THE WORLD BANK ECONOMIC REVIEW Volume 12 September 1998 Number 3 187 34 Credibility of Rules and Economic Growth: Evidence from a Worldwide Survey of the Private Sector Aymo Brunetti, Gregory Kisunko, and Beatrice Weder Does Economic Analysis Improve the Quality of Foreign Assistance? Klaus Deininger, Lyn Squire, and Swati Basu Demographic Transitions and Economic Miracles in Emerging Asia David E. Bloom and Jeffrey G. Williamson The Bolivian Social Investment Fund: An Analysis of Baseline Data for Impact Evaluation Menno Pradhan, Laura Rawlings, and Geert Ridder International Evidence on the Determinants of Private Saving Paul R. Masson, Tamim Bayoumi, and Hossein Samiei Unfair Trade? The Increasing Gap between World and Domestic Prices in Commodity Markets during the Past 25 Years Jacques Morisset A NEW DEVELOPMENT DATABASE A Database of World Stocks of Infrastructure, 1950-95 David Canning ISSN 0258-6770 THE WORLD BANK ECONOMIC REVIEW EDITOR Moshe Syrquin CONSULTING EDITOR Sandra Gain EDITORIAL BOARD Kaushik Basu, Cornell University and University of Delhi Stijn Claessens Francois Bourguignon, DELTA, Paris David Dollar Gerald K. Helleiner, University of Toronto Gregory K. Ingram Mark R. Rosenzweig, University of Pennsylvania Martin Ravallion Jacques van der Gaag John Williamson L. Alan Winters The World Bank Economic Review is a professional journal for the dissemination of World Bank- sponsored research that informs policy analyses and choices. It is directed to an international readership among economists and social scientists in government, business, and international agencies, as well as in universities and development research institutions. The Review emphasizes policy relevance and opera- tional aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but not necessarily proficient in advanced mathematical or econometric techniques. Articles will illustrate how professional research can shed light on policy choices. Inconsistency with Bank policy will not be grounds for rejection of an article. Articles will be drawn primarily from work conducted by World Bank staff and consultants. Before being accepted for publication by the Editorial Board, all articles are reviewed by two referees who are not mem- bers of the Bank's staff and one World Bank staff member; articles must also be recommended by at least one external member of the Editorial Board. The Review may on occasion publish articles on specified topics by non-Bank contributors. Any reader interested in preparing such an article is invited to submit a proposal of not more than two pages in length to the Editor. The views and interpretations expressed in this journal are those of the authors and do not necessar- ily represent the views and policies of the World Bank or of its Executive Directors or the countries they represent. The World Bank does not guarantee the accuracy of the data included in this publication and accepts no responsibility whatsoever for any consequences of their use. When maps are used, the bound- aries, denominations, and other information do not imply on the part of the World Bank Group any judgment on the legal status of any territory or the endorsement or acceptance of such boundaries. Comments or brief notes responding to Review articles are welcome and will be considered for publica- tion to the extent that space permits. Please direct all editorial correspondence to the Editor, The World Bank Economic Review, The World Bank, Washington, D.C. 20433, U.S.A. The World Bank Economic Review is published three times a year (January, May, and September) by the World Bank. Single copies may be purchased at $12.95. Subscription rates are as follows: Individuals Institutions 1-year subscription US$30 US$50 2-year subscription US$55 US$95 3-year subscription US$70 US$130 Orders should be sent to: World Bank Publications, Box 7247-7956, Philadelphia, PA 19170-7956, U.S.A. Subscriptions are available without charge to readers with mailing addresses in developing coun- tries and in socialist economies in transition. Written request is required every two years to renew such subscriptions. (© 1998 The International Bank for Reconstruction and Development / THE WORLD BANK 1818 H Street, N.W., Washington, D.C. 20433, U.S.A. All rights reserved Manufactured in the United States of America ISBN 0-8213-4112-X; ISSN 0258-6770 Material in this journal is copyrighted. Requests to reproduce portions of it should be sent to the Office of the Publisher at the address in the copyright notice above. The World Bank encourages dis- semination of its work and will normally give permission promptly and, when the reproduction is for noncommercial purposes, without asking a fee. Permission to make photocopies is granted through the Copyright Clearance Center, Suite 910, 222 Rosewood Drive, Danvers, MA 01923, U.S.A. This journal is indexed regularly in Current Contents/Social & Behavioral Sciences, Index to Interna- tional Statistics, Journal of Economic Literature, Public Affairs Information Service, and Social Sci- ences Citation Index'. It is available in microform through University Microfilms, Inc., 300 North Zeeb Road, Ann Arbor, MI 48106, U.S.A. THE WORLD BANK ECONOMIC REVIEW Volume 12 September 1998 Number 3 Credibility of Rules and Economic Growth: 353 Evidence from a Worldwide Survey of the Private Sector Aymo Brunetti, Gregory Kisunko, and Beatrice Weder Does Economic Analysis Improve the Quality of Foreign Assistance? 385 Klaus Deininger, Lyn Squire, and Swati Basu Demographic Transitions and Economic Miracles in Emerging Asia 419 David E. Bloom and Jeffrey G. Williamson The Bolivian Social Investment Fund: 457 An Analysis of Baseline Data for Impact Evaluation Menno Pradhan, Laura Rawlings, and Geert Ridder International Evidence on the Determinants of Private Saving 483 Paul R. Masson, Tamim Bayoumi, and Hossein Samiei Unfair Trade? The Increasing Gap between World and Domestic 503 Prices in Commodity Markets during the Past 25 Years Jacques Morisset A NEW DEVELOPMENT DATABASE A Database of World Stocks of Infrastructure, 1950-95 529 David Canning Index of Authors for Volume 12 549 Index of Titles for Volume 12 551 List of Referees 553 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 353-84 Credibility of Rules and Economic Growth: Evidence from a Worldwide Survey of the Private Sector Aymo Brunetti, Gregory Kisunko, and Beatrice Weder A business environment characterized by "incredible" rules such as unclear property rights, constant policy surprises and reversals, uncertain contract enforcement, and high corruption most Uikely translates into lower investment and growth. The litera- ture on growth and policies has suggested different ways to measure the relevant un- certainties. This article proposes a new measurement approach based on firmn-level surveys and an indicator of the "credibility of rules. " Using data from a private sector survey conducted in 73 countries and covering more than 3,800 enterprises, standard cross-country growth and investment analysis indicates that low credibility of rules is associated with lower rates of investment and growth. The survey was designed to capture local entrepreneurs' views of the predictability of changes in laws and policies, of the reliability of law enforcement, of the impact of discretionary and corrupt bu- reaucracies, and of the danger of policy reversals due to changes in governments. Con- fidence in the reliability of the survey results opens many avenues for further research that could exploit the micro dimensions of this data set. The general idea that an unstable political framework reduces growth is hardly controversial. It would be expected that a business environment characterized by "incredible" rules such as unclear property rights, constant policy surprises and reversals, uncertain contract enforcement, and high corruption would trans- late into lower investrment and growth. In such an uncertain environment, entre- preneurs are reluctant to commit resources especially in projects that are charac- terized by large sunk cost (see, for example, Dixit and Pindyck 1994 and Aizenman Aymo Brunetti is with the Department of Economics at the University of Saarland and the Department of Economics at the University of Base], Gregory Kisunko is with Country Department 6 in the Europe and Central Asia Region at the World Bank, and Beatrice Weder is on leave from the Fiscal Affairs Department of the International Monetary Fund and is currently with the Department of Economics at the University of Basel. The main parts of the survey were conducted in the context of the World Development Report 1997: The State in a Changing World. The authors thank Ajay Chhibber for his support of this research. Financial support for the surveys in the industrial countries was provided by the World Bank Research Support Budgets RPO 680-51 and RPO 681-52 and by the Wirtschafts- wissenschaftliches Zentrum F6rderverein, University of Basel. For valuable comments, the authors thank William Easterly, Markus Kobler, Guy Pfeffermann, participants at research seminars held at the World Bank, the University of Bern, the Max-Planck Institute in Jena, the Swiss National Bank, the Deutsche Bank, and the University ol the United Nations in Tokyo, and three referees who provided extremely helpful suggestions and comments. The authors also thank Michael Geller and Barbara Levy for excellent editorial and research assistance. © 1998 The International Bank for Reconstruction and Development/ THE WORLD BANK 353 354 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 and Marion 1993). This reaction of the private sector not only reduces aggre- gate investment but also distorts the allocation of resources and reduces eco- nomic growth. How to measure the relevant uncertainties is less clear. Early papers in the recent wave of empirical growth analysis include measures of political instabil- ity, proxied for instance by the number of coups and revolutions (see, in particu- lar, the influential paper by Barro 1991; Brunetti 1997 provides an updated survey). Such measures have the advantage of being universally observable and therefore objective, but they are also very crude measures of the kind of uncer- tainties that affect private entrepreneurs. Subjective measures have been used to proxy for property rights insecurity and corruption by relying on country risk indicators according to expert opinions (see Mauro 1995 and Knack and Keefer 1995). These indicators are likely to reflect the concerns of entrepreneurs more closely than the overall measures of political instability. However, they are based on the perceptions of country experts and not on those of local entrepreneurs themselves. In this article we propose a new measurement approach based on firm-level surveys, and we construct an indicator of the "credibility of rules" to be used in growth regressions. The data are from a private sector survey conducted in 73 countries and covering more than 3,800 enterprises.' The survey was designed to capture local entrepreneurs' views of the predictability of changes in laws and policies, of the reliability of law enforcement, of the extent of discretionary and corrupt bureaucracies, and of the danger of policy reversals due to changes in governments. We test this indicator and its various components in standard cross- country growth and investment regressions and find that low credibility of rules is associated with lower rates of investment and growth. Section I discusses how the existing measures of political uncertainty might be incomplete and why we designed a different measurement. Section II presents the survey approach, gives information on the firms surveyed, and discusses possible problems with selection bias and measurement error. Section III ex- plains the construction of the overall indicator of credibility and its various subindicators and gives some regional information on them. Section IV presents the empirical approach, and section V provides details on the results of growth and investment regressions for the credibility indicator and its components for the 51 countries with reliable data. I. A NEW APPROACH FOR MEASURING POLICY UNCERTAINTY At the most general level, we can distinguish two channels through which poli- cies may influence economic growth: efficiency and reliability. The first branch of 1. The data for 67 countries in the World Development Report 1997 survey can be downloaded from www.worldbank.org/html/prdmg/grlhweb/wdr.html. An expanded data set of the World Development Report survey, including surveys that were conducted after completion of the report, can be downloaded from www.unibas.ch/wwz/wifor/survey. Brunetti, Kisunko, and Weder 355 the existing literature on policies and growth focuses on the efficiency of policies. It explains differences in growth with differences in macro- and microeconomic policies. Many studies have found fiscal policy variables (for example, Easterly and Rebelo 1993), monetary policy variables (for example, Fischer 1993), or trade policy variables (for example, Edwards 1992) to be related to differences in cross- country growth performance. For a survey, see Barro and Sala-i-Martin (1995), and for a comparative analysis, see Levine and Renelt (1992). The second branch of the literature emphasizes the reliability of policies, that is, their stability and uncertainties surrounding their implementation. Within this branch, most studies use "objective" measures of political instability to proxy for uncertainties. For example, Alesina and others (1996) and Barro (1991) use average numbers of violent political events such as riots or political assassina- tions. Londregan and Poole (1990) and Cukierman, Edwards, and Tabellini (1992) use the number of orderly or disorderly changes in government or an estimated probability of a change in government. Aizenman and Marion (1993), Easterly and Rebelo (1993), and Hausmann and Gavin (1996) use the standard deviations of inflation or tax incomes as indicators of the volatility of macroeco- nomic policies. Clearly, these "objective" variables are incomplete proxies for the variety of institutional uncertainties that confront entrepreneurs in their daily business operations. For instance, these proxies disregard more micro aspects that entre- preneurs consider important such as uncertainties in tax legislation, large and unpredictable changes in labor regulations, uncertain and arbitrary decisions of courts, or unclear proceedings in the allocation of all sorts of licenses. Borner, Brunetti, and Weder (1995) provide reports on interviews on these issues con- ducted with private business owners in 10 developing countries. Two examples help make the point. 'The first case is Thailand (see Brunetti and Weder 1995). Indicators of political instability that are based on counting the number of coups would characterize Thailand as a country with high political uncertainty. But the interviews we conducted with entrepreneurs suggest that the coups did not affect the credibility of the institutional framework and that entrepreneurs did not fear wide-ranging, policy swings or reversals. The second and opposite case is Peru in the 1980s (see Keefer 1990). Despite the apparent stability of the government, legislation through executive and emergency decrees was so exten- sive that the private sector faced a much more uncertain environment than could be captured by measures of the number of changes in government. These ex- amples highlight two problems of all objective indicators of political instability as proxies for policy reliability. First, they concentrate on events that the private sector may not perceive as important. Second, they fail to capture many uncer- tainties that the private sector may perceive as crucial. In essence, the disadvantage of "objective" variables is that they measure in- stability and not uncertainty. Instability can be objectively observed, whereas uncertainty is subject ive to the individual investor. Because investment decisions are based on the subjective evaluations of entrepreneurs, a variable that captures 356 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 these perceptions would seem more promising for explaining investment and growth. The subjective measures of political uncertainty that have been used in the literature are based on the opinions of external experts (see Mauro 1995 and Knack and Keefer 1995). Companies that specialize in assessing a country's risks provide such indicators. The drawback of these indicators is that they are aimed at foreign firms, and the problems for foreign investors and local entrepreneurs may differ quite substantially. For instance, to a large degree these indicators reflect risks of nationalization and impediments to repatriation of revenues that do not arise in similar intensity for domestic entrepreneurs. Also, the degree to which investors are kept abreast of regulatory changes may differ significantly for multinational and domestic firms. Finally, multinationals may receive very different treatment from politicians and bureaucrats than the large majority of small local firms. Given that in most countries the development of the private sector depends mainly on local investors, an indicator based on their percep- tions would seem a promising way to obtain a more encompassing measurement of political uncertainty and its effects on investment and growth. In this article, we aim to fill this gap by constructing a measure of the credibil- ity of rules based on a survey of domestic entrepreneurs in the private sector. In this respect, our survey extends to the cross-country level the research based on country-level interviews with domestic firms done at the World Bank (see, for example, Stone, Levy, and Paredes 1992). II. THE PRIVATE SECTOR SURVEY This section gives a short overview of the survey. We discuss the structure of the questionnaire, explain how the survey was implemented, give an overview of the characteristics of responding firms, and discuss possible problems in the data. Questionnaire, Survey Implementation, and Sample Characteristics The survey instrument was developed over the past five years. It started with a large number of interviews of private entrepreneurs in different Latin Ameri- can countries that resulted in a short multiple-choice questionnaire and small- scale survey. Results are reported in Borner, Brunetti, and Weder (1995). Based on the results of this pretest, the survey instrument was refined and expanded. In preparation for the survey sponsored by the World Development Report 1997: The State in a Changing World, the expanded questionnaire was discussed with a number of country experts at the World Bank and the International Finance Corporation. After these discussions, the questionnaire was revised and final- ized, resulting in the survey presented in this article. The purpose of the questionnaire was to capture all relevant forms of policy uncertainties related to the development and enforcement of laws, regulations, and policies. In preparatory interviews and tests of this questionnaire, firms that were confronted with unpredictable state action usually came up with very dif- Brunetti, Kisunko, and Weder 357 ferent examples of such uncertainties. These answers ranged from surprising executive decrees to unpredictable court decisions, from uncertainty on the se- verity of tax audits to unpredictable customs procedures, and from policy rever- sals whenever a new minister was appointed to uncertainty about whether a bribe would lead to blackmailing by government officials. The questionnaire tries to cover the most important forms of such uncertainties. The main part of thae questionnaire consisted of 25 mainly multiple-choice questions. We use a subset of these questions in the empirical sections of this article. In addition, we asked respondents to judge the situation 10 years ago; this allows us to construct 10-year averages of the indicators used in the econo- metric work. The process of implementing the survey began in August 1996 and ended in June 1997. At the survey's conclusion, 73 countries had participated (see appen- dix A). In 60 of these countries, the questionnaires were distributed through World Bank missions or local consulting companies. The survey of industrial countries was undertaken at the end of 1996, as a separate exercise under our direction at the University of Basel. It covered nine European countries: Austria, France, Germany, Ireland, Italy, Portugal, Spain, Switzerland, and the United Kingdom. Because the coverage of Southeast Asian economies proved to be rather poor, we later conducted surveys using the same method in four additional econo- Mies: Hong Kong (China), the Republic of Korea, Singapore, and Thailand. Companies were selected based on stratification criteria including firm size, geographic location within the country, and foreign participation. The survey was conducted by direct mail where possible. In countries where mail delivery systems were unreliable, hand delivery was used. The average rate of return on the mailed survey in developing countries was 30 percent. The response rate in high-income industriaLl countries was considerably lower, at 18 percent on aver- age. For regional details on rates of response, see Brunetti, Kisunko, and Weder (1997). Due to various constraints, not all the country surveys were based on a random sample of companies. Nevertheless, the stratification criteria should ensure a reasonable coverage. Table 1 shows regional averages and some descriptive statistics on response patterns. Over all countries, the average number of questionnaires is 53, and the median is 50. The average number of responses is lowest in the industrial coun- tries. The minimum number of questionnaires is 13, and the maximum is 124. Appendix B provides regional information about the standard deviations and coefficients of variation. Table 2 gives regional information including the distribution of size, geo- graphic location, and ownership of responding firms. About 40 percent of firms are small (less than 50 employees), about 30 percent are large (more than 200 employees), and the remaining 30 percent are of medium size. The sample, there- fore, is reasonably diversified according to this criterion. The regional decompo- sition shows considerable variation in the percentage of firm size. This reflects differences in economic development and in the development of the private sec- 358 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 1. Descriptive Statistics of Returned Questionnaires, by Region (numbers) Surveyed Returned questionnaires in each region Region Countries Firms Average Median Minimum Maximum All countries 73 3,883 53 50 13 124 High-income industrial countries 11 254 23 20 14 56 South and Southeast Asia 7 337 48 43 29 88 Middle East and North Africa 3 109 36 42 15 52 Central and Eastern Europe 11 771 70 70 46 114 Latin America and the Caribbean 9 474 53 47 17 87 Sub-Saharan Africa 22 1,288 59 48 13 124 Commonwealth of Independent States 10 650 65 62 31 91 Note: See appendix A for the list of countries by region. Source: Authors' calculations. tor itself. Services and manufacturing are represented about equally, and agri- culture is underrepresented in all regions. Regarding geographical location, there is a bias toward the capital city; however, in many countries this reflects the distribution of firms. About two-thirds of the surveyed companies do not have any foreign participation-they are purely local. This contrasts with earlier sub- jective measurements of investment climate that concentrate entirely on the per- ceptions of multinational firms. Possible Problems with Selection Bias and Measurement Error Before turning to empirical results, we discuss possible selection biases and measurement errors in our approach. In most cases, we believe that they do not seriously affect the quality of the results. A possible source of selection bias is that for most of the 60 countries sur- veyed by World Bank contacts, governments had to be asked if firms in their country could participate in the survey. This introduces the problem that the countries with low credibility and low growth could choose not to participate in the survey because their governments might fear having this fact exposed. This bias would exclude the worst cases of low credibility. Not all countries were asked in the first place because the most important constraint in determining which countries were covered was the internal administrative capacity of the World Bank to organize the survey in a short time. Of the countries that were asked, in only five did the government explicitly choose not to participate and in five more there was no official response or the resident mission preferred not to conduct the survey; this bias, therefore, does not seem to be too strong. The fact that the questionnaire involved some delicate questions on the firm's relationship with the government might be another source of selection bias. There could be two possible problems. Entrepreneurs who are completely ex- Table 2. Description of Responding Companies, by Region (percentage of all responses in the region) High-income South and Middle East Central and Latin America Sub- Commonwealth All industrial Southeast and North Eastern and the Saharan of Independent Indicator countries countries Asia Africa Europe Caribbean Africa States Company size (number of employees) Fewer than 50 39 26 32 35 40 27 43 61 Between 50 and 200 31 45 28 35 28 29 31 23 More than 200 28 28 40 26 31 42 24 15 industry Manufacturing 49 69 52 51 48 41 46 35 Services 41 27 43 42 41 47 39 57 Agriculture 8 2 4 1 10 9 11 7 Location of management Capital city 48 23 49 42 37 59 58 61 Large city 28 29 24 40 36 25 26 21 Small city or countryside 21 48 11 12 27 13 13 18 Foreign participation Yes 35 33 47 34 26 30 42 25 No 63 65 52 62 73 67 56 73 Exports Yes 49 73 55 37 51 44 46 28 No 51 27 45 63 49 56 54 72 Note: See appendix A for the list of countries by region. Source: Authors' calculations. 360 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 asperated with their government might take the opportunity to vent their an- ger, while entrepreneurs who feel reasonably happy might choose not to an- swer the survey. In this case, the bias would be consistently to underestimate credibility. The other possibility is that entrepreneurs who are desperate have given up and do not even care to submit a questionnaire. This would lead to an overestimation of credibility. Similarly, some entrepreneurs might fear that their government could discover their responses and therefore present too rosy a picture. In order to temper this fear, we conducted the survey anonymously and asked for no company-specific data that would allow identification of the responder. All in all, the direction of a possible company-level bias is not evident: it could lead to under- as well as overestimation of our variables of interest. A more serious source of measurement error could be that purely local entre- preneurs might not have the experience to put their answers in relation to the situation in other countries. About 60 percent of the total sample of enterprises were purely local, that is, they had no foreign participation and did not export. Of course, entrepreneurs might still have had good knowledge of other coun- tries (through imports, or they might even have been nationals of other coun- tries), but in the smaller enterprises the entrepreneurs probably were purely lo- cal. On the one hand, this is exactly what we wanted, because the threat of uncertainty would affect a local entrepreneur's investment behavior in the coun- try. On the other hand, if serious, this problem would compromise the compara- bility of country surveys, and we would not find any association between uncer- tainty indicators and economic performance, even though such an association might exist. The fact that we do find strong associations between economic per- formance and indicators derived from the survey is, therefore, indirect proof of the validity of the instrument. Another possible measurement problem would occur if the survey were noth- ing but an indirect measure of the private investment rate. It is conceivable that entrepreneurs would respond to questions about the business environment with their general gut feeling, that their responses would reflect not their opinion about the institutional framework but rather whether they invested or not. In this case, we would expect that entrepreneurs would respond more or less the same to all questions. In other words, if the firm had recently invested in the country, the entrepreneur would answer all the questions positively and vice versa. It seems that this was not the case. Entrepreneurs seemed to distinguish clearly between, say, the perceived political stability and the level of corruption. The degree of differentiation in the answers of the same respondents to different questions is comforting. III. THE CREDIBILITY INDICATOR AND ITS COMPONENTS This section explains the construction of the credibility indicator and presents some regional statistics on the credibility indicator and its components. Brunetti, Kisunko, and Weder 361 Construction of the Credibility Indicator The multiple-choice questions used in the survey had six standardized re- sponses. For instance., in question number 1 entrepreneurs were asked whether they had to cope regularly with unexpected changes in rules and regulations that could seriously affect their business. The six answers ranged from changes in laws and policies are "completely predictable" to "completely unpredictable." Based on such standardized answers, we constructed indexes for every question. Entrepreneurs were also asked to rate the situation 10 years ago. We constructed a 10-year average using the average of the response for 10 years earlier and the value for 1996 (for the transition economies, only S-year averages were consid- ered). For the indicators of security of property, judiciary enforcement, and per- ceived political instability, we asked directly how the rating was 10 years ago. For the indicators of predictability and corruption, we asked one overall ques- tion on developments over time. The credibility indicator was designed as a broad measure of the reliability of the institutional framework averaging the information from many such ques- tions. It encompasses several different sources of uncertainty in the interaction between the governmient and the private sector and summarizes them into one global indicator. The credibility indicator is constructed as the simple mean of the average answers for five subindicators. For the individual questions used to construct these subindicators, see appendix C. The five subindicators are the following. * Predictability of rule making. This subindicator measures the extent to which entrepreneurs have to cope with unexpected changes in rules and policies and whether they expect their governments to stick to announced major policies. It encompasses the degree to which entrepreneurs are usually informed about important changes in rules and whether they can voice their concerns when planned changes affect their business. It is the average of questions 1-4. X Subjective perception of political instability. This subindicator reflects whether government changes (constitutional and unconstitutional) are perceived to be accompanied by far-reaching policy surprises that could have serious effects on the private sector. It is the average of questions 5 and 6. * Security of persons and property. This subindicator reflects whether entrepreneurs feel confident that the authorities would protect them and their property From criminal actions and whether theft and crime represent serious problems for business operations. It is the average of questions 7 and 8. * Predictability of judicial enforcement. This subindicator captures the uncertainty arising from arbitrary enforcement of rules by the judiciary and whether s uch unpredictability presents a problem for doing business. It is question 9. 362 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 - Corruption. This subindicator asks whether it is common for private entrepreneurs to have to pay some irregular additional payments to government agents to get things done. It is question 10. Regional Statistics for the Credibility Indicator and Its Components Figure 1 shows regional averages of the credibility indicator, which ranges in value from 1 (no credibility) to 6 (perfect credibility). The high-income indus- trial countries overall prove to have the most favorable institutional framework. Their firms clearly assign the highest credibility rating. The high-growth South and Southeast Asian countries have very good credibility ratings as well. At the lower end of the regional averages, we find the Commonwealth of Independent States and the Sub-Saharan African countries. Table 3 gives more detailed information by showing the regional averages for all of the five subindicators that are used to calculate the overall credibility indi- cator. Again, the high-income industrial countries have the best rating for all of these indicators. This is especially apparent for corruption, where this region has an extremely favorable rating of 5.04, which is much higher than the next- best value of 4.12 for the South and Southeast Asian region. For all of these indicators, the region of the Commonwealth of Independent States has very low ratings. In particular, political violence and the lack of a reliable judiciary seem to be major problems in this region. Figure 1. Regional Averages of the Credibility Index index 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 High-income South and Middle East Central and Latin America Suh-Saharan Commonwealth industrial Southeast and North Eastem and the Africa of Independent countries Asia Africa Europe Caribbean States Note: See appendix A for the list of countries by region. Source: Authors' calculations. Brunetti, Kisunko, and Weder 363 Table 3. Regional Averages of the Credibility Indicator and Its Components Components of the credibility indicator Reliability Credibility Political of Lack of Region indicator Predictability stability Violence judiciary corruption All countries 3.23 3.21 3.25 2.80 3.04 3.86 High-income industrial countries 4.15 3.85 4.27 3.64 3.98 5.04 South and Southeast Asia 3.69 3.55 3.56 3.28 3.94 4.12 Middle East and North Africa 3.28 3.36 2.86 3.57 2.61 4.01 Central and Eastern Europe 3.22 2.93 3.51 2.72 3.14 3.82 Latin America and the Caribbean 3.12 3.17 3.60 2.43 2.63 3.79 Sub-Saharan Africa 2.91 3.06 2.57 2.59 2.76 3.55 Commonwealth of Independent States 2.69 2.87 2.91 2.16 2.35 3.16 Note: See appendix A :For the list of countries by region. See text section III and appendix C for details on the components of the credibility indicator. Source: Authors' calculations. As instructive as such regional comparisons are, we cannot derive any strong conclusions because a lot of information on cross-country differences is aver- aged out. Therefore, in the next sections we turn to an econometric analysis of the data set that can take advantage of individual-country ratings. IV. SPECIFICATION AND DATA SOURCES FOR THE EMPIRICAL ANALYSIS In the empirical analysis, we use cross-sectional regressions to evaluate the hypothesis that high credibility is associated with higher growth rates and higher rates of investment. Starting with the contributions by Kormendi and Meguire (1985) and in particular Barro (1991), this has become the standard method for analyzing the sources of cross-country differences in economic performance. Our indicator and subindicators of credibility are added as additional ex- planatory variables in the most common specification of such growth regres- sions. This specification regresses the average rate of growth on the starting levels of per capita gross domestic product (GDP) and human capital. The start- ing level of per capita GDP controls for the convergence effect predicted by neoclassical growth theory. That is, the higher initial GDP per capita is, the lower the growth r;ate is, other things being equal, because decreasing returns to capital reduce the growth effects of additional capital. According to this argument, a country starting with a low level of GDP should grow faster and gradually converge to the levels of industrial countries. The problem with this approach is that it does not work for country samples that include developing and industrial countries (see Barro and Sala-i-Martin 1992). Mankiw, Romer, 364 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 and Weil (1992) have argued that the neoclassical growth model predicts not absolute but rather conditional convergence. Each country does converge, not to a common steady state but to its own steady state, which depends on coun- try characteristics, most prominently the level of human capital. As a conse- quence, more recent cross-country growth regression analysis has included, as we do, at least one measure of human capital as an additional right-hand vari- able in the basic specification. In addition to testing the credibility measure in this basic specification, we check whether the results are sensitive to adding individual additional explana- tory variables that are frequently used in the empirical growth analysis. The specification we test, therefore, has the following form: (1) Growth8O92 = ao + a1 GDP80 + a2 SEC80 + a3 Credibility + a4 X + u. Growth8092 is the average per capita growth rate for 1980-92 calculated from the updated data set provided by Summers and Heston (1991). GDP80 is per capita GDP in 1980 from the same data set. SEC80 is the enrollment ratio in secondary school in 1980 from the UNESCO Statistical Yearbook. Credibility is the average indicator calculated from our survey approach for the last decade. X is an additional variable that is drawn from a set of standard explanatory eco- nomic and political variables for economic growth; appendix D provides precise definitions and data sources. We use 1980-92 because this is the most recent period for which Summers and Heston data are available on the Internet. A potential problem with this approach is that this period includes the "lost decade" in the aftermath of the debt crisis. Other empirical growth studies have used averages for longer peri- ods, arguing that institutional variables are fundamental country characteristics that do not change much over time. To test if the time period has any effect on our results, we run regressions with macro variables for 1970-92. The fit of the regressions and the significance of the credibility indicator improve in every case. As in all cross-sectional growth analysis, this raises the issue of causality. Due to the notorious problem of finding adequate instruments, this issue is very hard to address. For a discussion, see Mankiw (1995). As control variables, we include the following frequently used measures: the average rate of inflation in 1980-92 calculated from World Bank data, the aver- age rate of government consumption as a percentage of GDP in 1980-92 pro- vided by Summers and Heston, the average degree of openness to international trade measured as the sum of exports and imports as a percentage of GDP in 1980-92 calculated from Summers and Heston, and the average level of liquid liabilities in GDP in the 1980s from King and Levine (1993). In addition we analyze how credibility affects economic growth. Credibility can influence growth either by affecting the accumulation of capital or by affect- ing the allocation of capital. We try to disentangle these effects by separately estimating investment and growth regressions. The investment regressions mea- sure the effect of credibility on accumulation. Brunetti, Kisunko, and Weder 365 V. REGRESSION RESULTS This section presents the results of the cross-country regression analysis in four steps. The first subsection presents the basic results of the overall credibility indicator in the growth and investment regressions for 51 countries for which we have reliable data (see appendix A for the country list).2 The second subsec- tion individually tests each of the five subindicators that together make up the credibility indicator. lThe third subsection comparies the credibility indicator with other political variables. Finally, in the last subsection, we do some exploratory analysis with data from 18 transition economies in our sample. Basic Growth and Investment Results We first test the relation between the aggregate indicator of credibility and average per capita growth rates for 1980-92. A highei value of this indicator means a more credible institutional framework. Therefore we expect a positive relationship. The simple scatter plot is shown in figure 2. Table 4 displays inultivariate regression results. The first regression shows that the sign of the coefficient is positive in the basic specification that contains GDP per capita and secondary school enrollment as additional right-hand vari- ables. The coefficient is significant at the 1 percent confidence level. Regressions 2 to 5 test whether this result is sensitive to the inclusion of additional explana- tory variables. Controlling for the rate of government consumption and the rate of inflation, the coefficient of the credibility indicator has the expected positive sign and remains highly significant. If we include the extent of international trade, the coefficient of the indicator is significant only at the 5 percent confi- dence level. In the last regression, we control for the level of financial depth, meaning the level of liquid liabilities of the banking system. In this case, credibil- ity is not significant. However, credibility and financial depth are highly corre- lated (simple correlation of 0.80). Appendix E provides the correlation matrix of the survey indicators with all economic variables. The high correlation between credibility and financial depth might be no co- incidence because the two variables might be measuring the same phenomenon. Clague and others ('1996) suggest that the depth of the financial system can also be interpreted as a variable of contract enforcement. The less confidence there is in contract enforcement, the less intermediation occurs through the banking sys- tem. If this interpretation is correct, the fact that credibility becomes insignifi- cant when controlling for liquid liabilities of the banking system is not a cause for great concern. Therefore, in the following regressions, we exclude this mea- sure as a control variable. 2. We drop one country (Jordan), although we have the necessary macroeconomic data because it is an extreme outlier in several dimensions. Including Jordan leads to specification problems because the main control variables (GDP in the base year and secondary school enrollment) are not significant. The significance of the credibility indicator is also reduced a bit when including this country, but all results nevertheless hold. 366 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 2. Growth in Per Capita GDP and the Credibility Indicatorfor 49 Countries, 1980-92 Average annual growth in per capita GDP (percent) 0.14 " O Sub-Saharan Africa A 0,12 . A South and South-east Asia * Latin America and the Caribbean 0.10 X High-incomer industrial countries o Other' 0.08- - A A 0.06*. 0.04- - * x 00 a -0.02 - * O O -0.04 * -/ / i e|1 2 3 4 5 6 Credibility indicator Note: Eact point represents a country. See appendix A for the list of countries by region; all the countries with asterisks are included except Morocco and Turkey. a. Other includes the Middle East and North Africa and the Commonwealth of Independent States. Source: Authors' calculations; see appendix D. We proceed to check whether credibility has a positive impact on growth through higher rates of investment. Figure 3 shows that investment and credibil- ity are highly correlated. Table S presents regression results for the impact on investment using the same set of variables as in the growth regression. Regression 1 in table 5 shows that the coefficient of credibility has the expected positive sign and is significant at the 1 percent level. Together with the variables for initial human capital and GDP per capita, this minimal specification explains 65 percent of the cross-coun- try variation in investment rates. The result proves to be quite robust, as can be seen in the extended specifications tested in regressions 2 to 4. When controlling for government consumption and inflation, credibility remains positive and sig- nificant at the 1 percent level of confidence. If we include the extent of interna- tional trade, the significance of the credibility indicator drops to the 5 percent level. The investment regressions test for the effect of credibility on the accumula- tion of resources. In order to test the allocation channel, we include investment as a control variable in the growth regression. Credibility keeps its positive sign Brunetti, Kisunko, and Weder 367 Table 4. The Impact of the Credibility Indicator on Growth in Per Capita GDP for 51 Countries, 1980-92 Variable 1 2 3 4 5 Constant -0.07- -0.06* -0.06* -0.06& -0.04' (-3.88) (-2.66) (-3.24) (-3.5) (-2.1) GDP per capita in 1980 -2.65 E-6' -2.69 E-6* -2.97 E-6** -2.2 E-6 -3.68 E-6- (-1.90) (-1.92) (-2.06) (-1.6) (-2.6) Secondary school enrollment 0.037* 0.029 0.043** 0.036' 0.05" rate in 1980 (1.85) (1.34) (2.06) (1.87) (2.5) Government consumption/ -0.0005 GDP average in 1980-92 (-0.94) Inflation rate average in -0.01 1980-92 (-1.55) Trade in 1980-92 0.0001't (2.03) Liquid liabilities in 1980-90 0.038** (2.28) Credibility indicator 0.022*'* 0.022- 0.021-' 0.017- 0.08 (3.35) (3.36) (2.95) (2.48) (1.04) Number of observations 51 51 49 51 47 Adjusted R2 0.35 0.35 0.37 0.39 0.38 Note: The dependent variable is growth in GDP per capita in 1980-92. The regressions are estimated using ordinary least squares. See appendix A for the list of countries by region and appendix D for variable descriptions. t-statistics are in parentheses. * Significant at 10 percent. Significant at 5 percent. S** Significant at 1 percent. Source: Authors' calculations; see appendix D. in these regressions but becomes insignificant (results not shown). This result suggests that higher credibility affects growth mainly through investment, that is, by raising the rate of capital accumulation. Subcomponents of the Credibility Indicator Table 6 presents results for the individual subcomponents of the credibility indicator for the basic growth regressions in 1980-92 and 1970-92. Each entry in this table presents the results from a separate equation. We only report the coefficients and t-statistics of the institutional variable, but in the regression we control for GDP per capita in 1980 and secondary school enrollment in 1980. We have shown that credibility affects growth mainly through the accumulation channel. Table 6 also shows results for the investment regressions. Three of the five subcomponents are significant in all regressions. The most robust one is the incdicator of predictability of judiciary enforcement, which is significant at the 1 percent level in all regressions. The indicator of the security of persons and property rights is highly significant in both growth regressions and remains significant in the investment regressions, albeit at a lower level of confidence. The reverse is true for the indicator of perceived political instabil- ity. The indicator of the predictability of rule making is the most fragile of all the subcomponents. It is only significant in the growth regression for the longer 368 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 3. Investment and the Credibility Indicatorfor 49 Countrties, 1980-92 Average annual investment/GDP (percent) 40 1 Sub-Saharan Africa 35 * South and South-east Asia 30 1* Latin America and the Caribbean A X High-incomer industrial countries * 25 ° Other' X 20 * A 15 S * 10 * a 5 ea O a/ I I II - 2 3 4 5 6 Credibility indicator Note: Eact point represents a country. See appendix A for the list of countries by region; all the countries with asterisks are included except Morocco and Turkey. a. Other includes the Middle East and North Africa and the Commonwealth of Independent States. Source: Authors' calculations; see appendix D. period. The opposite pattern applies to the indicator of corruption: it is signifi- cantly associated with investment but only weakly with growth. It is interest- ing to note that Mauro (1995) obtains the same result. He uses business expert data and a different sample of countries but also finds that corruption directly affects investment but not growth. The last row in table 6 shows that the re- sults for the overall credibility indicator are not specific to the 1980s. Credibil- ity is highly significant in both the investment and growth regressions for the longer period. Comparison with Other Political Variables Our next step is to compare the credibility indicator with other political vari- ables that are frequently used in cross-country growth analysis. Table 7 shows the results for the base growth regressions for 1980-92. Each regression in- cludes four right-hand variables: initial level of GDP per capita and secondary school enrollment (coefficients are not shown in the table), credibility, and one other political variable. Table 7 shows that credibility remains significant in all but one regression. The other political variables are all insignificant. The first political variable is Brunetti, Kisunko, and Weder 369 Table 5. Impact of the Credibility Indicator on InvestmentlGDP for 51 Countries, 1980-92 Variable 1 2 3 4 Constant -8.10 -4.18 -7.77 -6.34 (-1.78) (-0.76) (-1.50) (-1.40) GDP per capita in 1980 -2.5 E-4 -2.7 E-4 -2.4 E-4 -1.5 E-4 (-0.72) (-0.76) (-0.65) (-0.42) Secondary school enrollment 16.58''' 13.97"* 16.58"-' 16.44-- rate in 1980 (3.30) (2.59) (3.04) (3.36) Government consumption/GDP -0.15 ' average in 1980-92 (-1.26) Inflation rate average in 1980-92 -0.40 (-0.22) Trade in 1980-92 0.026- (1.89) Credibility indicator 5.38"- 5.40-' 5.29... 4.18- (3.28) (3.33) (2.93) (2.45) Number of observations 51 51 49 51 Adjusted R2 0.65 0.65 0.62 0.67 Note: The dependent variable is the average investment rate in 1980-92. The regressions are estimated using ordinary least squares. See appendix A for the list of countries by region and appendix D for variable descriptions. t-statistics are in parentheses. Significant at 10 percent. * Significant at 5 percent. * Significant at 1 percent. Source: Authors' calculations; see appendix D. Table 6. Coefficients for the Credibility Indicator and Its Subcomponents in Growth and Investment Regressions for 51 Countries, 1980-92 and 1970-92 Dependent variable Growth in per capita GDP Investment/GDP Indicator 1980-92 1970-92 1980-92 1970-92 Predictability of rule making 0.013 0.017- 2.50 3.73 (1.32) (2.17) (1.03) (1.54) Perceived political instability 0.008' 0.01'- 2.74- 3.57-' (1.77) (2.65) (2.51) (3.33) Security of persons and property rights 0.021-' 0.018-' 3.06" 3.40" (4.18) (4.13) (2.17) (2.39) Reliability of judiciary enforcement 0.013"'' 0.09'- 2.81-' 2.97'- (3.79) (3.16) (3.13) (3.27) Corruption 0.004 0.07 2.26'' 3.84-' (0.71) (1.60) (1.97) (3.21) Credibility indicator 0.023"' 0.019-' 5.36"' 6.16"' (3.35) (3.81) (3.28) (4.03) Note: Each coefficient :reported in the table is from estimation of a separate equation in which the dependent variable is regressed on the indicator, GDP per capita in 1980, and the secondary school enrollment rate in 1980. The regressions are estimated using ordinary least squares. See appendix A for the list of countries by region. See text section III and appendix C for details on the indicators. t-statistics are in parentheses. Significant at 10 percent. * Significant at 5 percent. * * * Significant at 1 percent. Source: Authors' calculations; see appendix D. 370 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO.3 Table 7. Impact of the Credibility Indicator and Other Political Indicators on Growth in GDP for Si Countries, 1980-92 Variable 1 2 3 4 s Political rights -0.002 (-0.85) Assassinations -120.8 (-0.92) Coups 0.01 (1.19) Wars -0.003 (-0.45) Institutional quality 0.007 (1.4) Credibility indicator 0.019... 0.021-' 0.02-' 0.022... 0.01 (2.89) (3.12) (2.97) (3.34) (1.43) Number of observations 50 49 49 51 42 Note: The dependent variable is growth in GDP per capita in 1980-92. Each regression includes two additional right-hand variables: GDP per capita in 1980 and the secondary school enrollment rate in 1980. The regressions are estimated using ordinary least squares. See appendix A for the list of countries by region and appendix D for variable descriptions. t-statistics are in parentheses. * * Significant at 1 percent. Source: Authors' calculations; see appendix D. the indicator of political rights compiled by Freedom House (on the internet). It is used as a measure of the level of democracy and ranges from a high of 1 to a low of 7. A negative sign on the coefficient therefore means that more democ- racy is associated with higher growth. From a theoretical point of view, it is not clear whether a more democratic system necessarily leads to more growth than a less democratic system. Empirically, the recent cross-country analysis has found no significant association between the level of democracy in a country and its long-term growth performance (see Brunetti and Weder 1995 for a survey of the literature on democracy and growth). This result is reproduced in our sample; the coefficient of the indicator is not significant at the 10 percent level. When we control for political rights, credibility remains significant and has the expected sign. The simple correlation between credibility and political rights is fairly high (-0.67), which indicates that democracy and credibility often go together. See appendix F for the correlation matrix of all political variables. The next three variables are objective indicators of political stability taken from Easterly and Levine (1997). We would expect negative signs on all of the coefficients. However, all three variables are not even close to significance. Cred- ibility, by contrast, remains significantly associated with growth. The simple correlation between credibility and these three indicators is very low, ranging from 0.07 to -0.17. The last political variable is a subjective indicator of institutional quality com- piled by International Country Risk Guide (ICRG), a professional risk-rating com- pany (see Knack and Keefer 1995 for details). The indicator we use is an average of corruption, rule of law, and quality of the bureaucracy. The institutional Brunetti, Kisunko, and Weder 371 quality indicator is the one most readily comparable with the credibility indica- tor. It tries to capture similar problems, albeit by asking country experts rather than the local private sector. The results of this regression are not fully compa- rable with the results of the other regressions in table 7 because the ICRG indica- tor was available only for 42 countries in our sample. The regression shows that institutional quality has the expected sign but is not significant. Credibility loses significance in this specification because the two indicators are very highly cor- related. The simple correlation is 0.83. This is an interesting result in itself be- cause it provides an indirect check on the validity of the survey method. It shows that local private sectors have on average expressed views similar to those of country experts. The advantage of the surveys are that they provide much more disaggregated information than risk-assessment companies do. Results for an Extended Sample of Transition Economies Here we present results for a set of transition economies for which data could be gathered. The figure shown should be regarded as tentative mainly because of data limitations in transition economies. The results are not directly comparable with the previous sections because we have to work with different growth data than for the sample ol: 51 countries. In addition, given that 10-year averages are not very sensible in the case of transition economies, we look at the average growth rate for 1990-95 (provided by the World Bank). The scatter plot of credibility and growth is shown in figure 4. The scatter plot indicates that credibility may also contribute to explaining differences in growth performance in the transition economies. However, the results for this sample have to be interpreted with caution mainly because of the short observed time period, as well as intrinsic problems of measuring and ex- plaining growth in countries that went through such a major systemic change. Therefore, we do not perform a more formal econometric analysis. VI. CONCLUSIONS AND DIRECTIONS FOR FURTHER RESEARCH This article has analyzed a new data set and some results that indicate a close association between indicators of institutional uncertainty derived from this data set and economic growth. The two premises of the research were that institutional uncertainty should be crucial in explaining cross-country differences in economic performance and that existing measures of this relationship are incomplete or crude. We suggested that indicators of the subjective perceptions of private entre- preneurs could be a promising way of measuring the relevant uncertainties. Our results seem to support the propositions. We constructed an overall indi- cator of credibility based on survey data and tested whether it contributes to explaining differences in growth and investment across countries. We found that credibility was significantly associated with cross-country differences in growth and investment in a sample of 51 countries for which comparable data were available. This result was strengthened when we looked at an extended time 372 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 4. Growth in Per Capita GDP and the Credibility Indicator for 18 Transition Economies, 1990-95 Average annual growth in per capita GDP (percent) Poland Albania / r Hungary Slovak Republic ;zch Republic c5 Uzbekistan * Bu 0 Russia / Estonia -10- ~~~~~~~~~Belarulls * Kazakhstan L-atvia -15- - Kyrgyz Republic * * Ukraine * Lithuania -20- - Azerbaijan *Armenia | Georgia _30 1 / I 2 3 4 5 Credibility indicator Source: Authors' calculations; see appendix D. period that might capture the long-term relationship between institutions and growth more adequately. In addition to testing whether institutions matter for growth, this study was also a first test of the quality of the new data. Although there are many potential problems with subjective indicators derived from surveys that are uniformly conducted in very diverse settings, the close association with economic perfor- mance suggests that these indicators are quite reliable. Furthermore, the cred- ibility indicator proves to be highly correlated with indicators of institutional quality provided by risk-assessment companies. This can be interpreted as a further confirmation of the quality of the survey data. Confidence in the reliability of the survey results opens many avenues for further research that could exploit the micro dimensions of this data set. Pos- sible research questions include whether we can say more about the effects of different uncertainties. For instance, is corruption more harmful than judiciary uncertainty? Are there important regional differences in uncertainties? Are firms of different sizes affected differently by uncertainties? How can we explain dif- ferences in institutional variables across countries? Brunetti, Kisunko, and Weder 373 APPENDIX A. COUNTRY LIST The 51 countries used in the econometric analysis are marked with an asterisk ( .) Sub-Saharan Africa High-income industrial countries Benin* Austria* Cameroon * Canada* Chad* France - Congo* Germany* Ghana* Ireland * Guinea* Italy* Guinea-Bissau* Portugal* C6te d'Ivoire* Spain* Kenya* Switzerland* Madagascar * United Kingdom-' Malawi* United States * Mali* Mauritius* Middle East and North Africa Mozambique-' Jordan Nigeria* Morocco* Senegal* West Bank and Gaza South Africa* Tanzania* Commonwealth of Independent Togo* States Uganda* Armenia Zambia* Azerbaijan Zimbabwe * Belarus Georgia South and Southeast Asia Kazakhstan Fiji* Kyrgyz Republic Hong Kong (China)* Moldova India'* Russia Korea, Rep. of* Ukraine Malaysia'* Uzbekistan Singapore* Thailand'* Central and Eastern Europe Albania Latin America and the Caribbean Bulgaria Bolivia* Czech Republic Colombia* Estonia Costa Rica * Hungary Ecuador* Latvia Jamaica* Lithuania Mexico'* Macedonia Paraguay* Poland Peru* Slovak Republic Venezuela'* Turkey'* APPENDIX B. REGIONAL VARIABILIIY OF POLITICAL INDICATORS Political Reliability Lack Region Credibility Predictability stability Violence of judiciary of corruption Standard deviation All countries 0.63 0.46 0.86 0.65 0.93 0.86 High-income industrial countries 0.57 0.46 0.88 0.48 0.79 0.72 South and Southeast Asia 0.56 0.47 0.76 0.52 0.83 1.03 Middle East and North Africa 0.49 0.23 0.65 0.72 1.02 0.67 Central and Eastern Europe 0.42 0.30 0.70 0.49 0.91 0.61 Latin America and the Caribbean 0.38 0.28 0.41 0.48 0.83 0.40 Sub-Saharan Africa 0.44 0.33 0.60 0.41 0.73 0.73 Commonwealth of Independent States 0.15 0.18 0.33 0.15 0.29 0.37 Coefficient of variation All countries 0.20 0.14 0.26 0.23 0.31 0.22 High-income industrial countries 0.14 0.12 0.21 0.13 0.20 0.14 South and Southeast Asia 0.15 0.13 0.21 0.16 0.21 0.25 Middle East and North Africa 0.15 0.07 0.23 0.20 0.39 0.17 Central and Eastern Europe 0.13 0.10 0.20 0.18 0.29 0.16 Latin America and the Caribbean 0.12 0.09 0.11 0.20 0.32 0.11 Sub-Saharan Africa 0.15 0.11 0.23 0.16 0.26 0.20 Commonwealth of Independent States 0.06 0.06 0.11 0.07 0.12 0.12 Note: See appendix A for the list of countries by region. See text section III and appendix C for details on the indicators. Source: Authors' calculations. Brunetti, Kisunko, and Weder 375 APPENDIX C. QUESTIONS USED TO CONSTRUCT THE CREDIBILITY INDICATOR Section III in the text explains the components of the credibility indicator. They were calculated by assigning a 1 for the least favorable and a 6 for the most favorable rating for the answer to each question. 