ThE WORLD BANK ECONOMIC REVIEW Volume 9 May 1995 Number 2 Does Participation Improve Performance? Establishing Causality with Subjective Data Jonathan Isham, Deepa Narayan, and Lant Pritchett Winners and Losers in Transition: Returns to Education, Experience, and Gender in Slovenia Peter F. Orazem and Milan Vodopivec An Eclectic Approach to Estimating the Determinants of Achievement in Jamaican Primary Education Paul Glewvwe, Margaret Grosh, Hanan Jacoby, and Marlaine Lockheed Natural Resource Management and Economywide Policies in Costa Rica: A Computable General Equilibrium (CGE) Modeling Approach Annika Persson and Mohan Munasinghe The Role of Infrastructure in Mexican Economic Reform Andrew Feltenstein and Jiming Ha The Current Account in Developing Countries: A Perspective from the Consumption-Smoothing Approach Atish R. Ghosh and Jonathan D. Ostry Comment on "Measuring the Independence of Central Banks and Its Effect on Policy Outcomes" by Cukierman, Webb, and Neyapti M. K. Anyadike-Danes THE WORLD BANK ECONOMIC REVIEW EDITOR Moshe Syrquin CONSULTING EDITOR Sandra Gain EDITORIAL BOARD Kaushik Basu, Cornell University and University of Delhi Sebastian Edwards Guillermo Calvo, University of Maryland Gregory K. Ingram Jonathan Eaton, Boston University John Page Alberto Giovannini, Columbia University Lant H. Pritchett Mark R. Rosenzweig, University of Pennsylvania Jacques van der Gaag The World Bank Economic Review is a professional joumal 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 marhemati- cal 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; 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 nor more than two pages in length to the Editor. The views and interpretations in articles published are those of the authors and do not necessarily repre- sent 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 boundaries, denomi- nations, 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 $10.95. Subscription rates are as follows: Individuals Institutions i-year subscription US$25 US$45 2-year subscription US$46 US$86 3-year subscription US$65 US$125 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 countries and in socialist economies in transition. Written request is required every two years to renew such subscriptions. 3 1995 The International Bank for Reconstruction and Development / THE WORLD BANK All rights reserved Manufactured in the United States of America ISSN 0258-6770 Material in this journal is copyrighted. The World Bank encourages dissemination of its work and will normally give permission promptly and, when the intended reproduction is for noncommercial purposes, without asking a fee. Permission to make photocopies is granted through the Copyright Clearance Center, 27 Congress Street, Salem, MA 01970 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 Sciences Citation Index(5. It is available in microform through University Microfilms, Inc., 300 North Zeeb Road, Ann Arbor, Michigan 48106, U.S.A. THE WORLD BANK ECONOMIC REVIEW Volume 9 May 1995 Number 2 Does Participation Improve Performance? Establishing 175 Causality with Subjective Data Jonathani Isham, Deepa Narayan, and Lant Pritchett Winners and Losers in Transition: Returns to Education, 201 Experience, and Gender in Slovenia Peter F. Orazein and Milan Vodopiwec An Eclectic Approach to Estimating the Determinants of 231 Achievement in Jamaican Primary Education Paul Gleunve, Margaret Grosh, Hananjacoby, ana' Marlaine Lockheed Natural Resource Management and Economywide Policies in 259 Costa Rica: A Computable General Equilibrium (CGE) Modeling Approach Annika Persson and Mohan Munasinghe The Role of Infrastructure in Mexican Economic Reform 287 Andreit Feltenstein andJiming Ha The Current Account in Developing Countries: A Perspective 305 from the Consumption-Smoothing Approach Atish R. Ghosh and.Jonathan D. Ostrv Comment on "Measuring the Independence of Central Banks 335 and Its Effect on Policy Outcomes" by Cukierman, Webb, and Neva pti M. K. Anvadikc-Danes F H I w () R I 1) B \ N K I ( O) N () NtlI( I( R 5 I I W al. V D I 9, N i 2 1 ' 5 11 Does Participation Improve Performance? Establishing Causality with Subjective Data Jonathan Ishamn, Deepa Naravan, and Lant Pritchett Data froni 121 dirersc rural water prolects provide stronig statistical findtings that in- creasing beneticiary participation directly causes better project ouitconmes. Three possi- ele econometric olbjections to these findings are addressed and answered. The subjective naturc of thbe data does not precluidle valid, cardinal measures of participation appropri- ate for statistical anralsysis. "Halo effects--changes in thet measurement of one varrable because of t/le observed State of canother viariable-lo not seem to inuduce a strong upward bais in thve mcasuremnent of participation or project performance. Reverse calu- satiotn is unizlikely: estimation ufsing instrunental 'ariablIs, data on project timinig, and documentation of case studies support the, cause-elfect relition betiveen participation and better project perf'orManIce. An increasing number of development practitioners agree that participation of the intended beneficiaries improves project performance. Participatory develop- ment, championed since the early 1970s by mostly noneconomic social scientists and grassroots organizations (Freire 1973); Korten 1980), is increasingly advo- cated by the largest and most influential aid agencies (UNDP 1993; World Bank 1991). The existence of consensus (or advocacy), however, does not imply the existence of clear and coInvinlcinig evidence that participation improves perfor- mance. Advocates of participation have most often relied on case studies to document the link between participation and performance (Briscoe and de Fer- ranti 1988; Korten and Siy 1988). These case studies, however, are easily dis- missed by skeptics as inconiclusive because the small number of cases and the informal method do not allow formal testing of the findings. In response, some studies have used the systematic case study method to establish statistically the relationship between participation and project performance (Esman and Uphoff 1984; Finsterbusch and Van Wicklin III 1987). lonathan Ishatim is witih the De)partment of Econoritics at the University of Mlaryland and the center of InstitUtiOintal Reform anid the Inforimial Sector (iRis); Dcepa Naravan is with the Environment D)epartment at the World Bank: arid Lint Pritchett is with the Policy. and Research Department at the World Bank. This article Wdas a background papel for Wkorld Developmenet Report 9)')4. It is based Upon a larger research project initiated arnd fulnded liv the Participatory [)evelopmient L.earning Group, the UNDP-World Bank Water aiid Sanitation Program, and the Social Polics anLd Resettlcicent Division ot the NVorld Bank. The aLutilOrs wiould like to tha.nk Jo hi Blaxall, Aitoiiio EStadLh., Peter I airo1LIW, and Jimon Tybout for helpful S ugge tir iis. Q' 199i The Inrrernatirornal Bank for RecoinstrUCtiil and D)vevlopment/ I THE WORLL) BANK 17/ 176 I11F \W.')RI 1) BANK 1 >N()M NM RIVIFW, Vol 4. Nol 2 Skeptics have raised three objections to the statistical evidence of a causal impact of participation on project performanice. First, participation and project ratings are not objective: subjective data are not appropriate for statistical analysis. Second, the subjectivity in the ranking of projects will lead to halo effects: if investigators believe participation is good, their subjective rankings will overstate the level of participation in highly successful projects and the level of success of highly participatory projects. Third, better project performance may cause increased beneficiary participation rathier than vice versa: a mere statistical association is not evidence of a causal impact of participation on performance. In this article we address and overcome each of the above objections. We start with a brief review of the construction of the data on project performance and beneficiarv participation for 121 rural water supply projects and a presentation of the basic statistical results. In answer to the first objection, we show that the subjective nature of the data does not preclude intersubjectively valid, cardinal measures of participation ap- propriate for statistical analysis. There is no necessary coniection between ob- jectivitv and quanitification withi cardinal numbers. Cardinal measurement using subjective criteria is common. Moreover, the cardinal rankinigs for each project were created by two different coders, and the degree of intercoder agreenment is very high. This agreement indicates that intersubjective reliability can be achieved even for intrinisically subjective concepts such as participation. Finally, we show that assuming cardinalitv or imposing linearity is not necessary to establish the basic result. We address the halo effect in coding performance and participation data from project documents in two wavs. First, we show that the results are the same if the first coder's performance indicators are regressed on the second coder's partici- pation scores (and vice versa); this test indicates that the coders' subjective ranking does not lead to a halo effect. The primary danger of subjective mea- surement, however, and one that we cannot address, is having the same individ- ual assess both project success and participation. If that person has strong views about the relationship, these views ma' iniduce a bias in the project documents themselves, because the performanice of participatory projects (and the partici- pation in successful projects) may be exaggerated. We show, however, that the strength of the performance-participation relationship does not depend on the objectiveness of the success indicator. The third and most difficult objection to answer is that the existence of an association does not imply causation. Although causality is nearly impossible to establish, we present three arguments in its favor. First, estimation using instru- mental variables allows the identification of the impact of exogenous changes in participation, and eliminates the effect on the participation estimates of reverse causation or simultaneity. Second, data on the timing of participation show that participation at early stages improves project performance at every stage, from implementation to maintenance. Third, case studies slhow that exogenous Isbayn, Naravan. and Prntchett 177 changes in participation in ongoing projects have strong impacts on performance. Even for this limited set of projects, this article is not intended to be compre- hensive. We focus on the econometric issues involved in drawing causal infer- ences from subjective project data. Narayan (1995), drawing on these data (and much more, including many case studies), discusses in greater detail the relation- ship between performanice and participation. the determinants of project success in addition to participation, the underlying determinants of participation, and the mechanisms wherebv participation increases overall effectiveness. 1. THE BASICS The systematic case review method (Finsterbusch 1990) is used to transform qualitative evaluations in a set of case studies into data suitable for statistical analysis. This method consists of the following basic steps: A conceptual frame- work for a system of related phenomena, usually in the form of a flow chart, is established, specifying the causal relationships between different subsets of the system; a questioiiaire is developed to delineate relevant elements of each of these subsets and to measure their magnitude on the basis of prior knowledge of the system and on the hypotheses to be tested; two independent coders then test this preliminarv questionnaire in a small number of case studies by transforming the relevant qualitative information for each question into a cardinal scale; finally, after refining the questionnaire-to match available information and to eliminate redundanicy-the coders independently review each of the cases and produce two sets of data that can be evaluated using established statistical techniques. Data onl Rural WY/ater Projects The data from the 121 rural water supply projects in this study were assembled from project documents, using the systematic case review method (for more on this and similar methodology, see Finsterbusch and Van Wicklin III 1987 and E.sman and Uphoff 1984). Ex post project assessments by development agencies typically combine limited quantitative evaluations with subjective judgments of project performance. Each project document was read by two independent readers who coded specific project variables (for example, overall success of project) onto a scale from 1 to 7, creating 144 distinct variables. Appendix table A-i shows the list of those coded variables that were used in the analysis along with the basic summary statistics. The variables can be di- vided into four groups: * Project performance indicators (for example, overall project effectiveness and the percentage of the water svstem in good condition) * Measures of participation (for example, overall beneficiary participation and participation in construction) 178 IHF WORI 1) BANK I CONOMI( R .IIW. Wt, 9, N\). I * Background characteristics or project characteristics that determine project performance (for example, the size of the project and the availability of spare parts) * Background characteristics or project characteristics that determine partici- pation (for example, the extent to which the agency made participation a goal, and consensus among users about project objectives). The projects, drawni from forty-nine countries in Africa, Asia, and Latin America, were financed by a range of external donors, including the U.S. Agency for International Development (USAID), the United Nations Development Pro- gramme (UN D P), the World Bank, and nongovernmental organizations (NGOS). The range of project costs was from $0.5 million to $250 million, and the water technologies ranged from spring captures to power-pumped piped water svs- tems. The primary objective as well as the common thread of all the projects was water supply, although many of the projects included other components, and the projects were assessed on1 their overall performance. The rural water sector was chosen primarily because community participation had became a central feature of the strategy for reaching scattered rural commu- nities with safe drinking water during the International Drinking Water Supply and Sanitation Decade of the 1980s. Given the local common property nature of the good, the difficulty in achieving economies of scale in rural areas, and the need to develop decentralized maintenance systems, local responsibility-gen- erated through participation-is thought to be particularly relevant to the sector. The participation variable merits some discussion. The measure of participa- tion was not simply a measure of whether potential beneficiaries were surveyed about tlheir preferences. Participation was scored on a continuum, progressing from information sharing, to more in-depth consultation, to shared decision- making, to control over decisionmaking. The participation of beneficiaries was considered at three different stages of the project cycle: project design, construc- tion, and operation and maintenance. THE BASIC MODEL AND RESULTS We specify the most general indicators-overall project effectiveness (OPE) and overall beneficiary participation (OBPi)-and estimate a simple linear rela- tionship between them. The bivariate relationship between OPE and OBP In the itb project is simply (1) OPE, =13*OBP, +e,1, which is functionally equivalent to the simple correlation of the two variables. I 1. In the hivariate linear regression, the regression coetficient is Q/CT2 ( is the covariance oh and x), where the correlation co-fficienit p is / The coirrelationi coefficienit p = i(al/) can he derived from Ishbam, Naravan. and Pritchett 179 The usefulness of this bivariate relationship is limited because other nonpar- ticipation determinants of project performance are excluded. In expanding the model, it proves useful to divide nonparticipation determinants into two groups: those that are fully exogenous and not affected by participation (for example, availability of spare parts), denoted by the matrix Z, and those that are poten- tiallv affected by participation (for example, responsiveness of managers), deno- ted by the matrix W. The multivariate equations are then (2) OPE, = * OBP, ±. * Z +3 , W, + ,, (3) W,=y*OBP, + *zZ,+Y *X,+ 1, where X is the matrix of nonparticipation determinants of the Ws. Both Zs and Ws are potential determinants of project performance; Zs, however, represent variables that are not influenced by participation, and Ws represent variables that may be determined (in part) by participation. As indicated in the second equation, Ws may also be determined in part by the Zs and by some other set of variables, Xs. In summary, these multivariate equations state that the perfor- mance of a water project depends on beneficiary participation, a set of inputs (Z) not related to participation, and a set of performance determinants (W) that may in turn be determined by participation and other inputs. The distinction between the Zs and the Ws is important for maintaining the distinction between the partial and the total impact of participation onl project performance. In the multivariate regression (equation 2), the ,B coefficient gives the direct impact of increasing participation, holding all included variables constant: (4) dOPE dOBP But participation may also influence performance indirectlv. The total impact of changing participation is the sum of the direct and indirect impacts: (5) dOPE dOPE dOPE dW dOBP dOBP 17Z'W=/ dW dOBP or, in this particular specification, (6) dBOPE Thus, the simple partial coefficient with all controls understates the total impact of participation, and the bivariate coefficient (which excludes the Zs and Ws) overstates the impact to the extent to which these determinants are positively correlated with participation. a simple rescalinig of either ot the possible hivariate linear regre,soion coefficienits (for example, v on x or x on v). 18() [HF. WO(RI ])SANK I:(:ONOMI(. RFVIFW. VOl.. 9. No. 2 Results Table I presents three estimates of the linear association of OBP with OPE: bivariate, limited multivariate (with Zs), and full muIltivariate (with Zs and Ws). For each regression we report the linear regression coefficient on OBP in the OPE regression. In all cases the results are strongly statistically significant and empiri- cally quite large. Thle estimiated impact of participation on project effectiveness ranges from 0.62 for the bivariate case to 0.24 in the full multivariate case. The multivariate impact is naturally lower than the bivariate effect because of the exclusion of positively correlated nonparticipation performance determinants, as discussed above. Although the sample sizes differ, sample size is not the determinative factor, as shown by the estimates of the hivariate impact in the smaller samples used for the multivariate regressions: (t-statistics in parentheses) 0.599 (7.31) for sample size 77, and 0.557 (6.09) for saniple size 68. The bivariate effect is an upward-biased estimate of the total effect. The full multivariate estimate of the partial effect is likely biased downward for the total effect. Therefore, the bivari- ate and full multivariate results create reasonable bounds for the total effect. How shoultd the coefficients be interpreted? The expected impact of increasing participation from a low level (OBP = 2) to a high level (OBP = 6) is to improve project performinance I to 2.5 points (oil a scale of I to 7). A one-standard- deviation increase in participation (in appendix table A-1, the standard deviation is 1.7) is associated with an increase in performance of between 0.41 points (for the full multivariate model) and 1.05 points (for the bivariate model). The interesting-and intuitively appealing-results of all the regressions (limited and full multivariate) are reported in appendix table A-2 and are discussed in Naravan (199.S). Table 1. Basic Estimncation Restilts from Regressing Overall Project Effectiveness on Overall Beneficiary Participation Lioited ,ntiltimariate izultivariate Item Biiiurinate on olel 117 Ide! mIo uel', Coefficienit 0)62 (0.28 0.24 (10.6! (S 3) (3.8) 1 1.4- <5. 3 <3.1 > Samplesize 121 77 68 R2 0.49 0.86 (.89 Note: Valies are from ordinary least squares (OLS) regressions in which the dependent variable is overall project effectiveness. The regressions include constanits for which results are not reported. t-statis- tics are in parentheses. Wlhite heteroskedasticitv-consistent t-statistics are in angle brackets. a The limited miultivariate model includes noniparticipationi determinants of project performance that are fully exogenous. the Z variables. Regression results for the Z variables are reported in table A-2. lb The full multivariate model incLides the W variables, nonparticipation determinanits of project peitolilialice that are potentially affected by participationi, and the Z variables. Regression results for the Z and W variables are relported in table A--' Source Authors calculations. Ishcam, Naravan, and Pritcbett 181 We note here that the choice of the Zs and Ws in this study was neither entirely straightforward nor driven by rules. However, all the results have proved robust to a number of variations of the model, and we feel that the choice of control variables is not of primary interest. Having 144 coded variables with only 121 projects meant that mechanical procedures for selecting variables would lack degrees of freedom and would not likely be of much help. Moreover, many of the variables were clearly overlapping and likely to be collinear. After some experimentation, we based variable inclusion on three criteria: decent intercoder reliability, prior judgments about the best choice among collinear sets of variables, and impact on the estimate of participation (we never dropped any variable that seriously affected the estimate of the participation coefficient). In none of the experiments were the results on participation substanltially different from the full multivariate case reported in table 1. We were more than generous in our inclusion of potential performance fac- tors, including a total of eighteen nonparticipation variables. The participation variable thus easily passes this kitchen-sink test of throwing all plausible vari- ables into the regression. The danger of inadequate controls for other determi- nants of project performance is not nearly as serious as the three problems we discuss in the following sections. Heteroskedasticity, a typical econometric problem that receives a fair bit of attention, is not a problem with the present results for two reasons. First, the White heteroskedasticity-consistent standard errors are roughly the same as those obtained using ordinary least squares (OLS). For the bivariate case, the t-statistics are 11.4 with White compared with 10.6 with OLS; for the full multi- variate case, they are 3.1 with White and 3.8 with OLS. Second, in scoring variables, the coders recorded their subjective assessment of the reliability of the score assigned. When these reliability measures were used to weight observa- tions, the results were roughly the same. 11. SUBJECTIVE CARDINAL. DATA The first objection to studies-and results-of this kind is that the data gener- ated by the systematic case review method are subjective. According to this skeptical view, subjective data are unreliable, ordinal, or both, and therefore inappropriate for statistical analysis. In this section we show that our data are subjective, yet reliable and cardinal. First, we argue that the automatic associa- tion of subjective phenomena with ordinal data is incorrect. Second, the degree of intercoder agreement on the scoring of the major variables reveals that the subjective measurement error, although present, is a minor source of variation. Third, using techniques appropriate for ordinal data does not dramatically change the results, and the constraints imposed by linear regressioni analysis are also not rejected by the data. 182 THIE+ WORLD BANK I (ONQMA( REVIEW. VO\ 4, _O ' Subjective anid Objectiv.e, Ordinal and Cardinial Economic theory often creates a presumption that objective phenomena (such as numbers of oranges or relative prices) have a natural cardinal metric (such as integers or real numbers) whereas intrinsically subjective phenomena (such as consumer utility) allow only ordinal comparisons that are better or worse, espe- cially intersubjectively. Ordinalitv stems directly from the basic theory of map- ping a binary preference relation into a utility index. With onlv the barest restric- tions imposed on the preference relation (complete, reflexive, and transitive), a numerical utility index can be derived, but any monotone transform of that index represents the same preferences equally well. This is not to say that all utility functionis are ordinal. Often additional assumptions are imposed that imply cardinal functions that are unique only up to an affine transformation (for