12.00 0.39 1.22 0.09 0.46 1. openness.10 0. growth Invest/GDP 1. The negative correlation between spatial GDP growth and changes in trade openness indicates that these two variables indeed measure two diﬀerent things: countries do not trade more just because their neighbors grow strongly. growth Invest/GDP d Hours p.20 0.04 0.56 0.13 0.12.00 0.00 0.34 0.31 -0.00 0. In particular. d Education d Openness d ln GDP sp Share 30-49 Pop.58 0.21 -0. .06 0.00 -0.66 0. Inv/Y d ln GDP p.c.00 skepticism about the neoclassical idea that high population growth depresses the available capital per worker.12 0.49 -0.38 0.02 0. all 40 countries d ln Y pc d ln K pc d Educ d Open d ln Ys 3049 Pop. In sum.14 0.13 1.34 0. years of education.c.00 0.00 0. p.00 0. spatial GDP and hours worked.68 0.18 -0.16 1.00 -0.28 -0.11 1.24 0. Correlations of changes between 1993 and 2003 21 rich countries 139 d ln Y pc d ln K pc d Educ d Open d ln Ys 30-49 Pop.11 -0. this ﬁrst look at the full dataset does not yield many surprises.14 -0.34 0.26 -0.19 -0.08 0.00 0. There is also a positive correlation between changes in human capital and changes in hours worked in the rich countries: quality and quantity of labor input do not appear to be substitutes. but rather tentative support for the theoretical considerations outlined earlier.00 -0.31 1.10 -0.00 0.01 0.15 0.05 1. Inv/Y d ln GDP pc d ln Cap.00 -0.19 0.c. d ln Cap. d Education d Openness d ln GDP sp Share 30-49 Pop.c.
Surprisingly. I will make use of this insight in the forecast competition in chapter 13. I used the program ub. The overall conclusion is that the capital stock per capita is an I(1) variable. The p-values are close to unity. The IPS-LM test also does not reject the null hypothesis. The ﬁrst diﬀerences are stationary both for the rich countries and for the emerging markets. the IPS-LM (and marginally also the UB test) reject the null hypothesis of I(1) in the speciﬁcation with intercepts for the emerging markets. with hardly any sign of a trend. The lag length was ﬁxed at 3.000. The two IPS tests which use heterogeneous alternative hypotheses now come close to rejecting the null. the overall conclusion on the years of education is not straightforward. although the t-statistic is relatively low. The overall conclusion is that the log of per capita GDP appears to be I(1). The conclusion is not quite as clear for the emerging markets. Therefore. For the natural logarithm of the stock of physical capital per capita over the 1971-2003 period. but not as conﬁdently. For the 19 emerging markets. eﬀectively ﬂat) while for many rich countries it might be 1 There were surprisingly large diﬀerences in reported statistics for the Breitung test. For the LLC and IPS tests I used the NPT 1. All tests were calculated in Gauss.140 12 Estimation results The age shares appear to be stationary. the two tests with the homogeneous alternative still cannot reject the null. In ﬁrst diﬀerences. For the ﬁrst diﬀerences of human capital the UB test does not reject the null of unit roots. For Breitung’s UB test. but no trends shows a unanimous result. but three of the four tests still indicate I(1) series.prg available at his website. Surprisingly. When a trend is added to the speciﬁcation. For many emerging markets the series might be I(0) (i. all four tests see no reason to reject the null of a unit root in the rich countries when a trend is not included. EViews gives similar results for the IPS test. e. EViews ﬁxed the problem that I alerted them to in March 2006. Allowing country-speciﬁc intercepts. The null hypothesis of a unit root in all series cannot be rejected at conventional thresholds. but there are some signs of non-stationarity in the emerging markets. When a trend is added. This indicates that a country-speciﬁc trend probably has signiﬁcant explanatory power for per capita GDP in many countries. the unit root hypothesis is strongly rejected for the rich countries.3. The LLC is somewhat less conﬁdent. LLC and IPS-t tests without and with time trends. . the LLC is now more conﬁdent in not rejecting the null. keeping in mind that the development is clearly country speciﬁc.439 in the rich countries and the IPS-t test with 7. the UB test still does not reject the null of I(1) series. The average years of education also are clearly not stationary series according to the UB. This is also the case when a trend is added.1 The natural logarithm of real GDP per capita in 2000 international dollars appears to be clearly I(1) over the 1971-2003 period. The UB test with a t-statistic of 5. the tests all indicate that the series are I(1). which is in line with visual inspection of the data. I will treat the series as I(1) below.342 are clearest with a p-value of 1.
Panel unit root tests 141 21 rich countries 19 emerging markets Homogeneous Heterogeneous Homogeneous Heterog.00) 7.566 (1.2 Panel unit root tests Table 12.56) -1.52) (0.31) ln GDP Space -2.889 (0.12.93) n/a n/a -3.929 (0.349 (1.019 (0.958 (0.99) Hours p.753 (0.00) -0.27) Share 30-49 -3.477 (0.86) -0.477 (1.00) n/a n/a 13.504 (1.601 3.00) 1.37) 3.773 (0.138 (0.145 (0.455 (0.815 (0.222 (0.052 0.32)7 -2.551 (0.00) n/a n/a 2.128 (0.97) 7.276 (0.88) -3.327 (0.541 (0.c.414 (0.99) 2.342 (1.96) -1.132 -1.723 (0.99) n/a n/a 1.66) 9.129 (0.530 (1.236 (0.09) -5.01) n/a n/a 2.c.38) ln Capital p.96) Sample length is 1971 to 2003.491 (0. I(2).102 (0.610 (0.159 (0.00) n/a n/a -3.522 (0.11) 4.81) 0.27) -0. -0.291 (1.117 (0.68) 0.732 (0.695 (0. Lag length ﬁxed at 3 for all tests.00) 3.086 (0.387 (0.019 (0.00) ln Capital p.195 (0.627 (0.91) 0.94) 0.68) Education 1.00) -2.00) (0.814 (1.035 (0.3.49) 1.356 (1.088 (0.530 (1.797 1.665 (0.121 (0.117 (0.08) 0.c.29) ln GDP Space 2.63) -2.00) 11.00) Share 50-64 1.23) -3.96) -0.439 (1. 0.316 (0.03) 1.153 (0.00) -2.58) time trend) 2.658 (0.692 (0.140 (1.409 (0.349 4.279 (1.00) 1.00) 0.349 (1.036 (0.99) 5.871 (0.897 (0.274 (1.c.45) -0. only intercept included.00) (1.025 (1.241 (0.462 (0.75) -4.329 (0.99) 14.560 (0.627 (0.886 (0.109 (0. .25) 0.00) 5.99) n/a n/a -0.046 (1.c.59) 0.37) -0.99) 4. 5.56) 1. 0.721 (1.00) (1.99) 4.519 (1.323 (0.886 5.13) -4.287 (0.42) 3.878 (1.87) Openness -0. The LLC tests take adjustment terms for ﬁxed time dimension T into account.122 (0.92) -3.087 (0.99) 0.98) (0.77) 0.00) (0.00) 3.99) -0.00) 3.98) 1.406 (0.000 4.66) -3. These are important insights that might help the forecasting model to diﬀerentiate between countries.24) -3.32) 9.746 (0. 2.708 (0.08) -1.00) -1. -0.149 (0.25) -0.05) 3.77) 1.341 (0.175 (0.00) 0.00) -0.69) -3.00) 1.666 (0.73) (1.00) 10.00) 3.616 (0.566 (0.00) -4.269 (1.666 (0.00) 7.00) 0.743 (0.847 0.184 (0.816 (1.74) 0.453 (0.45) 1.00) Share 50-64 -3.00) 5.27) 1.00) 0.341 (0.861 1.470 (0. Variable UB LLC IPS-t IPS-LM UB LLC IPS-t IPS-LM (only intercept) ln GDP p.00) (with intercept and ln GDP p.492 (0.882 (0.00) n/a n/a 0.00) -0.00) -3. P-values in parenthesis.99) (0.228 (0.75) 2.535 (1.00) 1.99) Education 2.486 (0.255 -0.13) 0.231 (0.91) 0.99) 3.297 (0.699 (0.00) -1. Bold: reject H0 at the 5% level and conclude that series are stationary.53) Share 30-49 3.929 (0.00) Hours p.198 (0.131 (0.656 (0.292 (0.87) 1.725 (0.59) 3.00) -1.98) -0.736 (1.99) Openness -0.c.384 (0.00) 0.
capital stock per capita. . this would explain why they are still relatively poor today. The hours worked per capita also seem to be I(1) in the rich countries.2 The UB test strongly rejects the null hypothesis that the ﬁrst diﬀerences are I(1). The Newey-West method is used to calculate the length of the kernel window for the gFMOLS. This is exactly what the panel cointegration analysis presented in this section ﬁnds. This is at odds with Osterholm (2004) who cannot reject the null hypothesis using the Levin-Lin test on the crosssectionally demeaned log-level of the age structure in 20 OECD countries. evolutionary or the augment Kaldor model) point to pair-wise cointegrating relationships among the main variables GDP per capita. three tests reject the null for the emerging markets. my main focus is on the group mean FMOLS estimator as proposed by Pedroni (2000) and on the two-step method proposed by Breitung (2005). Interestingly. it is not admissible to have two variables on the right-hand side which are themselves cointegrated. which is not surprising given the construction of the data.1). Lucas. All four tests with only the intercept are far away from rejecting the null. the share of 50 64-year olds seems to be I(1). Three of the four test statistics strongly reject the null hypothesis for the ¨ two age shares in the rich countries. The results from the Breitung test were 2 The UB test with trend suggests that there might be a common trend. In my framework. The test statistics drop signiﬁcantly when a trend is included in the two IPS tests with their heterogeneous alternatives. as that share has tended upward for most of the sample period .142 12 Estimation results The openness series for the 21 rich countries appear to be I(1) as well. However.3 Panel cointegration test As shown in chapter 2 on growth theory. The chart of hours worked indicates that the I(1) result is driven in particular by observations from the 1970s (see ﬁgure 4. When adding a trend the test statistics tend to go up signiﬁcantly. since the four test statistics unanimously do not reject the null hypothesis in the tests with intercepts. the natural logarithm of the constructed spatial GDP series is clearly I(1) with no signs of either a common or a country-speciﬁc trend in either the rich countries or the emerging markets. Importantly. most of the theories (augmented Solow. 12. education per capita and openness.but at low levels relative to the rich countries. This points to a signiﬁcant portion of poor countries having seen no major change in this driver of growth since the early 1970s. As outlined in chapter 11 on the empirical methods. In the emerging markets. Estimation is done in Gauss using the modiﬁed NPT kindly made available by Hlouskova and Wagner for the gFMOLS and with Breitung’s own code for the 2-step estimator. The downward trend stopped in most countries around 1980. The lag length is set at 3 throughout following Breitung (2005).
12.3 Panel cointegration test
used to recover the ﬁxed eﬀects and to calculate the cointegration errors for all 40 countries. These errors are pictured in ﬁgures 12.1 through 12.4 at the end of this chapter. GDP and capital stock move in parallel Some 50 years ago, Nicholas Kaldor established the stylized fact that GDP and the capital stock move in parallel: a 10% rise in GDP tends to go hand in hand with a 10% rise in the capital stock. In the cointegration analysis this implies a coeﬃcient of unity in the long-run relationship between the natural logarithm of GDP per capita and the natural logarithm of the capital stock per capita. This is almost exactly what the two panel tests ﬁnd as table 12.4 shows. The gFMOLS test sees a coeﬃcient of 1.07 for the 21 rich countries, while the 2-step test ﬁnds one of 0.96. Both t-statistics are very high. Not surprisingly, there is an almost 100% correlation between the capital-output ratios discussed in chapter 5 and the constants from the cointegration estimation. For the 19 emerging markets the cointegration vector diﬀers from the Kaldor values: a 10% rise in physical capital goes hand in hand only with a roughly 7% rise in GDP. However, the relationship is still estimated with a high statistical signiﬁcance. One possible explanation is that capital is not used very eﬀectively in emerging markets.