1. Do you regularly have to cope with unexpected changes in rules, laws, or policies that materially affect your business? Changes in laws and policies are (1) Completely predictable (2) Highly predictable (3) Fairly predictable (4) Fairly unpredictable (5) Highly unpredictable (6) Completely unpredictable 2. Do you expect the government to stick to announced major policies? (1) Always (2) Mostly (3) Frequently (4) Sometimes (5) Seldom (6) Never 3. "The process of developing new rules or policies is usually such that affected businesses are informed." This is true (1) Always (2) Mostly (3) Frequently (4) Sometimes (5) Seldom (6) Never 4. "In case of important changes in laws or policies affecting my business opera- tion, the government r1akes into account concerns voiced either by me or by my business association." This is true (1) Always (2) Mostly (3) Frequently (4) Sometimes (5) Seldom (6) Never 5. "Constitutional changes of government (as a result of elections) are usually accompanied by large changes ih rules and regulations that have an impact on my business." To what degree do you agree with this statement? (1) Fully agree (2) Agree in most cases (3) Tend to agree (4) Tend to disagree (5) Disagree in most cases 376 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 (6) Strongly disagree Does not apply 6. "I constantly fear unconstitutional government changes (i.e., coups) that are accompanied by far-reaching policy surprises with significant impact on my business." To what degree do you agree with this statement? (1) Fully agree (2) Agree in most cases (3) Tend to agree (4) Tend to disagree (5) Disagree in most cases (6) Strongly disagree Does not apply 7. "Theft and crime are serious problems that can substantially increase the costs of doing business." To what degree do you agree with this statement? (1) Fully agree (2) Agree in most cases (3) Tend to agree (4) Tend to disagree (5) Disagree in most cases (6) Strongly disagree 8. "I am not confident that the state authorities protect my person and my prop- erty from criminal actions." To what degree do you agree with this statement? (1) Fully agree (2) Agree in most cases (3) Tend to agree (4) Tend to disagree (5) Disagree in most cases (6) Strongly disagree 9. "Unpredictability of the judiciary presents a major problem for my business operations." To what degree do you agree with this statement? (1) Fully agree (2) Agree in most cases (3) Tend to agree (4) Tend to disagree (5) Disagree in most cases (6) Strongly disagree 10. "It is common for firms in my line of business to have to pay some irregular 'additional payments' to get things done." This is true (1) Always (2) Mostly (3) Frequently (4) Sometimes (5) Seldom (6) Never APPENDIX D. DESCRIPTION AND SOURCES OF NONSURVEY VARIABLES Variable Description Period Source Growth in GDP per capita Average annual growth of real GDP per capita 1980-92, 1970-92 Penn World Tables 5.6 GDP Reai GDP per capita in base year 1980, 1970 Peni world Tables 5.V Secondary school enrollment rate Secondary school enrollment in base year 1980, 1970 Penn World Tables 5.6 Government consumption Annual average of government consumption as a percentage of GDP 1980-92, 1970-92 Penn World Tables 5.6 Inflation Annual average of inflation 1980-92, 1970-92 World Bank data Trade Annual average of the sum of exports and imports as a percentage of GDP 1980-92, 1970-92 Penn World Tables 5.6 Investment Annual average of investment as a percentage of GDP 1980-92, 1970-92 Penn World Tables 5.6 9 Liquid liabilities Annual average of the ratio of liquid liabilities over GDP 1980-90 King and Levine (1993) Political rights Annual average of indicator for political rights: 1 (high) to 7 (low) 1984-93 Freedom House (various years) Assassinations Annual average number of assassinations per million population 1980-89 Easterly and Levine (1997) Coups Number of coups 1980-89 Easterly and Levine (1997) Wars Dummy for war in period 1980-89 Easterly and Levine (1997) Institutional quality Subjective expert opinion on corruption, rule of law, and quality of bureaucracy: 0 (low) to 6 (high) 1980-92 Knack and Keefer (1995) 378 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 APPENDIX E. CORRELATION MATRIX FOR SURVEY DATA AND ECONOMIC VARIABLES Variable Political or indicator Credibility Predictability stability Violence Judiciary Credibility 1.00 Predictability 0.83 1.00 Political stability 0.87 0.69 1.00 Violence 0.80 0.58 0.62 1.00 Judiciary 0.90 0.64 0.68 0.74 1.00 Corruption 0.88 0.78 0.69 0.54 0.73 Growth, 1970-92 0.47 0.35 0.41 0.50 0.49 Growth, 1980-92 0.52 0.35 0.43 0.59 0.58 Investment, 1970-92 0.75 0.60 0.70 0.57 0.63 Investment, 1980-92 0.75 0.60 0.69 0.60 0.66 GDP per capita, 1970 0.69 0.69 0.62 0.50 0.49 GDP per capita, 1980 0.76 0.74 0.67 0.57 0.56 School, 1970 0.71 0.69 0.67 0.52 0.52 School, 1980 0.74 0.69 0.69 0.56 0.57 Inflation, 1970-92 -0.23 -0.28 -0.03 -0.25 -0.27 Inflation, 1980-92 -0.26 -0.30 -0.10 -0.27 -0.28 Government, 1970-92 -0.49 -0.52 -0.51 -0.37 -0.33 Government, 1980-92 -0.47 -0.50 -0.50 -0.36 -0.33 Trade, 1970-92 0.33 0.35 0.22 0.17 0.36 Trade, 1980-92 0.35 0.37 0.23 0.19 0.38 Liquid liabilities 0.80 0.69 0.69 0.69 0.69 Note: See appendix D for descriptions of the variables and text section III and appendix C for details on the credibility indicator and its components. Correlations are for 51 countries for all variables with the exception of inflation in 1970-92 and 1980-92 (45 countries) and liquid liabilities (47 countries). Source: Authors' calculations; see appendix D. Brunetti, Kisunko, and Weder 379 Grcowth Investment GDP per capita Corruption 1970-92 1980-92 1970-92 1980-92 1970 1980 1.00 0.27 1.00 0.29 0.89 1.00 0.68 0.58 0.55 1.00 0.65 0.64 0.62 0.97 1.00 0.71 0.10 0.19 0.60 0.59 1.00 0.75 0.27 0.30 0.68 0.68 0.97 1.00 0.66 0.36 0.44 0.71 0.69 0.84 0.87 0.67 0.42 0.45 0.76 0.77 0.81 0.86 -0.17 -0.21 -0.23 -0.04 -0.09 -0.13 -0.17 -0.21 -0.25 -0.25 -0.10 -0.14 -0.15 -0.19 -0.42 -0.49 -0.39 -0.54 -0.59 -0.49 -0.56 -0.39 -0.49 -0.39 -0.53 -0.58 -0.49 -0.57 0.33 0.60 0.41 0.35 0.35 -0.01 0.13 0.34 0.64 0.45 0.37 0.38 0.00 0.14 0.68 0.45 0.49 0.76 0.77 0.67 0.71 (Table continues on the following pages.) 380 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 APPENDIX E. (continued) Variable School Inflation or indicator 1970 1980 1970-92 1980-92 Credibility Predictability Political stability Violence Judiciary Corruption Growth, 1970-92 Growth, 1980-92 Investment, 1970-92 Investment, 1980-92 GDP per capita, 1970 GDP per capita, 1980 School, 1970 1.00 School, 1980 0.91 1.00 Inflation, 1970-92 -0.10 0.01 1.00 Inflation, 1980-92 -0.14 -0.06 0.99 1.00 Government, 1970-92 -0.53 -0.64 -0.05 0.02 Government, 1980-92 -0.53 -0.65 -0.07 -0.01 Trade, 1970-92 0.15 0.16 -0.22 -0.22 Trade, 1980-92 0.17 0.18 -0.22 -0.22 Liquid liabilities 0.67 0.62 -0.31 -0.31 Brunetti, Kisunko, and Weder 381 Government Trade Liquid 1970-92 1980-92 1970-92 1980-92 liabilities 1.00 0.98 1.00 -0.17 -0.18 1.00 -0.22 -0.23 1.00 1.00 -0.44 -0.40 0.35 0.38 1.00 APPENDIX F. CORRELATION MATRIX FOR ALL POLITICAL VARIABLES Variable or Political Political Institutional indicator Credibility Predictability stability Violence Judiciary Corruption rights Assassinations Coups Wars quality Credibility 1.00 Predictability 0.83 1.00 Political stability 0.87 0.69 1.00 Violence 0.80 0.58 0.62 1.00 Judiciary 0.90 0.64 0.68 0.74 1.00 Corruption 0.88 0.78 0.69 0.54 0.73 1.00 Political rights -0.67 -0.57 -0.77 -0.53 -0.45 -0.54 1.00 Assassinations 0.07 -0.04 0.17 -0.10 0.02 0.15 -0.16 1.00 Coups -0.12 -0.20 -0.18 -0.02 -0.05 -0.07 0.05 0.11 1.00 Wars -0.17 -0.19 -0.11 -0.28 -0.09 -0.13 0.19 0.38 0.16 1.00 Institutional quality 0.83 0.77 0.67 0.66 0.73 0.81 -0.65 -0.14 -0.28 -0.33 1.00 Note: The first six indicators are calculated from the survey results. Correlations for these are for 51 countries. The institutional quality variable is only available for 41 countries, political rights and wars for 50, assassinations for 49, and coups for 48. See appendix D for descriptions of the variables and text section III and appendix C for details on the credibility indicator and its components. Source: Authors' calculations; see appendix D. Brunetti, Kisunko, and Weder 383 REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Aizenman, Joshua, and Nancy Marion. 1993. "Policy Uncertainty, Persistence, and Growth." Review of International Economics 1(2):145-63. Alesina, Alberto, Sule COezler, Nouriel Roubini, and Philip Swagel. 1996. "Political In- stability and Economic Growth." Journal of Economic Growth 1(2):189-211. Barro, Robert. 1991. "Economic Growth in a Cross Section of Countries." Quarterly Journal of Economics 106(2):407-43. Barro, Robert, and Xavier Sala-i-Martin. 1992. "Convergence." Journal of Political Economy 100(2):223-51. . 1995. Economic Growth. New York: McGraw Hill. Borner, Silvio, Aymo Brunetti, and Beatrice Weder. 1995. Political Credibility and Eco- nomic Development. London: Macmillan. Brunetti, Aymo. 1997. 'Political Variables in Cross-Country Growth Analysis." Jour- nal of Economic Surveys 11(2):163-90. Brunetti, Aymo, Gregory Kisunko, and Beatrice Weder. 1997. "Institutional Obstacles to Doing Business: Region-by-Region Results from a Worldwide Survey of the Pri- vate Sector." Policy Research Working Paper 1759. Policy Research Department, World Bank, Washington, D.C. Processed. Brunetti, Aymo, and Beatrice Weder. 1995. "Political Sources of Growth: A Critical Note on Measurement." Public Choice 82(1-2):125-34. Clague, Christopher, Philip Keefer, Stephen Knack, and Mancur Olson. 1996. "Prop- erty and Contract Rights in Autocracies and Democracies." Journal of Economic Growth 1(2):243-76. Cukierman, Alex, Sebastian Edwards, and Guido Tabellini. 1992. "Seigniorage and Political Instability." American Economic Review 82(3):537-55. Dixit, Avinash, and Robert Pindyck. 1994. Investment under Uncertainty. Princeton, N.J.: Princeton University Press. Easterly, William, and Ross Levine. 1997. "Africa's Growth Tragedy: Policies and Eth- nic Division." Quarterly Journal of Economics 1 12(4):1203-50. Easterly, William, and Sergio Rebelo. 1993. "Fiscal Policy and Economic Growth: An Empirical Investigation." Journal of Monetary Economics 32(3):417-58. Edwards, Sebastian. 1992. "Trade Orientation, Distortions, and Growth in Developing Countries." Journal of Development Economics 39(1):31-57. Fischer, Stanley. 1993. "The Role of Macroeconomic Factors in Growth." Journal of Monetary Economics 32(3):485-S12. Freedom House. Various issues. "Freedom in the World: The Annual Survey of Political Rights and Civil Liberties." New York. Processed. Hausmann, Ricardo, and Michael Gavin. 1996. "Securing Stability and Growth in a Shock- Prone Region: The Plolicy Challenge for Latin America." In Ricardo Hausmann and Helmut Reisen, eds., Securing Stability and Growth in Latin America: Policy Issues and Prospects. Paris: Organisation for Economic Co-operation and Development. Keefer, Philip. 1990. "Economic Growth When There Are No Rules for the Creation of Rules." Institute for Liberty and Democracy, Lima. Processed. 384 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 King, Robert, and Ross Levine. 1993. "Finance and Growth: Schumpeter Might Be Right." Quarterly Journal of Economics 108(3):717-37. Knack, Stephen, and Philip Keefer. 1995. "Institutions and Economic Performance: Cross- Country Tests Using Alternative Institutional Measures." Economics and Politics 7(3):207-27. Kormendi, Roger, and Philip Meguire. 1985. "Macroeconomic Determinants of Growth: Cross-Country Evidence." Journal of Monetary Economics 16(2):141-63. Levine, Ross, and David Renelt. 1992. "A Sensitivity Analysis of Cross-Country Growth Regressions." American Economic Review 82(4):942-63. Londregan, John, and Keith Poole. 1990. "Poverty, the Coup Trap, and the Seizure of Executive Power." World Politics 42(2):151-83. Mankiw, Gregory. 1995. "The Growth of Nations." Brookings Papers on Economic Activity 1:275-313. Mankiw, Gregory, David Romer, and David Weil. 1992. "A Contribution to the Empirics of Economic Growth." Quarterly Journal of Economics 107(2):407-37. Mauro, Paolo. 1995. "Corruption and Growth." Quarterly Journal of Economics 110(3):681-712. Stone, Andrew, Brian Levy, and Ricardo Paredes. 1992. "Public Institutions and Private Transactions: The Legal and Regulatory Environment for Business Transactions in Brazil and Chile." Policy Research Working Paper 891. Policy Research Department, World Bank, Washington, D.C. Processed. Summers, Robert, and Alan Heston. 1991. "The Penn World Table (Mark 5): An Ex- panded Set of International Comparisons, 1950-198 8." Quarterly Journal of Eco- nomics 106(2):327-68. Update to Penn World Tables 5.6. UNESCO (United Nations Educational, Scientific, and Cultural Organization). Various years. Statistical Yearbook. Paris. World Bank. 1997. World Development Report 1997: The State in a Changing World. New York: Oxford University Press. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 385-418 Does Economic Analysis Improve the Quality of Foreign Assistance? Klaus Deininger, Lyn Squire, and Swati Basu The World Bank undertakes an annual expenditure of around $60 million on country- specific economic anal,ysis and advice for its member developing countries. What is the impact of this economic and sector work on the quality of World Bank lending? It would be useful to know whether past analytical work has generated measurable eco- nomic benefits that would justify its continued provision in an environment of increas- ingly scarce resources. This article sets out an idealized model of decisionmaking in which a country man- ager makes a broad allocation of resources between lending services and economic and sector work. Given that decision, the task manager for each project makes project- specific decisions with respect to the allocation of resources between preparation and supervision. The analysis indicates that economic and sector work has a significant positive impact on the quality of World Bank loans. The results provide clear evidence of underinvestment in economic and sector work. And the analysis shows that re- sources could be switched from preparation and supervision to economic and sector work to the benefit of both the quality of programs and level of disbursements. What is the benefit of the millions of dollars of foreign assistance provided to developing countries in the form of economic analysis and advice? This is a difficult question to answer. In the absence of a clear market test, there is no simple way to value the contribution of analytical work; most of it is provided free of charge or embedded in an investment or program loan. And the impact of economic analysis on some ultimate objective, such as development or poverty reduction, is almost imapossible to identify because it is only one of many factors that determine outcornes. Yet, it would be useful to know whether past analytical work has generated measurable economic benefits that would justify its continued provision in an environment of increasingly scarce resources. Moreover, some analysts claim that with improvements in the access of developing countries to nonconcessional sources of finance, the comparative advantage of international institutions such as the World Bank will shift toward the provision of analytical services (see, for Klaus Deininger and Lyn Squire are with the Development Economics Research Group at the World Bank, and Swati Basu is with the Faculty of Management at McGill University. The authors would like to thank Bill Battaile, Shanta Devarajan, David Dollar, Christopher Kilby, Lant Pritchett, and participants at a World Bank workshop on Aid Effectiveness for detailed comments on an earlier draft. Support from the World Bank's research support budget (RPo 681-26) is gratefully acknowledged. © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 385 386 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 example, Rodrik 1995). But little empirical evidence exists that could be used to substantiate such a claim or to provide guidance on how and where to focus such services in the future. The issue tackled in this article faces the same problems that confront efforts to evaluate research more generally. The literature on this topic (including this article) takes the view that spending on research is an investment that should produce measurable economic returns. The difficulty, of course, arises with the word measurable. Traditionally, economists have based their analysis on the production function approach or the consumer surplus approach (Averch 1994). Both approaches rely on market outcomes to provide measures of the benefits of research. Often, however, it is not possible to use market outcomes. Researchers frequently try to create more basic knowledge that, like all public goods, proves difficult to value. This happens to be true in our case. Economic studies and policy advice create and disseminate knowledge, the ultimate impact of which can be far-reaching but is usually difficult to trace. The analysis could focus more directly on the immediate outputs of research, meaning the reports generated or articles published (Wise and Agranoff 1991). For example, the National Science Foundation identifies a set of indicators that includes publications, citations, data archiving and sharing, software develop- ment and sharing, and education and pedagogy (National Science Foundation 1995). The World Bank has also used a variant of this approach, specifically, the number of World Bank publications appearing in university reading lists (World Bank 1991). Apart from the incentive problems that these measures may create, they fail to provide any measure of the value of research. In short, the available literature on the evaluation of research does not provide much guid- ance on how to value the output from economic analysis and advice and hence how to estimate the return to such activity. This article provides a partial answer to this question. It does not encompass all the benefits of economic analysis and advice. Foreign assistance provided in the form of analysis and advice attempts, among other things, to maintain a policy dialogue with the client country in order to provide analytical support for critical macroeconomic decisions, examine the steps needed to generate and maintain a policy and regulatory environment that promotes private sector ac- tivity, provide analytical inputs to the design of sectoral strategies, and build local capacity for policy analysis and discussion of key policy issues in the broader context of civil society. In addition, economic analysis and advice often provide a critical underpinning for the design of individual projects and lending pro- grams. This last aspect is the focus of this article. As such, our results will be a minimum estimate of the full benefits from this form of foreign assistance. To our knowledge, the only data set providing measures of resources spent on analytical services as well as measures of the benefits of lending is for the World Bank. Thus the purpose of the article is to assess the impact of the economic and sector work (ESW) provided by the World Bank on the quality of World Bank lending. Although data availability forces us to focus on the World Bank's ESW, Deininger, Squire, and Basu 387 this is not insignificant either in coverage or in quantity. ESW is undertaken on all developing countries that are members of the World Bank, numbering well over 100 countries (see World Bank 1997). The work undertaken on each country is specific to that country but could include the following: a country economic memorandum (a comprehensive account of economic performance and pros- pects) and more topic-focused reports such as poverty assessments, public ex- penditure reviews, labor market studies, as well as a wide range of sectoral studies. In general, the more aggregate studies underpin the World Bank's policy advice, while the sectoral studies (reviews of the transport sector, health and education sector reports, and so on) provide the foundation for lending opera- tions in those sectors. Undertaking this range and magnitude of work is expen- sive. Since 1975, the World Bank has used about 22,000 staff-weeks a year on ESW. This amounts to an annual expenditure of around $60 million on country- specific economic analysis and advice. This almost certainly makes the World Bank one of the largest sources of such analysis for the developing world. In this article, we address three main questions. First, does ESW improve the quality (variously defined) of World Bank loans? Second, has the World Bank underinvested in ESW relative to lending services? Third, is there a trade-off be- tween quality and volume of lending? We set out an idealized model of decisionmaking in which managers are as- sumed to be concerned exclusively with development impact. We assume that managers have sufficiently long planning horizons and sufficient information to be able to allocate resources broadly in line with their objective. Of course, in reality managers may be influenced by factors other than development impact, they may have relatively short time horizons, and the information relevant to successful decisionmaking at best may be only partial. The intention then is not to treat the idealized model as a description of reality, but to see how close reality approximates the ideal given all the factors that inevitably beset decisionmaking in the real world. Where we observe departures from this ideal, we try to explain the difference. Our model of decisionmaking operates at two levels. We envisage a country manager who makes a broad allocation of resources between lending services and ESW. Given that decision, the task manager for each project makes project- specific decisions with respect to the allocation of resources between prepara- tion and supervision. So at the country level, decisions are assumed to be made by a country manager who can allocate staff time freely between lending ser- vices and ESW. In this model, therefore, ESW is endogenous. The country manager is assumed to maximize development impact subject to the overall availability of staff resources. ESW contributes to development impact through its contribution to the quality of the countrywide lending program (defined as the total number of projects in a country approved in a given year) and through its contribution to policy formulation more generally. Testing this model allows us to address all three questions posed above. The model also yields (implicitly) the shadow price of a unit of staff resources. This price enters the decisionmaking problem of the 388 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 task manager responsible for a single project, the second level of decisionmaking that we examine. At this level, we assume that ESW is exogenous and that the manager's objective is focused more narrowly on maximizing the quality of the project. This model provides additional evidence on the first two questions posed above. Section I places the study in the context of other work on foreign assistance and highlights how our data constitute an improvement over those used in ear- lier studies. Section II discusses the conceptual framework and estimating equa- tions. Section III describes the data used in the study and provides descriptive statistics. Section IV presents our empirical results. Concluding remarks are pre- sented in section V. I. RELATION TO THE LITERATURE Research related to foreign aid falls into three broad categories. Game- theoretic models focus on the negotiation and implementation of conditionality and on the credibility and commitment problems that arise if the donor or the recipient behave strategically (Mosley, Harrigan, and Toye 1995 and Svensson 1996). Empirical models analyze how donors allocate aid across recipient coun- tries and how aid may affect the decisionmaking and economic performance of recipient countries (see Burnside and Dollar 1996; Pack and Pack 1990, 1993; Feyzioglu, Swaroop, and Zhu 1995; and Trumbull and Wall 1994). Descriptive or quantitative efforts analyze the factors that affect the performance of indi- vidual donor-financed operations and countrywide programs (see, for example, Claessens, Pohl, and Quian 1996, Arias 1994, Cassen and Associates 1994, Isham and Kaufman 1995, Kilby 1994, and Lele 1992). This article is most closely related to the third category. It attempts to identify the quantitative contribution of economic analysis to the quality of individual investment projects and the size of overall resource flows. Unfortunately, data availability limits the scope of the study to assistance provided by the World Bank. The impact of economic analysis on the success of World Bank projects has been investigated in two other studies with conflicting conclusions. Arias (1994) defines ESW as the number of World Bank reports on the country published in the three years preceding project approval. He fails to find a significant effect of ESW on project quality. This finding is based on an analysis of all projects that were evaluated between 1991 and 1994. Analyzing projects approved between 1991 and 1994, and defining ESW the same way, Schneider (1995) finds that ESW has a significant quality-enhancing impact. These contrasting results are a fur- ther motivation for the present exercise. Although the basic issues to be investigated are the same, this article differs from those studies. We try to cast the problem in terms of a more explicit decisionmaking framework. And both the quality and coverage of data are sig- nificantly expanded. Using data on input measured in staff-weeks of analytical services rather than the number of written reports allows us to reduce measure- Deininger, Squire, and Basu 389 ment errors that may have affected other studies. And relying on all available information in the World Bank's databases facilitates the inclusion of projects that were approved after 1985. II. CONCEPTUAL FRAMEWORK In this section we describe the processes governing the allocation of resources to different activities vvithin the World Bank in the form of two decisionmaking models. We derive the equations for estimation and specify empirically testable hypotheses. We start with the model of a country manager responsible for all activities in a given country and then turn to that of a task manager at the project level. Model 1: Country Manager We assume that the country manager's objective is to maximize a weighted function of lending program quality, Q, and policy impact, P, subject to the total staff time, T, available for ESW and all lending services-both investment lending, INV, and adjustment operations, ADJ. Formally, the country manager solves the following problem: Max L= U[Q (ESW,INV..);P(ESW,ADJ ..)]-X(ESW+INV+ADJ-T). The solution to this problem will yield demand functions for ESW, INV, and ADJ contingent on the overall budget constraint as well as various exogenous fac- tors, such as country-level policies and institutional factors. In principle, these demand functions, together with the exogenous constraints, could be used to solve for the optimum values of Q and P. Empirically, however, it is impossible to distinguish between the parameters of the production func- tions for lending quality and policy impact (Q and P) and the relative weights that the decisionmaker assigns to each of these outcomes. Nor do we have a variable to measure policy impact. To answer the questions posed earlier, there- fore, we estimate a reduced-form equation for Q*, the optimum value of lending quality. Using a linear approximation, we estimate the following equation. ( 1 ) Q*jt = go + g,ESWjt + g2INVjt + !3POL1t + Vjt where j and t denote country and year, respectively. We include variables to capture policies (POL) and may adjust for other unobservable country-specific factors by using dummies. While the observed value of ESW can be taken to be a reasonably accurate estimate of the decisionmaker's choice, this is not so for INV. Instead, what is available in our data is the amount of resources that was actually spent on INV (preparation and supervision). This amount may differ from the country manager's plans if emergencies or external events cause management to allocate additional funds to redesigning projects during preparation or to restructuring the project during implementation. Furthermore, because task managers have 390 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 no incentive to reveal negative information about individual projects that is not available to the country manager, or because of a general desire not to cancel projects, more time may be systematically allocated to bad projects in practice than was intended by the country manager. To deal with this, we instrument for INV using anticipated project age, past supervision, and past preparation require- ments as well as past project success in the country. Equation 1 allows us to examine all three questions of interest. First, does ESW improve the quality of World Bank loans? We expect that, if ESW contributes to the quality of the lending program, g > 0 (hypothesis 1). Second, has the World Bank underinvested in ESW relative to lending services? Resources allocated to ESW compete with resources for iNV, both of which should in principle improve project quality. According to our definition, the lending program consists of all the projects approved in a given year; this would imply some kind of intertemporal reallocation, that is, higher expenditures on current ESW would be rewarded by lower spending on INV in the future. ESW also contrib- utes to policy impact. Through its impact on economywide and sector-specific policies and institutions, economic and sector work generates benefits beyond the projects supported by the World Bank. Therefore, if staff resources are allo- cated to maximize the contribution to development, the marginal contribution of ESW to the quality of the lending program plus its marginal contribution to policy impact should equal that of INV. As long as the impact of ESW on policy is positive, it follows that the marginal contribution of ESW to project quality should be less than that of INV. That is, if the World Bank has systematically maximized development impact, we should expect g, < 92 (hypothesis 2). Third, is there a trade-off between the quality and volume of lending? If the idealized model is supported by the data, we can conclude that managers have not compromised quality to increase the volume of lending. If the model is not supported by the data, then such a trade-off may underlie the observed out- come. We take up this issue in section IV. Model 2: Task Manager We now turn to the decision facing a task manager of a single project. The task manager's decision differs from that of the country manager in three re- spects. First, Esw is taken as exogenous. That is, we can reasonably assume that decisions about ESW (macroeconomic analysis as well as sector-specific analysis) are made by senior management rather than by task managers associated with individual projects. Moreover, we assume that those decisions occur well before individuals are assigned to the preparation of specific projects. Second, the task manager of an investment project has little control over economywide policies. We assume, therefore, that the manager at the project level focuses exclusively on project quality. To the extent that the project design includes some changes in policy, we assume that the task manager's primary impact is on the quality of the project itself. And third, we divide INV into preparation time (PRP) and super- vision time (SPN), both of which are assumed to contribute to project quality. We Deininger, Squire, and Basu 391 treat PRP and SPN as substitutes because a well-prepared project reduces the need for subsequent supervision, and more supervision is required to rescue a badly prepared project. The task manager chooses the level of PRP and SPN to maximize the quality of the project subject to the given shadow price of staff time. In the event that this price is high, the project task manager is obliged to cut corners and economize on staff input. In the event that the price is low, the manager is able to prepare and supervise the project more thoroughly. In general, the task manager's optimization problem is: Max L = Q (PRP, SPN; ESW, POL, C, S ...) - X (PRP + SPN) where X is the shadow. value of time derived from the country manager's optimal solution, and C and S represent country-specific and sector-specific factors, re- spectively. The solution to this problem yields estimable demand functions for PRP and SPN in terms of the exogenous parameters. We use linear approxima- tions and estimate the following two equations where both PRP and SPN are nor- malized by loan size: (2) PRP-,t = o + P1ESWj, + P2PPRPj, + P3 PQUALjl + 04POLjt + Mi + Vjit and (3) SPN,t = yo + 7yES)wj, + Y2LOANjt + 73PQUAL1 + Y4PSPNjt + 7YPO¼L + YsSi + Vjit where in the preparation equation PPRP represents average past preparation re- quirements in the country, PQUAL is past quality of the country portfolio, and i denotes a project. In the supervision equation, LOAN is loan size, and PSPN iS past supervision effort in the country and sector. Equations 2 and 3 provide a partial test of the assumed decisionmaking model. The coefficients of greatest interest are ,1 and yl. If P, < 0, average preparation requirements are reduced by the availability of prior ESW, implying that the task manager responds to the availability of ESW in a rational fashion consistent with the assumed decisionmaking process. Similarly, finding that yi < 0 provides em- pirical confirmation that managers respond to the higher quality of projects at approval (because of more prior ESW) by reducing the allocation of resources to supervision. We can also use these equations to test for several other possibili- ties. In particular, if only Po0 0 oryo ;* 0 and P,R yi = 0 for all i, then preparation and supervision time are allocated merely on the basis of loan size. If only P2 (or y4) are different from 0, preparation (supervision) are determined solely on the basis of past experience. We expect that unfavorable policies complicate both project preparation and implementation, that is, f4 < 0; ys < 0 (for policies where higher values indicate "better" policies). Similarly, higher past requirements for preparation and supervision at the country and sector level indicate the absence of institutional and administrative infrastructure, thus increasing the need for staff input (p2 > 0). I-ligher quality of the past country portfolio is expected to be associated with lower preparation requirements because existing infrastructure and talent can be drawn on (ft3 < 0; y3 < 0). 392 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 If the model posited is appropriate, lending program quality, Q, will depend on PRP, SPN, and ESW according to the following equation (where we again use a linear approximation): (4) Qiit = go + g,ESWjt + gt2PRPit + ,3SPNjt + N4POLt + gSSi + V1t. Estimating this equation using the predicted values of PRP and SPN from equa- tions 2 and 3 allows us to assess the marginal impact of ESW on project quality (hypothesis 1). In particular, if g, > 0, then we can conclude that ESW does im- prove the quality of projects. Indeed, using a quantitative indicator for project quality that can be expressed in dollar terms, we can measure the benefits of additional ESW and compare them with the nominal cost of staff input. In addi- tion, we can use project-level data to investigate whether reallocating resources between ESW and preparation or supervision would improve project quality. Recall that because ESW makes a contribution to nonproject objectives, we expect 1 < 2 and gi < R3 (hypothesis 2). If this hypothesis is rejected, we have further evidence that the country manager failed to achieve the appropriate allocation of resources across the three activities from the standpoint of development impact. III. MEASUREMENT ISSUES AND DESCRIPTIVE STATISTICS Before discussing our empirical results, in this section we describe data defini- tions and sources and present descriptive statistics. We discuss the variables used to measure lending services and ESW, project success, and policies. Then we dis- cuss bivariate relationships between the variables that will be of importance in the subsequent discussion. Lending Services and ESW Lending services are provided by the World Bank through projects that are associated with loans. In our analysis we consider a variety of project attributes such as the size of loan, the amount disbursed, and the amount of resources spent to prepare and supervise the project. We consider only loans that are ei- ther active or completed, thus discarding projects (about 10 percent of the total) that were dropped before becoming active. The amount of staff time (including long-term consultants) spent in preparation and supervision by project is avail- able from the World Bank's management information system for all years fol- lowing 1975. The variables for time spent in preparation and supervision are based on entries from the World Bank's time recording system, which classifies staff time devoted to specific activities by project and by task code. The only task codes encountered for lending projects are preparation, supervision, and preparation of the implementation completion report (which we have added to the time for supervision). Unless otherwise noted, preparation and supervision variables in all the regressions are normalized by loan size. That is, the short- hand terms preparation and supervision refer to staff-weeks of preparation or supervision time per $1 million in loans. Deininger, Squire, and Basu 393 ESW is defined as any use of resources by the World Bank (excluding research and other activities that are neither country-specific nor project-specific) that is not linked to the preparation or supervision of loans.' The main activities in- cluded in this category are macroeconomic and sector reports, public expendi- ture reviews, preparation of country assistance strategies, and donor coordina- tion. We distinguish between general macroeconomic or multisectoral analysis and sector-specific ESW, with the latter disaggregated into seven sectors-agri- culture and environment, education, health, finance, industry and energy, trans- port infrastructure and water, and public sector restructuring.2 The variable for ESW is measured in weeks of staff time and is available on a consistent basis beginning in 1985. Because the variable for staff cost that is available in the World Bank database does not add any information (its correlation with staff- weeks is 0.82) and is available for significantly fewer observations, we measure ESW in staff-weeks throughout. We link ESW, defined at the country or sector level, to specific lending projects by assuming that a project is affected by a simple average of the country and sector work undertaken during the four fiscal years preceding its approval by the World Bank's Board of Executive Directors. To account for the fact that macroeconomic work is likely to affect all projects in a given country while sector-specific ESW has a more limited impact, the ESW variable is defined as the sum of all macroeconomic ESW for the country in which the project is imple- mented and all sector-specific ESW for the sector in which the project falls using the seven sectors mentioned. Table 1 illustrates that since 1975 the World Bank has on average spent about 53,000 staff-weeks (equivalent to about 1,260 staff-years) annually on supervi- sion and preparation of loans, compared with about 22,000 staff-weeks invested in country-specific rmacroeconomic analysis (40 percent) or sector-related ESW (60 percent). Over time there have been considerable increases in the amount of resources devoted to supervision and sector-specific analysis, with much more modest increases in r esources devoted to preparation of loans and to macroeco- nomic work. While total processing costs associated with lending stayed more or less constant at 3.2 staff-weeks per $1 million in loans in 1977-92, the com- position changed considerably in favor of more supervision and less prepara- tion; the share of time spent in project preparation dropped from 68 to 58 per- cent. The share of ESW in total staff-weeks spent was consistently below 30 percent, and, even though a significant increase in spending on ESW occurred between 1989-92 and 1993-95, in absolute terms the increase was less than the expan- 1. Each year the World Bank spends about 26,000 staff-weeks on activities, mainly research and a few administrative functions, that are neither project-specific nor country-specific and that are therefore not included in our analysis. 2. Because the sectoral classification in the World Bank's management information system is quite poor (many sector-specific pieces are classified as general), we have reclassified the general category based on the task title. Fo:r example, an education sector review is placed into the education rather than the general category. Table 1. Evolution of World Bank Lending and Nonlending Services, 1975-95 (staff-weeks unless otherwise noted) Total lending Nonlending Annual Staff-weeks per Staff-weeks per commitments $1 million Sector- $1 million Period (1987 dollars) Preparation Supervision in loans Macroeconomic related in loans 1975-76 12.14 24,265 11,326 2.93 - - 'IO 1977-80 14.28 29,267 16,443 3.20 - - 48 1981-84 16.03 28,657 21,359 3.12 - - - 1985-88 17.32 31,819 22,783 3.15 6,917 8,604 0.90 1989-92 19.09 36,039 25,172 3.21 7,152 11,819 0.99 1993-95 18.26 36,702 34,621 3.91 9,310 17,225 1.45 Total 16.47 31,512 22,359 3.27 9,143 13,336 1.36 - Not available. Note: Values are annual averages. Source: Authors' calculations based on World Bank data. Deininger, Squire, and Basu 395 sion of spending on lending during the same period (0.46 compared with 0.7 staff-week per $1 million in loans), implying that in relative terms the impor- tance of ESW actually decreased slightly. A breakdown by sector for 1985-90 and 1991-95 highlights that the relative constancy of real commitments in the aggregate hides important intersectoral variations (table 2). Regarding the volume of lending, considerably reduced lev- els of commitments in industry and energy and in agriculture (down $0.94 bil- lion and $0.71 billion a year, respectively) were more than compensated by in- creases in lending for education, health, and public sector reform (up $0.92 billion, $0.67 billion, and $0.54 billion, respectively). The ranking of sectors was little affected; the largest volume of lending was for infrastructure, followed by industry and energy and by agriculture and environment. The relative impor- tance of education and health more than doubled. In 1985-90, direct lending costs, in staff-weeks per $1 million in loans, were above average in health (7.3), education (4.5), agriculture (4.4), and public sec- tor reform (3.9). They were below average in infrastructure (2.9), industry (2.3), and financial sector operations (2.0). Turning to intertemporal changes, the cost of lending (normalizecl by volume of total lending) decreased in health and edu- cation. It increased considerably in agriculture, where, in contrast to industry and energy, the decrease in overall commitments was not matched by a con- comitant reduction in staff-weeks. While ESW increased in every sector, the in- crease was, in quantitative terms, largest in infrastructure, where the amount of resources used for such services almost tripled. In 1991-95, the share of total sectoral resources (per $1 million in loans) spent on ESW ranged from a high of 28 percent in industry and energy to a low of about 16 percent in health and agriculture (divide column 4 by [column 4 + column 2] in table 2). Table 3 shows the regional disaggregation of lending and nonlending ser- vices. The largest increase in commitments was registered by Eastern and Cen- tral Europe followed by East Asia and the Pacific (up $1.44 billion and $0.96 billion per year, respectively). By contrast, annual real lending to South Asia dropped by almost $1.2 billion, with smaller decreases in Latin America and Africa. The share of ESW in total use of staff time during 1991-95 was about 20 percent for almost all, regions except Latin America (15 percent). This implies that staff time spent directly in both lending services and ESW varied widely across countries from a total of more than 10 staff-weeks per $1 million in loans in Africa to less than 3 staff-weeks in East Asia. Comparatively high costs of lend- ing were also observed in South Asia (a sharp increase from the 1985-90 period) and in the Middle East and North Africa. Measures of Project Success Direct indicators regarding the success of individual lending projects are avail- able from ex post evaluations conducted by the World Bank's Operations Evalu- ation Department (O0ED) and from the Annual Review of Portfolio Performance (ARPP). While the former are preferred insofar as they provide an assessment of Table 2. World Bank Resources Spent on Lending and Nonlending Services by Sector, 1985-90 and 1991-95 Lending services Nonlending services Annual commitments Staff-weeks per Staff-weeks per in billions of $1 million $1 million 1987 dollars Staff-weeks in loans Staff-weeks in loans Sector (1) (2) (3) (4) (5) 1985-90 Agriculture and environment 3.81 16,853 4.42 2,535 0.67 Public sector reform 0.25 972 3.89 255 1.02 Education 0.92 4,092 4.45 964 1.05 Health 0.39 2,815 7.28 664 1.72 Industry and energy 4.37 10,230 2.34 2,828 0.65 Infrastructure 4.22 12,083 2.86 1,170 0.28 Finance 2.08 4,052 1.95 1,045 0.50 Total 16.04 51,097 3.19 9,460 0.59 1991-95 Agriculture and environment 3.10 18,694 6.02 3,688 1.19 Public sector reform 0.79 3,282 4.14 696 0.88 Education 1.84 7,488 4.07 1,792 0.97 Health 1.06 6,543 6.19 1,220 1.15 Industry and energy 3.43 9,576 2.79 3,728 1.09 Infrastructure 4.39 14,491 3.30 3,117 0.71 Finance 1.48 3,675 2.48 1,080 0.73 Total 16.09 63,749 3.96 15,320 0.95 Source: Authors' calculations based on World Bank data. Table 3. World Bank Resources Spent on Lending and Nonlending Services by Region, 1985-90 and 1991-95 Lending services Nonlending services Annual commitments Staff-weeks per Staff-weeks per (billions of $1 million $1 million Region 1987 dollars) Staff-weeks in loans Staff-weeks in loans 1985-90 Africa 2.81 18,519 6.58 3,270 1.16 East Asia and the Pacific 3.60 8,785 2.44 1,982 0.55 Middle East and North Africa 1.25 5,103 4.09 880 0.71 w Latin America and the Caribbean 4.99 10,007 2.00 1,542 0.31 N3 Eastern and Central Europe 1.54 3,623 2.35 471 0.31 South Asia 3.77 9,261 2.45 1,314 0.35 Total 17.97 55,298 3.08 9,460 0.53 1991-95 Africa 2.55 21,152 8.29 4,975 1.95 East Asia and the Pacific 4.56 10,531 2.31 2,508 0.55 Middle East and North Africa 1.25 4,864 3.88 1,260 1.01 Latin America and the Caribbean 4.60 12,322 2.68 2,322 0.51 Eastern and Central Europe 2.98 9,026 3.03 1,983 0.67 South Asia 2.58 11,184 4.33 2,272 0.88 Total 18.52 69,079 3.73 15,320 0.83 Source: Authors' calculations based on World Bank data. Table 4. Indicators of the Success of World Bank Projects by Sector and Region, 1975-95 (mean values) OED indicator ARPP indicator Economic rate of Project Institutional Project Development Implementation Sector or region return at completion outcomea development sustainability objectives progress Sector All 16.81 71.71 2.69 2.40 3.39 3.10 Agriculture and environment 13.83 62.52 2.44 2.34 3.25 2.98 Public sector reform - 48.00 2.38 2.16 3.21 3.07 Education 82.33 3.19 2.53 3.44 3.17 Health - 60.32 2.59 2.22 3.25 3.01 Industry and energy 15.45 74.69 2.87 2.50 3.49 3.16 Infrastructure 20.88 77.97 2.78 2.38 3.50 3.14 Finance 17.14 75.08 2.76 2.54 3.42 3.13 Region All 16.81 71.71 2.69 2.40 3.39 3.10 Africa 14.73 61.82 2.32 2.18 3.25 3.04 East Asia and the Pacific 18.95 82.83 3.22 2.79 3.56 3.30 Middle East and North Africa 17.82 77.61 2.94 2.57 3.47 3.10 Latin America and the Caribbean 15.72 68.86 2.67 2.37 3.40 3.07 Eastern and Central Europe 16.95 75.49 2.94 2.60 3.36 3.04 South Asia 18.97 76.98 2.69 2.29 3.42 3.07 Number of observations 1,959 3,956 2,057 1,992 5,004 5,064 - Not available. Note: OED indicators are from the World Bank's Operations Evaluation Department. ARPP indicators are from the World Bank's Annual Review of Portfolio Performance. a. The share of projects that are rated "satisfactory." Source: Authors' calculations based on World Bank data. Deininger, Squire, and Basu 399 the overall performance of a completed project, the greater availability of data for ARPP indicators leads us to use them as an alternative variable to check the robustness of our results (although the corresponding regressions are not reported). OED indicators, which are established by OED staff, include an estimate of the economic rate of return at completion, a binary measure (satisfactory and unsat- isfactory) of the overall project outcome, an ordinal assessment (four levels from highly satisfactory to highly unsatisfactory) of the project's sustainability, and its contribution to institutional development in the client country. Although these indicators provide a succinct summary of overall project success, their disadvan- tage is that only completed projects undergo this evaluation. Because informa- tion on resources devoted to nonlending services is available on a consistent basis only after 1985, the number of projects for which both OED evaluations and data on nonlending services are available is substantially reduced. How- ever, this drawback rnay be more than compensated for by the fact that the evaluation is undertaken by an independent unit, presumably with some mini- mum consistency across projects. The ARPP rates projects annually according to several criteria relating to over- all project performance, that is, contribution to development objectives and imple- mentation progress. It: also rates projects on more specific intermediate charac- teristics, such as compliance with legal covenants, project management performance, availability of funds, procurement progress, training progress, tech- nical assistance progress, studies progress, environmental aspects, fiscal perfor- mance, and gender aspects. Not all of these criteria and characteristics are rated regularly, resulting in large variations in the number of observations available for different indicators. The analytical rigor and informational content of these measures are likely to be inferior to more precise indicators such as the eco- nomic rate of return. IHowever, the much increased coverage partly compensates for this shortcoming, with a total of more than 5,000 projects having been in- cluded in the ARPP at least once. At the project level, the advantage of greater data availability is less impressive than might appear at first sight. Although using the mean value of all available ratings for any given project to construct a project-specific perfo:rmance indicator increases the number of observations, it lumps together projects at very different stages of the project life cycle, whereas use of the final rating (for completed projects) considerably reduces the number of observations. Differences in average values for selected indicators of project success are illustrated in table 4. Data on economic returns after completion are available for about 2,000 projects in four sectors (agriculture and environment, industry and energy, infrastructure, and finance). Average ex post rates of return vary between 13.8 perceni: in agriculture and 20.9 percent in infrastructure and be- tween 14.7 percent in Africa and 19.0 percent in South and East Asia. Ex post ratings concerning contribution to institutional development and sustainability are highest in infrastructure and education projects. Contemporaneous (ARPP) Table 5. Simple Correlations between Different Indicators of Success of World Bank Projects, 1975-95 OED indicator ARPP indicator Project Economic rate of Project Institutional Development Implementation Sector or region outcome3 return at completion sustainability development objectives progress OED indicator Project outcomea 1 [3,957] Economic rate of return 0.3807*** 1 at completion [1,3761 [1,3761 Project sustainability 0.6381*** 0.3107*** 1 [2,024] [842] [2,031] 4a Institutional development 0.4829*** 0.1477*** 0.5084 1 [1,956] [799] [1,9511 [1,968] ARPP indicator Development objectives 0.2035*** 0.1522*** 0.1791*** 0.1355*** 1 [2,885] [1,371] [2,024] [1,960] [4,806] Implementation progress 0.2134*** 0.1741*** 0.2447** 0.2184*** 0.5719*** 1 [2,928] [1,375] [2,028] [1,965] [4,805] [4,855] * * Significant at 5 percent. *** Significant at 1 percent. Note: OED indicators are from the World Bank's Operations Evaluation Department. ARPP indicators are from the World Bank's Annual Review of Portfolio Performance. Values in square brackets are the number of observations. a. The share of projects that are rated "satisfactory." Source: Authors' calculations based on World Bank data. Deininger, Squire, and Basu 401 ratings, although characterized by limited variation across projects, are lowest in agriculture and in Africa and are consistently high in East Asia. Table 5 pro- vides simple correlations between these indicators, illustrating that ARPP indica- tors for achievement of development objectives and implementation progress, OED indicators for project outcome, sustainability, and institutional develop- ment, as well as the e-x post rate of return are significantly correlated with one another. Policy Variables The literature has lDng emphasized that policies are important factors influ- encing project implementation as well as eventual development impact. It is there- fore important to cont-rol for appropriate policy variables. Policies that have an important bearing on project success relate to fiscal policy, monetary policy, and external trade policy. These can be approximated by the public sector defi- cit, domestic inflation or the black market exchange rate, and openness, as in Burnside and Dollar (1L996). To avoid problems of endogeneity, we use the value of policy variables at project initiation throughout. Note that policy variables are generally available only up to 1992, causing us to omit them in the instru- mental equations for supervision. Preliminary Statistical Analysis Investigation of the bivariate relationship between OED success indicators for the lending program and our independent variables (table 6) leads to three broad conclusions. First, the correlation between ESW (both macroeconomic and sector-specific) and development impact is significant and positive except in the case of project outcome. Second, higher levels of resources spent on supervision and preparation are negatively, rather than positively, related to project success. Such a phenomenon could arise from selection bias whereby bad projects are allocated more resources in order to improve their performance. Alternatively, resources spent in either preparation or supervision may be correlated with an- other omitted variable (it could be, for example, an Africa dummy) that is asso- ciated with lower project performance. And third, concerning the impact of policy variables, we find thalt openness is consistently associated with improved project success, that sustainability and contribution to institutional development are better in countries with low public sector deficits, that inflation negatively af- fects project outcome and sustainability, and that projects appear to perform better in all dimensions except economic rate of return in countries that invest a higher proportion of their own resources. Table 7 uses descriptive statistics to summarize our hypotheses and to pro- vide empirical support for the conjecture of a positive link between ESW and project success. Splitting the sample of completed projects into groups based on the overall project outcome, the variable used in our analysis, we find that the satisfactory group of projects benefited from considerably higher levels of ESW but required significantly lower levels of preparation and supervision (per $1 402 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 6. Correlations between World Bank Program Level OED Indicators and Economic Sector Work as well as Policy Variables, 1975-95 OED indicator Economic rate of return Project Project Institutional Policy variable at completion outcomea sustainability development Analytical services (ESW) Total 0.1538*** 0.0715 0.1331** 0.1237*** [305] [3471 [601] [595] Macro only 0.13368** 0.05633 0.1269*** 0.1102*** [305] [347] [6011 [595] Sectoral only 0.1525*** 0.07473 0.1432*** 0.1215*** [305] [347] [601] [5951 Preparation -0.1139*** -0.0507* -0.0499* -0.0214 [1,299] [1,080] [1,500] [1,474] Supervision -0.1218*** -0.0751** -0.0903*** -0.0309 [1,299] [1,080] [1,500] [1,474] Openness 0.0495* 0.0914*** 0.1303*** 0.0978*** [1,225] [1,008] [1,391] [1,368] Public sector surplus 0.0361 0.0488 0.0993*** 0.0830** [706] [587] [894] [8811 Inflation -0.0438 -0.0955*** -0.0575** -0.0241 [1,095] [9271 [1,287] [1,269] Investment -0.01121 0.1174*** 0.1999*** 0.1554*** [1,225] [1,008] [1,391] [1,368] * Significant at 10 percent. * Significant at 5 percent. * * * Significant at 1 percent. Note: The lower number of observations is due to the use of portfolio-level data. Values in square brackets are the number of observations. Policy variables are from the Summers/Heston data set (openness and investment), the World Development Indicators (inflation), and William Easterly (pubic sector surplus and black market premium). a. The share of projects that are rated 'satisfactory." Source: Authors' calculations based on World Bank data. million in loans) than the rest. Key policies for the satisfactory projects were more favorable, with more open economies, lower public sector deficits, lower inflation and black market premia, and higher levels of investment than for the unsatisfactory group. However, the fact that unsatisfactory projects required 20 percent more preparation and 22 percent more supervision suggests that the ex ante design of a project, an activity to which ESW presumably contributes, is crucial. We obtained similar results (not reported) when we split projects into top and bottom groups for the other indicators discussed earlier. Although simi- lar descriptive results hold if ARPP indicators are used, we refrain from using them because the only advantage of ARPP over OED ratings-namely, the inclu- sion of projects that are not yet completed-is outweighed by the difficulty of appropriately aggregating ratings and by the fact that they may be subjective and prone to changes due to shifts in personnel. Deininger, Squire, and Basu 403 Table 7. Input and Po/icy Indicators for Satisfactory and Unsatisfactory World Bank Projects, 1975-95 Satisfactory Unsatisfactory Indicator outcome outcome Economic sector work (staff-weeks) 88.32 74.33 Preparation of loans (staff-weeks per $1 million in loans) 4.88 5.84 Supervision of loans (staff-weeks per $1 million in loans) 4.75 5.82 Openness 54.11 48.53 Public sector surplus -6.09 -6.57 Inflation 23.56 49.69 Black market premium 28.37 35.16 Investment 14.97 12.71 Note: Policy variables are from the Summers/Heston data set (openness and investment), the World Development Indicators (inflation), and William Easterly (public sector surplus and black market premium). Source: Authors' calculations based on World Bank data. IV. EMPIRICAL ANALYSIS OF THE IMPACT OF ESW In this section, we report the results of the reduced-form regressions discussed in section II. We examine whether, and if so by how much, ESW improves project quality. We evaluate whether the allocation of staff resources across three main uses-ESW, preparation, and supervision-has been consistent with the maximi- zation of development impact. And we analyze whether the allocation of staff resources to ESW reduces the volume of lending, presumably by reducing the amount of staff time available for the more immediate tasks of project processing. Estimating the Country Manager Model We first estimate the reduced-form equation 1. Because we are interested in the overall program outcome, we use the mean satisfactory/unsatisfactory rat- ing for all projects (weighted by loan size) as the dependent variable. (The re- sults do not change much if the unweighted average outcome of the lending program is used.) Where included, investment lending is instrumented using past preparation and supervision requirements in the country, past project success, and the mean project age in a given country and year. The results in the first column in table 8 suggest that ESW has a significant positive impact on the quality of the lending program, thus supporting hypoth- esis 1. This effect remains significant if policy variables are added to the regres- sion. Because our measure for ESW includes the years before project approval while policies here are measured at project approval, there is a potential endogeneity problem. Repeating the analysis with policies in earlier years re- duces the number of observations but does not alter the substantive conclusions (results not reported). Thus the use of better data provides empirical evidence in favor of a signifi- cant and positive impact of ESW on quality, in contrast with the findings of Arias (1994). Regarding the quantitative impact of ESW, we find that an additional 100 staff-weeks would increase the probability of a program being satisfactory 404 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 8. Determinants of Project Outcome and the Economic Rate of Return for World Bank Loans, 1975-95 Project outcomeA Economic rate of return Variable 1 2 3 4 5 6 ESW 0.12104** 0.1760** 0.1773* 1.933*** 2.240** 2.1605* (0.0529) (0.0753) (0.1008) -0.7132 (0.9319) (1.243) Openness 0.36271 0.3778 -2.262 0.6309 (0.3394) (0.4178) (4.647) (4.778) Inflation -0.2451 -0.3372** -0.0778 -0.7225 (0.1403) (0.01717) (0.5883) (0.5187) Public sector surplus 0.8937 0.8037 -2.917 -37.215 (1.S38) (1.707) (2.964) (0.31215) Investment lendingb 0.0054 0.0709 (0.0221) (0.2227) Log-likelihood/R2 -224.99 -136.52 -98.96 0.02 0.05 0.12 Number of observations 347 215 157 305 196 164 * Significant at 10 percent. Significant at 5 percent. * * Significant at 1 percent. Note: Policy variables are from the Summers/Heston data set (openness and investment), the World Development Indicators (inflation), and William Easterly (public sector surplus and black market premium). Standard errors are in parentheses. a. The mean satisfactory/unsatisfactory rating for all projects, weighted by loan size. b. Staff-weeks spent on supervision and preparation for investment lending. Source: Authors' calculations. by between 12 and 20 percent. While other resources spent on lending-super- vision and preparation-are insignificant, the point estimate for ESW is positive (the third column in table 8).3 And although the policy variables have the antici- pated sign (and can be shown to be significant in regressions when ESW is ex- cluded), they are not significant. Using other OED variables to represent project quality suggests that the results obtained are robust (not reported). While policy variables are rarely significant (only public sector surplus in the sustainability equation), ESW is highly significant throughout. The regressions using the economic rate of return (ERR) as the dependent vari- able have the added advantage of facilitating a simple cost-benefit analysis of assigning more resources to ESW. However, several problems arise with using ERRs. ERRS are only available for certain sectors (see table 4), and even within these sectors ERRS have not been computed consistently for all projects. This implies that whatever estimate is available is unlikely to be representative for the entire lending program, and problems of sample selection bias may be present. Also, aggregation of ERRS over projects (even if the latter are weighted by loan size) is at best a rough approximation of the true net present value (NPV) of the lending program. And the number of observations is quite limited (305 and 164 without and with policy variables, respectively). 3. The point estimate for INV is negative, though insignificant, in a specification that excludes policies (not reported). This suggests that, with the instruments available, we were not able to resolve fully the endogeneity problem associated with project preparation and supervision. Deininger, Squire, and Basu 405 These problems prevent us from reading too much into the figures obtained in the respective regressions. However, ESW is significant and positive even if policy variables are included (table 8). Introducing regional dummies (not re- ported) does not eliminate the significance of the coefficient on ESW. In fact, using the point estimate on the ESW variable in the regressions, we find that an increase of one staff-week in ESW would increase the NPV of the average lending program ($210 million) by between $36,000 and $44,000, similar to the higher end of the range of values ($18,000-$36,000) derived from the project-level regression reported here.4 Compared with the average cost of one staff-week (about $3,000) as reported in World Bank data, this suggests that an expansion in ESW would constitute an economically sound investment. When interpreting chis result, it is important to keep in mind the distinction between an improvement in the NPV for projects supported by the World Bank and an improvement in the NPV for the government's investment program. If ESW enables World Bank staff to identify and support new investment possibilities, thereby expanding the range of investment options open to a country, or if it leads to an improvement in the ex ante design of a project already in the invest- ment program and hence its ultimate development impact, then ESW yields a genuine return to the recipient country. Alternatively, conducting more ESW may enable World Bank staff to cherry-pick better projects from the existing portfo- lio, projects that would have been undertaken by the country anyway. In this event, ESW improves the performance of the World Bank's lending portfolio but not necessarily the quality of the public investment portfolio in general. How- ever, this explanation is not convincing. In most cases, countries are not in the position of having a well-defined investment program from which World Bank staff can pick and choose. World Bank staff are usually heavily engaged in identifying and developing new investment opportunities. And in most cases, projects are not close to being fully designed before World Bank staff become involved. In fact, World Bank staff usually play a major role in project design and conceptualization. A more formal test of these alternative explanations is possible. If the cherry- picking explanation underlies the observed impact of ESW on the NPV of the World Bank-supported lending program, it would be expected that World Bank portfolios supported by significant inputs of ESW would comprise exclusively "safe" projects. Cherry-picking implies that few, if any, projects would be dropped either during preparation or after approval. Thus, the cherry-picking hypothesis leads us to expect an inverse relationship between the amount of ESW and the 4. The transformation f rom ERR to NPV depends on the phasing of costs and benefits. Following Kilby (1995), the calculation reported in the text assumes that project costs occur in the first period and benefits in the second. With this assumption and an average lending program of $210 million, the increase in NPV lies between ($0.00019 x 210,000,000) / 1.1 and ($0.00023 x 210,000,000) /1.1, where the discount rate is assumed to be 10 percent. A more realistic assumption regarding the phasing of costs and benefits would tend to increase the calculated NPV. We therefore treat the value reported in the text as a conservative estirnate. 