example, vonI Neumann-Morgenstern utility functions). This distinction betweenA cardinal and ordinal is critical for empirical work. Although both cardinal and ordinal data can rank phenomena, only cardinal data can be used to tabulate numbers and to directly compare values of the phenomena being measured. Common statistical techniques such as correlations or linear regression cannot be applied to ordinal data. For instance, if x were an ordinal measure of participation, estimating the linear model 1y = ,B x could produce different results than estimating yv = , 1- f(x), where f(x) is a monotone transform of x, even though x and f(x) would represent exactly the same infor- mation. Therefore, any statistical procedure that relied on summing observa- tions (or any other comparison of the magnitude of the distance between obser- vations) would be invalid for ordinal data. Nevertheless, the data used in this analysis, created by applying the systematic case review method, are subjective, yet cardinal. Our data on rural water proj- ects are doubly subjective: the original project evaluator subjectively assessed and described the amount of participation in each project, and a coder later read the evaluator's report and subjectively assigned a level of participation to that project. If this process generated ordinal data, empirical analysis would be diffi- cult. But note that in everyday life we observe many events that generate subjec- tive, cardinal data. Contests are the most obvious example. When hogs, figure skaters, or bodybuilders compete, judges assign cardinal scores to subjective criteria: quality of coat for hogs; artistic impression for figure skaters; and mus- cle tone for bodybuilders. Grades for academic papers are another familiar example: a professor's subjective evaluation of a humanities paper is given a cardinal score. In each case, these subjectively assigned scores are added, aver- aged, and tabulated in ways only appropriate to cardinal data. Of course, the judging and grading criteria are created to achieve intersubjective consensus. This means that judging requires traininig to achieve this level of intersubjective agreement. For instance, judges of livestock contests are occasionally judged on the degree to which their subjective judgments conform to those of established judges (Herren 1984). Thus, the question for this data set on characteristics of Isham, Narayan, and Pritcbett 183 water projects is not whether the data are subjective, but whether the cardinal scores based on them are reliable. The notions of reliability and validity play a large role in the literature on educational and psychometric testing. Reliability typically refers to whether different versions of the same test on the same indi- vidual will produce the same result (for example, whether repeated I.Q. tests will produce the same result). Validity refers to the usefulness of the tests in some application (for example, the usefulness of scores on the Graduate Record Exam in identifying successful graduate students). Although we now show reliability, we have no external check on validity. Intersubjective Agreement Because project variables were scored from the same documents by two inde- pendent coders, the coherence of their separate scores illuminates the overall reliability of the variables. Table 2 presents two measures of the crosscoder agreement. The correlations between the scores of coder A and coder B are strikingly high: 0.95 for OPE and 0.92 for OBP. The average absolute value of the difference in the scores (on a scale of I to 7) is 0.36 for OPE and 0.55 for OBP. The difference is quite small: most scores either agree or differ by just one point. For each of the two major variables, the coders disagreed by two or more points on only one project. This high degree of intersubjective consensus between reasonably independent coders has two important implications: it implies that the characteristics of the project could be reliably gauged from the project documents (although the re- liability was much lower for some other variables in the data set) and places a relatively tight bound on the magnitude of measurement error. A correlation coefficient of 0.9 implies that the noise from measurement error is roughly 10 percent of the variance of the observed variable.2 Table 2. Crosscoder Reliability Correlation between Average absolute Variable coders A and B difference in scoresa Overall project effectiveness (OP'E) 0.95 0.36 Overall beneficiary participation (OBI') 0.92 0.55 a. The average absolute difference between the scores of the two coders. The scores are on a scale of l to 7. Source: A uthors' calculations. 2. If two observations differ by only measurement error, then the correlation between the two observa- tions is p= ./ ai+ \a + a-# where a2. is the variance of the true variable and a2A is the measurement error variance for coder A(B). If the measurement error variance for both coders is equal a7 = at'- - cr2 a correlation of 0.9 implies that the ratio of measurement error to true variance a2/ a'. is about . 1. 184 TEIL WKIR ) BANK I( IINOMMI[C RI VI[I. WVOl 4, NO. 2 Testing Lin earity or Cardinality We examine in several ways whether the assumption of cardinality affects the results of the analysis of the relation between project performance and participa- tion. We do a simple test for linearity of the relationship. We estimate the participation effect by treating the participation data as if they were ordinal, using dummy variables for each level of participation. We estimate the relation- ship by using the ordered probit estimation technique, which treats the project performance data as ordinal. Finally, we estimate the model with both participa- tion and performance treated as binary variables. Of course, these techniques do not prove cardinality of the data; they do, however, show that the basic results on beneficiary participation are unaffected either by allowing for the possibility that the data are ordinal or by our assumptions of functional form. The first approach argues that if the participation data were in fact ordinal, the relationship between OPE anld OBP would not be linear. That is, the true underlying relationship between the ordinal variables would not be invariant with respect to arbitrary transformations (for example, squaring) of the data. The second column of table 3 presents a test of linearity allowing for a slope shift depending on the value of participation. When participation is low (OBP less than 3.5), the slope is Al, and when participation is high (OBP greater than 3.5), the slope is ,BI + P2. The estimates suggest that the incremental impact of participation is larger at higher levels (the slope is 0.466 for OBP less than 3.5 compared with 0.781 for OBP greater than 3.5). This difference is not statistically significant; a test of the differ- ence resulted in a low t-statistic and a declining adjusted R2. The second approach treats participation as if it were ordinal while treating performanice as cardinal. Each discrete level of the participation variable is en- tered into the performance equation as a dummy variable. The first dummy equals I if OBP is less than or equal to 1.5; 0 otherwise. The second dummy equals I if OBP is greater than 1.5 and less than or equal to 2.5. And so on. Ranges for the variables were specified to generate these dummy variables be- cause the averages for the coders' scores were not always integers. This func- tional form imposes no a priori constraints on the effect of the independent variable. The results in the third column of table 3 show a strong participation effect-performance increases for each performance category-without any strong indications that this statistically unconstrained fit is tremendously supe- rior to the imposition of linearity. The adjusted R2 is lower with the series of dummy variahles, but this combines the effects of relaxing linearity with impos- ing a discrete step function. The implied slope from category to category (from the differences in the coefficients and means of participation) ranges from 0.46 to 0.87. The mean is 0.642, roughly equivalent to the overall linear slope of 0.623. Increases in participation have roughly the same impact along the range from low to high participation. The third approach uses ordered probit estimation by creating a categorical variable for each range of performance. Again, because the averages of the Ishba.n Narayvan, ,and Pritcbett 185 Table 3. Estimationi Resuilts for Alternative Funtictional Forms Binar7>y Binary participatI01n perlormance mieasure ?4 Linear Implied Ordierecd participation Va7riable Line7ar itith kink Coeflhco'nt slopel probit5 menasures, Participation 0.(623 0.466 0.552 0.552 coefficient (10.6) (3.007) (7.906) (7.34) Participation 0.3 15 coetticient after the (1.33) kink point"l Constant 1.79 2.09 ( .5 ) (6.13) Second constant (0.027 (0.066 Cuatof porints tor ordered probmi First 0.957 Second 1.39 Third 2.16 Fourtih 2.9 3 Prtzicipation dummy' v'ariable <1.5 , 55 I. <-5< 5 .3.06 0.46(0 2.5< < 3.5 3.59 (0.624 3.5< < .5 4.25 0.609 45< -< 5.5 5.1 6 0.877 >.5. -5 5.74 0.6.39 k2, o0.481 0.480 0.459 0.1828 0.306 Note: Estimates are based on ordinarv least squares (ol.s) regression unless orherwise noted. t-statis- tics are in parentheses. a. The implied slope is the difference in the coefficienit across participation categories divided by the difference in the means of the participation variable across categories. b. The niagnILide of the coefficient in the ordered probit model is inot directly comparable to the OLS models or the binary nmodel in the sixth column. The fact that both values are 0.5 52 is a pure coincidence. c. The binary model is estimated as a linear probabilitv iiiodel. d. For valLies of the participationi variable greater than 3.5. e. The R2 is not comparable between the linear regressions and the ordered probit pseudo AR2 (or the binary model because the dependent variable is transformed. Source: Authors' calculations. coders' scores are used, the variables are not just levels. Five categories were created by dividing at OPE levels of 2.5, 3.5, 4.5, and 5.5. Ordered probit estima- tion uses only the information that performance categories are different and that higher performance categories represent better levels of performance. It does not use the magnitude of the differences and hence does not assume that the perfor- mance variable is cardinal. The fifth column of table 3 reports the ordered probit results, which again show a strongly significant effect of participation. This comes through in the multivariate ordered probit results as well. The slope coefficient in an ordered probit estimation is not the marginal effect on the probabilities. Calculationis from the results, however, implv that an increase of 186 THiE WORI I) BANK [( ONOMIC: RVIE, VOl _9, NO. 2 participation from 3.5 to 4.5 would reduce the probability of failure (OBP falling in the lowest performance category) by 62 percent (from 0.164 to 0.063) and increase the probability of excellent performance (being in the highest perfor- mance category) by' more than 100 percent (from 0.157 to 0.325).3 Again, these results are broadly consistent with the results of simple linear OLS.4 The final approach, which checks on cardinality, treats both the performance and the participation data as ordinal. For both project performance and partici- pation, a binary variable takes a value of I if the score is high (OBP greater than 3.5) and 0 otherwise. This procedure is valid even if the data are ordinal; binary variables would be unaffected by monotone transforms. If the data are in fact cardinal, however, this procedure is very inefficient because it throws away all of the information about variation within each of the two performance categories. The final column of table 3 reports the results of this linear probability regres- sion. The performance-participation effect remains evident with this crude trans- formation of the data. The subjective nature of the data per se appears to have no impact on the results. High intersubjective reliability of the measures was achieved. The results appear to be broadly consistent with a simple linear model, and treating either performance or participation data as if they were ordinal produces similar results. III. HALO EFFECTS A potentially more serious problem than the intrinsically subjective nature of the data is that either the initial evaluator of the projects or the coders them- selves succumbed to the plausible assumption that all good things go together: the halo effect. This psychological tendency to associate all good things has been discussed in a number of fields. In particular, there is a large literature in human resource management about the halo effect problem in assessing performance. 3. The formulation for calculating the incremental change in probability ot observing the dependent variable in one of I categories (where the J categories are defined by whether they fall between endo- genously determined cutoff points: 0 < pr < p ....< p1 ) with respect to a change in a dependent variable inan ordered probit model iseaProb[xy= 01/X = -1c/P3X),for rle lowest category. For the higlhest category the formula isaProbly =]l/aX = ((p, - P'X)P, where P)(.) is the value ofthe standard normal probability density function (Greene 1990). Using these formulas at 0)BI' = 3.5, the impact of changing OBP for the probability of OPE's occurring in the lowest category is - 0.034, and for the highest category the niarginal impact of OBP oni the probability is 0. 132. 4. UJsing the simple OLS model, we can calculate the change in probability of project failure as the change in the probability that OPE iS less than 2.5 when OPE equials 3.5 compared with the probability that OrE is less than 2.5 when OPE equals 4.5 (which is not exactly comparable as the cutoff points are endogenously estimated in an ordered probit). The first probability for instaince wouild be Pr(a + ,B3.5 + e< 2.5), which, given our estimates of a = 1.79, ,B = 0.623, and c*' = 1.246 and assumilIg that the error term is nornial. is the same as the probability thatxis less than - 1.316, where zis a stanidard iuormal. A similar calculation can he performed for oPF greater than 5.5. With our linear estimates the probability of oPE less than 2.5 falls by 0.064 (froni 0.094 to 0.03(0) and the probability of OPF greater than 5.5 rises by 0.123 (from 0.085 to 0.208). Isham, Naravrn, andl Pritchett 187 Outstanding performance in one dimension or characteristic (even a potentially irrelevant characteristic, such as physical attractiveness) may tend to bias up- ward the evaluation of other dimensions or characteristics. Hammermesh and Biddle (1994) find that plain people make about 5 percent less and attractive people 5 percent more than persons of average attractiveness. However, for a recent dissenting view on the importance of halo effects in performance evalua- tion, see Murphy, Jako, and Anhalt (1993). The halo effect occurs when the measurement of one variable is affected by the observed state of another variable. This systematic measurement error will in- duce an association between two subjectively measured variables even in the absence of any true relation between the underlying variables. In our study, the halo effect may occur at two stages. The evaluators may have falsely attributed participation to successful projects (or vice versa), or the coders-searching the project documents for evidence of project participation-may have been affected by their simultaneous assessment of project success despite their efforts to re- main objective. The second possibility is particularly dangerous. In this study the two coders knew the purpose of the empirical exercise and may have had some strong prior beliefs as to the expected outcome.i There is nothing we can do about the potential halo effect of the original evaluations. We know that the project reports were regular parts of the institu- tional evaluation cycle and that it is doubtful that the financing agencies had a particular stake in promoting participation. It can also be expected that the many individuals writing the project documents would have widely different beliefs about the importance of participation, so that a uniform bias in the firsthand assessments would be unlikely. We explore three methods of addressing the problem of the potentially serious halo effect in the coding process. Note that the results in tables 1 and 3 are based on the average of the two coders' assessments. In the first method we estimated the same models using only data from coder A and from coder B. Differences in these two assessments may reflect differences in the halo effect between the coders. In the second, we estimated the same models using coder A's assessment of the explanatory variables (including OBP) with coder B's assessment of the dependent variable (oPiE). Because coder A's assessments of participation and the other potential determinants are not affected by coder B's performance assess- ment, the halo effect bias should be reduced (although the confounding effect of pure measurement error in the coders' assessments will be important). In the third method, we use project performance indicators-created bv the coders- that, by their nature, are more objective than others. If halo effects were present, they would be more likely to appear for the more subjective indicators. 5. In fact, one of the coders had participated in a previous similar empirical study that had found signifi- canr effects of participation. The other coder was a graduate student who was hired and trained to code for this exercise but was new to) the field and to the topic. 188 THI W('OR1I) BANK I ( ONONII RFVII W, Vol 9 NO) _ Table 4. Estimationi Results bv and across Coders Scores uised Average for Coder A's Coder B's coders Coder A OPE and coder Coder B OPE and coder Model A ,iyid B only, B's OBP 0nl A's OBP Bivartate model Coefficient 0.62 (.6() 0.62 0.60 (.57 (10.6) ((.1) l10.3) (9.7) (9.3) Samplesize 121 111 116 III 116 R' )0.49 0.49 0.48 0.46 0.43 Full multiu'ariate mode(l Coefficient 0.24 0.23 0.26 0.21 O.25 (3.8) (2.6) (2.1) ))) (2.7) Sample size 68 37 46 46 37 R2 (0.89 0.94 0.85 0.89 0.94 Note: Results are presenited for OLS regression of overall project effectiveness (oPr.) on1 overall henefici- ary participation (OBP). t-statistics are in parentheses. Source: Authors calculations. Results by, Coder and across Coders Table 4 shows the results of OLS estimation using the average scores of the two coders (A and B) and using each coder's scores. The table also shows the results of regressing coder A's score for OPE on1 coder B's score for OBP and of regressing coder B's score for OPE on coder A's score for OBP. The differences for both the bivariate and multivariate models are very small. In both models the coefficient does not systematically change, whether we use the average of the coders' scores, each coder's owni scores, or one coder's dependent variable scores on the other coder's independent variable scores. How reassuring are these crosscoder results? Suppose that A's observation on project performance is the truth (OPE*) plus some random noise (e.), plus an upward bias based on A's observation of participation. Then (7) OPEA = OPE + ±A * OBPA + EA. Coder A's observation on participation is just the truth plus random error: (8) OBPl = OBP* + rlA- In this case, if we assume that the pure observational errors are uncorrelated, the coefficient of regressing performance on participation will still be biased upward by the halo effect. If the true structural relationship were (9) OPE=/3*OBP±+, the estimated coefficient would he (10) A+ 3. So even if there were no structural relationship between the true variables (,B = 0), the estimate of the participation effect could be spuriously positive be- cause of halo effects. Isbl'Iom, Narcivai I_7d Pratcbett IS9 Table 5. Resuilts ot Monite Cairlo Simulations of the Combined Effects of Measurement Error and(i Halo Effects Using Crosscoder Informiiation Degree o" measu roenent error (k) Zero, k = i', k = (9. 27 High. k = (.5 (oder Coder A's C(oder ( odcr A's Coder Coder A's A's ()PE on A', (Pif on A's OPE on Degree ofl-/7io s(ores coder B'-s SCoreS codesr Bs scorcs cocer B's effect (61 onlly (BP on/v OBP (onI/, OBP Zero, 6 0 0.5 9.5 0.4(0 0.40 (3 .3 0.33 Low, 6 = 0.25 (. .7 5 9.( i 0.60 (.58 0.50 MN1odera te, S= 0.5 I I 0(.91) t).80 0.8.3 0.67 High, S= I I.5 1.5 1.4 1.2 1. . I0 Note: Thl results are thl- ave rage cstitnates fr-oim 1,0I)() rpi ariotii osf cacieh o the 129 observati ons of tilhe miodel: y= fix + g, [ = (O.S, x., - N(0),l). The valtles iI thel tirst two coiiinmns Of rthe ftirst tow are/ = (.5, tile rtr]e ValN wiril 1th n11CoAsU reICIelt errior Aoild io hatlo ieffet. A;BUi obserationis on the x varablhie are sLIject to me,sorermien t e-rroir of the foriri: y kio;' = \kii . XY IS the tuLeC valuLIC, alnd 1 - N'O. II \Nhere A'1 B! IndiC-tS tha.t A and B, haK e ind.epli(lendent randotmi meaCsuISLreeli t trr ot pro i puotrtil oi. The ohserv.trion ion tile deepe nl dent vi ri ahie ! .1 re demterI edh ! =! + I ciii So that the Te.llsaLlrelnlrolt e--or (If A (r- B n11flUellieS rile the m sLii reMCe ot v he a m0111111iim "hi i0'' factor of 6. Souree: Atitltors' calculiatioi. Given this background, why does havinig another coder matter? If the observations-and scorinig-of participation are comiipletely objective, usiSlg a second coder's data will have no effect: As and B's observations on participa- tion will be identical (OBP4 = OBP6). And if the degree of halo effect is similar 6= 0), the bias on performanice will be equivalenit. If the observations of participation are subjective, then the halo effect bias should be less using cross- coder data because the pure subjective componient of B's assessment does not affect the bias in the performanlce measure. However, to the extent the measure of performanice is truly subjective, this measurement error argues against the intersubjective reliability above and induces downward bias in the estimates caused by classical measuremenit error. Table 5 shows Mionte Carlo simulations of the combined effects of the halo effect hias and of piure me3suremenit error, using different assumptions about the relative strengths of the two effects. ULifortuniately the simulations show that both underestimation and overestimation of the true coefficient are possible when coders' scores are crossed, depending on the ratio of the halo and measure- ment error effects. With high measurement error (fifth and sixth columns), crossing the coder rankings should produce lower estimates than using the rank- ings of a single coder for all possible strengths of halo effect. For inputs and outputs with no measuremilenit error (the first and second columns) and high degrees of the halo effect, crossing the coder rankings does not help because it produces the same estimates (with a similar upward bias). Evidence of the relatively high intersubjective reliability (as well as the instru- mental variable results below) suggests low but nonzero measurement error. The 190 IHF F.RI 1) BANK F ONO MI C REVIF W, \V I. Q. No. 2 Table 6. The Impact of Participation on Various Indicators of Performance Percentage Percentage Overall Objective of uater of target project value of sTstem in good population Model effectiveness beozefrts conditiont reached Bivariate model Coefficient on 0.62 0.5 3 0.54 0.29 participation (10.6) (10.3) (6.4) (5.30) Sample size 121 120 98 118 R2 50.49 0.47 0.29 0.19 Full multitariate mtiodel Coefficient on 0.24 0.27 0.29 0.25 participation (3.8) (.36) {2.4) (2.50) Sample size 68 68 60 68 R2 (.89 0.79 0.77 0.47 Note: Each colulmn presents the results of OLS regression of the bivariate and full multivariate models using different objective indicators of proiect success. I-statistics are in parentheses. Source: Authors calculations. ratio of the measurement error variance ((X2) to the true variance (us), that is, C2IC2., is between 0.1 and 0.25. In this range of measurement error (third and forth columns), when the coders' OPE and OBP rankings are crossed there should be a modest but significant change in estimates if the halo effect is strong. The lack of a consistent downward movement of the estimates (in the multivariate case they actually rise) suggests at least that the halo effect is not dominating the results. Results by Performlance Indicator The third method of evaluating the halo effect is to examine the impact of participation on project performance indicators that vary in their objectivity. In addition to the omnibus measure OPE, several more objective indicators of proj- ect success were coded, including the percentage of the water system in good condition and the percentage of the population target reached. To the extent that these more objective phenomena are relatively less susceptible to halo over- estimation, the possible halo effect bias should be reduced. If the true coefficient is equal across models using different dependent vari- ables (which is not clear-see table 6), the more objective indicator should be systematically lower than the upwardly biased subjective indicator. Table 6 pre- sents these results. There