Table 12.4. Panel cointegration tests, 1971 to 2003 21 rich countries 19 emerging markets gFMOLS 2-step gFMOLS 2-step 1.07 (76.3) 0.23 (38.8) 0.02 (17.2) 0.98 (64.4) -1.08 (-4.2) 0.22 (40.3) 0.04 (15.0) 0.12 (17.0) 0.96 (29.2) 0.20 (31.5) 0.04 (7.9) 0.91 (34.0) -0.59 (-1.5) 0.20 (23.0) 0.03 (9.8) 0.10 (7.4) 0.66 (47.1) 0.03 (13.9) 0.003 (5.8) 1.17 (36.4) n/a n/a 0.33 (29.3) 0.004 (5.7) 0.01 (5.3) 0.71 (39.2) 0.23 (24.7) 0.017 (4.8) 1.23 (26.7) n/a n/a 0.35 (27.2) 0.012 (3.0) 0.02 (4.1)
Left side ln GDP p.c. ln GDP p.c. ln GDP p.c. ln GDP p.c. ln GDP p.c.
Right side ln Capital p.c. Education Openness ln GDP Space Hours p.c.
ln Capital p.c. Education ln Capital p.c. Openness Education Openness
Table shows coeﬃcients from bi-variate panel cointegration tests. Null hypothesis is no cointegration. T-statistics in parenthesis.
While the common cointegrating vectors are estimated with a high degree of statistical signiﬁcance in both samples, there are some countries that do
12 Estimation results
not show that average relationship. Most notably, the capital stock in Ireland rose much less quickly than GDP over the last two decades, leading to a large positive cointegrating error in 2003 as ﬁgure 12.1 at the end of this chapter shows. By contrast, the capital stock in Japan rose signiﬁcantly more quickly then GDP, leading to a large negative cointegrating error in 2003. Neither result comes as much of a surprise. The case of Ireland may be linked to a rising share of Irish GDP going to foreigners who provide intellectual capital that produces GDP in Ireland without much physical capital. Irish GNP in 2005 was almost 20% lower than GDP. Excluding these two countries would raise the signiﬁcance of the estimated coeﬃcients even further. One more year of education: 20% higher GDP The cointegrating relationship between GDP per capita and the average years of education is also estimated with high statistical signiﬁcance. The coeﬃcients of 0.23 (gFMOLS) and 0.20 (2-step) in the 21 rich countries indicate that an additional year of education goes hand in hand with a 20% rise in GDP per capita in the long run. Compared with the microeconomic returns of less than 10%, this appears high at ﬁrst sight. However, physical capital (and openness) will also rise with higher education. The 20% therefore captures both direct and indirect eﬀects of an increase in human capital over time, while the microeconomic studies only uncover partial eﬀects. Since one year of education is roughly a 10% increase in the 21 rich countries, there is no support for the prediction of the augmented Solow model that human capital and GDP grow at the same rate. The augmented Kaldor model is a more appropriate speciﬁcation, with ω in equation 2.23 equal to roughly 0.2. As in the case of physical capital, the common cointegrating slope does not ﬁt all countries, as ﬁgure 12.2 on page 149 illustrates. GDP rose more quickly than the rise in human capital would have indicated in Portugal and Ireland. By contrast, GDP rose more slowly in Spain and Italy. Measurement error and changes in the country-speciﬁc eﬃciencies in the usage of human capital may help explain these discrepancies. But these large errors may also point to slow GDP growth ahead in Portugal as GDP may move back into line with one of its fundamental determinants. In the emerging markets sample there is also a highly signiﬁcant cointegrating relationship between GDP and human capital - with the 2-step coeﬃcient slightly higher than in the rich countries. Not surprisingly, the heterogeneity of the cointegrating vectors is even larger than in the rich countries. China and Taiwan saw GDP grow much faster than indicated by human capital throughout the sample period. Nigeria, South Africa and Brazil are the mirror cases. Measurement error for human capital may play an important role here. Alternatively, other factors such as inward foreign direct investment may lead to diﬀerent levels of “imported” human capital. These results support the model of Azariadis and Drazen (1990, p. 504), who argue that rapid economic growth cannot occur ”without a high level of
12.3 Panel cointegration test
human investment relative to per capita income.” They focus on the ratio of literacy rates to per capita output, but already indicate that average years of education would be a more appropriate measure. Openness and spatial GDP cointegrate with GDP as well Not surprisingly, the cointegrating relationship between per capita GDP and trade openness produces relatively low t-statistics and there is some disagreement between the two tests about the size of the coeﬃcient. However, the coeﬃcient has the expected positive sign and is still estimated with a high level of statistical signiﬁcance. The coeﬃcients of 0.025 (gFMOLS) and 0.036 (2-step) indicate that a ten percentage point increase in the trade share leads to a roughly 30% rise in GDP per capita in the long run. In the emerging markets sample the estimated coeﬃcients are much smaller, with only the 2-step estimate of 0.018 close to the numbers from the rich countries. Figure 12.3 shows the errors for this cointegrating relationship in the rich countries. Both tests agree that there is a very close relationship between a country’s per capita GDP and the per capita GDP of its neighbors (with Ireland again an outlier as ﬁgure 12.4 shows). A 10% rise in neighbors’ GDP goes hand in hand with a roughly 9.5% rise in home country GDP in the rich country sample. In the emerging markets sample a 10% rise in neighbors’ GDP goes hand in hand with a 12% rise in home GDP. Possibly this can be interpreted as a larger dependence of the emerging markets on developments abroad. The two tests do not ﬁnd a strong cointegration between GDP and hours worked per capita. The t-statistics are quite low given the high power of the panel tests. Adding hours worked to the cointegration between GDP and human capital in the rich country sample is an economically meaningful experiment. The Breitung test ﬁnds a signiﬁcant role for hours worked and a slightly higher (and more signiﬁcant) coeﬃcient on human capital than without hours worked. However, the gFMOLS does not ﬁnd a signiﬁcant role. Because of this result and the additional complexity from adding more variables, I do not pursue this route further and stick to the pair-wise relationships. So far, all cointegration tests had GDP on the left-hand side. Other permutations conﬁrm the theoretical insight that all major variables are pair-wise cointegrated with transitivity holding as table 12.4 shows. The relationship between physical capital on the left-hand side and human capital is estimated with high signiﬁcance and the coeﬃcients are almost exactly equal to those from the relationship between GDP and education. This makes sense given the coeﬃcient of unity between GDP and physical capital. The same transitive result holds for the cointegration between GDP and openness. The coeﬃcients are very similar in magnitude to those from the GDP-capital relationship. Here again, the coeﬃcients are estimated with relatively low precision. Finally, the cointegration relationship between education and openness is also signiﬁcant with the coeﬃcients roughly equal to the ratio of the coeﬃcients from the individual relationships with GDP.
4 The short-run forecasting models Having established the signiﬁcant cointegration between the main variable of interest.04 and a median of -0. insigniﬁcant (10% level) coeﬃcients were eliminated. For the changes of the non-stationary variables the index l runs from 0 to 1.1)yt = c + ˆ j=1 k=1 πkj CIk. and the four variables physical capital per capita. the four stationary cointegrating errors for all 40 countries were calculated using the ﬁxed eﬀects and slope coeﬃcients from the Breitung test (and multiplied by 100 for easier reading). The augmented technological progress function 2. Since the models will be used to calculate out-of-sample forecasts for the period 2001-05. the general form of the empirical equation for each country i = 1 to 40 is 2 4 1 5 (12. the second stage in my two-stage approach is to estimate short-run models using the stationary series. Among the 21 rich countries. First diﬀerences of the ﬁve nonstationary variables as well as the stationary age shares complete the list of variables for the stationary models. Also. so estimation starts from a small model with j = 1. these 21 coeﬃcients are too much to estimate. Also. As indicated. allowing the error in the previous one or two years to aﬀect GDP growth today. From here. with a range from -0. and γ = 0. In the end. In addition. In addition.86 to -0. and then spatial GDP. the change in hours per capita is added to take care of possible changes in work eﬀort that were not accounted for in equation 2. The cointegrating errors capture the possibility that adjustment is not instantaneous. allowing contemporaneous changes or last year’s changes to aﬀect today’s GDP growth. l = 0. The selected country-speciﬁc models generally conﬁrm the theoretical model outlined above. The index j runs from 1 to 2 for the cointegrating vectors. age shares and lagged GDP tested for inclusion. positive coeﬃcients on the cointegrating errors were rejected because they would lead to unstable models. With just 30 observations for each country. human capital. the sample length is just 1971 to 2000. GDP per capita.21 is the basis for the models for GDP growth. other lag speciﬁcations tested.38.t−l + δaget + γ yt−1 + ˆ ˆ t Where c is the country-speciﬁc constant. but they diﬀer signiﬁcantly across countries. changes in spatial GDP enter with one additional lag. Therefore.21. .t−j + l=0 s=1 βsl xs. However. openness and spatial GDP. the age structure and lagged own GDP growth may enter the model. πkj and βkl are coeﬃcients on the k = 1 to 4 cointegrating errors and on the s = 1 to 5 changes of the non-stationary variables. the average country-speciﬁc model chosen includes ﬁve variables on the right-hand side. spatial GDP was not included in the ﬁrst estimation through either the cointegrating error or the changes of spatial GDP. δ = 0.146 12 Estimation results 12. 15 detect a signiﬁcant negative coeﬃcient on the error from physical capital.
12.4 The short-run forecasting models
Also, 10 countries show a signiﬁcantly negative coeﬃcient on the error term from human capital with an average of -0.14. The error from trade openness plays a role in just ﬁve countries, with an average coeﬃcient of just -0.07. The error from spatial GDP appears in only four equations. Among the other variables, current growth in the capital stock appears in 16 of the 21 countries and the change in hours worked in 15. The changes in openness appear in 6 and the changes in human capital in four countries. The share of people aged 30 to 49 is signiﬁcant in only ﬁve countries. The models for the 19 emerging markets show similar results, with the error from the capital stock cointegration and the concurrent change in the capital stock playing the biggest roles. The errors from openness and spatial GDP appear to be slightly more important in the emerging markets sample. And the age shares have a higher explanatory power. Table 12.5 shows four examples of short-run models.
Table 12.5. Models for per capita GDP growth 1971-2000 Panel Rich France USA Panel EM Indonesia Turkey constant CI yk (-1) CI yh (-1) CI yo (-1) CI ys (-1) d capital d education d open d y space d hours d ypc (-1) adj. R2 Schwarz cr. -0.05 -0.06 1.52 -0.27 -0.19 -0.05 -0.04 1.50 9.28 0.12 0.46 0.19 0.46 0.97 0.10 0.47 4.13 0.60 0.83 2.94 2.95 0.32 5.65 0.54 5.21 0.60 4.98 -3.62 -0.05 -3.03 -0.13 -0.42 -0.03 0.60 -0.36 1.72 1.24
All coeﬃcients signiﬁcant at least at the 10% level. “CI yk” is the cointegration error between GDP and physical capital per capita
In addition, ﬁxed eﬀects panel models were estimated for both country groups (columns in boldface in table 12.5). The cointegrating errors from physical capital and human capital have a signiﬁcant inﬂuence on GDP growth: the negative sign is in line with the expectation from theory. It indicates that a 10% excess of GDP per capita beyond the level indicated by physical or human capital in the previous period will be reduced by 0.5% in the current period. The current period change in physical capital has the expected large inﬂuence on GDP growth. Likewise, a 1% rise in hours worked boosts GDP growth by 0.45% in the same period. Changes in openness also have a signif-
12 Estimation results
icant eﬀect on GDP growth. These three variables mostly capture business cycle phenomena. To test whether the levels of human capital and openness have an inﬂuence on GDP growth as suggested by the models with scale eﬀects, the deviations from group-speciﬁc period means were included as well. The deviation of human capital was insigniﬁcant in the panel for the 21 rich countries and the deviation of openness was just signiﬁcant at the 5% level. If a country has above-average openness, then its GDP tends to grow slightly faster. However, this is not necessarily evidence for the impact of the level of openness because the deviation from average is highly persistent and constants are also included in the panel regression. Models for physical capital A similar search was performed for the models for physical capital, the empirical equivalent to equation 2.25. In this model, only the lagged cointegrating error and lags of changes in physical capital and GDP were allowed to enter. (12.2) ˆ kt = c +
πj CIykt−j +
ˆ βj kt−j
+γj yt−j + ˆ
The coeﬃcient on the cointegrating error between GDP and physical capital now has to enter with a positive sign: if GDP is above the level indicated by the long-run link, then physical capital should react and accumulation should pick up. This is the case in 7 of the 21 rich countries and in 8 of the 19 emerging markets. Lags of the percentage changes of physical capital and GDP play an important role in most models. Overall, the results indicate that capital contributes signiﬁcantly to narrowing any deviation of capital and GDP from their long-run relationship in line with the Kaldor model.