406 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 share of the program dropped. No such relationship is found in the data. Con- trolling for policies and the level of gross domestic product per capita, there is no statistically significant relationship between the share of the lending program dropped prior to the loan becoming effective and the amount of ESW, nor be- tween the share of the portfolio canceled after approval and the amount of ESW (results not reported). This result suggests that, rather than enabling task man-- agers to pick the cherries out of a limited portfolio of investments, ESW contrib- utes by either expanding the available investment opportunities or improving the ex ante design of the available projects. Both avenues result in a genuine increase in the development impact of foreign assistance. Regarding a possible trade-off between ESW and investment lending (INV) as stated in hypothesis 2, we find that, contrary to expectation, the marginal con- tribution of ESW exceeds that of INV. Indeed, the marginal contribution of INV, while positive, is not significantly different from 0 (the third and sixth columns in table 8). This suggests that it would have been possible to shift resources from INV to ESW and increase program quality. And because ESW also (we assume) contributes to policy impact, we conclude that the observed allocation of staff time between these two uses has not been optimal from the point of view of maximizing development impact. In other words, our ideal model of decision- making at the level of the country manager may not be a good approximation of reality. Of course, the representative manager may not have been concerned exclu- sively with development impact as we have assumed. Other objectives are clearly possible. For example, managers may have given weight to resource flows or lending volume. This could reflect a concern with the overall level of financial transfer (a development concern) or, if the volume of commitments is the basis for rewards within the institution, a career concern. Claessens, Pohl, and Quian (1996) find a significant impact of task managers on project quality but note that the quality of task managers is not correlated with promotion decisions within the World Bank. Either way, the objective function outlined in section II needs to be expanded to include a third element-lending volume-in addition to the quality of lending and policy impact. Corresponding to equation 1, we now have an additional reduced-form equation for the optimal lending volume V* consistent with our revised objective function for the country manager. Again assuming a linear approximation, the new equation to be estimated is: (5) V# = Xo + X1ESWI, + X2INV1t + X3POLit + X4ADJjt + Vjt where, as before, ESW represents time spent on economic and sector work, INV is staff time devoted to lending services excluding adjustment operations, POL rep- resents policy variables, and ADJ is the staff time for adjustment operations. Fo- cusing on the difference between X2 and Xl, we can examine whether a realloca- tion of staff resources in favor of ESW would have a negative impact on lending volume (hypothesis 3). If X2 < k1, managers would have been operating within the frontier-that is, they could have increased both quality and volume of the Deininger, Squire, and Basu 407 lending program by shifting resources out of investment and into ESW. If k2 > 1, reallocating staff from lending services to ESW, regardless of its impact on qual- ity, would have come at the cost of decreased lending volume. This would imply a trade-off between quality and quantity, indicating that managers did indeed place some value on the volume of lending as well as the quality. We use as a depenclent variable both the volume of commitments and the volume of actual disbursements for a given lending program. The former is the better variable if the representative manager believed that approval of a large lending program by the World Bank's Board of Executive Directors was a key factor in promotions and career development. The latter is more appropriate if the manager was interested in achieving a large transfer of resources. Our results (the first column in table 9) support the hypothesis of a significant difference between the effect of ESW and INV on the volume of commitments. INV appears to be between 40 and 50 percent more effective in increasing commit- ments than is ESW. If the observed relationship is indeed the outcome of an opti- mization process on the part of the decisionmaker, we can assess quantitatively the trade-off between volume and quality. Ignoring for the moment the possibil- ity that ESW has a positive policy impact, at the optimum the decisionmaker will be indifferent to a marginal shift of staff time between INV and ESW. Using the estimated results for equations 1 and 5, we calculate that the average manager seems to be indifferent between a 1 percent decrease in the program's ERR (or a $2 million decrease in the program's NPV) and an additional $4 million in com- mitments (compared wvith arn average lending program of $210 million). This result suggests the presence of a strong trade-off. That is, World Bank managers are willing to accept a moderately large decline in the quality of the lending program in return for only a modest increase in the level of commitments. Of course, when making a decision about this trade-off, the country manager should Table 9. Determinants of World Bank Lending Volume, 1975-95 Variable Commitments Disbursements ESW 0.1899K** 0.2118*** (0.039) (0.04465) Investment lending, 0.2709* '*i 0.0727* * (0.029) (0.0314) Adjustment lendinga 0.1401 ** 0.1432* ** (0.0338) (0.0324) Adjusted R2 0.9090 0.7248 Number of observations 406 302 * Significant at 5 percent. Significant at 1 percent. Note: The dependent variable, lending volume, is commitments or disbursements in millions of dollars. Coefficients of policy variables and country dummies are included but not reported. Standard errors are in parentheses. a. Staff-weeks spent on preparation and supervision. Source: Authors' calculations. 408 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 also take account of the contribution of ESW to policy, an impact that is addi- tional to its immediate effect on the quality of the lending program. If it is true that the contribution of ESW to policy impact is quantitatively important, then our unfavorable interpretation of managerial behavior is strengthened. The picture with respect to disbursements is very different. In table 9 the evidence suggests that managers were operating within their frontier. That is, resources could have been switched from INV to ESW to the benefit of program quality and the volume of disbursements. In this case, managers could have achieved a 1 percentage point increase in the program's ERR (an increase in the NPV of about $2 million) and increased disbursements by almost $7 million. The different outcomes are consistent with the view that managers did give weight to the level of commitments, even though this reduced the quality of the lending program and reduced the level of disbursements. Estimating the Task Manager Model We turn now to the decisionmaking problem faced by the task manager of a single project. In this case ESW is treated as an exogenous variable, and the manager is assumed to maximize project quality given the (shadow) price of staff inputs. PROJECT PREPARATION. As discussed in section II, the assumed decisionmaking process yields demand functions for preparation and supervision contingent on existing ESW. According to the model, a higher level of ESW is expected to reduce the need for preparation resources because, as a result of ESW, the basic project rationale and design will be on a much firmer footing. Using total staff time spent on preparation (normalized by loan size) as the dependent variable and estimating equation 2 confirms this notion. In particular, we find that the representative manager clearly and significantly reduces the amount of resources used to prepare any given project as ESW increases (table 10). This result is consistent with the view that managers see a benefit from ESW and adjust their resource allocations accordingly. Regarding the magnitude of this impact, we find that, at the mean level of ESW across all projects (136 staff-weeks), one additional staff-week of ESW causes the task manager to reduce the time allocated to preparation by 0.015 staff-week per $1 million in loans or 1.05 staff-weeks for the mean $70 million loan (the third column in table 10). We also note that at empirically observed levels of input, the impact of ESW on preparation is concave, that is, it is subject to slowly decreasing returns (the first three columns in table 10). Returns from ESW start declining at almost 1,000 staff-weeks on average per project, a value that is well beyond the mean (136 staff-weeks). Because we are dealing with individual projects, we add dummies to control for unobservable sector-specific or country-specific characteristics. Their inclusion leaves the statistical results unchanged. PROJECT SUPERVISION. According to our decisionmaking model, by improving the design of projects, ESW should enable managers to reduce the amount of time Deininger, Squire, and Basu 409 Table 10. Determinants of the Preparation Requirements for World Bank Projects, 1975-95 Variable 1 2 3 4 ESW .-1.1784*** - 1.1907*** -2.0273*** -.0.2847*** (0.143) (0.222) (0.171) (0.076) ESW2 0.0O11*** 0.0012" 0.0019*** (0.000) (0.000) (0.000) Public sector surplus -5.2725* (2.876) Inflation -0.0408 (0.040) Openness 0.9206* (0.485) Country performance 65.5442 (relative to region) (51.208) Past project successa -40.8768*** -39.9245* -60.7319*** -42.7433 (15.185) (22.790) (14.018) (32.492) Past preparationa 55.9327** 52.0307'** 67.2006*** (3.753) (5.789) (6.239) Adjusted R2 0.1984 0.1606 0.1115 0.2224 Number of observations 2,355 1,239 3,194 737 S Significant at 10 percent. i** Significant at 1 percent. Note: The dependent variable is total staff time spent on preparation (normalized by loan size). Sector dummies are included but not reported. Standard errors are in parentheses. Coefficients and standard errors are multiplied by 100. a. At the country level. Source: Authors' calculatiDns. allocated to supervision and thus have a negative effect on resource requirements for supervision (equation 3). Table 11 provides partial support that this is indeed the case; higher levels of ESW are clearly associated with lower inputs into supervision except when preparation is included as an explanatory variable. The other regressions reported in table 11 support the notion that, whether consciously or not, task managers; recognize the benefits from ESW and, in line with the proposed maximization model, adjust the allocation of staff time to supervision accordingly. The quantitative magnitude of the effect of ESW on supervision is lower than in the case of preparation; an increase of 1 staff-week in ESW is associated with a decrease of 0.009 staff-week in supervision per $1 million in loans. The decrease is equivalent to a reduction in supervision requirements by 0.64 staff-week for the mean $70 million project (the first column in table 11). Furthermore, it is much cheaper to implement projects in countries with an appropriate policy environment. A 3 percentage point decrease in the public sector deficit, for exarnple, reduces the amount of supervision needed by 0.27 staff-week per $1 million in loans or 56.7 staff-weeks for the mean lending program of $210 million (the second column in table 11). 410 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 11. Determinants of the Supervision Requirements for World Bank Projects, 1975-95 Variable 1 2 3 ESW -1.5812*** -1.0446*** -0.2464 (0.300) (0.320) (0.250) ESW2 0.0024*** 0.0014** 0.0004 (0.001) (0.001) (0.001) Public sector surplus -9.0527*** -7.6614*** (2.872) (2.222) Inflation 0.0769 0.0781 (0.062) (0.048) Openness -1.2403** -0.8083* (0.543) (0.420) Loan size -2.0057*** -1.8145*** -0.7596*** (0.145) (0.151) (0.123) Preparation 63.4220*** (2.410) Past project successa -73.1693*** -60.5990** -57.4011 * * * (19.044) (23.937) (18.519) Adjusted R2 0.2134 0.228 0.538 Number of observations 1,639 1,045 1,045 Significant at 10 percent. * * Significant at 5 percent. * * * Significant at 1 percent. Note: The dependent variable is total staff time spent on supervision (normalized by loan size). Sector dummies are included but not reported. Standard errors are in parentheses. a. At the country level. Source: Authors' calculations. Taken together with the results for project preparation, this implies that as- signing one more staff-week to ESW would lead task managers to reduce the time allocated to project preparation and supervision per project by about 1.7 staff- weeks over the entire life of the project. Moreover, ESW actually benefits more than one project. If we assume that macroeconomic work benefits all projects in the lending program (3 on average for each country), and that sectoral analysis only benefits projects in a particular sector, we calculate that each staff-week of ESW benefits 1.5 projects on average. For the average investment project, the ratio of sector-specific to macroeconomic ESW is about 3 to 1. Thus the saving in preparation and supervision for the entire lending program generated by 1 staff- week of ESW amounts to 2.5 staff-weeks. Therefore, a focus on volume of lend- ing that comes at the cost of project quality (as evidenced at the country level) does seem to increase the costs associated with administering World Bank loans. PROJECT QUALITY. This evidence implies that task managers recognize the positive impact of ESW on lending quality. This hypothesis can be tested directly by estimating equation 4 at the project level for a wide range of independent variables and specifications. The variables of greatest interest are the project Deininger, Squire, and Basu 411 outcome rating (satisfactory-unsatisfactory) and the ERR (estimated after project completion). We also use ARPP ratings (either the average over the whole project life cycle or the final rating for a completed project) to construct a large number of quality indicators that we then use as dependent variables. Because the results from these regressions do not add any substantively new insights, we do not report them separately. From the probit equation of project outcome, ESW does significantly increase the probability of a satisfactory project outcome, even if other policy variables are included (the first three columns in table 12). An increase of 100 staff-weeks allocated to ESW increases the probability of a successful project by between 9 and 13 percent, comparable to the estimate obtained at the program level. In line with the literature, we find that better policies-lower levels of inflation and a more open economy-also contribute significantly to greater project success and are quantitatively quite important. ERRS are not availalble for projects in the social sectors. And a project's benefit to the borrowing country is determined by the size of the loan as well as the ERR, so the net present value would be a more appropriate measure of the benefit to the borrower countryT. Unfortunately, the cost and benefit streams underlying the ERR calculation are available only for very few projects, so we cannot repro- duce NPVS. That said, our results suggest that ESW causes a considerable increase in the ERR for a typical project (columns 4-6 in table 12), regardless of whether the specification includes policy variables, instrumented preparation and super- vision, or country or sector dummies (hypothesis 1). An increase of one staff- week in ESW before project initiation increases a project's rate of return by be- tween 0.02 and 0.04 percentage point. For the mean loan of $70 million in the sample of completed projects, this result translates into an increase in NPV of between $12,000 and $25,000.5 And because each staff-week of ESW includes macroeconomic and sectoral analysis, it should, as noted, benefit more than one project in the lending program. Using the earlier calculation that 1 staff-week of ESW benefits 1.5 projects implies an increase in NPV for each addit:ional staff-week of ESW of between $18,000 and $36,000. This result is very similar to the one obtained when the lending program was the unit of analysis. Based on the reported cost of a staff-week, slightly below $3,000, if the sample of projects available here is any guide, putting additional resources into ESW would be an economically attractive option.6 The fact that ESW iS simi- larly significant in a probit regression of project outcomes suggests that this result is not driven by the sectoral limitations of the availability of ERRs. 5. Again following Kilby (1995), with an average project cost of $70 million, a staff-week increase in ESW before project initiation increases net present value of an average project by between ($0.0002 x 70,000,000) / 1.1 and ($0.0004 x 70,000,000) / 1.1, where the discount rate is assumed to be 10 percent. 6. Although the cost of a staff-week should include travel costs, the latter may not have been reported consistently in our data sources. Even a generous adjustment for this does not change the substantive results. Table 12. Probit Analysis of the Determinants of Project Outcome and the Economic Rate of Return for World Bank Loans, 1975-95 Project outcomea Economic rate of return Variable 1 2 3 4 5 6 ESW 0.0973'* 0.0900'> 0.1302' 2.3737* 4.2285* * 2.3439* (0.387) (0.407) (0.063) (1.4047) (1.6638) (1.3990) Public sector surplus 0.7192 0.8727 14.9735 42.0955 (0.9632) (1.365) (53.6540) (43.8989) Inflation -0.0371 ' -0.06002"* 1.9165 4.90144 (.0214) (.0214) (10.2482) (5.798) Openness 0.7391 0.4729" -3.4949 -1.0494 (0.018) (0.2564) (21.8876) (5.241) Preparation (instrumented) -0.56119 42.0955 (0.3192) (139.257) Supervision (instrumented) 0.02102 5.1640 (0.2649) (103.772) Log-likelihood/R2 -825.56 -534.62 -394.05 0.146 0.142 0.1109 Number of projects 1,324 873 643 481 302 200 Significant at 10 percent. * Significant at 5 percent. * Significant at 1 percent. Note: Sector dummies included but not reported. Standard errors are in parentheses. Coefficients and standard errors are multiplied by 100. a. The mean satisfactory/unsatisfactory rating for all projects, weighted by loan size. Source: Authors' calculations. Deininger, Squire, and Basu 413 Because equation 4 captures our central result (that lending quality depends on ESW), we subjected iit to three tests of robustness. In the first, we changed the specification from linear to log-linear. With or without policy variables, the co- efficient on ESW in the log-linear specification with either project outcome rating or ERR as the dependent variable was positive and significant at least at the 5 percent level (results not reported). In the second, we took advantage of new information that became available late in our study. The results reported thus far cover projects that were evaluated before July 1996. Having completed the estimation for projects up to this date, information became available on addi- tional projects evaluated up to February 1998. Armed with this additional infor- mation, we reran equation 4. The results (not reported) are even stronger. They show that with or without policy variables, the coefficient on ESW remains posi- tive and highly significant in the equation with project outcome rating as the dependent variable. In the equation with ERR as the dependent variable, the co- efficient is positive and significant at the 5 percent level without policies and at the 10 percent level with policies. The third test entails expanding the right-hand-side variables to include some others that have been found in the literature to have an impact on project out- comes. These include indexes of country risk, institutional quality, and bureau- cratic delays (Knack and Keefer 1995) and the number of revolts, coups, and assassinations to capture political stability (King and Levine 1993). The results (not reported) for both the program-level and project-level regressions reveal that the coefficient on ESW never changes sign and remains significant in the majority of cases at the 10 percent level. Thus the results are robust to these tests of specification, sample selection, and choice of independent variables. Although the task manager does not control the time allocated to ESW, we can still use equation 4 to test whether the country manager achieved the right allo- cation between lending services and ESW by exploring whether a shift in resources from preparation and supervision to ESW would have resulted in a better project outcome (hypothesis 2). Table 12 (the third and sixth columns) indicates that the marginal contribution of resources spent on supervision (instrumented) is positive but not statistically significant. The contribution of resources spent on preparation is also positive but insignificant in the equation for the rate of re- turn and is negative and marginally significant in the equation for project out- come. We interpret this last result as an indication that, with the instruments available, we were not able to control for the endogeneity of these two variables, and not as an indication of a negative impact of preparation on project success. The general result is consistent with our results at the program level that an increase in the allocation of resources to ESW would have been associated with an overall improvement in project quality and presumably with a positive im- pact on policy. As we saw earlier, there is a negative raw correlation between project quality and the amount of resources used in supervision (or preparation). We also note that supervision requirements are highly correlated with resource use in prepa- 414 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 ration (preparation alone explains about 55 percent of the variation in alloca- tions to supervision). This correlation suggests that managers do not behave as though preparation and supervision are substitutes, as we have assumed. To the contrary, projects that consume substantial amounts of resources during prepa- ration are generally associated with more than proportional allocations to su- pervision, possibly reflecting a different decisionmaking model. Instead of a simultaneous decisionmaking process undertaken by a single task manager, we may be observing a sequential process undertaken by two or more managers. In the first step, staff time is allocated to preparation, given the avail- ability of ESW to achieve maximum project quality at the time the project is approved by the board. This step reflects the fact that the project and its quality are subject to considerable scrutiny at this time and that staff are rewarded for performing well. In the second step, supervision resources are allocated, often by a different task manager, based on the project's quality at approval to maxi- mize project quality by the time project implementation is completed. In this variant, the task manager who is preparing an ill-conceived project is obliged to allocate more time to preparation than would be the case for a project with a better conceptual foundation. But the task manager is not willing to allocate enough time to reduce future supervision requirements because someone else will be responsible for implementation. Hence, both preparation and supervi- sion tend to be larger than normal for such projects. This suggests that the re- duction in preparation that occurs when more ESW has been undertaken will carry over to a reduction in supervision as well. There are certainly cases where higher spending on preparation (and supervi- sion) may be justified, for example, in the development and piloting of untested or relatively complex approaches. But in general we interpret high preparation costs as indication of a failure to have carried out the necessary analytical back- ground work or to have engaged in a constructive process of consensus building with the main stakeholders involved and, as a consequence, a lack of govern- ment commitment. Experience as well as our results suggest that in such situa- tions it might be advisable first to prepare the ground with ESW before moving ahead with a lengthy and prolonged process of project preparation. Or, where this has not been done, the expected high costs of supervision should be borne in mind when decisions are being made on whether to drop or to continue prepar- ing projects that have incurred large costs during preparation. The fact that, even if instrumented, project preparation and supervision are never significantly different from 0 at the 5 percent level of significance suggests that, at the program as well as the project level, higher spending on preparation or supervision does not automatically increase quality. This failure to identify a positive contribution of preparation and supervision to project success does not mean that, at any given point in a project's life, higher spending on supervision will not increase the probability of an improvement in project quality as demon- strated by Kilby (1995). The difference is that Kilby looks at changes between successive ARPP ratings, thus eliminating project-specific fixed effects, while we Deininger, Squire, and Basu 415 focus on the quality of the project as a whole. We therefore conclude that greater attention to ESW in the form of elaborating the broader context, clarifying the rationale for a specific type of intervention, and assessing its feasibility and eco- nomic desirability against a broader set of potential alternatives could result in the design of better projects and thus less need for spending time on preparation or supervision. V. CONCLUSION At the start of this article, we posed three questions. In these concluding re- marks, we report our answers to each one. We also try to explain why we get the results that we do regarding the behavior of managers. And we make sugges- tions for improving the impact that World Bank projects have on development. First, does ESW improve the quality of World Bank loans? We find that ESW has a significant positive impact on various measures of the quality of World Bank projects. For example, an increase of one staff-week in the amount of time devoted to ESW before project initiation is associated with an increase in the economic rate of return for an individual project of between 0.02 and 0.04 per- centage point. That translates into an increase of between $12,000 and $25,000 in the project's net present value for a cost of no more than $3,000; $1 of ESW yields $4 to $8 in development impact. Indeed, because each staff-week of ESW benefits more than one project, this is an underestimate. Examining the impact of ESW for the entire lending program for a country, we find that $1 of ESW yields between $12 and $15 of development impact. Even this is an underestimate because our calculation fails to capture any benefit that ESW may have in terms of influencing policy formulation in developing countries, a key objective of ESW. We therefore conclude that there has been a high payoff to ESW. Second, has the World Bank underinvested in ESW relative to lending services? Our results at both the program and the project level provide clear evidence of underinvestment in ESW. Assuming that staff resources are fungible across time and among different uses, the marginal contribution of ESW to the quality of lending should be lower than that of lending services (preparation and supervi- sion), because ESW has benefits beyond its immediate impact on lending whereas lending services do not. Indeed, ESW iS often undertaken to provide the basis for policy advice to governments and is not necessarily tightly linked to a particular project or lending program. However, we find the reverse, that ESW has a sys- tematically positive effect on the quality of the lending program, whereas nei- ther preparation nor supervision turns out to be significant. This result suggests that reallocating staff time from lending services to ESW would increase the qual- ity of the lending program. Consistent with this, we find that task managers at the project level are able to reduce the time allocated to lending services by about 2.5 staff-weeks for every staff-week expended on ESW. We infer that ESW helps staff to identify and support new investment options and expands the set of feasible projects. ][t also helps staff to design better projects ex ante and im- 416 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 proves the quality of projects already in the investment program. Preparation and supervision, by contrast, can only improve the quality of a project (whether good or bad) ex post. Third, is there a trade-off between the quality and volume of lending? We find that time allocated to preparation and supervision is between 40 and 50 percent more effective in increasing commitments than ESW, but that resources could be switched from preparation and supervision to ESW to the benefit of both the quality of program and level of disbursements. Even though the results suggest that higher levels of ESW improve the quality of the lending program, it is still possible that, within a given resource envelope, shifting staff time from lending services to ESW reduces the overall volume of lending. Addressing this question, we find that this is indeed the case. Lending services are between 40 and 50 percent more effective in increasing total commitments than is ESW. However, when we analyze disbursements (resource transfer) instead of commitments, we find that managers could increase both the quality of lending and disbursements by switching resources from lending services to ESW. This result, together with our conclusion that there has been underinvestment in ESW from the standpoint of project quality, suggests that at least to some degree the volume of commit- ments has been an additional objective guiding the disposition of staff resources. Our results provide some insight into the trade-off between quality and quan- tity. The analysis suggests that on average a manager is indifferent between a decrease of $2 million in the NPV of a lending program and an additional $4 million of lending volume. If this estimate is broadly accurate, it suggests that managers are prepared to allow a substantial reduction in program quality in return for only a small increase in commitments relative to the average program size. Identifying the behavioral relationships underlying any set of data is a diffi- cult task. Our analysis finds evidence that broad allocations of World Bank staff resources across different activities are influenced by a concern about commit- ments. Country managers choose to allocate resources in favor of preparation at the expense of ESW. And we find evidence that task managers clearly perceive the benefit of ESW and respond by reducing the amount of time allocated to prepara- tion and supervision. Our results suggest that managers reduce preparation and supervision by about 2.5 staff-weeks if one more week is allocated to ESW. To- gether, these results are consistent with the view that task managers factor avail- ability of ESW into their decisions regarding the allocation of resources at the project level at least to some extent, but that country managers systematically underinvest in ESW at the program level. The compensating behavior of task managers (partial maximization in the small) is not able to offset the fact that country managers fail to maximize at the global level. There is good news and bad news in these results. The good news is that our analysis points to areas of improvement that could yield substantial benefits for the development impact of World Bank lending and the total resource transfer that the World Bank provides to developing countries. The bad news is that our Deininger, Squire, and Basu 417 analysis of behavior suggests that in the past managers have had a strong inter- est in the level of commitments. This suggests a need for action. The World Bank should continue to strengthen the reward system such that managers pay even more attention to development impact. And it should educate managers about the importance of ESW in improving project quality. REFERENCES The word "processed"' describes informally reproduced works that may not be com- monly available throu;gh library systems. Arias, Omar. 1994. "Results of the Country Econometric Analysis for the Annual Review of Evaluat'ion Results 1994 for the World Bank, Operations Evaluations Department." Operations Evaluations Department, World Bank, Washington, D.C. Processed. Averch, Harvey. 1994. "Economic Approaches to the Evaluation of Research." Evalua- tion Review 18(1): 77-89. Burnside, Craig, and D)avid Dollar. 1996. "Aid, Policies, and Growth" Policy Research Department, World Bank, Washington, D.C. Processed. Cassen, Robert, and Associates. 1994. Does Aid Work? Report to an Intergovernmental Task Force. Oxford: Clarendon Press. Claessens, Stjin, Gerhard Pohl, and Xuefeng Quian. 1996. "Nonlending Services and Managerial Incentives." World Bank, Washington D.C. Processed. Feyzioglu, Tarhan, Vinaya Swaroop, and Min Zhu. 1995. "The Impact of Foreign Aid on Public Spending." Policy Research Department, World Bank, Washington, D.C. Processed. Isham, Jonathan, and Daniel Kaufman. 1995. "The Forgotten Rationale for Policy Re- form." Policy Research Working Paper 1549. Policy Research Department, World Bank, Washington, D.C. Processed. Kilby, Christopher. 1994. "Risk Management: An Econometric Investigation of Project- Level Factors." Background paper for the Annual Review of Evaluation Results 1994, World Bank, Washington, D.C. Processed. -. 1995. "Superfision and Performance: The Case of World Bank Projects." Cen- ter Working Paper 9545. Tilburg University, Tilburg, Netherlands. Processed. King, Robert G., and Ross Levine. 1993. "Finance, Entrepreneurship, and Growth: Theory and Evidence." JoUrnal of Monetary Economics 32(3):513-42. Knack, Stephen, and Plhilip Keefer. 1995. "Institutions and Economic Performance: Cross- Country Tests Using Alternative Institutional Measures." Economics and Politics 7(3, November) :207-27. Lele, Uma, ed. 1992. Aid to African Agriculture: Lessons from Two Decades of Donors' Experience. Baltimore, Md.: Johns Hopkins University Press. Mosley, Paul, Jane Harrigan, and John Toye. 1995. Aid and Power: The World Bank and Policy-Based Lending. New York: Routledge. National Science Foundation. 1995. "Report of NSF Stakeholders' Panel in Linguistics." Washington, D.C. Processed. Pack, Howard, and Janet Rothenberg Pack. 1990. "Is Foreign Aid Fungible? The Case of Indonesia." The Economic Journal 100(March):188-94. 418 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 -. 1993. "Foreign Aid and the Question of Fungibility." Review of Economics and Statistics 75(2, May):258-65. Rodrik, Dani. 1995. "Why Is There Multilateral Lending?" NBER Working Paper 5160. National Bureau of Economic Research, Cambridge, Mass. Processed. Schneider, Robert. 1995. "Economic Sector Work and Results on the Ground." World Bank, Washington, D.C. Processed. Svensson, Jakob. 1996. "When Is Foreign Aid Policy Credible? Aid Dependence and Conditionality." Policy Research Department, World Bank, Washington, D.C. Pro- cessed. Trumbull, William N., and Howard J. Wall. 1994. "Estimating Aid-Allocation Criteria with Panel Data." The Economic Journal 104(July):876-82. Wise, Lois Rescasino, and Robert Agranoff. 1991. "Organisational Characteristics and Productivity Measurement in Research Organisations." Public Productivity and Man- agement Review 15(1):77-85. World Bank. 1991. "Report on the World Bank Research Program." Research Admin- istrative Department, World Bank, Washington, D.C. Processed. . 1997. World Development Report: The State in a Changing World. New York: Oxford University Press. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 419-55 Demogralphic Transitions and Economic M[iracles in Emerging Asia David E. Bloom and Jeffrey G. Williamson The demographic transition-a change from high to low rates of mortality and fertil- ity-has been more dramatic in East Asia during the twentieth century than in any other region or historical period. By introducing demographic variables into an em- pirical model of economic growth, this article shows that this transition has contrib- uted substantially to East Asia's so-called economic miracle. The miracle occurred in part because East Asia's demographic transition resulted in its working-age popula- tion growing at a much faster rate than its dependent population during 1965-90, thereby expanding the per capita productive capacity of East Asian economies. This effect was not inevitable; rather, it occurred because East Asian countries had social, economic, and political institutions and policies that allowed them to realize the growth potential created by the transition. The empirical analyses indicate that population growth has a purely transitional effect on economic growth; this effect operates only when the dependent and working-age populations are growing at different rates. These results imply that future demographic change will tend to depress growth rates in East Asia, while it will promnote more rapid economic growth in Southeast and South Asia. This article has two objectives. The first is to estimate an empirical model that isolates the impact of demographic variables on economic growth. The second is to use these results to infer how much of the East Asian miracle can be explained by the region's spectacular demographic transition.1 1. We define East Asia to include China, Hong Kong (China), Japan, the Republic of Korea, Singapore, and Taiwan (China); Southeast Asia to include Cambodia, Indonesia, Laos, Malaysia, Myanmar (Burma), the Philippines, Thailand, and Vietnam; and South Asia to include Afghanistan, Bangladesh, Bhutan, India, the Maldives, Nepal, P'akistan, and Sri Lanka. David E. Bloom and Jeffrey G. Williamson are with the School of Public Health and the Department of Economics, respectively, at Harvard University and with the Harvard Institute for International Development. The authors are grateful for the comments of participants at several seminars at the Harvard Institute for International Development and conferences at the Asian Development Bank, the World Bank, the East-West Center, Columbia University, Duke University, FEpADE (Fundacion Empresarial para el Desarrollo Educativo), Harvard University, Johns Hopkins University, the Harvard Institute for International Development, Massachusetts Institute of Technology, Princeton University, Tsukuba University, the Population Council, the University of Pennsylvania, and the Universidad Torcuato di Tella. They appreciate the comments by Neil Bennett, Eric Bettinger, John Bongaarts, David Canning, Mark Gersovitz, Frank Harrigan, Allen Kelley, Gerald Keusch, Ronald Lee, Pia Malaney, Andrew Mason, Jacob Mincer, Steven Radelet, Larry Rosenberg, Jeffrey Sachs, Warren Sanderson, and Andrew Warner; the excellent research assistance provided by Eric Bettinger, Taku Imagawa, Lysander Menezes, Karthik Muralidharan, Andrew Noyimer, and Sze-Tien Quek; and the helpful comments of three anonymous reviewers. The research reported in this article was begun in connection with the Asian Development Bank's project Emerging Asia. © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 419 420 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 The article begins by revisiting the debate on the impact of population growth on economic growth. "Population pessimists" believe that rapid population growth is immiserizing, because it tends to overwhelm any induced response by technological progress and capital accumulation (Coale and Hoover 1958 and Ehrlich 1968). "Population optimists" believe that rapid population growth al- lows countries to capture economies of scale and promotes technological and institutional innovation (Boserup 1981, Kuznets 1967, and Simon 1981). Re- search culminating in the 1980s cast doubt on both views: investigators showed that population growth has neither a significant positive nor a significant nega- tive impact on economic growth (Bloom and Freeman 1986 and Kelley 1988). These studies were typically based on cross-country regressions of per capita income growth on population growth, controlling for a variety of other influ- ences. As Kelley and Schmidt (1995: 543) put it recently, Possibly the most influential statistical finding that has shaped the "population debates" in recent decades is the failure, in more than a dozen studies using cross-country data, to unearth a statistically significant association between the growth rates of population and of per capita output. This "population neutralist" finding is surprising, but whether it arose because population has no positive or negative effects on economic growth, because it has no net effect on economic growth, or because both the pessimists and the optimists have misspecified the test remains unclear. More recent work has decomposed population growth into its fertility and mortality components and examined their independent effects on economic growth (Barlow 1994, Bloom and Freeman 1988, Brander and Dowrick 1994, Coale 1986, and Kelley and Schmidt 1995). These studies find that measures of fertility, specifically past birth rates, are negatively and significantly associated with economic growth, whereas the effect of mortality is insignificant. This more recent work is the direct precursor of this article, insofar as it justifies the de- composition on the grounds that changes in fertility and mortality imply very different changes in the age distribution and points toward our hypothesis that population growth affects economic growth insofar as it affects the ratio of working-age population to dependent population. Population growth attribut- able to improvements in longevity among the elderly should have an immediate negative effect on economic growth, because this implies a greater number of elderly to support. Population growth attributable to a general decline in mor- tality has no effect, because the ratio of the economically active population to dependents stays the same. Population growth attributable to a rise in fertility should have an immediate negative effect on economic growth, given the pres- ence of more mouths to feed, and so should population growth stemming from a fall in infant mortality. These latter demographic effects will, however, have a delayed positive impact on economic growth, because the economically active population will boom two decades later. Bloom and Williamson 421 This article contributes to the population debate in four ways. First, like Kelley and Schmidt (1995), it uses the new empirical models of economic growth to isolate the effects o:F demography. It does this by incorporating demographic variables into a growth model similar to the one used in Asian Development Bank (1997) and Barro and Sala-i-Martin (1995). Second, it explores the pos- sibility of reverse causality between economic growth and demographic change by using a two-stage specification in which instruments for the growth rate of the population are used to correct for possible endogeneity. Third, it intro- duces demography into the growth equations in a theoretically more appeal- ing way-by adding the growth rates of the total population and the economi- cally active population rather than by simply including birth and death rates. This allows population growth to affect economic growth both by its overall rate of increase and by its effect on the age structure. The distinction matters. Finally, the article highlights how changes in the growth of labor force per capita, changes in the savings rate, and changes in the investment rate are three plausible channels through which a changing age structure might affect the rate of economic growth (Bloom and Williamson 1997 and Higgins and Williamson 1997). The article uses the econometric results to assess the extent to which popula- tion dynamics may account for a significant portion of East Asia's economic miracle. East Asia is an excellent context in which to examine this effect for several reasons. It has experienced a more rapid demographic transition than any other region at any time in history. We argue that the initial fall in infant mortality, which set the demographic transition in motion, was likely to have been exogenous in laice twentieth century East Asia. East Asia has also experi- enced higher sustained rates of economic growth over the past 30 years than any other region at any other time in history. East Asia is often compared with South- east and South Asia, whose demographic transitions either began later or pro- ceeded more slowly and whose recent economic progress has not rivaled that of East Asia. And analysts have badly neglected the potential role of population change in economic performance in the region, a neglect illustrated best by the World Bank's oft-quoted work The East Asian Miracle (1993). In redressing this imbalance, the article compares Asia with the rest of the world and Asia's subregions with one another. Section I describes the demographic transition in more detail, focusing on the difference between the experiences of Western Europe and Asia to show that demographic effects have been much more pronounced in Asia. Section II de- scribes the model and the recent literature on economic growth on which it is based. Section III presents the econometric results, and section IV uses those results to estimate just how much of the East Asian miracle may be accounted for by demographic dynamics. Section V discusses labor supply and capital ac- cumulation, the most likely channels through which population dynamics affect economic growth. Section VI concludes with an agenda for future research. 422 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 I. THE DEMOGRAPHIC TRANSITION AND ECONOMIC GRowTH The demographic transition describes the change from preindustrial high fer- tility and mortality to postindustrial low fertility and mortality. Figure 1 offers a stylized view of the transition. Declines in mortality mark the beginning of al- most all demographic transitions, and changes in the age structure are exacer- bated because infants and children enjoy most of these early declines in mortal- ity. True, the improved survivor rates for children induce parents to reduce their fertility. If parents adjusted completely and immediately, there would be no youth glut and no acceleration in population growth. But they do not: they adjust slowly, and the youth glut is large and persistent. After a lag, however, fertility begins to decline, which marks the next stage of the transition. The population growth rate is implicit in the first panel of figure 1 as the difference between fertility and mortality. The second panel makes the population dynamics ex- plicit: the demographic transition must be accompanied by a cycle in population growth and the age structure. Figure 1 and the rest of this article treat the demo- graphic system as if it were closed and thus ignore external migration. If it were quantitatively important and responded to cohort gluts and scarcities, external migration might very well mute the impact of demographic transitions. In the late twentieth century, international migrations are simply not great enough to matter except, perhaps, for the United States and some oil-producing countries in the Middle East (Bloom and Noor 1997). They mattered a great deal, how- ever, in the age of relatively unrestricted mass migration prior to World War I (Williamson 1998). These components of the demographic transition might have separate influ- ences on economic growth. The population growth rate could influence eco- nomic growth for the reasons cited by population pessimists or optimists. The demographic transition could also affect economic growth through the age dis- tribution, as we emphasize. Coale and Hoover (1958) made the dependency rate the centerpiece of their analysis of the impact of large youth cohorts on savings, investment, and educational capital deepening. Because they were, by virtue of the decade in which they conducted their analysis, constrained to study only the first-"burden"-phase of the Asian demographic transition, they could not devote attention to the "gift" phase from the mid-1960s to the present that drives this analysis. Overall, the age distribution effect will operate first to lower, then to raise, then to lower again the ratio of the economically active population to the total population and thus will have a transitional impact on growth of the labor force per capita. Note that the demographic "gift" in the middle phase of the transition may or may not be realized. It represents a growth potential whose realization depends on other features of the social, economic, and political environment. Like industrial revolutions, demographic transitions take many decades to complete, but in the case of postwar East Asia it has been much faster than it was in nineteenth century Europe. Over a century and a half, Europe slowly Bloom and Williamson 423 Figure 1. Demographic Transition and Population Growth Demographic transition Population Birth rate growth rate Death rate \ < t ~~~~~~~~~B irt~~~Brh rate Death rate Time Population growth and the age structure Share | ......... --- ~~~~~~working Bfirth rateI /\\ death ~~~~~~~~~~~~~~~~~~Percent in rate ~~~~~~~~~~~~~~~~workforce minus p/ ~~Growth -`-- erckfOr Time improved its understanding of and practices with respect to basic sanitation, management of solid waste, provision of clean drinking water, and the elements of sound nutrition. It invested in these measures to reduce mortality and chronic malnutrition and eventually eliminated famines (Fogel 1994). It cleaned up what Victorian reformers called "killer cities" (Williamson 1990). These factors, to- 424 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 gether with the advent of antibiotics and vaccines and recognition of the impor- tance of preventive medicine, led to a gradual decline in mortality in Europe. Infant and child mortality led the decline because the very young, like the el- derly, are most vulnerable to infectious disease, and because children are far more numerous than the elderly at early development stages, the decline in in- fant and child mortality matters most. The fertility rate also declined, but more slowly, and the European demographic transition stretched out for more than 100 years (Coale and Watkins 1986). The health investments and medical technologies that had been developed and put into practice in Europe did not exist in Asia until relatively recently. There was a large gap between the best health practices prevailing in industrial- ized Europe and local health practices prevailing in Asia, and by 1940 the scope for the transmission of health technologies was enormous, having been pent up by deglobalization, two world wars, the Great Depression, and wars of colonial liberation. When the postwar transfer of this pent-up health technology finally took place, it happened in a rush. The process was sped up even further by investment in health-improving social overhead, which was heavily financed by world funding agencies that did not exist prior to the 1940s. In short, the possi- bilities for Asia to catch up with the West in terms of health and demography were enormous in the late 1940s, and they were driven by factors external to Asia itself. In the half century since then, Asia has exploited the catch-up poten- tial with such enthusiasm that it has produced one of the fastest and most dra- matic demographic transitions ever. The language we use in this section is pur- posely similar to that used in the debate about economic catch-up and convergence (Abramovitz 1986, Barro 1991, Baumol 1986, and Sachs and Warner 1995), be- cause we think that exactly the same reasoning applies to the demographic tran- sition in Asia. Asia's demographic transition followed the stylized model by starting with a decline in mortality rates. By the late 1940s, the crude death rate had begun to decline very rapidly throughout much of Asia. The decline proceeded most rap- idly in East Asia (figure 2) and was accompanied by an increase in life expect- ancy from 61.2 to 74.6 years from 1960 to 1992. Similar declines occurred in Southeast and South Asia, where life expectancy improved from 51.6 to 67.2 years and from 46.9 to 60.6 years, respectively. In the 1950s and 1960s, most of the aggregate decline in mortality was driven by declines in mortality among the youngest cohorts (Bloom and Williamson 1997). There are a number of possible explanations for the rapid decline in child mortality in Asia in the middle of this century. One possibility has already been suggested: that is, in the 1940s Asia escaped from some four or five decades of relative isolation, ushering in an era of transfer and diffusion of new public health programs, technologies, and techniques. For example, the medical ad- vances that were implemented in postwar Asia had been accumulating on the technological shelf for at least two decades: penicillin was discovered in 1927, sulfa drugs in 1932, and bacitracin in 1943; streptomycin was isolated in 1943 Bloom and Williamson 425 Figure 2. Crude Death Rate, by Subregion in Asia, 1950-2020 Number of deaths per 1,000 population 40 --t-.--- _ 30 - - - - - - -… 25……IF 20 15-- East s;tN L ast 1950 1960 1970 1980 1990 2000 2010 2020 Source: United Nations (1594). and its curative value against tuberculosis demonstrated; the efficacy of chloro- quine in treating malaria was demonstrated in 1943; and 1945 saw the nonmili- tary use of penicillin and 1948 the introduction of tetracycline. With the advent of these and other drugs, diseases that had once killed hundreds of thousands, and even millions, became treatable at low cost. In addition, the pesticide DDT became available in 1943. To cite just one example, DDT spraying in the late. 1940s dramatically reduced the incidence of malaria in Sri Lanka: the crude death rate declined from 21.5 to 12.6 between 1945 and 1950, with the most precipitous drops in the most malarial areas (Livi-Bacci 1992). Figure 3 illus- trates the effect by plotting changes in mortality in the most and least malarial zones of Sri Lanka between 1930 and 1960. While the least malarial areas show a gradual decline during the period, the decline is dramatic and steep between 1943 and 1949 in the most malarial zone. Another possibility is that increased agricultural productivity and trade in food both improved nutrition sufficiently to lower infant mortality dramatically in less than a decade and did so everywhere in Asia. This may be true, but it seems unlikely given that the magnitude and timing of the decline in mortality were so similar everywhere in Asia, regardless of level of development and agri- cultural productivity. 426 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 3. Tbe Effect of DDT Usage on Mortality in Sri Lanka, 1930-60 Deaths (per 1,000 population) 60 50 Beginning of DDT eradication 40 35 30 25 Anuradhapura 20 15 _ \5o sas _ _ Kalutara 10 5 I . . .. . . 1. . 1 .. 1 . . 1930 1935 1940 1945 1950 1955 1960 Source: Livi-Bacci (1992). Resolving the debate between the view that favors an exogenous supply-side- driven fall in infant mortality in the 1940s and 1950s and one that favors an endogenous demand-side-driven fall matters because it influences the extent to which the demographic transition in East Asia was mostly exogenous to the economic miracle itself. Future research must resolve this issue. It must be stressed that whether and how fertility responds to economic events (and to rising child survivor rates) is irrelevant to the discussion of whether these demographic shocks were exogenous to the economic miracle in the first place. The decline in fertility is, of course, largely endogenous, but that response simply serves to mute the impact on population growth of the exogenous decline in child mortality that sets the whole demographic transi- tion in motion. Although the timing of the decline in mortality was remark- ably similar across rich and poor Asia-suggesting that exogenous forces were at work-the lag between the drop in mortality and fertility, as well as the size of the ensuing fall in fertility, varied-suggesting that endogenous forces were at work (Bloom and Williamson 1997 and Feeney and Mason 1997). Figure 4 plots the decline in the crude birth rate for East, Southeast, and South Asia. Although the crude birth rate fell much more rapidly in East Asia than in Southeast or South Asia, the timing was not so different. In most countries, like Korea, Malaysia, and Singapore, fertility began to decline about 15 years Bloom and Williamson 427 after the drop in child mortality. In other countries, like Thailand, the delay was closer to 25 years. What is remarkable about the onset of the decline in Asian fertility is thatt it occurred in such a short period and that it was so dramatic everywhere, even where the pace of economic development was slow (Caldwell and Caldwell 1996). There are, of course, a number of possible explanations for the decline in fertility. Contraceptive use rates vary across Asia (Bloom and Williamson 1997: table 5); government intervention accounts for some of this variance, while family demand, responding in part to economic events, accounts for the re- mainder. The big debate is over which factor matters most. Two well-known demographers argue that government intervention matters a great deal and that the intervention is distinctly Asian (Caldwell and Caldwell 1996). Another even offers an estimate: examining the decline in the total fertility rate from 1965 to 1975 for 683 developing countries, Boulier (1986) concludes that 27 percent was due to economic change and 40 percent to government-supported family planning, with the remainder representing a continuation of long-term trends. By contrast, Gertler and Molyneaux (1994) and Pritchett (1994) both find that socioeconomic variables such as income and education play a much more significant role in fertility decline than family planning does. The general view, however, seems to be that family planning programs helped to trigger the decline in Asian fertility, beginning with India in 1951. But, as Sanderson and Tan (1995) point out, diminishing marginal returns may imply a reduced ben- Figure 4. Crude Birtb Rate, by Subregion in Asia, 1950-2020. Number of births per 1,000 population 45 40 35 Ute As a 30 25 20 atMa 15 10 1950 1960i 1970 1980 1990 2000 2010 2020 Source: United Nations (1994). 428 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 efit to incremental government investments in family planning in countries with well-established programs. The pace and timing of the demographic transition have led to enormously divergent trends in population growth and age structure across Asia. Figure 5 plots the ratio of the working-age population to the nonworking-age population for the three subregions in Asia. With only two precocious exceptions, Japan and Sri Lanka, Asia's surge to peak youth dependency rates occurred in the 1960s and 1970s, reflected in figure 5 by the low ratio of working-age popula- tion to nonworking-age population in those decades. As figure 5 demonstrates, the ratio of working-age population to nonworking- age population has been rising in Asia since 1970, but this increase was espe- cially dramatic in East Asia between 1975 and 1990. According to the United Nations (1991) projections, the ratio of working-age population to nonwork- ing-age population will peak in East Asia in 2010, ending the second phase of its demographic transition. With the exception of Japan, the elderly dependency rate has been mostly irrelevant to Asia in this century, even to the more eco- nomically mature East Asia. It will, of course, become very relevant to these older tigers as they enter the next century. Indeed, figure 5 projects a decline in the ratio of the working-age to the nonworking-age population after 2010 (the third phase of the demographic transition). This reflects the increase in the el- derly dependency rate as the bulge in the age distribution works its way through East Asia's population pyramid. However, the elderly dependency rate is not expected to become a dominant demographic force anywhere else in Asia even as late as 2030. Figure 5. Ratio of Working-age to Nonworking-age Population in Asia, 1950-2030 Ratio 1.8 i Southep1t Asia 1950 1960 1970 1980 1990 2000 2010 2020 2030 Source: United Nations (1994). Bloom and Williamson 429 In this article, we seek to measure the effects on economic performance of population growth and of changes in age structure. Population growth is ex- pected to influence economic growth through the channels discussed in the stan- dard debate between optimists and pessimists, such as economies of scale or reductions in the capital-labor ratio. This article, however, argues that in the early stages of the demographic transition, rising youth dependency burdens and falling shares of working-age adults diminish the growth of per capita in- come. As the transition proceeds, falling youth dependency burdens and rising shares of working-age adults promote the growth of per capita income. The early burden of having few workers and savers becomes a potential gift: a dis- proportionately high share of working-age adults. Later, the economic gift dissi- pates, as the share of elderly rises. If this framework is correct, then some of the slower growth prior to 1970 can be attributed to E ast Asia's very heavy youth dependency burden, which, by itself, was depressing growth rates. Without the youth dependency burden, so the argument goes, East Asia would have had higher growth rates prior to 1970. As East Asia graduated from demographic burden to gift, the youth dependency burden decreased and the proportion of working-age adults increased. The re- sult was an acceleration of the economic growth rate due to demographic forces. This and other transitional forces-productivity gains from "borrowing" for- eign technologies, from shifting labor from sectors with low productivity (agri- culture) to sectors with high productivity (industry and services), from exploit- ing the potential of globalization-pushed the growth rate far above its pre-1970 level to the "miraculous" rates of the past quarter century. The demographic transition accounts foir a decrease in the growth rate associated with high youth dependency burdens and a subsequent rise in the growth rate stemming from the emergence of the dernographic gift. Some time in the near future, however, East Asia's demographic gift will dissipate, and, consequently, economic growth will tend to slow as the share of elderly in the population increases. Once the demo- graphic transition is complete and the population age structure stabilizes, popu- lation growth will affect economic growth only insofar as it operates through level effects. Hence, any economic effect due to the changing age distribution will be temporary. Figure 6 offers a stylized version of the economic hypothesis in which the sustainable growth rate is taken to be about 2 percent a year. Note, however, that the contribution of the demographic transition to the East Asian miracle will also depend on how the miracle is defined. If it is defined as a share of per capita GDP growth between 1960 and 2010 (as in figure 6), then the demo- graphic transition accounts for about one-third of the miracle. If it is defined as the surplus over the sustainable rate, then the transition accounts for almost half, while if it is defined as the increase in growth rates from 1945-60 to 1960-2010, then thLe transition accounts for almost three-quarters. What fol- lows is a test of the hypothesis and a defense of the magnitudes suggested by figure 6. 