is no evidence that the more subjective indicators (such as OPE) have systematically larger estimated impacts. IV. JOINT DETERMINATION AND CAUSALITY The previous two sections have answered possible skepticism about the strong statistical association between performance and participation. Other skeptics may accept the statistical association between participation and performance but deny that this association reveals cause and effect. In this view, the data do not Isatin, Narayan, and Pritchett 191 show that greater participation causes better project performance but simply that participation and performance happen to be related. Indeed, there are at least two good reasons to believe that an association between participation and performance may not be causal. First, there could be reverse causation: projects that are exogenously better might induce greater beneficiary participation. Reverse causation makes sense, especially when performance and participation depend on a sequence of actions. Once it is clear that a project is failing, potential beneficiaries may be less likely to participate because they perceive a relatively low benefit from their own participation, which is unlikely to alter the project outcome. Second, joint determination of project success and beneficiary participation may be driven by a third local or project attribute. For example, if dynamic leaders induce both project performance and participation, performance and participation data will be strongly associated with each other-even without an independent causal effect of participation on1 performance. Although we have tried to address this concern over spurious correlation with the inclusion of possible performance determinants, it would not take too clever a skeptic to name a large list of excluded variables (some of which are unobservable even in principle) that could affect both performance and participation. We use three approaches to resolve the problems of reverse causation and joint association and to demonstrate a causal impact: estimation with instrumental variables, timing, and case studies. Instriinmental Variables One econometric solution to the problem of identifying a structural relation- slhip is estimation with instrumental variables (IV). This estimation avoids the problem of the joint determinlationl of the indepenidenit and dependent variables: in estimating the coefficients, it eliminates that part of the variation in the inde- pendenit variable that is caused by variation in the dependent variable. The vehicle for eliminating that variation is a third variable (the instrument), which affects only the independent variable and not the dependent variable. Estimation witlh instrumental variables requires a variable that affects partici- pation but that neither affects performance directly, nor is affected by it. This variable is used as the instrumenit for purging the participation variable of any performance-related component. When the participation effect is estimated using only the part of participation variation that is correlated with the variation in the instrument, the resulting estimate is free of reverse causation. Because better performanice does not affect the instrument, the reverse effect of better performance on participation is eliminated and cannot bias the results. For the following model (11) OPE = ,B * OBPP+ . * Z, +3, * W, +±3, * V, +E, (12) O BP= az * Z, +a,, * W, +a,, * V, +,7, Tahle 7. Com77parison of OLS and InIstrumz7len7tal Variables Istizmates of the Participation7-Performance Relation I nIstrume'ntal variables Perce'ntage Fxtenlt to tW bich cof jnV'stmo7'nt Net benefits ()rganinationz Prior participationz c-osts paitl b'v o0/ based ., local oMnnitment Model ()IS wais a goal uisers participation collectives of clients All instrum1ents Bivariate mnodel Coefficient 0.6.3 0.70 0.59 0.77' 0.74 0.97 0.86 (10.6) (10.2) (7.3) (10.6) (6.3) (7.54) (10.4) Samplesize 120 120 113 120 98b 105 90 R 2 0.488 0.482 0.476 0.453 0.507b1 0.378 0.521 xc First-stage R2 i na. 0.763 0.573 0.701 0.364 0.326 0.816 > Lunimited multiv'ariatc' model Coefficient ().28 0.34 0.32 0.36 0.15 0.39 0.37 (5.25) (5.2) (3.6) (5.4) (1.28) (3.00) (3.57) Sampic size . 5 7 66 .2 6.3 R2 0.862 0.860 0.86 1 0.8.58 0.X8.5. 0.863 (.865 First-stage R2 0.401 0.826 0.643 0)80.3 (.719 (.559 (.857 na. Not applicable. a. Unadjusred R2 of the firsr-srage regression of participation on the instruments (which in the mulrivariate case includcs all variables in the performnance equtioin). b. Because rhe sample siWes are not the samc, the results are iiot strictly comparable in all columnis. In particular. the IV R2 *alues are less rhiia the OLS R2 when run for the samc samilple. c. This is the R2 of pa rticipation regrcssed on all the Z variables thar ar-c inclided in the performanlce equation. lhe increment to the R2 for each instrumllent can he calculated as the difference with this colimn. Source: Authors calculations. Isb/in, Naravln, and Pritchett 193 all Vs that are included in the participation equation (a,, t 0) but excluded from the performance equationi (3,, = 0) are legitimate instruments. The Vs provide a source of variation in participation that is exogenous to performance. Neither the Zs nor the Ws are valid instruments, because they directly affect both partici- pation and performance. To choose appropriate instruments, we need a positive model for participation. Hypotheses based on the larger study of these water projects (Naravan 1995) and on other statistical work (Basu and Pritchett 1994), as well as theoretical literature on the determinanits of participation, generated a set of equations for estimating the effects of participation (appendix table A-3 shows the full first-stage regres- sions). We identify five variables as legitimate instruments: the extent to which participation was a project goal, the percentage of investment costs paid by users, beneficiaries' overall net beniefits of participation, the extent to which organization was based on local collectives, and prior commitment of clients. We hypothesize that each of these phenomena may directly affect participation but should have n1o independent, direct effect on project performance after controlling for participation. In table 7 the OLS and IV results are compared-for both the bivariate and the limited multivariate case. The full multivariate case is excluded because it loses too many degrees of freedom; although the results are empirically similar, they are less precise. The estimated impact of participation increases with IV esti- mates. For instance, when the extent to which participation was a project goal is used as an instruLmlent, the bivariate impact rises from 0.63 to 0.70, and the multivariate rises from 0.28 to 0.34. The IV estimation produces a higher and statistically significant estimate for each of the instruments used individually. The only exceptions are for the bivariate case in which percentage of investment cost paid by users is used as an investment and for the multivariate case in which organization based on local collectives is used. In the second case the coefficient drops to 0.15 and is statistically insignificant, probably because of the low power of the instrument. When all instruments are used together, the impact rises to 0.86 in the bivariate case and to 0.37 in the limited multivariate case. What do these IV estimates tell us? The basic statistical relationship-high correlation between participation and performance-would also occur if better project performance caused greater participation. As clean water is delivered in the early stages of a project, more potential beneficiaries may want to get involved. But if this were the causal story, the IV technique would cause the estimates to fall by removing this upward simultaneity bias. The rising coefficients reported in table 7 are consistent with causality running from higlher participation to better project performance in the presence of some measurement error. The IV results allow US to compute an independent estimate of the magnitude of pure measurement error. Even if the OLS estimator is inconsistent with measurement error, the IV estimator is consistent, and the ratio of the estimates converges to plim (aOLS/I ) I a I/ + ( E ). 194 THF. WORID IBANK F( ONOMIC REVIEW, Vl 9, No. ' With the reported estimates, this ratio is between 0.8 and 0.9; for example, 0.63/0.70 (bivariate) or 0.28/0.34 (limited multivariate). The ratio of the vari- ance with measurement error to the true variance, q,21,, is between 0.11 and 0.25. This is consistent with (although somewhat higher than) the estimates of measurement error from the correlations of crosscoder reliability in table 2. In using this technique, it is useful to test whether the assumptions made in obtaining the IV results are valid. Note that because one variable-beneficiary participation-may be endogenous, at least one instrument must be used to identify the model.6 But if it can be unambiguously accepted that one instrument directly affects only the independent variable, then the validity of any other instrument can be tested. Indeed, the entire set of instruments can be tested. We believe that the extent to which participation was a goal is the most plausibly exogenous variable among the individual instruments. There is no reason to believe that participation as a goal would by itself lead to better performance, except insofar as it actually raised participation. When each of the other instruments is tested, conditional on the validity of this variable (using a Hausman-Taylor test), we fail to reject the exogeneity of the other instruments in every case. When the entire set of instruments is tested, we do not reject the validity of the instrument set in either the bivariate or multivariate case. The value of the Sargan test with the full set of instruments is 7.03 (significance level 0.133) in the bivariate and 5.24 (significance level 0.263) in the multivariate estimates. Our set of instruments stands up to the available tests for instrument validity. Of course, the major objection to these tests is that they tend to be of very low statistical power (that is, these tests will often fail to reject a hypothesis that is false). Therefore a failure to reject the instruments cannot be taken as compelling evidence for accepting the instruments. Timning Evidence on causality also can be observed from the timing of the project cycle. If the association between participation and project performance was not causal, we would see no association between events that occur before project completion-proximate determinants of project performance-and beneficiary participation. Table 8 reports the impact of participation on quality of imple- mentation, effectiveness of operations and maintenance, and maintenance after one year. We find that in all but one (multivariate) case, beneficiary participation is a statistically significant input of these proximate determinants. If project effectiveness was causing participation rather than vice versa, we would not expect to see this result. Roughly these same results also hold true when, rather 6. Heuristically, the problem is that we need to rest the exclusion of the instrument fronT the performance equarion. However, we cannot test the exclusion directly (say by a t-test of the inclusion of the instrument) because in the presence of endogeneity the coefficient on the potenitially endogenious variable is inconsistent when it is not instrLimented and hence the t-test onl the instrument woould be biased. Ishain, Naravan, and Pritchett 195 Table 8. The Impact of Benieficiary Participation otl the Proximate Determinants of Project Performance Limited Full Bzl 3ariate multivaricate multivariate Independent variable Model mnodel model Quality of implementation 0.5 3 0.1 7 0.21 (9.1) (2.7) (2.7) Effectiveness of operation and maintenance 0.49 0.14 0.11 (7.4) (2.0) (1.1) Nlaintenance after one year 0.43 0.16 0.18 (6.6) (2.0) (1.8) Note: Values are OLS estimates. t-sratistics are in parentheses. Source: A Utthors' calculatioiis. than overall participation, the participation at various stages (which therefore precede outcomes) is used as the independent variables. Case Studies Case studies of individual projects also help to resolve questions about causal- ity between beneficiary participation and project effectiveness. Narayan (1995) documents the specific effects of participation on performance in many of the projects included in this study. Two of these case studies-in which exogenous shifts in participation during implementation changed project outcomes-clearly illustrate the direction of causality that we have explored econometrically above. Phase I of the Bank's Aguthi Rural Water Supply Project in Kenya was con- ducted without community participation. The project, which involved piped water systems, was so plagued with problems-construction delays, cost over- runs, and disagreements over consumer payment methods-that it came to a standstill. At this point, there was a substantial shift in the level of beneficiary participation. The project was redesigned, and local leaders organized them- selves into the Aguthi Water Committee. Working with project staff, they mo- bilized community support for the project. After public stakeholder conferences, community members began to contribute labor and finances. Phase 11 of the project was completed on schedule and within budget. The communities contin- ued to pay monthly tariffs for the new water service, and operation and mainte- nance of the system were handled successfully, in cooperation with the govern- ment parastatal. The goal of the Wanita, Air dan Sanitasi (WAS) program in Nusa Tengarra Timur, Indonesia, was to help community groups launch and manage their own water system. A water group in the village of Silla was formed in 1986 as WAS began. Initially, the group waited for the arrival of a government team to dig a borehole, but none came. When the group members realized that they could not rely on immediate government assistance, they increased their level of participa- tion in the project. They negotiated water rights with a neighboring water group, collected building materials, and built three water tanks, with only a small 196 THF WORI D BANK [CONOMI(: RFVIF'W, VI()I NO/. 2 amount of outside technical assistance. By 1988 a new well was under construc- tion, financed by their own contributions. This increased level of participation was maintained and led to the project's sustained success. V. CONCLUSION We began by showing the existence in project-level data of a strong associa- tion between project performance and beneficiary participation. We then ad- dressed and answered the three econometric objections to acknowledging this association. The subjectivity of the data is not an overwhelming problem. The halo effect does not appear to induce a strong upward bias. Most important, strong arguments support the cause-effect relation between participation and project performance. This article, together with the more comprehensive work of Narayan (1995) in particular, provides development practitioners-including early and recent converts to the participatory approach-with strong statistical findings that increasing participation directly causes better project outcomes. Four questions that are important for practice and policy are not explored here. First, does participation directly cause better project performance across all sectors? We cannot blindly apply the results in this study across all sectors, because these data are limited to rural water supply projects. The economic characteristics of rural water as a good would seem to promote the importance of direct beneficiary participation; these economic characteristics vary across goods provided by projects in other sectors. Second, what policy instruments help to achieve more effective participation? Project beneficiaries, staff in project agencies, and other suppliers respond to incentives, but there is little documented experience on creating incentives in public sector agencies for promoting and incorporating participation. Third, is the use of participation justified-even if it is costly-simply because participation improves outcomes? Although a full estimate of the costs and benefits is beyond the scope of this article, it is a vital step in the research agenda. Finally, can experiences with participation help to clarify the analysis of the deficiencies inherent in either a purely individualistic market or a purely statist, governmental approach to development? An analytic approach that incorpo- rates participation might examine the various mechanisms whereby cooperative action by groups can overcome the inefficiency of individualistic solutions-for example, from free riding or strategic (mis)revelation of private information- while avoiding the limitations of centralized government. These informal methods of cooperation have been explored by several authors (Ostrom, Schroeder, and Wynne 1993; de Soto 1989; Wade 1988), but much remains to be learned. Isham, Narayan, and Pritcbett 197 Table A-I. Coded Variables and Sunmmarv Statistics Standard Variable Samlple size Mean deviation Per formnance indicators Overall project effectiveness 121 4.1 1.6 Percentage of water svstem in good condition 98 4.8 1.8 Objective value of benefits 12) 4.2 1.3 Percentage of target population reached I 11 4.9 1.1 Participation variable Overall participation l21 3.7 1 7 Fuill-C exogeni is pert ormnaece detcrin inants (Z) ;Ni per capita 114 519.8 389.3 Project comilplexity 121 3.3 1.2 Total cost (In) 104 IS 4 1.5 Adequacy of facilities 121 4.5 1.3 Difficultles in staff recruiting 92 3.8 1.7 Availability of parts 115 4.2 1 .5 Target objectives I 1- 1 4.4 1.2 (Jtber perlrmrmance determinants (W) Appropriateniess of techbiology 121 4.5 1.3 Support of government I 18 4.6 1.1 Agency understanding 118 2.8 (0.9 Conduciveness of political cointext 1 2 1 3.2 0.7 Conducivenless of econiomiiic conitext 121 3.2 0.7 Conduciveness of sociocultural conitext 121 1 3.5 0.7 Conduciveniess of geological andL environmelital context 121 3.2 0.9 Average nunmber of users I 17 3.2 1.1 (oiiipetition from otber sour.es 11)9 3.4 1 .5 Skill of staff II1 4.6 1.2 Overall quality of manageiment 120 4. 2 1.3 Note: All qualitative variable, are raniked on a scale of I to . (;NP per Capita andl total cost are in clollars. Soiurce: Au]thorb' calculations. 198 THF WORI I) BANK ECONOMIC RE,VIEW, Vol 9, No. I Table A-2. Results of Bivariate, Limited Multivariate, and Full Multiv'ariate Regressions of Project Performance, Using Various Models Full Limnitecd nultivariate mnultitvariate inclusive of all inclusive of all Z Z and W Variable Bivartate variables (n1o Ws.) variables Overall beneficiary participation 0.6 0.28 0.24 (10.6) (5.3) (3.8) Availability of parts 0.57 0.44 (9.6) (5.6) Target objectives 0.22 0.04 2. 9) (0.4) Adequacv of facilities 0.14 0.03 1.9) (0.4) GNP per capita -0.0003 -0.00006 (-1.5) (- 0.3) Project complexity -(0(8 -0.07 (-1.3) (-0.9) Difficulties in staff recruiting -(.05 0.006 (-1.1) (0.1) Total cost (In) 0.04 (.08 (0.8) (1.3) Appropriateness of technology 0.19 (2.3) Overall quality of management 0.21 (1.6) Support of government 0.10 (1.1) Average number of users -0.08 (-1.06) Conduciveness of economic context 0. 1 ((0 9) Conduciveness of geological and -0.1 environlmlental context (-0.9) Skill of staff 0.08 (0.8) Agencv understanding (0(5 ((.5) Conduciveness of political context 0.03 (0.24) Competirion from other sources -0.01 (-0.2) CondIIciveness of soCiocultural (.02 context (0. 1) Sample size 121 77 68 R2 0.49 0.86 0.89 Note: All qualitative variables are ranked on a scale of I to 7. GNP per capita and total cost are in dollars. The values are from OL9 regressions in wvhich the dependent variable is overall project effective- ness. t-statistics are in paren theses. Source: Authors' calculatioIs. Table A-3. First-Stage Regressions of Project Participation Percentage Extent to which otmzvesttmenit Net benefits Organization Prior participation costs paid by of based onz commitment Instrument and variable All instruments was a goal users participation local collectit'es otclients Instruments Extent to which participation 0.24 0.682 ws a goal (2.89) 1(2.9) Percentage of investment 0.07 ( O'S costs paid by users (0.87) (6.51) Net benefits of participation 0.248 0.658 (3.06) (11.8) Organization based on 0.1 1 0.44 local collectives (1.47) (5.83) Prior commitment of clients 0.044 0.49 (0.452) (4.03) Exogenous variables in the limited multivariate regressio(n 1'rojectcomplexit. -0.013 -0.034 ((.158 0.014 0.068 0.151 (0.191) (0.431) (1.40) (0.176) (0.73) (1.19) Total cost (In) - 0.098 -0.019 -0.1 69 -0.0)29 0.14 -0.286 (1.73) (0.329) (1.93) (0.474) (1.99) (2.86) Adequacy of facilities 0.1 9 0.1 8 1 0.381 0.22.3 0.3-54 0.382 (2.22) (2.01) (2.98) (2.33) (3.1 5) (2.53) Difficulties in staff - 0.016 0.041 (0.004 -(.011 -0.052 0.009 recruitinig (0.323) (0.75) (0.051) (0.192) (0.783) (0.108) Targetobjectives 0.144 0.12 0.142 0.128 (.187 0.129 (1.88) (1.36) (1.09) (1.38) (1.82) (0.901) Availabilitv of - 0.048 -0.001 0.009 -0.086 0.00(3 -0.028 parts (0.727) (0.016) (0.09) (1.11) (0.035) (0.238) GNPpercapita 0.(0051 0.00073 0.00034 0.00083 0.00043 0.00067 (2.68) (3.51) (1.12) (3.76) (1.76) (1.98) Sample size 6.3 77 75 77 66 72 R 2 0.865 0.826 0.643 0.803 0.719 0.559 Note: All qualitative variables are ranked on a scale of I to 7. GNP per capita and total cost are in dollars. The values are from oi.s regressions in which the dependent variable is over all project effectiveness. t-stahistics are in parentheses. Source: Authors' calculations. 200 rHI WOkR 1) BANK F: (ONO MNi RFVI IF\. VOl 'NO. (2 REFERENCES The word "processed" describes informally reproduced works that may not he com- monly available through library systems. Basu, Ritu, and lIant Pritchett. 1994. "The Determiniants of the Magnitude and Effective- ness of Participation: Evidence from Rural Water Supplv Projects." Backgrotind note for WYZorld Dev'elopment Report 1994. World Bank, Washington, D.C. Briscoe, John, and David de Ferranti. 1988. Water for Rural Communities: Helping People Help Themnseltves. Washington, D.C.: World Bank. de Soto, Hernando. 1989. The Other Path: The Invisible Revolution in the Third Wy'orld. New York: Harper and Row. Esman, Milton J., and Norman T. Uphoff. 1984. Local Organizations: Intermediaries in Rural Development. Ithaca, N.Y.: Cornell University Press. Finsterbusch, Kuirt. 1990. "Studying Success Factors in Multiple Cases Using Low-Cost Methods." University of Maryland, Department of Sociology, College Park, Md. Processed. Finsterbusch, Kurt, and Warren Van Wicklin 111. 1987. "Conitribution of Beneficiary Participation to Developpment Project Effectivenress." liublic Adminiistration and Devel- opment 7(January/March):1-23. Freire, Paulo. 1973. Education /or Critical Consciousness. New York: Seabury Press. Greene, Williarn. 1990. Econometric Analysis. New York: Macmillani. Hammermesi, D)anicl S., and Jeffrey E. Biddle. 1994. "Beautv and the Labor Market." The American Econoinic Reuieu 84 (5, December): 1174-94. fHerren, Hans R. 1984. "Factors Associated with the Success of Participants in the Na- tional Future Varmers of America Livestock Judging Contest." Journal of American Association of TeaLchinig anid Education in Agricultutre 25(1):1-19. Korten, David C. 1980. "Comimutnity Organization and Rural Development: A Learning Process Approach." Pu1blic Administration Rcvietv 40(5, September-October):480- 510. Korten, Frances F., and Robert Siy, Jr., eds. 1988. Transforming Bureaucracy: The Expe- rience of the Philippine National Irrigation Admninistration. West Hartford, Conn.: Kiumariani Press. Murphy, Kevin R., Robert A. Jako, anid Rebecca Anhalt. 1993. "Nature and Conse- quenices of Halo Error: A Critical Analvsis." Journal of Applied Psvchology 78(2)):218- 25. Narayan, Deepa. 1995. "The Contributioni of People's Participation: Evidence from 121 Rural Water Supply Projects." ESD) Occasional Paper Series 1. World Bank, Environ- mentally Sustainable Developmenit Department, WXashington, D.C. Processed. Ostroin, Elinor, ILarry Schroeder, and Susani Wynne. 1993. Institutional Incentives and .Sustainable Development: lnfrastructuire Policies in Perspective. Boulder, Colo.: Westview. UNDI' (United Nations Development Programme). 1993. The Humian Development Re- port 1993. New York: Oxford University Press. Wade, Robert. 1988. Village Republics: Econiomyiic Conditions for Collective Actioni in South Indiia. Cambridge, U.K.: Cambridge Liniversity P'ress. World Bank. 1991. World Development Report 1991. New York: Oxford University Press. I Hii W I R I I) R A N K F ti N O\ 1 ( R I F V. VO ( , N O . 