12.4 The short-run forecasting models
100* Deviation of ln (Y pc) from constant + 0.9629 * ln(K pc) Ireland
how far is Y above (+) or below (-) the long-run level indicated by K
-20 1971 1976 1981 1986 1991
Fig. 12.1. Errors from cointegration between GDP and capital stock
100* Deviation of ln (YPCP) from constant + 0.2009 * Education
-30 1971 1976 1981 1986 1991 1996 2001
Fig. 12.2. Errors from cointegration between GDP and years of education
0357 * Openness 40 Norway 20 0 −20 −40 Ireland −60 1971 1976 1981 1986 1991 1996 2001 Fig. Errors from cointegration between GDP and spatial GDP .4. 12. 12.3. Errors from cointegration between GDP and trade openness 50 100* Deviation of ln (Y pc) from constant + 0.150 12 Estimation results 60 100* Deviation of ln (YPCP) from constant + 0.9133 * ln(Y space (t-1)) Ireland 40 30 Norway 20 10 0 -10 Switzerland -20 -30 1971 1976 1981 1986 1991 1996 2001 Fig.
13 Forecast competitions and 2006-2020 forecasts One purpose of this study is to derive a set of forecasts for the growth rates of GDP for the years 2006 to 2020. hours worked and age structures. Once values for each year until 2005 are calculated. I use half of the annual own country change over the previous 10 years and half of the average change in the peer group. Forecasts for the year 2002 are based on forecasted values of GDP for 2001 which are used to calculate the cointegration errors for 2001 and so on. it is now straightforward to construct forecasts for per capita GDP growth. . 13. The insights from this competition will be used to derive models that are likely to provide the most robust forecasts until 2020. the United Nations provide forecasts until 2050. The exercise uses only data available as of approximately spring 2001. For the age structure. the geometric average of the ﬁve annual observations is used for comparison with realized average values over 2001-05. This use of 5-year averages avoids comparison with annual realizations that are mostly driven by business cycle inﬂuences. Since the growth rate of physical capital depends only on past changes of GDP and capital as well as on the cointegration error as estimated in the models just shown. For the more volatile trade openness and hours worked. The forecasts are at an annual frequency. no further assumptions are needed. trade openness. I simply extrapolate the slow-moving series using the average annual change over the years 1995-2000. The only missing ingredients are forecasts for human capital.1 Forecast competition 2001-2005 Using the short-run models for per capita GDP growth and for growth of the capital stock estimated in the previous chapter. An important intermediate step is to see how well the fundamental models presented in the previous chapter perform in an out-of-sample forecasting competition. For human capital.
Marcellino (2006) shows that ”in general linear time series models can hardly be beaten if they are carefully speciﬁed.1 show. No intercept correction is done. However. As the fourth data row in table 13.152 13 Forecast competitions and 2006-2020 forecasts Two sets of forecasts are derived for average annual per capita GDP growth for the period 2001-05: One set uses the country-speciﬁc short-run models. Not surprisingly. A simple average of the two fundamental models reduces the RMSE. the RMSE is still higher than in the two fundamental models for both sub-samples. this candidate performs well for the USA (and Switzerland and the UK). Any overshoot of GDP in 2000 relative to the time trend is assumed to correct by 2005. but it performs worst of all for the full sample with an RMSE of 2.49 for the panel models as shown in the last column in the table. This competitor clearly underperforms relative to the two fundamental models especially in the emerging markets sub-sample. (4) The fourth candidate is a Hodrick-Prescott (HP) ﬁlter ﬁtted to the data from 1981 to 2000 (multiplier = 100). this model produces a larger root mean squared error of 2. and therefore still provide a good benchmark for theoretical models of growth and inﬂation”. By contrast. six competitors were constructed: (1) The na¨ steady-state forecast.1 shows. The literature usually takes this exogenous rate g as 2%. with 1981-2000 as the base period. (2) The second candidate is a simple linear trend ﬁtted to the natural logarithm of GDP in each country. the largest errors stem from the forecasts for Nigeria and Colombia.1 shows. Indeed. The country-speciﬁc models did not foresee the growth slowdown in the Netherlands and in Belgium. Using data for the USA. Comparing these forecasts with the actual outcome using growth rates published in the IMF’s April 2006 World Economic Outlook shows a positive bias of 0.05 for the emerging markets group.49 percentage points for the country-speciﬁc model in the rich countries and a negative bias in the emerging markets for both models as the ﬁrst and fourth data columns in table 13. This indicates that Marcellino’s conclusion may not be valid for most countries other than the USA. The root mean squared error (RMSE) across all 40 countries is 1. while the other set uses the coeﬃcients from the two panel estimations. the forecasting performance improves signiﬁcantly relative to the trend model without intercept correction. To see how the fundamental models perform relative to other forecasting approaches. The increased ﬂexibility relative to the simple time . (3) The third candidate uses the same country-speciﬁc linear trend as candidate 2. even if this 2% is mainly derived from US experience. Indeed. This candidate does not include an intercept correction. the panel model wrongly expected a signiﬁcant growth slowdown in China.47 percentage points for the country-speciﬁc models and 1. where all countries’ per capita GDP ıve grows at the same rate because the same technology is available everywhere and is improving at a constant rate. but performs an intercept correction in line with the insights from the forecasting literature reviewed in chapter 10.39 as the ﬁfth data row in table 13.
The RMSE in that group is slightly below that of the fundamental models.00 0.90 -0. markets All countries Bias MAE RMSE Bias MAE RMSE Bias MAE RMSE Fundamental c-spec.41 0.35 1.84 -0. Consensus Economics is a private company that surveys ﬁnancial institutions and research institutes.90 0.05 1.14 1.91 0.13 1. although still above the two fundamental models.84 0.15 2.49 1. In percentage points. MAE: Mean absolute error.37 1.48 1. By contrast.00 0.25 1.78 Panel & trend w/ IC 0.82 0.69 2.16 1.44 1.28 1.26 1.49 1.41 1.01 -0.28 0. RMSE: Root mean squared error.28 1.68 0.19 -0. (5) The HP ﬁlter with intercept correction surprisingly performs worse than the alternative without intercept correction in the rich country group.36 0.1 Forecast competition 2001-2005 Table 13.48 1.00 -0.15 1.54 1.18 0.13.97 1.45 1.96 0.47 1.47 0.91 1.1.33 0. similar to the method used for constructing the forecasts for openness and hours.13 1.72 0.01 1.93 -0.39 1.98 0.27 1.39 0.98 0.60 Memo: combinations Country & HP w/o IC 0.67 0.64 1.94 0.14 1.84 0.95 1.82 0.31 1. Fundamental panel Simple average Na¨ 2% forecast ıve Trend without IC Trend with IC HP ﬁlter without IC HP ﬁlter with IC HP ﬁlter w/o IC 50-50 Consensus for 12 Private organization 0.04 0.21 1.35 1.52 0. Several private and international organizations regularly publish medium and long-run growth forecasts and it is worth comparing their performance with that of the fundamental models and the simple time series models.41 1.35 1.31 0.84 1. this candidate performs quite well with an overall RMSE below the previous three competitors.22 0. Each April edition of their publication Consensus Fore- .56 1.00 0.12 1.35 1.26 0. Indeed.18 1. but intuitive competitor is a simple average of the countryspeciﬁc forecasts from the well-performing HP ﬁlter without intercept correction and the peer-group average forecast using the same ﬁlter. This average indeed shows the best forecasting behavior and even outperforms the fundamental models.64 -0.07 -0.36 1.63 2. bias and RMSE fall in the emerging markets group.28 1.25 Table compares forecasts for annual average per capita GDP growth over 2001-05 with realizations.34 1. (6) An ad-hoc.28 1.05 -0.12 -0.02 2.67 1.33 1.10 1.82 1.82 0.37 1.98 0.28 1.22 0.90 0.98 1. trend suggests a better forecasting performance.65 0. Results of the forecast competition 2001-2005 153 21 rich countries 19 em.21 0.91 1.82 1.75 0.28 0.13 1.42 1.31 0.19 1.23 1.00 -0.
the regression has to be along the cross-section dimension. combining them allows researchers to diversify across forecast errors . weights have to be derived.67. The RMSEs are higher than for my fundamental models.154 13 Forecast competitions and 2006-2020 forecasts casts includes averages for the long-term outlook for 12 large economies. As long as forecasts are not perfectly correlated. The regression chooses an intercept of 0. In order to generate forecast combinations for the 2006-20 period. Since I have only one forecast per country.2 Forecast combination As indicated in chapter 10. Lower forecast errors and variances would be the result. combining forecasts from diﬀerent models has repeatedly been found to lead to improved overall characteristics relative to individual forecasts. the larger the likely gain from combining them. This indicates that combining the forecasts derived from growth theory with very simple time-series forecasts should be a promising endeavor. The more the properties of the individual models diﬀer.1 The forecasts for 2001 to 2005 from the April 2001 edition show a very large bias (see table) and a root mean-squared error that is similar to the other competitors but slightly above the fundamental panel models. I subtracted actual population growth to generate per capita GDP growth. two were selected based on good performances in the forecast competition in table 13.9 percentage points. a combination of the country-speciﬁc fundamental models and the well-performing HP ﬁlter without intercept correction is calculated.just as combining diﬀerent assets in a ﬁnancial portfolio improves the characteristics of the overall portfolio. First.1) ˆf undam ˆtime y01−05 = a0 + a1 y01−05 + a2 y01−05 ˆactual Among the large number of possible combinations. The intercept a0 is included to allow the possibility that the underlying forecasts may be biased: (13. 13. . The March 2001 edition of their forecasts for per capita GDP growth for 2001-05 shows a larger bias than the six competitors for both the 21 rich countries and for the 19 emerging markets of 0. This is done with a simple regression of the realized values on the forecasts from the diﬀerent models that generated forecasts for 2001-05.1 and on the idea that combining a fundamental and a time series model is likely to generate the largest beneﬁt. Among the time series models (leaving aside the simple 2% model and the ad-hoc 50-50 Hodrick-Prescott model) the trend with intercept correction and the HP ﬁlter without intercept correction showed the best performance in the rich countries. a weight on the fundamental forecast 1 The survey refers to headline GDP growth. A US-based private research institute regularly publishes long-run forecasts for most economies around the world.
The cointegration vectors and the short-run models for GDP and the capital stock were re-estimated over 1971-1995 (although without spatial GDP and the age shares. especially when the simple average is used where the RMSE is lowest at 1. This is not too surprising given the high output volatility in these countries and because the Asian crisis happened during the forecast horizon. these results underline the strength of the fundamental models and the value of HP ﬁlters especially for the emerging markets. which were shown to be of minor importance in the earlier models). The exception is that hours worked are assumed to stay unchanged from 2010 onwards. for the two samples. a second competition was conducted over the 1996-2005 horizon.37 and 0. However. in alphabetical order six sets of forecasts . The regression chooses an intercept of 0.18. the best performer is the 50-50 HP ﬁlter with an RMSE of 1.37 and a weight on the HP forecast of 0. In the emerging markets the forecast errors are very large across all competitors.64 and weights of 0. The forecasts for the exogenous variables (human capital. only the consensus and the ”private organization” forecasts are not available here. openness and hours worked) were derived using the same simple rules as for the 2001-05 competition.13.3 shows.4 Forecasts for 2006-2020 155 of 0. the na¨ 2% forecast performs quite well also in the emergıve ing markets. Again. By construction the bias disappears and the RMSE for both sub-samples is much lower than for the individual models in table 13. Table 13. 13. The outline of forecasting theory in chapter 10 indicates that fundamental models do not necessarily have to be better over a longer forecasting horizon.35.4 Forecasts for 2006-2020 The insights of the previous sections are now used to generate forecasts for the years 2006 to 2020.1. the RMSEs are generally higher than in the case of the shorter forecast horizon. The forecasts for the exogenous variables are derived in the same way as for the two out-of-sample exercises above. The na¨ 2% forecast also does well ıve and so does the 50-50 HP ﬁlter with an RMSE of 1.28. 13. by construction the bias disappears and the RMSE declines. Together with the results from the 2001-05 competition.2 shows the results using the same competitors as for the 2001-05 competition.65.3 Forecast competition 1996-2005 To check the robustness of the results from the 5-year forecasting competition. Table 13. Not surprisingly. For the 21 rich countries. More surprisingly. the two fundamental models perform quite well. The second model combines the forecasts from the fundamental panel model and from the simple trend forecast with intercept correction.26.