430 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 6. Stylized Model of Economic Growth and the Demographic Transition in East Asia, 1945-2025 Growth rate of real GDP per capita Ecnmircl'| Economic Demographic "gift" ,,,"~~~~., . ". . .. .. . .. . . . . .. . . .. .. .. ............. demiograph Sustainable 'burden" growth c. 1945 c. 1960 c. 2010 c. 2025 Time II. THE THEORETICAL FRAMEWORK The cross-country growth equations estimated in the next section are derived from a conventional Solow-Swan model of economic growth (Barro and Sala-i- Martin 1995). The Solow-Swan model is a special case of the Ramsey model with fixed savings rates. However, the empirical estimation equation derived by the log-linear approximation around the steady state is identical in both models. Competitive firms take wages and the interest rate as given and produce the same good. The savings rate is fixed and exogenously determined. Workers are identical. If we assume that production per worker (y) takes the form y = Aka, where A is an index of total factor productivity, oc is the output elasticity of capital, and k represents capital stock per worker, then we can derive equation 1 for the growth rate of y. Equation 1 will be familiar to anyone who has read a current advanced macroeconomics textbook (for example, Barro and Sala-i- Martin 1995). It is also consistent with the empirical growth literature, espe- cially that which focuses on conditional convergence (Barro 1991, Barro and Lee 1994, Mankiw, Romer, and Weil 1992, and Sachs and Warner 1995).2 In 2. For an alternative framework within which to model the demographic transition, see Ehrlich and Lui (1991). Using an overlapping generations model, they show that utility-maximizing individuals will Bloom and Williamson 431 the Solow-Swan model, the average growth rate (gy) of output per worker be- tween any time T1 and T2 is proportional to the natural logarithm of the ratio of income per worker in the steady state (y*) and income per worker at time T1 as follows: (1) T2 -= lnT y(T)] = _ Y ] We add two modifications to this model. The first involves the formulation of steady-state output. As in Asian Development Bank (1997), we assume that y# is formed as (2) y X where X is a matrix with k determinants of the steady state. We also follow Asian Development Bank (1997) in our selection of the variables to include in X. These variables are average years of secondary schooling in the initial period (in natural logs), life expectancy in the initial period, a measure of natural resource abundance, a measure of openness, an index of institutional quality, average government savings, and geographic variables indicating the ratio of coastline to land area, whether there is access to major ports, and whether the country is located in the tropics. The second modification involves changing the model from output per worker (y) to output per capita (y). We note that ( Y YL L N LN N where N is the total population, L is the number of workers, and y is output per capita. This expression can easily be converted to growth rates, (4) 9y = gy + gworkers - gpopulation- When equations 1 and 2 are substituted into 4 and a stochastic term is added, the estimation equation 5 emerges: (5) X1 = 1 fl +y(Tl) 12 +orkers n13 +gpopulation 14 +E- choose to have fewer children in response to an exogenous decline in mortality rates. The resultant investment in the quality, instead of in the quantity, of children can push a country onto an endogenous growth path, leading to higher growth rates. Meltzer (1995) includes health along with education as a factor of production in a standard Ramsey growth model. When fertility rates are endogenous, an exogenous decline in mortality can be shown once again to set the economy on a path of sustained economic growth with a parallel decline in population growth. Table 1. Variable Definitions and Selected Descriptive Statistics Variable Source Mean Standard deviation Minimum Maximum Population growth rate, 1965-90 World Bank data 1.88 1.00 0.17 3.49 Growth rate of economically active population, 1965-90 World Bank data 2.17 1.03 0.25 3.63 Growth rate of population under age IS World Bank data 1.11 1.53 -1.43 3.69 Growth rate of population over age 64 World Bank data 2.62 0.98 0.79 5.73 Growth rate of the dependent population World Bank data 1.46 1.17 -0.40 3.55 Average birth rate, 1967-87 World Bank data 30.89 12.56 t3.7 53.9 Average death rate, 1967-87 World Bank data 11.68 5.04 5.15 28.85 Average infant death rate, 1967-87 World Bank data 2.54 2.52 0.12 9.70 Average noninfant death rate, 1967-87 World Bank data 9.03 3.21 3.87 19.55 Log GDP per capita as a ratio of U.S. GDP per capita, 1965 World Bank data -1.65 0.91 -3.34 0.00 Log years of secondary schooling, 1965 (average years of secondary school for population age 25 or older) Barro and Lee (1994) -0.70 1.15 -4.83 1.26 Log life expectancy, 1960 World Bank data 4.02 0.22 3.47 4.30 Natural resource abundance (share of primary product exports in GDP in 1971) World Bank data 0.10 0.09 0.00 0.51 Access to ports dummy (indicating if the country is landlocked) 0.13 0.34 0.0 1.0 Openness Sachs and Warner (1995) 0.45 0.45 0.0 1.0 Tropics dummy (indicating if country is located between the tropics) World Bank data 0.51 0.48 0.0 1.0 Ratio of coastline to land area World Bank data 0.30 0.96 0.0 7.33 Government savings as a share of GDP, 1970-90 World Bank data 1.44 3.43 -5.24 12.57 Quality of institutions (index of quality of governmental institutions) Keefer and Knack (1995) 6.11 2.42 2.27 9.98 Note: The database is available from the authors upon request. Bloom and Williamson 433 Theoretically, one would expect that H13 = -114 = 1, which implies that for a stable population, where the growth rate of the work force equals the growth rate of the population, net demographic effects should vanish. If the population is unstable, as it is during a dynamic transition, then demography might matter. This formulation takes age structure into account by focusing on both the total population and the working-age population. Bloom, Canning, and Malaney (forthcoming) extend this approach to account for the age structure of the working-age populatiorn, which may be important insofar as productivity varies over the working life cycle. Their model can be generalized to take account of other productivity-related characteristics as well. It is possible as well that both population growth and growth in the labor force might affect the steady-state rate of income growth. The Solow-Swan model posits an exogenous rate of growth of workers n. This is presumed to have a negative effect on the steady-state level of income per worker through reduc- tions in the capital-labor ratio and hence on the rate of growth of income per worker. However, once demographic factors are incorporated, an increase in n relative to population growth will also reduce the dependency ratio. According to Coale and Hoover's (1958) hypothesis, this leads to increases in the per capita rate of savings, which will offset, and possibly even reverse, the negative effect of labor growth on the capital-labor ratio. If the increase in savings is more than proportional to the growth in labor, then increases in n will lead to a rise in the steady-state rate of growth. Alternatively, if it is less than proportional, the capi- tal-labor ratio will decline and the steady state will fall. These effects on growth rates are not identified separately in our model and will be absorbed into the coefficient on gworkers , potentially causing it to deviate from 1 in magnitude. The coefficient on gpopulation will also absorb any influences of population growth on the steady-state rate o0: economic growth, as discussed in the debate between population optimists and pessimists. To the extent that these influences are im- portant, the coefficient on gpopulation may deviate from -1. III. ECONOMETRIC RESULTS The econometric analysis is based on 78 Asian and non-Asian countries cov- ering the quarter century from 1965 to 1990. It includes every country for which all the data exist. Table 1 provides a complete description of the data with sources, and the appendix provides a list of the countries. We start by asking whether the level of population growth affects economic growth, because the population debate has always been couched-erroneously we believe-in those terms. The results appear in table 2. Most of the recent research on economic convergence has focused on the sign of the coefficient on logged initial income. If the coefficient is negative, the model predicts condi- tional convergence, that is, after controlling for factors that determine the steady- state level of income, poor countries tend to grow faster and approach their steady-state level more quickly than rich countries. Consistent with recent re- 434 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 2. The Impact of Population Growth and Other Factors on Economic Growth, 1965-90 Variable a b Population growth rate, 1965-90 0.16 0.56 (0.20) (0.16) GDP per capita as ratio of U.S. GDP per capita, -1.50 -2.30 1965 (logged) (0.25) (0.22) Log life expectancy, 1960 5.81 (0.98) Log years of secondary schooling, 1965 0.82 0.37 (0.18) (0.15) Natural resource abundance -4.68 -2.40 (1.35) (1.17) Openness 2.23 1.88 (0.47) (0.36) Quality of institutions 0.21 0.22 (0.10) (0.07) Access to ports dummy -0.68 -0.87 (0.39) (0.29) Average government savings rate, 1970-90 0.18 0.15 (0.04) (0.03) Tropics dummy -1.09 (0.33) Ratio of coastline to land area 0.29 (0.12) Constant -2.11 -27.38 (0.92) (4.3) Adjusted R2 0.69 0.83 Note: The dependent variable is the growth rate of real GDP per capita in 1965-90 in purchasing power parity terms. Estimates are from ordinary least squares. The sample size is 78 economies (see the appendix). Standard errors are in parentheses. Source: Authors' calculations. search on economic convergence, we also find conditional convergence in our sample. Our focus, however, is on the rate of population growth. In the first specification in table 2 (column a), there is no significant relationship between population growth and growth of gross domestic product (GDP) per capita, thereby supporting the neutralist position. However, this result is sensitive to the speci- fication. As soon as log life expectancy in 1960 and two variables controlling for economic geography are added, population is shown to have a positive and sig- nificant impact on growth of GDP per capita (table 2, column b), thereby sup- porting the optimists' position. Throughout this section, and specifically in tables 2, 3, 4, 5 and 6, we report both specifications. Specification a always refers to a model that excludes initial life expectancy and two geographical variables-a tropics dummy and a ratio of coastline to land area. Specification b always in- cludes these three variables. Table 2 illustrates the kind of analyses that economic demographers have undertaken to examine the connection between demography and economic Bloom and Williamson 435 growth. It seems plausible, however, that both the sources of population growth and the stage of the demographic transition do matter: both a decline in child mortality and a baby boom raise the share of young dependents in the popula- tion; a decline in mortality among the elderly increases the share of the retired dependent age cohort; immigration raises the working-age population (because it self-selects young adults); and improved mortality among the population at large has no impact on age structure at all. Because an economy's productive capacity is linked directly to the size of its working-age population relative to its total population, distinguishing between the two components when exploring the impact of demographic change on economic performance seems natural and worthwhile. Table 3 conforms to these notions: the growth rate in the economically active population joins population growth in the regression. The growth rate of the working-age population measures the change in the size of the population ages 15 to 64 between 1965 and 1990. Table 3 confirms that the growth of the working-age population has a powerful, positive impact on growth of GDP per capita, while growth of the total population has a powerful negative impact after controlling for other expected influences. Consider the results reported in column lb of table 3. The coefficient on the growth rate of the working-age population is positive, statistically significant, and large in magnitude: an in- crease of 1 percent in the growth rate of the working-age population is associ- ated with an increase of 1.46 percent in the growth rate of GDP per capita. The coefficient on the growth rate of the total population is negative, statistically significant, and almost as large: an increase of 1 percent in the growth rate of the overall population (effectively, the dependent population, since the empirical specification holds fixed the growth rate of the working-age population) is asso- ciated with a decrease of 1.03 percent in the growth rate of GDP per capita. The coefficients of the other variables are similar to those found in Asian Develop- ment Bank (1997) and Radelet, Sachs, and Lee (1997). Columns 2a and 2b of table 3 show what happens when the impact of the growth rate of the working- age population and that of the entire population are constrained to be equal, but of opposite sign. In steady state, when the age distribution is stable, population growth will not matter in either of these two specifications. In transition, when the age distribution changes, population growth does matter. The coefficient here is large, positive, and significant. Thus, in our sample, where the growth rate of the economically active population exceeds that of the overall popula- tion, higher growth r ates of GDP per capita have appeared (ceteris paribus). The opposite is true if the growth rate of the total population exceeds that of the economically active population. If the dependent population is growing more rapidly than the work force, the estimates provide evidence of slower growth. Previous contributions to the population debate typically failed to explore the possibility of reverse causality between population growth and economic growth, despite a literature suggesting that economic events can induce demographic responses. While table 3 uses ordinary least squares (OLS), table 4 reports the 436 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 3. The Effects of Growth in the Population and the Economically Active Population on Economic Growth, 1965-90 Variable la lb 2a 2b Growth rate of economically 1.95 1.46 active population, 1965-90 (0.38) (0.34) Population growth rate, -1.87 -1.03 1965-90 (0.43) (0.40) Difference in growth rates' 1.97 1.68 (0.38) (0.35) GDP per capita as a ratio of U.S. -1.36 -2.00 -1.39 -1.97 GDP per capita, 1965 (logged) (0.21) (0.21) (0.21) (0.22) Log life expectancy, 1960 3.96 2.94 (0.97) (0.97) Log years of secondary 0.50 0.22 0.50 0.28 schooling, 1965 (0.16) (0.14) (0.16) (0.14) Natural resource abundance -4.86 -2.35 -4.86 -2.57 (1.2) (1.0) (1.1) (1.1) Openness 2.06 1.92 2.00 1.72 (0.40) (0.32) (0.38) (0.33) Quality of institutions 0.23 0.20 0.22 0.15 (0.08) (0.07) (0.08) (0.07) Access to ports dummy -0.35 -0.64 -0.31 -0.40 (0.34) (0.27) (0.32) (0.27) Average government savings 0.14 0.12 0.14 0.13 rate, 1970-90 (0.03) (0.03) (0.03) (0.03) Tropics dummy -1.31 -1.20 (0.30) (0.31) Ratio of coastline to 0.24 0.23 land area (0.11) (0.12) Constant -2.46 -19.5 -2.28 -14.3 (0.79) (4.3) (0.69) (4.1) Adjusted R2 0.76 0.86 0.78 0.85 F(1, 68)b 0.22; Prob > F = 0.64 F(1, 64)b 9.03; Prob > F = 0.003 Note: The dependent variable is the growth rate of real GDP per capita in 1965-90 in purchasing power parity terms. Estimates are from ordinary least squares. The sample size is 78 economies (see the appendix). Standard errors are reported in parentheses. a. Growth rate of the economically active population minus growth rate of the total population, 1965-90. b. Test of the null hypothesis that the population growth rate equals the negative of the growth rate of the economically active population between 1965 and 1990. Source: Authors' calculations. results when an instrumental variables (iv) estimator is used to account for pos- sible reverse causality. The instruments include lagged population growth, log life expectancy in 1960 (in columns la and 2a), population policy indicators, and information on the religious composition of the population. Because the instruments chosen are available only for a smaller sample of countries, the OLS estimates corresponding to this sample are also included in the table. The coun- tries excluded from the smaller sample can be found in the notes to table 4. In column lb of table 4, the coefficients on the growth rates of the working-age Bloom and Williamson 437 Table 4. Instrumental Variables Estimates of the Effects of Population Growth on Economic Growth, 1965-90 Variable la lb 2a 2b Growth rate of economically active population, 1965-90 Instrumental variables 3.83 1.37 (0.82) (1.71) Ordinary least squares 1.95 1.41 (0.40) (0.37) Population growth rate, 19655-90 Instrumental variables -4.19 -0.92 (0.96) (2.12) Ordinary least squares -1.93 -0.97 (0.45) (0.43) Difference in growth ratesa Instrumental variables 3.28 2.96 (0.65) (0.75) Ordinary least squares 1.95 1.60 (0.40) (0.38) R2 from first-stage regression 0.96 0.96 0.54 0.54 F-test for joint significance of instruments in first-stage regression (ndf, ddf) 2.59 0.26 2.59 0.26 (9, 52) (8, 50) (9, 52) (8, 50) Hausman specification test 7.13 0.00 6.58 4.47 (chi-square with d) (10 d (13 df) (9 d) (12 df) Note: The dependent variable is the growth rate of real GDP per capita in 1965-90 in purchasing power parity terms. Instruments in the first-stage regression include average growth of population in 1950-60, log life expectancy in 1960 (in columns la and 2a), urban share of the population in 1965, population policy variables including attitudes toward fertility and population growth and whether a government agency exists to design and implement population policy, and dummy variables for economies where a majority of the population is Islamic or is Judeo-Christian. The regressions in columns la and 2a also include the additional variables in column a of table 2. The regressions in columns lb and 2b include those in column b of table 2. The sample size is 70 economies (see the appendix). The following economies are not included in the data set used to calculate the estimates reported in this table due to missing data: Botswana, Haiti, Hong Kong (China), Niger, Singapore, Taiwan (China), Tanzania, and Zaire. Standard errors are in parentheses. a. Growth rate of the economically active population minus growth rate of the total population, 1965-90. Source: Authors' calculations. and the total population are similar to the OLS estimates: an increase of 1 per- centage point in the growth rate of the working-age population is associated with an increase of 1.37 percentage points in growth of GDP per capita, and an increase of 1 percentage point in the growth rate of the total population is asso- ciated with a decrease of 0.92 percentage point of growth in GDP per capita. In column Ib, the IV estimates, with high standard errors, lack the precision of the OLS estimates, but in columns la, 2a, and 2b, the IV estimates yield more precise estimates. As can be seen from columns 2a and 2b in table 4, when the coeffi- cients on the growth rate of the economically active population and total popu- lation growth are constrained to be equal and opposite in sign, the estimated IV 438 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 coefficients are almost twice as large as their OLS counterparts. All of the con- strained estimates are statistically significant at all conventional levels. The simi- larity of the signs and significance of these estimated coefficients across the al- ternative specifications and estimation techniques speaks well for the robustness of the result. Hausman specification tests (Hausman 1978) were performed to test for con- sistency of the OLS estimates. The test statistics, reported in each column of table 4, suggest that, in both the constrained and unconstrained versions of the model, one cannot reject the null hypothesis that the Iv and OLS estimates are statisti- cally equivalent. Although the data do not provide evidence of an endogeneity problem, F-statistics reported in table 4 reveal that the instruments are not jointly significant in columns lb and 2b. Lagged life expectancy is the instrument doing all the work in columns la and 2a, but its validity as an instrument is itself not beyond question. Thus the iv estimates do not permit us to dismiss the possibil- ity of reverse causality. Further analysis is clearly warranted. Table 5 reports the results when interaction terms and regional controls are included. The table addresses two issues. Is the effect of demographic change on Table 5. Alternative Specifications of the Effects of Population Growth on Economic Growth, 1965-90 Variable la lb 2a 2b 3a 3b Growth of economically 1.94 1.36 2.03 1.43 1.91 1.24 active population, 1965-90 (0.66) (0.55) (0.43) (0.39) (0.45) (0.40) Population growth rate, -1.87 -1.01 -1.88 -1.02 -1.72 -0.78 1965-90 (0.45) (0.41) (0.43) (0.40) (0.49) (0.45) Interaction between growth 0.002 0.01 of economically active (0.07) (0.06) population and institutional quality Interaction between growth -0.12 -0.05 of economically active (0.31) (0.25) population and openness Asia dummy 0.81 0.60 (0.44) (0.35) North America dummy 0.36 0.67 (0.67) (0.55) South America dummy 0.08 0.35 (0.49) (0.42) Europe dummy 1.00 0.53 (0.60) (0.50) Constant -2.43 -19.3 -2.62 -19.6 -2.89 -20.19 (1.35) (4.3) (0.89) (4.3) (1.20) (4.4) Adjusted R2 0.77 0.86 0.77 0.86 0.79 0.86 Note: The dependent variable is the growth rate of real GDP per capita in 1965-90 in purchasing power parity terms. Estimates are from ordinary least squares. The regressions in columns la, 2a, and 3a also include the additional variables in column a of table 2. The regressions in columns lb, 2b, and 3b include those in column b of table 2. The sample size is 78 economies (see the appendix). Standard errors are in parentheses. Source: Authors' calculations. Bloom and Williamson 439 economic performance conditioned by key policy variables such as "institutional quality" and "openness"? And does growth in Asia respond differently to de- mographic and economaic conditions than growth in other regions? In columns la, lb, 2a, and 2b the unconstrained versions of the model are reestimated by including interactions between the growth rate of the economically active popu- lation and a measure of the quality of institutions (Keefer and Knack 1995), on the one hand, and the growth rate of the economically active population and a measure of openness (Sachs and Warner 1995), on the other. Columns 3a and 3b explore whether any regional effect remains. There is no evidence supporting the view that the policy environment influences the linkage between population dynamics and economic performance. Further work will be required to examine the conditions that promote enjoyment of the demographic gift. See, for ex- ample, Higgins and Williamson (1997) or Bloom, Canning, and Malaney (1998). There is some weak evidence that Asia grew faster than the omitted region, Africa, even after controlling for all of these forces, but there is no strong evi- dence to suggest that Asia, after controlling for all of these forces, grew any faster than North America or Europe. (See Bloom and Sachs forthcoming for an analysis of the economic performance of Africa that highlights the importance of demographic and geographic factors.) There is also no evidence of a nonlin- ear relationship, ceteris paribus, between initial income and income growth. We have established that growth of the dependent population slows economic growth. However, does a growing young, dependent population have the same impact as a growing elderly, dependent population? In table 6 we modify the estimation equation by inserting the growth rates of the population under 15 and over 65 in place of the growth rate of the population as a whole. The results sharpen our understanding of how dependent populations contribute to the slow- down. Table 6 reports only the coefficients on the demographic variables. The coefficient on the population under the age of 15 is negative and significant in both specifications: thus an increase of 1 percentage point in growth of the popu- Table 6. Effects of Growth in the Economically Active Population and the Dependent Population on Economic Growth, 1965-90 Variable a b Growth of economically active population, 1965-90 0.82 0.81 (0.21) (0.18) Growth rate of population under 15, 1965-90 -0.71 -0.37 (0.16) (0.16) Growth rate of population over 64, 1965-90 0.11 0.08 (0.10) (0.08) Adjusted R2 0.78 0.86 Note: The dependent variable is the growth rate of real GDP per capita in 1965-90 in putchasing power parity terms. Estimates are from ordinary least squares. Only the coefficients on the demographic variables are reported in the table. The specification used in columns a and b also include the other variables in table 2 in columns a and b, respectively. The sample size is 78 economies (see the appendix). Standard errors are in parencheses. Source: Authors' calculations. 440 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 lation under age 15 is associated with a decrease in growth of GDP per capita of about 0.4 percentage point (column b). In contrast, a small, statistically insig- nificant, but positive, coefficient emerges for the elderly population. We conjec-. ture that because the elderly continue to make important economic contribu- tions by tending the young, by working part-time, and perhaps by continuing to save, they are a smaller net drag than are the very young, whose labor participa- tion and savings rates are trivial in magnitude. Because the elderly are currently a small minority of the total dependent population in Asia (11 percent in 1990), the relationship between the dependent young and growth in GDP per capita dominates, accounting for the negative effect that the dependent population as a whole exerts on the growth rate of GDP per capita. Our findings regarding the economic impact of the demographic transition can be summarized as follows. Economic growth is less rapid when the growth rate of the working-age population falls short of that of the population as a whole (an event that characterized the first phase of East Asia's postwar demo- graphic transition prior to 1970). Economic growth is more rapid when the growth rate of the working-age population exceeds that of the population as a whole (an event that characterized the second phase of East Asia's postwar de- mographic transition after 1970 and overlapped with the economic miracle dur- ing the past quarter century). And economic growth is somewhat less rapid when the growth rate of the working-age population once again falls short of that of the entire population (an aging phenomenon-the graying of East Asia-that will dominate this subregion during the next quarter century). Our interpretations of the econometric results reported in this section are, of course, limited by the validity of the underlying theoretical framework and cor- responding empirical specifications and by the quality of the data. Although the estimates are wholly consistent with the predictions that emerge from the theo- retical model, firmly establishing causality requires further empirical and theo- retical research. For example, it would be useful to conduct analyses that (a) incorporate indicators of the quality of data contained in the Penn World Tables, (b) exploit time-series variation in the data, and (c) explore the use of alternative regressors and functional forms. Working directly with figures on growth in employment in place of growth of the working-age population would also be useful, although this would reduce the sample of countries considerably. Recent extensions of this analysis to the late nineteenth century by Williamson (forth- coming) show quite clearly that other episodes of dramatic growth have been significantly influenced by demographic transitional events, and they need more attention too. IV. EXPLAINING THE EAST ASIAN MIRACLE So far, these results seem consistent with the stylized characterization in fig- ure 6. But they concern only hypothesis testing and statistical significance. What Bloom and Williamson 441 about economic significance? Can our improved understanding of population dynamics explain a significant part of the East Asian miracle? Between 1965 and 1990, the working-age population in East Asia grew 2.39 percent a year, dramatically faster than the 1.58 percent rate for the entire popu- lation and the 0.25 percent rate for the dependent population (table 7). The working-age population also grew faster than the entire population in Southeast Asia, but the differences were almost half of those in East Asia, while in South Asia they were only a quarter. These demographic differences explain at least some of the variation in eco- nomic growth across ihe subregions of Asia between 1965 and 1990. Combining the coefficients from the estimated growth equations in table 5 and the growth rates of the working-age and total population, table 7 indicates that population dynamics can explain between 1.37 and 1.87 percentage points of growth in GDP per capita in East Asia or as much as one-third of the miracle (1.9 / 6.11 = 0.31). If, instead, the miracle is defined as the difference between current growth of GDP per capita (a transitional rate where population dynamics matter) and the esti- mated steady state of 2 percent (when population is also in steady state and has no impact), then population dynamics can explain almost half of the miracle (1.9 / [6.11 - 2] = 0.46). In Southeast Asia, where the decline in fertility took place a little later and the decline in infant mortality was a little less dramatic, population dynamics still account: for 0.9 to 1.8 points of economic growth or, again, as much as half of its (less impressive) miracle (1.8 /3.8 = 0.47). In South Asia, the incipient demographic transition accounts for only 0.4 to 1.3 percentage points of eco- nomic growth, but still as much as three-quarters of a poor growth performance (1.3 / 1.7 = 0.76). The economies that benefited most from these demographic changes were Hong K:ong (China), Malaysia, Republic of Korea, Singapore, Tai- wan (China), and Thailand, all of which are old or new fast-growing tigers. The biggest demographic contribution seems to have been in Singapore, at 1.9 to 2.3 percentage points, but Thailand is close behind, at 1.5 to 2.3 percentage points. It is no coincidence that these tigers attracted most of Krugman's attention when he asserted that the East- Asian miracle was driven mainly by high rates of capital accumulation and labor force growth (Krugman 1994).3 Compared with the rest of the world, East Asia was the largest beneficiary of the population dynamics associated with demographic change. Europe received only a small post-baby boom boost of 0.33 to 0.52 percentage points. Even South America's demographic impact, 0.74 to 1.54 percentage points, was smaller than East Asia's, although the demographic contribution was almost identical to 3. Krugman relied on the findings of Young (1994a, 1994b) and Kim and Lau (1994). In a recent study, however, Hsieh (1998) uses a price-based approach to calculate total factor productivity growth (TFPG) and gets much higher estimates, especially for Singapore, than those resulting from the conventional, quantity-based approaches, He attributes the differences between "primal" (quantity-based) and "dual" (price-based) estimates of TFP to errors in national accounts data on quantities of output and capital. Hsieh argues that factor price data are more reliable, because they can be observed in a marketplace. Table 7. Contribution of Demographic Change to Past Economic Growth, by Region, 1965-90 Average Average Average Average growth rate growth growth rate of growth rate of real GDP rate of economically of dependent Estimated contribution Region per capita population active population population la l b 2a 2b Asia 3.33 2.32 2.76 1.56 1.04 1.64 0.86 0.73 East Asia 6.11 1.58 2.39 0.25 1.71 1.87 1.60 1.37 Southeast Asia 3.80 2.36 2.90 1.66 1.25 1.81 1.07 0.91 46 South Asia 1.71 2.27 2.51 1.95 0.66 1.34 0.48 0.41 tl> Africa 0.97 2.64 2.62 2.92 0.14 1.10 -0.07 -0.06 Europe 2.83 0.53 0.73 0.15 0.43 0.52 0.39 0.33 South America 0.85 2.06 2.50 1.71 1.03 1.54 0.87 0.74 North America 1.61 1.72 2.13 1.11 0.94 1.34 0.81 0.69 Oceania 1.97 1.57 1.89 1.00 0.74 1.14 0.62 0.53 Note: The averages in the first four columns are unweighted country averages. The estimated contribution is created by multiplying the coefficients on the growth rate of economically active population and the population growth rate (table 5) by the regional averages and adding the two for each of the reported specifications. Source: Authors' calculations. Bloom and Williamson 443 that of Asia as a whole. Furthermore, these figures for late twentieth century East Asia are far bigger than for nineteenth century Europe and the New World (Williamson forthcoming). The future will look quite different. Table 8 offers a forecast based on the coefficients of the estimrated growth model and the United Nations (1991) de- mographic projections up to the year 2025. In East Asia, the growth in GDP per capita attributable to clemographic influences is projected to be negative be- tween 1990 and 2025, cleclining from a positive gain of 1.37 to 1.87 percentage points between 1965 and 1990 to a loss of 0.14 to 0.44 percentage point up to 2025. This projected slowing of economic growth of 1.5 to 2.3 percentage points is caused solely by demographic forces. The demographically induced loss in growth is projected to be even larger in some parts of East Asia. If nothing happens to offset them, demographic events will induce a decline of 2.0 to 2.4 percentage points in the growth of GDP per capita in Hong Kong, a decline of 2.5 to 3.0 percentage points in Singapore, a decline of 1.9 to 2.2 percentage points in Korea, and a decline of 0.9 to 1.1 percentage points in Japan. In contrast, South Asia will potentiatlly enjoy a gain of 0.77 to 1.38 percentage points in the growth rate as it leaves the early "burden" stage of the demographic transition and enters the "gift" stage, with the largest potential gains being for Pakistan and Bangladesh. Southeast Asia should register a slightly smaller demographic gift of 0.62 to 1.10 percentage points, with considerable variance across coun- tries. The biggest potential gainer is projected to be the Philippines, and the biggest potential losers are projected to be Malaysia and Thailand. While demographic divergence contributed to Asian economic divergence during the past quarter century, with South Asia falling behind East Asia, the demographic indicators most important to economic performance will converge across Asia from now tco 2025. If our hypotheses survive further scrutiny, demo- graphic convergence should contribute to economic convergence during the next 30 years in the region. The East Asian connection between demographic transi- tion and economic miracle is now being replayed in South Asia and even more so in Southeast Asia. Although demographic divergence contributed to economic divergence in Asia over the past three to four decades, demographic convergence will contribute to economic convergence over the next three to four decades. Figure 7 offers a stylized characterization of those events. V. POSSIBLE CHANNELS OF IMPACT Macro evidence supports the hypothesis that demographic events matter in explaining the East Asian economic miracle. Theory seems to explain the corre- lation, but the hypothesis will be strengthened further if we can show evidence that the channels of impact have been working the way theory predicts. What follows suggests that demographic factors were driving not only the labor force but also a good portion of the high and rising savings and investment rates. Table 8. Contribution of Demographic Change to Future Economic Growth, by Region, 1990-2025 Projected Projected growth Projected growth Estimated contribution growth rate of rate of economically rate of ______________________ Region population active population dependent population l a l b 2a 2b Asia 1.36 1.61 0.99 0.61 0.99 0.50 0.43 East Asia 0.43 0.20 0.87 -0.40 -0.14 -0.44 -0.38 Southeast Asia 1.29 1.66 0.63 0.83 1.10 0.73 0.62 South Asia 1.65 2.11 0.90 1.02 1.38 0.90 0.77 Africa 2.40 2.78 1.88 0.98 1.63 0.73 0.68 -tl Europe 0.17 -0.004 0.48 -0.32 -0.16 -0.34 -0.29 South America 1.50 1.87 0.94 0.82 1.15 0.71 0.60 North America 1.28 1.33 1.21 0.21 0.645 0.11 0.10 Oceania 1.08 0.93 1.37 -0.22 0.24 -0.31 -0.26 Note: The averages in the first three columns are unweighted country averages. The estimated contribution is created by multiplying the coefficients on the growth rate of economically active population and the population growth rate (table 5) by the regional averages and adding the two for each of the reported specifications. Source: Authors' calculations. Bloom and Williamson 445 Figure 7. Stylized Model of Economic Growth and the Demographic Transition in Asia Growth in GDP per capita , ~ ~ .9 a East Asia / Southeast Asia / South Asia Demographic divergence Demogrsphic convergence -* economic divergence? -No- economic convergence? I c. 1950 c. 1995 Future Time The Impact of Demography on Labor Force Growth How much of the fast-growth transition in Asia can be explained by the im- pact of demography on labor inputs? Elsewhere we offer some answers that are only summarized here (Bloom and Williamson 1997: table 6). Our interest, of course, is in labor inputs per person. Growth in labor input per person (working hours per capita, or h' / P) can be separated into three parts: changing hours worked per worker (H I L), changing labor participation rates among persons of working age (L / EAP), and changing shares of the population of working age (EAP / P), the pure demographic effect. Thus per capita hours worked can be decomposed into H / P= (H / L) (L / EAP) (EAP / P). How much of Asia"s economic growth can be explained by a rise in labor input per capita brought about by purely demographic forces? The answer for the period between 1965 and 1975 is very little, but for the period between 1975 and 1990, quite a lot. The rising working-age share served to augment the growth of labor input per capita by about 0.75 percentage point a year between 1975 and 1990. This implies that about 0.4 percentage point of Asia's transitional growth since 1975 (or about a tenth of growth in GDP per capita) can be ex- plained by pure demographics.4 The results are more striking for East Asia. 4. Calculated by multiplying the growth of labor input per capita by the output elasticity of labor. The output elasticity is taken to be 0.56, the average for the 1960s and 1970s of Japan, Hong Kong, India, Republic of Korea, Singapore, and Taiwan (Chenery, Robinson, and Syrquin 1986: table 2-2). 446 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Growth in labor input per capita brought about by pure demography was more than 1.1 percentage points a year, equivalent to 0.6 percentage point of eco- nomic growth. The previous section estimated that demographic forces could account for 1.37 to 1.87 percentage points of the East Asian miracle. The demo- graphic impact on labor input per capita appears to account for about 30 to 40 percent of the total demographic effect. The results for Southeast Asia are more modest-a little more than 0.6 percentage point a year and thus a little less than 0.4 percentage point of economic growth is explained by pure demographics. The results are less striking for South Asia. By itself, the pure demographic effect implies a reduction of 0.5 percentage point in growth of GDP per capita in South Asia compared with East Asia, thus contributing to economic divergence be- tween the two regions since the early 1970s. How much of faster growth in East Asia compared with the industrial countries was due simply to these demographic forces of labor input per capita? The answer is almost 0.5 percentage point, or about four-tenths of the gap between the two. These demographic forces of labor input per capita do not, of course, exhaust all influences on labor supply. Nor do they exhaust all demographic transitional influences on the growth rate. But are they likely to persist in the future? It depends on where in Asia we look. The fall in the pure demographic effect will be a huge 1.13 percentage points a year in East Asia, causing growth to slow by about 0.6 percentage point. In sharp contrast, the fall will raise South Asia's growth rate of GDP per capita, although not by much. The demographic influ- ence on labor inputs will, by itself, foster a convergence of GDP per capita be- tween the poor South and the rich East, favoring growth in the South by 0.7 percentage point. Whether South Asia will actually realize this potential is, of course, a different matter. Will the demographic factors that are slowing economic growth be offset by Asians working harder and by participating more actively in the labor force? A more likely outcome is that Asians will work less hard as their incomes rise, just as workers before them have done in the more industrially mature countries. And fewer prime-age Asians will work, because they will be able to afford ear- lier retirement and will invest more in schooling. In any case, even if Asians work just as hard in the future, this will contribute nothing to growth of labor input per capita; Asians would have to work harder and harder to maintain rates of labor input per capita growth (as opposed to levels) in the future. The Impact of Demography on Savings Almost 40 years ago Coale and Hoover (1958) proposed their famous depen- dency hypothesis. It was based on a simple, but powerful, intuition: rapid popu- lation growth from falling infant and child mortality and high or rising fertility swell the ranks of the dependent young, thereby increasing consumption re- quirements at the expense of savings. Eventually, the youth dependency burden evolves into a young adult glut, and the resulting savings boom contributes to an economic miracle. Finally, the demographic transition is manifested by a large Bloom and Williamson 447 elderly burden, low savings, and a deflation of the miracle. The Coale-Hoover hypothesis suggests that some of the impressive rise in Asian savings rates can be explained by the equally impressive decline in dependency burdens, that some of the difference in savings rates between sluggish South Asia and booming East Asia can be explained by differences in their dependency burdens, and that some of the savings rate gap.s between the two regions should diminish as the youth dependency rate falls in South Asia and the elderly dependency rate rises in East Asia over the next three decades. Empirical tests of the Coale and Hoover (1958) hypothesis have yielded mixed results. Leff's (1969) study appeared to place the youth dependency hypothesis on a solid empirical footing. But later research by Goldberger (1973), Ram (1982), and others failed to confirm the dependency hypothesis and cast doubt on the validity of the empirical methods employed in the earlier studies. Theoretical developments also seemed to undermine the foundations of the dependency hy- pothesis. Tobin's (1967) life-cycle model held that the national savings rate should increase with faster population growth. The reason is simple, at least in that model: faster population growth tilts the age distribution toward young, saving households and away from older, dissaving ones. The representative-agent elabo- ration of Robert Solow's neoclassical growth model pointed in the same direc- tion as Tobin's, with faster population growth resulting in higher savings rates in response to heightened investment demand (Cass 1965, Phelps 1968, and Solow 1956). However, the models just described failed to deal adequately with the dynamics implied by the demographic transition. The "age tilt" in Tobin's steady-state model occurs because the model describes a world restricted to ac- tive adults and retired dependents; it would imply a very different tilt if it also acknowledged youth dependency. Similarly, the neoclassical growth models as- sume fixed labor participation rates and by implication assume no change in the dependency rate, which is exactly what one would assume in a model of steady- state behavior but is inconsistent with the facts of demographic change. In ef- fect, both models sacrifice the rich population dynamics implicit in Coale and Hoover's (1958) predictions about the Asian demographic transition. In the 1980s Fry and Mason (1982) and Mason (1988) addressed the tension between the dependency rate and life-cycle models. These authors developed what they called a "variable rate-of-growth effect" model to link youth depen- dency and national savings rates. Their model rests on the premise that a decline in the youth dependency rate may induce changes in the timing of life-cycle consumption. If consumption is shifted from child-rearing to later, nonchild- rearing stages of the life cycle, aggregate savings rise with a strength that de- pends directly on the growth rate of national income. As a result, the model suggests that the savings rate depends on the product of the youth-dependency ratio and the growth rate of national income (the "growth-tilt effect"), as well as on the dependency ratio itself (the "level effect"). Under the aegis of i:his new model, the dependency hypothesis has enjoyed something of a renaissance. The Coale-Hoover theory has evolved into explicit 448 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 economic models that, now revised, account very well for cross-country savings variations in macro time series. Almost all recent analyses of macro data con- firm the Coale-Hoover effects (Collins 1991; Harrigan 1996; Higgins 1994, 1998a; Kang 1994; Kelley and Schmidt 1995, 1996; Lee, Mason, and Miller 1997; Masson 1990; Taylor 1995; Taylor and Williamson 1994; Webb and Zia 1990; and Williamson 1993), even if they receive weak support at best in house- hold cross-sections (Deaton and Paxson 1997), a difference that future research needs to reconcile.5 Higgins and Williamson (1996, 1997) have estimated the largest macro im- pacts, and what follows uses their results. They estimate the effect of changes in population age distribution on changes in, not levels of, the savings rate as it deviated around the 1950-92 mean. Thus East Asia's savings rate in 1990-92 was 8.4 percentage points above its 1950-92 average because of its transition to a much lighter dependency burden. Similarly, East Asia's savings rate in 1970- 74 was 5.2 percentage points below its 1950-92 average because of its heavy dependency burden at that time. The total demographic swing was an enormous 13.6 percentage points, which would appear to account for the entire rise in the savings rate in East Asia during these 20 years. The figures for Southeast Asia are similar, but not quite so dramatic. Southeast Asia's savings rate was 7.9 percentage points higher in 1990-92 than the average in 1950-92 because its dependency burden was lighter and was 3.6 percentage points lower in 1970- 74, when its burden was heavier. The total demographic swing was 11.5 per- centage points, a smaller figure than for East Asia, but still apparently account- ing for the entire rise in the savings rate in Southeast Asia after 1970. The region with the slowest demographic transition has been South Asia, so its far more modest changes in the savings rate are predictable. To the extent that domestic saving constrains accumulation, falling depen- dency rates have played an important role in East Asia's economic miracle since 1970. Indeed, assuming the increase in investment to have been equal to the increase in savings, and assuming a capital-output ratio of 4, it follows that demographic changes raised accumulation rates in East Asia by 3.4 (13.6 / 4) percentage points, thus augmenting the growth in GDP per capita by an esti- mated 1.5 percentage points. Given that demographic forces raised East Asian growth rates by as much as 1.87 percentage points, about three-quarters of this growth seems to have been due to capital accumulation responses. The figure is 5. Higgins (1998b) also points out that the results from analysis of the micro data and the macro data might not agree, because the data are not consistent. Specifically, "household survey data typically do not correspond to the appropriate concept of personal saving as measured by the national income accounts." He notes, as well, that data on the components of national savings (personal, corporate, and government) are difficult to find. Lee, Mason, and Miller (forthcoming) point out that Deaton and Paxson's (1997) analysis is essentially a comparative steady-state analysis rather than a dynamic analysis, because they assume that "the age profiles of saving and income are invariant to changes in the rate of population growth." Because the effects of the demographic transition depend crucially on dynamic changes, Lee, Mason, and Miller argue that the Deaton and Paxson analysis misses such effects. Bloom and Williamson 449 too high, of course, because of the unsupported assumption that domestic sav- ings fully constrained investment. The Impact of Demography on Investment To the extent that East Asia was able to exploit global capital markets during the past quarter century, the supply of domestic savings is far less relevant than the demand for investraent in determining accumulation performance. As the children of the baby boom became young adults, did the increase in new work- ers imply the need for investment in infrastructure to get them to work, to equip them while at work, and to house them as they moved away from their parents? When Higgins and Williamson (1996,1997) test this augmented Coale-Hoover hypothesis on Asia's past, changing age distributions seem to have had the pre- dicted impact. For East Asia, demographic effects have raised investment shares by 8.8 percentage points since the late 1960s. Using the same assumptions made in the previous section on savings, this implies a rise of 1 percentage point in the growth rate of GDP per capita. In sum, demographic forces appear to have con- tributed 0.6 percentage point to the East Asian miracle via labor inputs per capita and 1 percentage point via capital accumulation per capita-quite consis- tent with the total demographic impact estimated using macro growth equa- tions: 1.37 to 1.87 percentage points. Thus labor force growth responses might account for about one-third of the positive demographic contribution to the miracle (0.6 1 1.9), capital accumulation responses for about one-half (1 / 1.9), and other forces for the small remainder. VI. DIRECTIONS FOR FUTURE RESEARCH AND CONCLUSION The findings presented herein are revisionist, and thus the methods and data used are likely to come under close scrutiny. Future studies will no doubt refine and revise our arguments. Already we can suggest five ways to further this line of research. First, other theoretical approaches might be explored. The standard Solow-Swan model has, after all, been criticized. New ways of thinking about growth could provide other models in which demographic dynamics and eco- nomic growth could be assessed jointly. Second, as further advances in the growth literature define the steady state more effectively, the robustness of our results can be tested and the analysis extended. Third, far more work needs to be done to establish the sources of the demographic transition in Asia after World War II: how much of the transition was due to exogenous factors and how much to endogenous? Fourth, economists and demographers may search for other dra- matic episodes of growth or decline, such as the age of mass migration (Taylor and Williamson 1994 and Williamson forthcoming), the AIDS epidemic (Bloom and Mahal 1997), or the Russian mortality crisis of the early 1990s (Bloom, Canning, and Malaney forthcoming) to see if this model proves equally appli- cable. Fifth, other approaches to understanding simultaneity in the effect of eco- nomic growth on population need to be developed, perhaps by analyzing the 450 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 effects on demographic variables of large and unanticipated variations in in- come per capita such as those caused by oil price shocks in the 1970s and 198 Os. With those caveats, our first major finding is that population dynamics mat- ter in the determination of economic growth. But the overall population growth rate is not the mechanism driving economic performance. Rather, age distribu- tion is the mechanism by which demographic variables affect economic growth. These age distribution effects seem to be purely transitional-although a full transition can take more than 50 years-and operate only when the growth rates of the working-age and dependent populations differ. The demographic transition is induced by an initial decline in infant and child mortality, which swells the youth dependency cohort until fertility rates begin to fall. It thereby helps to trigger an economic transition in which growth performance passes through three phases: initially it is impeded when the youth dependency cohort swells; it is abetted in the next phase about two decades later when the swollen cohort reaches working age; and it is modestly impeded again some decades later when this swollen cohort becomes elderly. The second major finding is that population dynamics account for a substan- tial share of East Asia's economic miracle. Population dynamics account for somewhere between 1.4 and 1.9 percentage points of East Asia's annual growth in GDP per capita from 1965 to 1990, or as much as one-third of observed eco- nomic growth during the period. The economic miracle can, of course, be de- fined differently. Assume that the steady-state growth rate in East Asia is about 2 percent a year, in which case the "miracle" is everything in excess of that, or about 4.1 percent (6.1 percent - 2 percent = 4.1 percent). Under this definition, population dynamics could account for almost half of the miracle. One-third or one-half is certainly not everything, but it suggests that population dynamics may have been the single most important determinant of growth. Within Asia, the evidence also suggests that demographic divergence contributed to economic divergence during the same period. If Southeast and South Asia can use their midphase demographic "gift" in the same way that East Asia did earlier, demo- graphic convergence within Asia may contribute to economic convergence in the coming decades. In any case, although the results presented here certainly do not prove that population dynamics affect economic growth during transitions, they are suffi- ciently robust to justify additional research on the economic-demographic con- nection. That research, we suggest, should focus not just on aggregate popula- tion growth but also on population dynamics as they affect the age distribution. Bloom and Williamson 451 APPENDIX. LIST OF ECONOMIES INCLUDED IN THE DATA SET 1. Botswana 27. Mexico 53. Philippines 2. Cameroon 28. Nicaragua 54. Singapore 3. The Gambia 29. Trinidad and Tobago 55. Sri Lanka 4. Ghana 30. United States 56. Syria 5. Guinea-Bissau 31. Argentina 57. Taiwan (China) 6. Kenya 32. Bolivia 58. Thailand 7. Malawi 33. Brazil 59. Austria 8. Mali 34. Chile 60. Belgium 9. Niger 35. Colombia 61. Denmark 10. Senegal 36. Ecuador 62. Finland 11. Sierra Leone 37. Guyana 63. France 12. South Africa 38. Paraguay 64. Germany, Rep. of 13. Tanzania 39. Peru 65. Greece 14. Tunisia 40. Uruguay 66. Ireland 15. Uganda 41. Venezuela 67. Italy 16. Zaire 42. Bangladesh 68. Netherlands 17. Zambia 43. China 69. Norway 18. Zimbabwe 44. Hong Kong (China) 70. Portugal 19. Canada 45. India 71. Spain 20. Costa Rica 46. Indonesia 72. Sweden 21. Dominican Republic 47. Israel 73. Switzerland 22. El Salvador 48. Japan 74. Turkey 23. Guatemala 49. Jordan 75. United Kingdom 24. Haiti 50. Korea, Fed. Rep. of 76. Australia 25. Honduras 51. Malaysia 77. New Zealand 26. Jamaica 52. Pakistan 78. Papua New Guinea REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Abramovitz, Moses. 1986. "Catching Up, Forging Ahead, and Falling Behind." Journal of Economic History 46(2, June):385-406. Asian Development Bank. 1997. Emerging Asia. Manila. Barlow, Robin. 1994. "Population Growth and Economic Growth: Some More Corre- lations." Population and Development Review 20(March):153-65. Barro, Robert J. 1991. "Economic Growth in a Cross Section of Countries." Quarterly Journal of Economics 106(2, May):407-43. Barro, Robert J., and J. W. Lee. 1994. "Sources of Economic Growth." Carnegie- Rochester Conference Series on Public Policy 40(June):1-46. Barro, Robert J., and Xavier Sala-i-Martin. 1995. Economic Growth. New York: McGraw-Hill. Baumol, William J. 1986. "Productivity Growth, Convergence, and Welfare: What the Long-Run Data Show." American Economic Review 76(December):1072-85. Bloom, David E., David Canning, and Pia N. Malaney. 1998. "Population Dynamics and Economic Growth." Paper presented at the conference on economic aspects of 452 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 demographic transition: the experience of Asian-Pacific countries, Taiwan (China), 19-20 June. Processed. . Forthcoming. "Macroeconomic Consequences of the Russian Mortality Cri- sis." World Development. Bloom, David E., and Richard B. Freeman. 1986. "The Effects of Rapid Population Growth on Labor Supply and Employment in Developing Countries." Population and Development Review 12(3, September):381-14. . 1988. "Economic Development and the Timing and Components of Population Growth." Journal of Policy Modeling 10(1, April):57-82. Bloom, David E., and Ajay S. Mahal. 1997. "Does the AIDS Epidemic Threaten Eco- nomic Growth?" Journal of Econometrics 77(1, March):105-24. Bloom, David E., and Waseem Noor. 1997. "Is an Integrated Regional Labour Market Emerging in East and Southeast Asia?" In Duncan Campbell, Aurelio Parisotto, Anil Verma, and Asma Lateef, eds., Regionalization and Labour Market Interdependence in East and Southeast Asia, pp. 15-39. New York: St. Martin's Press. Bloom, David E., and Jeffrey D. Sachs. Forthcoming. "Geography, Demography, and Economic Growth in Africa." Brookings Papers on Economic Activity. Bloom, David E., and Jeffrey G. Williamson. 1997. "Demographic Change and Human Resource Development." In Asian Development Bank, Emerging Asia. Manila. Boserup, Ester. 1981. Population and Technological Change: A Study of Long-Term Trends. Chicago: University of Chicago Press. Boulier, Bryan. 1986. "Family Planning Programs and Contraceptive Availability: Their Effects on Contraceptive Use and Fertility." In Nancy Birdsall and others, eds., The Effects of Family Planning Programs on Fertility in the Developing World. World Bank Staff Working Paper 677. Washington, D.C.: World Bank. Brander, J. A., and Steve Dowrick. 1994. "The Role of Fertility and Population in Eco- nomic Growth: Empirical Results from Aggregate Cross-National Data." Journal of Population Economics 7(1):1-25. Caldwell, John C., and Bruce K. Caldwell. 1996. "Asia's Demographic Transition." Background paper prepared for the Asian Development Bank, Manila. Processed. Cass, David. 1965. "Optimum Growth in an Aggregative Model of Capital Accumula- tion." Review of Economic Studies 32(3, July):233-40. Chenery, Hollis B., Sherman Robinson, and Moshe Syrquin. 1986. Industrialization and Growth: A Comparative Study. New York: Oxford University Press. Coale, Ansley J. 1986. "Population Trends and Economic Development." In Jane A. Menken, ed., World Population and U. S. Policy: The Choices Ahead. New York: Norton. Coale, Ansley J., and Edgar Hoover. 1958. Population Growth and Economic Develop- ment in Low-Income Countries. Princeton, N.J.: Princeton University Press. Coale, Ansley J., and Susan C. Watkins. 1986. The Decline of Fertility in Europe. Princeton, N.J.: Princeton University Press. Collins, Susan. 1991. "Saving Behavior in Ten Developing Countries." In B. Douglas Bernheim and John B. Shoven, eds., National Saving and Economic Performance. Chicago: University of Chicago Press. Deaton, Angus, and Christina Paxson. 1997. "The Effects of Economic and Popula- tion Growth on National Saving and Inequality." Demography 34(1, February):97- 114. Bloom and Williamson 453 Ehrlich, Paul R. 1968. The Population Bomb. New York: Ballantine. Ehrlich, Isaac, and Francis T. Lui. 1991. "Intergenerational Trade, Longevity, and Eco- nomic Growth." Journal of Political Economy 99(October):1029-59. Feeney, Griffith, and Andrew Mason. 1997. "Population in Eagt Asia." Paper presented at the conference on population and the Asian economic miracle, East-West Center, Honolulu, Hawaii, 7-10 January. Processed. Fogel, Robert W. 1994. "Economic Growth, Population Theory, and Physiology: The Bearing of Long-Terra Processes on the Making of Economic Policy." American Eco- nomic Review 84(3, june):369-95. Fry, Maxwell, and And:rew Mason. 1982. "The Variable Rate-of-Growth Effect in the Life-Cycle Saving Model: Children, Capital Inflows, Interest, and Growth in a New Specification of the Life-Cycle Model Applied to Seven Asian Developing Countries." Economic Inquiry 20(3, July):426-42. Gertler, Paul, and John Molyneaux. 1994. "How Economic Development and Family Planning Programs Combined to Reduce Indonesian Fertility." Demography 31(1, February):33-63. Goldberger, Arthur. 1973. "Dependency Rates and Savings Rates: Further Comment." American Economic Review 63(1, March):232-33. Harrigan, Frank. 1996 "Saving Transitions in Southeast Asia." Asian Development Bank, Manila. Processed. Hausman, Jerry. 1978. "Specification Tests in Econometrics." Econometrica 46(6, No- vember):1251-71. Higgins, Matthew. 1994. "The Demographic Determinants of Savings, Investment, and International Capital Flows." Ph.D. diss., Harvard University, Cambridge, Mass. 1998a. "Demography, National Savings, and International Capital Flows." In- ternational Economic Review 39(2, May):343-69. 1998b. Processed. Higgins, Matthew, and Jeffrey G. Williamson. 1996. "Asian Demography and Foreign Capital Dependence." NBER Working Paper 5560. National Bureau of Economic Re- search, Cambridge, Mass. Processed. . 1997. "Age Structure Dynamics in Asia and Dependence on Foreign Capital." Population and Development Review 23(2, June):261-93. Hsieh, Chang-Tai. 1998. "What Explains the Industrial Revolution in East Asia? Evi- dence from Factor Markets." Department of Economics, University of California, Berkeley. Processed. Kang, Kenneth. 1994. "Why Did Koreans Save So Little and Why Do They Now Save So Much?" International Economic Journal 8(4, winter):99-111. Keefer, Philip, and Stephen Knack. 1995. "Institutions and Economic Performance: Cross- Country Tests Using Alternative Institutional Measures." Economics and Politics 7(3, November):207-27. Kelley, Allen C. 1988. "Economic Consequences of Population Change in the Third World." Journal of Economic Literature 26(4, December):1685-728. Kelley, Allen C., and Robert M. Schmidt. 1995. "Aggregate Population and Economic Growth Correlations: The Role of the Components of Demographic Change." De- mography 32(4):54-3-55. . 1996. "Savings, Dependency, and Development." Journal of Population Eco- nomics 9(4, November):365-86. 454 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Kim, Jong-Il, and Lawrence Lau. 1994. "The Sources of Economic Growth of the East Asian Newly Industrialized Countries." Journal of the Japanese and International Economies 8(3, September):235-71. Krugman, Paul. 1994. "The Myth of Asia's Miracle." Foreign Affairs 73(6, November- December):62-78. Kuznets, Simon. 1967. "Population and Economic Growth." Proceedings of the Ameri- can Philosophical Society 111 (June):170-93. Lee, Ronald, Andrew Mason, and Timothy Miller. 1997. "Saving, Wealth, and the De- mographic Transition in East Asia." Paper presented at the conference on population and the Asian economic miracle, East-West Center, Honolulu, Hawaii, 7-10 Janu- ary. Processed. . Forthcoming. Response in Andrew Mason, Tom Merrick, and Paul Shaw, eds., Macroeconomics and Population Momentum. Washington, D.C.: World Bank. Leff, Nathaniel. 1969. "Dependency Rates and Savings Rates." American Economic Review 59(5, December):886-96. Livi-Bacci, Massimo. 1992. A Concise History of World Population. Cambridge, Mass.: Blackwell. Mankiw, N. Greg, David Romer, and David N. Weil. 1992. "A Contribution to the Empirics of Economic Growth." Quarterly Journal of Economics 107(2, May):407- 37. Mason, Andrew. 1988. "Saving, Economic Growth, and Demographic Change." Popu- lation and Development Review 14(1, March):113-44. Masson, Paul. 1990. "Long-Term Macroeconomic Effects of Aging Populations." Fi- nance and Development 27(2, June):6-9. Meltzer, David. 1995. "Mortality Decline, the Demographic Transition, and Economic Growth." Brigham and Women's Hospital and National Bureau of Economic Re- search, Cambridge, Mass. Processed. Phelps, E. S. 1968. "Population Increase." Canadian Journal of Economics 1(3, Au- gust):497-518. Pritchett, Lant H. 1994. "Desired Fertility and the Impact of Population Policies." Popu- lation and Development Review 20(1, March):1-55. Radelet, Steven, Jeffrey D. Sachs, and Jong-Wha Lee. 1997. Economic Growth and Trans- formation. Development Discussion Paper 609. Cambridge, Mass.: Harvard Institute for International Development. Ram, Rati. 1982. "Dependency Rates and Aggregate Savings: A New International Cross- Section Study." American Economic Review 72(3, June):537-44. Sachs, Jeffrey D., and Andrew Warner. 1995. "Economic Reform and the Process of Global Integration." Brookings Papers on Economic Activity 1:1-118. Sanderson, Warren C., and Jee Peng Tan. 1995. Population in Asia. World Bank Re- gional and Sectoral Studies. Washington, D.C.: World Bank. Simon, Julian. 1981. The Ultimate Resource. Princeton, N.J.: Princeton University Press. Solow, Robert. 1956. "A Contribution to the Theory of Economic Growth." Quarterly Journal of Economics 70(1, February):65-94. Taylor, Alan. 1995. "Debt, Dependence, and the Demographic Transition: Latin America into the Next Century." World Development 23(5, May):869-79. Bloom and Williamson 455 Taylor, Alan, and Jeffrey G. Williamson. 