2 t I -' *) Winners and Losers in Transition: Returns to Education, Experience, and Gender in Slovenia Peter F. Orazem and Milan Vodopivec This article, using an unusually rich data set on7 Slovenian wvorkers over the 1987-91 period, explores changes in the structure of /vages and employment produced by transi- tion to a market economny. Employmentt and real uvages fell dramatically over the period, but the losses wvere bornie disproportionately by the least skilled. Across all sectors of the econmny. relative u'ages and employment rose for the most-educated w^orkers. Women gained inz comparison with nien, primarily hecause rwomen occiupiedl sectors less adiersely affected by the transition. Pensionl policies, ivhich encouraged retirem7xentt, are show n to have dlrastically redce(il employmient of experienced nvorkers anlld helped contribute to risinig returnis to skill. Increases in retuinis to education and experience contributed to risinig wvage inequality, bn t the, variance ot wvages increased for ivorkers ivith identical skills as i ell. Drawing on the writings of Marx, socialist governments sought to suppress the labor market. Labor supply mechanisms were constrained because work was regarded as a sacred dutv of all members of the socialist society. Those who did not want to work were stigmatized and sometimes sentenced to compulsory jobs. Occupational and educationial choices were centrally rationed, and geo- graphic migration was restricted by rationing housing and restricting the sale of property. On the labor demand side, jobs were provided for virtually everyone, and firing was effectively forbidden. Hiring and promotions were influeniced by ideological criteria. To ensure egalitarian income distribution, and perhaps to discourage individLIals from switching jobs or localities, economvwide wage rates were assigned for all classes of jobs. Commodity prices were also issued centrally, with an emphasis on stability over time. Consumer, producer, and worker expectations were formed in a world of stable prices, stable wages, stable labor demand, and stable labor supply. The ultimate collapse of the socialist economies resulted in disequilibriulm1 of epic proportion. Transition economies faced a complete disruptioll of their Peter F. Orazem is in tht Departmenit of Economics at Iowa State Unliversity, and N1ilan Vodopivec is in the Policy Research De partnieit at the World BaLik. The arthors are grateful to the Statistical Office, the Pensioni and Invalid Fulid, and the Emplovnyent Office of Slovenia for providing the data Lised in this analysis; to Dehahrara Dj. and Ruitlh wLi for -areftl- research assistance; aLnd to serminar participants at the World Bank and threc anonymoron refcrees for usCfil commelibe(1t,. IX 1995 The International Bank for Reconstr-Ltion 1and Developnmenit /THE WORLD RANK 20(1 202 T HF WORI 1) BANK Fi()NOMI( REVi1 W, VOL. 9,N() ' pretransition economic system, and therefore of price and wage determina- tion, production and consumption decisions, and expectations formation. To varying degrees since the collapse, the former socialist economies have been dis- mantling the mechanisms that held the labor market in check. Workers have been assigned the responsibility for finding work and have been given the right to make occupational and geographical choices. Wages and prices have been allowed to adjust to market forces. Rewards have been granted, and failures have been allowed to occur. In general, the costs of transition from a planned economy are unlikely to be borne equally by all. Several factors may cause the least skilled to face particular hardships under transition. These factors include short-term changes in returns to skill associated with the process of transition itself. They also include long- term changes associated with corrections of distortions in the wage structure in the old regime and changes in the mix of final production demands. Moreover, the process of transition creates a need for entrepreneurial skills that may raise relative returns to human capital. Schultz (1975: 832) argues forcefully that economic systems characterized by constancy "place no premium for the human ability to deal with secular eco- nomic change." Systems in flux require this ability to perceive disequilibrium and to reallocate resources accordingly. The gains from such reallocations are observable in wages and profits. Schultz stresses that presumed gains to entre- preneurial ability during periods of change are relative gains. "For people to have gains from their resource allocations does not imply that they are neces- sarily better off than they were prior to the disequilibrium, but it does imply that their economic position has been improved relative to what it would be if they had stayed in disequilibrium" (Schultz 1975: 834). If, as Schultz posits, entrepre- neurial ability is complementary with education and work experience, then rela- tive returns to education and experience should rise in the newly emerging market economies in comparison with pretransition returns. There may be long-term forces that would raise relative returns to skill as well. Pretransition wages for low-skilled industrial workers were often set near or even above wages for workers in occupations and sectors that required more education. (See Redor 1992 for a discussion of wage setting in socialist econ- omies.) Market wages would not support such low or even negative returns to skill. The increase in returns to skill could be reinforced by disruptions in trade patterns that limit markets for industrial goods if low-skill-intensive sectors were most adversely affected by shifts in final demand. There are many other ways in which the structure of earnings might change as a result of transition to a market economy. Competitive forces might be expected to remove arbitrary wage-setting mechaniisms that tend to discriminate against women or minorities. Centrally planned economies, however, were charac- terized by high female labor force participation rates and female-male wage ratios that were comparable to those in Western economies. If women did rela- tively well under socialism, it might be presumed that they would lose in transi- Orazem and Vodopwvec 20.3 tion. Egalitarian, centrally directed wage setting tends to reduce inequality. Removal of these egalitarian policies would be expected to increase wage inequality. Their removal might also reduce relative earnings for minority groups and women to the extent that a common wage formula might prevent discrimination. This study examines the winners and losers in Slovenia's transition by tracing out changes in returns to education, experience, and gender and changes in wage inequality from 1987 to 1991. The main finding is that returns to human capital increased dramatically during transition. Rising returns to education and experi- ence contributed to rising wage inequality, but the variance of wages increased for individuals with identical skills as well. Women gained over men, primarily because women occupied sectors less adversely affected by the transition. Efforts to use pension policies to encourage retirements have drastically reduced the labor supply of experienced workers of pensionable age, contributing to large relative wage increases for workers of pensionable age and work experience. The Slovenian results are in sharp contrast to those in East Germany (Krueger and Pischke 1992; Bird, Schwarze, and Wagner 1994); East German workers have had decreasing returns to education and experience. However, it is not clear that the East German labor market has lessons for other transition econ- omies, because of West Germany's intervention to lessen the adverse effects of transition in East Germany. Perhaps the best indicator of the uniqueness of the East German transition is that real wages rose dramatically in East Germany, although they have fallen in every other formerly socialist economy. The massive transfers of capital from West Germany to East Germany cannot bIe replicated in other transition economies. Using grouped data, Flanagan (1993) finds that returns to education rose in the Czech Republic during transition but that returns to potential experience fell. Although Brown (1992) and Knight and Lina (1993) do not examine changes over time, they find results similar to ours for other economies in transi- tion. Brown finds that returns to education in Estonia are larger in the emerging private sector than in the state sector. Knight and Lina (1993) find greater re- turns to education in the less institutionalized, rural labor market than in the urban labor market in China. If we set East Germany aside as a unique case, the Slovenian data base offers the first chance to use data for individuals, rather than aggregate group data, to examine the dynamics of labor market returns in an economy in transition. Although there are respects in which Slovenia has been atypical (being among the richer and more Western-oriented of transition economies), lessons from the Slovenian experience should be applicable elsewhere. The pretransition Slove- nian labor market shared key features with other former socialist economies, most notably, social ownership, full employment coupled with substantial hid- den unemployment, and an egalitarian wage structure. In addition, Slovenia has introduced labor market reforms and has experienced dislocations that are simi- lar to those of other European transition econonies. 204 1 IF WORI I BANK 1( ONONII( RI lVI l:W. \V()I q No. ' Section I reviews the labor market institutionis that were in place before and during transition. The data are described in section 11. Empirical outcomes for wages and employment are reported in sections III and IV. Section V concludes with policy implications derived from the analysis. 1. INSTITUTIONAL BACKGROUND To put in perspective changes in wages and other labor market outcomes, it is necessary to review the features of the previous self-management system in Yugoslavia. Subsequent labor market reforms can be viewed as relaxing con- straints on labor market outcomes that were imposed under the earlier system. The effectiveness of the reforms canl be judged by the extent to which labor market outcomes changed accordingly with the relaxation of these constraints. Labor Market uinder Self-Management in Yuigoslavia Until achieving independence in October 1991, Slovenia was one of the six constituent republics of Yugoslavia. From the early 1950s until 1988-89, Yugoslavia maintained a uLique social and economic system known as worker self-management. The absence of explicit property rights under social ownership and the commitment to self-management dictated a specific wage-setting mecha- nism. Both government and workers had clearly delineated roles in the wage- setting process. In the absence of an advocate for capital, the government set the firm's wage bill. The government's objective was to even out differences in pay among firms. Special regulations termed social agreements laid out a detailed methodology for determining each firm's "socially warranted" wage bill. In the late 1980s the methodology involved computation of a business success index for each firm. The index depended positively on firm income per worker and per unit of capital stock. The indexes were then adjusted by a special correction factor. The adjusted indexes dampened the measured business success for above-average firms and raised measured success for below-average firms. The socially warranted wage bill was then computed on the basis of the adjusted indexes for firm success. Other things being equal, wages in a firm with an unadjusted success index 60 percent above average would be only 25 percent above average, and wages in a firm with an unadjusted success index 40 percent below average would be only 19 percenlt below average (Vodopivec 1993). Further leveling of wages across firms was achieved through a massive firm income redistribution policy implemeiited by discretionary taxation and subsi- dization of enterprises. Paying wages strictly according to the socially warranted wage bill meant that proportionally more income was retained by above-average firms than by below-average firms. Government control of the distribution of firm incomiie created numerous channels by which income could be shifted from above-average to below-average firms. For example, below-average firms whose socially warranted wage bills exhausted their earnings were exempted from taxes and qualified for subsidies (chief among them, concessiotnary financing). Orazemz anid Vodopzvec 20s These subsidies were financed by discretionary taxes on firms with above- average business success. The magnitude of these discretionary redistributive flows was staggering, exceeding by several times formal taxes and formal sub- sidies (see Vodopivec 1993). Once the government set the wage bill, the workers' role was to set individual wages within the firm. The wage scale was determined by a referendum of employees. Theoretically, pay depended also on supervisors' assessments of work quality and fulfillment of work norms, but in reality the wage scale was decisive. In comparison with capitalist firms, Yugoslav firms had extremely com- pressed wage scales. For example, in one establishment with several thousand workers, the pay of the highest-paid manager was 4.54 times that of the lowest- paid laborer. To put that range in perspective, the pay range of entry-level posi- tions in U.S. state governments is of similar magnlitude to the pay range across all experience groups in the self-managed Yugoslav firm, and U.S. state govern- ments have relatively compressed pay at the upper range of skills (see Vodopivec 1992 on Yugoslavia and Orazem, Mattila, and Weikum 1992, table 3, on U.S. state governments). In the same Yugoslav firm, a worker with twenty years of job tenure earned just 6.2 percent more than a worker with no job tenure. In contrast, Topel (1991) found that a U.S. worker with twenty years of job tenure was paid almost 34 percenit more than a novice worker. As in other socialist economies, Yugoslav workers had strong job security. In fact, job security in Yugoslavia was constitutionally guaranteed. Except for ex- tremely rare cases of bankruptcy, workers could he fired only for breaching work discipline or refusing job reassignment. Even such dismissals were hin- dered by a judicial system that clearly sided with workers. Moreover, to prevent uniemploymenit, governiment pressed firms to hire. Unlike in the rest of Yugoslavia, where hiring was mandated by law, only informal hiring pressure was applied to Slovenian firms. Similar to other socialist economies and in contrast to the rest of Yugoslavia, unemploymllenlt was negligible in Slovenia until the late 1980s. Labor Market Reforms during the Transition in Slovenia While Slovenia was still a constituent republic, Yugoslavia made a decisive step toward the creation of a market economy in late 1988. The Law on Enter- prises transferred decisionmaking rights from workers to equitv owners, thus formally ending the era of self-management. Implementation of privatization finally was passed in November 1992 after a two-year delay (see Pleskovic and Sachs 1994). The l aw of Enterprises caused importanit changes both in wage and employment policies, as well as in labor market programs. Maany measures adopted in Slovenia are strikingly similar to measures that have been adopted in other former socialist countries undergoing transition to a market economy. WVage settinzg. The old wage-setting mechanism was replaced by a system with three components: the Labor Code, collective bargaining, and incomes policy. The Yugoslav Labor Code was instituted in October 1989 and amended by the 206 THE WORI W ) BANK EC:ONOMI(: REVIFW, VOI. 9, NO. ' Slovenian Assembly in February 1991. The Labor Code removed administrative constraints and collective decisionmaking from wage determination, leaving wage determination as a managerial responsibility. This disabled both the gov- ernmental mechanism, which served to equalize pay across establishments, and the worker referendum mechanism, which served to equalize pay within estab- lishments. Managerial discretion to set pay was not absolute, however. The code established a minimum wage, to be adjusted at least biannually to reflect changes in living costs. The law also introduced collective bargaining, a genuinely new component of wage setting. The first general collective agreement for Slovenia was ratified in August 1990. It classified workers into nine categories and prescribed the mini- mum basic wage for each category. The basic wage for the highest category was three times that of the lowest category. However, firms in bad financial standing had the right to reduce the minimum basic wage levels by up to 20 percent. The code also liberalized returns to job tenure. The suggested wage increased at the rate of half a percentage point a year for each year of job tenure. Thus, the suggested return to twenty years of job tenure rose from 6.2 percent to 10 percent, still well below the average return to job tenure in Western economies but much higher than returns under socialism. The wage-setting system continued incomes policies from the previous system. These policies were imposed twice after 1990. The 1990-91 incomes policy law required paying a portion of the wage bill in internal shares of the enterprise. The 1991 law tied growth of the wage bill to growth in the cost of living and called for wages to grow less than the rate of inflation if prices rose faster than 5 percent a month. It also limited managerial salary to be no greater than fifteen times the minimun wage. Employment policies and practices. In October 1989, legislation gave em- plovers the right to lay off workers. However, significant constraints on firing were still limiting the firms' use of the law. The firm was required to provide twenty-four months' notice of a layoff, place the worker in another firm, retrain the worker with pay, or purchase pension credits to allow the worker to retire early. In February 1991 advance notification was shortened to six months, but the cost of a layoff was still high. Detailed discussion of these and other Slove- niani employment, unemployment compensation, and redundancy policies is contained in Vodopivec and Hribar-Milic (1993). Before April 1992 men qualified for old-age pensions at sixty years of age or forty years of work experience, and women at fifty-five years of age or thirty-five years of work experience. Pensions were set at 85 percent of the pension base. The base was the average of the ten highest annual inflation-adjusted incomes in the pensioner's career. Unlike wages, pensions were fullv indexed to inflation. To reduce the inflow into unemployment, the government in 1990 began to pro- mote early retirements by reimbursing enterprises for a fraction of the costs associated with the purchase of retirement credits for early retirees. Early retire- Orazem and Vodopiv'ec 207 ment at reduced pension levels was made available to men fifty-five years of age and with thirty-five years of work experience, and to women fifty years of age and with thirty years of work experience. As a consequence of both the increase of the flow into retirement and falling employment and production, the burden of pensions has escalated enormously. The share of public pension expenditures in gross domestic product (GDP) sky- rocketed from 9.3 percent in 1989 to 11.1 in 1991 to 13.9 percent in 1993 (Government of Slovenia 1994). Following the surge in retirements, in 1992 a law was enacted to institute a gradual increase in pensionable age to sixty-three for men and fifty-eight for women by 1997. Comparison witb Otber Socialist Economiiies Although Slovenia's wage controls differed from those in other socialist econ- omies, they produced very similar wage distributions. As Slovenia, other socialist economies leveled wages across firms and maintained job and wage security through interfirm redistribution; manipulation of prices, taxes, and subsidies; and concessionary finance. Leveling effects of redistribution have been found, for example, for Hungary (Kornai and Matits 1987) and Poland (Schaffer 1990). In contrast to Slovenia, however, determination of relative wages in other econ- omies was centrally imposed through the so-called tariff system. jobs were clas- sified into different skill grades (the "skill-exertion matrices") to which econ- omywide wage rates were centrally assigned. Because central planners imposed an egalitarian pay structure, both preferences of planniers in other socialist econ- omies and democratic determination of wages in Slovenia contributed to the same outcome-egalitarianism. Indeed, a comparison of earnings distributions for economies in Central Europe shows that Slovenia's net earnings distribution in 1989 did not deviate much from those in other planned economies. In 1989, just before the transition started, Slovenia's earninigs distribution was more egali- tarian than Hungary's but less egalitarian than Czechoslovakia's or Poland's (see Vodopivec 1993). As a key component of the current reforms, other transition economies have also overhauled their labor market legislation. Their measures were similar to Slovenia's: they enacted redundancy legislation and ended the era of virtually complete job security; allowed enterprises to be independent in hiring and wage setting; and introduced unemiiployment insurance and active labor market pro- grams to cope with the growing number of unemployed. In the area of wage setting, the reforms in other economies are quite similar to those in Slovenia. Above all, as in Slovenia, the authority to set wages has been transferred to managers. All other Eastern European countries have also intro- duced incomes policies. These policies-usually' variants of tax-based incomes policies-have typically been in place for short durations and later renegotiated, often in somewhat different form. Collective bargaining has also emerged, with similar confrontation among unions for worker representation, as in Slovenia. 208 111i Wo(!1I 1) BANK FL(ONO NI I( R[VIEW. Vol ". No ' Moreover, minimum wage scales for different categories of labor, such as those in Slovenia, have been agreed on in Bulgaria, Poland, and Romania. In many respects, other economies have chosen less generous and less protec- tive policies than Slovenia has. One example is job security. Czech and Polish workers are given up to three months of advanice notice of layoffs, and Roma- nian workers one month-compared with Slovenia's six months (Burda 1993; Scarpetta, Boeri, and Reutersward 1993); and Earle and Oprescu 1993). Sim- ilarly, although Slovenian workers are entitled to severance pay of one month's wage for each two years of job tenure, workers in most Central and Eastern Europeaii economies receive severance pay of Lip to two months' pay. Moreover, the maximum duration of unemploymenit benefits in all other transition econ- omies is twelve mointhis (only six in the Czech Republic and former Soviet Union economies), conipared withi the maximum dUration of twentv-four months in Slovenia. The pension replacement rate in Slovenia has been the largest among transi- tion economies. In 1992 the average pension was 76 percent of the average wage, compared with 64 percenit in Poland and 50 percent in Bulgaria and Hungarv. Slovenia's public expenditures on pensions as a percentage of GDP have been among the highiest in transitioni economies. Only l'oland, at 14.7 percent of GDP, has a higher pensioni hurden. Governmenr-sponsored early retirement programs, similar to Slovenia's, have been introduced in Poland, Ro mania, and, to a lesser degree, in Hun-ary. 11. DATA S(OURCLS To assess how these myriad changes affected the Slovenia labor market, it is necessary to ha1ve both pre- and postreform data. Our main data base is taken from earnings records of the Slovenian Pension and Invalid Fund (slIF), col- lected for pension benefit calculation. Three additional data sources are used: the work history of employees covered by the social security system, data on unemploymenit spells of registered unlemployed, and the register of enterprises. The data provide a cross-section of employees for each year from 1987 through 1991. rhe SPlF collects data on earnings for all workers who are paying contribu- tions to the fund. As in other transition economies, old-age insurance is manda- tory, so virtually all workers are covered. In our analysis, however, we focus only on wages paid in social enterprises, and thus we exclude employers who are able to "adjust" reported earnings of their employees. The main two groups that are excluded from the data set are the self-employed and workers in private enter- prises. Self-employed workers may opt either to underreport their earnings so as to reduce their old-age contribution or to overreport their earnings so as to increase the pension level. Similar opportullities to adjust reported income mav exist in private enterprises. Moreover, in oui data set only jobs for which contri- butions to social security are paid are included. In the case of multiple job ()razcm 71d VlodopivZec 209 holders, moonlighting or secondary jobs rarely pay into the social security fund and are thus excluded. For such multiple jobholders, our measure of the wage is the wage in the primary job alone. The data set does not have information on benefits, but only on wages. To the extent that many benefits-health insurance, pensions, and unemployment insurance-are universal, they tend to reduce the variance in compensation in relation to the variance in wages. Fixed benefits would not affect the marginal returns to schooling, experience, and gender that underlie the other analysis, however. Nor is it clear that the inclusioni of benefits would reduce measured inequality. Deregulation of compensation may have allowed firms to increase the variety and amount of perquisites offered to employees, particularly to those at the upper bound of allowable wages. Observations on work spells were drawn from a random sample of employees representing about 5 percent of the social sector. The social sector covered 92.6 percent of nonagricultural employment in 1987 and 89.8 percent in 1991. Of the 0.028 decline in the social sector share, 0.016 is attributable to lost employment in the social sector and 0.012 to increased employment in the private sector. Usable observations on work spells ranged from 30,474 in 1987 to 21,198 in 1991. The decline in usable observations was due primarily to the 15 percent decline in social sector employment from 1987 through 1991. A lag in SPIF resolution of irregularities in data provided by firms contributed to a smaller usable sample in 1991. As the measurement errors are apparently random, the 1991 usable sample should still be representative. D)etails on sample selection and data base construction are reported in Orazem and Vodopivec (1994). The earnings data include information on earnings, regular and overtime hours, and the starting and ending date of the employment spell within the year. Hourlv wage is computed as earnings divided bvy hours. For workers who switch employers during the year, hourly wage is computed separately for each employ- ment spell. 