13 Forecast competitions and 2006-2020 forecasts Table 13.2. Forecast competition over 1996 to 2005 21 rich countries 19 em. markets All countries Bias MAE RMSE Bias MAE RMSE Bias MAE RMSE
Fundamental c-spec. Fundamental panel Simple average Na¨ 2% forecast ıve Trend without IC Trend with IC
0.23 1.01 -0.42 0.96 -0.09 0.84 0.25 0.96 0.18 1.21 0.12 1.11
1.34 -0.01 1.90 1.29 -0.60 2.40 1.18 -0.30 1.87 1.33 -1.06 1.38 1.59 -0.56 1.99 1.39 -0.20 1.93 1.44 0.42 2.19 1.33 0.51 2.26 1.28 0.42 1.42
2.30 0.12 1.43 2.79 -0.50 1.64 2.12 -0.19 1.33 1.89 -0.37 1.16 2.44 -0.17 1.58 2.37 -0.03 1.50 2.61 0.13 1.65 2.64 0.18 1.64 1.65 0.15 1.22
1.86 2.14 1.69 1.62 2.04 1.92 2.08 2.06 1.47
HP ﬁlter without IC -0.13 1.17 HP ﬁlter with IC -0.13 1.09 HP ﬁlter w/o IC 50-50 -0.09 1.03
Table compares forecasts for annual average per capita GDP growth over 1996 to 2005 with realizations. In percentage points. MAE: Mean absolute error, RMSE: Root mean squared error.
for per capita GDP growth: from the country-speciﬁc fundamental model, the panel model, the HP ﬁlter without intercept correction, the trend with intercept correction, a combination of the HP ﬁlter and the country-speciﬁc fundamental model and a combination of the panel model and the trend with intercept correction. For comparison, the last column of the table shows the average realized growth rate over the 15 years to 2003.
13.4 Forecasts for 2006-2020 Table 13.3. Forecasts for annual growth of per capita GDP 2006-2020 Country Panel HP w/o Trend IC Cty&HP Panel&Tr 1989-03 Australia Austria Belgium Canada Denmark Finland France Germany Greece Ireland Italy Japan Netherlands New Zeal. Norway Portugal Spain Sweden Switzerland UK USA Argentina Brazil Chile China Colombia Egypt India Indonesia Israel Korea Malaysia Mexico Nigeria Philippines Singapore S. Africa Taiwan Thailand Turkey 5.3 1.7 2.3 0.7 -0.3 2.7 1.8 1.5 2.1 6.2 2.0 2.6 2.5 1.5 3.0 2.0 2.4 2.1 1.9 2.7 2.9 2.3 1.4 1.3 9.9 2.0 1.8 3.7 3.1 2.4 0.1 2.5 2.3 14.8 2.5 0.8 1.4 -0.6 2.3 2.7 2.0 1.3 2.0 -0.1 3.7 1.2 2.3 1.3 1.7 5.1 2.0 1.2 2.2 1.1 1.6 0.7 6.4 1.2 0.9 2.3 2.1 0.7 1.4 2.6 6.4 3.0 0.1 4.0 5.0 0.5 2.4 2.4 3.6 0.8 3.5 0.7 2.3 -0.6 1.6 4.4 2.6 1.9 1.9 2.4 1.9 3.0 1.8 1.3 3.0 7.3 1.5 0.8 2.2 1.8 2.6 2.5 3.0 2.4 0.9 2.3 2.1 -0.5 0.9 3.2 7.4 -0.1 2.4 3.9 0.9 1.7 4.1 2.8 1.8 0.1 1.6 2.6 1.6 3.8 1.6 1.0 2.1 2.0 1.9 1.5 1.6 1.7 1.6 1.8 1.7 5.7 1.7 1.9 2.1 1.3 2.6 3.0 2.7 1.5 0.7 2.1 1.9 1.0 0.7 4.8 7.9 1.2 2.0 3.7 3.4 2.1 5.6 4.4 1.1 0.7 1.1 4.4 0.3 5.5 4.7 2.0 3.3 1.8 2.0 1.5 1.0 2.4 1.8 1.6 2.2 4.9 1.8 1.9 2.2 1.7 2.5 2.1 2.3 2.1 1.6 2.3 2.3 1.4 1.4 2.0 6.3 1.4 2.0 3.1 2.1 2.0 1.8 2.3 2.0 6.2 2.0 1.6 1.6 1.5 1.9 1.9 2.1 1.8 2.1 1.1 2.6 1.7 2.1 1.7 1.9 4.6 2.0 1.8 2.2 1.5 2.2 2.0 4.0 1.6 1.2 2.3 2.1 1.3 1.4 3.3 5.8 2.2 1.4 3.4 3.7 1.6 3.5 3.1 2.4 1.2 2.3 2.5 1.6 2.4 2.9 3.0 2.1 1.9 1.8 1.4 1.5 1.5 1.4 1.6 2.2 6.1 1.4 1.6 1.9 1.5 2.4 2.4 2.6 1.5 0.6 1.9 1.8 0.5 0.4 4.3 7.8 0.9 2.0 3.8 3.2 1.7 5.1 4.3 1.4 1.0 1.2 3.6 0.6 4.8 4.3 1.8
13 Forecast competitions and 2006-2020 forecasts
The results in table 13.3 indicate that the models roughly agree on the outlook for Belgium, Germany, Italy and the UK, while they draw diﬀerent conclusions for Spain, Denmark and Australia. Across the emerging markets, agreement is strongest for Brazil, India and South Africa, while uncertainty is particularly high for Nigeria, Taiwan and South Korea. So far, all forecasts have been for GDP per capita. Table 13.4 on page 159 shows population growth rates as forecast by the United Nations and the forecasts for the growth rates of total GDP which are the simple sum of population growth and the per capita growth rates. The countries are ranked by total GDP growth from the model that combines the panel estimates and the trend model with intercept correction.
0 2.5 2.2 3.1 3.6 -0.7 4.1 3.3 2.3 0.8 3.7 4.6 3.2 1.3 1.1 2.1 0.1 2.0 8.6 2.1 1.4.5 2.8 1.1 -0.8 2.3 0.7 1.0 3.1 1. Forecasts for annual growth of total GDP 2006-2020 in % Population growth Country & HP ﬁlter Panel & trend China Ireland India Indonesia Malaysia Chile Turkey Spain Korea Philippines Thailand Colombia Singapore Mexico Nigeria Egypt USA Israel Australia Denmark Taiwan UK Brazil Norway Netherlands France Argentina Belgium New Zealand Portugal Greece Finland Canada Austria Sweden Germany Japan South Africa Italy Switzerland 0.1 -0.3 0.6 2.5 0.3 2.4 4.7 2.3 1.6 2.0 3.1 0.6 3.1 3.2 2.6 -0.4 -0.9 0.7 1.7 1.7 1.2 0.1 1.7 2.4 2.3 3.6 3.2 1.9 3.2 3.0 -0.6 4.7 2.6 1.1 0.5 0.8 1.0 1.9 2.2 2.6 2.7 2.1 0.8 2.4 6.9 1.4 Forecasts for 2006-2020 Table 13.5 2.1 3.4 1.8 0.7 1.4 3.2 4.13.3 5.5 0.3 2.5 2.7 1.6 4.7 1.1 0.1 0.9 3.4 0.4 3.5 2.5 2.3 1.8 5.7 1.4 3.4 4.8 3.0 2.5 2.0 159 .0 2.9 1.1 2.1 0.7 0.1 0.2 6.1 2.3 0.7 1.6 1.1 -0.9 1.0 1.9 1.8 1.1 0.
Batista and Zalduendo (2004) show that cross-section regressions can outperform the IMF’s staﬀ projections for emerging markets in terms of both bias and standard errors . . On average.2 percentage points higher than those in Bergheim (2005). For example.2%.160 13 Forecast competitions and 2006-2020 forecasts 13. They ﬁnd ”large discrepancies between forecasts and actual outcomes” and see some gains from combining forecasts. labeled PWC 2006) and from the Economist Intelligence Unit (2006. Table 13. The private sector also appears to use mostly cross-section models and the idea of absolute convergence. Kraay and Monokroussos (1999) employ time series methods and cross-country regressions for 5-year ahead forecasts for a large number of emerging markets. while Wilson and Purushothaman (2003) use a 2% rate. Ianchovichina and Kacker (2005) build on the Kraay/Monokroussos framework to calculate forecasts for the period 2005-14 for 55 developing countries and to derive policy conclusions.5% to 1. Ireland and Thailand. Both use neoclassical models as the theoretical basis and use investment ratios and growth rates of population among the explanatory variables. With the exception of China the forecasts from Ianchovichina and Kacker (2005) are generally lower than those from the other models.but not those for rich economies. but are similar enough for a valid comparison.5 page 161 compares the results for per capita GDP growth from diﬀerent models. the forecasts from this study are 0. The combined panel and trend model from this study is used to rank the countries and the results are compared with those from Bergheim (2005. labeled EIU 2006).5%. The IMF and the World Bank are institutions with a strong interest in long-run forecasts.5% and 4. it uses a large number of explanatory variables and does not outline how it forecasts these into 2020. The forecast horizons are not always identical. forecasts for Indonesia’s per capita growth vary between 1.5 Other long-run forecasting models Systematic long-run forecasts for GDP growth are few and far between. There is a large variance of forecasts. Hawksworth (2006) uses ad-hoc country-speciﬁc rates of absolute convergence ranging from 0. labeled DB 2005). For example. labeled WB 2005) from Hawksworth (2006. However. The Economist Intelligence Unit (2006) uses pooled cross-section estimates for 86 countries for three 10-year periods over 1971 to 2000. The forecasts from the Economist Intelligence Unit (2006) are signiﬁcantly higher than the numbers in the present study for the Asian economies. with the largest diﬀerences in the cases of Australia. from Ianchovichina and Kacker (2005. while they are lower for most rich countries.
9 Data from diﬀerent studies: DB 2005 is Bergheim (2005).0 1.9 Japan 1.6 Italy 2.9 1.5 Singapore 2.1 1.0 Nigeria 1.4 Philippines 2.9 Finland 1.5 Brazil 1.1 1.2 Indonesia 3.8 Israel 1.9 Greece 1.7 2.9 1.2 1.5 2.6 1.5 France 2.0 2.9 3.8 1.2 UK 2.3 3.9 2.9 0.4 2.7 2.1 2.5 1.2 Sweden 1.7 2.1 3.6 South Africa 1.9 Belgium 2.5 1.6 Chile 3.4 2.7 1.7 Spain 4.13.5 Other long-run forecasting models Table 13. Comparisons for annual per capita GDP growth in % This study DB 2005 WB 2005 PWC 2006 2006-20 2006-20 2005-14 2005-50 EIU 2006 2006-20 161 China 5.8 Norway 2.1 2.6 2.2 0.2 2.4 Ireland 4.9 3.2 1.5 4.3 2.1 0.6 0.9 3.8 2.3 1.3 1.4 3.0 USA 2.1 Denmark 2.1 3.8 1.9 1.4 1.5 New Zealand 1.0 1.9 3.9 0.9 Germany 1. EIU 2006 is Economist Intelligence Unit (2006) .8 4.2 Portugal 2.3 1.0 2.4 Thailand 2.8 2.6 3.5 Mexico 2.5.6 0.6 Australia 2.4 1.0 2.4 3.2 1.6 1.4 6.8 4.7 Switzerland 1.8 1.1 1.5 3.8 5.9 Netherlands 2. WB 2005 is Ianchovichina and Kacker (2005). PWC 2006 is Hawksworth (2006).4 3.0 Korea 3.4 Canada 1.0 1.3 1.8 1.8 Taiwan 2.5 3.4 3.1 1.0 2.7 3.4 1.9 0.0 1.2 Turkey 3.4 1.2 4.3 4.7 Malaysia 3.7 1.1 Colombia 2.1 Argentina 1.7 Austria 1.5 1.3 Egypt 1.9 1.6 India 3.5 4.3 1.2 0.