1994. "Capital Flows to the New World as an Intergenerational Transfer." Journal of Political Economy 102(2, April):348-69. Tobin, James. 1967. "Life-Cycle Savings and Balanced Economic Growth." In William Fellner, ed., Ten Essays in the Tradition of Irving Fischer. New York: Wiley Press. United Nations. 1991. Global Estimates and Projections of Populations by Age and Sex. New York. 1992. The Sex and Age Distributions of Population. New York. 1994. "World Population Prospects 1950-2050. (The 1994 revision)." Demo- graphic Indicators 15950-2050. (The 1994 revision): electronic data. New York. Webb, Steven, and Heidi Zia. 1990. "Lower Birth Rates = Higher Saving in LDCS." Finance and Development 27(2, June):12-14. Williamson, Jeffrey G. 1990. Coping with City Growth during the British Industrial Revolution. Cambridge, U.K.: Cambridge University Press. .1993. "Human Capital Deepening, Inequality, and Demographic Events Along the Asia-Pacific Rim." In Gavin Jones, Naohiro Ogawa, and Jeffrey G. Williamson, eds., Human Resources and Development along the Asia-Pacific Rim. New York: Oxford University Press. . 1998. "Growth, Distribution, and Demography: Some Lessons from History." Explorations in Economic History 35(3, July):241-71. World Bank. 1993. The East Asian Miracle: Economic Growth and Public Policy. New York: Oxford University Press. Young, Alwyn. 1994a. "Lessons from the East Asian NICS: A Contrarian View." Euro- pean Economic Review 38(3-4, April):964-73. 1994b. "The Tyranny of Numbers: Confronting the Statistical Realities of Asian Growth Experience." NBER Working Paper 4680. National Bureau of Economic Re- search, Cambridge, Mvass. Processed. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 457-82 The Bolivian Social Investment Fund: An Analysis of Baseline Data for Impact Evaluation Menno Pradhan, Laura Rawlings, and Geert Ridder The Bolivian Social Investment Fund (SIF) is a financial institution that promotes sus- tainable investments in the social sectors, principally in the areas of health, education, and sanitation. This article shows how to use preintervention data collected for evalu- ating the SIF to improve the targeting of a program, to test the quality of the evaluation design, and to define corrective measures if necessary. It finds that among SIF interven- tions the benefits in education are distributed relatively equally over the population, while the investments in health and sanitation favor better-off communities. The article contributes to the methods used to evaluate social investment funds and similar programs. It compares two types of evaluation designs to assess social invest- ment fund interventions in the education sector. The authors demonstrate that a simple matched-comparison design introduces a bias in the estimate of the program effect, whereas an experimental design based on random assignment does not. With preintervention data, the analyst can select a quasi or indirect experiment, where the choice of the indirect. experiment coincides with the selection of valid instrumental variables. The availability of preintervention data makes it possible to compare the two types of evaluation designs as well as to test the validity of the instruments and to determine the loss of efficiency due to the use of quasi-experimental techniques instead of random treatment assignment. The Bolivian Social Investment Fund (SIF) is a financial institution designed to promote sustainable investments in the social sectors, notably in the areas of health, education, and sanitation. The SIF cofinances initiatives providing infra- structure, training, and equipment by making funds available to requesting agen- cies that then subcontract the implementation of the projects. The SIF is charac- terized by rapid disbursement, institutional efficiency, and a demand-driven approach, and it is among the first of its kind in the world. The SIF has been a Menno Pradhan is with the Department of Economics at the Vrije Universiteit in Amsterdam, Laura Rawlings is with the Development Economics Research Group at the World Bank, and Geert Ridder is with the Department of Economics at The Johns Hopkins University. The analysis presented in this article was made possible by a grant from the Dutch Trust Fund at the World Bank. The collection of the baseline data was financed by the Social Investment Fund Project (1991-94). This study is part of the World Bank-financed tesearch project Evaluation of Social Sector Investments, which has provided support to the Bolivian Social Investment Fund for the design, application, and analysis of the impact evaluation. The authors wish to thank John Newman and Ramiro Coa for their support and the referees for their comments. © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 457 458 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 key player in funding social sector projects in Bolivia. Since its inception in 1991, the SlF has provided $187 million worth of financing to more than 3,000 projects.' It has caught the attention of policymakers worldwide, and similar funds have been introduced in Africa, Asia, and other parts of Latin America. In Latin America and the Caribbean alone, 16 SIFs have been created (Glaessner and others 1994). Despite the growing popularity of social funds, few have been subject to em- pirical investigation to assess the effects of these interventions. This question is of central concern to the World Bank, which has been a principal supporter of social funds, and to policymakers in the social sectors more broadly. The Boliv- ian SIF provides an opportunity to assess this impact, a task that will be com- pleted in 1999 when the follow-up data will be available and analyzed. This article makes an initial contribution by exploring a special feature of the evalu- ation made possible by the presence of baseline data and the use of different evaluation methodologies. For the evaluation of the sIF, preintervention (baseline) data were collected in select rural provinces in 1993, and a follow-up survey is being completed in 1998. This article is based on an analysis of the baseline data, which were collected in two stages. In June 1993, a survey was conducted in five provinces in the Chaco region.2 In October and November 1993, this survey was extended to 17 other provinces in select rural areas throughout Bolivia, hereafter named the Resto Rural region.3 The survey gathered information on the communities, fa- cilities, and individuals expected to benefit from projects financed by the SIF. The survey also collected information for control groups not expected to receive SIF interventions. In the education sector, two methods were used to construct the control groups. In the Chaco region, an experimental randomized design was used, while in the Resto Rural region, the control group was constructed by matching on observed characteristics. A matched comparison was also used in the sanitation sector, whereas the evaluation in the health sector was based on a reflexive comparison before and after the SIF intervention. Analyzing the baseline data before collecting the follow-up data is useful for two reasons. First, information on the facilities that will be upgraded and benefit incidence analysis allow for midcourse corrections in implementing the SIF projects, particularly with respect to targeting. Second, in the case of experimental designs or matched-comparison designs, the evaluation methodology can be tested by as- 1. As of 31 March 1998 (the Bolivian Social Investment Fund, Program and Policy Support Unit, La Paz, May 1998). 2. The provinces in the Chaco region are O'Connor and Gran Chaco in the department of Tarija; Cordillera in the department of Santa Cruz; and Luis Calvo and Hernando Siles in the department of Chuquisaca. 3. The provinces in the Resto Rural region are Bernando Saaverdra, Camacho, Murieca, and Franz Tamayo in the department of La Paz; Capinota, Tapacari, Quillacollo, and Arque in the department of Cochabamba; Saucari, Cercado, Carangas, Sur Carangas, and Nor Carangas in the department of Oruro; and Linares, Nor Chicas, Saavedra, and Modesto Omiste in the department of Potosi. Pradhan, Rawlings, and Ridder 4S9 sessing the comparability of the treatment and comparison or control group. Noncomparability may have implications for the statistical methods used to de- termine the impact and the required sample size of the follow-up survey. This article focuses on two issues. First, we look at the target population of the SIF in education, health, and sanitation projects and compare it with the total population in the Chaco and Resto Rural regions. We investigate the effects of the institutional design of the SIF on its targeting. The demand-driven approach of the SIF does not necessarily guarantee that the poorest in society will benefit from the investments. Second, we take advantage of the two types of evaluation designs in the education component of the SIF to investigate the adequacy of each of the evaluation designs employed: randomization and matched comparison. In the evaluation literature, randomized designs are widely regarded as the most methodologically robust evaluation approach. The process of randomiza- tion ensures that before the interventions take place the treatment and control groups are statistically equivalent, on average, with respect to all characteristics, observed and unobserved. (For a general discussion of experimental and quasi- experimental methods for causal inference, see Holland 1986.) Therefore, any differences observed after the intervention takes place can be causally attributed to the effect of the intervention itself (Grossman 1994). The standard alternative, in the absence of randomized assignment, is matched comparison using a nonrandom process to select a comparison group that re- sembles the treatment group already assigned to receive an intervention. The matching is obviously restricted to (readily) observable characteristics, and match- ing does not guarantee that the comparability extends to unobserved or unob- servable characteristics. Indeed, Fraker and Mayard (1984) and LaLonde (1986) have shown that matching on observables may induce a large bias in estimates of program effects. Researchers have made considerable progress in the improve- ment of matching techniques, in particular by concentrating on the selection probability (or propensity score) as the relevant quantity for assessing the qual- ity of the match (Rosenbaum and Rubin 1983, Heckman and others 1998, and Dehejia and Wahba 1995). These methods require a good understanding of the selection process and high-quality data to identify and use the variables that are relevant in the matching, that is, the variables for which the matching makes the treatment and control groups comparable. In evaluating the Bolivian SIF, we cannot benefit from experience with non- random selection for similar programs, nor do we have access to the type of high-quality data often available in other countries, such as for evaluating job training programs in the United States. Under these circumstances, an indirect experiment is the best ;alternative. In an indirect experiment, some feature of the selection process for the intervention is assumed to provide exogenous variation in the selection probability. If this feature can be measured by a variable, then this variable is a valid instrument and can be used as an instrumental variable (Iv) estimator of the program effect (Moffitt 1991). 460 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 An instrumental variable satisfies two conditions: first, the variable affects the probability of selection, and second, it does not affect the outcome or re- sponse variable that is used to evaluate the program. The first condition implies that the instrumental variable must be included among the variables that are used to match treatment and control groups. An important reason for preferring the Iv method over matching is that the set of matching variables, and in particu- lar the degree of comparability achieved by the set of observed variables, is criti- cally important for the quality of the matching estimator. For the iv approach, the instrument only needs to be exogenous (the second condition); it does not need to balance the treatment and control group. The amount of exogenous variation in the selection probability induced by the instrument is important. A "weak" instrument may result in an Iv estimator with poor small-sample perfor- mance. With randomized assignment, all variation in selection is exogenous. Iv and matched-comparison estimates may be similar (Friedlander and Robins 1995), but there is no reason to expect this to be true in general. There is still considerable controversy over the interpretation of the iv estima- tor when the program effect varies in the population and, in particular, when that variation cannot be explained by observable characteristics of the popula- tion units (Imbens and Angrist 1994 and Heckman 1997). We are not concerned with this discussion but simply assume that all effects of heterogeneity can be captured by observable (and observed) variables. In general, the iv method is based on untestable assumptions. In particular, the second condition usually cannot be verified empirically. If the instrumental variable and the response or outcome variable are uncorrelated, then we do not know whether this is due to no program effect or to lack of induced variation in the selection probability. If the correlation is nonzero, then we do not know whether there is a direct effect on the outcome variable. (Additional functional form assumptions as in the normal sample selection model help to make the distinction, but these assumptions are untestable.) Hence, the choice of an in- strument has to be justified with ad hoc arguments, which may be more or less convincing. Sometimes the instrument is obtained by randomization, as in ran- domized experiments with noncompliance (see, for example, Angrist 1990). With preintervention data, we can test the second condition under much weaker assumptions. If we assume that selection for the program does not change preintervention behavior-for example, if the preintervention response is mea- sured before selection for the program is made public-then a zero (partial) correlation between preintervention response and instrument is evidence that the instrument is valid. We shall use this procedure to identify valid instruments for evaluating the Bolivian SIF. Of course, with only preintervention data, we cannot assess program impact. For this, we need the postintervention data that will be collected. However, as we shall argue, we can compare the efficiency of the different designs and predict the accuracy of the estimated program effects. This article does not contain an evaluation of the SIF. Such an evaluation is not possible using only preintervention data. A full evaluation of the SIF invest- Pradhan, Rawlings, and Ridder 461 ments in education, health, and sanitation will be carried out in 1998-99 once the follow-up data are available. This article contributes to the methods that can be used to evaluate the SIF and similar programs. We show that preintervention data contain valuable information for choosing an appropriate evaluation method. Section I presents the data and sample design. Section II describes the SIF in greater detail and looks at the targeting mechanisms used by the SIF. Section III compares the evaluation designs for the education component. Section IV pro- poses the instrumental variable method that eliminates the bias in the matched- comparison design and studies the loss of efficiency due to this method. Section V concludes. I. DATA The SIF impact evaluation considers investment projects in health care, educa- tion, and sanitation. The data collected for the impact evaluation were based on surveys applied to both the institutions that receive funding (schools and health centers) and the households and communities that benefit from the investments. Similar data were also collected from comparison institutions and households. The household survey gathered information on a range of household charac- teristics including consumption, access to basic services, and each household member's health and education status. The household survey consisted of three subsamples. The first was a random sample of all households in the Chaco and Resto Rural regions, the second was a sample of households that live near the schools in the treatment or control group for the education component, and the third consisted of households that will benefit from the sanitation component. The surveys can be merged easily. Por example, each school has a unique code that is recorded in the household survey if a child attends that school. The sur- veys for the Chaco and Resto Rural regions differed slightly. The baseline data collected in the Resto Rural region are more extensive because shortcomings discovered in the surveys in the Chaco region, which were conducted first, were corrected. Sample sizes are given in table 1. The health facility survey gathered information on the quality of infrastruc- ture, staffing, and visits to the center. Because the SIF planned to intervene in all health centers in the Chaco and Resto Rural regions, all were included in the survey. The survey distinguished between health clinics at the sector, area, and district levels. Sector health clinics are typically very small, providing basic health care. Area health clinics provide more sophisticated care and serve a larger geo- graphical region. District health clinics are hospitals, the largest type of facility. The larger the health clinic, the more detailed the questionnaire that was ad- ministered. The questlionnaires are, however, comparable and collected similar types of information on infrastructure, equipment, the availability of medicines, staffing, and the services provided. The school survey used two questionnaires, one for the director and one for each teacher separately. It gathered information on infrastructure, equipment, 462 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 1. Sample Sizes in the Baseline Survey, by Region in Bolivia, 1993 Category Chaco Resto Rural Health centers Sector 82 119 Area 16 24 District 3 4 Total 101 147 Schools Treatment group 35 37 Control group 37 33 Total 72 70 Households Random sample 2,029 2,138 Education component 995 902 Sanitation component 666 569 Total 3,670 3,609 Source: Authors' calculations based on survey data. teaching methods, and dropout and repetition rates of students. For the Chaco region, the sampling frame consisted of all primary and secondary schools that qualified for SIF interventions and were subject to the random-selection process. As explained in section III, for equity purposes, the worst-off schools were all selected to receive active promotion for an SIF project and none of the best-off schools received active promotion. Only those schools in the middle of the qual- ity distribution were subject to random selection because the financing did not allow for all schools to be reached. For the Resto Rural region, the sampling frame consisted of schools already designated to receive SIF interventions. The sample was augmented by a comparison group of schools not receiving an SIF intervention. Section III provides more detail on the requirements for eligibility for schools to receive SIF investments and the construction of comparison groups. The community survey collected data from community leaders on a range of topics, including the quality of the infrastructure, the distance to facilities, and the presence of local organizations. II. TARGETING OF SIF INTERVENTIONS The SIF has traditionally funded, but not executed, project proposals received from the private, public, and not-for-profit sectors. The SIF is a demand-driven institution because it does not initiate projects but responds to outside initiatives by providing cofinancing for investments in infrastructure, equipment, and train- ing. The cofinancing provided by the SIF generally accounts for approximately 80 percent of project costs, and the requesting institution provides the remain- ing 20 percent. Regional SIF offices assist communities in preparing proposals. The decision on whether to fund a project is made at the SIF central offices in La Pradhan, Rawlings, and Ridder 463 Paz. The final outcome thus depends on both the preferences and capabilities of the local communities, in particular the local authorities and local nongovern- mental organizations (NGOS), with respect to preparing and cofinancing projects and on centrally defined targets (see also Newman, Grosh, and Jorgensen 1992). Central objectives, such as targeting the poor, may not always be reflected in the final outcome of the program. In particular, the project approval process has historically favored more well-organized groups that have access to counterpart financing and are rarely found in poorer areas. This is an inherent conflict of targeted, demand-driven projects. With the introduction of the Ley de Participaci6n Popular in 1994, the preferences of the local population are be- coming more influential. Under this law, a proportion of the government budget is allocated directly to communities, municipal elections ensure accountability of the leaders, and cojmmunities are given discretion over budget allocations. An analysis of the SIF's targeting mechanisms using the available preintervention data cannot deal with behavioral responses that may result from SIF interven- tions. Baseline data can only be used to characterize households that are using the facilities in which the SIF is planning to invest. Of course, changes in house- hold behavior as a result of changes in the supply of public services may be an important factor in determining the net impact of the project (Jimenez 1995). With the current data, however, we cannot deal with these effects. Figure 1 shows the nonparametric estimate of the density function of log per capita consumption based on the random sample of all households.4 The figure is included to enable a comparison with the other figures relative to the distribu- tion of consumption in the population and can be used as a guide for assessing the levels and concentration of poverty in the population. Figure 2 shows a nonparametric regression estimate of a function that is pro- portional to the conditional probability that a child in a household attends a school that will receive SIF funding, given the log per capita household consump- tion. The estimate is based on the equality Pr[SIF school I log(consumption)] = f [log(consumption) I SIF school] Pr(SIF school) f [log(consumption)] where consumption is per capita household consumption. The probability that a child in the household attends a school that will receive support from the SIF is not known, but from the random sample of all households and the sample of households near schools that have been selected to receive SIF projects, we can estimate the densities in the numerator and denominator of the first term on the right-hand side of the equation. The estimated ratio is depicted in figure 2. Figure 3 presents a nonparametric regression estimate of the probability that an individual visited a government health clinic in the month before the survey date as a function of log per capita household consumption. This estimate is 4. The nonparametric density and regression estimates are kernel smooth based on a Gaussian kernel. The bandwidth was chosen as suggested by Silverman (1986: 45). 464 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Figure 1. Density Function of the Log per Capita Consumption ofHousehold, in the Chaco and Resto Rural Regions of Bolivia, 1993 Density 0.5 0.4 0.3 0.2 0.1 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 Log per capita consumption Source: Authors' calculations based on survey data. based on the random sample of all households. The households that have a high probability of visiting a government health clinic are more likely to benefit from the SIF investments. The nonparametric regressions estimate in figure 4 was obtained in the same way as for figure 2; it indicates the probability that the household lives in a commtnity that is selected for an investment in sanitation. Figure 2. Targeting of SIFInterventions in Education in Bolivia Relative probability 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 Log per capita consumption Note: Values are nonparametric estimates of the probability of benefit given household welfare. See the text for details on the calculation of the estimates. Source: Authors calculations based on survey data. Pradban, Rawlings, and Ridder 465 Figure 3. Targeting cf SIF Interventions in Health in Bolivia Probability 0.25 0.20 0.15 0.10 0.05 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 Log per capita consumption Note: Values are nonparametric estimates of the probability of benefit given household welfare. See the text for details on the calculation of the estimates. Source: Authors' calculations based on survey data. The figures reveal that there is no relation between household welfare, as measured by log consumption per head, and benefits from SIF interventions in education. However, SIF investments in health facilities and basic sanitation benefit households that are relatively better off. District, area, and sector health clinics are not the only providers of medical care. Private doctors and particularly traditional healers are extensively con- sulted for medical care. Table 2 was constructed to determine who will benefit Figure 4. Targeting qf SIF Interventions in Sanitation in Bolivia Relative probability 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 o -s i I I I I 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 Log per capita consumption Note: Values are nonparametric estimates of the probability of benefit given household welfare. See the text for details on the calculation of the estimates. Source: Authors' calculations based on survey data. Table 2. Health Status and Actual Health Care Consumption in the Month Preceding the Survey in Bolivia, 1993 (percent) If sought medical care, went to Quartile level of per Reported Sougbt Government Private doctor Traditional If went to government clinic, type capita consumption3 good health medical care clinic or midwife careb Sector Area District 1-poor 84 44 44 10 47 48 33 19 2 83 S5 58 10 32 42 35 23 3 80 58 59 20 21 38 34 28 4-rich 75 58 63 22 15 38 31 31 All 81 54 57 16 27 40 33 27 a. Quartiles correspond to the following levels of per capita yearly consumption: quartile 1, up to Bs746; quartile 2, Bs747-Bsl,224; quartile 3, Bs1,225- Bs2,158; quartile 4, above Bs2,1S8. Bolivia's currency is the boliviano. b. Includes traditional healers, neighbors, family, and others without formal medical training. Source: Authors' calculations based on survey data. Pradhan, Rawlings, and Ridder 467 from the SIF investments in health centers. To this end, the population has been divided into four groups ranging from richer to poorer, depending on per capita consumption. The table presents results on the health status of future SIF ben- eficiaries and their use of health services prior to the SIF intervention. Health status is measured by asking the respondent whether he or she was in good health during the past month. Richer households tend to report a health con- dition that is worse than that of poorer households. However, this self- reported condition does not necessarily reflect the actual health status of the respondent, because richer people may be prepared to admit to being in bad health more readily than poorer people. Being in bad health is often associated with "not being able to work" or "seeking medical care,'" which is more readily affordable for the rich. The data in table 2 show that, if ill, rich households seek health care more fre- quently and go more often to government health clinics. Poor households seek medical care less frequently and visit traditional healers more often. If the poor visit government health clinics, they mostly go to small (sector) health clin- ics. The results suggest that if the SiF wants to target investments in health facili- ties to poor communities, those investments should be concentrated in sector- and community-level health clinics. The results also show a need for information and outreach programs to encourage poor households to seek medical care when ill and to visit public health-care providers. With respect to investments in sanitation, we find that households that will benefit from SIF investments in basic sanitation already have better sanitation facilities than most of the rural population. For example, 47 percent of the tar- geted households have access to piped water compared with only 26 percent of rural households in general. However, this is not necessarily inconsistent with SIF policy. Investments in basic sanitation are made most effectively in areas that already have access to a water system and are located in populated areas so that the project is able to take advantage of economies of scale. Constructing these facilities in remote rural areas may lead to better targeting but would be ex- tremely costly relative to the per capita benefits achieved. The results in this section show that evaluation of the impact of the SIF invest- ments may be problematic. The selective targeting of investments in health care and sanitation biases a direct comparison of beneficiaries and nonbeneficiaries, because the latter were already worse off before the intervention. An impact evaluation should take this selectivity into account. Under the assumption that changes that occur in the time between the baseline survey and the date of mea- surement of the response variables affect all relevant units (facilities, regions, and communities) in the same way, we can estimate the impact of the SIF with a difference-in-differences estimator, as in Ashenfelter (1978). A direct compari- son may be possible for the education component. In the next section, we inves- tigate whether an unbiased estimate of the effect of SIF investments in education can be obtained. 468 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 III. RANDOMIZATION AND MATCHED COMPARISON IN EDUCATION The German Institute for Reconstruction and Development had earmarked funding for education interventions in the Chaco region in 1991, yet the process for promoting SIF interventions in select communities had not been initiated. This situation provided an opportunity for assessing the need in schools for an SIF intervention and for applying a random-selection process. The school quality index for the Chaco region assigns each school a score from 0 to 9 based on the sum of five school infrastructure and equipment indicators: electric lights (1 if present, 0 if not); sewerage (2 if present, 0 if not); water source (4 if present, 0 if not); at least one desk per student (1 if so, 0 if not); and at least 1.05 square meters of space per student (1 if so, 0 if not). Schools were ranked according to this index, with a higher value reflecting more resources. Only schools with an index below a particular cutoff value were eligible for an SIF intervention, and the worst-off schools were automatically designated to receive SIF-financed promotions and investments because of their extremely low quality. Both the worst-off and best-off schools were excluded from the randomization and sample, so that the restriction to eligible schools implies that the effects of the SIF cannot be generalized to all schools. The eligible schools in the middle of the quality distribution were allocated to the treatment or control group at random, creating the basis for an experimental evaluation design. In the Resto Rural region schools had already been selected for SIF interven- tions, making randomization impossible. Therefore, treatment schools were sampled from the list of all schools designated for SIF interventions, and a com- parison group was constructed by matching the sampled schools to non-SIF schools based on several observable characteristics. The matching procedure used in the Resto Rural region consisted of two steps. First, using the 1992 census, cantons in which the treatment schools were lo- cated were matched to similar cantons with respect to population size and distri- bution by age, education level, gender, infant mortality rate, language, and lit- eracy rate. Second, once the control cantons were identified, control schools were selected from the cantons to match the treatment schools with respect to the school quality index as developed for the Chaco region. No other data on schools, households, or communities were available for use in the matching ex- ercise given the paucity of data in rural Bolivia. Thus two distinct evaluation designs were used in the two regions: a classical experimental design with randomized assignment in the Chaco region and a matched-comparison design in the Resto Rural region. In recent years, there has been a controversy over the validity of various evaluation designs. As illustrated by Grossman (1994) in a review of the theory and practice of evaluation re- search, much of the controversy over evaluation techniques can be traced back to a seminal study by Fraker and Mayard (1984). They assess the impact of the National Supported Work (NSw) demonstration (a major employment program Pradhan, Rawlings, and Ridder 469 in the United States) using both a matched-comparison methodology and an experimental design made possible by the random assignment present in the NSW program. Fraker and Mayard calculate the impact estimates of NSW using both types of methodologies and find that the estimates based on the matched- comparison methodology do not come close to the impact estimated using the experimental design. In his own review of the NSW data, LaLonde (1986) sup- ports this conclusion arguing that matched-comparison designs can be severely biased and that randornized assignment is the only design that can produce un- biased estimates of the effect of some intervention. In a reaction, Heckman and Hotz (1989) argue that careful modeling of the selection effect can remove most of this bias. However, because of the uncertainties in this approach, it seems that a more secure bas.is for identification of the intervention effect is needed. The improved matching estimators of Dehejia and Wahba (1995) and Heckman and others (1998) cannot be used, because the selection for the SIF is not suffi- ciently understood to identify the relevant variables in the matching procedure and the current data are limited in their description of the project promotion and selection process. Hience, the most promising approach is to look for a vari- able that affects the selection into the treatment group, but not the relevant response variable. Such a variable is a valid instrument, and if such an instru- ment is available it can be used to obtain an unbiased, be it less efficient, esti- mate of the intervention effect (Angrist 1990 and Imbens and Angrist 1994). An instrument corresponds to an indirect experiment as opposed to a direct experi- ment due to randomizat:ion. Indirect experiments may be the only available evalu- ation design in many instances because, as in the Resto Rural region, random assignment often is not politically feasible and the information required for an unbiased matched comparison is rarely available. In this section we clheck whether the assignment done in the Chaco region was indeed random, and we test whether the matching done in the Resto Rural region was selective and will give a biased estimate of the effect of the SIF with postintervention data. We test for random assignment because of the need to verify that the program administrators who were involved in selecting schools did not alter the planned evaluation design. Because the SIF cofinances invest- ments in schools, it is natural to take the school as the unit in the evaluation. The goal of the investment is to improve the quality of education, and hence we need variables that measure this quality. An obvious choice would be the (average) score(s) on a standardized test of the achievement of pupils, but unfortunately such a test could not be administered during the baseline because of a major reform in the education sector. For that reason, we use two indirect measure- ments of education quality: repetition and dropout rates. The repetition rate is defined as the fraction of pupils repeating a grade in the year of the survey; the dropout rate is defined as the fraction of students who dropped out of school in the same year. Both variables are indirect measures of school quality, and in particular a high repetition rate may result from either high standards or low-quality education. Although there is a lively debate over 470 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 3. Descriptive Statistics for Schools in the Chaco and Resto Rural Regions in Bolivia, 1993 Treatment schools Control schools P-value of Sample Mean Sample Mean test of equala Region and variable size value size value Means Distribution Chaco region Response variable Repetition rate 35 0.13 36 0.13 0.95 0.86 Dropout rate 35 0.13 36 0.09 0.20 0.51 School resources Blackboards per classroom 35 0.12 37 0.07 0.43 1.00 Desks per student 35 0.60 36 0.44 0.14 0.06 Students per classroom 35 24.3 36 22.1 0.36 0.12 Books per student 27 0.48 28 0.31 0.15 0.09 Students per teacher 35 19.9 36 20.0 0.97 0.90 Proportion of teachers with professional degrees 35 0.37 37 0.28 0.34 0.33 Characteristics of students Log per capita consumption of householdb 31 7.05 34 7.03 0.81 0.98 Education of mother (years) 31 2.49 34 2.05 0.32 0.37 Education of father (years) 31 3.49 33 2.71 0.15 0.22 Community characteristics Number of nongovernmental organizations 28 0.54 28 0.14 0.08 0.44 Knowledge of the Social Investment Fund 28 0.46 28 0.46 1.00 1.00 Population 28 454.1 28 468.4 0.93 0.90 Distance to main road (kilometers) 28 17.8 28 13.3 0.40 0.44 Resto Rural region Response variable Repetition rate 24 0.08 21 0.08 0.92 0.98 Dropout rate 24 0.13 21 0.08 0.21 0.62 School resources Blackboards per classroom 37 0.53 33 0.22 0.00 0.02 Desks per student 37 0.72 32 0.45 0.03 0.04 Students per classroom 37 24.1 33 24.5 0.81 0.56 Books per student 37 0.38 33 0.28 0.27 0.66 Students per teacher 32 23.2 27 23.3 0.97 0.83 Proportion of teachers with professional degrees 32 0.55 27 0.57 0.86 0.88 Characteristics of students Log per capita consumption of householdb 32 6.57 32 6.61 0.80 0.53 Education of mother (years) 32 1.35 32 1.43 0.80 0.94 Education of father (years) 32 3.85 31 3.75 0.86 0.95 Community characteristics Number of nongovernmental organizations 33 1.67 29 0.72 0.00 0.01 Knowledge of the Social Investment Fund 33 0.79 29 0.45 0.01 0.03 Population 33 293.4 29 344.7 0.52 0.10 Distance to main road (kilometers) 33 13.6 29 9.56 0.33 0.69 a. The test of equal means is the t-tesr; the test of equal distributions is the Kolmogorov-Smirnov test. b. Household consumption is in bolivianos. In 1993 1 U.S. dollar = 4.31 bolivianos. Source: Authors' calculations based on survey data. Pradhan, Rawlings, and Ridder 471 the determinants of school effectiveness, most authors consider the resources available at the school to be an important determinant of the quality of educa- tion (Velez, Schiefelbein, and Valenzuela 1993). We are particularly interested in this dimension because the SIF aims to improve these resources. Hence, we also compare available resources at the schools, so that we use both the indirect measures and an indicator of the resources of the school as outcome variables. The results of the comparison are reported in table 3. Table 3 shows compliance with the experimental design in the Chaco region. There are no significant differences between the treatment and control schools either in a comparison of the response variables or in a comparison of the school, student, or community characteristics. By contrast, in the Resto Rural region the matching of treatment schools with comparison schools on observable charac- teristics does not eliminate all differences between the schools. Although there are no significant differences for the response variables, the selected schools have significantly more resources in terms of the number of blackboards per class- room and desks per student. Moreover, they are located in communities with a larger number of NGOs and greater knowledge of the SIF. Table 3 contains urnivariate comparisons. As a further check on treatment assignment, we estimal;e a probit that relates the probability of being a control or comparison school to the resources of the school and the characteristics of the student population.5 The probit estimates in table 4 confirm the results of the univariate comparisons in table 3. There is weak evidence that even in the Chaco region schools with more resources have a higher probability of being selected for the SIF. However, the likelihood ratio test does not reject the hypoth- esis of random selection. That hypothesis is rejected for the Resto Rural region because better-off schools have a higher probability of selection. Note that the dummy for missing student data is not significantly different from 0, so that the selectivity is indeed due to nonrandom selection and not to nonrandomly miss- ing observations. Because assignment to the SIF seems to be related to the resources of the school and the resources affect the response variables, we estimate a linear regression in which the repetition rare and the dropout rate are related to school characteris- tics, student characteristics, and an indicator of being a comparison or control school. Besides the repetition and dropout rates, we also use an indicator of the resources of the school-the number of desks per student-as a response vari- able. Table 5 shows that for the Chaco region the control group indicator is not significantly different from 0 for all three response variables. Again this con- firms compliance with the random assignment in the Chaco region. For the Resto Rural region, we find that before the intervention, controlling for other observables, the comparison group had significantly lower dropout rates and fewer desks per student (table 5). As noted, comparison schools have significantly fewer resources but have repetition and dropout rates that do not 5. Control groups are randomly generated, whereas comparison groups are not. 472 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 4. Probit Estimate of the Probability of Selection for a Social Investment Fund Project in Bolivia, 1993 Variable Chaco region Resto Rural region School resources Blackboards per classroom -0.36 -1.56 (-0.53) (-3.17) Desks per student -0.80 -1.17 (-1.98) (-2.27) Students per classroom -0.023 -0.042 (-1.10) (-1.43) Books per student -1.10 0.29 (-2.30) (0.45) Students per teacher -0.026 0.0011 (-0.73) (0.026) Proportion of teachers with professional degrees -0.62 0.12 (-1.46) (0.23) Dummy for missing school data 0.18 -0.097 (0.42) (-0.082) Characteristics of students Log per capita consumption of household -0.29 0.32 (-0.66) (0.73) Education of mother (years) -0.081 0.29 (-0.42) (1.17) Education of father (years) -0.12 0.036 (-0.78) (0.30) Dummy for missing student data 2.40 -2.27 (0.78) (-0.78) Constant 2.05 1.72 (1.96) (1.45) Likelihood ratio test of significance (p-value) 14.1 20.0 [0.23] [0.046] Number of observations 71 69 Note: t-statistics are in parentheses; p-values are in square brackets. Source: Authors' calculations. differ significantly from treatment schools. For that reason, we find that, con- trolling for differences in resources, schools that perform poorly with respect to repetition and dropout rates are selected to receive SIF funding. The difference is significant for the dropout rate and the number of desks per student. None of the indicators for missing observations has a coefficient that is significantly dif- ferent from 0. The results give an indication of the quality of the response vari- ables. The repetition rate is weakly correlated with the resources of the school. That correlation is stronger for the dropout rate. This confirms our earlier doubts about the use of the repetition rate as a response variable. We conclude that the matched-comparison design for the Resto Rural region does not yield directly comparable treatment and comparison groups. Compari- son group schools have fewer resources but make better use of their resources, which results in lower repetition and dropout rates. The finding that there are Pradhan, Rawlings, and Ridder 473 Table 5. The Impact of School and Student Characteristics on Repetition Rate, Dropout Rate, and Desks per Student in the Chaco and Resto Rural Regions of Bolivia, 1993 Region and variable Repetition rate Dropout rate Desks per student Chaco region School resources Blackboards per classroom 0.026 0.032 (0.43) (0.61) Desks per student -.055 0.035 (-1.53) (1.14) Students per classroom 0.0022 0.0020 (1.23) (1.28) Books per student 0.010 0.062 (0.24) (1.70) Students per teacher 0.00075 -0.006 (0.26) (-2.59) Proportion of teachers wiz:h 0.013 -0.021 professional degrees (0.33) (-0.62) Dummy for missing school data 0.019 -0.019 (0.47) (-0.54) Characteristics of students Log per capita consumption 0.046 0.040 -0.19 of household (1.20) (1.24) (-1.42) Education of mother (years) -0.034 -0.013 0.0037 (-1.89) (-0.86) (0.063) Education of father (years) 0.022 0.0070 0.043 (1.71) (0.642) (0.94) Dummy for missing student data -0.32 -0.29 0.99 (-1.23) (-1.28) (1.08) Control group 0.0039 -0.014 -0.13 (0.123) (-0.53) (-1.24) Constant 0.076 0.19 0.741 (0.81) (2.37) (4.42) K2 0.21 0.22 0.10 Number of observations 71 71 71 Resto Rural region School resources Blackboards per classroom -0.014 -0.056 (-0.38) (-1.37) Desks per student -0.067 -0.084 (-1.76) (-2.02) Students per classroom 0.0014 -0.0010 (0.39) (-0.28) Books per student 0.011 -0.060 (0.18) (-0.91) Students per teacher 0.00052 -0.00046 (0.15) (-0.12) Proportion of teachers with 0.0025 0.084 professional degrees (0.051) (1.58) Dummy for missing schoo.l data 0.020 -0.20 (0.20) (-1.84) (Table continues on following page.) 474 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 5. (continued) Region and variable Repetition rate Dropout rate Desks per student Characteristics of students Log per capita consumption 0.018 0.027 0.063 of household (0.40) (0.53) (0.56) Education of mother (years) -0.008 -0.024 0.11 (-0.49) (-1.31) (1.86) Education of father (years) -0.001 0.020 0.041 (-0.12) (1.88) (1.09) Dummy for missing student data -0.15 -0.37 -0.87 (-0.49) (-1.12) (-1.18) Comparison group -0.026 -0.094 -0.26 (-0.78) (-2.61) (-2.30) Constant 0.12 0.54 0.85 (1.02) (4.21) (4.77) R2 0.28 0.58 0.26 Numbers of observations 44 44 69 Note: Results are from ordinary least squares. t-statistics are in parentheses. Source: Authors' calculations. no significant differences in the response variables between the comparison and treatment group schools ex ante does not imply in general that a direct compari- son of the response variables between the two groups ex post yields an unbiased estimate of the true SIF effect. The resulting bias can be derived from a linear regression equation that relates the response variable y to the vector of school resources x (we omit student characteristics, which do not differ between treat- ment and comparison schools and are assumed to be the same before and after the intervention). The bias is derived for a linear regression, but this restriction is for ease of exposition only. The results can be generalized to arbitrary nonlin- ear relations. Consider Ystk = as't + fxstk + s'tk t = 0,1; k = 1, .., K;,V = C, T where k is the school, s indicates the assignment to the treatment (s = T) or comparison (s = C) group, v denotes the outcome with (v = T) or without (v = C) the SIF investment, and t is 0 if the data are preintervention or 1 if postintervention. We introduce the superscript v to stress that for a particular school the program effect is obtained by comparing the outcome variable for that school with (v = T) and without (v = C) the intervention. Of course, only one of these variables can be observed: the outcome where superscript v coincides with the assigned treatment s, that is, s = v. The true SIF effect, that is, the difference in the expected outcomes with and without the SIF intervention, can be defined either for the treatment or the com- parison schools. If we choose the first option, the average SIF effect is (1) E(YT1) - E(yc1) = (XTf, _ 1aC,1) + p,(-T - Pradhan, Rawlings, and Ridder 475 where an overbar incdicates the mean value over all schools. The counterfactual E(yc 1), which is the expected outcome of the treatment group had they not received an SIF inves'tment, is not observed. Under several assumptions, how- ever, the average treatment effect can be derived directly from the observed preintervention and postintervention outcomes. Using the (observecd) postintervention difference between comparison and treat- ment groups, we obtain (2) E(y}7) - E(yc,I) = (aT, - ac l) + ,6((XT - Equation 2 is an estimate of the average SIF effect for the treatment schools, as defined in equation 1. The estimated effect in equation 2 results in a bias equal to (subtracting equation 1 from equation 2) (aTC I - ac,1) + '(XT,l - XC 1)- This bias is the average difference in outcomes between the treatment and com- parison schools for the counterfactual case that the treatment schools do not receive the SIF support. Under the strong assumption that the difference in the observed (x) and unob- served (a) characteristics between comparison and treatment groups in the ab- sence of an intervention remains the same before and after the intervention, aC aC C 0 C -c - c -c -c -T, CC, =a, -aC,0 and XT1 - XC, = XT,O - XC,O and under the weak assumption that the intervention has no effect for the treat- ment group before the start of the program (this assumption is also made with the iv estimator below), aTo= CT 0 and C =-T we find that the bias is equal to the difference between the preintervention aver- age responses: (XT,O - XC,O) + P'(XT,O - XCSO)- Thus the difference between the observed postintervention and preintervention averages gives an unbiased estimate of the SIF effect for the schools that are selected for the SIF. The estimator is the difference-in-differences estimator, as used by Ashenfelter (1978). Because of the long time interval between pre- intervention and postintervention surveys-five years-and the marked differ- ences found between the two groups before the intervention, it seems very un- likely that the two groups would have evolved similarly in the absence of an intervention. Therefore, it is unlikely that the assumptions that support difference-in-differences estimation will hold. 476 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 From the estimates in tables 3-5, we have aXT 0 < at ,0, P iTT,o > P XC'0o The preintervention difference between the average responses is not significantly different from 0. This does not imply that the postintervention difference in the average response is an unbiased estimator of the SIF effect. The strong assump- tions introduced above are needed to guarantee that this estimator is unbiased. As this example demonstrates, the fact that the difference between the pre- intervention average response of the treatment and comparison groups is 0 is neither a necessary nor a sufficient condition for using postintervention differ- ence as an unbiased estimator of the SIF effect for the schools that receive SIF support. In the next section, we explore how to address the potential bias. IV. QUASI-EXPERIMENTAL EVALUATION Besides matching and IV, economists have suggested other methods to deal with the bias due to selective treatment. These other methods are based on mod- eling the selection decision and are appealing if the assignment is the result of choice by an economic agent who has an interest in the outcome of the choice. The estimates in table 4, where the assignment is related to observable charac- teristics of the schools and the students, are a first step in this direction. How- ever, in our case, because several actors are involved in selecting schools in the Resto Rural region to receive SIF projects, a detailed economic model is hard to identify. A model of the assignment process will be at best a reduced-form model. The results in table 5 show that the bias is not induced by the observable vari- ables in the model for the treatment assignment, but by the unobservables. Hence, we should allow for a correlation between the error of the regression for the response variable and the error of the binary response model for treatment as- signment (Heckman 1979). To apply this method, we must specify the joint distribution of the unobservable errors. It is well known that this method gives sensible results only if there are variables that affect the treatment assignment, but not the response variable. Such variables are also essential for the quasi- experimental iv approach, which does not require arbitrary distributional as- sumptions and for that reason is more robust. The quasi-experimental IV approach starts from the observation that random- ized assignment as used in a classical experiment induces exogenous variation in the intervention indicator, and this variation is not correlated with the response variable. If we can find a variable that affects the treatment assignment, but not the response variable, we have exogenous variation that mimics the type of varia- tion induced by randomization. A variable with these properties is called an instrumental variable, and the corresponding experiment is referred to as a quasi or indirect experiment. The instrumental variable estimator of the intervention Pradhan, Rawlings, and Ridder 477 effect is consistent, but it is less accurate, that is, it has a larger variance, than the estimator of the intervention effect that could be used if the treatment assign- ment had been random. In this section, we use preintervention data to show that some community characteristics provide valid instruments, because they affect the selection into the SIF, while they do not have an effect on the preintervention response vari- ables. Hence, these instruments can be used in a quasi-experimental evaluation of the SIF. We also study the efficiency of this design. We show that the relative efficiency, that is, relative to a randomized design, is independent of the true treatment effect and hence can be estimated using baseline data only. This al- lows us to determine the sample size that will compensate for the loss of effi- ciency due to the quasi-experimental design. The comparison in table 3 suggests that for the Resto Rural region the num- ber of NGOs and the community leaders' knowledge of the SIF have a significantly positive effect on selection into the SIF. This is confirmed by a linear regression- the first step in a two-stage least squares estimation procedure for the SIF ef- fect-of the indicator of selection into the program on the exogenous variables and the set of potential instruments in table 6. The number of NGOs and knowl- edge of the SIF have a significant negative effect (on the selection into the control group) in this linear probability model with coefficients -0.23 and -0.19, re- spectively (see table 6). This result is expected because NGOs often act as subcon- tractors for the implementation of projects and because of the role that local leaders have in the selection process. The variables are valid instruments if they have no effect on the response variables. The evidence for the Resto Rural re- gion is in table 7. Because the regression coefficients for the potential instru- ments are not significantly different from 0, we conclude that they are valid instruments. The instrumental variables allow us to obtain a consistent estimate of the effect of the SIF intervention from the postintervention data from the regression equation6 Yslk = a01 + 1 + D 'Xslk + eslk k = l,...,K,;s = C,T where d is the vector of observations on the SIF indicator, and d k = 1 for the schools that receive SIF funding. X is the matrix with the observations on the independent variables including the constant but excluding the SIF indicator, and Z is the matrix with the observations on the instrumental variables. We define M = I - X(X'X)-1X' with I as the identity rnatrix and M as the least-squares projection matrix. The 6. Because the denominator of an tv estimator is random, it is biased in small, but not in large, samples. 478 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 6. Linear Probability Model of Membership in the Control or Comparison Group in Bolivia, 1993 Chaco Resto Rural Variable region region School resources Blackboards per classroom -0.2037 -0.3133 (-0.93) (-2.51) Desks per student -0.2553 -0.1541 (-1.96) (-1.44) Students per classroom -0.0054 -0.0086 (-0.82) (-1.15) Books per student -0.4108 -0.0512 (-3.15) (-0.35) Students per teacher -0.0130 0.0069 (-1.29) (0.74) Proportion of teachers with -0.2747 0.1362 professional degrees (-1.82) (1.13) Dummy for missing school data 0.0788 -0.1149 (0.48) (-0.40) Characteristics of students Log per capita consumption of household -0.1009 0.0485 (-0.84) (0.38) Education of mother (years) -0.0368 0.1174 (-0.53) (2.15) Education of father (years) -0.0161 -0.0163 (-0.33) (-0.55) Dummy for missing student data 0.8254 -0.1057 (0.98) (-0.13) Potential instruments Knowledge of the Social Investment Fund 0.0455 -0.1877 (0.38) (-1.30) Number of nongovernmental organizations -0.1221 -0.2252 (-1.37) (-4.15) Population (thousands) 0.0683 0.0359 (0.52) (0.26) Distance to main road (kilometers) -0.0043 0.0006 (-1.11) (0.17) Dummy for missing instruments data -0.1149 0.2934 (-0.72) (1.68) Constant 1.3374 0.5865 (4.17) (1.77) F-test instruments (p-value) 1.53 8.07 (0.19) (0.00) R2 0.24 0.43 Number of observations 71 69 Note: Values are from a linear probability model regressing membership in control group or school and student characteristics and potential instruments. Huber-corrected standard errors are in parentheses. Source: Authors' calculations. Pradhan, Rawlings, and Ridder 479 Table 7. The Impact of Response Variables on Repetition Rate, Dropout Rate, and Desks per Student in the Resto Rural Region in Bolivia, 1993 Variable Repetition Dropout rate Desks per student School resources Blackboards per classroom -0.013 -0.053 (-0.32) (-1.07) Desks per student -0.036 -0.078 (-0.88) (-1.59) Students per classroom 0.0015 -0.0012 (0.45) (-0.30) Books per student 0.0037 -0.065 (0.060) (-0.88) Students per teacher 0.00052 0.00006 (0.14) (0.013) Proportion of teachers with -0.043 0.082 professional degrees (-0.82) (1.32) Dummy for missing school data 0.033 -0.21 (0.32) (-1.68) Characteristics of students Log per capita consumption 0.040 0.030 0.045 of household (0.81) (0.51) (0.37) Education of mother (years) -0.021 -0.022 0.14 (-1.11) (-1.00) (1.91) Education of father (years) 0.00072 0.019 0.032 (0.065) (1.44) (0.80) Dummy for missing student data -0.27 -0.38 -0.81 (-0.83) (-1.00) (-1.00) Potential instruments Knowledge of the Social -0.035 -0.0026 0.067 Investment Fund (-0.84) (-0.053) (0.44) Number of nongovernmental 0.022 -0.011 0.0024 organizations (1.21) (-0.49) (0.031) Population (thousands) 0.038 0.010 -0.26 (0.70) (0.16) (-1.20) Distance to main road (kilometers) -0.0010 -0.00034 0.0017 (-0.80) (-0.23) (0.41) Dummy for missing instruments data -0.097 -0.00064 0.20 (1.44) (-0.008) (0.90) Constant 0.19 0.55 0.71 (1.52) (3.77) (3.14) F-test instruments (p-value) 1.22 0.10 0.60 [0.32] [0.99] [0.70] R2 0.42 0.59 0.29 Number of observations 44 44 69 Note: Values are ordinary least squares estimates. t-statistics are in parentheses; p-values are in square brackets. Source: Authors' calculations. 480 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 variances of the ordinary least squares and iv estimators of ac,1 are given respec- tively by C2 d'Md and 62 d'MZ(Z'MZ)-hZ'Md The ratio of these variances does not depend on the variance of the disturbance of the postintervention regression, a2, and hence can be computed with pre- intervention data. For the Resto Rural region, we find that the ratio of the standard errors of the estimates of the SIF effect is 3.97. Hence the standard error is four times larger than could have been obtained if the matched comparison had succeeded or if the SIF assignment had been random. As a consequence, 242 schools instead of 61 are needed to estimate the SIF effect with the same precision as with random- ized assignment. Because the assignment was random for the Chaco region, we find that the ratio is 11.20 for that region. Using an instrumental-variable esti- mator with randomized assignment gives a very inaccurate estimate of the SIF effect. V. CONCLUSIONS We used preintervention data to study the targeting of Bolivian SIF interven- tions in the education, health, and sanitation sectors and to examine selection of SIF interventions. For the health and sanitation components, we found that households that are better off are more likely to be beneficiaries of SIF invest- ment. For these components, the selectivity of the SIF complicates the impact evaluation. For the education component, the random selection of a group of schools eligible for SIF interventions that then received active promotion for SIF educa- tion projects facilitated evaluation of the impact of the SIF. In the Resto Rural region, an attempt was made to mimic randomized assignment by matching treatment and comparison group schools on observable characteristics. We found that this attempt was not fully successful. The matched-comparison approach in the Resto Rural region yielded less comparable treatment and comparison groups than the random-selection process used in the Chaco region. We proposed using an alternative indirect procedure to evaluate the intervention in the Resto Rural region using an instrumental variable approach to control for nonrandom selec- tion. The preintervention data allowed us to verify that our instrumental vari- able proposal will produce an unbiased estimate of the SIF effect but remained a less efficient approach compared with randomization. We computed the loss of accuracy due to the indirect experiment and estimated the number of schools Pradhan, Rawlings, and Ridder 481 that will be needed to obtain an estimate with a precision that is comparable to that obtained from randomized assignment. Our analysis demonstrated that a simple matched-comparison design intro- duces a bias in the estimate of the program effect. We showed that with preintervention data we can select a quasi or indirect experiment that coincides with the selection of valid instrumental variables and that preintervention data allow us to test the validity of the instruments. We revealed that the loss of efficiency due to the use of quasi-experimental techniques instead of random treatment assignment can be determined from preintervention data. With post- intervention data, we will be able to compare estimates of the program effects obtained using the different methods. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Angrist, Joshua D. 1990. "Lifetime Earnings and the Vietnam Era Draft Lottery: Evi- dence from Social Security Administrative Records." American Economic Review 80(3, June):313-36. Ashenfelter, Orley. 1978. "Estimating the Effect of Training Programs on Earnings." Review of Economics and Statistics 60(1, February):47-57. Dehejia, R. H., and S. M. Wahba. 1995. "Causal Effects in Non-Experimental Studies." Working Paper. Department of Economics, Harvard University, Cambridge, Mass. Processed. Fraker, Thomas, and Rebecca Mayard. 1984. An Assessment of Alternative Compari- son Group Methodologies for Evaluating Employment and Training Programs. Princeton, N.J.: Mathematica Policy Research, Inc. Friedlander, Daniel, and Philip K. Robins. 1995. "Evaluating Program Evaluations: New Evidence on Some Commonly Used Nonexperimental Methods." American Economic Review 85(September):923-37. Glaessner, Phillip J., Kye Woo Lee, Anna Maria Sant'Anna, and Jean-Jacques de St. Antoine. 1994. Poverty Alleviation and Social Investment Funds: The Latin Ameri- can Experience. World Bank Discussion Paper 261. Washington, D.C.: World Bank. Grossman, Jean. 1994. "Evaluating Social Policies: Principles and U.S. Experience." The World Bank Research Observer 9(2, July):159-80. Heckman, JamesJ. 1979. "Sample Selection Bias as a Specification Error." Econometrica 47(1, January):153-61. . 1997. "Instrumental Variables: A Study of Implicit Behavioral Assumptions Used in Making Program Evaluations." Journal of Human Resources 32(3, sum- mer):441-62. Heckman, James J., and V. J. Hotz. 1989. "Choosing among Alternative Nonexperimental Methods for Estimating the Impact of Social Programs: The Case of Manpower Train- ing." Journal of the American Statistical Association 84(408):862-80. Heckman, James J., Hidehiko Ichimura, Jeffrey Smith, and Petra Todd. 1998. "Charac- terizing Selection Bias Using Experimental Data: A Study of Adult Males in JTPA." Econometrica 66(5, September):1017-98. 