111. EFFECTS OF EDUCATION, EXPERIENCE, AND GENDER ON TRANSITION WAGES To summarize how the earnings structure in Slovenia changed during transi- tion, we first apply the standard earnings function approach pioneered by Mincer (1974) to the Slovenia data described above. The estimates yield infor- mation on returns to education, experience, and gender. We then report how wage inequality within and between skill groups changed durinig transition. The dependent variable is average hourly earnings over an employmenit spell. The vector of independent variables includes a set of dummy variables indicating different levels of formal educationi, years of employment, non-Slovenian eth- nicity, temporary or internship position, sector of employment, and the monthis in which the individual worked. Given that wages are computed over the employmenit spell, monthly dummy variables are used to conitrol for chianges in conisumer prices over the spell. 210 I1HF WORLI) RANK H ONOMR R1 VIEW, VOL. 9', NO. 2 Because of inflation, nominal wages and prices in the first half of the year differed significantly from nominal wages and prices in the last half of the year. This problem was particularly acute during the hyperinflation of the last quarter of 1989. The use of monthly dummies controls for within-year inflation, allow- ing the coefficients of the earnings function to be interpreted in real terms. The wage equations will be used to explore four issues related to the effects of economic transition on the labor market: first, the effect of transition on returns to education; second, the effect on returns to job experience; third, whether wage differentials between men and women narrowed or increased as a result of the economic transition; and fourth, how earnings inequality changed in general and when controlling for levels of human capital. As we will argue in more detail below, the observed changes in wage structure for Slovenia are quite similar to those observed in market-oriented economies over the past twenty years. In addition, as in the Western economies, the main causes for observed changes appear to be associated with shifts in labor demand. The difference between Slovenia and the Western economies is that the changes in wage structure oc- curred much more rapidly in Slovenia and seem to be related to Slovenia's transition to a market economy. The wage fLnctions for men and women (designated by subscripts M and F, respectively) in year t can be written as (1) WMt = XMAf3f11 + eMT W = XFtJ3P, + eF,t where W is the natural logarithm of average hourly earnings over an employ- ment spell, X is the vector of indepenident variables, and ei, is an error term. Changes in the earnings structure over time are measured by changes in the coefficients, ,B,,. The joint restriction that over two periods t and t', f3j, = f3, can be tested to establish whether changes in the earnings structure are statistically significant. Such tests can also be performed for subsets of the coefficients that are of particular interest, namiiely, those on the education and experience vari- ables. Estimates of the coefficients and the associated tests of structural change are reported in table 1. The regressions reveal dramatic changes in the structure of earnings. The overall explanatory power of the regression falls as wage setting becomes less standardized across firms and sectors. Individual coefficients change dramati- cally, and the general pattern reveals a sharp increase in returns to human capi- tal. There are no significant changes in returns to ethnicity or term of employ- ment. Nevertheless, the null hypothesis that the eleven reported coefficients would be unchaniged from 1987 to 1991 is easily rejected in both the male and female wage equations. In the United States and Western Europe, the intersec- toral pattern of relative wages has been nearly constant over time (Krueger and Summers 1987). By contrast, in Slovenia there are sharp changes in the coeffi- cients on sector dummy variables. These changes are identical in sign and similar in magnitUde across the male and female equations. Holding human capital Orazem and Vodopivec 211 Table 1. Estimation of the Wage Funtctionz for Men anzd Women, Sloven7ia, 1987 and 1991 Men Women Va~riabl1e 1987 1991 (hange 1987 1991 Change Ediucatiou1' Elementary .044 .107 .063 .079 .112 .033 (..28) (6.17) (1.641 (10.10) (6.16) (1.87) Vocational .163 .201 .()38 .164 .183 .018 (23.5) 113.8) (2.58) (19.91 (9.60) (l1.98) Middle school .319 .406 .i)87 .370 .465 .095 (40.0) 24.6) 05.22) (44.9) (24.9) (5.16) University (2 years) .520 .67. .1 56 .569 .685 .116 (43.3) (27.7) (6.28) (50.9) (28.5) (4.80) University (4 vears) .715 .943 .228 .768 .940 .171 (61.8) (41.5) 19.75) (.59.3) (35.2) (6.26) Experience .(19 .018 -.((i .019 .l)1 -.008 (22.2) (1().2) 1.571 (19.6) (5.14) (3.60) Experience2/100 -.027 -.019 .(7 -.(019 .012 .032 (11.3) (3.35) (1.35) (6.55) (1.57) (4.31) Non-Slovenian .023 .(003 -.2)70 -.0)4 -.019 -.014 (3.41) ).24) (1.43) (0.55) (1.17) (0.89) Fixed term - .038 -.003 .008 .007 - .017 -.023 (1.48) ) 1.1.3) (.20) (0.37) (6.69) (0.74) Permanient internship -.035 -.178 - .144 -.05,7 -.0 57 -.000 (.40)) (2.09) (1. (I) (0.66) (0.69) (0.00) Fixed-term internship -.149 -.197 - .0)47 -.111 -.2(18 -.096 (-3.93) (6.05) (.X5) (2.58) (5.88) (1.46) Month dummies inc. inc. n.j. imc. inc. n.a. Industry dummies Inc. inc. II. inc. inc. n.a. Sample size 15,884 10,822 n .,. 14,590) 10,376 n.a. RI '.428 .342 n.l. .46 1 .313 n.a. F) 1, n)b n.a. 11.6 n.a. n.a. 8.74 n.a. na. Not applicahle. inc. Included. Note: t-statistics are in parentheses. a. Workers with less than an elemenitarv education are not included. h. Test of the null hypothesis that the human capital, ethnicity, and terimi of employment coefficients are equal in 1987 and 1991. i represenits the degrees of freedom in the Unconstrained equation. 5ousrce: Autho-s' calcul,itioins. measures fixed, relative wages have risen in agriculture, services, health, and government, but have fallen in construction and educationl. It must be emphasized that the period uLider consideration is one of declininig earnings, and that the dynamics of returns to human capital and gender may change when earnings start to rise. For example, it is conceivable that better indexation of minimum wages in the period of increasing earnings reverses the trend of increasing returns to human capital. Larger profits, however, may in- duce firms to further increase the rewards for the most productive, skilled labor and thus contribute to furthier accentuationi of wage inequality. Retuirnis to Education The coefficients on the education dummy variables tell a very consistent story. Average returns to years of education have risen dramatically following transi- 212 TIIF WORI L) BANK I t ONO(IK RFVIEW, VOlI. , NO. 2 Figure 1. Inde-x of Relative Wages by Education, Slovenia, 1987-91 Index ( 1987 100) 130- 4 years of universitv ,' 125 - 120 - , 2 vears of 115- ,' ~~~~~~~~~~~~~~~~~universihv 115 105- , Mcational school ._ -*E lementary school 100- 1987 1988 1989 1990 1991 Note: Values are reltrive io values for workers with less than in elementary education. All values Ire for full-time. vear-round workers. The sample sizes are 28,176 for 1987. 29,613 for 1988 28 188 for 1989, 25.432 tor 1990 and 19,880 for 1991. Source. Authors CACUlations base(d on unpublishiedi coluntry dlatal. tion in comparison with earnings of the least-educated group. The changes in relative returns to education are virtually identical for men and women; individ- uals with four years of university education gained the most in relative earnings, followed by those with two years of university training. These findings nicely complement those of Abraham and Vodopivec (1993) on changes in workers' ability to switch jobs, as well as to find a job if unemployed, during Slovenia's transition. They also find that the relative advantage enjoyed by highly educated workers has strongly increased. The educated group that gained the least in comparison with the least educated during this period were holders of voca- tional degrees. This finding is consistent with Flanagan's (1993) argument that vocational training was overemphasized in the controlled economies of Central and Eastern Europe. It is important to emphasize that the gains reported here are not absolute but relative gains. As shown below, the increase in relative returns to schooling occurs because the most educated faced the lowest proportional decline in real wages. Figure I charts the relative wages of full-time, year-round workers by educa- tional group, using those with less than an elementary degree as the base. Part- year workers were paid very different nominal wages, depending on the timing of their employment within the year; using the sample of year-round workers removes the artificial increase in wage variation caused by inflation. In figure 1 Oraze,n and Vodopivec 213 each ratio is normalized to one in 1987 so that changes in the height of the ratio can be interpreted as percentage changes in relative earnings for the group. The figure shows that relative earnings for the most educated were rising slightly in 1988 but then changed dramatically after the transition began in 1989. Those with four-year universitv degrees gained 27 percent over those with less than an elementary education. Equally remarkable is that the proportional changes in relative earnings in Slovenia over this period are larger than the dramatic increases in relative re- turns to college graduates in the United States observed over the same period. Using Current Population Survey (cps) data for the United States durilng 1984-91, relative annual earnings for those with sixteen years of education rose 25 percent in comparison with those of elementary school graduates. In fact, real salaries for U.S. college graduates were declining over the period, but salaries for the elementary school graduates were declining even more. More detailed analv- sis of changes in inequality of earnings across skill groups in the United States are in Juhn, Murphy, and Pierce (1993) and Katz and Murphy (1992). Levy and Murnane (1992) provide an extensive review of the topic. Retuirns to Experience Consistent with the results on education, returns to the most experienced rise in comparison with returns to the least experienced. As with returns to educa- tion, returns to experience change in similar ways for men and women. The linear term becomes less positive and the quadratic term becomes less negative, meaning that marginal returns to a year of experience fall for the least experi- enced but rise for the most experienced. In fact, the wage-experience profile for women turns convex in 1991. This outcome is not just a fiction of the quadratic approximation. In figure 2, the coefficients on a series of dummy variables representing progressive four-year experience increments pooled across men and women reveal the same pattern. The wage-experience profile in 1991 is flatter than in 1987 for the first eight years of experience but steeper thereafter. In 1991 there is a dramatic increase in returns to experience beginning at twenty-eight years. This sharp increase in returns may reflect the general increase in relative returns to human capital discussed thus far, but additional reasons related to pension policy may also have an effect. These reasons will be discussed in more detail in the next section. Male-Female Earnings Differentials Compared with women in Western economies, women in Slovenia had high labor force participation rates and high relative wages. In 1987 the female labor force participation rate was 0.75, compared with 0.54 in Austria, 0.55 in West Germany, and 0.68 in the United States. Women in Slovenia were paid 88 per- cent of what men were paid, a much higher ratio than in Austria (0.73), West Germany (0.69), or the United States (0.68). Fong and Paul (1992) report that relative wages for women in Central and Eastern Europe were comparable to 214 TIHE WO(RIl1) ANK F( ONONM F(V1FWV0V. VOL '. Nil . Figure 2. Returns to Years of Fxperience, Slouenia, 1987 and 1991 Change in real wages (percentage) o.6 - 0.5- 1991 0.4- 0.3 - 0.3- 0.1 - 0) 4 .. 12.5 17.5 22.5 27.5 32.5 37.5 Years of experience Note. Values are calculated holding fixed education, ethnicity, gender, and intern status. Sample sizes are 30,473 for 1987 and 21,19-7 for 1991. Source. Authors calculations hased on unpublished c(lUntry data. those in Western Europe. Relative wages for Yugoslav women were higher than those for women in all other European countries included in their study. Given that women did well under socialism in Slovenia, it may seem natural to presume that women could only do worse under a market system. This presumption will be examined in this section. Table 2 reports estimates of the differences in coefficients between the male and female wage equations for each year from 1987 to 1991. Positive differences imply that the coefficient in the female equation is greater than the coefficient in the male equation. The results indicate that womeni have had higher marginal returns to education and steeper returns to experience than men. After the tran- sition, the male and female earnings structures became much more similar. The only exception is that women's wage experience profiles became steeper. The F-statistic of the null hypothesis that the male and female earnings functions have identical coefficients falls monotonically from 11.3 in 1987 to 4.2 in 1991. The F-statistic on the null hypothesis that the eleven human capital, ethnicity, and term of emplovment coefficients are the same falls monotonically from 8.2 to 3.2. The implication is that transition led to lower differences in the pricing of male and female characteristics, an outcome consistent with theories of how increased labor market mobility and market competition would affect wages. (Drazei nand Vodopivcc 21l5 Table 2. Differences in Returns to Characteristics betwveen Men and Women, Slovenia, 1987-91 Variable 1987 198S 1989 1990 1991 Educationt Elementary .035 .013 .007 .020 .005 (3.09) (1.16) (0.41) (1.02) (0.19) Vocational .()01 -.014 -.00 -.012 -.018 (0.l1) (1.39) (0.3() (0.63) (0.77) Middle school .l)51 .032 .i26 .018 .059 (4.43) (2.89) (1.44) (0.94) (2.36) Universitv (2 years) .048 .(I5 .(19 .024 .008 (2.93) (0.96) (().73) 0(.87) (0.23) University (4 years) .054 .057 .055 .006 -.004 (3.08) (3.49) (2.05) (0.21) (0.10) Experience -.000 -.001 .006 -.006 -.007 (0.06) (0.8 S) (3.22) (2.81) (2.44) Experience2/01 0 .007 .(I I -.004 .031 .032 (1.98) (3.0 1) i0.61) (4.54) (3.27) Non-Sloveniani -.02 8 -.o 11 .01)3 .007 -.022 (2.71) (). 6) ((.19) (0.4,3) (1.0) Fixed term .045 .057 .045 .117 .013 (1.43) (2.22) (1.33) (3.88) (0.37) IPermanent internship -.022 -.0(14 -.026 .141 .121 (0.18) ((.18) (0.26) (1.41) (1.02) Fixed-tern internship .038 -.0(5 -.006 -.0 16 -.011 (0.66) (1.23) (0.13) (0.38) ().2 3) F(11, n)' 8.2 8.8 7.6 6.0 3.2 F1(21, n) 1 11.3 9.7 9.4 7.8 4.2 Note: Positive differences imply that the coefficient in the female equatioin is greater than the coefficient in the male equation. t-statistics are in parentheses. a. Test of the joint hYpothesis that the eleveii coefficients .are equal acros- the niale and female wage equations. b. Test of the joint hypothesis that the cleven coefficienits and the ten industrv dumimiiy variables are equal across the male and female wage equations. Soiurce: Authors' ca(cularions. Following the methodology developed by Juhn, Murphy, and Pierce (1993), the effects of economic transition on earnings differences between males and females can be explored in more detail. As shown in Orazem and Vodopivec (1994), the change in the male-female wage differential between year t and t' is (2) I(XAl,' - XA1,) - (XFt' - XFtH AMtW + (XA1, - XF)(A1, - JAt) + [(O9Mt" - (I) - (QMt - eFd)]qM, + (O.Mt -- /Ft)(gMt,' - UM) where a, is the standard deviation of the residual of the male earnings func- tion in year t, Om, = eM Iat/S is the standardized residual of the regression, and OF, = (WF, - XFlPM)/OM4- The first term in expression 2 captures how the wage gap changes in response to changes in characteristics between men and women. The second term measures how changes in the returns to these characteristics affect the wage gap. The third term represents how the change in women's relative position in the male residual earnings distributioni affects the wage gap. The fourth term shows how increases in the standard deviation of the 216 IHLWORI DR IANK I e(()NO 11 VIt . Vol. . o.12) Table 3. Chuanges in Women's Wages in Comparison uwithA Men's Wages, Sloven ia, 1987-91 St,indardized residual Standardized residual controlling for humoan controlling /or all )bserved Capital/ variables Yeajr Difteren e.' Positionllh Di/ferencee Position"1 Differcnce, Positionil1 1987 -.1. 35 -.4 l -.5(1 26 1988 -.12 36 -.47 3 0..4 2.7 1989 -.13 31 -.38 3 1 -.4.5 28 1990 -.1 1 39 -.26 3 -.34 2 8 1991 -.I( 40 -.25 3.3 -.30 30 a. Estinmated as rhe rnaturiral log of average wziges ft(o- wvonmeln minLus the natLiral log of average wages for mc(li. h. Percentile posi till (of ian r an fermale earnings (wages) in the male earninigs distribuition. c. Average value of 0,, as defined it equiarion 2 in thL text. Source: A uthors' calcu lations. residual earnings distribution affect the wage gap. Blau and Kahn (1994) label these four effects "Observed X's," "Observed Prices," "Gap," and "Unobserved Prices," respectively. Summary statistics for the five years of data are reported in tables 3 and 4. Before the transition, the female-male wage ratio in Slovenia was very high. Nevertheless, women's wages rose in relation to men's wages during transition. The log wage gap fell from 0.13 to 0.10, implying that the female-male ratio rose from 0.88 in 1987 to 0.90 in 1991. In 1987 median female earnings were at the 35tlh percentilc of the male wage distriblution, but they were at the 40th percen- tile in 1991 (table 3). In contrast, the mediani female wage in the United States is at the 31st percentile of the male wage distribution (Blau and Kahn 1994). Examining residual inequality after controlling for human capital, ethnicity, part-time status, and sector of employmenit lowers the relative position of women in the male residual earnings distribution. Although in 1991 women were at the 40th percentile in the observed earnings distribution, they were only at the 30th to 33rd percentile of the residual earnings distribution. Nevertheless, both the observed and the residual wage data indicate that women rose from 4 to 5 percentage points in the male wage distribution over the five-year period span- ning the Slovenian transition. Blau and Kalhn's estimates for the sample of West- ern economies show female percentile status in male residual earnings distribu- tions varying from 0.16 in West Germany to 0.31 in Australia. In Western economies, male advantages in work experience, job tenure, and college training serve to explain some of the pay difference between men and women. In Slovenia, the opposite holds. Women do worse in the residual distri- bution than in the observed distribution. The main reason appears to be that women have superior education. Women are less likely than men to have less than an elementary education or a vocational degree, the two educational cate- gories that have performed worst throughout the transition. Women, however, Orazcom and Vodopiu'ec 217 Table 4. Decomiipositioni of Changes in Women's Waiges in Comparison with Ment's Waiges. Sloveonia, 1987-91 O)bserved chbangct 1987-89 1989-9 1 1987-91 1. Oserl-vedi X's' - .005 - .014 .012 2. Observed prices" - .024 .016 -..014 .3. G a p -A.22 .073 --.1(() 4. Unobserved pricesJ .049 .043 .(97 Toral -..002 --.028 --.030) Geender specific (I + 3) - .026 --.087 -.1 1 3 Wage struICture () + 4) .(_1 S .(S9 .08 3 Explained ( I + ) .()09 .((2 .026 Unexplained (3. + 4) .277 -.0 (0 -.003 Note: Estimlates are based oni .1 tilly specifiedl iVage equ.arioni, ncluding sector dummy variables. Negative numbers indicaite factors thart increase women Is pay ii c omp.arison withi meln's pay. a. How the wage gial changed in response to changes in Jiaractcristics bietweeni iileni anid woni. 1). HI-ow changes in the' returns to Lhara1cteristics aftected the wage gap. Ho w the Ji,.i nge wi in wo 's rclative piosition in the rinale residual earninigs distrhibution affected the wage gaip. d. How iricreases in the st,andard des aontoi ot the residual earninigs dlistributition affected the wage gap. Sourcet: Authors' -SICL`latiiii. were more likely to hold middle school or university degrees, the educational categories that have done best in the transition. In table 4 the clhanges in male-female wage differentials are decomposed into the four comiiponents. Negative numbers indicate factors that increase women's pay in relation to men's pay. Over the five-year period, the first three terms serve to reduce wage differentials between meni and womiien, and the fourth term, which captures the effect of increasing inequality in the residual wage distribu- tion, raises the gender wage gap. Almost all of the changes in the wage gap occur after 1989. Although there are nontrivial values for the last three terms in the 1987-89 period, they cancel each other out. For the full period, the greatest impact occurs through the gap effect, with women gaining by moving up the male residual earnings distribution. The next largest effect is through the unob- served prices effect, from increasing inequality in the residual wage distribution. Because women are in the lower tail of the distribution, they lose in relation to men from the increase in inequality. Much smaller relative gains to women come from narrowing differenices in labor market characteristics and from chalnges in returns to characteristics that favor women more than men. In terms of more traditional wage decompositions, the first two terms sum to explained differences, and the last two sum to unexplained differences in wages between men and women. Based on that decompositionl, 87 percent of the im- provement in women s earnings in relation to men's is attributable to changes in observed characteristics and observed returns to those characteristics. There is evidence of a decline in wage discrimilnation durinlg 1989-91, but the increase in the uniexplained component from 1987 to 1989 is of nearly equal magnitude. 218 THF W'ORI 1) IANK ( Ot)NOIM RFVI FW, %:. 9. No Figure 3. Real Wage Distribution, Sloventia. 