Growth accounting exercises are therefore of little help for understanding the growth process. . There are no scale eﬀects. capital stock) and the growth rate of income.14 Conclusion and outlook This study makes three contributions to the growth literature.it requires changes at home. population size. (3) ”Standard” variables such as the investment ratio or population growth should not be used when explaining income levels. there is no major link between the levels of relevant variables (e. because they vary too little across countries and because they are largely endogenous. (6) Economic growth does not simply spill over from neighboring countries . Second. Some empirical ﬁndings in the literature stem from the inappropriate use of investment data from the Penn World Tables. Some controversial conclusions are either introduced or conﬁrmed in this study: (1) The main drivers of growth are complementarities rather than substitutes. And third. Therefore. it shows that the hypothesis of pair-wise cointegration derived from a number of theoretical models is present in the data for both rich countries and emerging markets. (2) Unlike in many endogenous growth models. Nonstationary panel methods are a better choice. g. human capital and trade openness in the technical progress function. substituting the investment ratio for changes in the capital stock in cross-section estimates is not appropriate. it suggests that the most appropriate theoretical model of the long-run growth process is an augmented Kaldor model that includes physical capital. (5) The still widely used cross-section estimation techniques are not very helpful when modeling a dynamic process such as economic growth. human capital. (4) Investment ratios are not proportional to either the level or the changes in the capital stock. First. it presents a set of forecasts for GDP growth for 40 countries until 2020 using weights on diﬀerent models derived from a forecast competition.
(D) The short-run models used here are rather small and may beneﬁt from the inclusion of other variables (e. (B) Higher quality data or better proxies could be constructed to measure the drivers of GDP growth. (F) The framework and results presented here could be used for the analysis of national economic policies and to derive priorities for policy action. (C) The forecasting exercise used simple rules to derive the paths for the exogenous variables until 2020. but based on a diﬀerent empirical framework. e. More eﬀort could go into improving the quality and reliability of these forecasts. e. i.variables. leaving unexplained why one country’s human capital or openness develops more strongly than that of another country. What might this mean for policy priorities? Also. For example. And trade shares may not be the perfect measure for institutional quality. Despite these contributions much remains to be done in future research. For example. (E) This study has focused largely on the proximate sources of economic growth. These are important questions. However. What matters is the utilization of labor as measured by the hours worked per capita. which countries are currently experiencing GDP above the level indicated by human capital according to the estimated cointegrating relationship? Would that indicate a stronger policy focus on human capital accumulation? Such an analysis would be similar in spirit to Hausmann. one could try to account for research and development in a similar way to Bottazzi and Peri (2007). this would also require forecasts for these . (A) The list of variables could be extended as indicated in chapter 9. where does the estimated common cointegrating relationship not ﬁt the national data. g. Which countries are most eﬃcient at using their physical or human capital i. interest rates). Rodrik and Velasco (2005). measurement error may be large especially in the number of years of education in emerging markets. oil price. have the highest average ratios of output to capital? Why might this be the case? Which countries diﬀer in the time paths of the determinants relative to GDP. but were not the main focus here given the ultimate goal of deriving long-run GDP forecasts. as in the analysis of structural breaks that was employed in Bergheim (2005).usually volatile . .164 14 Conclusion and outlook (7) A population’s age structure does not have a major impact on the growth rate of per capita GDP.
. . . . . . . . . 55 57 58 59 60 61 62 64 64 Average years of education in 21 rich countries . .2 5. . . .6 5. . . . . .1 7. . . . . . . . .2 8. . . . . . . . . . . . 93 Trade openness in 19 emerging markets . . . . . .List of ﬁgures 2. . . . . . . . . . . 38 GDP per capita in 21 rich countries . . . . 109 . . . . . 100 Spatial GDP p. . . . . . . . . . . . . . . .2 3. . . . . . . . .7 5. . . . Predictable diﬀerences in investment shares . . . . . . . .5 5. . . . . . . . . . 78 Average years of education in 19 emerging markets . . . . . . . . . .2 5. . . . . . . . . . . . . . 93 Moran plot for the levels of GDP per capita in 2000-2003 . . . . 99 Moran plot for the growth rates of GDP p. . . . . Capital stock per capita in 19 emerging markets . . . . . . .4 5. . . . . .1 4. . . . . . . . . . . . . . . . . . . . .4 Two economies converging to parallel steady-state paths . . 41 GDP per capita in 19 emerging markets . . 14 Large diﬀerences in distance to frontier . .3 8. . . . . . . . Investment ratios and changes in the capital stock .1 Intercept correction for a simple trend model . .c. . . . .c. . . .c. . . . Capital stock per capita in 21 rich countries . . . . . . . . . . . . . . . . . . No link between investment ratios and capital stocks . . . . . . Nominal investment shares over time . . . . . . . . .8 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 7. .3 4. . . . . . .c. . 90 Trade openness in 21 rich countries . . 48 Share of 30 to 49-year age group . .3 5. . . . . . . . . . . .1 3. . . 78 Small countries with large trade shares . . . . . . . .1 5. . . . . . . . . .9 6. . Investment ratios and the level of GDP p. . . of 21 rich countries . . . . 102 10. . . . . . . . . . . . . . .1 8. . . . . . . . . 41 Annual hours worked per capita in 21 rich countries . .1 6. . . . . . . . .2 7. . . 102 Spatial GDP p. . . . . . . . . . .3 8. . . . . . . . . . No link between GDP per capita and capital-output ratios . . . . . . . . . . . .1 3. . . . . . 49 Nominal and real investment shares in 21 rich countries . . . . . . . . . . . 1994-2003 . . . of 19 emerging markets . . . . . . . . .
.2 12. . . . . . 150 spatial GDP . .1 12. . . 149 years of education 149 trade openness . . . . . . .3 12.4 Errors Errors Errors Errors from from from from cointegration cointegration cointegration cointegration between between between between GDP GDP GDP GDP and and and and capital stock . . .166 List of ﬁgures 11. . . . .1 Diﬀerence between cross-section and panel . . 122 12. 150 . . . . . . .
. . . . . . 1971 to 2003 . . . . . 65 Overview of the empirical literature on human capital . .1 Empirical techniques used in the literature . . . . . . . . . . 79 Overview of the empirical literature on openness . 139 Panel unit root tests . . . . . . . . . . .List of tables 3. . . . .2 13. . . . . .2 5. . . 161 .1 6. . 37 Databases and sample sizes . . . . .) . . . . . .1 7. . .1 12. . . . . . . . 157 Forecasts for annual growth of total GDP 2006-2020 in % . . . 147 Results of the forecast competition 2001-2005 . . . . . . . . . . . . . . . 138 Correlations of changes between 1993 and 2003 . . 156 Forecasts for annual growth of per capita GDP 2006-2020 . . . . . . . . .3 12. . . . . . . . . . . . . . . . . . . . . . . . . . . .4 13. . . . .1 3. . .3 13. . . . 115 12. . . . 153 Forecast competition over 1996 to 2005 . . . . . . . . 40 Overview of the empirical literature on physical capital . . . . . . . . . . . . . . . . . . . . . .2 12. . . . . .4 12. . . . . . . . . . . . . . .1 The weights in world GDP of the 40 countries considered . . . . . .5 Correlations of the levels of key variables (1994-2003 avg. . 143 Models for per capita GDP growth 1971-2000 . . . . . . . . . . . . . . . . . 94 11. . . . . . 159 Comparisons for annual per capita GDP growth in % . . . . . . . . . . .1 13.5 13. . . . . . . . . . . 141 Panel cointegration tests. . . . . .
Aghion. pp.M.): Handbook of Economic Growth. Daron. 67-109. Elsevier. Philippe and Peter Howitt (2005a): Growth with quality-improving innovations: an integrated framework. Robinson (2001): The colonial origins of comparative development: An empirical investigation. pp. In: Philippe Aghion and Steven Durlauf (eds. Elsevier. Econometrica 60(2). Acemoglu. Daron and Joshua Angrist (2000): How large are human capital externalities? Evidence from compulsory schooling laws. .F. Philippe and Peter Howitt (1998): Endogenous growth theory. Maria. Florax (2004): Space and growth: A survey of empirical evidence and methods. Philippe and Peter Howitt (2005b): Appropriate growth policy: a unifying framework. 323-351. The 2005 Joseph Schumpeter Lecture. NBER Macroeconomics annual. 9-59. pp. Aghion.G. Alcal´. pp. Philippe and Steven Durlauf (eds.) (2005): Handbook of Economic Growth. 612-645. Aghion. Simon Johnson and James A. pp. 1369-1401. Philippe and Peter Howitt (1992): A model of growth through creative destruction. Quara terly Journal of Economics 119. Acemoglu. de Groot and Raymond J. Cambridge Massachusetts: The MIT Press. Francisco and Antonio Ciccone (2004): Trade and productivity. Aghion.References Abreu. American Economic Review 91(5). Tinbergen Institute Discussion Paper 129. Volume 1A. Henri L. Aghion.
pp. pp. Manuel and Stephen Bond (1991): Some tests of speciﬁcation for panel data: Monte Carlo evidence and an application to employment equations. Quarterly Journal of Economics. Costas and Allan Drazen (1990): Threshold externalities in economic development. Robert E. Vivek and Athanasios Vamvakidis (2005): How much do trading partners matter for economic growth? IMF Staﬀ Papers 52(1). (2001): Econometric Analysis of Panel Data. 106. Kenneth (1962): The economic implications of learning by doing. Quarterly Journal of Economics 105. Journal of Economic Growth 4(2). 237-277. Review of Economic Studies 58(2). Barro. Baltagi. Swedish Economic Policy Review 6(2). Cambridge University Press. Oxford: Basil Blackwell. pp. New York: Elsevier Science. Esben Sloth (2004): From Schumpeter’s failed econometrics to modern evometric analysis. 277-297. Anselin. cointegration in panels and dynamic panels: A survey. 2nd edition.): A Companion to Theoretical Econometrics. Review of Economic Studies 29. Massimiliano Marcellino and Chiara Osbat (2004): Some cautions on the use of panel methods for integrated series of macroeconomic data. Azariadis. (1997): Human capital and growth in cross-country regressions. pp. In: Badi H. Barro. . Arora. pp 407-443. 155-173. Robert J. and Chiwa Kao (2000): Nonstationary panels. In: Badi Baltagi (ed. Luc (2001): Spatial Econometrics. 310-330. Paper for the Conference of the International Schumpeter Society. Econometrics Journal 7. Anindya. Baltagi (ed. Badi H. pp. Arellano. Barro.): Advances in Econometrics (Vol. Badi H. (1999): Notes on growth accounting. Baltagi. (1991): Economic growth in a cross section of countries. pp. panel cointegration and dynamic panels. pp. (2003): Openness and growth: What’s the empirical relationship? NBER Working Paper 9578.170 References Andersen. Arellano. 15): Nonstationary panels. Arrow. 501-526. Manuel (2003): Panel Data Econometrics. Robert J. 24-40. pp. Robert J. Banerjee. 119-137. Baldwin. pp. 7-51. 322-340.
Becker. Bassanini. Massachusetts: The MIT Press. American Economic Review 86(2). 145-149. Andrea. pp. Becker. Robert J. 541-563. (1964): Human capital: A theoretical and empirical analysis with special reference to education. Oxford Economic Papers 53(3). Beaugrand. Murphy (1999): Population and economic growth. Cambridge. Batista. und Jong-Wha Lee (1993): International comparisons of educational attainment. pp. Robert J. pp. Barro. Edward L. Carnegie-Rochester Conference Series on Public Policy 40. pp. Glaeser and Kevin M. 218223. Geert. und Jong-Wha Lee (2001): International Data on Educational Attainment: Updates and Implications. pp. Philippe (2004): And Schumpeter said ”this is how thou shalt grow”: the further quest for economic growth in poor countries. Harvey R. Gary S. Campbell and Christian Lundblad (2005): Does ﬁnancial liberalization spur growth? Journal of Financial Economics 77. Robert J. OECD Economics Department Working Papers 259. Andrea. 3-56. technology and economic growth: recent evidence from OECD countries. OECD Economics Department Working Papers 283. Bassanini. IMF Working Paper 04/40. und Jong-Wha Lee (1996): International measures of schooling years and schooling quality. 1-46. Bassanini. Chicago: University of Chicago Press. Journal of Monetary Economics 32(3). Barro. Barro. Papers and Proceedings. Gary S. 363-394. American Economic Review 89(2). Stefano Scarpetta and Ignazio Visco (2000): Knowledge. Robert J. Catia and Juan Zalduendo (2004): Can the IMF’s medium-term growth projections be improved? IMF Working Paper 04/203. Stefano Scarpetta and Philip Hemmings (2001): Economic growth: The role of policies and institutions. Robert J. pp.References 171 Barro. Bekaert. and Jong-Wha Lee (1994): Sources of economic growth. . and Xavier Sala-i-Martin (2004): Economic growth (2nd edition). Andrea and Stefano Scarpetta (2001): Does human capital matter for growth in OECD countries? OECD Economics Department Working Papers 282.. Barro.