482 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Holland, P. W. 1986. "Statistics and Causal Inference." Journal of the American Statis- tical Association 81(408):945-70. Imbens, Guido, and Joshua D. Angrist. 1994. "Identification and Estimation of Local Average Treatment Effects." Econometrica 62(2, March):467-76. Jimenez, Emmanuel. 1995. "Human and Physical Infrastructure: Public Investments and Pricing in Developing Countries." In Jere Berman and T. N. Srinivasan, eds., Hand- book of Development Economics, vol. 3B, chap. 47. Amsterdam: North-Holland. LaLonde, R. J. 1986. "Evaluating the Econometric Evaluations of Training Programs with Experimental Data." American Economic Review 76(4, September):604-20. Moffitt, Robert. 1991. "Program Evaluation with Nonexperimental Data." Evaluation Review 15(3):291-314. Newman, J. L., Margaret Grosh, and Steen Jorgensen. 1992. "Demand-Driven Funds: Managing Their Conflicts." In Steen Jorgensen, Margaret Grosh, and Mark Schacter, eds., Bolivia's Answer to Poverty, Economic Crisis, and Adjustment: The Emergency Social Fund, chap. 7. A World Bank Regional and Sectoral Study. Washington, D.C.: World Bank. Rosenbaum, P. R., and D. B. Rubin. 1983. "The Central Role of the Propensity Score in Observational Studies for Causal Effects." Biometrika 70(1):41-55. Silverman, B. W. 1986. Density Estimation for Statistics and Data Analysis. London: Chapman and Hall. Velez, Eduardo, Ernesto Schiefelbein, and Jorge Valenzuela. 1993. Factors Affecting Achievement in Primary Education: A Review of the Literature for Latin America and the Caribbean. Washington, D.C.: Human Resources and Development Opera- tions Policy, World Bank. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 483-501 Internationral Evidence on the Determinants of Private Saving Paul R. Masson, Tamim Bayoumi, and Hossein Samiei A broad set of possible determinants of private saving behavior is examined using data for a large sample of industrial and developing countries. Both time-series and cross- sectional estimates are obtained. Results suggest that there is a partial offset on private saving of changes in public saving and (for developing countries) in foreign saving, that demographics and growth are important determinants of private saving rates, and that interest rates and terms of trade have positive, but less robust, effects. Increases in per capita gross domestic product seem to increase saving at low income levels (rela- tive to the United States) but decrease it at higher ones. Despite an extensive literature on saving behavior, several empirical issues have not been resolved conclusively, including the effects of real interest rates, demo- graphic factors, and per capita income on private saving; the relationship be- tween growth and saving; and the extent to which private saving offsets move- ments in public (dis)saving (Aghevli and others 1990 and Deaton 1992). This article extends the empirical knowledge of private saving behavior by exploiting data for a large sample of industrial and developing countries and by looking at a broad set of possible determinants of private saving. It uses both time-series and cross-sectional information because the variability of potential explanatory variables differs in those two dimensions. In particular, some variables seem to explain persistent country differences (for example, dependency ratios or rela- tive per capita income), while others are correlated with year-to-year fluctua- tions (for example, the terms of trade or growth in gross domestic product- GDP). Fiscal variables seem to explain both some persistent long-term differences and short-term fluctuations. The existing literature tends to be limited to one of these two dimensions, one of the few exceptions being Schmidt-Hebbel, Webb, and Corsetti (1992), who use panel data to study behavior across developing countries.1 Conclusions con- 1. After the first version of this article was drafted, we discovered a study by Edwards (1996), covering related issues and corning to similar conclusions, but with a somewhat different empirical and policy emphasis. Paul R. Masson, Tamim Bayoumi, and Hossein Samiei are with the Research, Asia and Pacific, and European I departments, respectively, at the International Monetary Fund. The authors are grateful to colleagues and seminar participants at the World Bank, to Klaus Schmidt-Hebbel and to David Weil for advice and comments, and to Toh Kuan and particularly to Claire Adams for data and computing assistance. (D 1998 The International Bank for Reconstruction and Development/ THE WORLD BANK 483 484 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 cerning the significance of one or another factor have often depended impor- tantly on the choice of time-series or cross-sectional estimation, as well as the country or countries included. For instance, time-series estimation has typically found evidence of demographic effects on private saving in Japan but not in the United States, whereas cross-sectional estimates have yielded large effects (see Horioka 1993 on Japan, Carroll and Summers 1991 on the United States, and Modigliani 1970 and Graham 1987 for cross-sectional estimates). By exploiting both dimensions and using data for 61 industrial and developing countries, we examine the robustness of more limited studies. I. OUTSTANDING EMPIRICAL ISSUES This section provides a selective survey of unresolved issues. Does Private Sector Saving Offset Government Dissaving? The empirical literature on the private saving offset to government deficits (or dissaving) has generally concluded that a full offset (Ricardian equivalence) is rejected by the data, with some dissenters. According to Bernheim (1987), exist- ing evidence for industrial countries indicates that a unit increase in the govern- ment deficit would be associated with a decrease in consumption of 0.5 to 0.6. He presents new empirical results tending to confirm this range. Others have obtained similar results for developing countries. Corbo and Schmidt-Hebbel (1991), in a typical estimate, find a roughly 50 percent offset on private saving of changes in government saving. Haque and Montiel (1989) overwhelmingly reject Ricardian equivalence for their sample of 16 developing countries. They also conclude that the presence of liquidity constraints affecting at least some households causes the nonequivalence. Hayashi (1985), Flavin (1981), and Campbell and Mankiw (1989) find evidence that households in industrial coun- tries face liquidity constraints. By contrast, Seater (1993) argues that much of the empirical work is inadequate and concludes that the evidence supports the hypothesis of Ricardian equivalence. Nevertheless, he recognizes that different government behavior than in the past could imply Ricardian nonequivalence in the future. An increase in the government deficit as a result of lower taxes or higher government spending can have different effects on private saving, so the esti- mation in section II allows these variables to have separate coefficients. In- creased government spending may lower the resources available to the private sector and hence have a negative effect on private saving, regardless of whether it affects the deficit. The composition of government spending may also be important. Public investment, to the extent that it is viewed as productive, is not expected to require further taxes and should not generate a private saving response. Its coefficient in a saving equation should be smaller than the coeffi- cient of government consumption. In contrast, investment that does not gener- Masson, Bayoumi, and Samiei 485 ate revenues for the government (and is considered equivalent to government consumption) would involve future taxes and might induce a larger private saving offset. D)oes Income Growth Raise Saving? Modigliani (1966) argues that a higher growth rate (whether due to popula- tion or productivity growth) would, with unchanged saving rates by age group, raise aggregate saving because it would increase the aggregate income of those working relative to those not earning labor income (that is, retired persons liv- ing off their accumulated assets). This view is based on the life-cycle hypothesis, which relates saving behavior to successive stages of schooling, increased earn- ings, and retirement (Modigliani and Brumberg 1954 and Modigliani and Ando 1957). In fact, saving cloes seem to be positively correlated with income growth, because high-growth countries such as Japan or Korea also have high saving rates (Modigliani 1970). However, Tobin (1967) points out that unchanged individual saving rates are only consistent in this context with myopic expectations of future income. If workers correctly expect that their income will grow in the future, according to the life-cycle model, they should want to consume more today. Thus, saving rates for working individuals could fall by a sufficient amount to offset the ag- gregate effects of higher growth, a hypothesis confirmed by back-of-the- envelope calculations given the length of working lives relative to retirement. Thus the empirical positive correlation of saving with income growth is not, on the face of it, consistent with the life-cycle hypothesis, unless the higher income growth is at least partly transitory. Carroll and Weil (1994) confirm that lagged values of increases in income growth seem to explain higher saving rates; they argue that the usual consump- tion models with either uncertainty or liquidity constraints are not sufficient to explain this result and advance instead the hypothesis of habit persistence, ac- cording to which higher consumption associated with temporarily higher in- come takes some time to be reduced when income falls back. If growth leads to higher saving, for whatever reason, then these are important implications for countries like Japan whose growth has slowed. However, another explanation for the correlation may be that a high growth rate is a proxy for a high rate of return on capital, which may be reflected inadequately in domestic interest rates (especially if financial markets are not liberalized). Do Higher Interest Rates Lead to Higher Saving? The effect of interest rates on consumption is ambiguous theoretically, being subject to potentially offsetting negative substitution and positive income ef- fects, the latter reflectijng the fact that the private sector is a net creditor in finan- cial assets. It is true that human wealth (that is, discounted future labor income) is much larger than financial wealth for a typical individual and that human 486 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 wealth varies inversely with the rate of interest-suggesting that the negative substitution effect should dominate. However, consumers may not plan their lifetime consumption but instead respond primarily to current income. The sav- ing behavior of pension plans enhances the empirical importance of the income effect on private saving. For defined benefit plans, higher interest rates increase the income available to pay pensions, allowing lower contributions (Bernheim and Shoven 1988). Empirical research has reported mixed results, paralleling the theoretical ambiguity. For instance, using data on saving for industrial coun- tries, Bosworth (1993) finds a positive interest rate coefficient in a time-series estimation for individual countries, but a negative coefficient in a panel (cross- country) estimation. For developing countries, Giovannini (1985) concludes that in most cases the real interest elasticity is zero, while Schmidt-Hebbel, Webb, and Corsetti (1992) also find no clear effects on saving. Ogaki, Ostry, and Reinhart (1995) find positive interest rate effects that vary with income but are still small. Given that financial liberalization may have changed the interest rate effects, it is not too surprising that results are not robust. The effect of liberalization on saving behavior can operate through at least two channels. First, financial devel- opment may provide outlets for financial saving, thereby raising saving rates, a channel that has been emphasized in the development literature (McKinnon 1973 and Shaw 1973). However, although financial liberalization generally affects the form that saving takes and also the efficiency of investment, it need not raise the level of saving (De Gregorio and Guidotti 1994). The second aspect involves the liberalization of consumer access to bank credit, as occurred in a number of industrial countries in the 1980s. Regulatory changes have allowed banks to lend more freely to individuals, for instance for purchase of a house or for con- sumption, and this may lead, at least initially, to a significant decline in saving. Empirical evidence supports this effect in countries that have liberalized access to consumer credit (Jappelli and Pagano 1989, Bayoumi 1993, Lehmussaari 1990, and Ostry and Levy 1995). Financial liberalization may involve one or another of these aspects, each of which will tend to increase the sensitivity of saving to interest rates. Financial liberalization in a given country may also expand the international diversification possibilities of other countries, making their saving more responsive to foreign interest rates. Does Saving Vary with a Country's Income Level? Differences in per capita income could be one of the factors that explain the wide range of saving rates in developing countries. At subsistence levels, the potential for significant saving is small. A rise in per capita income may therefore lead to higher saving rates. The size of this effect is likely to decline as per capita income rises and may even become negative for rich countries where investment opportunities and growth are relatively lower. It seems to be a stylized fact that the process of development involves initially low sav- ing rates, a period of high growth accompanied by high saving rates, and Masson, Bayoumi, and Samiei 487 lower saving rates in rnore mature economies (see Ogaki, Ostry, and Reinhart 1995). Is the Age Structure a Significant Influence on Saving? The life-cycle hypothesis highlights the importance of the age structure of the population. If a high proportion of the population is of working age-especially if at peak earning years-then the economy should have a high rate of private saving, as workers provide for their retirement. Conversely, when this cohort reaches retirement age and dissaves (or, at least, consumes a greater fraction of its income), then the aggregate saving rate should decline. An extensive litera- ture attempts to link demographic variables to saving behavior. Studies using cross-country data (either as cross sections or as panels) have been more success- ful than time-series studies for individual countries in finding significant demo- graphic effects, probably because the variation over time of demographic vari- ables is relatively small. In particular, Leff (1969), Modigliani (1970), Modigliani and Sterling (1983), Graham (1987), and Masson and Tryon (1990) find that higher proportions of the young and elderly in relation to persons of working age-dependency ratios-are associated with lower saving rates. These estimates, and the projections of population aging in coming decades, would produce quite large falls in private saving in many industrial countries, especially in Japan. Koskela and Viren (1989) question the robustness of the cross-country demo- graphic effects identified by Graham (1987). And macroeconomic results (in- cluding across countries) conflict with studies using micro data for consumers by age cohort. Kennickell (1990) and Carroll and Summers (1991), for instance, argue that age-consumption profiles do not differ enough to explain why aggre- gate consumption should be very much affected by demographic factors. The discrepancy may be explained, however, by interactions between generations that are picked up by the macro data but ignored by the micro data studies: bequests may lower the saving of the young, and hence aggregate saving, even if the elderly do not themselves dissave (Weil 1994). Therefore, the thought ex- periment of changing the age structure of the population while keeping age- specific saving profiles unchanged may not be legitimate. Nevertheless, it must be acknowledged that studies using macro data have also found diverse results. Is There a Terms-of-Trade Effect on Saving? Another aspect of saving behavior that has appeared in the literature is the possible relationship between the terms of trade and saving (the Harberger- Laursen-Metzler effect). That is, an improvement in the terms of trade leads to an increase in saving and an improvement in the trade balance. The modern literature integrates this effect into intertemporal models and stresses the dis- tinction between transitory and permanent changes in the terms of trade. A tran- sitory improvement, because it causes only a transitory change in income, should lead to higher saving r ather than higher consumption, confirming the direction of the Harberger-Laursen-Metzler effect (Obstfeld 1982 and Svensson and Razin 488 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 1983). Permanent shocks to the terms of trade would have ambiguous effects that should be small in magnitude. The empirical literature tends to confirm a positive correlation between transitory terms-of-trade shocks and saving (Ostry and Reinhart 1992). Other Potential Determinants Other possible explanatory factors include inflation, wealth, and foreign sav- ing. Inflation may affect saving for several reasons: higher inflation tends to lead to higher nominal interest rates and hence higher measured household income and saving. However, higher inflation may also lower saving by increasing un- certainty. Financial wealth should negatively affect saving in a life-cycle model, because it increases the resources available for consumption. Foreign saving be- comes a potential exogenous determinant of national saving when foreign bor- rowing is rationed, as often is the case in developing countries. Some empirical evidence supports such a negative relationship between national and foreign saving (Fry 1978, 1980 and Giovannini 1985) and between household and for- eign saving (Schmidt-Hebbel, Webb, and Corsetti 1992). II. EMPIRICAL RESULTS We regressed saving rates for industrial and developing countries on several potential explanatory variables that could be collected on a reasonably compa- rable basis across all countries. For the industrial countries, the panel data set contains 21 countries over 1971-93. It contains 23 industrial countries, as de- fined by the International Monetary Fund, excluding Iceland and Luxembourg. See the appendix for data sources.2 Measurement issues are discussed by Blades and Sturm (1982), Lipsey and Kravis (1987), and Elmeskov, Shafer, and Tease (1991). In addition to the ratio of private saving to GDP, the data set consists of the general government budget surplus, government current expenditure, govern- ment investment, and beginning-of-period private sector wealth (all measured as ratios of nominal GDP); growth rates of real output, consumer prices, and the terms of trade; the real short-term interest rate; GDP per capita relative to that in the United States (measured using purchasing power parities); and the depen- dency ratio (the ratio of persons under 20 and over 64 to persons ages 20-64). Separating the overall dependency ratio into dependency ratios for the young and the old gave coefficients that were not significantly different from each other. The private wealth variable includes the stock of government debt. To the ex- tent that individuals are Ricardian, however, this debt should not be included in private wealth. Results when the stock of government debt was included in the specification as a separate variable were very similar to the main case and are not reported. 2. The data are available from the authors, either on a diskette or by e-mail (pmasson@imf.org). Masson, Bayoumi, and Samiei 489 We collected the same variables for a sample of 40 developing countries over 1982-93. Several variables in the developing-country data had to be constructed due to limitations of the data. We calculated national saving as domestic invest- ment plus the current account surplus, which means that foreign transfers are included as part of national saving. We calculated private saving as national saving minus the central government fiscal surplus and minus central govern- ment expenditure on capital goods. Hence, private saving includes saving by lower levels of government. In addition, we derived private wealth as the cumu- lative sum of nominal private savings. Because most developing countries face constraints on their external borrowing, foreign saving is also likely to be a determinant of domestic saving. Therefore, we included the current account sur- plus (equal to minus foreign saving) as a determinant of saving in the case of developing countries. Because the current account includes net private and offi- cial transfers, it exclucles foreign aid from foreign saving. Data on foreign aid were not available on a balance of payments basis. Thus the estimations re- ported here did not test for the effect of foreign aid on national saving. Panel data provide variation both across countries and over time. Table 1 provides information on some of the characteristics of the underlying data. It divides the total variance of each of the series into the part ascribed to changes over time within countries (the time-series variation) and the part ascribed to long-term differences across countries (the cross-sectional variation). Briefly, the variation over time was calculated by summing the individual variances across countries assuming that each country has a different mean. The cross-sectional variation was calculated as the variance across these country means multiplied Table 1. Cross-Sectional and Time-Series Variance in the Variables for Industrial and Developing Countries (percentage of total variance) Industrial countries Developing countries Across Over Across Over Variable countries time countries time Private saving/GDP 65.6 34.4 77.2 22.8 Government budget surplu!;/GDP 60.5 39.5 53.6 46.4 Government current expenditure/GDP 67.3 32.7 90.5 9.5 Government investment/GDP 62.1 37.9 72.5 27.5 GDP growth rate 8.2 91.8 20.7 79.3 Real interest rate 13.2 86.8 36.7 63.3 Wealth/GDP 66.7 33.3 82.1 17.9 Inflation rate 24.5 75.5 67.5 32.5 Percent change in terms of trade 1.1 98.9 4.4 95.6 Per capita GDP relative to the United States 94.7 5.3 97.0 3.0 Dependency ratio 62.3 37.7 95.7 4.3 Current account/GDP - - 35.7 64.3 - Not available. Note: The analysis uses 1971-93 data for 21 industrial countries and 1982-93 data for 40 developing countries. See the appendix for countries, variable definitions, and data sources. Source: Authors' calculations. 490 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 by the number of time periods. The two measures sum to the total variation. Note that the greater number of observations for industrial countries implies that a larger proportion of the total variance for these countries is accounted for by variance over time. See Kessler, Perelman, and Pestieau (1993) for a more detailed description of this approach. From table 1, private saving, the dependent variable, contains significant amounts of variation in both dimensions across both data sets, with cross- sectional differences explaining 60-80 percent of the total variance and changes over time explaining the remainder. The importance of the cross-sectional dif- ferences presumably reflects the persistence of differences in saving behavior across countries. For example, countries such as China, Italy, Japan, and Korea had relatively high private saving ratios throughout the sample period, while Kenya, the United Kingdom, the United States, and Uruguay had relatively low ratios. Cross-sectional differences are also more important than changes over time for the fiscal variables, the dependency ratio, the wealth ratio, and per capita GDP relative to the United States. By contrast, most of the variation in real short-term interest rates, output growth, the change in the terms of trade, and the current account is across time, presumably reflecting the greater impor- tance of cyclical variation in these cases. Inflation in industrial countries also shows more variation over time, but in developing countries the reverse is true. Most variables have significant variation across both countries and time, indicating that useful information can be extracted in both dimensions, the main exceptions being relative per capita GDP and the change in the terms of trade. The panel regressions focus on four principal explanatory factors as determinants of private saving: fiscal variables; demographics; GDP per capita and GDP growth; and interest rates, inflation, and changes in the terms of trade. A Combined Panel of Industrial and Developing Countries We combined the industrial- and developing-country data sets to produce an unbalanced panel involving a total of 61 countries: 21 industrial countries with 23 years of data (1971-93) and 40 developing countries with 12 years of data (1982-93). The data were treated identically across all countries except for the current account, which was eliminated from the estimation for industrial countries. In order to allow for different intercepts for each of the countries, "fixed effects" estimation was performed by including separate country dummies in the initial ordinary least squares (OLS) regressions. These coefficients (or rather n - 1 of them) were always jointly significant. We included time dummies for each year to account for the possible common effect of excluded variables on all countries' saving rates. However, the estimates for the other coefficients were not affected in a substantial way, so that only the estimates without time dum- mies are reported here. Masson, Bayoumi, and Samiei 491 The first column in table 2 reports the results from a general specification including all the variables, estimated using OLS with country dummies. The coef- ficients generally have signs that accord with intuition and are significant. In- creases in the general. government budget surplus (the fiscal position), govern- ment current and capital expenditure, and the dependency ratio all lower private saving, while increases in the GDP growth rate and wealth raise it. Table 2. Determinants of Private Saving: Results from the Combined Industrial- and Developing-Country Panel Ordinary Corrected First-difference, least for serial instrumental Variable squaresa correlationa,b variables Government budget surplus/GDP -0.60 -0.75 -0.81 (13.1) (19-9) (1 0. 1) Government current expe:rditure/GDP -0.32 -0.27 -0.14 (10.2) (7.2) (2.2) Government investment/GDP -0.24 -0.35 -0.42 (3.8) (5.9) (6.5) GDP growth rate 0.11 0.092 0.22 (3.5) (4.1) (1.5) Real interest rate 0.03 -0.026 0.10 (1.1) (1.1) (0.6) Wealth/GDP 0.012 -0.0038 -0.016 (3.0) (0.7) (1.3) Inflation rate 0.003 -0.042 0.089 (0.1) (1.9) (0.6) Percent change in the terms of trade 0.011 0.0098 0.005 (1.3) (1.9) (0.8) Per capita GDP relative to U.S. 0.51 0.53 0.076 (4.6) (3.8) (0.2) Per capita GDP relative to U.S. squared -0.004 -0.0039 -0.001 (4.7) (3.6) (0.5) Current account/GDPC 0.44 0.53 0.71 (12.4) (16.9) (4.0) Dependency ratio -0.14 -0.13 -0.14 (6.7) (4.2) (1.6) Fit statistics Adjusted R2 0.89 0.69 0.43 Standard error of regression 2.99 2.27 2.24 Durbin-Watson 0.90 1.91 1.85 Panel autocorrelation coefficient, p 0.64 Number of observations 963 963 780 Note: The dependent variable is the private saving/GDP ratio. Regressions are estimated using 1971- 93 data for 21 industrial countries and 1982-93 data for 40 developing countries. Absolute t-ratios are in parentheses. See the appendix for countries, variable definitions, and data sources. a. Country dummies were included, but their coefficients are not reported. b. Dependent and independent variables are quasi-differenced to eliminate residual serial correlation, using the method of Bhargava, Franzini, and Narendranathan (1982). c. Developing countries only. Source: Authors' calculations. 492 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 The results also support the hypothesis of a quadratic relationship between the national saving rate and per capita income. The estimated coefficients sug- gest that the turnaround is mild and occurs at around 60 percent of U.S. per capita income. We included the current account ratio only for developing coun- tries to measure the availability of financing. It has a strong positive effect on private saving. Finally, the coefficients on inflation, the real interest rate, and the percent change in the terms of trade are small and insignificant. However, a problem with the OLS results is the presence of serial correlation of the residuals, as evidenced by the low Durbin-Watson statistic calculated across the panel. (See Bhargava, Franzini, and Narendranathan 1982 for a generaliza- tion of the usual time-series statistics to panel data.) In order to avoid biasing estimates of standard errors, a small-sample estimate of the panel autocorrelation coefficient (p) was then calculated iteratively, using the formula in Bhargava, Franzini, and Narendranathan (1982), and the original series for the dependent and independent variables were then quasi-differenced. In particular, each ob- servation X, was replaced by X-X, - ,, except for the first observation, which was multiplied by '1p. The regression results reported in the second column in table 2 are based on the quasi-differenced series. In many respects, the estimates resulting from a serial correlation correction confirm the OLS results, with a few notable exceptions. The fiscal variables are all still significant, but a change in the government budget surplus now pro- duces a larger offset from private saving, about three-quarters. The GDP growth rate has a strong positive effect on private saving, the dependency ratio has a negative effect, and the same quadratic relationship between per capita in- come and saving emerges. The current account ratio has an even stronger and more significant effect for developing countries. The coefficient of the terms of trade continues to be (barely) insignificant, as does that for the inflation rate, which, however, changes sign, and the coefficient for the wealth variable is now insignificant. A further potential problem with these results is that saving may be deter- mined simultaneously with some of the other variables, in particular GDP growth, the real interest rate, and developing countries' current account ratios, causing the estimated coefficients to be biased. Accordingly, we reestimated the model after using first differencing and using as instrumental variables only the second lags of the above variables plus the other explanatory variables. (For the use of first differencing in the more general context of dynamic panel models, see, for example, Anderson and Hsiao 1982.) First differencing removes fixed effects and hence does not require the inclusion of country dummies (or de-meaning). It also allows lagged levels of endogenous variables to become valid instruments. This method deals with serial correlation of the residuals in a rough way but has the disadvantages of over-differencing (see the estimate of p in table 2) and of ignoring all information on level effects. In particular, variables that have only cross-sectional variation have no effect in this specification. Therefore, standard errors of coefficients (in parentheses) may be excessively high. Masson, Bayoumi, and Samiei 493 Qualitatively, the results of the first-difference instrumental variables method tend to be consistent with the earlier results, although there are now more insig- nificant variables (third column in table 2). The coefficients of the fiscal vari- ables and the current account ratio of developing countries remain significant. The Ricardian offset is slightly greater than before, while government invest- ment (for a given level of the deficit) has a larger negative effect on private saving than does government consumption. About seven-tenths of external fi- nancing extended to developing countries displaces private saving. The depen- dency ratio retains its coefficient, implying that a 7 percentage point rise in the dependency ratio lowers the private saving by 1 percent of GDP, but it is now only significant at about the 10 percent level. Relative per capita GDP now has no explanatory power, no doubt because levels have been discarded and changes in this variable occur only slowly. The R2 statistics indicate that even in first differ- ences, these regressions explain a relatively large proportion of the variation in the private saving ratio. Separate Panels of Industrial and Developing Countries It is interesting to consider the industrial- and developing-country panels sepa- rately, for at least three reasons. First, we were not able to use the same sources for the two groups. In particular, as mentioned, we calculated the fiscal vari- ables for general government (the preferred concept) for industrial countries, but central government for developing countries (for which general government data are not readily available). Differences in data may be associated with differ- ent coefficients, for instance on the Ricardian offset. Second, behavioral differ- ences may explain difl:erences in coefficients (in addition to variation captured in separate country intercepts). For instance, financial development may change the sign and significance of interest rate effects, depending on whether savers or borrowers or both are insulated from market interest rates. Third, our focus on the nonlinearity of the relative income relationship suggests that dividing the sample into higher-income and lower-income countries may be useful. The results for OLS and a correction for serial correlation are given in table 3 for the two subgroups. Using the OLS results, a test for equality of the coeffi- cients across the two subgroups gives an F-statistic (with 11 and 890 degrees of freedom) of 7.12, well above the 1 percent critical value of 2.30. Therefore, it is of interest to look in more detail at differences between the two sets of estimates. From table 3, some notable differences can be seen in the signs and magnitude of coefficients on GDP growth, the real interest rate, wealth, the inflation rate, the percentage change in the terms of trade, and relative in- come. GDP growth is weakly associated with saving for industrial countries, but much more strongly and significantly so for developing countries, perhaps re- flecting the greater importance of liquidity constraints and subsistence consider- ations. The real interest rate has a significant positive effect for industrial coun- tries, but a negative, though insignificant, effect for developing countries, perhaps as a result of differences in financial liberalization. 494 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 3. Determinants of Private Saving: Panel Estimates with Separate Country Constant Terms Industrial countries Developing countries Ordinary Corrected Ordinary Corrected least for serial least for serial Variable squares correlationa squares correlation' Government budget surplus/GDP -0.57 -0.77 -0.63 -0.73 (9.4) (16.0) (9.4) (11.4) Government current expenditure/GDP -0.47 -0.49 0.0074 0.041 (11.3) (10.5) (0.1) (0.5) Government investment/GDP -0.60 -0.60 -0.23 -0.36 (5.7) (5.1) (2.7) (4.4) GDP growth rate -0.060 0.011 0.14 0.10 (1.1) (0.3) (3.6) (3.3) Real interest rate 0.21 0.11 -0.021 -0.052 (4.4) (2.7) (0.6) (1.6) Wealth/GDP 0.23 0.012 0.0072 -0.0067 (5.1) (1.9) (1.1) (0.8) Inflation rate 0.18 0.083 -0.049 -0.064 (4.6) (2.2) (1.6) (2.1) Percent change in the terms of trade 0.047 0.039 0.0067 0.0057 (3.1) (4.7) (0.6) (0.8) Per capita GDP relative to U.S. 0.59 0.28 0.82 0.94 (3.4) (1.5) (3.7) (3.3) Per capita GDP relative to U.S. squared -0.0048 -0.0024 -0.0083 -0.0086 (3.9) (1.8) (2.7) (2.2) Current account/GDP 0.46 0.55 (11.2) (14.2) Dependency ratio -0.12 -0.18 -0.20 -0.18 (4.1) (3.6) (6.3) (4.4) Fit statistics Adjusted R2 0.77 0.72 0.84 0.71 Standard error of regression 2.33 1.54 3.36 2.70 Durbin-Watson 0.64 1.92 1.04 1.96 Panel autocorrelation coefficient, p 0.75 0.59 Number of observations 483 483 480 480 Note: The dependent variable is the private saving/GDP ratio. Regressions are estimated using 1971- 93 data for 21 industrial countries and 1982-93 data for 40 developing countries. Absolute t-ratios are in parentheses. See the appendix for countries, variable definitions, and data sources. a. Dependent and independent variables are quasi-differenced to eliminate residual serial correlation, using the method of Bhargava, Franzini, and Narendranathan (1982). Country dummies are not reported. Source: Authors' calculations. The coefficient for the wealth variable is (surprisingly) positive for industrial countries, but insignificant (and even negative) for developing countries. A simi- lar contrast emerges for the inflation rate. The percentage change in the terms of trade is only significantly positive for industrial countries. This series is domi- nated by large changes in oil and other commodity prices, but developing coun- tries include gainers as well as losers from these changes. Finally, because the industrial countries are all within about 50 percent of U.S. per capita income (near to the point at which the effect identified in the panel changes sign), this Masson, Bayoumi, and Samiei 495 panel is unable to identify the two coefficients. In fact, including only the linear term gives a negative coefficient, as would be expected from the derivative of the quadratic function estimated in the combined panel, when evaluated in this interval. The variables that are remarkably similar in effect in the two panels are the fiscal surplus, government investment, and dependency ratio. In particular, there is no evidence of difference in the Ricardian offset, despite the use of central government data for developing countries. Cross-Sectional Results Cross-sectional results are of interest for comparison with the first-difference specification and with earlier studies. Therefore, we estimated a simple OLS re- gression on the sample means for the variables that have significant cross- sectional variation. The industrial country regressions involve 21 observations, one for each country, while the developing country and combined results in- volve 40 and 61 observations, respectively. The industrial country regressions use data averaged over the full 1971-93 period, while the other regressions use 1982-93 averages. The variables included are the government balance, real out- put growth, the dependency ratio, relative per capita GDP, and (in the case of the developing-country and combined estimates) the square of relative GDP. The first column in table 4 reports the results for industrial countries from a restricted regression using these variables. A comparison with the results in table 3 indicates that the estimated coefficients tend to be greater in the cross- sectional regression than in the time-series results. The cross-sectional analysis Table 4. Determinants of Private Saving: Cross-Sectional Estimates Industrial Developing Variable countries countries All countries Government budget surpluS/GDP -0.71 -0.61 -0.53 (4.6) (2.0) (2.6) GDP growth rate 2.77 1.73 1.25 (3.9) (3.1) (3.2) Per capita GDP relative to U.S. -0.06 0.72 0.16 (1.7) (2.1) (1.3) Per capita GDP relative to U.S. squared -0.014 -0.0015 (2.1) (1.3) Dependency ratio -0.28 -0.05 -0.10 (3.8) (1.0) (2.5) Fit statistics Adjusted R2 0.74 0.37 0.41 Standard error of regression 2.06 5.95 5.18 Number of observations 21 40 61 Note: The dependent variable is the private saving/GDP ratio. Regressions are estimated using OLS using 1971-93 averages for 21 industrial countries, 1982-93 averages for 40 developing countries, and 1982-93 averages for all 61 countries. Absolute t-ratios are in parentheses. See the appendix for countries, variable definitions, and data sources. Source: Authors' calculations. 496 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 attempts to explain the cross-country variation in saving on the basis of the independent variables, while the panel "explains" some of the variation through separate intercepts. The most dramatic difference is in the case of real growth, which has a coefficient of more than 2 in the cross-sectional regression. The time-series regressions measure the sensitivity of saving to changes over the eco- nomic cycle, while the cross-sectional regressions measure the impact of long- term differences in behavior. Thus the difference in the real growth coefficient may indicate that saving is more sensitive to long-term differences in output growth than to shorter-term movements in these variables. Alternatively, both saving and growth may be affected by some third variable that is not included and whose effect shows up in separate constant terms in the panel regressions. A somewhat larger coefficient is also estimated for the dependency ratio, although here the difference is less striking. The second column in table 4 shows the results from running the same speci- fication on the developing countries, except that the square of per capita relative GDP is included in the specification. As in the case of the industrial country re- gressions, the coefficient on growth is considerably higher in these cross- sectional regressions than in the panel estimates reported earlier. The coeffi- cients on the fiscal position and on the dependency ratio are actually lower in the cross-sectional regression than in the panel estimation, in contrast to both our own and others' results using industrial country data. Both the level and square of per capita relative GDP are significant. The coefficients are generally similar to those found in the panel estimation, although the peak value for sav- ing implied by these point estimates occurs at around one-quarter of U.S. per capita GDP, which is lower than that found using the time-series estimates. The last column in table 4 shows the results from the combined industrial and developing country data. As in the other cross-sectional regressions, the coeffi- cient on growth is much higher than in the equivalent panel regression; how- ever, the coefficients on the fiscal balance and dependency ratio are somewhat lower. The coefficients on the relative level of GDP and its squared value are also somewhat smaller than in the equivalent panel regression and are not very well determined. At around 60 percent of U.S. GDP, the implied peak level of saving is very similar to that found earlier. Comparing the overall results from the cross-sectional regressions with those found using panel estimation provides a number of interesting insights. First, the two approaches provide reasonably similar estimated coefficients (for those vari- ables that are included in both regressions), except in the case of output growth. This contrasts with results using only industrial country data, where several authors have pointed to the very different coefficients, in particular for demo- graphic variables, produced by the two estimation techniques (see, for example, Bosworth 1993). Second, the results confirm the quadratic relationship between saving and per capita income. Finally, the strong relationship between saving and growth in the cross-sectional results may well imply a joint response to a third variable that affects the long-run values of each in the same direction. Masson, Bayoumi, and Samiei 497 Changes in the rate of growth in output over the cycle, by contrast, appear from the panel estimates to have a much smaller impact on the saving rate. III. CONCLUDING REMARKS Several conclusions emerge clearly from the regressions, despite some hetero- geneity in the results. First, there seems to be a substantial offset of changes in the government fiscal position from private saving, averaging 75 percent, de- pending on whether those changes are due to changes in government spending or in taxes. Although this offset is large, it is considerably below unity, implying that changes in the government's fiscal position can have a significant impact on national saving, especially if they result from reductions in spending. The estimates for both country groups show that demographic effects are an important determinant of private saving rates. However, the size of the effect of the dependency ratio on private saving is somewhat lower than in most previous studies that found a significant saving impact from demographic variables. The coefficients on the demographic variables are similar across different estimation techniques. The results identify a number of channels through which growth influences saving. A direct positive association between GDP growth and private saving emerges from most of the specifications, especially for developing countries, al- though it is unclear whether there is a causal effect in either direction or a joint response to a third factor. A suggestive result concerns the level of per capita income (relative to the United States) and saving. For developing countries, the analysis finds a generally significant positive effect of the level, but a negative effect of the squared level, of per capita income. This result implies that beyond a certain point higher income has a negative effect on the private saving rate. The combined panel and the cross sections are consistent with this implication. The real interest rate has a positive, and significant, coefficient for industrial countries, but the results are not very robust. Measurement problems related to the choice of the appropriate interest rate and measure of inflation may, in par- ticular, affect the results for developing countries, for which the coefficient is negative, but insignificant. Different levels of financial development may ex- plain the results, but it may also be the case that financial reforms changed the relationship during the sample period. Changes in the terms of trade have a significantly positive effect on saving for industrial countries (for which a longer sample was available), but not for devel- oping countries or the combined panel. The terms of trade in many countries deteriorated due to the oil price shocks of 1973 and 1979, and the deterioration had large effects in reducing their saving rates. Conversely, the terms of trade improved in oil-exporting countries and increased their saving, at least for a time. However, the effect is transitory, and because terms-of-trade changes bal- ance out at the world level, there is no presumption that this variable will dura- bly affect world saving. An additional external factor that negatively affects 498 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 private saving in developing countries is the level of foreign saving. As in the case of the government fiscal position, however, the offset, though large, is only partial, implying that foreign saving raises domestic investment. APPENDIX. DATA SOURCES Industrial Country Data The 21 industrial countries for which data were available are the following: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, United Kingdom, and United States. Most of the data came from the International Monetary Fund's World Eco- nomic Outlook (WEo) Database, supplemented in some cases by Organisation for Economic Co-operation and Development (OECD) sources (mainly the OECD Analytical Database). Specifically, OECD values were used for the private saving rate in Portugal and for some general government fiscal surplus and investment series. The dependency ratio data came from United Nations (1992). In some cases, the central government fiscal surplus was used to infer historical general government values. General government current expenditures were calculated as total general government expenditures less general government investment. The real interest rate was calculated as the short-term rate minus current infla- tion. Private wealth was calculated as the sum of the beginning-of-period capital stock (from the OECD Analytic Database where available; otherwise from cumu- lated investment), government debt, and net foreign assets. Some of the histori- cal values for net foreign assets were calculated by cumulating current account values backward from the earliest available figures. Developing Country Data The data source for developing countries was the WEO Database, except for the interest rate, for which data from the International Monetary Fund's Inter- national Financial Statistics were used for some countries to supplement the WEO Database (specifically, China, Paraguay, and Uruguay). The regressions include the following 40 countries (dictated by data availabil- ity): Algeria, Bangladesh, Benin, Burkina Faso, Burundi, Cameroon, Central African Republic, Chile, China, Colombia, Costa Rica, Cyprus, Ecuador, Egypt, El Salvador, Gabon, The Gambia, Honduras, India, Indonesia, Islamic Republic of Iran, Jamaica, Kenya, Republic of Korea, Lesotho, Oman, Malaysia, Mali, Malta, Mauritius, Morocco, Nepal, Nigeria, Panama, Paraguay, Rwanda, Tur- key, Uruguay, Venezuela, and Zimbabwe. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Masson, Bayoumi, and Samiei 499 Aghevli, Bijan, James Boughton, Peter Montiel, Delano Villanueva, and Geoffrey Woglom. 1990. The Role of National Saving in the World Economy. Occasional Paper 67. Washington, D.C.: International Monetary Fund. Anderson, T. W., and 'Cheng Hsiao. 1982. "Formulation and Estimation of Dynamic Models Using Panel Data." Journal of Econometrics 18(1):47-82. Bayoumi, Tamim. 1993. "Financial Deregulation and Household Saving." The Eco- nomic Journal 103 (November):1432-43. Bernheim, B. D. 1987. "Ricardian Equivalence: An Evaluation of Theory and Evidence." In NBER Macroeconomics Annual 1987. Cambridge, Mass.: MIT Press. Bernheim, B. D., and J. B. Shoven. 1988. "Pension Funding and Saving." In Zvi Bodie, J. B. Shoven, and D. A. Wise, eds., Pensions in the U.S. Economy. Chicago: University of Chicago Press. Bhargava, A., L. Franzini, and W. Narendranathan. 1982. "Serial Correlation and the Fixed-Effects Model." Review of Economic Studies 49(October):533-49. Blades, D. W., and Peter Sturm. 1982. "The Concept and Measurement of Savings: The United States and Other Industrialized Countries." In Saving and Government Policy. Proceedings of a con ference sponsored by the Federal Reserve Bank of Boston, Melvin Village, N. H., October. Boston: Federal Reserve Bank of Boston. Bosworth, B. P. 1993. Saving and Investment in a Global Economy. Washington, D.C.: Brookings Institution. Campbell, J. Y., and N. G. Mankiw. 1989. "Consumption, Income, and Interest Rates: Reinterpreting the Time-Series Evidence." NBER Working Paper 2924. National Bu- reau of Economic Research, Cambridge, Mass. Processed. Carroll, C. D., and Lawrence Summers. 1991. "Consumption Growth Parallels Income Growth: Some New Evidence." In B. D. Bernheim and John Shoven, eds., National Saving and Economic Performance. Chicago: University of Chicago Press. Carroll, C. D., and David Weil. 1994. "Saving and Growth: A Reinterpretation." Carnegie-Rochester Conference on Public Policy 40(June):133-92. Corbo, Vittorio, and Klaus Schmidt-Hebbel. 1991. "Public Policies and Saving in Devel- oping Countries." Journal of Development Economics 36(July):89-115. Deaton, A. S. 1992. Understanding Consumption. Oxford: Clarendon Press. De Gregorio, Jos6, and Pablo Guidotti. 1994. "Financial Development and Economic Growth." Research Department, International Monetary Fund, Washington, D.C. Processed. Edwards, Sebastian. 1996. "Why Are Saving Rates So Different across Countries? An International Comparative Analysis." NBER Working Paper 5097. National Bureau of Economic Research, Cambridge, Mass. Processed. Elmeskov, Jorgen, Jeffrey Shafer, and Warren Tease. 1991. "Saving Trends and Mea- surement Issues." Economics and Statistics Department Working Paper 105. Organisation for Economic Co-operation and Development, Paris. Processed. Flavin, Marjorie A. 1981. "The Adjustment of Consumption to Changing Expectations about Future Income." Journal of Political Economy 89(5):1020-37. Fry, M. J. 1978. "Money and Capital or Financial Deepening in Economic Develop- ment?" Journal of Money, Credit, and Banking 10(November):464-75. . 1980. "Saving, Investment, Growth, and the Cost of Financial Repression." World Development 8(4):317-27. 500 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Giovannini, Alberto. 1985. "Saving and the Real Interest Rate in LDCs." Journal of Development Economics 18(August):197-218. Graham, J. W. 1987. "International Differences in Saving Rates and the Life-Cycle Hy- pothesis." European Economic Review 31(December):1509-29. Haque, N. U., and Peter Montiel. 1989. "Consumption in Developing Countries: Tests for Liquidity Constraints and Finite Horizons." The Review of Economics and Statis- tics 71(3):408-15. Hayashi, Fumio. 1985. "The Permanent Income Hypothesis and Consumption Durabil- ity: Analysis Based on Japanese Data." Quarterly Journal of Economics 100(Novem- ber):1083-113. Horioka, C. Y. 1993. "Saving in Japan." In Arnold Heertje, ed., World Savings: An International Survey, pp. 238-78. Oxford: Blackwell Publishers. Jappelli, Tullio, and Marco Pagano. 1989. "Consumption and Capital Market Imper- fections: An International Comparison." The American Economic Review 79(5):1088- 105. Kennickell, Arthur. 1990. Demographics and Household Savings. Finance and Econom- ics Discussion Series 123(May). Washington, D.C.: Federal Reserve Board, Division of Research and Statistics. Kessler, Denis, Sergio Perelman, and Pierre Pestieau. 1993. "Savings Behavior in 17 OECD Countries." Review of Income and Wealth Series 39(1):37-49. Koskela, Erkki, and Matti Viren. 1989. "International Differences in Saving Rates and the Life-Cycle Hypothesis: A Comment." European Economic Review 33(Septem- ber):1489-98. Leff, N. H. 1969. "Dependency Rates and Savings Rates." American Economic Review 59(5):886-96. Lehmussaari, 0. P. 1990. "Deregulation and Consumption: Saving Dynamics in the Nordic Countries." IMF Staff Papers 37(March):71-93. Lipsey, R. E., and I. B. Kravis. 1987. Saving and Economic Growth: Is the United States Really Falling Behind? Report 901. New York: American Council of Life Insurance and the Conference Board. Masson, Paul R., and R. W. Tryon. 1990. "Macroeconomic Effects of Projected Popula- tion Aging in Industrial Countries." IMF Staff Papers 37(September):453-85. McKinnon, R. I. 1973. Money and Capital in Economic Development. Washington, D.C.: Brookings Institution. Modigliani, Franco. 1966. "The Life-Cycle Hypothesis of Saving, the Demand for Wealth, and the Supply of Capital." Social Research 33(summer):160-217. -. 1970. "The Life-Cycle Hypothesis of Saving and Intercountry Differences in the Saving Ratio." In W. A. Eltis, M. F. Scott, and J. N. Wolfe, eds., Induction, Growth, and Trade, pp. 197-225. Oxford: Clarendon Press. Modigliani, Franco, and Albert Ando. 1957. "Tests of the Life-Cycle Hypothesis of Savings: Comments and Suggestions." Bulletin of the Oxford Institute of Statistics May:99-124. Modigliani, Franco, and Richard Brumberg. 1954. "Utility Analysis and the Consump- tion Function: An Interpretation of Cross-Section Data." In E. E. Kurihara, ed., Post- Keynesian Economics, pp. 388-436. New Brunswick, N.J.: Rutgers University Press. Masson, Bayoumi, and Samiei 501 Modigliani, Franco, and Arlie Sterling. 1983. "Determinants of Private Saving with Spe- cial Reference to the Role of Social Security: Cross-Country Tests." In Franco Modigliani and Richard Hemming, eds., The Determinants of National Saving and Wealth. New York: St. Martin's Press. Obstfeld, Maurice. 1982. "Aggregate Spending and the Terms of Trade: Is There a Laursen-Metzler Effect?" The Quarterly Journal of Economics 97(1):251-70. Ogaki, Masao, Jonathar Ostry, and C. M. Reinhart. 1995. "Saving Behavior in Low- and Middle-Income Developing Countries: A Comparison." Working Paper WP/95/ 3. International Monetary Fund, Washington, D.C. Processed. Ostry, J. D., and Joaquira Levy. 1995. "Household Saving in France: Stochastic Income and Financial DereguLation." IMF Staff Papers 42(June):375-97. Ostry, J. D., and C. M. ]Reinhart. 1992. "Private Saving and Terms-of-Trade Shocks." IMF Staff Papers 39(September):495-517. Schmidt-Hebbel, Klaus, S. B. Webb, and Giancarlo Corsetti. 1992. "Household Saving in Developing Countries: First Cross-Country Evidence." The World Bank Economic Review 6(3):529-47. Seater, John. 1993. "Ricardian Equivalence." Journal of Economic Literature 31(March):142-90. Shaw, E. S. 1973. Financial Deepening in Economic Development. New York: Oxford University Press. Svensson, L. E. O., and Assaf Razin. 1983. "The Terms of Trade and the Current Ac- count: The Harberger-Laursen-Metzler Effect." Journal of Political Economy 91(1):97- 125. Tobin, James. 1967. "Life-Cycle Saving and Balanced Growth." In William Fellner, ed., Ten Economic Studies in the Tradition of Irving Fisher, pp. 231-56. New York: John Wiley and Sons. United Nations. 1992. World Population Prospects, 1992 rev. New York. Weil, David N. 1994. "The Saving of the Elderly in Micro and Macro Data." The Quar- terly Journal of Economics 109(February):55-8 1. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 503-26 Unfair Trade? The Increasing Gap between World and Domestic Prices in Commodity Markets during the Past 25 Years Jacques Morisset The early literature on international trade has debated extensively the constraints fac- ing exporting countries in the major commodity markets. This article goes one step further by suggesting that the final demand for these products could not have increased because the declines in world commodity prices were not transmitted or were trans- mitted imperfectly to domestic consumer prices. In contrast, upward movements in world prices were clearly passed on to domestic prices. As a result, the spread between world and domestic pf'ices almost doubled in all major commodity markets during 1975-94. This asymmetry, seldom discussed in the literature, does not seem to be caused, at least systematically, by changes in trade and tax policies or factors such as transport, processing, and marketing costs. This article argues, therefore, that a special effort should be made to better understand the transmission from world to domestic prices, especially the role of large international trading companies that may have the capability to influence such spreads through one or several stages of processing in most major commodity markets. Commodity prices have fallen in international markets since the 1970s. During the same time, however, prices for consumers in industrial countries have risen. For example, the price of coffee declined 18 percent on world markets but in- creased 240 percent for consumers in the United States between 1975 and 1993. Such diverging patterns can be generalized across a wide sample of commodities and countries-from crude oil to coffee, from Italy to the United States-but remain largely unexplored in the current economic literature. This article looks at the spreads between international and domestic com- modity prices and explains why these spreads have increased over time. The main finding is that the spreads between world and domestic wholesale prices as well as between domestic wholesale and consumer prices have increased dra- matically because domestic consumer prices have responded asymmetrically to movements in world prices. In all major consumer markets, decreases in world commodity prices have been transmitted to domestic consumer prices much less Jacques Morisset is with the Foreign Investment Advisory Service of the International Finance Corporation and the World Bank. The author would like to thank Marc Bacchetta, Joel Bergsman, Stijn Claessens, Antonio Estaci e, Michael Finger, Alejandro Izquierdo, Cheikh Kane, Marcelo Olarreaga, and Neda Pirnia, as well as three anonymous referees for their valuable comments. © 1998 The International Bank for Reconstruction and Development /THE WORLD BANK 503 504 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 than have increases. This asymmetry does not seem to be explained, at least systematically, by changes in trade and tax policies across consumer markets or in one individual market over time. Similarly, factors such as transport, process- ing, and marketing costs, as well as changes in standard quality, do not appear to have played a major role in the increasing spreads over time. A special effort should be made, therefore, to better understand the determinants of world, whole- sale, and consumer prices and their relationships in commodity markets. Over the past 25 years, the increasing spreads have certainly cost several billion dol- lars every year to countries producing and exporting commodities by restraining the expansion of the final demand for these products. Section I provides empirical evidence on the evolution of the spreads between world and domestic consumer prices, between world and domestic wholesale prices, and between wholesale and consumer prices for several commodities over the past 25 years. The section also discusses the data used throughout the ar- ticle. Section II examines the response of domestic prices to variations in world prices, paying special attention to the possible asymmetric relationship between these two prices. Section III explores the reasons for this asymmetry, ranging from changes in trade policies to variations in transport, marketing, and pro- cessing costs. Section IV offers concluding remarks and suggests possible direc- tions for future research. I. MEASURING THE VARIATIONS IN SPREADS BETWEEN WORLD AND DOMESTIC COMMODITY PRICES Consumers in industrial markets can easily observe that prices of coffee, rice, beef, and gasoline have risen almost continuously over the past two decades. When these prices have declined, it has only been because of short-term correc- tions related to episodes such as the oil price shocks in the 1970s. This general- ized increase in consumer prices can be contrasted with the long-term trend of declining world commodity prices. For example, the World Bank's nonfuel com- modity index declined 11 percent in nominal dollars or 42 percent in constant dollars between 1980 and 1994 (World Bank 1996). It is not surprising, there- fore, to find that the spreads between international and domestic commodity prices increased dramatically during this period. This section shows how to measure the changes in these spreads and then gives the results for a sample of commodities and countries during 1970-94. The changes in the spread between world and domestic consumer prices can be measured by the following standard equation expressed in changes of logs of prices: (1) Agii = APcij - A(e1p';) where jlit is the log of the spread (or markup) associated with product i in coun- try j, pci is the log of the domestic consumer price of product i in country j, ej is the log of the nominal exchange rate (dollar/local currency) in country j, and p ; Morisset 505 is the log of the world price of commodity i. To account for the influence of changes in the inflation rate on the measurement of the spread over time, all variables are expressed in logs so that their sample variations represent relative changes. Introducing domes tic wholesale prices into equation 1 can further decompose the spread between world and domestic consumer prices. Consequently: (2) ALtii = [Apc4, - Apwii I + [Apwij - A(e1p,')] where pwij denotes the log of the domestic wholesale price of commodity i in country j. The first expression in brackets on the right side in equation 2 repre- sents the spread (margin) between domestic consumer and wholesale prices, while the second expression represents the spread between domestic wholesale and world prices. I am particularly interested in the total spread because it captures the impact of the spread on the final demand for these commodities. The decom- position may provide additional information on how prices are transmitted through the stages of processing (even though the decomposition remains highly simplified). It is worth noting that equations 1 and 2 reflect the evolution of the spreads over time but do not provide information on the size of the spreads at any given point in time. The equations are based on the assumption that the exchange rate is neither under- nor overvalued. And they ignore differences in product quality and in transportation, storage, and marketing costs as well as other nontradable inputs. The influence of these factors will be examined more closely in the next section. Equations 1 and 2 were applied to a sample of six commodities: beef, crude oil, coffee, rice, sugar, and wheat. (Bananas were initially included in the sample, but data were only available for the United States and Japan.) The six commodi- ties were selected with. several factors in mind. I chose commodities that have as little processing as possible in order to limit the influence of exogenous factors. Another goal was to provide variation in the types of products. For this reason, five of these commodities are produced in both industrial and developing coun- tries, while one is a tropical product (coffee). Only one mineral commodity (crude oil) was selected because it is hard to match one specific final product with mineral commodities. The seven following pairs of commodities/consumer products were associaLted: beef/beef; coffee/coffee; crude oil/fuel oil; crude oil/ gasoline; rice/rice; sugar/sugar; and wheat/bread. The data on domestic consumer and wholesale prices were compiled on an annual basis for six countries: Canada, France, Germany, Italy, Japan, and the United States. The choice of an annual frequency primarily reflects the need to economize on data collection efforts. (For Canada and the United States, pro- ducer rather than wholesale prices were used because the latter series were not available; see appendix A for details.) All data were obtained from government publications or databases of these respective countries. This sample was con- strained by unequal access to comparable national sources for all countries at a 506 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 fairly disaggregated level (see appendix A). Nevertheless, these countries should capture a large portion of worldwide consumption. In addition, the differences in their trade and tax policies as well as their production structures should guar- antee enough diversity for the sample. International commodity prices were drawn from the World Bank's database (see appendix A). Finally, the exchange rate for every country was defined as the annual average rate reported in International Monetary Fund (various issues) even though in many countries this variable is volatile because of inflation and changes in exchange rate regimes. The next section uses several alternative econometric approaches to reduce the effect of this volatility. The results show an unambiguous positive long-term trend in the spreads between world and domestic consumer prices (as measured in equation 1) and between world and domestic wholesale prices as well as between wholesale and consumer prices (as measured by equation 2). For presentation purposes, the results are reported in index values rather than in percentage variations in figure 1 and tables 1 and 2. The base year is 1990 for all variables (1990 = 100), and the index values are derived from the yearly percentage variations in the spreads. Figure 1 shows that the (simple arithmetic) average spread between world and domestic consumer prices for all commodities (and all countries) followed a positive trend over the past two decades, accelerating during the 1980s. To ac- count for the annual volatility produced by seasonal and climatic factors in com- modity markets, the trend is best captured by the five-year moving average of the spread index, which almost doubled from a value of 62 to 118 between 1975 and 1994. The decline in the early 1970s is explained principally by the behav- ior of oil prices because the average index, which excludes this commodity, ac- tually increased during this period. The trends of increasing spreads between world and domestic consumer prices have been robust across countries and commodities. The spreads surged in all industrial countries between 1975 and 1994, ranging from an increase of 83 percent in the United States to 166 percent in Japan (table 1). Among the Euro- pean countries, the strongest increase was observed in Italy, followed by France and Germany. Similarly, the spreads rose in all commodity markets (table 2). A few spreads declined in the first half of the 1970s due to unexpected booms in commodity prices, but they more than recovered during the 1980s. As a result, only the spread for crude oil/gasoline was still lower in 1994 than in the begin- ning of the 1970s. The secular increase in the spreads is also demonstrated when the coverage period is extended to the 1960s, at least for countries where the data were readily available (France, Italy, and the United States). The decomposition of the spreads between world and domestic consumer prices reveals that rising spreads have been caused by a rise in domestic whole- sale prices compared with world prices and by a rise in consumer prices com- pared with wholesale prices (see tables 1 and 2). Again this finding seems homo- geneous across commodities and countries, with the exception of oil/gasoline, where the spread between the world and domestic wholesale prices declined 6 Morisset 507 Table 1. Spread Index for All Commodities, by Country, 1970-94 (annual average, 1990 = 100) Variation, 1975-94 Country and indicator 1970-74 1975-79 1980-84 1985-89 1990-94 (percent) Canadaa Consumer price/wholesale price 75.9 78.3 83.7 101.3 102.5 31 Wholesale price/world price 67.2 71.4 73.8 100.3 103.5 45 Consumer price/world price 92.7 53.4 62.7 98.5 105.0 97 France Consumer price/wholesale price 81.0 78.3 82.3 99.9 102.0 30 Wholesale price/world price 84.1 69.5 S9.9 94.0 107.2 54 Consumer price/world price 70.0 53.9 49.2 95.9 109.3 103 Germanyb Consumer price/wholesale price 104.3 95.5 86.4 100.0 109.1 14 Wholesale price/world price 85.0 65.7 71.2 96.3 109.2 66 Consumer price/world price 92.6 61.8 58.7 95.1 119.5 93 Italyc Consumer price/wholesale price 50.3 50.0 78.2 90.5 106.6 113 Wholesale price/world price 81.0 78.3 82.3 99.9 102.0 30 Consumer price/world price 65.4 52.6 53.0 90.1 117.9 124 Japan Consumer price/wholesale price 72.6 63.9 72.1 102.9 113.0 77 Wholesale price/world price 78.6 77.6 79.5 102.7 118.5 53 Consumer price/world price 56.5 50.5 58.9 110.7 134.2 166 United States Consumer price/wholesale price 90.5 74.1 83.1 99.7 102.5 38 Wholesale price/world price 97.5 82.9 90.1 103.6 105.5 27 Consumer price/world price 79.0 59.1 79.1 111.0 108.2 83 Note: The commodities are beef, coffee, oil, rice, sugar, and wheat. a. Wholesale prices for coffee, gasoline, rice, and sugar are only available from 1980 onward. b. Wholesale and consumer prices for rice are excluded. c. Wholesale prices for fuel are excluded. Source: Author's calculations. percent between 1975 and 1994 (although this negative trend is corrected if 1980 instead of 1975 is considered as a starting date for the comparison). The simultaneous increase in these two indicators suggests that reasons for the in- creasing spreads will have to encompass the successive stages of processing, from producers to consumers, and both international and domestic transactions. This observation will be kept in mind in the explanations explored in the next sections. Finally, an interesting aspect of these results is that the changes in the spreads of each commodity appear to have moved jointly across countries. An increase, say in the spread of oil in France, is likely to occur simultaneously in the other industrial countries surveyed here. Specifically, the spreads between world and 508 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 2. Spread Index for All Countries, by Commodity, 1970-94 (annual average, 1990 = 100) Variation, 1975-94 Commodity and indicator 1970-74 1975-79 1980-84 1985-89 1990-94 (percent) Beef Consumer price/wholesale price 77.8 77.8 87.3 97.2 107.1 38 Wholesale price/world price 69.8 81.5 76.0 90.2 98.0 20 Consumer price/world price 53.0 63.0 65.7 86.1 105.3 67 Coffeea Consumer price/wholesale price 74.9 67.8 73.7 85.4 98.3 45 Wholesale price/world price 65.9 62.3 63.2 75.5 102.4 64 Consumer price/world price 48.0 36.5 43.8 59.8 104.4 186 Oillfuelc Consumer price/wholesale price 85.0 74.7 73.8 104.7 100.8 35 Wholesale price/world price 115.3 80.2 77.4 101.0 115.2 44 Consumer price/world price 122.9 64.1 58.5 109.5 126.8 98 Oil/gasolinel Consumer price/wholesale price 95.6 75.9 78.8 100.6 113.4 50 Wholesale price/world price 167.4 117.2 73.6 102.9 110.2 -6 Consumer price/world price 174.8 81.7 51.1 103.9 128.4 57 Riceb Consumer price/wholesale price 67.4 65.9 77.9 90.0 104.6 59 Wholesale price/world price 55.6 73.4 72.0 96.8 98.5 34 Consumer price/world price 38.8 44.7 57.3 87.9 99.7 123 Sugara Consumer price/wholesale price 71.3 75.6 89.6 122.4 112.1 48 Wholesale price/world price 73.3 78.3 100.0 137.2 121.6 55 Consumer price/world price 55.2 55.1 97.8 171.2 136.1 147 Wheat' Consumer price/wholesale price 61.2 60.7 78.8 107.0 111.7 84 Wholesale price/world price 68.8 68.2 70.1 89.0 99.7 46 Consumer price/world price 37.1 38.8 47.5 82.0 105.9 173 Note: The countries are Canada, France, Germany, Italy, Japan, and the United States. a. Wholesale prices for Canada are only available from 1980 onward. b. Wholesale and consumer prices for Germany are not available. c. Wholesale price for Italy is excluded and for Canada is only available from 1980 onward. Source: Author's calculations. domestic consumer prices appear correlated from a minimum of 0.53 in the fuel market to a maximum of 0.95 in the gasoline market. These high correlation values indicate that the causes for the changes in the spreads of each commodity should be found in all markets simultaneously rather than in each individual consumer market or country. On the contrary, as reported in Morisset (1997), the variations in the spread of different commodities are correlated weakly, or even negatively, within each country. Morisset 509 Figure 1. Average Spread Index between World and Consumer Prices for Six Countries, 1970-94 Index, 1990= 100 120 - - All commodities 100 Excluding oil 80 - (--All commodities (five-year moving average) 60 40 20 l l l l l l l I l l l l 1970 1975 1980 1985 1990 Source: Author's calculations. Note: The six countries are Canada, France, Germany, Italy, Japan, and the United States. "All commodities" includes beef, coffee, oil, rice, sugar, and wheat. {I. ASYMMETRIC RESPONSE OF DOMESTIC CONSUMER PRICES TO CHANGES IN WORLD PRICES Why did the spreads of most commodity prices increase dramatically over the past two decades? The answer lies in the asymmetric response of domestic con- sumer prices to changes in world prices. If increases in world prices are well transmitted to domestic prices, while decreases are not, the spread between these two prices will rise automatically over time. The asymmetry in the price trans- mission will also be evidenced through the different stages of processing- between world and domestic wholesale prices and between wholesale and con- sumer prices. Several authors have shown that the changes in world commodity prices are well transmitted to domestic wholesale and consumer prices, but none has ex- plored the possibility that upward and downward movements in world prices are asymmetrically transmitted to domestic prices (see, for example, Mundlak and Larson 1992 and Anderson and Tyers 1992). In practice, however, increases in world prices generally have been forwarded to domestic prices more gener- ously than have decreases. For example, the surge in the price of oil in the early 1970s was almost perfectly passed on to domestic consumer fuel prices. By con- trast, the decline of 30 percent observed in the early 1990s was not transmitted to domestic consumer gasoline prices, which actually rose on average 5 percent 510 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 in the six countries surveyed. This section takes an empirical look at the elastici- ties of transmission between world and domestic prices and between domestic wholesale and consumer prices and explores the possibility of asymmetry in the transmission of prices. Explaining the growing spreads and the asymmetric price transmission is clearly a matter of investigating the determinants of world and domestic prices. Ideally, the relationship between these two prices should be explored empirically, prod- uct by product and country by country, to account for the unique features in the policy and institutional environments as well as the market structure of each of them. Yet the quantity of data and policy information required goes beyond the scope of this article. I have selected another, more global approach. That is, a general model can capture the changes in the spreads between world and domes- tic consumer prices of each commodity. In my view, such an approach is justi- fied by the homogeneous movements in the spreads, which were depicted across countries in the preceding section. The model used in this article is based on the approach developed by Mundlak and Larson (1992) and so is summarized briefly here. As in equations 1 and 2, this model assumes that the changes in world prices (Ap,tk ) and in nominal ex- change rates (Aej,) play a significant role in setting domestic (both wholesale and consumer) prices but that exporters and wholesalers can influence prices by us- ing their monopolistic power at different stages of processing. This approach is similar to the one followed by authors interested in the transmission of exchange rate variations to domestic prices, the so-called pass-through literature (see Knetter 1993 for a good summary). The impact of world prices on domestic prices is likely to vary across commodities. Within this approach, domestic prices are also influenced by other explanatory variables, including changes in domestic inputs, transport and marketing costs, as well as trade and tax policies observed in the destination countries. Domestic inputs principally reflect processing costs that are concentrated in consumer markets, which can be captured by changes in their nominal wages (Awi,). Other costs such as marketing, transport, and storage are difficult to observe. Knetter (1992) shows that they can be incorpo- rated into the model using fixed-time effects (0,) when their changes are com- mon to all the destination markets. Finally, as emphasized by Mundlak and Larson (1992), it is certainly a too-strong assumption to believe that tax and trade policies are uniform across countries, leading to a possible bias in the estimated transmission elasticities. This assumption can be weakened by consid- ering that the differences in these policies are captured by country time-invariant variables (0X) in the empirical tests. I also tested random effects in an earlier version of the article, and the estimated results are quite similar to those pre- sented here (see Morisset 1997). The general model of domestic price adjustment I propose to estimate for the seven pairs of commodities in the six main consumer markets covered here can be written as the three following equations: Morisset 511 (3a) Apcit = 0 1t + Olj + lAp it + yiAe,t + plAwj, + elij (3b) Apwi,t = 02t +0 r2Ap t + 72Aejt + p2Awjt + 62#t (3c) Apceit = 03t + 03, + I3ApWit + p3Awit + E3ijt. Equation 3a captures the relationships between world and domestic consumer prices, while equations 3b and 3c capture those between world and domestic wholesale prices and between domestic wholesale and consumer prices, respec- tively. I dropped the exchange rate as an explanatory variable from equation 3c because it only involves domestic prices. All variables have been described in the text and are expressed in logs (their variations represent relative changes). The 0, coefficients are on the time effects, which capture common movements in do- mestic prices over time across all destinations. The 0; coefficients reflect time- invariant changes in trade and tax policies across destinations. The D coeffi- cients represent the elasticity of the change in domestic prices with respect to the change in world prices; (or equivalently in wholesale prices in equation 3c), re- ferred to as the elasticily of transmission. A value of 1 implies that the variations in world prices are transmitted fully to domestic prices. However, a perfect cor- relation should not be expected because the commodity price is unlikely to ac- count for 100 percent of the wholesale or consumer prices for a variety of rea- sons that range from omitted variables to measurement errors (see Mundlak and Larson 1992 for a fuller discussion). A lag structure does not seem necessary because one-year sales are rare for the commodities analyzed here. The esti- mates obtained with the lagged dependent variable confirm that most of the price transmission seems to be made within one year (see Morisset 1997). This may reflect the emergence of large commodity funds in the 1980s, which have increased arbitrage opportunities and possibly shortened the transmission time between world and domestic prices. Finally, the error term e£1j is assumed to be independent and identically distributed. In order to test specifically for asymmetries in the response of domestic prices to changes in world prices, the full period is divided into upward and downward changes in world prices and in domestic wholesale prices. Rather than divide the full period between the years with upward and downward movements, which obviously would limit the number of observations and reduce the quality of the empirical results, I have multiplied the world commodity price in equations 3a and 3b and the wholesale price in equation 3c by two zero-one dummy vari- ables. The first variable takes the value 1 for the upward movements in world prices and zero elsewhere; the second takes the value 1 for the downward move- ments in world prices and zero elsewhere. The three above equations can be rewritten as follows: (4a) APCijt = Olt + Ol + I11AP1 it + P12AP2*it + ylAe1t + plAw1j + s1Iit 512 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 (4b) Apwiit = 02, + 02, + 321AP1 lt + 022AP2t + y2Ae1, + P2Aw1j + E2iit (4c) APcit- 03t + 03; + 331APW1.t + 332APW2it + P3Awit +E3ijt- If upward movements in world commodity prices (Ap1it) have been transmit- ted to domestic prices more systematically than have downward movements, (Ap 't) then it should be expected that f11 > 1312 and 121 > 322 in equations 4a and 4b. Identically, in equation 4c, the asymmetric transmission from changes in wholesale to consumer prices should result in 331 > 1332. Equations 3a, 3b, and 3c and equations 4a, 4b, and 4c were estimated for each commodity/product chain described in the preceding section during 1976- 94. The data on prices and exchange rates were the same as those described earlier. Labor costs were measured as the average unit labor cost in each indus- trial country covered in the sample. These data were extracted from databases of the International Monetary Fund or the United Nations Industrial Develop- ment Organization. Because the volatility of the exchange rates may affect the estimated transmission elasticities, domestic prices were expressed either in dol- lars or in local currencies. The regressions including fixed-time effects eliminate the exchange rate because they use variables expressed in changes in the price differences between countries for each year. F-statistic tests were used to deter- mine whether the data accept the restrictions on time and country effects. Although the response varies by commodity, the results of estimating equa- tions 3, which do not distinguish the direction of the changes in world prices, confirm the positive and significant relationships between world and domestic commodity prices, in line with the findings of other authors. Table 3 reproduces the estimated elasticities obtained for the changes in consumer prices with re- spect to the changes in world prices, those for the changes in wholesale prices with respect to world prices, and those for the changes in consumer prices with respect to wholesale prices (for detailed results, see appendix B). The estimated elasticities are reported only for the regressions including fixed-time effects be- cause the results from testing the homogeneity of time effects seem to indicate that the changes in domestic prices include critical time-correlated elements com- mon to all consumer markets. In contrast, fixed-country effects were not signifi- cant and thus were omitted from the regressions. The values of the elasticities of domestic consumer prices with respect to changes in world prices are relatively low, with a median value of 0.15 as derived from the first column in table 3. Such low values can be expected with regressions in variations rather than lev- els. The higher values of the elasticities presented in the second and third col- umns in table 3, with their respective median values of 0.40 and 0.24, reflect the closer connections between world and domestic wholesale prices as well as be- tween wholesale and consumer prices than between world and domestic con- sumer prices. The most interesting aspect of the empirical results concerns the asymmetry of the price transmission, which is almost always supported by the empirical Morisset 513 Table 3. Elasticity of Price Transmission, Upward and Downward Movements, 1976-94 Commodity/product Equation 3a Equation 3b Equation 3c Beef/beef 0.204 0.365 0.378 (2.54) (4.44) (4.26) Coffee/coffee 0.103 0.754 0.393 (3.78) (5.02) (6.08) Oil/fuel 0.073 0.203 0.329 (0.48) (1.47) (2.97) Oil/gasoline 0.228 0.459 0.243 (2.10) (2.92) (3.61) Rice/rice 0.071 0.403 0.214 (0.45) (2.89) (1.94) Sugar/sugar 0.229 0.691 0.187 (2.96) (3.85) (4.69) Wheat/bread 0.190 0.052 0.194 (3.20) (0.48) (3.32) Note: The countries are Canada, France, Germany, Italy, Japan, and the United States. Equation 3a estimates elasticities of consumer prices with respect to changes in world prices. Equation 3b estimates elasticities of wholesale prices with respect to changes in world prices. Equation 3c estimates elasticities of consumer prices with respect to changes in wholesale prices. See equations 3 in the text. Values are estimated elasticities for the regressions using domestic prices expressed in U.S. dollars and including fixed-time effects but not fixed-country effects. t-statistics are in parentheses. Source: Author's calculations. results given in tables 4 and 5 (the two exceptions are noted). The estimated transmission elasticities appear higher with upward than with downward move- ments in world prices when the regressions include time-differences variables to reduce the omitted-variable bias. Because an increasing positive trend in world commodity prices would automatically bias the empirical results toward higher elasticities with upward changes than with downward changes, I used variables in first differences to reduce the possibility of spurious correlation associated with time-series data when measured in levels. I also verified that world com- modity prices have not shown an increasing trend during 1970-94. The inclu- sion of fixed-time effects should correct for this eventual bias. Comparison of the first columns in tables 4 and 5 indicates that the median value of the elasticities of consumer prices with respect to changes in world prices exceeds 0.25 when the world prices were increasing and reaches only 0.05 when those prices were decreasing. Asymmetric elasticities were also evidenced between world and domestic wholesale prices (as revealed by the median values of 0.39 and 0.26 for upward and downward changes, respectively) and between wholesale and consumer prices (0.30 and 0.11). By comparison, Knetter (1993) finds the opposite behavior for a sample of manufacturing products. Prices ad- justed more rapidly to exchange rate depreciation (equivalent to a decline in world prices), suggesting that exporters of manufactured goods chose to increase their market shares rather than their markups. Similar behavior could not be shown in commodity markets. 514 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 4. Elasticity of Price Transmission, Upward Movements, 1976-94 Commodity/product Equation 4a Equation 4b Equation 4c Beef/beef 0.281 0.394 0.464 (2.31) (5.62) (4.08) Coffee/coffee 0.151 0.832 0.014 (0.79) (3.90) (0.13) Oil/fuel 0.292 0.155a 0.325 (1.29) (0.75) (2.15) Oil/gasoline 0.385 0.521 0.301 (2.37) (2.20) (2.07) Rice/rice 0.147 0.445 0.389 (0.62) (2.08) (2.71) Sugar/sugar 0.115 0.020a 0.189 (0.98) (0.07) (2.06) Wheat/bread 0.244 0.287 0.198 (2.64) (1.74) (2.54) Note: The countries are Canada, France, Germany, Italy, Japan, and the United States. Equation 4a estimates elasticities of consumer prices with respect to changes in world prices. Equation 4b estimates elasticities of wholesale prices with respect to changes in world prices. Equation 4c estimates elasticities of consumer prices with respect to changes in wholesale prices. See equations 4 in the text. Values are estimates for the regressions using domestic prices expressed in U.S. dollars and including fixed-time effects but not fixed-country effects. t-statistics are in parentheses. a. The elasticity for upward changes was lower than that estimated for downward changes. Source: Author's calculations. The results pertaining to the other variables also deserve a brief explanation. First, as indicated earlier, fixed-time effects have influenced significantly the changes in domestic prices in all commodity markets, suggesting that cost or productivity variations over time can produce fluctuations in the domestic con- sumer prices of commodities. Second, by contrast, country discrimination in the behavior of domestic prices was rejected by the F-statistics tests, except perhaps for sugar. Although Japan, the United States, and Europe have systematically followed distinct trade and tax policies, their spreads increased homogeneously during the past 25 years, as already evidenced by the results presented in the previous section. Third, the domestic price response equals one-third of the changes in the nominal exchange rate in most commodity markets, which is in the lower range of the elasticity values reported by Goldberg and Knetter (1997) for a sample of industrial products.1 Finally, the changes in domestic prices have been associated positively with those in domestic wages (see appendix B for details). This correlation was higher for wheat/bread and oiVlfuel because the labor and processing costs are certainly greater for those commodities than for the others. III. EXPLANATIONS FOR THE ASYMMETRIC RESPONSE OF DOMESTIC PRICES What are the sources of the asymmetric response of domestic prices to changes in world commodity prices? There are multiple possible explanations that should 1. These results are available from the author upon request. Morisset 515 Table 5. Elasticity of Price Transmission, Downward Movements, 1976-94 Commodity/product Equation 4a Equation 4b Equation 4c Beef/beef 0.113 -0.069 0.112 (0.85) (-0.56) (0.48) Coffee/coffee 0.048 0.663 -0.230 (0.23) (2.87) (-1.71) Oil/fuel -0.159 0.255a 0.274 (-0.68) (1.18) (1.80) Oil/gasoline 0.061 0.394 0.287 (0.36) (1.60) (2.76) Rice/rice 0.008 0.369 -0.127 (0.04) (1.90) (-0.60) Sugar/sugar -0.325 0.248a 0.062 (-3.03) (5.31) (0.10) Wheat/bread 0.141 -0.159 0.186 (1.63) (-1.02) (1.51) Note: The countries are Canada, France, Germany, Italy, Japan, and the United States. Equation 4a estimates elasticities of consumer prices with respect to changes in world prices. Equation 4b estimates elasticities of wholesale prices with respect to changes in worLd prices. Equation 4c estimates elasticities of consumer prices with respect to changes in wholesale prices. See equations 4 in the text. Values are estimates for the regressions using domestic prices expressed in U.S. dollars and including fixed-time effects but not fixed-country effects. t-statistics are in parentheses. a. The elasticity for upward changes was lower than that estimated for downward changes. Source: Author's calculations. rely on a frictionless competitive model of trade. These explanations should also encompass every stage of processing because the increasing spreads have been evidenced between not only world and domestic wholesale prices but also do- mestic wholesale and consumer prices. The three most popular explanations are the presence of trade restrictions in the main consumer markets, rising process- ing costs that act as bottlenecks in the trade of commodities, and differentiated changes in productivity across the stages of processing over time. This section describes these three explanations and discusses their limitations. The first explanation is based on the existence of trade restrictions in most industrial countries and has been used by many authors interested in explaining the asymmetric transmission of exchange rates (see Knetter 1993). It suggests that, in the presence of binding quantity constraints in export markets, the de- cline in world commodity prices will not be transmitted to domestic prices be- cause there is no incentive for exporters to stimulate the final demand by reduc- ing their selling prices. Exporters will instead increase their margins. At first sight, empirical support to this theory is provided by the increasing spread be- tween world and domestic wholesale prices reported earlier (up 45 percent dur- ing 1975-94) and by the numerous import barriers faced by commodity export- ers in consumer markets. Using instruments specifically designed to insulate domestic producers from lower world prices has also enhanced the asymmetric transmission of world commodity prices. Perhaps the most notorious examples are the levies and variable tariffs adopted as part of the European agricultural 516 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 policy, but examples can be found in other industrial countries and commodity markets as well. Nearly every major grain-producing country has used a two-tier scheme to separate domestic consumer prices from international prices. Since 1967 the European community has followed a two-tier pricing system for grains that maintains domestic prices above world market levels. A similar policy has been adopted by Canada. Japan maintains a differential between rice prices paid by consumers and prices received by producers by means of large subsidies to con- sumers. Even the United States has followed a two-tier pricing system at certain times (see Mitchell and Duncan 1987 for a fuller description). The second explanation for the asymmetric response of domestic prices is that exporters and wholesalers face a series of binding internal constraints when they want to increase their sales. For example, Foster and Baldwin (1986) intro- duce an approach using a fixed-proportion marketing technology that is required to sell products in foreign markets. This approach predicts that declines in world prices will be only imperfectly transmitted to domestic prices because, if existing sales are constrained by marketing capacity, exporters will compensate for ris- ing marketing costs by raising their selling prices. This increase will partially offset the initial impact of declining world prices on domestic prices. Because there is no similar constraint on higher world prices, it might be expected that more domestic price adjustments would occur with rising than with declining world prices. Potentially, this bottleneck approach can apply to a variety of costs, such as processing, distribution, marketing, and transportation, all of which play a significant role in setting domestic prices in commodity markets. The third popular explanation assumes systematic differences in productivity gains between producing, wholesaling, processing, and retailing activities. It could be argued, for example, that foreign competition has led to greater productivity gains in the export process than in the transactions between domestic wholesal- ers and retailers over the past 25 years. Under these conditions, the downward trend in world prices would have been transmitted imperfectly to domestic con- sumer prices, leading to increasing spreads over time. This explanation is obvi- ously very close to that based on the bottleneck approach because both assume that costs that are increasing (or not decreasing fast enough) will constrain the expansion of commodity trade at some stages of processing. Can these three explanations really provide an answer to the asymmetric re- sponse of domestic prices to changes in world commodity prices? Evaluating their explanatory power is difficult in the absence of an analytical model ca- pable of nesting competing explanations for the asymmetric behavior. Short of such a framework, I will simply confront the generalized increase in the spreads observed in most commodity markets over the past 25 years with basic empiri- cal evidence, in an attempt to determine a general pattern in the causes of the asymmetric transmission in all commodity markets. The hypothesis based on restrictive trade policies might not be as important as it appears at first sight. Indeed, the variations in trade barriers seem to ex- Morisset 517 plain relatively well thEe persistent deviations in the levels of commodity prices across countries, but not their relative changes over time. This lack of explana- tory power can be evidenced by several arguments. First, the changes in the spreads between world and domestic wholesale prices (which would best cap- ture the impact of trade policies) followed the same patterns in all consumer countries, included in the sample, in spite of systematic differences in trade poli- cies in these countries. Second, the cross-country differences in trade policies, as captured by the fixed-country effect, were not significant in almost all the re- gressions presented in the above section. Third, if these policies were the major cause for the asymmetric response of domestic prices, a close and positive corre- lation would be expected between the changes in the spreads and those in the levels of protection, both across countries and in one individual country over time. The weakness of this correlation is most apparent in Europe, Japan, the United States, and Canada where, despite distinct patterns of trade protection, the spreads moved alrmost simultaneously over the past 20 years. The flaws of the hypothesized link are further exposed by the weak correla- tion between the effective rates of protection and the spreads. Effective rates of protection present the advantage of capturing the effects of both tariffs and nontariff barriers. Obtaining exact measurements of the effective rate of protec- tion is always difficult, even for relatively homogeneous products such as food- stuffs. Differences in the quality of products to which available price data refer and the presence of dlata on marketing margins are but two of the problems associated with using even the simplest indicator of the extent of distortions. As reported in table 6, only in the case of sugar did these two variables-the effec- tive rate of protection and the spread-move in the same direction in all con- sumer markets between 1986 and 1994. Finally, it is certainly audacious to think that movements in trade barriers contributed significantly to the surge in the spreads of coffee and rice in the United States, up 85 and 112 percent, respec- tively, during 1975-94, when their effective rates of protection were on average below 2 percent during this period. The explanatory power of the bottleneck approach seems higher than that of protection policies. Indeed, binding costs on commodity sales would bias the price transmission in all countries simultaneously and, thus, be consistent with the significant fixed-time effects found in the regressions. Yet, the statistical find- ings do not necessarily imply that commodity exporters were constrained by higher transport, marketing, and processing costs over the past few decades. Transportation and insurance costs, which may contribute up to 10-20 percent of the final value of commodities, followed a descending trend over the past 20 years. Atkin (1992) reports that transportation costs may account for 10 per- cent of the landed price of grain on a trade route between efficient ports used by large vessels (for example, from New Orleans to Rotterdam) and 20 percent on a less efficient route. For example, Amjadi and Yeats (1995) report that the share of these costs in the total exports of developing countries declined from 7.8 percent in 1970 to 5.8 percent in 1991. The international evidence on mar- 518 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Table 6. Changes in Spreads and Effective Rates of Protection in Europe, Japan, and the United States between 1986-88 and 1989-93 (percent) Commodity and indicator Europe' Japan United States Beef Effective rate of protection 17 -54 -33 Spread 7 6 6 Coffee Effective rate of protection n.a. n.a. 0 Spread 23 33 45 Rice Effective rate of protection -33 -20 100 Spread 6 -1 4 Sugar Effective rate of protection -38 -16 -49 Spread -13 -16 -34 Wheat Effective rate of protection -36 -24 0 Spread 9 1 7 n.a. Not applicable. a. France, Germany, and Italy. Source: Ingco (1995) for the effective rates of protection and author's calculations for the spreads between world and domestic consumer prices. keting and distribution costs is more limited, but the trend in the United States has also been clearly negative, down from 18 percent of gross domestic product (GDP) in 1980 to 10 percent of GDP in 1994 (Council of Logistics Management 1996). Technological progress and new management techniques clearly have contributed to this direction. Among many examples, electronic data interchanges have powered up market-clearing activities, and just-in-time techniques as well as new hedging instruments (for example, warehouse bonds) have reduced con- signment and inventory costs. Given these declining trends, it is highly unlikely that these costs have been a binding constraint on commodity sales and, thus, the major cause for the asymmetric response of domestic prices. The bottleneck approach can partially explain the asymmetric transmission of world commodity prices through rising processing costs, even though their influence was limited by the kind of commodities selected in this article. Unlike transportation and marketing costs, processing costs have certainly increased over time due to higher wages in processing facilities, explaining to some extent the increasing spreads between world and domestic prices observed in the major consumer markets over the past two decades. Higher processing costs can also be explained by the improved quality of consumer products such as unleaded gasoline and high-quality coffee (robusta compared with arabica). This associa- Morisset 51 9 tion can also partially reflect the positive and significant estimated relationship between wages and domestic prices presented in appendix B. Nevertheless, la- bor costs have to play a disproportionate role in sales to fully explain the asym- metric response of consumer prices. As an illustration, the weight of processing costs-measured as the average industrial labor costs in the six countries sur- veyed-in domestic consumer prices would need to be four times greater than that of world prices to compensate for the 100 percent increase in the average spread between world and consumer prices in the commodity markets during 197S-94.2 The differential in the productivity gains through the stages of processing may also partially explain the increasing spreads, especially those between do- mestic wholesale and consumer prices. The argument would be similar to the one presented above because it may reflect faster productivity growth in trad- able activities compared with labor over the past few decades. As argued before, however, this argument would remain partial because it would hardly explain the 45 percent increase in the spreads between world and domestic wholesale prices that have approximately the same nontradable and tradable contents. In light of the caveats of the previous explanations, additional reasons have to be found for the increasing spreads. Among a few possible alternatives, it might be tempting to consider the influence of large international trading com- panies. Although this article does not provide a definitive answer, the strategic position of these companies between buyers and sellers and their concentration in a few companies make it possible for them to affect spreads in most commod- ity markets. Morgan (1979) and Brown (1993) report that six or fewer trading companies control about 70 percent of the total international trade of the com- modities covered in this article. As an example, cereal exports are controlled by five companies: Cargill, Continental, Andre, Dreyfuss, and Bunge-Born. Fur- thermore, many of these companies are vertically integrated and thus capable of influencing both wholesale and retail margins. For example, Cargill-the world's largest trading company of cereals-owns plantations, storage facilities, and vessels around the world. Similarly, many oil companies carry out not only mining and refining but also a complex set of activities involving distribution to wholesalers, transportation, inventories, and pricing to consumers. Those effects would be consistent with the significant fixed-time effects found in the regressions presented earlier be- cause these companies were generally active in all the commodity consumer markets over the period considered and thus were capable of influencing them simultaneously. Finally, preliminary evidence for the oil market shows a high 2. Equation 1 was modified as follows: Alsi! = Api1 - aA(e1p,) -(1- a)Aw; where w, is defined as the average unit labor cost in the recipient country j and ox is the weight of the world commodity price in the production function. The value of the parameter oc must be on average as low as 0.2 to eliminate the increase in the spread between world and domestic prices in commodity markets observed during 1975- 94. These results are available upon request. 520 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 and positive correlation between the profits of the largest U.S. oil companies and the spread between world and domestic consumer prices during 1979-94 (Fortune various issues; the eight companies are Amoco, Ashland Oil, Atlantic Richfield, Chevron, Exxon, Mobil, Philips Oil, and Texaco). Unfortunately, this indicator could not be tested for other commodities due to the absence of basic data on most international trading companies, which are generally not publicly traded and so do not publish their results. The market power of trading companies could be a factor in explaining the increasing spreads observed over the past 25 years, but additional hard evidence is needed to support this explanation. In particular, it remains unclear why those companies would exercise their power asymmetrically in the way suggested by the econometric results presented earlier. There is therefore a need to develop an analytical model including those effects. Empirical evidence is also lacking, prin- cipally due to the general reluctance of these companies to share information. Surprisingly, policymakers, economists, and consumers seem to remain largely unaware of these companies, even though they are often bigger than developing economies and play a determinant role in most commodity transactions world- wide. The current academic literature as well as the international institutions have traditionally ignored their presence. Such insufficient attention partially explains why the debate over these companies lacks focus and clarity and why there are various misconceptions about what these companies actually do and whether their activities are a legitimate cause for public concern. For some ideas along these lines, see Morisset (1997). IV. CONCLUDING REMARKS Prebisch (1950) and Singer (1950) emphasized the relatively low income and price elasticities of demand for commodities about 45 years ago. This article goes one step further by suggesting that the final demand for these products could not have increased in the major consumer markets because the declines in world commodity prices were not transmitted or were transmitted imperfectly to domestic consumer prices. In contrast, upward movements in world prices were clearly passed on to domestic prices. This asymmetry was apparent in all major commodity markets and consumer countries surveyed in this article over the past 25 years. As a result of this asymmetry, the spread between world com- modity prices and domestic consumer prices has increased over time, by about 100 percent on average for the seven commodities analyzed here for 1975-94. Explanations for such patterns remain largely unexplored in the current eco- nomic literature. This article has reviewed several possible explanations for the asymmetry in the price transmission, including changes in trade policies, trans- port and insurance costs, and marketing and processing costs. This was the most logical approach in the absence of a general analytical framework. However, none of these explanations can fully account for the asymmetric response of domestic prices, even though changes in trade policy and in processing costs Morisset 521 may have exerted some influence over time. Many countries have used a two- tier pricing strategy to protect their producers against declines in world com- modity prices, and processing costs have been rising due to higher wages in destination countries. Nevertheless, the country location of the transaction be- tween buyers and sellers does not seem to matter much because the spreads have increased homogeneously in all markets over the past two decades. This article should be viewed as a starting point for future research. Although some progress has been made, the sources for the increasing spreads are poorly understood. Possible directions for understanding better the asymmetric response of domestic prices should include a closer look at the possible differential re- sponse between temporary and permanent changes in world commodity prices. Froot and Klemperer (1989) show that a model with consumer switching costs will lead producers to respond differently to temporary and permanent changes in costs. Additional attention should also be given to changes in the transactions between producers and wholesalers, and between wholesalers and consumers, product by product and country by country. Such detailed analysis represents a challenge but may be necessary for understanding the sources of market power and, if any, the stage of processing at which they are likely to predominate. Finally, this effort should include an analysis of the international trading com- panies that remain largely ignored by the mainstream academic literature on international trade and the actual preoccupations of the multilateral agencies such as the World Trade Organization and the World Bank. Understanding the role and functions of intermediaries in international commodity trade clearly represents an area in need of much more research. APPENDIX A. DATA SOURCES AND DEFINITIONS Domestic Consumer Prices Canada: consumer price index, 1970 and 1975-94. Source: Statistics Canada. France: consumer price index, 1964-94 except fuel (1971-94). Source: INSEE a and b. Germany: consumer price index,1966-94, except for rice, which is not avail- able. Source: Statistisclhes Bundesamt. Italy: consumer price index, 1960-94. Source: ISTAT a and b. Japan: consumer price index, 1970-94. Source: Bank of Japan and Govern- ment of Japan Statistics Bureau. United States: consumer price index, 1960-94, except for coffee (1969-94), rice (1978-94), and sugar (1970-94). Source: U.S. Bureau of Labor. Domestic Wholesale/Producer Prices Canada: producer price index, 1970-94 except for coffee, fuel, gasoline, rice, and sugar (1980-94). Source: Statistics Canada. France: wholesale price index, 1970-94. Source: INSEE a and b. 522 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Germany: producer price index, 1970-94, except for rice, which is not avail- able. Source: Statistisches Bundesamt. Italy: wholesale price index, 1970-90, except for coffee (import price index). Source: ISTAT a and b. Japan: wholesale price index, 1970-94, except for coffee, gasoline, and sugar (import price index). Source: Bank of Japan and Government of Japan Statistics Bureau. United States: producer price index, 1970-94, except for coffee (1969-94), rice (1978-94), and sugar (1970-94). Source: U.S. Bureau of Labor. International Commodity Prices Source for all international commody prices: World Bank data. Beef. All origins, U.S. ports, U.S. cents/pound. Coffee: All coffee, New York, U.S. cents/pound. Crude oil (petroleum): Average crude price, U.S. dollars/barrel. Rice: United States, New Orleans, U.S. dollars/metric ton. Sugar: Caribbean, New York, U.S. cents/pound. Wheat: United States, U.S. gulf ports, U.S. dollars/bushel. APPENDIX B. REGRESSION RESULTS OF THE PANEL OF SIX COUNTRIES, 1976-93 Model with dummy variables for upward and down- General model ward changes in prices Commodity/product Eq. 3a Eq. 3b Eq. 3c Eq. 4a Eq. 4b Eq. 4c Beef/beef Changes in world prices (Ap*) 0.204 0.365 (2.54) (4.44) Upward changes in 0.281 0.394 world prices (Ap1,*) (2.31) (5.62) Downward changes in world 0.113 -0.069 prices (Ap,) (0.85) (-0.56) Changes in wholesale 0.378 prices (Apwil) (4.26) Upward changes in wholesale 0.464 prices (Ap,w,t) (4.08) Downward changes in wholesale 0.112 prices (AP2Wit) (0.48) Changes in wages (Aw,u) 0.269 0.229 0.204 0.253 0.153 0.182 (2.35) (1.95) (1.86) (2.18) (1.42) (1.66) Adjusted R2 0.0141 0.240 0.235 0.137 0.369 0.239 Coffee/coffee Changes in world prices (Ap*) 0.103 0.754 (3.78) (5.02) Upward changes in world 0.151 0.832 prices (Ap1 ) (0.79) (3.90) Morisset 523 APPENDIX B. (continued) Model with dummy variables for upward and down- General model ward changes in prices Commodity/product Eq. 3a Eq. 3b Eq. 3c Eq. 4a Eq. 4b Eq. 4c Downward changes in world 0.048 0.663 prices (AP2 ) (0.23) (2.87) Changes in wholesale 0.393 prices (Aptit) (6.08) Upward changes in wholesale 0.014 prices (Aplw,5) (0.14) Downward changes in whole- -0.230 sale prices (Ap2wit) (-1.71) Changes in wages (Awit) 0.287 0.189 0.355 0.283 0.183 0.320 (1.51) (0.88) (1.90) (1.49) (0.85) (1.86) Adjusted R2 0.028 0.243 0.032 0.019 0.237 0.043 Oil/fuel Changes in world prices (Ap?t) 0.073 0.203 (0.48) (1.47) Upward changes in world 0.292 0.155 prices (Ap,*) (1.29) (0.75) Downward changes in world -0.159 0.255 prices (Ap2,) (-0.68) (1.18) Changes in wholesale 0.329 prices (Apw,t) (2.97) Upward changes in wholesale 0.325 prices (Ap,wit) (2.15) Downward changes in wholesale 0.338 prices (AP2wit) (1.80) Changes in wages (Aw,t) 0.477 -0.316 0.582 0.462 -0.312 0.581 (2.20) (-1.60) (2.87) (2.14) (-1.58) (2.75) Adjusted R2 0.052 0.031 0.136 0.059 0.021 0.126 Oil/gasoline Changes in world prices (Apt) 0.228 0.459 (2.10) (2.92) Upward changes in world 0.385 0.521 prices (AP, *) (2.37) (2.20) Downward changes in world 0.061 0.394 prices (Ap2 it) (0.36) (1.60) Changes in wholesale 0.243 prices (Apwjt) (3.61) Upward changes in wholesa.le 0.301 prices (Aplw,t) (2.07) Downward changes in whol,sale 0.281 prices (AP2W"i) (2.76) Changes in wages (Aw,) 0.241 -0.073 0.296 0.230 -0.78 0.300 (1.55) (-0.32) (2.05) (1.49) (-0.35) (2.01) Adjusted R2 0.080 0.079 0.159 0.087 0.070 0.152 (Continued on following page.) 524 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 APPENDIX B. (continued) Model with dummy variables for upward and down- General model ward changes in prices Commodity/product Eq. 3a Eq. 3b Eq. 3c Eq. 4a Eq. 4b Eq. 4c Rice/rice Changes in world prices (Apt) 0.071 0.403 (0.45) (2.89) Upward changes in world 0.147 0.445 prices (Apl*) (0.62) (2.08) Downward changes in world 0.008 0.369 prices (AP2*z) (0.04) (1.90) Changes in wholesale 0.214 prices (Apwit) (1.94) Upward changes in wholesale 0.389 prices (Aplwi,) (2.71) Downward changes in whole- -0.127 sale prices (Ap2wit) (-0.60) Changes in wages (Aw,r) 0.243 0.364 0.161 0.270 0.362 0.084 (1.11) (1.83) (0.74) (1.32) (1.81) (0.38) Adjusted R2 0.08 0.135 0.047 0.08 0.126 0.073 Sugar/sugar Changes in world prices (Apt) 0.229 0.691 (2.97) (3.85) Upward changes in world 0.115 0.020 prices (Apl*) (0.98) (0.08) Downward changes in world -0.325 0.248 prices (Ap,2) (-3.03) (5.31) Changes in wholesale 0.187 prices (Apwil) (4.69) Upward changes in wholesale 0.189 prices (Aplwut) (2.06) Downward changes in wholesale 0.062 prices (AP,wif) (0.10) Changes in wages (Aw,,) 0.249 0.034 0.276 0.233 -0.064 0.274 (2.26) (0.13) (2.72) (2.11) (-0.26) (2.69) Adjusted R2 0.159 0.144 0.261 0.166 0.237 0.254 Wheat/bread Changes in world prices (Apt) 0.190 0.052 (3.20) (0.48) Upward changes in world 0.244 0.287 prices (Apl,) (2.64) (1.74) Downward changes in world 0.141 -0.159 prices (AP2 *) (1.63) (-1.02) Changes in wholesale 0.194 prices (Apwvi) (3.32) Upward changes in wholesale 0.198 prices (Ap,tvt) (2.54) Downward changes in wholesale 0.186 prices (AP2Wdt) (1.51) Morisset 525 APPENDiX B. (continued) Model with dummy variables for upward and down- General model ward changes in prices Commodity/product Eq. 3a Eq. 3b Eq. 3c Eq. 4a Eq. 4b Eq. 4c Changes in wages (Aw,) 0.322 0.223 0.337 0.316 0.195 0.337 (3.81) (1.45) (4.05) (3.70) (1.128) (4.03) Adjusted R2 0.258 0.020 0.264 0.255 0.046 0.256 Note: The six countries a.re Canada, France, Germany, Italy, Japan, and the United States. Equations 3a and 4a are for the elasticity of consumer prices with respect to changes in world prices. Equations 3b and 4b are for the elasticity of wholesale prices with respect to changes in world prices. Equations 3c and 4c are for the elasticity of consumer prices with respect to changes in wholesale prices. See equations 3 and 4 in the text. For each commodity/product and equation there are 108 observations. t-statistics are in parentheses. Source: Author's calculations. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Amjadi, Azita, and Alexander Yeats. 1995. "Have Transport Costs Contributed to the Relative Decline of Sub-Saharan African Exports?" Policy Research Paper 1559. Policy Research Department, World Bank, Washington, D.C. Processed. Anderson, Kym, and Rod Tyers. 1992. Disarray in World Food Markets. New York: Cambridge University Press. Atkin, Michael. 1992. The International Grain Trade. Cambridge, U.K.: Woodhead Publishing Ltd. Bank of Japan. Various issues. Monthly Statistics of Japan. Tokyo. Brown, M. B. 1993. Fair Trade. London: Zed Books. Council of Logistics Management. 1996. "Annual Conference Proceedings." Oak Brook, Ill. Processed. Fortune. Various issues. 'The Fortune 500 Largest U.S. Industrial Corporations." Chi- cago: Time. Foster, Harry, and Richard Baldwin. 1986. "Marketing Bottlenecks and the Relation- ship between Exchange Rates and Prices." Department of Economics, Massachusetts Institute of Technology, Cambridge, Mass. Processed. Froot, Kenneth, and Paul D. Klemperer. 1989. "Exchange Rate Pass-Through When Market Share Matters." American Economic Review 79(4, September):637-54. Goldberg, Pinelopi K., and Michael Knetter. 1997. "Goods Prices and Exchange Rates: What Have We Learned?" Journal of Economic Literature 35(3, September):1273- 319. Government of Japan Statistics Bureau. Various years. Japan Statistical Yearbook. Tokyo. Ingco, Melinda D. 1995. "Agricultural Trade Liberalization in the Uruguay Round: One Step Forward, One Step Back." Policy Research Paper 1500. Policy Research Department, World Bank, Washington, D.C. Processed. 526 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 International Monetary Fund. Various issues. International Financial Statistics. Wash- ington, D.C. INSEE (Institut National de la Statistique et des Etudes Economiques). Various issues a. Bulletin mensuel de statistique. Paris. . Various issues b. Annuaire statistique de la France. Paris. ISTAT (Instituto Nazionale di Statistica). Various issues a. Boletin mensile di statistica. Rome. . Various issues b. Annuario statistico italiano. Rome. Knetter, Michael M. 1992. "Is Price Adjustment Asymmetric? Evaluating the Market Share and Marketing Bottlenecks Hypothesis." NBER Working Paper 4170. National Bureau of Economic Research, Cambridge, Mass. Processed. . 1993. "International Comparisons of Pricing-to-Market Behavior." American Economic Review 83(3, June):473-86. Mitchell, Donald O., and Ronald C. Duncan. 1987. "Market Behavior of Grain Export- ers." The World Bank Research Observer 2(1, January):3-21. Morgan, Dan. 1979. Merchants of Grain. New York: Viking Press. Morisset, Jacques. 1997. "Unfair Trade: Empirical Evidence from Commodity Markets Over the Past 25 Years." Policy Working Paper 1815. Policy Research Department, World Bank, Washington, D.C. Processed. Mundlak, Yair, and Donald F. Larson. 1992. "On the Transmission of World Agricul- tural Prices." The World Bank Economic Review 6(3, September):399-422. Prebisch, Raoul. 1950. "The Economic Development of Latin America and Its Principal Problems." Economic Bulletin for Latin America 7. New York: United Nations. Singer, H. W. 1950. "The Distribution of Gains between Investing and Borrowing Coun- tries." American Economic Review 15(May):473-85. Statistics Canada. Various issues. Consumer and Price Indexes. Ottawa. Statistisches Bundesamt. Various issues. Statistiches Jahrbuch. Wiesbaden. U.S. Bureau of Labor. Various issues. Producer and Consumer Prices Index: Bureau of Labor Statistics Data. Washington, D.C. World Bank. 1996. "Commodity Markets and the Developing Countries." World Bank Quarterly (May). A NEW DEVELOPMENT DATABASE The following article is one in an occasional series introducing new databases. The series intends to rnake new development databases more widely available and to contribute to diiscussion and further research on economic development issues. The databases included in the series are selected for their potential useful- ness for research and policy analysis on critical issues in developing and transi- tion economies. Some are drawn from micro-level firm or household surveys; others contain country-level data. The authors describe the data contents, crite- ria for inclusion or exclusion of values, sources, strengths and weaknesses, and any plans for maintenance or updating. Each database is available on the Inter- net, at the address pro vided in the article. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: S29-47 A. Database of World Stocks of Infrastructure, 1950-95 David Canning This article describes an annual database of physical infrastructure stocks for a cross- section of 152 countries for 1950-95. The database includes estimates of six measures of infrastructure: the number of telephones, the number of telephone main lines, kilo- watts of electricity-generating capacity, kilometers of total roads, kilometers of paved roads, and kilometers of railway lines. Both raw and manipulated data sets, in which series have been linked to overcome changes in definition and coverage, are reported. Some measures of infrastructure quality, such as the percentage of roads in poor con- dition, the percentage of local telephone calls that do not go through, the percentage of diesel locomotives available for use, and the percentage of electricity lost from the distribution system, are included. The data on all series except total roads are of rea- sonably good quality and should prove useful to researchers. The article also presents regression results relating stocks of infrastructure to popu- lation, per capita gross domestic product, land area, and level of urbanization. It shows that stocks of telephones, electricity-generating capacity, and paved roads tend to in- crease proportionately with population and more than proportionately with per capita gross domestic product. Both the length of total roads and the length of total rail lines rise with country size and are relatively insensitive to population and income. Physical infrastructure has long been considered an important determinant of economic growth. Aschauer (1989), for example, finds very large returns to public capital in the Ulnited States. Canning, Fay, and Perotti (1992, 1994) esti- mate large growth effects of physical infrastructure. Easterly and Rebelo (1993) find that public investment in transportation and communication is consistently correlated with economic growth. Lee and Anas (1992) find lack of infrastruc- ture, particularly lack of a consistent supply of electricity, to be a major con- straint on firms in Nigeria. Antle (1983) finds a significant role for infrastruc- ture in agricultural productivity in developing countries. (For a review of the literature on the importance of infrastructure to economic development and an evaluation of empirical results estimating the contribution of public capital and infrastructure to gromwth, see World Bank 1994, Gramlich 1994, and Jimenez 1995.) David Canning is with the Department of Economics at Queen's University of Belfast and is currently visiting Harvard University. The author gratefully acknowledges research funding from the World Bank and the comments and assistance with data collection of Esra Bennathan. © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK s29 530 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 A recurrent problem in this literature, as noted by Jimenez (1995), is the lack of data. The main aim of this article is to provide and describe a data set on the stock of physical infrastructure-telephones, telephone main lines, electricity- generating capacity, roads, paved roads, and railway lines. Physical measures of infrastructure stocks are used because of the problems associated with using investment data to estimate infrastructure capital. Because of differences in the efficiency of the public sector and the price of infrastructure capital, the same level of investment in infrastructure may yield very different results across coun- tries (Pritchett 1996). Moreover, investment figures represent annual flows into infrastructure. Deriving estimates of infrastructure stock at a point in time re- quires the use of perpetual inventory methods, which may introduce systematic errors in stock estimates. The annual time series for infrastructure stocks presented here goes back as far as 1950, the earliest year for which the Summers and Heston (1991) Penn World Tables data on purchasing power parity gross domestic product (GDP) are available.' The 152 countries reported are the same as those covered in the Penn World Tables. An earlier data set constructed by Canning and Fay (1992) and published by the World Bank (1994) gives infrastructure data on a quinquennial basis back to 1960. Although for many purposes quinquennial data are suffi- cient, a full set of annual data allows more detailed investigation of the time- series properties of the data. In addition, the data reported here have been im- proved by cross-checking different sources, reconciling differences by referring to more detailed national sources, and linking series in which breaks occur as a result of changes in definition or coverage. This data set should therefore be seen as superseding the original data set. Two types of data files are included. The first reports raw data. All of the data in this set are reported exactly as they appear in the original sources (except for road lengths in miles, which have been converted to kilometers, and adjustments to deal with border changes). Where different sources report different figures, the series that seems closest to our definition of the relevant variable is used. In many cases, disaggregated national sources have been studied to determine which figures best represent a particular stock of infrastructure. When different sources appear equally plausible, official national sources are used first, international sources collected from government agencies second, and international sources collected from nongovernmental agencies last. The only manipulation of these raw data is the suppression of what appear to be misprints in data sources or implausibly large year-to-year changes in reported stocks. A problem with these raw data is that within a single time series, the report- ing source, the definition of the stock variable, or the coverage of the variable may change. It nevertheless seems desirable to make the raw data available to researchers, so that they can interpolate or transform the data as desired. 