1987 and(t 1991 Frequeney (percentage) 9 8 I ' \1987 6 X 5 4 1991 3 I 2- 1 0- 0 500 1,(0( 1,900 2.000 2,5(( 3,000 3,500 4,00() 4,500 Real hourly wage (1987 dinars) Note: Values are for full-time, vear-round workers. The 1991 wages were deflated hv the ratio of median wages in 1991 to median wages in 1987. Sample sizes are 29.613 for 1987 and 19,880 for 1991. Source: Authors' CalCulations based on unpublished country data. Inequalitv To the extent that controlled economies were successful in suppressing in- equality, relaxation of central government controls would be expected to in- crease the dispersion of income in the economy. This expected dispersion has occurred in Slovenia. Figure 3 contains a mapping of the distribution of wages for year-round, full-time workers in 1987 and 1991. To correct for the large changes in currency value, the 1991 wages have been deflated by the ratio of median wages in 1991 to median wages in 1987. This deflation has the effect of forcing the median of the two distributions to be equal, making it easier to visualize changes in the distribution of wages. It is clear that the variance of wages has increased. The distribution is much less peaked in 1991 than in 1987, and the distribution has become further skewed to the left. Moreover, there is a larger number of workers at the upper tail of the distribution as well. Given our earlier results on experience and education, it seems that the upper tail is dispro- portionately populated by those with greater skills. Not only has the distribution become more unequal, but the gap between the richest and the poorest has increased. Figure 4 shows the percentage change in Orazemn and Vodopivec 219 Figure 4. Change in Real Wages, Slovenia, 1987-91 Change in real wages (percentage) -0.2- o.4 -~~ 0,6 -~~ -0.6- I I III I I II 0 10 20 30 40 50 60 70 80 90 100 Real wage percentile Note: Sample sizes are 28,176 for 1987 and 19,880 for 1991. Source: Authors' calculations based on unpublishecd country data. real wages from 1987 to 1991 by percentile in the wage distribution. All percent- age changes are negative, indicating that real wages declined at all points in the wage distribution. However, the largest percentage reductions are for those at the bottom of the wage distribution. As percentile position in the wage distribu- tion increases, the percentage wage reduction decreases monotonically. By 1991, wages at the 10th percentile were 56 percent of the wage earned by those at the 10th percentile in 1987. In contrast, those at the median in 1991 earned 64 percent of the wage earned by the median workers in 1987, and those at the 90th percentile earned 70 percent of the wage earned by those at the 90th percentile in 1987. Wage earners in the upper tail gained relatively because they lost less in relation to 1987. An issue that has been studied extensively in the United States and in Western Europe is the rising inequality within narrowly defined skill groups as well as between these groups. Coefficients of variation for education groups, experience groups, and gender in Slovenia are reported in table 5. These statistics are computed over a sample of full-time, year-round workers. The results show increases in the variance of earnings for the least- and most-educated groups and for the most-experienced groups. Measures of rising inequality were nearly iden- tical for men and women. However, there was no evidence of rising inequality 220 1 l IIWoR[ I) BANK I-(fONOM I(1 RFVIFW, Vo[ (. N(). ' Table 5. Coefficienits of Variation of Real WVage by, Subgrou-p, Sloveenia, 1987-91 Subgroup 1987 1989 1991 Education Less than elemeitary .36 .44 .43 Eleinentary .51 .49 .s/ Vocational .48 .47 .38 Middle schoiol .43 .47 .42 Universirv (2) .37 .39 .47 University (4) .28 .40 .45 Expericnce 0-S years .61 _53 .7 I 1-2() vears .52 .56 .52 >2(1 years .42 .48 .56 (,ezder Mell .49 .57 .55 NY/omen .49 .5( .56 Rcsidual standard de,ization Male, a, 30 .40 .49 Femlale, , .29 .47 .5 1 Note: The statistics are comilIpLted over a sample of tLIll-time, vear-routnd work-ers. Source: AuLthor,s' Calcla=tionent by Gender Women gained in relative pay and relative employment. In table 6 it seems clear that the relative gains in women's employment were due to women's rela- tive concentration in sectors that were less adversely affected hy the transition. Men were concentrated in manufacturing, agriculture, construction, transporta- tion, communicationi, and services, all of which lost employment more than average. Women were concentrated in sectors that did better. Within sectors, women were not more likely to retain emplovment than men were. In fact, male employment shares rose in seven of the eleven sectors, although most changes were small. These results suggest that the gains made by women were due to sector-specific demand-side factors that adversely affected sectors with predomi- nantly male workers. Selection The most serious shortcoming of our analvsis is the lack of information on those exiting the social sector. For example, to the extent that movement into the private sector is a result of individual choice based on expected earnings in the social and private sectors, the estimates of returns are biased. To examine the extent of the bias, a longitudinal sample of workers employed in 1990 was drawn. By 1991 many of these individuals had no reported wages, some having retired, others having become unemployed, and still others having moved either into the private sector or out of the labor force. We then estimated equations of the form (3) In (W9i/W9,) = X9(!J + p ±AK + P2'AZ + PAR,, + e where W911W9() is the ratio of the wage in 1991 and the wage in 1990, X9( is the vector of human capital and sector variables for 1990 used in the analvsis above, (OIra,cem and Vlodopnvec 227 and AR, i,,, and A,, are estimates of the instantaneous probability of exiting the social sector for retirement, unemployment, and other reasons, respectively. Fol- lowing Heckman (19'79), the A terms correct for potential selection bias. The outcomes were generally consistent with those reported above. Those individuals with twelve years or less of education lost real wages, and those with more than twelve years gained. The largest gains were for the most educated. Women gained in relation to men. However, changes in returns to experience were insignificant, albeit showing patterns similar to those already reported. The implication is that the increase in returins to education and gender we have found is not driven by selection, hLut that selection plavs the dominant role in changing returins to experienice. I V. WAGE AND EMPLOYMENT CHANGES ANt) LABOR MARKET l)OLICY The disruption of what had been a very stable economic system created large shifts in labor demand across sectors, skills, experience groups, and geographic areas in the formerlv socialist econiomies. These large shifts occurred over a very short period. The fortunate people were occupying sectors of the earlier system that faced smaller disruptions in labor demand. Workers in those sectors re- ceived a relative quasi-renit from the positions they occupied. The unfortunate, those in sectors that shrank or collapsed, faced a relative loss. Identifying the relative winners and losers durinlg transitioni will aid in the shift of labor toward its most productive uses, those sectors in whicih relative quasi-rents have risen. The stylized facts regardinig changes in wages and employment during the Slovenian transition to a more market-oriented economy can be summarized very briefly as follows: * Relative wages and employment rose for the most educated and fell for the least educated. The apparent shift in relative labor demand toward the most educated occurred in all sectors. * Relative wages and employment rose with years of work experience until pensionable age. These results are consistenit with shifts in relative labor demand toward more-experienced workers. * At pensionable age, relative wages increased very rapidly and relative em- ployment plummeted. The effect is consistent with a labor supply shock for workers of pensionable age. * Women gained over men in both wages and emplovment. Relative returns to women's characteristics became m)ore similar to men's returns to those char- acteristics. Women's gains are attributable to the fact that women were more educated and occupied in sectors that were treated more favorably by the transition, not to econo)mywide reductions in discrimination against women. 1. We have begunl setting up a inure conmplete longitudinal data base on earnings for Slovenian workers thar will enahle a more rhoroigh ex ploraion (if the role ot labor market transitiotis on wages 228 1 li- WORI I) VANK I S(SNOMIU R-V11`W, VO)l 9, \NO * Wage inequality increased. Wage variation increased between skill groups, within skill groups, and within groups with identical sectoral and human capital characteristics. The litmus test for the impact of labor market policy is whether the policy changes are consistent with these labor market outcomes. In Slovenia, the policy story seems clear. Disabling the tax or transfer policy from relatively profitable to relatively unprofitable firms and eliminating worker referenda on wage scales removed mechanisms that served to compress wage variation. As a result, re- turns to human capital rose rapidly, both for education and for experience. Additionally, relative labor demand grew for the more skilled, both because of economywide factors, such as the need for more huLmllan capital in order to cope with uncertainty, and because of sector-specific factors that lowered demand, particularly in low-skill-intensive sectors. It remains to be seen whether these increases in inequality will cause long-termn adjustments in supply toward high- wage sectors, or if they will represent a permanent shift toward a less egalitarian wage structure, with the least skilled being the permanent losers. Policies that actively encouraged retirements were tremendously effective in lowering employment for womeni over fifty and men over fifty-five. Such policies were more successfUl at reducing employment than was legalization of layoffs, because (in the period studied) costs assessed to firms for layoffs were severe. The policy was so effective that it may hiave caused firms to bid tIp wages for workers of pensioniable age to prevent them from retiring. Alternatively, it may have caused a selection process by which oinly the highest paid workers of pensionable age remained in the labor force. Preliminary exploration is Conl- sistent with the selection viewpoint. In either case the outcomes suggest that the pension policy has proven very costly for Slovenia, both because of the drain on GDP needed to meet pension obligations and because of lost productioll from retirees. At the same time, the retirements did not "make room" for young workers: employment shares for the least experienced fell, and employment shares rose for those who were the closest substitutes to the retirees-those just under pensionable age. Incomes policies that set minimum wages, fixed ranges of pay, and partially indexed wages to inflation did not prevent increases in wage variation from occurrinig. Wage minimums did not appear to have an effect, presumably be- cause inflation reduced real minimum wages so quickly that most workers were paid above the minimum. In fact, wage distributions shoowed no signs of massing at the lower tail, indicating no evidence that minimum wages were effective.2 Partial indexation did not work because inflationl was frequently above the maximum required adjustment. Since 1991 inflationi has been more or less con- trolled, so these incomes policies may have become more effective in recent years. 2. These wage ListribuLtions are available tromi the authors oin requlest. Orazein and Vodopivec 229 REFERENCES The word "processed" describes informally reproduced works that may not he com- monly available through library systems. Abraham, Katharine, and Milan Vodopivec. 1993. "Slovenia: A Study of Labor Mvarket Transitions." World Bank, Policy Research Department, Washington, D.C. Processed. Bird, Edward J., Johannes Schwvarze, and (Gert G. Wiagner. 1994. "Wage Effects of the Move Toward Free MIarkets in East C,ermany." Industrial an11d Labor Relations Review 47(3, April):390-400. Blau, Francine D., and Lawrence M. Kahni. 1994. "Rising Wage Inequality and the U.S. Gender Gap." Amneric-an EconomniC Revieu, Papers and Proceedings 84(May):23-28. Brown, Aninette N. 1992. "Private Sector 'Wage Differentials in Estonia." University of Michigan, Departmeint of Econiomics, Ann Arbor. PJrocessed. BUrda, Michael C. 1993. "Unemployment, Labour Markets and Structural Change in Eastern Europe: A ELiropean Forum." Economic l'olicy 8(1, April):101-28. Earle, John E., and Gheorghe Oprescu. 1993. "Employment and Wage Determination, Unemiiploymenit, and Labor Policies in Romania1." Paper presented at the World Bank conferenice, Unemploymenit, Restructuring, and the Labor Market in East Europe and Russia, Washington, D.C., October. Flanagani, Robert J. 1993. "Were Coommunists Good Huilani Capitalists? The Case of the C'zech Republic." Stanford University, Graduiate School of Business, Stanford, Calif. Processed. Fonig, Monica S., and Gilliani Paul. 1992. "The Chaniging Role of WVomen in Employment in Eastern Europe." Report 8213. World Bank, Europe and Central Asia, Middle East anLId North Africa Regions Techinical Department, Washingtoll, D.C. Processed. (Government of Slovenia. 1994. Statistical Yearbook of Slovenia. 1994. Ljubliana: Statisti- cal Office. Heckman, James J. 1979. "Sanple Selection Bias as a Specification Error." Econiometrica 47( 1, January): 153-61. Juhn, Chinhui, Kevini M. Murphy, and B3rooks Pierce. 1993. "Wage Inequality and the Rise in Returns to Skill." lournal of Political Economy 101(3, June):410-42. Katz, Lawrence, and Kevin M. Murphy. 1992. "Changes in Relative Wages, 1963-1987: Supply and Demanld Factors." Quarterly ]ournal of- Economics 107(1, February):35- 78. Knight, John, and Song Lina. 1993. "Workers in China's Rural Industries." In Keith Griffin and Zhao Renvwel, eds., The Distribution of Income in China. Londoni: St. .Martin's Press. Kornai, janus, and Agnes Matits. 1987. "The Softnless of Budgetary Constraints: An Analysis of Enterprise l)ata." ELastern. European Economics 25(4): 1-34. Krueger, Alan B., and Jorn-Steffeni Pischke. 1992. "A Comparative Analvsis of East and 'West German Labor Markets: Before and After Unificationi." Working Paper 307. Princeton University, Industrial Relations Section, Princeton, N.J. Processed. Krueger, Alan B., and Lawrence H. Summers. 1987. "Reflections on the Inter-Industry Wage Structure." In Kevin Lang and Jonathani S. Leonard, eds., Unemployment and the Strubctufre of Labor Markets. New York: Basil Blackwell. Levy, Frank, and Richard J. Murnane. 1992. "U.S. Earnings Levels and Earnings Inequal- itv: A Reviexv of Recent Trends and Proposed E'xplanations." journal of Econlomilic Literature 30)(September): 1333-8 1. 230 rYiii W(5RI I) BANK l` ONOMI(: RV OI>'. \(O 9, No. ' Mincer, Jacob. 1974. Schooling, Experience, and Earnings. Studies in Human Behavior and Social Institutions No. 2. N.Y.: Columbia University Press. Orazem, Peter F., J. Peter Mattila, and Sherry K. Weikum. 1992. "Comparable Worth and Factor Point Pay Analysis in State Government." Induistrial Relationis 31(1, Winter):1 95-228. Orazem, Peter F., and Milan Vodopivec. 1994. "Winners and Losers in Transition: Re- turns to Education, Experience, and Gender in Slovenia." Policy Research Working Paper 1342. World Bank, Policy Research Department, Washington, D.C. Processed. Pleskovic, Boris, and jeffrev Sachs. 1994. "Political Indepenldence and Economic Reform in Slovenia." In Olivier Blanchard, Kenieth A. Froot, and Jeffrey D. Sachs, eds.. Transition in Ecasternz Euirope. Country Studies (1). Chicago: ULniversity of Chicago Press. Redor, Dominique. 1992. Wage Inequalities in Fast aindc WVest. New York: Cambridge University Press. Scarpetta, Stefano, Tito Boeri, and Ander-s Reutersward. 1993. "Unemiiployment Benefit Systems and Active Labour Market Policies in Central and Eastern Europe: An Over- view." Organization for Economic Cooperation anLd Developnient and Center for Co- operation with Econornies in Transition (OECD-CC:. r), Directorate for Educationi, Em- ployment, Labour, and Social Affairs, Paris. Processed. Schaffer, Mark E. 1990. "How Polish Enterprises Are Subsidized." University of StIssex, School of ELuropeani Studies, Falmer, Brighton., 1.K. Processed. Schultz, Theodore W. 197.5. "The Value of the Ability to Deal with Disequilibria." louir- nial of Economic Literature 13(3, September):825-46. Topel, Robert. 1991. "Specific Capital, Mobility, and Wages: Wages Rise with Job Se- niority." journal of Political Economy 99(FebrLuarv): 14.5-76. Vodopivec, Milan. 1992. "The Effects of Deilmocratic Deteriniiationi of Wages: Theory and Evidence from Self-Managed Firms." wi's 971. World Bank, Country Econoinics Department, Washi ngton, D.C. Processed. - . 1993. "Determination of Earinigs in Yugoslav Firms: Can It Be Squared with Labor Managemient?" Economic Development and Cultural Change 41(3, April):623- 32. Vodopivec, Milan, and Samo Hribar-Milic. 1993. The Slovenian l abor Niarket in Transi- tion: Issties and Lessonis Learned. wps 1162. World Bank, P'olicy Research Department, Washington, D.C. Processed. T H i- ix} R 1. 1) B A N K I ( i) N o M I L R I V I E W \ O v L 9 , NO 2: 2 3 1 -2 8 An Eclectic Approach to Estimating the Determinants of Achievement in Jamaican Primary Education Paul Glewwe, Margaret Grosh, Hanan Jacoby, and Marlaine Lockheed This article estimates the determinants ot cognitive skills in Jamaican prinmary eduica- tion. W/e take an eclectic approach, integrating the production funizctionz framework favored by economists with the concerns of educators about pedagogical processes and those of sociologists regarding school organization and management. At the same time, wve correct for selectivity biases induced by school choice. We iuse an unufsually rich data set, the 1990 Jamaican So rv'ey of Living Conzditiozs, wehich includes not only scores on cognitive achievement tests but also deta7ied information on each child's household and the primary school he or she attends. WVe find that all three components-physical and pedagogical inputs, pedagogical practices, and school organization2 and climate- influence student achievement. Our policy siniuilations suiggest that a focus on inputs alone miay he misplaced in school systems with input levels as high as those found in Jamaica: school reforms that concentrate on just a fit'w pedagogical practices cozuld lead to substantial inprovemcents in student achievement. Primary education is a cornerstone of social development and a principal means of improving the welfare of individuals. It promotes economic growth-thereby reducing poverty-and enhances political, economic, and scientific institutions. Primary education also ameliorates the health and nutritional consequences of poverty (World Bank 1990, 1993). The benefits of primary education for devel- opment stem largely from the iniproved cognitive skills it imparts: literacy, nu- meracy, and problem-solving ability. Yet, a large body of evidence suggests that primary educationi in many developing countries is inefficiently organized and that school-level implementation of effective strategies is particularly weak Patil Glewwe and Margaret Grosh are with the Policy Research Department at the World Bank, Hanan Jacoby is with the Department of Economilics at the University of Rochester, and Marlaine Lockheed is with the Educationi and Social Policy Departnieit at the World Bank. Financial support was provided by the World Bank Research Committee I Ri'o 676-87). The Jamaican Ministrm of Education collaborated exten- sively in designinig the survey iistrunments, implementing the surveys, and providing access to adimiinistra- tive records. The Planning Institute of Jamaica provided further support, and the Statistical Institute of Jamaica carried out the extremely demandinig survey work. The authors would also like to recognize the many contributionIs of their late lamnaican collaborator, Dr. Derek Gordon. Henri Jeancard and Qlighua Zhao provided com1pLItatiolnal assistancC. C© 1995 The International BaLiik for Reconstruction and Developimienit /THE WVORLD BANK 231 2.32 THF I 'ORI I IBANK F( ()O)NM Il IVIFWX'. VOL. , N(1. 2 (Lockheed and Verspoor 1991). Thus, a key issue in education and development policy is how priniarv schools can more (cost-) effectively promote the acquisi- tion of cognitive skills. In this article we estimate the determinants of cognitive skills in Jamaican primary schools and use these estimates to evaluate alternative efficiency- enhancinig policies. Our approach is eclectic, going beyond the production func- tion approach favored by economists (for example, Hanushek 1986) that focuses largely on inputs. We examine pedagogical processes, which are of concern to educators (for example, Wittrock 1986), as well as the organizational and man- agement characteristics considered by sociologists and political scientists (for example, Coleman, Hoffer, and Kilgore 1982; Chubb and Moe 1990). This is not to say that economists are uliconcerned with anything but educational inputs; Harbisoni and Hanushek (1992), for example, examinle the impact of teaching practices in Brazil. Rather, one could interpr-et our analysis as an "expanded productioll function" approach that examines many more characteristics of schools and teachers than are usually conisidered in the standard function ap- proach. At the same time, because we have detailed household-level data, we are able to control for differences in home background as well as for selectivity biases induced by school choice. Concer-ni with such selectivity issues-a main- stay in empirical ec!)nomics-has been conIspicuously absent from most studies of school effects (for exceptions, see Glewwe and Jacobv 1994; Jimenez and Lockheed 1994; and Harbison and Hanushek 1992). Section I puts our approach in perspective by reviewing the literature on school effectiveness and the determinanits of cogiitive skills. Section ll describes the education systemil in Jamaica and sets the context for the present study. Section III describes the data, and section IV discusses the selectivity issue and the statistical method used to deal withi it. The estimation results are presented in section V, and our policy simulationis are performed in section VI. Section VII summarizes the article. 1. SCHOOL EFFECIIVENESs: THE ISSUES A major debate in the literature on schiool effectiveness (see Purkev and Smith 1983) concerns whilch school-level factors-inputs, teaching, and management- best promote cognitive skill formation, and which do so most cost-effectively. This question has implications for educational budget allocations, specifically for whether governimients should provide schools with a particularly effective set of inputs (for example, school facilities and instructional materials); concentrate resources on improving pedagogical processes (for example, through teacher traininig and perforimianice evaluation); or reform the management of the entire school (for example, through decentralization, administrator training, and school supervision). In this section, we review the current state of knowledge on the influence of schiool-level physical and pedaigogical inputs, pedagogical pro- cesses, and organizationi, climate, and control on the process of learninig in developing coun1tries. Glewte, nd otbcrs 233 Internationally, there is ample evidence that primary schools are differentially effective in producing learning (Creemers, Peters, and Reynolds 1989). Yet, re- cent reviews of the impact of physical and pedagogical inputs on student achievement in the lJnited States have found that variations in these inputs are rarelv related to variationis in student achievement (Hanushek 1986). However, in developing countries some physical and pedagogical inputs have been found to be related to studenit achievemeit. Two recent reviews of this literature are summarized in table 1. Each review examined two types of physical inputs: school facilities and instructional materials. Both concluded that school facilities have an impact on student achievement, although Harbison and Hanushek (1992) conicluded that the effect was more consistentiv positive (twenty-two of thirty-four studies) thani did Velez, Schiefelbein, and Valenzuela (1993) (twenty- three of sevenity studies). The Velez, Schiefelbeini, and Valenzuela study was restricted to researchi from Latin America, a regionl that generally has higher- quality facilities and less quality variation across schools; this could account for the reduced impact of facilities on achievemilent. The latter study also found that variations in the availabilitv of instructional materials were positively related to achievemenit (fourteen of thirty-four studies), but not as strongly as suggested in previous reviews (for example, Fuller 1987; Hevneman and Loxlev 1983). The pedagogical inputs most often examined are teacher quality, curriculumII, and teaching time. Neither review found evidence that the last two were associ- ated with greater learning, but about half of the studies examined found that teacher education had a positive impact on studenit learniig. The other indicator of teacher quality-experience-was related to learining in about one-third of the Harbison and Hanushek studies and about 4() percent of the studies in Velez, Schiefelbein, and Valenzuela. Two-thirds of studies of curriculum (reform) effects found negative relations to achievement, and two-thirds of studies of the effect of student-to-teacher ratio (a proxy for teacher-student contact time) found no effect or actually a positive effect. Velez, Schiefelbein, and Valenzuela found a positive effect for increased instructionial time in only eighit of sixty studies. Aniother area of research has sought to identify specific pedagogical processes related to student achilevenienit (for example, Walherg 1986). Teaching qualitv, in contrast to teacher quality, has been found to affect student learning in the United States and other industrial countries (see, for example, Walberg 1982; Johnson and others 1981; and Postlethwaite and Ross 1992). Variations in teach- ing practices in developing countries, however, are only rarelv fouLid to be asso- ciated with variation in student learning (Anderson, Ryan, and Shapiro 1989). One recent review notes that studenit learning was associated withi the teacher's class preparationi time in five of eight studies, and frequency of homework in nine of eleven studies (FLuller and Clark 1994). Postlethwaite and Ross (1992) found that teacher assessment of student performance was consistently associ- ated with higher levels of studenit learning in the three developing countries that participated in the Interinational Association for the Evaluation of Educational Achievement (IEA) study of reading literacy. Table 1. Inpuit Effects on Primi-ary ScboolAcbievernenzt in Developing Colunitries (numher of studies) Hari7ison and Hanushek Veiez, Sch-iefelbhein, and VIailenzuela Number o/ Positive NQgative No Number of Positive Negative No Input category studies effect ef'fect effect studies effect effe Ct effeCt Physical inputs Facilities .34 22 3 9 70 23 2 45 Iinstructional nidacrials - - - 34 14 3 17 Pedagogical inputs CLuririculuml - - - - 45 13 30 2 lime: stident-to-teacher ratio 30 8 8 14 29 3 10 16 Time: hours - - - - 60 8 34 18 Teacher qualitv: experience 63 35 2 26 68 31 4 33 Teacherqualitv: education 46 16 2 28 62 25 2 35 -Not available. Soturce: Harblison and Haiiushek (1992); Velez, Schiefelhein, and Valenzuela (1993). Gleivwt ,and others 2,35 A final set of school factors influencing learning is that to which we refer collectively as "organization, climate, and control." In industrial countries, vari- ations in management, organization, and "school climate" are correlated with gains in student achievement. Specific characteristics found to distinguish "effec- tive" from "ineffective" schools include principal leadership, emphasis on basic skills, orderly school environment, high teacher expectations for students, staff stability, parental involvement, and collegial relationships (Edmonds 1979; Rut- ter and others 1979; Purkev and Smith 1983). These differences in the manage- ment and organization of effective and ineffective schools have also been found in developing countries. Effective schools promote communitv and parental in- volvement, accord school-based professionals autonomy and accountability, and are flexible with respect to their service delivery (see Levin and Lockheed 1993; Psacharopoulos, Rojas, and Velez 1993; and Raudenbush and others 1992). 