Collins (2003): The empirics of growth: an update. Journal of Monetary Economics 34(2). .W Temple (2001): GMM estimation of empirical growth models. and Refet S. Bond. Bosworth. Jennifer. Romer and Weil seriously. (2006): European unemployment: The evolution of facts and ideas. World Economic Forum. Blanchard. Brookings Papers on Economic Activity. pp. 1160-1183. G¨rkaynak (2001): Is growth exogenous? Taku ing Mankiw. pp. Benhabib. Blanke. Xavier Sala-iMartin and Augusto Lopez-Carlos (eds. Fiona Paua and Xavier Sala-i-Martin (2003): The Growth Competitiveness Index: Analyzing key underpinnings of sustained economic growth.1. 143-173. Asli Leblebicioglu and Fabio Schiantarelli (2004): Capital accumulation and growth: a new look at the empirical evidence. Economics Papers 2001-W21. 935-966. 653-679. Volume 1A. Chapter 1.): The Global Competitiveness Report 2003-2004. Bergheim. Stephen R. Nuﬃeld College.172 References Ben-David. pp. NBER Macroeconomics Annual 16. Spiegel (2005): Human capital and technology diﬀusion. In: Philippe Aghion and Steven Durlauf (eds. Stefan (2005): Global Growth Centres 2020. Elsevier. Bernanke. Mark und Peter J. pp. Bond. Blomstr¨m. Economic Policy.. Barry and Susan M. pp. 269-276. Ben S. 5-59. Lipsey and Mario Zejan (1996): Is ﬁxed ino vestment the key to economic growth? Quarterly Journal of Economics 111(1). Quarterly Journal of Economics 1008(3). in: Michael E. pp. Robert E. Benhabib. Jess and Mark M. Magnus. Dan (1993): Equalizing exchange: Trade liberalization and income convergence. Klenow (2000): Does schooling cause growth? American Economic Review 90(5).). pp. Anke Hoeﬄer and Jonathan R. Stephen R. Boston College. Olivier J. Deutsche Bank Research Current Issues. Klaus Schwab. Jess and Mark M. Bils. Handbook of Economic Growth. 11-57. Spiegel (1994): The role of human capital in economic development: evidence from aggregate cross-country data. Porter. Working Papers in Economics 591..
J¨rg and M. and Dynamic Panels. pp. pp. 486511. 03/37. Dahlman (2004): Knowledge and development. Chen.References 173 Bottazzi.C. Francesco. 151-173. University of Bristol Discussion Paper 00/509. pp.): Nonstationary Panels. 161-178. Laura and Giovanni Peri (2007): The dynamics of R&D and innovation in the short run and in the long run. John Y. Matyas and P. Durlauf (2001): Growth economics and reality. o In: Badi Baltagi (ed. 1-20. J¨rg and Samarjit Das (2005): Panel unit root tests under cross o sectional dependence. Panel Cointegration. Derek H. Roberto. William A. and Carl J. World Bank Policy Research Working Paper 3763. and Pierre Perron (1991): Pitfalls and opportunities: what macroeconomists should know about unit roots. Breitung. Breitung. Hashem Pesaran (2006): Unit roots and cointegration o in panels. 15. and Hiau Looi Kee (2005): A model on knowledge and endogenous growth. Cannon. Brunner. J¨rg (2000): The local power of some unit root tests for panel data. Allan D. David Demery and Nigel W. Advances in Econometrics. NBER Technical Working Paper 100. World Bank Economic Review 15 (2). Journal of Economic Growth 1(3). Campbell. World Bank Policy Research Working Paper 3539.. Derek H. 229-272.C. Breitung. 363-390. Edmund S. pp. Linda Kaltani and Norman Loayza (2005): Openness can be good for growth: The role of policy complementarities. In : L. (2003): The long-run eﬀects of trade on income and income growth. J¨rg (2005): A parametric approach to the estimation of cointegrao tion vectors in panel data. Vol. Statistica Neerlandica 59(4). Gerardo Esquivel and Fernando Lefort (1996): Reopening the convergence debate: a new look at cross-country growth empirics. pp. pp. Chen. Breitung. IMF Working Paper No. and Steven N. Kluwer Academic Publishers. Sevestre (eds): The Econometrics of Panel Data (Third Edition). Duck (2000): Does distance matter for economic performance? Evidence from European regions. Caselli. Econometric Reviews 24. Brock. Economic Journal 117. World Bank Policy Research Working Paper 3366. Chang. .
Journal of Economic Methodology 12(3). Cohen. Cambridge. . and C. pp. and Hendry. David F. Chick. Victoria and Sheila Christine Dow (2005): The meaning of open systems.174 References Chiang. 363-381. Daniel and Marcelo Soto (2001): Growth and human capital: good data. Cohen. Michael P. Avi J.maxwell. Angel and Rafael Dom´nech (2002): Human capital in growth e regressions: how much diﬀerence does data quality make? An update and further results. Colander.edu/maxpages/faculty/cdkao/working/npt.A User Guide. Victoria (2003): On open systems.syr. Jean-Francois Tremblay and Sylvie Marchand (2004): International Adult Literacy Survey: Literacy scores. (2003): Economic forecasting: some lessons from recent research. (1999): Forecasting non-stationary economic time series. Mimeo.C. Harcourt (2003): Whatever happened to the Cambridge Capital Theory Controversies? Journal of Economic Perspectives 17(1). Paper presented at the Conference of the International Network for Economic Method. pp. 301-329. M-H. and G. Coulombe. OECD Development Centre Technical Paper 179. good results. Statistics Canada. Journal of the History of Economic Thought 22(2).3 . David F. Angel and Rafael Dom´nech (2000): Human capital in growth e regressions: how much diﬀerence does data quality make? OECD Economics Department Working Papers 262. Daniel and Marcelo Soto (2002): Why are some countries so poor? Another look at the evidence and a message of hope. Serge. Clements. Michael P. pp.html Chick. April 2001. September 2003. and Hendry. De la Fuente. Kao (2002): Nonstationary Panel Time Series Using NPT 1. Economic Modelling 20(2). 199-214. http://www. Clements. Massachusetts: The MIT Press. Cohen. OECD Development Centre Technical Paper 197. Leeds. David (2000): The death of neoclassical economics. Syracuse University. pp. human capital and growth across fourteen OECD countries. 127-143. Center for Policy Research. Consensus Economics (2001): Consensus Forecasts: A digest of international economic forecasts. De la Fuente.
Steve (2004): Ideas and education: level or growth eﬀects and their implications for Australia. Hans-Eggert Reimers and Barbara Roﬃa (2006): Long-run money demand in the new EU Member States with exchange rate eﬀects. Journal of Economic Perspectives 12(2). pp. Dollar. Hans-Eggert Reimers (2005): Health care expenditures in OECD countries: A panel unit root and cointegration analysis. Dufrenot. Dreger. Dowrick. pp. T. Chicago: Chicago University Press. Dowrick. H. Steve and Mark Rogers (2002): Classical and technological convergence: beyond the Solow-Swan growth model. Steve (2002): G-20 comparisons of incomes and prices: What can we learn from the International Comparison Program? RBA Annual Conference Volume. pp. Diebold. Gernot. and growth. present. (1998): The past. and Rose.References 175 De la Fuente. Christian. pp. Bradford and Lawrence H. 9-37. Christian. Dreger. IMF Working Papers 05/164. Francis X. Oxford Economic Papers 54 (3). pp. 155-175. Journal of Monetary Economics 50(1). J. 3rd edition. Ronald I. In: Ito. 369-385. 813-835. Dowrick. (2001): Forecasting in macroeconomics: A practitioner’s view. 1-36. Francis X. Diebold. (2004): Elements of Forecasting. . De Economist 149. 445-502. DeLong. Quarterly Journal of Economics 106. David and Aart Kraay (2003): Institutions. 175-192. trade. Doppelhofer. Reserve Bank of Australia. American Economic Review 94(4). Don. pp. AK. Miller and Xavier Sala-i-Martin (2004): Determinants of long-term growth: A Bayesian Averaging of Classical Estimates (BACE) approach. Thomson South-Western. (eds): Growth and Productivity in East Asia. pp. ECB Working Paper 628. F. and Etienne B. pp. International Journal of Applied Econometrics and Quantitative Studies 2(2). and future of macroeconomic forecasting. Gilles J. Yehoue (2005): Real exchange rate misalignment: A panel co-integration and common factor analysis. Summers (1991): Equipment investment and economic growth. Angel and Rafael Dom´nech (2006): Human capital in growth e regressions: how much diﬀerence does data quality make? Journal of the European Economic Association 4(1). 133-162.
Robert F.176 References Durlauf. pp. International Economic Review 39(2). Elsevier. Fisher (2003): Aggregation in production functions: What applied economists should know. Easterly. 143-159.): Economic growth and the structure of long-term development. Johnson and Jonathan Temple (2005): Growth econometrics. Felipe. Mimeo. Edwards. . Ross Levine and Sergio Rebelo (1994): Policy. Steven N. Felipe. Wiley. William and Ross Levine (2001): It’s not factor accumulation: stylized facts and growth models.): Handbook of Economic Growth. and Byung Sam Yoo (1987): Forecasting and testing in cointegrated systems. Engle. Volume 1A. Steven N. Metroeconomica 54(2-3). 235-308. Evans. Easterly. Sebastian (1998): Openness. Enders. pp. Econometrica 55. pp. J. pp 295-306. World Bank Economic Review 15(2). pp. technology adoption and growth. industry and corporate trends. productivity and growth: what do we really know? Economic Journal 108. Economist Intelligence Unit (2006): Foresight 2020 . Engle. 208-262. Jesus and John McCombie (2005): Why are some countries richer than others? A skeptical view of Mankiw-Romer-Weil’s test of the neoclassical growth model. Durlauf. Elsevier. Quah (1999): The new empirics of economic growth. William. pp. Volume 1A. William (2001): The elusive quest for growth: economists’ adventures and misadventures in the tropics. Robert F. Cambridge. Jesus and Franklin M. 177219. In: Philippe Aghion and Steven Durlauf (eds. pp. pp. Metroeconomica 56(3). In: John B. Granger (1987): Cointegration and errorcorrection: Representation. estimation. 383-398. In: Robert Solow and Luigi Pasinetti (eds. 555-677. Fagerberg.Economic. Macmillan. and testing. Paul (1998): Using panel data to evaluate growth theories.. pp. 2nd edition. Walter (2004): Applied econometric time series. Taylor and Michael Woodford (eds): Handbook of macroeconomics. Massachusetts: The MIT Press. 360-392. and Clive W. Journal of Econometrics 35. 251276. Jan (2002): A layman’s guide to evolutionary economics. Paul A. and Danny T. Robert King. Easterly.