1. The database is available on the Internet at http:l/www.worldbank.org/html/dec/Publications/ Workpapers/WPS1 900series/WPS1929/canningl .xls. Canning 531 For some variables, additional data sets have been created by systematically linking the data series where breaks occur as a result of changes in definition or coverage. Overlapping data are used where available to construct proportional indexes; that is, all of the data before the break have been changed proportion- ately to bring them into line with data after the break. Where it is not possible to link series, the less appropriate series has been deleted. Some countries measure infrastructure stocks only at infrequent intervals, reporting the same figure for a number of years. In cases where this practice is documented, the constructed data set deletes repeated instances of the same value, so that the values given relate only to the actual year of measurement. Infrastructure stoclks usually move very slowly, and there is scope for interpo- lation in some series with very few gaps in the data. Most gaps are short (one to two years), but we have interpolated over gaps of up to five years in the con- structed data. All interpolation is carried out so as to be linear in the logs of the infrastructure stock (that is, growth is assumed to be exponential over the inter- vening period). The data for telephones, telephone main lines, length of paved road, and length of railway line have been manipulated in this way to produce constructed data sets. For these variables, the constructed data rather than the raw data are recommended for use, and the results reported here are based only on the manipulated data. In the case of electricity-generating capacity, the raw data appear so good tChat no manipulation has been carried out. For total roads, the definition and coverage of the data vary too much over time and across countries to produce a consistent series. If, for a particular purpose, interpola- tion and linking of series are not desirable, it is possible to derive a data set that contains only high-quality reported data by imposing the condition that the same figure be reported in both the raw and the manipulated data sets. The provision of infrastructure reflects the forces of demand and supply and the effect of public policy. Public policy may, in fact, play a very large role, because in many cases the price mechanism works imperfectly or is absent in the provision of infrastructure. In the absence of price variables for infrastructure, a complete model of infrastructure provision is not possible. It is nonetheless pos- sible to note the positive relationships between infrastructure provision and eco- nomic development. For example, Queiroz and Gautam (1992) find a signifi- cant correlation between kilometers of paved roads and GDP in a cross-country study. Ingram and Liii (1997) find stable relationships between provision of roads and indicators cf economic development, both at the municipal and na- tional levels. A problem with usinIg measures of physical infrastructure capital is that they do not reflect the quality of the services provided. Although figures for paved roads as well as total roads and for telephone main lines as well as telephones are given, they may stillI conceal possibly large differences in quality across coun- tries. Hulten (1996) argues that the management and efficient use of infrastruc- ture may be more important than the quantity. To address this issue, some mea- 532 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 sures of infrastructure quality and efficiency of use are included. These measures are available for only a limited number of years, however. The data reveal a very strong relationship between infrastructure and mea- sures of economic development and geography. For telephones and electricity- generating capacity, the stock of infrastructure rises proportionately with popu- lation and more than proportionately with per capita income. Urbanization is positively correlated with the number of telephones in poorer countries, while country size is negatively correlated with the number of telephones in richer countries. For transportation infrastructure, a different pattern emerges. The provision of total roads and rail lines rises less than proportionately with population and level of income per capita but more than proportionately with area. In contrast, on average paved roads increase almost proportionately with population and more than proportionately with income, with the association between income and paved roads being stronger in rich countries. Area appears to have no significant relationship with the provision of paved roads. These results suggest that rail lines and unpaved roads operate below capacity and represent nonrival public goods, while, at least in rich countries, paved roads generally operate near full capacity. Although paved roads are generally consid- ered to be nonexcludable public goods, congestion effects may transform them into rival goods. These interpretations assume, of course, that the level of infra- structure provision is close to the efficient level, an assumption that may not be valid. Given the large-scale involvement of government, patterns in infrastruc- ture stocks may be explained better by political economy than by economic efficiency. However, all these relationships are probably equilibrium outcomes and do not directly reflect either the demand or supply functions. The positive relation- ship between infrastructure provision and income cannot be interpreted as re- flecting income elasticity of demand unless the price of infrastructure is held constant across countries. Preliminary price data for road construction show that prices vary systematically across countries, with real prices in middle- income countries averaging about two-thirds those in rich countries and poor countries. This suggests that the stable relationship between per capita GDP and infrastructure stocks may be the result of a complex interaction of demand and supply effects. The determinants of infrastructure growth rates during 1965-85 are also examined. These growth regressions test the robustness of the cross-sectional infrastructure relationships and their stability over time: if the cross-country relationships represent equilibrium conditions, the growth rates of the infra- structure stocks should respond to disequilibrium in the relationship. In fact, significant disequilibrium adjustment is found for every type of infrastruc- ture. In this article, the emphasis is on describing the construction of the data set and some correlations between the infrastructure data and economic develop- Canning 533 ment. The central questions of the direction of causation and the size of any causal effects are left to future research. 1. THE DATA Data on six measures of physical infrastructure stocks-the number of tele- phones, the number of telephone lines, kilowatts of electricity-generating capac- ity, kilometers of road, kilometers of paved road, and kilometers of railway line-from various international publications were collected for 152 countries for the period 1950--95. (The main international data sources for each variable are given in the appendix.) These data were then supplemented with data from national sources as necessary. Telephones and Telephone Main Lines For telephones the various data sources are consistent with one another and pro- duce series that appear continuous over time. The basic sources of data on tele- phones are the International Telecommunications Union's Yearbook of Common Carrier Statistics and the American Telephone and Telegram Company's publica- tion The World's Telophones. These sources provide almost identical figures and are in agreement with data from the United Nations' Statistical Yearbook. The only difference between the data sources appears to be the point within the year at which the stock of telephones is measured. The World's Telephones measures stocks as of 1 January, while the Yearbook of Common Carrier Statis- tics measures stocks at different dates in different countries. For many countries, the figure that appears in the Yearbook of Common Carrier Statistics is the same figure that appears in The World's Telephones for the following year. I have used the Yearbook of Common Carrier Statistics as my main source, be- cause its coverage is more comprehensive. Where other sources are used to ex- tend the series or fill the gaps, the year reported is adjusted so that the overlap- ping part of the series; agrees with data from the Yearbook of Common Carrier Statistics. Provision of telephone services is measured by both the number of telephone sets and the number of main lines connected to local telephone exchanges. The number of telephone main lines seems to be the better measure of the capacity of a telephone system (although in practice the two measures are almost perfectly correlated). In theory, a better measure of infrastructure stock might be the capacity of telephone exchanges. To reflect the use of cellular phones, which have been in use since 1982, the infrastructure stock should reflect the area over which cellu- lar calls are possible, as well as the number of cellular phones. The data reported here include cellular phones; detailed data on this topic are available from the International Telecommunications Union. Data on the numbe.r of telephone sets are fairly comprehensive for the period 1950-95; data on telephone main lines are sparse in the earlier years. In addi- 534 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 tion to the raw data, a data set with interpolation of periods up to five years is also provided. Interpolation is carried out linearly in the log of the variable. The percentage of local calls that were unsuccessful in 1990 is used as an indicator of quality. Where data for 1990 are not available, data for the nearest available later year (up to 1995) are used. The data on quality indicate high variations in the percentage of calls that fail, from none up to 98 percent, indicating that a pure quantity measure may be deficient as a proxy for telecommunications services. Electricity The main sources of data on electricity-generating capacity are the United Nations' Energy Statistics and Statistical Yearbook. The time series seem good and are reported without adjustment, yielding a fairly complete data set for the period 1950-9S. Data for Botswana, Lesotho, Namibia, and Swaziland are in- cluded in the data for South Africa. The raw data are not manipulated because there appear to be no breaks in the series and there is little to be gained by inter- polation. These data on capacity do not take into account the extent of the electricity distribution system. The percentage of generated electricity lost in the system is used as an indicator of quality for 1971, 1980, and 1990. Total Roads The data on roads come from two international sources, the International Road Federation's World Road Statistics and the statistical yearbooks of the regional commissions of the United Nations, as well as various national sources. World Road Statistics is based on data supplied by the contracting industry in each country. The earliest data available are for 1958, and the coverage of the data set expands with time, particularly during the 1970s. Where the two inter- national sources disagree, I tend to use the United Nations data, which are re- ported from official government sources. The international data on total roads are patchy, with frequent gaps and many large changes that are often quickly reversed. Different countries define roads differently, and the definition of a road often changes within countries over time. The definition of minimum quality standards for roads varies, and differ- ences in reporting reflect the functional split of road management between cen- tral and local government. Large changes in the series occur, for example, when the reporting source changes the nature of its coverage (from coverage only of roads controlled by the central government, for example, to roads controlled by central and provincial governments). Intraurban roads above a certain quality threshold are often centrally controlled, while urban roads are controlled by municipal authorities, leading to an underreporting of urban and low-quality rural roads controlled by the central authority. National sources are used to supplement the data from international sources. Use of these sources increases the coverage of the data substantially and pro- duces more consistent numbers. When the national sources agree broadly with Canning 535 the international sowtrces, the national sources are used as the primary data source, with gaps filled from international sources where possible. Study of national sources also reveals, to some extent, the causes of discontinuities in the interna- tionally reported data. As far as possible, the reported time series for total roads includes urban roads and reflects total length of public roads in the country, regardless of the controlling authority. Some results using total roads are reported here for com- parison, but for the most part the raw data seem too unreliable to be useful. Even with the use of national sources, it appears impossible to construct data that are consistent either across countries or over time. No attempt is made to manipulate the raw data since there is a lack of documentation on basic ques- tions of definition. Paved Roads Paved roads are cefined as concrete or bitumen-surfaced roads. This defini- tion excludes stone, gravel, water-bound gravel, oil-bound gravel, and earthen roads and is consistent with the definition used by most countries. Where a country defines paved roads differently, the data are adjusted to fit our narrow definition. China, for example, defines paved roads much more broadly. The data reported here for China are therefore taken from a World Bank country report rather than from official Chinese sources. In many cases, detailed data on the type of road are not available, and the national definition is used. When national sources give figures for "paved" or "hard-surfaced" roads and no other information is available, these categories are accepted as equivalent to the nar- row definition of paved roads used here. For many countries, international sources report only rural roads. Using na- tional sources, we are able to construct, for some countries, the total stock of paved roads by adding urban and rural paved kilometers. However, for many countries, it is unclear even in national sources exactly what the coverage is. In countries for which both urban and rural data are available, urban roads make up about 15-30 percent of the total stock of paved roads. Even using national data, however, the series still exhibits large changes and splits, making the raw data for paved roads too inconsistent for practical use. To make the series consistent over time, it is necessary to link series following changes in definition and coverage where these are documented. The processed data are then interpolated over gaps of up to five years. The resulting processed data are the best currently available. Within the category of paved roads, there may be large variations in quality. In particular, the data do not reflect the width of the road, which varies from single-lane roads to multiple-lane highways. The percentage of the main paved and unpaved road network that the World Bank considers to be in good, fair, and poor condition in 1984 and 1988 is used as a measure of quality. These data cover most developing countries. The quality measures refer to the main road network and may not be representative of the total road network. In addition, 536 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 they make no allowance for the age of the road stock and thus may not be good indicators of maintenance levels. The data file on infrastructure quality contains information on the data source used for paved roads in each country and on the coverage of the data, where this information is available. Railways The main sources of data on length of railway lines are Mitchell's Interna- tional Historical Statistics (1992, 1993, 1995) until 1980 and World Bank data thereafter. National sources are also used to supplement these data. The data refer to length of line, where a line may consist of two or more tracks. The only problem with these data seems to be changes in coverage due to the treatment of rail lines owned by companies for industrial use and not open to the public (such as railways owned by the sugar industry in Latin America). To produce a consis- tent series, the data are limited to railroads open to the public. The processed data link the series over breaks in the data. Interpolation is also carried out, although the gaps in the railway line series tend to be very minor. Countries reporting no railways in national or international sources at any time during 1950-95 are assumed to have zero line length in the processed series. The percentage of the stock of diesel locomotives available for use mea- sured between 1990 and 1995 is given as a measure of quality. Assessment and Recommendations The data sets for telephone sets, telephone main lines, and electricity- generating capacity seem excellent, and both the raw data sets and the manipu- lated (interpolated) data can be used without worry. The data for railways seem good, although for a number of countries the series have to be linked to achieve consistency over time. For this series, the processed data should be used, par- ticularly for time-series work. The data for total roads are unreliable. Unexplained fluctuations in the series seem too numerous to allow the construction of consistent series. Because most of the large changes reflect changes in definition rather than changes in the stock of roads, these data are not suitable for time-series work. They may be used for cross-sectional work if allowance is made for the fact that different definitions across countries mean that the series contain large measurement errors. This problem also affects the raw data series for paved roads. There, however, most of these problems are overcome by processing, and the series are consistent over time. For cross-sectional work, countries that report only an administrative subcategory of paved roads should not be used. Using the remaining countries- countries that report either total paved roads or rural paved roads or do not define clearly what they are reporting-in cross-sectional studies introduces an under- reporting error of about 15-30 percent of the total stock of roads for countries reporting only rural roads. Measurement error of this magnitude should be accept- able if the data are to be used for cross-sectional studies. Canning 537 II. CROSS-COUNTRY RELATIONSHIPS: INFRASTRUCTURE, INCOME, AND GEOGRAPHY For each data set, the mean, standard deviation, minimum, maximum, and number of observat:ions per 1,000 population are reported for each country. Table 1 provides summary descriptive statistics. As we shall see, this may not be an appropriate normalization for all the variables in the data set. The year 1985 is used as a base year for comparison, because it has good coverage of all the relevant measures of infrastructure. The correlation between stocks of infrastructure is also examined (table 2). As expected, all of the infrastructure stocks are positively correlated. The corre- lation between the number of telephones and the number of telephone main lines is greater than 99 percent, indicating that the number of telephones can be used as a proxy for the number of telephone main lines. Paved roads and total roads are correlated only loosely, suggesting that one series cannot be used as a proxy for the other. Table 1. Infrastructure per 1,000 Population, 1985 Standard Number of Type of infrastructure Mean deviation Minimum Maximum observations Number of telephone main lines 99.1 144.3 0.45 627.8 146 Number of telephones 137.9 196.5 0.59 840.1 129 Kilometers of paved roads 2.82 4.38 0.06 24.51 116 Kilometers of total roads 6.71 7.99 0.11 54.13 137 Kilometers of rail lines 0.29 0.48 0.00 3.74 150 Kilowatts of electricity-generating capacity 0.62 0.95 0.002 5.59 146 Source: Author's calculations. Table 2. Correlation of Infrastructure Levels, 1985 Kilowatts of Number of electricity- Kilometers Number of telephone generating of paved Kilometers Kilometers Type of infrastructure telephones main lines capacity roads of roads of rail lines Number of telephones 1.00 Number of telephone main lines 0.99 1.00 Kilowatts of electricity- generating capacity 0.96 0.96 1.00 Kilometers of paved roads 0.69 0.70 0.83 1.00 Kilometers of roads 0.54 0.53 0.71 0.83 1.00 Kilometers of rail lines 0.36 0.37 0.59 0.79 0.84 1.00 Source: Author's calculations. Table 3. Cross-Country Patterns of Infrastructure, 1985 Log Log telephone Log electricity- Log Log Log Item telephones main lines generating capacity paved roads total roads rail lines Constant -7.93 -8.22 -13.0 -8.10 -0.771 -5.58 (11.0) (13.2) (21.0) (11.1) (11.1) (7.63) Log population 1.001 0.997 0.931 0.798 0.544 0.552 (17.5) (21.5) (21.5) (9.98) (9.14) (9.36) Log per capita GDP 1.479 1.442 1.398 1.243 0.583 0.820 x0 (13.9) (14.8) (17.6) (10.8) (6.57) (8.42) Percentage of population 0.891 1.221 1.167 -0.661 -0.217 -0.634 living in urban area (2.07) (3.10) (3.96) (1.25) (0.56) (1.08) Log area -0.107 -0.095 0.081 0.108 0.335 0.352 (2.50) (2.60) (2.41) (1.50) (6.72) (5.24) Number of observations 126 144 142 112 133 106 Adjusted R2 0.925 0.938 0.932 0.842 0.866 0.721 Note: Heteroscedastic-consistent t-ratios are in parentheses. Source: Author's calculations. Canning 539 It seems likely that the stock of infrastructure in a country varies with popu- lation and per capita GDP. These variables affect the demand for infrastructure, as well as the cost of providing it. Geography may also matter. Hong Kong (China) and Singapore, for example, both have very low stocks of paved roads relative to the size of their population and the level of per capita GDP. It may be that in such city states, where population density is high, the need for roads is low. This relationship can be tested by developing a measure that reflects the percentage of the population living in urban centers and the total area of the country. Such a measure may not be a good proxy for geography, however, because infrastructure often has network effects, and the precise shape of a coun- try, as well as the location of mountain ranges and rivers and the distribution of population, may affect the provision of infrastructure. For lack of a better mea- sure of geography, however, this measure is used here. Ordinary least squares regressions have been run for infrastructure levels on these factors using data from a cross-section of countries in 1985 (table 3). All variables other than the ratio of people in urban centers are in logarithms, so the coefficients can be interpreted as elasticities. The t-ratios given are heteroscedastic consistent. For nonitransportation infrastructure, the coefficient on population is significant and close to 1, indicating that holding other factors constant, infra- structure rises proportionately with population. The coefficient on GDP per capita is greater than 1, indicating that stocks of nontransportation infrastructure rise more than proportionately with income. The most interesting variables in the regression are the geographical factors- urbanization and area-which have different effects for different types of infra- structure. For example, holding constant population, per capita GDP, and per- centage of the population living in urban areas, a country with more land area has proportionately more electricity-generating capacity and fewer telephone main lines than a smaller country. To understand why these relationships may hold, consider a country with two population centers that have to be linked by communication infrastructure. In a large country, the two centers are likely to be farther apart, so the length of each link will be longer and the cost of each link will be higher. Because of the higher cost per link, the total number of links in a large country is likely to be smaller, while the total length of the links is likely to be greater. It is possible that while large countries have fewer telephone main lines, the length of their main lines (which is not measured) is greater than in small countries. In the case of electricity, generating capacity, not the number of connections or the total length of the distribution system, is measured, so the effects of geog- raphy are likely to be quite different from the effects on the other infrastructure measures. One explanation for the rise in generating capacity with area is that electricity distribution systems suffer leakage, which depends on the length of the connection. In large countries with low population density, leakage can be avoided by operating small local plants. Doing so, however, may reduce the Table 4. Cross-Country Patterns of Infrastructure in Less- and More-Developed Countries, 1985 Log Log telephone Log electricity- Log Log Log Item telephones main lines generating capacity paved roads total roads rail lines Less-developed countries Constant -6.23 -7.49 -14.3 -5.16 1.32 -2.70 (4.57) (5.88) (9.54) (3.39) (1.17) (1.67) Log population 0.906 0.946 0.940 0.807 0.503 0.490 (12.3) (15.4) (13.9) (7.11) (6.07) (5.45) Log per capita GDP 1.291 1.358 1.473 0.739 0.318 0.395 (6.72) (7.3) (6.42) (3.36) (1.96) (1.89) Percentage of population 1.953 2.051 2.563 0.762 1.002 0.129 living in urban area (2.60) (3.10) (3.34) (1.03) (1.69) (0.101) Log area -0.086 -0.093 0.141 0.120 0.292 0.429 (1.30) (1.68) (1.95) (1.19) (4.29) (3.79) Number of observations 63 72 69 57 66 52 Adjusted R2 0.863 0.893 0.867 0.765 0.826 0.640 More-developed countries Constant -6.81 -6.94 -13.2 -11.0 -2.16 -7.73 (5.27) (5.32) (16.6) (6.60) (1.60) (5.04) Log population 1.179 1.124 0.941 0.832 0.642 0.619 (16.4) (17.9) (18.7) (7.59) (7.97) (6.98) Log per capita GDP 1.277 1.263 1.472 1.606 0.707 1.076 (7.97) (7.25) (16.5) (7.12) (4.23) (6.01) Percentage of population 0.093 0.409 0.396 -1.337 -0.893 -1.095 living in urban area (0.22) (0.71) (2.21) (1.68) (1.94) (1.46) Log area -0.177 -0.140 0.039 0.086 0.313 0.281 (3.27) (3.00) (1.15) (0.89) (4.72) (3.27) Number of observations 63 72 73 55 67 54 Adjusted R2 0.925 0.931 0.964 0.840 0.886 0.724 F-test for parameter equality F(5,116) = 4.31 F(5,134) = 3.64 F(5,132) = 3.39 F(5,102) = 1.70 F(5,123) = 2.15 F(5,96) = 1.21 between subsamples p = 0.004 p = 0.001 p = 0.006 p = 0.141 p = 0.063 p = 0.311 Note: Less-developed countries are those with annual per capita income in 1985 less than $2,500; more-developed countries are those with annual per capita income in 1985 of $2,500 or more. Heteroscedastic-consistent t-ratios are in parentheses. Source: Author's calculations. Canning 541 scope for economies of scale and increase the need for reserve capacity for peak periods if transfers within the system are difficult. The provision of electricity and telephones tends to increase with urbaniza- tion, a result that is consistent with the lower cost of providing these services in cities, where the cost of connecting a consumer is lower. Alternatively, the large urbanization effect may reflect the fact that urbanization is a proxy for indus- trial structure: higher rates of urbanization are associated with more manufac- turing and less agricultural production. If manufacturing output has a greater need for electric power than agriculture, industrial structure may be very im- portant. Turning to transportation infrastructure, we see somewhat different relationships. The results for paved roads and total roads differ significantly from each other, possibly reflecting differences in the nature of the services provided by the two types of infrastructure. One reason for this difference may be the fact that paved roads handle large volumes of traffic, while unpaved roads link places together that have low levels of traffic flowing between them. Holding other factors constant, an increase in area significantly increases the length of road, while it has a statistically insignificant impact on the provision of paved roads. Because the average distance between two points increases by the square root of the increase in area, a coefficient of 0.5 on log area is expected to correct for distances. The fact that paved roads increase proportionately (or more than pro- portionately) with income and almost proportionately with population while total roads increase less than proportionately with both factors may reflect the fact that paved roads are rival goods due to congestion effects, while unpaved roads typically have excess capacity. The results for railways are similar to those for total roads, suggesting that they may serve similar functions. The regression results for railways cover only those countries with a positive stock of railways, however; a more sophisticated approach would also include those countries with zero rail length. Regression results for countries with 1985 per capita income of more than and less than $2,500 dollars a year are reported separately to reveal how these relationships differ across the two groups of countries (table 4). Results of an F-test for parameter stability across the two subsamples are also shown. For nontransportarion infrastructure, parameter stability is rejected at the 1 percent level of significance. The main difference in parameters between the two subsamples appears to be that urbanization has a stronger positive effect on infra- structure in less-developed countries than in more-developed countries. The num- ber of telephones and telephone main lines seems to decline with area in richer countries, but not in poorer ones. In the case of transportation infrastructure, the parameter stability cannot be rejected, even at the 5 percent level of significance, although there does seem to be some indication that the elasticity of transporta- tion infrastructure with income is higher in more-developed countries. While the R2 coefficients for each of the regressions in tables 3 and 4 are high, the regressions may not explain infrastructure levels. In particular, they have Table 5. Infrastructure Growth Regressions, 1965-85 (growth rates) Telephone Electricity-generating Paved Total Rail Item Telephones main lines capacity roads roads lines Constant -2.929 -1.978 -5.071 -2.668 -2.015 -0.492 (4.70) (2.41) (5.44) (2.27) (3.35) (0.88) Growth of population 1.034 1.072 1.341 -0.326 0.967 0.254 (4.03) (3.22) (4.10) (0.96) (3.34) (0.94) Growth of per capita GDP 0.926 0.859 0.982 0.887 0.040 0.389 (8.29) (6.23) (6.74) (5.72) (0.36) (3.74) Change in urbanization ratio 2.645 0.028 2.030 -0.813 1.228 -0.964 (4.33) (3.84) (2.77) (0.86) (1.97) (1.59) Log population, 1965 0.320 0.316 0.323 0.305 0.276 0.008 (5.30) (4.17) (4.63) (3.87) (6.72) (0.18) Log per capita GDP, 1965 0.572 0.341 0.592 0.619 0.323 0.085 (5.88) (2.69) (5.39) (4.02) (4.31) (1.17) Percentage of population -0.200 0.000 0.011 -1.372 -0.292 -0.192 living in urban area, 1965 (0.787) (0.08) (0.03) (2.87) (1.12) (0.74) Log area -0.094 -0.128 0.036 0.201 0.041 0.106 (3.45) (3.91) (1.04) (4.70) (1.04) (2.83) Log 1965 stock of -0.271 -0.202 -0.343 -0.513 -0.304 -0.117 relevant infrastructure (4.75) (2.58) (5.71) (9.00) (5.27) (2.09) Number of observations 105 79 113 80 79 85 Adjusted R2 0.682 0.642 0.553 0.712 0.574 0.182 Note: Heteroscedastic-consistent t-ratios are in parentheses. Source: Author's calculations. Canning 543 nothing to say about the directions of causation; it may be that per capita GDP and urbanization rates depend on the provision of infrastructure. The important point is that the provision of infrastructure is significantly correlated with geog- raphy, particularly for poorer countries, probably because the costs and benefits of infrastructure vary with geography. This implies that the impact of infra- structure on economic growth may depend on geography and that geographical considerations should be taken into account when analyzing these effects. In terms of applications, it is usual to normalize quantity variables in order to make them independent of the size of the country. For telephones, telephone main lines, electricity-generating capacity, and perhaps paved roads, it seems reasonable to norrnalize by population, because each of these variables appears to increase proportionately with population. In other words, the regressions in tables 3 and 4 could be rerun using infrastructure stock per capita on the left and excluding population on the right without changing the other coefficients. Total roads and railways do not increase proportionately with population, however. For rival goods, normalization by population seems appropriate, because the quantity of the good divided by the population indicates average consumption. For nonrival goods, however, normalizing by population does not yield average per capita consumption; with a fixed stock of nonrival infrastructure, increases in population neecl not reduce average consumption. If transportation infra- structure is really nonrival, then normalizing by population is unlikely to be appropriate. Ingrain and Liu (1997) normalize by area, but doing so yields a coefficient of less than 1 in our regressions; explaining total roads per square kilometer of area requires that area appear as an explanatory variable. A case could be made for using the square root of area on theoretical grounds, but the data appear to support a figure nearer to the cube root of area. We are thus left with no obvious normalization for stocks of transportation infrastructure. If the cross-sectional regressions generate stable relationships, they may re- flect cointegrating mechanisms for the data. That is, a country that is out of line given the cross-sectional relationship might be expected to move into line in the long run. To address this question, while avoiding the pitfalls of panel estima- tion, regressions are run in which the dependent variable is the growth rate of infrastructure stock during 1965-85 (table 5). (The same level variables are used as in table 3.) These regressions are similar to those of Barro (1991) for eco- nomic growth; a neg,ative coefficient on the initial stock of infrastructure indi- cates convergence of infrastructure stocks to an equilibrium level that depends on the other variables in the regression. Infrastructure stocks in each country appear to be converging to the same long-run equilibrium relationship (defined as the steady state, where all growth rates are zero), conditional on the values of the other initial condition included in the regressions. The long-run steady state can be determined by first postulat- ing that all growth rates on the right-hand side of the regression in table 5 are zero. Doing so yields a growth rate of telephones of 544 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 (1) 0.27(1.18 log population + 2.11 log per capita GDP - 0.74 percent urbanized - 0.35 log area - log telephones) (where the parameters have been reduced by a common factor of 0.27). About 27 percent of the gap between actual telephone stocks and the "equilibrium" telephone stock (when this expression is zero) is closed in a 20-year period. Setting this expression to zero (that is, assuming the growth rate of telephones is zero), it appears that telephones in each country are converging to the following relationship: (2) log telephones = 1.18 log population + 2.11 log per capita GDP - 0.74 percent urbanized - 0.35 log area. The steady-state relationships for all of the infrastructure variables in table 5 can be determined in a similar way. Estimated in this way, the long-run steady- state coefficients are remarkably similar to those found in table 3. III. CONCLUSIONS This article describes a new set of panel data on stocks of infrastructure in a cross-section of countries over time. The stock of infrastructure across countries varies significantly with their population size, income level, and geography, and these relationships appear stable over time. The data sets for telephones, telephone main lines, electricity-generating ca- pacity, and railway lines for 1950-95 are fairly complete. Some minor additions could be made if data become available to fill some gaps, but the data presented here are not likely to change. The data for total roads and paved roads may change substantially for a number of countries if better national data sources are found. In addition, it is hoped that a more definite classification of the variable for coverage of the stock of paved roads will eventually be possible. APPENDIX. DESCRIPTION AND SOURCES OF NONSURVEY VARIABLES Variable Source GDP per capita Penn World Table 5.6 (Summers and Heston 1991) Land area of country World Bank (1995) Percentage of population living in World Bank (1995) urban areas Kilowatts of electricity-generating United Nations, Energy Statistics (various years); United capacity Nations, Statistical Yearbook (various years) Number of telephones International Telecommunications Union (various years); American Telephone and Telegram Company (various years) Number of telephone main lines International Telecommunications Union (various years); American Telephone and Telegram Company (various years) Canning 545 Kilometers of paved roads International Road Federation (various years); United Nations Economic Commission for Africa (various years); United Nations Economic and Social Commission for Western Asia (various years); United Nations Economic and Social Commission for Asia and the Pacific (various years); United Nations Economic and Social Commission for Latin America and the Caribbean (various years); United Nations Economic Commission for Europe (various years); national sources Kilometers of roads International Road Federation (various years); United Nations Economic Commission for Africa (various years); United Nations Economic and Social Commission for Western Asia (various years); United Nations Economic and Social Commission for Asia and the Pacific (various years); United Nations Economic and Social Commission for Latin America and the Caribbean (various years); United Nations Economic Commission for Europe (various years); national sources Kilometers of rail lines Mitchell (1992, 1993, 1995); World Bank Rail Statistics Database Percentage of local telephone calls International Telecommunications Union (1996) unsuccessful Percentage of paved main roads in United Nations Economic Commission for Africa (1993) poor condition Percentage of total main roads in United Nations Economic Commission for Africa (1993); poor condition World Bank (1988) Percentage of diesel locomotives World Bank Rail Statistics Database available for use Percentage of system electricity World Bank (1997) losses REFERENCES The word "processed' describes informally reproduced works that may not be com- monly available through library systems. American Telephone and Telegram Company. Various years. The World's Telephones. New York. Antle, J. M. 1983. "Infrastructure and Aggregate Agricultural Productivity: Interna- tional Evidence." Economic Development and Cultural Change 3(April):609-19. Aschauer, D. A. 1989. "Is Public Expenditure Productive?" Journal of Monetary Eco- nomics 23(2):117-200. Barro, Robert J. 1991. "Economic Growth in a Cross-Section of Countries." Quarterly Journal of Economics 106(2):407-43. Canning, David, and Marianne Fay. 1992. "Infrastructure and Economic Growth." Department of Economics, Columbia University, New York. Processed. Canning, David, Marianne Fay, and Roberto Perotti. 1992. "Dotazioni infrastrutturali e crescita economica." Revista di Politica Economica 82(November):117-54. 546 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 .1994. "Infrastructure and Growth." In Mario Baldassarri, Luigi Paganetto, and Edmund Phelps, eds., International Differences in Growth Rates, pp. 113-47. New York: Macmillan Press. Easterly, William, and Sergio Rebelo. 1993. "Fiscal Policy and Economic Growth: An Empirical Investigation." Journal of Monetary Economics 32(3):417-58. Gramlich, E. M. 1994. "Infrastructure Investment: A Review Essay." Journal of Eco- nomic Literature 32(3):1176-96. Hulten, C. R. 1996. "Infrastructure Capital and Economic Growth: How Well You Use It May Be More Important Than How Much You Have." University of Maryland, College Park. Processed. Ingram, G. K., and Zhi Liu. 1997. "Motorization, Road Provision, and Economic Growth." World Bank, Washington, D.C. Processed. International Road Federation. Various years. World Road Statistics. Geneva. International Telecommunications Union. 1996. "World Telecommunications Indica- tors [on diskette]." Geneva. Processed. . Various years. Yearbook of Common Carrier Statistics. Geneva. Jimenez, Emmanuel. 1995. "Human and Physical Infrastructure: Investment and Pric- ing Policies in Developing Countries." In Jere Behrman and T. N. Srinivasan, eds., Handbook of Development Economics, vol. 3B, pp. 2773-843. Amsterdam: Elsevier Science, North-Holland. Lee, K. S., and Alex Anas. 1992. "Costs of Deficient Infrastructure: The Case of Nige- rian Manufacturing." Urban Studies 29(7):1071-92. Mitchell, B. R. 1992. International Historical Statistics, Europe, 1750-1988. New York: Macmillan Press. .1993. International Historical Statistics: The Americas, 1750-1988. New York: Stockton Press. .1995. International Historical Statistics: Africa, Asia, and Oceania, 1750-1988. New York: Stockton Press. Pritchett, Lant. 1996. "Mind Your P's and Q's, The Cost of Public Investment Is Not the Value of Public Capital." Policy Research Working Paper 1660. Policy Research De- partment, World Bank, Washington, D.C. Processed. Queiroz, Cesar, and S. P. Gautam. 1992. "Road Infrastructure and Economic Develop- ment: Some Diagnostic Indicators." Policy Research Working Paper 921. Policy Re- search Department, World Bank, Washington, D.C. Processed. Summers, Robert, and Alan Heston. 1991. "The Penn World Table (Mark 5): An Ex- panded Set of International Comparisons, 1950-88." Quarterly Journal of Econom- ics 106(2):327-68. United Nations. Various years. Energy Statistics. New York. . Various years. Statistical Yearbook. New York. United Nations Economic and Social Commission for Asia and the Pacific. Various years. Statistical Yearbook for Asia and the Pacific. Bangkok. United Nations Economic and Social Commission for Latin America and the Caribbean. Various years. Statistical Yearbook for Latin America and the Caribbean. Santiago, Chile. United Nations Economic and Social Commission for Western Asia. Various years. Sta- tistical Abstract of the Region of the Economic and Social Commission for Western Asia. Baghdad. Canning 547 United Nations Economic Commission for Africa. 1993. Survey of Economic and Social Conditions in Africa, 1988-89. New York. . Various years. African Statistical Yearbook. Addis Ababa. United Nations Economic Commission for Europe. Various years. Annual Bulletin of Transport Statistics for Europe. New York. World Bank. 1988. Road Deterioration in Developing Countries. Washington, D.C. . 1994. World Development Report 1994: Infrastructure for Development. New York: Oxford University Press. 1995. World Tables. Washington, D.C. 1997. World Development Indicators 1997. Washington, D.C. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 549-50 Index of Authors for Volume 12 Aw, Bee Yan, and Geeta Batra, "Technological Capability and Firm Efficiency in Taiwan (China)'' (1, January):59-79 Basu, Swati (see Deininger, Klaus) Batra, Geeta (see Aw, Bee Yan) Bayoumi, Tamim (see Masson, Paul R.) Bloom, David E., and Jeffrey G. Williamson, "Demographic Transitions and Eco- nomic Miracles in Emerging Asia" (3, September):419-55 Brunetti, Aymo, Gregory Kisunko, and Beatrice Weder, "Credibility of Rules and Economic Growth: Evidence from a Worldwide Survey of the Private Sector" (3, September):353-84 Burcroff II, Richard (see Wei, Anning) Canning, David, "A Database of World Stocks of Infrastructure, 1950-95" (3, Sep- tember):529-47 Deininger, Klaus, Lyn Squire, and Swati Basu, "Does Economic Analysis Improve the Quality of Foreign Assistance?" (3, September):385-418 Fernandez, Raquel, atnd Jonathan Portes, "Returns to Regionalism: An Analysis of Nontraditional Gains from Regional Trade Agreements" (2, May):197-220 Feyzioglu, Tarhan, V'inaya Swaroop, and Min Zhu, "A Panel Data Analysis of the Fungibility of Foreign Aid" (1, January):29-58 Guba, Waldemar (see Wei, Anning) Jolliffe, Dean, "Skills, Schooling, and Household Income in Ghana" (1, January):81- 104 Kisunko, Gregory (see Brunetti, Aymo) L6pez, Ram6n, "The Tragedy of the Commons in C6te d'Ivoire Agriculture: Em- pirical Evidence and Implications for Evaluating Trade Policies" (1, January): 105- 31 Masson, Paul R., Tamim Bayoumi, and Hossein Samiei, "International Evidence on the Determinants of Private Saving" (3, September):483-501 Morisset, Jacques, "Unfair Trade? The Increasing Gap between World and Domes- tic Prices in Commodity Markets during the Past 25 Years" (3, September):503- 26 Olarreaga, Marcelo, and Isidro Soloaga, "Endogenous Tariff Formation: The Case of Mercosur" (2, May):297-320 Portes, Jonathan (see Fernandez, Raquel) 549 550 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 Pradhan, Menno, Laura Rawlings, and Geert Ridder, "The Bolivian Social Invest- ment Fund: An Analysis of Baseline Data for Impact Evaluation" (3, Septem- ber):457-82 Puga, Diego, and Anthony J. Venables, "Trading Arrangements and Industrial De- velopment" (2, May):221-49 Rawlings, Laura (see Pradhan, Menno) Ridder, Geert (see Pradhan, Menno) Samiei, Hossein (see Masson, Paul R.) Schiff, Maurice, and L. Alan Winters, "Dynamics and Politics in Regional Integra- tion Arrangements: An Introduction" (2, May):177-95 Schiff, Maurice, and L. Alan Winters, "Regional Integration as Diplomacy" (2, May):271-95 Soloaga, Isidro (see Olarreaga, Marcelo) Squire, Lyn (see Deininger, Klaus) Swaroop, Vinaya (see Feyzioglu, Tarhan) Vamvakidis, Athanasios, "Regional Integration and Economic Growth" (2, May):251-70 Venables, Anthony J. (see Puga, Diego) von Amsberg, Joachim, "Economic Parameters of Deforestation" (1, January):133- 53 Waelbroeck, Jean, "Half a Century of Development Economics: A Review Based on the Handbook of Development Economics" (2, May).323-52 Weder, Beatrice (see Brunetti, Aymo) Wei, Anning, Waldemar Guba, and Richard Burcroff II, "Why has Poland Avoided the Price Liberalization Trap? The Case of the Hog-Pork Sector" (1, January):155- 74 Williamson, Jeffrey G. (see Bloom, David E.) Winters, L. Alan (see Schiff, Maurice) Yeats, Alexander J., "Does Mercosur's Trade Performance Raise Concerns about the Effects of Regional Trade Arrangements?" (1, January):1-28 Zhu, Min (see Feyzioglu, Tarhan) THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 551-52 Index of Titles for Volume 12 "The Bolivian Social Investment Fund: An Analysis of Baseline Data for Impact Evaluation," Menno Pradhan, Laura Rawlings, and Geert Ridder (3, Septem- ber):457-82 "Credibility of Rules and Economic Growth: Evidence from a Worldwide Survey of the Private Sector," by Aymo Brunetti, Gregory Kisunko, and Beatrice Weder (3, September):353-84 "A Database of World Stocks of Infrastructure, 1950-95," David Canning (3, Sep- tember):529-47 "Demographic Transitions and Economic Miracles in Emerging Asia," by David E. Bloom and Jeffrey G. Williamson (3, September): "Does Economic Analysis Improve the Quality of Foreign Assistance?" by Klaus Deininger, Lyn Squire, and Swati Basu (3, September):419-55 "Does Mercosur's Trade Performance Raise Concerns about the Effects of Regional Trade Arrangements?" by Alexander J. Yeats (1, January): 1-28 "Dynamics and Polit:ics in Regional Integration Arrangements: An Introduction," by Maurice Schiff and L. Alan Winters (2, May):177-95 "Economic Parameters of Deforestation," byjoachim von Amsberg (1, January):133- 53 "Endogenous Tariff Formation: The Case of Mercosur," by Marcelo Olarreaga and Isidro Soloaga (2, May):297-320 "Half a Century of Development Economics: A Review Based on the Handbook of Development Economics," by Jean Waelbroeck (2, May):323-52 "International Evidence on the Determinants of Private Saving," Paul R. Masson, Tamim Bayoumi, and Hossein Samiei (3, September):483-501 "A Panel Data Analysis of the Fungibility of Foreign Aid," by Tarhan Feyzioglu, Vinaya Swaroop, and Min Zhu (1, January):29-58 "Regional Integration and Economic Growth," by Athanasios Vamvakidis (2, May):251-70 "Regional Integration as Diplomacy," by Maurice Schiff and L. Alan Winters (2, May):271-95 "Returns to Regionalisrn: An Analysis of Nontraditional Gains from Regional Trade Agreements," by Raquel Fernandez and Jonathan Portes (2, May):197-220 "Skills, Schooling, and Household Income in Ghana," by Dean Jolliffe (1, Janu- ary):81-104 "Technological Capability and Firm Efficiency in Taiwan (China)," by Bee Yan Aw and Geeta Batra (1, January):59-79 551 552 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3 "Trading Arrangements and Industrial Development," by Diego Puga and Anthony J. Venables (2, May):221-49 "The Tragedy of the Commons in Cote d'lvoire Agriculture: Empirical Evidence and Implications for Evaluating Trade Policies," by Ram6n L6pez (1, Janu- ary):105-31 "Unfair Trade? The Increasing Gap between World and Domestic Prices in Com- modity Markets during the Past 25 Years," by Jacques Morisset (3, Septem- ber):503-26 "Why has Poland Avoided the Price Liberalization Trap? The Case of the Hog-Pork Sector," by Anning Wei, Waldemar Guba, and Richard Burcroff II (1, Janu- ary):155-74 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 3: 553 List of Referees The Editorial Board of The World Bank Economic Review thanks the following referees for their contribution to the editorial process during the past year. Richard Adams James E. Foster Diana McNaugthton Pierre-Richard Agenoir Joseph F. Francois Jean Mercenier John Akin Roman Frydman Ashoka Mody Harold Alderman Bruce L. Gardner Peter G. Moll Carol Alexander Paul Glewwe Peter Montiel Gershon Alperovich Jose de Gregorio Jonathan Murdoch Julian M. Alston Margaret Grosh Phil Musgrove Jock R. Anderson J. Luis Guasch John Nash John Antle Mona Haddad Jonathan D. Ostry Edward B. Barbier John C. Haltiwanger Howard Pack Jere R. Behrman Daniel S. Hamermesh Marco Pagano Linda Bell Eric Hanushek Roberto Perotti Jean-Marc Benabou John M. Hartwick Mark Pitt Dan Ben-David Campbell Harvey Solomon W. Polachek Eli Berman Norman Hicks Graham Pyatt William March Boal Lawrence E. Hinkle Martin Ravallion Eric W. Bond Susan Horton Charles Re Velle Jeannie D. Braithwaite Stephen Howes Philip K. Robins Philip L. Brock Charles R. Hulten David Sahn Drusilla K. Brown Guido Imbens David Sapsford Donald A. P. Bundy Jyotsna Jalan Luis Serven Craig Burnside Emmanuel Jimenez Robin C. Sickles Gerard Caprio Timothy Josling T. N. Srinivasan David Card Harry Kaiser Richard H. Steckel Eliana Cardoso Graciela Kaminsky John Strauss Anne Case Ravi Kanbur V. Sundararajan Kenneth Chomitz Tony Killick Guido Tabellini David T. Coe John B. Knight Hong Tan Carlos E. Cuevas Michael R. Kremer Linda Tesar Janet Currie Anne 0. Krueger Duncan Thomas Roy Darwin Sanjaya Lall Vinod Thomas Angus S. Deaton Peter Lanjouw Erik Thorbecke Lionel Demery Victor Lavy Thomas A. Timberg Shanta Devarajan Brian Levy Rodney Tyers Sumana Dhar David Lindauer Dominique van de Walle Steve Dowrick Michael Lipton Thierry Verdier Sebastian Edwards Jennie I. Litvack Jeffrey Vincent Robert E. Evenson Elio Londero Joachim von Amsberg Bruce Fallick Florencio Lopez-di-Silanes T.G. Weyman-Jones Robert Feenstra Annamaria Lusardi Kenneth I. Wolpin Francisco Ferreira Sergio Margulis Shlomo Yitzhaki Price V. Fishback F. Desmond McCarthy 553 The YMimr Bank is proud to announce 1ii;i- exciting publications: THE WORLD DEVELOPMENT REPORT 1998/99: KNOWLEDGE FOR DEVELOPMENT The information revolution makes understanding knowledge and develop- ment more important than ever before. This year's World Development Report, the twenty-first in this annual series, looks at the role of knowledge in advancing economic and social well-being. Knowledge is the defining issue for today's developing countries. The Report will help increase understanding of the complex relationship between knowledge and devel- opment thus helping to better apply the power of knowledge to the great challenge of eradicating poverty and improving people's lives. October 1998. 264 pages. Paperback: Stock no. 61118. ISBN 0-19-521118-9. $25.95 October 1998. 264 pages. Hardcover: Stock no. 61119. ISBN 0-19-521119-7. $49.95 GLOBAL ECONOMIC PROSPECTS 1998/99 What is ahead for developing countries after the East Asian crisis? How can 0 j future crises be prevented? Global Economic Prospects 1998/99 provides answers to these questions and much more. The results of this crisis and the short-term and long-term prospects are explored for their effect on the future of the global econtmy and private and public institutions. December 1998. 150 pagles. Stock no. 14123. ISBN 0-8213-4123-5. $24.95 EAST ASIA: THE ROAD ._l T TO RECOVERY Did the "miracle" in East Asia ever really occur? Last year, the world wit- _ nessed catastrophic events in East Asia-markets crashing in a matter of days and businesses going bankrupt in just hours. After three decades of _ remarkable expansion, the East Asian economy went into a tailspin. Why has East Asia faltered? East Asia: The Road to Recovery asks this very important question. This comprehensive World Bank study is designed to be a snap shot of where the region stands, a progress report on the enor- mous changes that have taken place in the last year, and an analysis of the remaining obstacles to establishing a firm recovery. September 1998. 170 pages. Stock no. 14299. ISBN 0-8213-4299-1. $25.00 z World Bank Publications US customers, contact The World Bank, P.O. Box 960, Herndon, VA 20172-0960, Phone: (703) 661-1580, Fax: (703) 661-1501. Shipping and handling: US$5.00. Payment US$ check drawn on a US bank payable to the World Bank or by VISA, MasterCard, or American Express. Customers outside ot the US may also contact their local World Bank distributor. WBRV Visit our !II1 at www.worldbank.org and click on Publications. DISTRIBUTORS OF Chinese Corporation for Promotion of FINLAND IRAN WORLD BANK Humanities Akateeminen Kirakauppa Ketab Sara Co. Publishers 52 You Fang Hu Tong, PO. Box 128 Khaled Eslamboli Ave., 6th Street PUBLICATIONS Xuan Nei Da Jie FIN-00101 Helsinki Delatrooz Alley No. 8 Prices and credit terms varyfrom Beijing Tel: (358 0) 121 4418 PO. Box 15745-733 country to country. Consult your Tel: (8610) 660 72 494 Fax: (358 0)121-4435 Tehran 15117 local distributor before placing an Fax: (86 10) 660 72 494 E-mail: akatilaus@stockmann.ti Tel: (98 21) 8717819; 8716104 order. URL: httpl/www .akateeminen.com Fax: (98 21) 8712479 COLOMBIA E-mail: ketab-sara@neda.net.ir ARGENTINA Intoenlace Ltda. FRANCE Oficina del Libro Intemacional Carrera 6 No. 51-21 Editions Eska; DBJ Kowkab Publishers Av. Cordoba 1877 Apartado Aereo 34270 48, rue Gay Lussac PO. Box 19575-511 1120 Buenos Aires Santafe de Bogota, D.C. 75005 Pads Tehran Tel: (541) 815-8354 Tel:(571)285-2798 Tel: (33-1) 55-42-7308 Tel: (98 21) 258-3723 Fax: (54 1) 815-8156 Fax: (57 1) 285-2798 Fax: (33-1) 43-29-91-67 Fax: (98 21) 258-3723 E-mail: olilibro@satlink.com COTE D'IVOIRE GERMANY IRELAND AUSTRALIA, FIJI, PAPUA NEW Center d'Edition et de Diffusion UNO-Verlag Govemment Supplies Agency GUINEA, SOLOMON ISLANDS, Afncaines (CEDA) PoppelsdorferAllee 55 Oifig an tSolathair VANUATU, AND SAMOA 04 B.R 541 53115 Bonn 4-5 Harcourt Road D.A. Information Services Abidjan 04 Tel: (49 228) 949020 Dublin 2 648 Whitehorse Road Tel: (225) 246510; 246511 Fax: (49228) 217492 Tel: (3531) 661-3111 Mitcham 3132 Fax: (225) 25 0567 URL: hnp://www.uno-verdag.de Fax: (3531) 475-2670 Victoria E-mail: unovedagOaol.com Tel: (61) 3 9210 7777 CYPRUS ISRAEL Fax: (61) 3 9210 7788 Center for Applied Research GHANA Yozmot Literature Ltd. E-mail: service @dadirect.com.au Cyprus College Epp Books Services PO. Box 56055 URL: httpi/www .dadirect.com.au 6, Diogenes Street, Engomi P.O. Box 44 3 Yohanan Hasandlar Street RO. Box 2006 TUC Tel Aviv 61560 AUSTRIA Nicosia Acora Tel: (972 3) 5285-397 Gerold and Go. TTel: (357 2) 59-0730 Tel: 223 21 778843 Fax: (972 3) 5285-397 Weihburggasse 26 Fax: (357 2) 66-2051 Fax: 223 21 779099 A-1011 Wien R.O.Y Intemational Tel: (43 1) 512-47-31-0 CZECH REPUBLIC GREECE PO Box 13056 Fax: (431) 512-47-31-29 USIS, NIS Prodejna Papasotinou S.A. Tel Aviv 61130 URL: http://mww.gerold.co/at.online Havelkova 22 35, Stournara Str. Tel: (972 3) 649 9469 130 00 Prague 3 106 82 Athens Fax: (972 3) 648 6039 BANGLADESH Tel: (420 2) 24231486 Tel: (301) 364-1826 E-mail: royil@netvision.net.il Micro Industries Development Fax (4202) 24231114 Fax: (301) 364-8254 Assistance Society (MIDAS) URL: httpl/www .nis.cz/ Palestinian Authorty/Middle East House 5, Road 16 HAITI Index Information Services Dhanmondi R/Area DENMARK Culture Diffusion ROB. 19502 Jerusalem Dhaka 1209 SamfundsLitteratur 5, Rue Capois Tel: (972 2) 6271219 Tel: (880 2) 326427 Rosenoems A116 11 C.P 257 Fax: (972 2) 6271634 Fax: (880 2) 811188 DK-1970 Frederiksberg C Port-au-Pnnce Tel: (45 35) 351942 Tel: (509) 23 9260 ITALY, LIBERIA BELGIUM Fax: (45 35) 357822 Fax: (509) 23 4858 Licosa Commissionaria Sansoni SPA Jean De Lannoy URL: httpJlwww.sl.cbs.dk Via Duca Di Calabda, 1/1 Av. du Pol 202 HONG KONG, CHINA; MACAO Casella Postale 552 1060 Brussels ECUADOR Asia 2000 Ltd. 50125 Firenze Tel: (322) 538-5169 Libri Mundi Saes & Circulation Department Tel: (39 55) 645-415 Fax: (32 2) 538-0841 Libreria Intemacional 302 Seabird House Fax: (39 55) 641-257 PO. Box 17-01-3029 22-28 Wyndham Street, Central E-mail: licosa@hbcc.it BRAZIL Juan Leon Mera 851 Hong Kong, China URL: http//www .ftbcc.iVlicosa Publicacbes Tecnicas Intemacionais Quito Tel: (852) 2530-1409 Ltda. Tel: (593 2) 521-606; (593 2) 544-185 Fax: (852) 2526-1107 JAMAICA Rua Peixoto Gomide, 209 Fax: (593 2) 504-209 E-mail: sales@asia2000.com.hk Ian Randle Publishers Ltd. 01409 Sao Paulo, SR E-mail: libnmul @librimundi.com.ec URL: httpi/www .asia2000.com.hk 206 Old Hope Road, Kingston 6 Tel: (55 11) 259-6644 E-mail: libnmu2@librimundi.com.ec Tel: 876-927-2085 Fax: (55 11) 258-6990 HUNGARY Fax: 876-977-0243 E-mail: postmaster@pti.uol.br CODEU Euro Into Service E-mail: irpl@colis.com URL: httpi/ww .uol.br Ruiz de Castilla 763, Edit. Expocolor Margitszgeti Europa Haz Prmer piso, Of. #2 H-1 138 Budapest JAPAN CANADA Quito Tel: (36 1) 350 80 24, 350 80 25 Eastern Book Service Renouf Publishing Co. Ltd. TeVFax: (593 2) 507-383; 253-091 Fax: (36 1) 350 90 32 3-13 Hongo 3-chome, Bunkyo-ku 5369 Canotek Road E-mail: codeu@impsat.net.ec E-mail: euroinfo@mail.matav.hu Tokyo 113 Ottawa, Ontaro K1J 9J3 Tel: (81 3) 3818-0861 Tel: (613) 745-2665 EGYPT, ARAB REPUBLIC OF INDIA Fax: (81 3) 3818-0864 Fax: (613) 745-7660 Al Ahram Distribution Agency Allied Publishers Ltd. E-mail: orders@svt-ebs.co.jp E-mail: order .dept@ renoufbooks.com Al Galaa Street 751 Mount Road URL: http://www .bekkoame.orjp/-svt- URL: http://www.renoutbooks.com Cairo Madras -600 002 ebs Tel: (20 2) 578-6083 Tel: (91 44) 852-3938 CHINA Fax: (20 2) 578-6833 Fax: (91 44) 852-0649 KENYA China Financial & Economic Publishing Africa Book Service (E.A.) Ltd. House The Middle East Observer INDONESIA Quaran House, Mfangano Street 8, Da Fo Si Dong Jie 41, Sherit Street Pt. Indira Limited PO. Box 45245 Beijing Cairo Jalan Borobudur 20 Nairobi Tel: (86 10) 6401-7365 Tel:(202)393-9732 PO. Box 181 Tel: (2542)223641 Fax: (86 10) 6401-7365 Fax: (20 2) 393-9732 Jakarta 10320 Fax: (254 2) 330 272 Tel: (62 2t) 390-4290 China Book Import Centre Fax: (62 21) 390-4289 PO. Box 2825 Beijing Legacy Books Auckland ROMANIA SWITZERLAND Loita House lbl: (64 9) 524-8119 Compani De Librani Bucuresti S.A. Libraire Payot Service Institutionnel Mezzanine 1 Fax: (64 9) 524-8067 Str. Lipscani no. 26, sector 3 C6tes-de-Montbenon 30 P.O. Box 68077 Bucharest 1002 Lausanne Nairobi Oasis Official Tel: (40 1) 613 9645 Tel: (41 21) 341-3229 Tel: (254) 2-330853, 221426 FO. Box 3627 Fax: (401) 312 4000 Fax: (41 21) 341-3235 Fax: (254) 2-330854, 561654 Wellington E-mail: Legacy@form-net.com 'el: (644)4991551 RUSSIAN FEDERATION ADECO Van Diermen Fax: (64 4) 499 1972 lsdatelstvo eVes Mir> EditionsTechniques KOREA, REPUBUC OF E-mail: oasis@actdx.gen.nz 9a, Kolpachniy Pereulok Ch. de Lacuez 41 Dayang Books Trading Co. lJRL: http://ww.oasisbooks.co.nz/ Moscow 101831 CHI 807 Blonay Intematfonal Division Tel: (7 095) 917 87 49 Tel: (41 21) 943 2673 783-20, Pangba Bon-Dong, Socho-ku NIGERIA Fax: (7095)9179259 Fax: (41 21) 943 3605 Seoul tJniversity Press Limited Tel: (82 2) 536-9555 T'hree Crowns Building Jericho SINGAPORE; TAIWAN, CHINA THAILAND Fax (82 2) 536-0025 F'rivate Mail Bag 5095 MYANMAR; BRUNEI Central Books Distribution E-mail: seamap@chollian.net Ibadan Hemisphere Publication Services 306 Silom Road 'el: (234 22) 41-1356 41 Kallang Pudding Road #04-03 Bangkok 10500 Eulyoo Publishing Co., Ltd. F:ax: (234 22) 41-2056 Golden Wheel Building Tel: (66 2) 2336930-9 46-1, Susong-Dong Singapore 349316 Fax: (66 2) 237-8321 Jongro-Gu NORWAY Tel: (65) 741-5166 Seoul S)WETS Norge AS Fax: (65) 742-9356 TRINIDAD & TOBAGO Tel: (82 2) 734-3515 Book Department, Postboks 6512 E-mail: ashgate@asianconnect.com AND THE CARRIBBEAN Fax: (82 2) 732-9154 Etterstad Systemalics Studies Ltd. N-0606 Oslo SLOVENIA St. Augustine Shopping Center LEBANON 1Tel: (47 22) 97-4500 Gospodarski vestnik Publishing Group Eastem Main Road, St. Augustine Librairie du Liban f-ax (47 22) 97-4545 Dunajska cesta 5 Trinidad & Tobago, West Indies PO. Box 11-9232 1000 Ljubljana Tel: (868) 645-8466 Beirut l'AKISTAN Tel: (386 61) 133 83 47; 132 12 30 Fax: (868) 645-8467 Tel: (961 9) 217 944 Mirza Book Agency Fax: (386 61) 133 80 30 E-mail: tobe@trinidad.net Fax: (961 9) 217 434 ri5, Shahrah-e-Quaid-e-Azam E-mail: repansekj@gvestnik.si E-mail: hsayegh@librairie-du- Lahore 54000 UGANDA liban.com.lb 'rel: (92 42) 735 3601 SOUTH AFRICA, BOTSWANA Gustro Ltd. URL: http://www.libraine-du-liban.com.lb Fax: (92 42) 576 3714 For single titles: PO Box 9997, Madhvani Building Oxford University Press Southern Aftca Plot 16/4 Jinja Rd. MALAYSIA Oxford University Press Vasco Boulevard, Goodwood Kampala University of Malaya Cooperative ti Bangalore Town PO. Box 12119, Nt City 7463 Tel: (256 41) 251 467 Bookshop, Limited Sharae Faisal Cape Town Fax: (256 41) 251 468 PO. Box 1127 '0 Box 13033 Tel: (2721) 5954400 E-mail: gus@swiftuganda.com Jalan Pantai Baru Karachi-75350 Fax: (27 21) 595 4430 59700 Kuala Lumpur 'rel: (92 21) 446307 E-mail: oxford@oup.co.za UNITED KINGDOM Tel: (60 3) 756-5000 Fax: (92 21) 4547640 Microinfo Ltd. Fax: (60 3) 755-4424 E-mail: ouppak@TheOffice.net For subscnption orders: PO. Box 3, Omega Park, Alton, E-mail: umkoop@tm.net.my International Subscription Service Hampshire GU34 2PG lPak Book Corporation PO. Box 41095 England MEXICO Aziz Chambers 21, Queen's Road Craighall Tel: (44 1420) 86848 INFOTEC Lahore Johannesburg 2024 Fax: (441420) 89889 Av. San Femando No. 37 'rel: (92 42) 636 3222; 636 0885 Tel: (2711) 880-1448 E-mail: wbank@microinfo.co.uk Col. TorieJo Guerra Fax: (92 42) 636 2328 Fax: (27 11) 880-6248 URL: http://www.microinfo.co.uk 14050 Mexico, D.F E-mail: pbc@brain.net.pk E-mail: iss@is.co.za Tel: (52 5) 624-2800 The Stationery Office Fax: (52 5) 624-2822 IPERU SPAIN 51 Nine Elms Lane E-mail: intotec@rtn.net.mx Editorial Desarrollo SA Mundi-Prensa Libros, S.A. London SW8 5DR URL: http:/rIn.net.mx Apartado 3824, Ica 242 OF 106 Castello 37 Tel: (44 171) 873-8400 1-ima 1 28001 Madrid Fax: (44171) 873-8242 Mundi-Prensa Mexico S.A. de C.V. Tel: (51 14) 285380 Tel: (34 91) 4 363700 URL: http:l/www.theso.co.uk/ c/Rio Panuco, 141 -Colonia Cuauhtemoc Fax: (51 14) 286628 Fax: (34 91) 5 753998 06500 Mexico, D.F E-mail: libreria@mundiprensa.es VENEZUELA Tel: (52 5) 533-5658 PHILIPPINES URL: http:/Iwww .mundiprensa.com/ Tecni-Ciencia Libros, S.A. Fax: (52 5) 514-6799 Intemational Booksource Center Inc. Centro Cuidad Comercial Tamanco 1127-A Antipolo St, Barangay, Mundi-Prensa Barcelona Nivel C2, Caracas NEPAL Venezuela Consell de Cent, 391 Tel: (58 2) 959 5547; 5035; 0016 Everest Media Intematonal Services Makafi City 08009 Barcelona Fax: (58 2) 959 5636 (P) Ltd. Tel: (63 2) 896 6501; 6505; 6507 Tel: (34 3) 488-3492 GPO Box 5443 Fax: (63 2) 8961741 Fax: (34 3) 487-7659 ZAMBIA Kathmandu E-mail: barcelona@mundiprensa.es University Bookshop, University of Tel: (9771) 416 026 POLAND Zambia Faxr (9771) 224 431 Intemational Publishing Service SRI LANKA, THE MALDIVES Great East Road Campus Ul. Piekna 31/37 Lake House Bookshop PO. Box 32379 NETHERLANDS 00-677 Warzawa 100, Sir Chittampalam Gardiner Lusaka De LUndeboom/lntemationale Tel: (48 2) 628-6089 Mawatha Tel: (2601) 252 576 Publicaties b.v.- Fax: (48 2) 621-7255 Colombo 2 Fax: (2601) 253 952 P.O. Box 202, 7480 AE Haaksbergen E-mail: books%ips@ikp.atm.com.pl Tel: (941) 32105 Tel: (31 53) 574-0004 URL: http:/www .ipscg.waw.pllips/ Fax: (94 1) 432104 ZIMBABWE Fax: (31 53) 572-9296 exporU E-mail: LHL@sn.lanka.net Academic and Baobab Books (Pvt.) Ltd. E-mail: lindeboo¢woridonline.nl 4 Conald Road, Graniteside URL: httpJ/www.woridonline.nl/ PORTUGAL SWEDEN PO. Box 567 -lindeboo Livraria Portugal Wennergren-Williams AB Harare Apartado 2681, Rua Do Carmo 70-74 P 0. Box 1305 Tel: 263 4 755035 NEWZEALAND 1200Lisbon S-171 25Solna Fax:2634781913 EBSCO NZ Ltd. Tel: (1) 347-4982 Tel: (46 8) 705-97-50 Private Mail Bag 99914 Fax: (1) 347-0264 Fax: (46 8) 27-00-71 New Market E-mail: mail@wwi.se Coming in the next issue of THIEV WORLD BANK ECONOMIC REVIEW January 1999 Volume 13, Number 1 A symposium on public sector downsizing, including ... * Public Sector Downsizing: An Introduction by Martin Rama * Matching Severance Payments with Worker Losses in the Egyptian Public Sector by Ragui Assaad * Cross-Country Evidence on Public Sector Retrenchment by John Halt-iwanger and Manisha Singh * The Efficient Mechanism for Downsizing the Public Sector by Doh-Shin jeon and Jean-Jacques Laffont * Earnings and Welfare after Downsizing: Central Bank Employees in Ecuador by Martin Ramna and Donna MacIsaac * Labor Earnings in One-Company Towns: Theory and Evidence from Kazakhstan by Martin Ranma and Kinnon Scott * The Algerian Retrenchment System: A Financial and Economic Evaluation by Elizabeth Ruppert