11. PRIMARY EDUCATION IN JAMAICA Table 2 presents key indicators for Jamaican primary schools. Enrollment is nearly universal, at 98 percent of the cohort of children six to twelve years old in 1987, and the average number of years required to produce a graduate from the six-year system is 7.2, which compares favorably with international norms. Daily attendance, however, is rather low, averaging about 70 percent. As will be seen in section IV, this low attendance leads to a difficult estimation issue. Table 2. Quantitative Indicators oflthe Janiaican Primary School System, 1987 (percent unless otherwise noted) Indicator Measuire Enrollment of students age 6 ro 12 98 Teacher-to-student ratio 1:43 Teachers withi three or more sears of college training 95 Student attendance 70 Repetition rate in primarv schools 4.1 Repetition rate in all-age schools 3.0 C ompletion rate for grade 6 8 5 Tiime to produce a graduate (years)' 7.2 Net efficienicv 91.4 Students funLtionallV literate at grade 6-.1 7( Students reading at grade 6l1'1" 29 Cost per student per year (U.S. dollars) 96 Government expeniditure on primary educationi as a percentage of total spending on education 36.0 a. 1989. b. American norms used. Source: Years to prodLce a graduate and net efficiencN are World Bank statistics; all others are from Trevor (1990M Vol. 11, p. 1. 1. L.ockheed and Verspoor (1991, tab. A-12) show that the medianl difference in ibmher of studenr-years needed to produce a prnimary graduate and the numhier of grades in primnary school is 1.2 for upper-middle- income countries, 1.8 for lower-middle-inconme coluitries, andt 4.) for low-income countries. 236 Tlil WORI 1) BANK H ONONIKI REVIIW, Vol. 1, NO. 2 In Jamaica there are three types of schools for children in grades I to 6: prep schools, primary schools, and all-age schools. The private prep schools account for less than 5 percent of enrollment (Lockheed and Verspoor 1991), usually charge substantial tuitions, and are attended by the economic elite. About 60 percent of students in grades I to 6 are enrolled in public sector primary schools; the remainder attend the lower section (that is, grades I to 6) of the nine-grade all-age public schools. Both types of public school offer the same curriculum and use the same textbooks. Teacher pay scales and per-student subsidies are also the same across primary and all-age schools. Students not awarded a place in sec- ondary school through passage of the Common Entrance Examination continue their studies in all-age schools. There are, however, historical differences between primary and all-age schools. The latter are much more common in rural areas, especially in remote rural areas, and so tend to be less well equipped and to be used more by poorer children (Planning Institute of Jamaica, 1992: 14-15). Formerly, primary sclhool students were expected to attend academic high schools, and all-age students were expected to enter the labor force or continue their education in normal (teacher training) schools. Perhaps as a result, primary school students still gen- erally outperform all-age school students on the Common Entrance Examina- tion and have greater functional literacy (Trevor 1990, chap. 4). The Jamaican primary school system provides substantial educational inputs, despite tight financial constraints. In 1990 per-student expenditures were about $181 for public primary education (World Bank 1993, ann. 1, tab. 5). Nonsalary recurrent expenditures are low and have fallen over the past five years. After subtracting out utility payments over which there is little control, the govern- ment funds over which principals have ready discretion amount to only $2 per student per vear. As to physical and pedagogical inputs, the nearly universal enrollment shows that, at least at a minimal level, capacity exists to handle all children. Critics of that level cite that the system comes up 30 percent short on space and that 40 percent of available school places are in either bad or very bad condition (Trevor 1990: 24). Student-to-teacher ratios, which averaged 43:1 in 1987, are also rela- tively high. But, about 90 percent of these teachers have received training (Trevor 1990, vol. II, tab. 1.11, p. 25). Regarding pedagogical practices, the curriculum has been reformed since In- dependence to reflect more Jamaican and Caribbean content, illustration, and images. Textbooks in basic subjects have been provided annually to all students under externally funded projects since 1982. Finally, with respect to school mani- agement, schools are relatively autonomous in Jamaica. For example, the local Board of Education and principal hire and fire teachers and nanme them to internal school leadership positions such as vice principalships, grade coordina- tors, and senior teachers (James-Reid 1989: 60). Miller (1990) reports that the degree of autonomivy has led to substantial innovation in work-study, school feeding, curriculum reform, school organization, and school financing. Glewve and otbers 2,37 111. DATA COLLECTION AND VARIABLE SPECIFICATION The data used in this article come from the 1990 Jamaican Survey of Living Conditions (SLC), which collects detailed information on household structure, health, education, nutrition, consumption, and housing (see Grosh 1991 for a general description). The education data contain unusually detailed schooling information for all individuals age twenty-five and under, including 1,151 chil- dren in primary-level schools. This section provides details of the data collection and the choice of variables for the analysis. The 1990 jamaican Survey of Living Conditions The 1990 sL.c identified the school attended by all children in primary or secondary school in each household.2 These 212 primary and all-age schools and 110 secondary schools became the sample for the school survey. The school administrator's questionnaire collected information on physical characteristics, school feeding programs, expenditures on schooling, instructional materials, admission, completion, dropout and repetition, teachers and parents, school organization, and comlmlullication. In addition, a teacher questionnaire was ad- ministered to up to ten teachers randomly; selected in each sample school. All in all, 1,640 primary and all-age teachers and 972 secondary teachers were sur- veyed. The teacher questionnaire requested information on the teachers' teach- ing history, use and opinioni of textbooks, and use of class time. Both the school and teacher variables reflect the average environment in the school rather than that of the particular classroom in which each sample child was enrolled in 1990. Our indicators of cognitive skills for this study are the California Achievement Test (CAT) measures of mathematics computation and reading comprehension. The small household-based sample of 1,151 children in primary school meant that no more than about 200 students coLlId be expected to be in each grade. Thus a test was required that would yield comparable scores across all grade levels (that is, a vertically equated test). Because there are no vertically equated Jamaican tests and there was no possibility of developing one witlin the scope of this research project, other tests conmmercially available in English were consid- ered. Several options were reviewed with the lamaican Ministry of Education. The CAT was selected because of its technical quality and because it provides the fewest apparent concerns over cultural biases. It has proven to be a valid and reliable instrument to use in studying variation in achievement among jamaican school children (see Harris 1993). Children from the households in the SLC were traced to their schools for testing, then were taken out of their regular classrooms and tested together. The mean number of children tested per school was 2.3. A screening test was given to determine which level of the battery of complete tests was appropriate to a given child; thus each child took a test appropriate to his or her achievement. 2. The liSting was avrLiallV carried out as part of the October- 1990 Labour Force Survey I LFS). The sLc Is designed to cover a raidom sushample of 2,592 households of the 7,566 hloseholds covered in the LFS. 2.38 THIl N(lKI.I) LANK H 0N()MI( \ RIVILW., V0l 9, NO. ' Unfortunately, it was difficult to locate a large portion of the children in their schools simply by using the information gathered from the households because of a series of factors: the chronically low attendance rates observed among Jamaican students, differences between the colloquial names used for children in the household and the legal names used on school enrollment rosters, the time lapse between the October 1990 listing of children (and their schools) and the school surveys in March and September of 1991 (especially for the second wave of schools surveyed), and the impossibility for the interviewers to visit the schools more than two or three times. In the end, test scores are available for less than half of the children (508 of 1,152). Although many children were not tested for reasons clearly unrelated to their potential performanice on the CAT, absenteeism is not a random event and thus creates a problein. Table 2 shows that on an average day, 30 percent of primary school students in Jamaica are absent and thus would not be tested. Because interviewers made repeat visits to schools to retest absent children, this 30 per- cent figure certainly overstates the loss of sample caused by absenteeism, al- though by how much is difficult to say. The tested children had the following characteristics: 56.5 percent were female, their average age was 9.4, household log per capita expeniditure was 5.1, 55 percenit were from ordinary primary schools and 41 percent from all-age schools, and 38 percent came from urban areas. The childreni not tested had the following characteristics: 46.7 percent were female, average age was 9.8, log per capita consumption was 5.5, 54 percent were in ordinary primary schools and 38 percent in all-age schools, and 48 percent came from urban areas. Thus, clildreni who were not tested were more likely to be male and from urban areas, wlich suggests that unobservables, such as motivation, might differ across samples as well. We return to this selec- tivity issue in section IV. Choice of School Variables In addition to student characteristics (sex, grade, and age) and home hack- ground information (parental education and household per capita expenditure), we consider a wide range of school variables, which we divide into three catego- ries: school physical and pedagogical inputs, school pedagogical processes, and school organization and climate. The inclusion of variables such as household income (per capita expenditure) is warranted because the achievement regres- sion specified in the next section is not a structuLral education production func- tion. The underlying structural production function includes school attendance, an endogenous variable that depends on household income; we have substituted out this variable in the reduced form equationi that we estimate (see Glewwe and Jacoby 1994). It is also worth noting that the estimated impact of school vari- ables on achievemiienlt may in part be attributable to their influence on school attendance. Potentially our data would allow us to construct hundreds of school variables. Unfortunately,, our degrees of freedoml are constrained by the relatively small Gletutwe and others 239 number of students in our sample, which limits the number of school variables we can use. We eliminate variables that do not show sufficient variation (for example, dummy variables where 95 percent or more of the responses are in one of the two categories). We also try to avoid redundancy as much as possible, as well as characteristics that seem only tangentially related to the learning environ- ment. Three years, since the design of the questionnaire, spent working with the Jamaican government on education policy questions gave us a sense of which variables were good measures of the object or process of interest. After applying these selection criteria, we were left with forty-two school and teacher variables. The remaining school and teacher variables include each set of factors thought to be important by the different sides of the debate on school efficiency. More specifically, the fortv-two variables that were retained measure school quality in terms of physical inputs (the school facility, instructional materials, health- related services); pedagogical inputs (curriculum, instructional time, and teacher quality); pedagogical processes (teaching practices in the classroom); and school organization, climate, and control (school autonomyv, work-centered environ- ment, community involvemeit, orderly environment and school tvpe). In the Jamaican school systemii the central ministrv is responsible for providing the school facility, instructionial materials, and personnlel; teachers largely determine which specific practices they employ in the classroom; and the principal may have some control over the "climate" of the school. Although ildicators of these differenit aspects of school quality could potentially be highly correlated, empiri- cally we did not find any of the correlations to be very large.' IV. ECONOMETRIC Sl'EC(IFI(CATION Parenlts in jamaica have considerable freedom in choosing which school their children attend. Consequently, standard regression estimates, for example, ordi- nary least squares (oLs), of the determilants of student achievement may give biased estimates of the imiportance of school characteristics. If better schools tend to attract better studenits, these estimates would confound the direct effects on cognitive skills of changing school characteristics with the indirect effect of differences in the composition of the student body. Another potential selectivity bias arises when some students are not tested hecause they are absent on the day(s) the test is administered.4 This sectioni discusses both types of selectivity bias. Although we are able to address the problem of school choice, the problem of absenteeismii defies an econonmetric solution-. 3. As poiintd oiut bh. a referee, Jorrelations betwecen the difterenit sclhool variables raise the issuie that schoo() = 0 other-wise, where wi is another vector of observable characteristics. The error term Ui, repre- sents the unobserved preference for primary over all-age schools. This error term is likely to be correlated with the e's in equation 3, because both may reflect the importance the houselhold places on educationi. If Eil, Ej,, and u,I are jointly 5. P-ivare (pleparator)y school stndent, are droplped hICCJUSe thlec are only ten SLIChI students in the s;ample. 6. Before estimilating this modlel. we allol tht r to also doiffer across sciool types and tested this model against thc more r-estrictive specification given hy eqUations I and 2. Althougil we rejectedt the restrictions at the 5 percent signiticance levcl, we toLlund that the parantetcr estimlates of the unrestricted modiel were extremelc imprecise, reflectin1g the large 11111111nhel ot pnats in relation to sample size. Xe thus decided to rest rict the y to h,e the sane. C;lewU'Cve and others 241 normal with an unrestricted covariance matrix, then we cani write the condi- tional expectation of equation 3 as (5) LICAr,|x, Z, I=aD,, +x,fi+ zyI + D,,Eie,i|D,1 = 1J+ (I - D,1)EI,,jBD,i = 01 aoD , + x',B+ z,y+ 5D,,A,, + 5.(I -D,I)K,, where 6, = -aIjaI; 3Ž = aJ21,1ai,, = 0(w,;o)1)(w;o), and i,, = O(w,0)IF1 - F)(vw;o)1. In the latter two expressions, 0 and 4) denote the nornmal density and cumulative density functions, respectively. Equation 5 shows that, in general, the composite error in equation 3 has nonzero mean. Omitting the terms involving A,, and A,, could lead to biased OLS estimates of equatioi 3, the determinants of achievement. In particular, not taking into account the correlation betweenl the dumilmy variable Di, and the composite error term in equationl 3 could give Lis a misleading assessment of the efficacy of primary conmpared with all-age schools. To correct this potential bias, we use the now standard two-step method of Heckman (see Maddala 1983, ch. 8), first estimating equation 4 as a probit and then, in the second step, inserting estimates of the conditional mean correction terms (the As) into equation 3 as additional regressors. So as not to rely strictly on the norimiality assumption to identify selection effects, we include a variable in w, that is not in x,. This variable is the difference betweeni the distances to the all-age and primary schools that the househiold reports as being nearest. In Jamaica it can be reasonablv argued tlhat the cost differential of sending a child to one public sclhool over another arises mainly from this difference in distanices from the household. Another way to achieve statistical identification is to include school quality characteristics of both types of sclhools in equation 4. The characteristics of the type of school not attended can theni serve as ideitifyiig variables (see llewvwe and jacoby 1994). However, for reasons explailied in section V we do not use school characteristics in esti- mating equationl 4. Now turn to the selectivity bias induced by studenit absenteeism. Children who are often absenit from school are less likely to be tested on the day(s) set aside for that purpose. These children are likely to belong in the lower tail of the motivation distribution anid to test lower than the average child. Hence, the sam)ple of tested childrein may be noniranidom, leadinig to biased estimates of y Intuitively, we would expect this bias to attenuate the estimated impact of school quality on achievemeit. The reason is that, all else being equal, higher-quality schools would tend to attract less-motivated studenits on any' giveln day (that is, better schools encourage suich studenits to attend), and these students would bring down the average test score in the school. Although the selection bias problem resultinlg from poor attendance is poten- tially serious in the Jamaican context, it is virtually' impossible to correct. Indeed, most (perhaps all) existing studies of school achievenmenit in developing countries, where absenteeisimi is typically high, are subject to the same criticism. The identi- 242 I HF WOKRLI) ANK I ( ONOMIR R1 VIF, Vol.. 9, NO. 2 fication problem in this case is that any variable that affects the child's decision to attend school on the day of the test, such as travel time to school, is bound to affect whether the child attends school on any other day and thus should influence scholastic achievement directly. We do in fact find that children who live farther away from their school are significantly less likely to have taken the test. The selection bias effect can thus be identified only from functional form (distri- butional) assumptions. The fact that test administrators missed some children because, for example, they tended not to revisit more remote schools, does not ameliorate the identification problem, even though these reasons are uncorrelated with child-specific unobservables. We still require variables that influence the child's attendance decision but not his or her test scores directly. For a rigorous discussion, see Maddala (1983: 278-82), who shows that when multiple criteria for selectivity are present each criterion requires an exclusion restriction to achieve (nonparametric) identification for the equation of interest. It is also worth point- ing out that even data on changes in school achievement over time, or "value added," would not necessarily attenuate this selectivity problem, because school attendance is a choice variable rather than an individual fixed effect. To summarize, the school choice issue can be dealt with econometrically, provided that we have a variable, such as distance between school types, which determines the choice of schools but not student achievement within a school. Absenteeism, on the other hand, creates a potentially serious selection problem that is intractable econometrically. On the bright side, however, this selection bias militates against finding significant effects of all three types of school vari- ables (inputs, pedagogical processes, and school management) on achievement. So our estimates, by placing lower bounds on the effects of individual variables, would still be useful for policy analysis. More generally, significant impacts from all three types of school variables would justify our eclectic approach. V. DETERMINANTS OF ACHIEVEMENT This section presents estimates of the determinants of school choice and the determinants of student achievement. Although the school choice estimates are not of primary interest, they offer some insights on primary education in Ja- maica; therefore, we begin with a discussion of these results. Scbool Choice and Selection Bias We estimate the determinants of the choice between the two kinds of primary- level schools in jamaica-all-age and primary-using a probit model. Before turning to these results, however, we must point out that as an empirical matter it is impossible to model the full school-chioice decision. In Jamaica, particularly in urban areas, most parents face a wide variety of schooling choices. Our data do not provide us with information on the characteristics of, and distances to, all the schools a child may potentially attend. We do know the distance of the household from both the nearest primar y school and the nearest all-age school, Gleuwi'e and others 243 but we do not know the other characteristics of these schools. We therefore assume that parents consider schools oniy in their parish (Jamaica is divided into fourteen parishes), and we use thirteen parish dummy variables to control for the differences in school quality (parish averages) between all-age schools and primary schools. Table 3 presents our probit results for the school choice model on the full sample-with or without CAT scores-of 1,067 children (85 observations are lost because of missing data). The distance variable is strongly significant and has the expected sign. The farther away the nearest all-age school is in relation to the nearest primary school, the less likely the child is to be sent to an all-age school. Recall that this variable is the one plausible identifying variable in our school- choice model, so its statistical significance is heartening. Seven of the thirteen parish dummy variables are also significantly different from zero at the 5 percelit level, suggesting that importanit differences may exist in relative school quality between primary and all-age schools in different parishes. All-age schools appear to compare unfavorably in St. Catherine's (the omitted dummy variable), St. Thomas, and St. Elizabelth but compare relatively favorably in Portland, St. Mary, Trelawny, and Manchester. We also include several household variables that may influence schooling choice: parental education, houselhold per capita expenditure,7 a female child dummyv variable, and the age of the child, but none, save father's education, are statistically significant. Better-educated fathers appear more likely to send their children to a primary, rather than an all-age, school. With the schooling chioice parameter estiniates in hand, we construct the conditional meani correction terms as described in section V and estim.ate the cognitive skills regressions. These regressions, along with the descriptive statis- tics for the variables, are presented in table 4. Only one of the four school-choice selectivity correction terms in the math and reading regressions is statistically significant above the 10 percent level. The coefficient estimates are not apprecia- blv affected by the exclusion of these terms, which suggests that selectivity hias as a result of school ciloice mav not be importanlt in our data. Of course, as discussed already, there may be other sources of selection hias that are statis- tically important but that we cannot explore econometrically for lack of identi- fying information. Student Cbaracteristics andiit Achievement As expected, children in higher grades outscored those in lower grades (see table 4). We tested whether the linear relationship between grade and achieve- ment was overly restrictive by entering dummy variables for each grade in an alternate specification. The F-test could not reject the linearity assumption at the 5 percent level. The meani achievement gain per grade is about 47 points for math and 37 points for reading, a useful standard for comparing the magnitude 7. We trear this vari ahle as exogetnous to the choict of school. Fhis assitioiption is rea.sonait becaILIe trasel time and transport costs are likcl\ to be small LOmpMCrLd Wvitl houMIsehold Iiiconme. 