Tinbergen Institute Discussion Papers 02-040/3.References 177 Felipe. Oded and David N. and David Romer (1999): Does trade cause growth? American Economic Review 89(3). 189-199. In: Dieter Bender and Hans-Rimbert Hemmer (eds): Entwicklungsl¨nder im Zeitalter der Globa alisierung.G. o Fernandez. M¨nchen: Vahlen. Weil (2000): Population. (2004): Two centuries of economic growth: Europe chasing the American frontier. Henri L. Fisher. NBER Working Paper 10662. Anthem Press. u Fullbrook. de Groot and Reinout Heijungs (2002): The empirical growth literature: robustness. 283-299. Ralf and Michael Frenkel (1999): Convergence. Clive and Paul Newbold (1974): Spurious regressions in econometrics. pp. pp. (1969): The existence of aggregate production functions. Journal of Applied Econometrics 16. pp. 563-576. Frenkel. American Economic Review 90(4). Jesus and John McCombie (2006): The tyranny of the identity: growth accounting revisited. 553-577. Felipe. pp. Journal of Econometrics 2. Limburg-Seminar ”Wissenschaft und Praxis der Entwicklungs¨konomik”.F. 195-229. 2. Jeﬀrey (2003): Comments on Bosworth and Collins.. International Review of Applied Economics 20(3). pp. Florax. and the role of openness in the process of economic development. 111-120. Gordon. Jeﬀrey A. Robert J. Edward (2004): A guide to what’s wrong with economics. pp. 379-399. The Empirics of Growth: An Update Brookings Papers on Economic Activity Vol. Galor. Frankel. Econometrica 37(4). Carmen. Franklin M. technology and growth: From Malthusian stagnation to the demographic transition and beyond.M. Eduardo Ley and Mark F. divergence. Tagungsband zum 2. Fendel. Steel (2001): Model uncertainty in cross-country growth regressions. Granger. Michael and Hans-Rimbert Hemmer (1999): Grundlagen der Wachstumstheorie. Frankel. Jesus and John McCombie (2007): Is a theory of total factor productivity really needed? Metroeconomica 58. pp. 806-828. pp. signiﬁcance and size. Raymond J. .
James D. Gwartney. Dani Rodrik and Andr´s Velasco (2005): Growth diage nostics. Harvard University. Jan. Jones (1999): Why do some countries produce so much more output per worker than others? Quarterly Journal of Economics 114. David F. Lant Pritchett and Dani Rodrik (2005): Growth accelerations. (1994): Time Series Analysis. Kimko (2000): Schooling. (1993): Econometric Analysis. Charles Skipton and Robert Lawson (2001): Trade openness. Grossman. Chapter 3 in Economic Freedom of the World: 2001 Annual Report. Princeton: Princeton University Press. Jan J. Oxford Bulletin of Economics and Statistics 66(5). u Hendry. Groen. labor-force quality. PricewaterhouseCoopers. 2nd edition. Hamilton. and Elhanan Helpman (1990): Comparative advantage and long-run growth. American Economic Review 90(5). Journal of Economic Growth 10(4). New York: Macmillan. Dana Hajkova and Randall K. Hemmer. 1980-1998. Hall. Hausmann. and Frank Kleibergen (2003): Likelihood-based cointegration analysis in panels of vector error-correction models. 303-329. Hausmann. Hauk. and the growth of nations. 796-815. John F. and Romain Wacziarg (2004): A Monte Carlo study of growth regressions. Robert E. M¨nchen: Vahlen. Hanousek. Kennedy School of Government. Journal of Business and Economic Statistics 21(2). William R. and Hans-Martin Krolzig (2004): We ran one regression. William H. Hans-Rimbert and Andreas Lorenz (2004): Grundlagen der Wachstumsempirie. Ricardo. pp. NBER Technical Working Paper 296. income levels and economic growth. pp. James. 799-810. American Economic Review 80(4). and Dennis D. Filer (2004): The other side of the moon: The data problem in analyzing growth determinants.J. 1184-1208. Eric A. Ricardo. pp. Gene M. John (2006): The world in 2050. pp. and Chad I. pp. . 295-318. pp. 83-116. Hanushek. William Davidson Institute Working Paper 682. Hawksworth.178 References Green.
References 179 Heston. . Papers and Proceedings 86(2). 1127-1170. pp. Douglas and Marko Tervi¨ (2002): Does trade raise income: Evidence o from the twentieth century. Irwin. Islam. 2nd edition. Jaroslava and Martin Wagner (2006): The performance of panel unit root and stationary tests: Results from a large scale simulation exercise.A reply. Kevin D. Oxford Bulletin of Economics and Statistics 66(5). Baltagi (ed. pp. pp. pp. Journal of Econometrics 115. In: Badi H. New York: Elsevier Science. 765-798.1. Hlouskova. Quarterly Journal of Economics 113. Peter (2000): Endogenous growth and cross-country income diﬀerences. 1-18. Cambridge University Press. Im. Nazrul (2000): Small sample performance of dynamic panel data estimators: a Monte Carlo study on the basis of growth data. World Bank Policy Research Working Paper No. pp. 53-74. 3775. Kyung So. Robert Summers and Bettina Aten (2002): Penn World Table Version 6. Economics Bulletin 10(3). Alan. panel cointegration and dynamic panels. Econometric Reviews. Nazrul (1998): Growth empirics: A panel data approach . Islam. Hondroyiannis. 1-8. Hsiao. and Stephen J. Howitt. pp. 317-339. American Economic Review 90(4). M. 85-116. pp. George and Evangelia Papapetrou (2002): Demographic transition in Europe. Center for International Comparisons at the University of Pennsylvania (CICUP). Ianchovichina. Quarterly Journal of Economics 110(4). pp. Journal of International Economics 58. Hashem Pesaran and Yongcheol Shin (2003): Testing for unit roots in heterogeneous panels. pp. 20-24. Alan and Robert Summers (1996): International price and quantity comparisons: Potentials and pitfalls. Hoover. 15): Nonstationary panels. Perez (2004): Truth and robustness in crosscountry growth regressions.): Advances in Econometrics (Vol. Heston. Cheng (2003): Analysis of Panel Data. pp. Elena and Pooja Kacker (2005): Growth trends in the developing world: Country forecasts and determinants. 829-846. Nazrul (1995): Growth empirics: A panel data approach. American Economic Review. Islam. 325-329.
W. Charles I. 220-239. Quarterly Journal of Economics 110. pp. 1063-1111. 139-144. . New York: Elsevier Science. pp. Wolfgang (2002): Geographic localization of international technology diﬀusion. (2005): Growth and ideas. Soren (1995): Likelihood-based interference in cointegrated vector autoregressive models. (1999): Growth: with or without scale eﬀects? American Economic Review 89(2). (2002b): Sources of U. (1997): Comment on Klenow-Rodriguez. pp. 759-784. Charles I. 15): Nonstationary panels. (2002a): Introduction to economic growth. 495-526. New York: W.S. (1995a): Time series tests of endogenous growth models. Jones. Ruth A. Jones. ”The Neoclassical Growth Revival: Has it Gone too Far?”. 2nd edition. The Economic Journal. Keller. Volume 1B. Kaldor. In Badi H. Judson. pp. Owen (1999): Estimating dynamic panel data models: a practical guide for macroeconomists. Elsevier. The MIT Press. 591-624. pp. Charles I. Jones. Jones. Oxford University Press. (1995b): R&D-based models of economic growth. Nicholas (1957): A model of economic growth. panel cointegration and dynamic panels.). pp. Chiwa and Min-Hsien Chiang (2000): On the estimation and inference of a cointegrated regression in panel data.): Advances in Econometrics (Vol. NBER Macroeconomics Annual. and Ann L. 915. 179-222. pp. Charles I. Journal of Political Economy 103(4). American Economic Review 92(1). 347-357. Charles I. Jones. American Economic Review 92. Nicholas (1975): What is wrong with economic theory. pp. Norton.180 References Johansen. Jones. Economics Letters 65. Charles I. Quarterly Journal of Economics 89. economic growth in a world of ideas. 120-142. Jones. Kao. Charles I. Baltagi (ed. pp. In: Philippe Aghion and Steven Durlauf (eds. Kaldor. pp. Handbook of Economic Growth.
Le. 109-142. Landes. 817-863. 53-78. 2224. . Johan Gibson and Les Oxley (2003): Cost. (1999): The wealth and poverty of nations. Norton & Co. Klenow. and eﬃciency of various estimators in dynamic panel data models. 59-70. World Bank Economic Review 15(2). Alan B. pp. pp. pp. Elsevier. pp. Lee. Aykut and Selahattin Dibooglu (2001): Long-run economic growth: an interdisciplinary approach. Peter J. Paul (1996): What economists can learn from evolutionary theorists. pp. Econometrics Journal 4(1). Johan Lyhagen and Mickael L¨thgren (2001): Likelihood based o cointegration tests in heterogeneous panels. Peter J. and Andr´s Rodriguez-Clare (2005): Externalities and e growth. Stephen (2001): Are the Penn World Tables data on government and investment being misused? Economics Letters 71. Jan F. 271ﬀ. Technology & Policy 13(4). (1990): Why are we so rich and they so poor? American Economic Review 80. and Andr´s Rodriguez-Clare (1997): The neoclassical revival e in growth economics: has it gone too far? NBER Macroeconomics Annual. (1995): On bias. World Bank Policy Research Working Paper No. and Mikael Lindahl (2001): Education for growth: why and for whom? Journal of Economic Literature 39. inconsistency.References 181 Kibritcioglu. New York: W.and income-based measures of human capital.): Handbook of Economic Growth. David S. p. Krueger. Knowledge. Aart and George Monokroussos (1999): Growth forecasts using time series and growth models. David S. 293-298.W. Klenow. In: Philippe Aghion and Steven Durlauf (eds. Trinh. pp. Ha Yan. Landes. Journal of Economic Surveys 17. 1101-1136. pp. is openness good for growth? Journal of Development Economics 75(2). 221-224. Why some are so rich and some so poor. pp. Knowles. Klenow. Rolf. Kraay. Krugman. Kiviet. Journal of Econometrics 68. 451ﬀ. Larsson. 1-13. (2001): Comment on ”It’s not factor accumulation: stylized facts and growth models”. pp. Volume 1A. Peter J. Luca Antonio Ricci and Roberto Rigobon (2004): Once again. Talk given to the European Association for Evolutionary Political Economy.
Ross (2005): Finance and growth: Theory and evidence. Thomas (2004): Medium-term forecasts of potential GDP and inﬂation using age structure information. 1850-1990. Andrew. Volume 1A. GS and In-Moo Kim (1999): Unit roots. 942-963. Planning Forum 9. Malmberg.182 References Lee. Elsevier. Loayza. Up (2003): A spatial analysis of regional income convergence. pp. pp. Bo and Thomas Lindh (2002): Population change and economic growth in the western world. Maddala. 19-49. Journal of Monetary Economics 22. 357-392. 319-323. cointegration and structural change. Robert (1988): On the mechanics of economic development. pp. In: Philippe Aghion and Steven Durlauf (eds. Levine. pp. Lim. M. 1-24. Lee. Kevin. Mortensen (2005): An empirical model of growth through product innovation. Bo and Thomas Lindh (2004a): Forecasting global growth by age structure projections. Levin. Lentz. Norman V. World Bank Policy Research Working Paper 1333. 865-933. Paper presented at workshop in Rostock.42. NBER Working Paper 11546. Hashem Pesaran and Ron Smith (1997): Growth and convergence in a multi-country empirical stochastic Solow model. Journal of Econometrics 108(1). Levine. Chien-Fu Lin and Chia-Shang Chu (2002): Unit root tests in panel data: asymptotic and ﬁnite-sample properties. Journal of Forecasting 23(1). Cambridge University Press. Lindh. Malmberg. Kevin.A comment. Arbetsrapport 2004:5. . Quarterly Journal of Economics 113. Ross and David Renelt (1992): A sensitivity analysis of cross-country growth regressions. pp. pp. American Economic Review.): Handbook of Economic Growth. (1994): A test of the international convergence hypothesis using panel data. Lucas. Journal of Applied Econometrics 12. pp. M. 66-80. Institute for Futures Studies. Rasmus and Dale T. 3 . pp. Hashem Pesaran and Ron Smith (1998): Growth empirics: A panel data approach .
Mimeo. David Romer and David N. pp. pp. Moran. Moreno. 3. pp. Mimeo. and Donggyu Sul (2003): Cointegration vector estimation by panel DOLS and long-run money demand. Benoit Perron and Peter C. and Ashok Dhareshwar (1993): A new database on physical capital stock: Sources. pp. Hyungsik Roger. Richard R. Mayer. . Nelson C. pp. pp. Ramon and Bharat Trehan (1997): Location and the growth of nations. 17-23.. Harvard University Press.B. Thierry and Soledad Zignago (2006): Notes on CEPII’s distances measures. Maeso-Fernandez. 407437. Gregory N. Journal of Economic Growth 2(4). Arbetsrapport 2004:7. Bocconi University. Moon. Quarterly Journal of Economics. Biometrica 37. paperback edition. pp. (1981): Research on productivity growth and productivity diﬀerences: Dead ends and new departures.References 183 Malmberg. P. Marcellino. Chiara Osbat and Bernd Schnatz (2006): Towards the estimation of equilibrium exchange rates for CEE acceding countries: methodological issues and a panel cointegration perspective. Rivista de Analisis Economico 8 (1). Weil (1992): A contribution to the empirics of economic growth. methodology and results. and earnings. Massimiliano (2006): A benchmark for models of growth and inﬂation. Francisco. Mark. Richard R. Journal of Comparative Economics 34(3). 655-680. 1029-64. Nehru. Journal of Economic Literature 19(3). (2005): Where are we now on an evolutionary theory of economic growth and where should we be going? CCS Working Paper No. 37-59 Nelson. Richard R. pp. experience. Mimeo.(1950): Notes on continuous stochastic phenomena. Institute for Futures Studies. 399-418. Phillips (2005): Incidental trends and the power of panel unit root tests. Mankiw. Jacob (1974): Schooling.A. 499-517. Oxford Bulletin of Economics and Statistics 65(5). 9-51. Mincer. Columbia University Press. Bo and Thomas Lindh (2004b): Demographically based global income forecasts up to the year 2050.P. Reprinted in: Nelson. Vikram. Nelson. (2005): The sources of economic growth.
Philippe and Steven N. Stephen L. pp. Princeton University Press. OECD (2004. (2001): The failure of endogenous growth. Prescott (2005): A uniﬁed theory of the evolution of international income levels. and Sidney G. Parente. 49-58. Nelson. 93-130. 69-75. Journal of Economic Perspectives 16(2). Parente. Massachusetts: The MIT Press. 2005): Education at a Glance. New York: Elsevier Science. pp. Nelson. Cambridge. Massachusetts: The MIT Press. ¨ Osterholm. Obstfeld. Stephen L. Daniel A. Mimeo. Baltagi (ed.184 References Nelson. pp. and Edward C. In: Badi H. Nuxoll. and Edmund S. American Economic Review 56(2). P¨r (2004): Estimating the relationship between age structure e and GDP in the OECD using panel cointegration methods. Technology. panel cointegration and dynamic panels. Marco (2006): Why the openness-growth linkage is so ambiguous: answers based on Ricardo and Laspeyres. Stephen L. (2005): Understanding the process of economic change. Richard R. Knowledge. and Sidney G.): Advances in Econometrics (Vol. 1371-1417. Phelps (1966): Investment in humans. Elsevier. Neuhaus. In: Aghion. Winter (1982): An evolutionary theory of economic change. pp. 298-321. Winter (2002): Evolutionary theorizing in economics. and Policy 13(4). Durlauf: The Handbook of Economic Growth. 15): Nonstationary panels. Prescott (2000): Barriers to riches. North. Parente. Volume 1B. pp. Stephen L. 1423-36. Cambridge. Uppsala University. and Edward C. Peter (2000): Fully modiﬁed OLS for heterogeneous cointegrated panels. . Maurice and Kenneth Rogoﬀ (1996): Foundations of international macroeconomics. Paris. Richard R. Department of Economics. Douglas C. Harvard University Press. pp. Journal of Political Economy 102(2). Richard R. Pedroni. Prescott (1994): Barriers to technology adoption and development. American Economic Review 84(5). 23-46. and Edward C. Parente. (1994): Diﬀerences in relative prices and international differences in growth rates. Working Paper 2004:13. pp. technological diﬀusion and economic growth.
The Review of Economics and Statistics 83(4). pp. Vol. Hashem (2007): A simple panel unit root test in the presence of cross section dependence. 367-391. Miguel. (2004): Why do Americans work so much more than Europeans? Federal Reserve Bank of Minneapolis Quarterly Review 28. 2-13. Edward C. Pedroni. 111-134. George and Harry Anthony Patrinos (2004): Returns to investment in education: a further update. Education Economics 12(2). p. Pritchett. 621-634. Mimeo. Rob Alessie and Coen Teulings (2005): Measurement error in education and growth regressions. Peter (2001): Purchasing power parity tests in cointegrated panels. 11(3). Williams College. Lant (2001): Where has all the education gone? The World Bank Economic Review 15(3). Peter and Tim Vogelsang (2005): Robust unit root and cointegration rank tests for panels and large systems. 426-434. Journal of Economic Perspectives. (1998): Needed: A theory of total factor productivity.References 185 Pedroni. M. European Economic Review 37. Lant (1997): Divergence. M. Journal of the American Statistical Association 94(446). pp. Peter (2004): Social capital. pp. pp. Yongcheol Shin and Ron P. Mimeo. 3-17. barriers to production and capital shares. Pritchett. big time. 525-551. Quah. Psacharopoulos. pp. pp. Smith (1999): Pooled mean group estimation of dynamic heterogeneous panels. International Economic Review 39(3). and Neville Francis (2006): A century of work and leisure. 265-312. Pedroni. pp. Valerie A. Edward C. Pesaran. Prescott. Pesaran. Prescott. Danny (1993): Empirical cross-section dynamics in economic growth. University of San Diego. 727-731. Tinbergen Institute Discussion Paper 040/03. . Ramey. Version dated November 2005. Portela. pp. William University Economics Department Working Paper. Journal of Applied Econometrics 22. Hashem.
old theory and the welfare costs of trade restrictions. Journal of Political Economy 98(5). Finance and Economics Discussion Series 2000-46. 1002-1037. American Economic Review 87(2). pp. Arvind Subramanian. Brookings Papers on Economic Activity 1995:1. Quarterly Journal of Economics 106(2). pp. Maria Teresa and Ana Maria Loboguerrero (2005): Spatial dependence and economic growth: evidence from a panel of countries. pp. Robinson. Sala-i-Martin. Finley (ed. American Economic Review 51. Policy Options July-August 1994. Rudd. pp. Paul (2001): Comment on ”It’s not factor accumulation: stylized facts and growth models. Rodrik. Jeﬀrey and Andrew Warner (1995): Economic reform and the process of global integration.A skeptic’s guide to the cross-national evidence. A. Paul (1994b): Beyond classical and Keynesian macroeconomic policy. Journal of Political Economy 94(5). 5-38. pp. 131-165. S71-S102. 531-555.” World Bank Economic Review 15(2). New York: Nova Science Publisher. 360-369. Rodriguez. Paul (1994c): The origins of endogenous growth. . 178-183. pp. and Paul M. Rivera-Batiz. Journal of Economic Perspectives 8(1).186 References Ramirez. Board of Governors of the Federal Reserve System. Francisco and Dani Rodrik (2001): Trade policy and economic growth . Romer. Dani. pp.a review article. 225-227. Sachs. Paul (1994a): New goods. Romer. Jeremy (2000): Empirical evidence on human capital spillovers. 3-22. NBER Macroeconomic Annual 2000. Romer. Romer (1991): Economic integration and endogenous growth. pp. Paul (1990): Endogenous technological change. Romer. Romer. pp.): Economic Growth Issues. Paul (1986): Increasing returns and long-run growth. Journal of Economic Growth 9(2). Xavier (1997b): I just ran two million regressions. and Francesco Trebbi (2004): Institutions rule: the primacy of institutions over geography and integration in economic development. Joan (1961): Equilibrium growth models . In: L. Journal of Development Economics 43. Romer. Luis A.
Solow. spillovers and growth in the G7 economies: An empirical investigation. An update on post-Keynesian growth theories. pp. Singer. pp. Reprint of the 4th edition of 1934. 157-200. Robert M. 45-54. Solow. Edward Elgar. Philip and Martin Weale (2004): Education and economic growth.References 187 Sarno. Bulletin of the American Academy of Arts and Sciences 51(2). Robert M. Journal of Economic Surveys 17(2). Maxine (1997): Thoughts of a Nonmillenarian. pp. Robert M. Lucio (2001): Nonlinear dynamics. American Economic Review 88(5). Journal of Economic Perspectives 8(1). Engelbert (1999): Robinsonian and Kaleckian growth. Department of a Economics Working Paper wuwp067. 1994 Modern Library Edition. Paul A. Schumpeter. (1994): Perspectives on growth theory. pp. Adam (1776): An inquiry into the nature and causes of the wealth of nations. (1987): Nobel Prize Lecture.): The International Handbook on the Economics of Education. Solow. Quarterly Journal of Economics 70. Sianesi. Solow. (1956): A contribution to the theory of economic growth. 401-426. 1290-1310. (1957): Technical change and the aggregate production function. Robert M. Smith. (2001): What have we learned from a decade of empirical research on growth? Applying growth theory across countries. . Marcelo (2002): Rediscovering education in growth regressions. 65-94. Stockhammer. 283-288. Stevens. Segerstrom. Robert M. World Bank Economic Review 15. 36-51. The Review of Economics and Statistics 39. OECD Development Centre Technical Papers 202. 312-320. Soto. pp. Universit¨t Wien. pp. Economica 68. Solow. Joseph (1912): Theorie der wirtschaftlichen Entwicklung. (Includes an add-on from 2001). Barbara and John van Reenen (2003): The returns to education: Macroeconomics. (1998): Endogenous growth without scale eﬀects. pp. In: Geraint Johnes and Jill Johnes (eds.
United Nations (2005): World population prospects: The 2004 revision. 433-455. pp. p. pp. Waldo R. 112-156. Tools and Applications. Bulletin of Economic Research 52(3). Eindhoven Centre for Innovation Studies. Temple. 1950-1988. distance to frontier and composition of human capital. IMF Working Paper 06/59. 181-205. Verspagen. pp. Moreno and J. Thirlwall. 97-127. Jonathan (2000): Growth regressions and what the textbooks don’t tell you. L´pez-Bazo. An Empirical Analysis using Spatial Econometrics. Robert and Alan Heston (1988): A new set of international comparisons of real product and price level estimates for 130 countries.327-368. pp. Bart (2000): Economic growth and technological change . 1950-1985. Journal of Economic Growth 7. e European Economic Review 45. Jonathan (1999): The new growth evidence. 57-80. Philippe Aghion and Costas Meghir (2006): Growth. Anselin. Surinach (2004): Growth and a o Externalities Across Economies. Journal of Economic Growth 11. Edward Elgar. Robert and Alan Heston (1991): The Penn World Table (Mark 5): An expanded set of international comparisons. Temple. New York.J. Berlin: Springer. Vay´. Review of Income and Wealth 34. pp. Florax and S. Tony (2002): The nature of economic growth: an alternative framework for understanding the performance of nations. Rey (eds. E. 1341-1378. Economic Geography 46. 1-25.. Tobler.): Advances in Spatial Econometrics: Methodology. Summers. Journal of Economic Literature 37(1). R. . In: L. E. R. Vamvakidis.an evolutionary interpretation. Collins.M. 234-240. pp. Lester (2003): Fortune favors the bold.G. Athanasios (2002): How robust is the growth-openness connection? Historical evidence. Quarterly Journal of Economics. pp. Thurow. Jerome.188 References Summers. (1970): A computer movie simulating urban growth in the Detroit region. pp. Vandenbussche. Jos´ and Romain Wacziarg (2001): How democracy aﬀects growth. Timmerman. Allan (2006): An evaluation of the World Economic Outlook forecasts. Tavares.
pp. Wacziarg. NBER Working Papers 10152. South-Western. World Bank Economic Review 15(3). (2003): Introductory Econometrics: A Modern Approach. Romain and Karen Horn Welch (2003): Trade liberalization and growth: New Evidence. pp. 907-918. Working Paper Series in Economics and Finance 435. Goldman Sachs Global Economics Paper No. Wacziarg. Wacziarg. Wooldridge. Wilson. 99.References 189 Verspagen. Stockholm School of Economics. Jeﬀrey M. Quarterly Journal of Economics 106(2). Romain (2002): Review of Easterly’s ”The elusive quest for growth”. Bart (2002): Evolutionary Macroeconomics: a synthesis between neo-Schumpeterian and post-Keynesian lines of thought. 369-405. Young. The Electronic Journal of Evolutionary Modeling and Economic Dynamics No 1007. Romain (2001): Measuring the dynamic gains from trade. Yudong and Johan Lyhagen (2001): Using a trade-induced catch-up model to explain China’s provincial economic growth 1978-97. Yao. Alwyn (1991): Learning by doing and the dynamic eﬀects of international trade. pp. 393-429. Journal of Economic Literature XL. Dominic and Roopa Purushothaman (2003): Dreaming with the BRICs: The path to 2050. .