244 TiiF WORI D BANK F.( OON(M RFVIW. Vol . 9. No. ' Table 3. Probit Estimates of Scbool Choice, Jamaica, 1990 Variable Coefficient Intercept 0.2330 (1.03) Distanice to nearest school (miles) -0.0945 ( I 1.34 '! Parish) dumnm-y variablesil Kingston 0.4294 1.66) St. Andrew 0.0834 (0.59) St. Thomas -0.3298 (-0.93) Portland 0.6937 (3.07* *) St. Mary 0.8845 (3.81' Sr. Ann1 0.3915 (2.01 -) Trela wny 0.8442 (3.64* !) St. James 0.0464 (0.22) Haniover 0.4230 (1.87) Westnioreland 0.4253 (2.24*) St. lFlizaheth -().2601 (- l.3.5i Manchester 1.0140 (S. 26* ) Claredoni 0.3511 (2).12*) Other variables Fathier's schooling (years) -0.0300 (- 2.11 -) FIather's education missing -0.0128 (-0.10) Mother's schooling (years) -0.0222 (-1.36) Mother's education missing -0.2785S -1.49) Per capita household expenditure (Jamaica dollars) -0.0218 (-1.60) Age of studeint (years) -0.i189 (-1.45) Femnale student 0.0160 (-0.19) indicates significanice at the 5 percent level. indicates significance at the I percent level. Note: The sample size is 1,067. The log likelihood value is -60(0.17. t-statistics are in parentheses. Dependent variahle equals o01e if all-age school attended, zero otherwise. a. St. Catherine's is rhe omitted parish dummy variahle. Source: Authors' calculations. Gler,vu'e a.nd others 245 Table 4. Descriptive Statistics and Estimates of the Determinants of Math Computation and Reading Comprehension Scores in jamaican Primary Schools, 1990 Math Reading Standard lcomputation comprehension Variable Nfean deu'ration estimate estimate Achieuemnent Mathematics computation scaled score 605 140 n.a. ii.a. Reading comprehension scaled score 512 159 n. n.a. Basic studtent characteristics Age of students I months) 124 21 -0.21 0.51 (-0t).3.5) (0.72) Grade of students 3.33 1.70 46.98 37.23 (6.45*) (4.34 -$) Sex (1 = female; 0 = male) 0.59 0.49 41.66 77.64 (3.19 *) (.03 Homne background Household per capita expenditures 4.70 3.(19 8.74 4.47 (thousands of JamaiLan dollars) (3.6(5* ., (1.60) Father's education (years otschool 7.35 3.12 -0.41 1.90 completed) (-0.18) (0.70) Father's education data missing (I = ves; (.10 0.30 28.94 58.58 0 = no) j1.22) (2.10 ') Mother's education (years of school 8.45 2.6 2.80 1.29 completed) (1.02) (0.40) MNorler's educationi data missing I1 = yes; 0.(5 0.21 - 11.84 - 25.05 0 = no) (-0.36) (-0.65) First selectivity correctioni term n.a. n.a . 4.72 -9.0( (-0.28) (-0.45) Second selectivitv correctioni term n.a. na.3. 33.23 2.15 (1.83:-) (0.10) Schoof-lev'el physical inputs School /acility Classrooms not separated hy walls 69.0 30.0 -- 16.54 --6.64 (percentage) (-0.54) (-0.18) Students witl desks (percentage) 85 15.5 0.72 1.16 (1.23) (1.69*) Index of equipmentet (ranige: low = 0 to 2.84 1.78 - 8.42 -5.48 high = 7) -1.41) (-0.78) Number of specialized instructional 0.81 0.84 0.48 2.26 roomsh (0.()4) (0.15) Reliahilirv of electric service (range: 2.15 0.96 -3.37 -6.52 had =0 to good = 3) (-0.33) (-0.55) Piped water ( 1 = yes; 0 = no) 0.73 0.45 - 1I5.99 30.35 (-(0.63) (1.02) Instrictionial mnaterials Classrooms with usable hlackboards 92.0 12.4 0.21 -0.67 (percentage) (0.33) (-(.85) Index of instrtuctional materialsc 9.56 2.63 -2.45 -6.16 (ranige: low = 0 to high = 20) (-0.65) (-1.45) Index of writing materialsd (range: 3.85 1.71 -1.28 6.58 lo%v = 0 to high = 9) (-0.24) (1.04) (Table continues on the following page.) 246 I HI: Wio I) RANK HI :ON()MIN : RiV1FW. VOI .9, No.? Table 4. (continued) Math Reading Standard computation comprehension Variable Mean deviation estimate estimate Health NLurse available or visits (1 = yes; 0.70 0.46 20.11 - 15.73 0= no) (-1.16) (-0.77) School conducts eve tests ( I = yes; (.17 0.38 51.79 33.42 0 = nl) (2.26*) (1.23) School-level pedag'i)gical inputs Textbooks arrived two months or more 0.76 0.43 36.40 14.92 late (I = yes; 0 = no) (-1.78x (-0.63) Studenit-to-teachier ratio 39.4 7.8 2.06 1.30 (1.56) (0.90) Teachers withi traininig in the last three 32.6 2.3.7 - 3.87 104.88 years (percentage) (-0.I0) (2.26*9) Teachers with diploma or certificate 84.4 22.1 48.70 --9.01 (percentage) (0.85 (-0.13) Teachers' average primary teaching 13.6 3.3 0(.88 1.92 experience (yearsl) (0.39) (0.72) Pedagogical processes Average intensity of testing studenits 2.48 0.66 29.85 40.73 (0 = never, 5 = almost everv lesson (1 .85*) (2. 14e * Time spent on whole-class instruction .31.2 7.5 - 3.63 -- 2.83 (percentage) - 1.67-) (-1.12) Class time teacher copies n(otes onto 12.6 5s5 -0.31 2.53 blackboard (percent) (- 0.0) (0.72) Class time spent instrUctilIg small groups 15.4 4.7 -0.64 0. 1 7 (percentage) (-0.21)) ) 0.04) Individual writtenl assigilnmeits done dotriig class-time (O = never; 5 = almost 3.69 0.77 -.34.03 -35.50 every lesson) (-3.06*', (>-2.81 ) Time spent providing individual instruc- 12.07 3.43 -0.29 3.08 tion in class (percenitage) (-0.0)9) (0.79) UIse of textbooks ii instruction 8.19 1.17 10.14 21.69 (0 = never; 20 = almost every lesson) (1.29) (2.371*) Lectures to whole class (O = niever; 4.19 0.62 1. -).s 7.80 5 = almost every lesson) (0.81) (0.45) In-class homework review (O = never; 3.85 0.45 22.02 -27.60 5 = almost every lesson) (0.95) (- 1.01) Students copy notes from blackboard 4.05 0.52 -0.96 -16.78 (i) = never; 5 = almost every lesson) (-0.05) (-0.78) School-level organizatiion, clinmate, and control School auitononzy Relative influence of the Ministrv of Education compared with the 0.13 (.18 - 11.49 - 3.03 principal on1 the school's -0.21 ) (-0.(5) organization' (rario) Principal's influenice on the curriculum 1.62 (.77 20.94 13.19 (range: none = 0 to high = 6) (1.12) (0.60) Teachers' influence oi the curriculum 2.14 0.79 9.04 11.42 (range: none = 0 to high = 6) (0.47) (0.51 GleCt'u'e and others 247 Math Reading Standard co;npiitutolol Cr07mprebension Variable Mean deviation estimate estimate Work-centered environment Curriculum or pedagogy first or second most common theme in staff 0.16 0(.37 38.88 43.58 meetings ( I = yes; 0 = no) (1.71 1 (1.61) Instructional assistance and leadership 9.29 -.46 0.88 2.35 by principal (hours per week) (0.72) (1 .65) Index of commitment to teaching hasic skills and critical thinking (range: L.79 (0.41 1 1.7 I 0.55 none = 0 to high = 1) (0.48) (0.02) Average frequency of receiving help from other teachers to improve 1.29 0.82 --8.23 21.20 teaching skills (range: never = 0 to (0--.84) (1 .86;) daily = 5) Comimunintv involuement Parents who attend P-rA meetinigs 27.9 18.9 -0.66 0.16 (percentage) (-1.49) (-0.30) index of Communllity ivolvemilent 0.96 0.47 3.99 9.79 (ranige: none = 0 to highi = 2) (-0.22) (-0.47) Orcderly, environment Uniform requirement strictly enforced (.32 '(.47 29.51 28.11 (I = yes; 0 =no) (1.61) (1.3 1) Class time spent on disciplinie 11.3 3.7 1.03 2.10 (percentage) (0.28) (0.48) S.chool tvpe School operated in shifts I = yes, 0.88 (.33 -23.1 1 -39.70 0 = no) (-0.87) (- 1.29) Students grouped by abilitv (I = yes; 0.63 0.4 . . 31 24.46 0 = no) (0.16) (1.07) All-age school (I = ses; 0 = no) 0.45 0.5( -42.20 -46.00 (-1.62) (-1.56) City school, Kingston excluded 0.18 0.38 -M1).69 3. 22 (1 = yes; 0 = no) (-0.38) ((.1)() Rural schtool (I = yes; 0 n,o) (.7(0 (.46 6.48 -2.76 (0.24) (-0.09) Adjusted R2 n.a. n.a. 0.36 0.32 F-statistic n.a. n.a. 5.912 4.218 n.a. Not applicable. -Not available. Sigiiificant at tht 10 percent level. Signlificaint at the 7 percent level. * Significant at the I percent level. Note: The sample size is 355. Of the 508 children for whom test scores were available, 10 were dropped because thev attended private (preparatory) schools, 85 were dropped because they could not be matched with any school for whichi we had school and teacher questionnaires, and 58 were dropped because of incomplete data from the school and teacher questionnaires. t-statistics are in parentheses. a. Available equipment in the sclool, including telephone, typewriter, television, computer, radio, and copying or duplicating machine. b. Includes libraries, labs, and studios. c. Includes maps, charts, science kits, and dictionaries. d. Includes pens, pencils, paper, notebooks, complete set of required text books, and dictionaries. e. The intiuenice of rhe Ministry of Education on che organizarion of the school is scaled on a range fromn nonte = 0 to high = 12. The influence of the principal is scaled on a range fromii none = 0 to high = IS. Sourer: Authors' calculatinrls. 248 rHF W')RI 1) BANK I ( t)NoNM1 RFVIFW, VO'L 9, N(o. I of effects discussed later. After controlling for grade, the student's age had no significant effect. This is not surprising in Jamaica, where delayed enrollment, repetition, and dropout in the primary grades are all relatively rare. Gender also influences achievement, with girls showing much higher cognitive achievement than boys in both mathematics computation and reading compre- hension. For the math scores, age and grade being equal, girls' scores exceed boys' by about the amount learned in one grade of schooling. For the reading scores, the girls' advantage is two grades higher. Internationally, it is not unusual for girls to outperform boys at the primary level. For example, in a 1983 interna- tional study of math achievemenit in nineteen education systems, boys outper- formed girls in ten systems, whereas girls outperformed boys in nine systems (Robitaille and Garden 1989). In a 1991 study of reading achievement in twenty- seven education systems, girls outperformed bo's in all systems (Elley 1992). And it is not surprising in the Jamaican context that girls outperform bo's. After all, about 59 percent of the placements into highi schools (the most elite track of the secondary system) are awarded to girls (Miller 1988, p. 4). Students from "good" homes may do better in school for a number of reasons: more resources that promote cognitive learninig at home, a higher value placed on school performance, or a generally calm, stable environment. We include several variables to measure these effects. We also tried to include variables related to children's school behavior: days attended, number of times late for school, and hours spent on homework, all referring to the seven days previous to the household interview. Only the attendance variable was significant (positive) in the mathematics regressioni and only the homework variable was significalit (again positive) in the reading regression. Because these variables cannot be considered as exogenous, we left all of them out of the regressions reported here. Their absencc has virtually no effect on the other parameter estimates. Economic welfare, as measured by household per capita expenditure (again treated as exogenous), is strongly and positively significant for mathematics, but not for reading; children from poorer houselholds have lower mathematics achievement. An increase in household per capita expenditure by one standard deviation implies a 43-point increase in mathematics-almost the same effect as an additional year of schooling. Exploratory regressions found little evidence that this effect varied across expenditure levels. Similarly, we found no signifi- cant interaction terms between houselhold expenditure levels and school charac- teristics. Parental education, however, has no significant influence on achieve- ment. One possible explanation for this is that the economic welfare variable used here is more comprehensive than what is usually available in other data sets. Thus, once its effects are sorted out from those of parental educationi, the impact of education per se disappears. Furthermore, Jamaican society is culturally rather homogeneous, and parental education levels are generally high when compared with those in other developing countries; therefore, even relatively less-educated parents may place importance on their children's education. CIcu'u e afnd (others 249 School Characteristics anid Achievement We now turn to the effects of the three categories of school characteristics, beginning with the physical and pedagogical inputs. Table 4 shows that only one such input, vision testing of students, has a significant (beneficial) effect on math achievement at the 5 percent level. Significant at the 10 percent level in the math regression is the negative effect of textbooks arriving late in the school year. Only one input variable has a significant impact on reading achievement at the 5 percent level-teacher training received within the past three years, which causes students' reading skills to improve. In addition, availability of desks for students is significant at the 10 percent level. Overall, these results suggest that physical and pedagogical inputs may play oniv a marginial role in explaining cross- sectionial differences in cognitive skills in Jamaica. Of course, the weak perfor- manlce of these variables may also reflect downward bias in parameter estimates as a result of sample selectivity with respect to the students who were tested. Nonethieless, the differences in effects across subjects admit plausible explana- tions. If math teaching relies more on blackboard exercises than reading does, then vision tests are likelv to be more important to math achievement than to reading. (In Jamaica, vision testing may also be a proxy for communiity involve- ment in the school, because these tests are often carried out by local charitable organizations, although this would not explain their lack of impact on reading scores.) Likewise, we speculate that the content of recent teacher-training pro- grams has been geared more to reading than to math skills, hence the stronger effect of teacher traininlg in the reading regressioni. Next consider the effects of pedagogical processes. One pedagogical process variable has a strongly significant (I percent level) impact on mathematics scores. Doing writteni assignments in class (seat work) has a strong negative effect on achievement. Two pedagogical process variables are significant at the 10 percent level: testing students has a positive effect on mathematics achieve- ment, but time spent in whole-class instruction has a negative effect. On the reading side, one pedagogical process variable is strongiv (I percent level) related to achievement, and two more are significant at the 5 percent level. Intensity of textbook use and the percentage of teacher time spent testing students have positive effects on achievement, but time spent doing written assignments in class detracts from learning. The results on written assignmients in class and whole-class instruction do not, of course, imply that these practices are directly harmful. Rather, they may divert class time away from other uses that better stimulate learning. Finally, we turn to school organization, climate, and control. No variable in this category emerges as significant at the 5 percent level in the mathematics regression, although the positive effect of disciussing curriculum and pedagogy issues at staff meetings shows marginal (10 percent level) significance. In the reading regression, none of the school organization variables is significant at the 5 percent level, but two are weakly significant at the 10 percent level: hours of 2,50 [ HF WORI[ 1) 1NK F ( )N IC RI( 01 VH. V V .. 9,St instructional assistance by the principal and the average frequency with which teachers help each other lead to higher scores. Finally, note that the all-age school dummy variable is insignificant in both the mathematics and reading regressions, meaning that, once school and student characteristics observed are controlled for, there is no appreciable difference in average achievement between these two types of schools. Overall, we find that variables measuirinig pedagogical processes are more often significantly related to student achievemenit thani are physical and ped- agogical input variables and school organizationi variables. It should be noted, however, that because Jamaican schools have such high levels of some inputs generally believed to be important in increasling learning (textbooks, writing materials, school feeding programs, high levels of teacher certification), there may have been inadequate variation in the sample to measure their effects on achievement. Alternatively, Jamaica mav have reachied the point of diminishing returns withi respect to these inputs. Anothier, less sangLuille, interpretation of our estimates is that selection bias resulting from school absenteeism has a greater impact on the school input coefficients than it does oni the process and organiza- tion coefficients, yet there is no a priori reason to think that this would be the case. VI. POLICY SIMULATIONS This section presents the results of policy simulations to explain the differ- ences between all-age schools and primary schools and between ineffective and effective schools. Closing the Gap betueen All-Age School and Primiiarv School Characteristics The gap in student performance betweeni all-age schools and primary schools has long been a focus for those concerned with equity in the Jamaican primary school system. We can use our estimates from the previous section to shed some light on the following question: By how mucih would the learning of a child with a given set of home chiaracteristics improve if he or she were transferred from an all-age school to a primary school? To do so, we predict the difference in compo- site learning scores in three steps. First, we predict math and reading scores using just the school characteristics, by simply multiplying the school characteristic coefficients from table 4 by the mean values of their respective variables for both all-age schools and primary schools. Then we take the differences between the predicted scores in each subject for all-age schools and primary schools. Finally, we add the two together to get a composite score differential. Of course, this amounts to giving equal weights to literacy and numeracy skills. Ultimately, policymakers must decide how much weight to put on each skill, and if unequal weights are chosen, the results given here may change somewhat. In the absence of such information, we simply weight math and reading equally. The overall difference in the predicted composite score for an average child attending an average primary school compared withi an average all-age school is Glewtc' and otbers 2,51 139 points, as seen at the bottom of table 5. This difference is equivalent to about one and a half grades' worth of learning. Differences in measured school characteristics account for 50 points of the difference in composite scores. Of these 50 points, 37 are due to differences in physical and pedagogical inputs, 12 to differences in pedagogical practices, and only less than 2 to differences in school organizationi and climate variables. The relatively small differences in scores attributable to defined school characteristics is an indication of the small differences in meanl characteristics between the two kinds of schools, especially for the latter two grouIps of variables. Of the total difference in predicted composite scores, 88 points' worth is due to the dummy variable for all-age schools. This is a residual difference (although it is not statistically significant) that is not explained by either the measured school characteristics or by selectivity effects. OCur estimates suggest, therefore, that changing the chiaracteristics of all-age schools may close only about 40 percent of the gap in predicted composite scores between children in the two types of schools. SLibject to the limitations of our data, this exercise does not indicate that sizable improvements in studenit achievement can be obtained by converting all-age schools into primary schools. Clositng the Gap between hieffectiue and Effective Schools To show the differences between "good" and "bad" schools, we perform an analogous simulation. We do so by predicting math and reading scores and producing a composite score for each school. We then rank schools from best to worst. Schools in the top quartile, we define as "good" schools, and those in the bottom quartile, as "had" schools. The simulationi results are presented in table 5. The range in school perfor- mance is very large: the mean predicted score for schools in the bottom quartile is 309 points lower than that of schools in the top quartile, and this difference is equivalent to alnost four grades' worth of learning. Because the difference be- tween the average predicted score for good schools and for bad schools is So much greater than the difference between the average for all-age and for primary schools, resources would appear to be better spent on improving bad schools from either track rather than on equalizing the averages for the two tracks. Indeed, a few of the bad schools are primary schools, and 31 percent of the good schools are all-age schools. Table 5 shows the sources of the 309-point difference between the good and the bad schools: differences in physical and pedagogical inputs account for 85 points; in pedagogical processes, 112 points; and in school organization and climate, 60 points. In this case, the 51-point difference arising from the effect of the dummy variable for all-age schools accounts for a much smaller proportion of total predicted differences. Assuming no serious problems of bias in our parameter estimates, bringing the bad schiools to the standard of the good Table 5. Predicted Coomposite Score andt Characteristics for All-Age and Primary Schools and fo1r the Bottol anld Top 25. Percent of All Primary-Level Schools, Jamaica, 1987 All-age and primarly scbools Bottonm andl top 25 percent o/ all primary-level scbools Dillrernce in Mean tor Ditference in total Mean /or Mean fi° total s-ores of bottom 25 Mean for top 2.5 sc ores of bottom 25 all-age primary all-ageand percent o/ percent of percentand top 25 Variable schools scbools primary schools schools scbools percent of scbools School level physical inputs School facility Classrooms not separated by walls 0.61 0.66 1.18 0.64 0.57 1.66 (percentage) (3.42) (4.78) StLdents with desks (percentage) 81.15 87.88 12.63 88.26 82.81 -10.22 (8.56) (6.93) Index ofequipmenta (range: low = 0 to 2.58 3.17 - 8.21 2.48 3.2.3 - 1 0.32 high=7) (7.69) (9.66) Number ofspecialized instructional 0.83 0.74 -0.25 0.90 0.8 -0.0(9 toomsb (2.57) (0.88) Reliability otelectricservice (ranige: bad 2.00 2.26 - 2 58 2.19 2.13 0.64 0 to good = 3) (5.74) (1.42) P'iped water (I = yes; 0 = no) 0.58 (.90 4.53 0.58 0.81 .3.24 (17.36) (12.44) Instructional materials Classrooms with usable blackboards 90.12 91.46 -(0.61 94.48 87.90 2.99 (percentage) (1.94) (9.49) Index of instructional materialsc (ranige: 9.40 9.82 -3.60 9.77 9.35 3.61 low = 0 to high = 20) (3.35) (3.36) Index of writing imaterialsd (range: 4.22 3.64 - 3.07 4.42 3.94 -2.56 low = ( to high = 9) (6.76) (5.65) Health Nurse available or visits ( I = ves; 0 = no) 0.60 0.57 1.25 0.74 0.45 10.40 (1.32) (10.98) School conducts eve tests ( I = yes; 0 = no) ((.17 0.25 6.86 0.13 0.42 24.74 (4.0(3) (14.55) School-le-el pedagogical in puits Textbooks arrived two months or more late 0.77 0.71 2.80 0.87 0.61 13.24 (I = yes; 0 = no) (2.40) (11.36) Student-to-teacherratio 36.18 42.72 21.93 32.48 41.X7 31.56 (18.01) (25.91) Teacherswith trainingin tie lastthree years 0.31 0.32 1.67 0.27 0.34 6.12 (percentage) (1.42) 5.21) Teachers with diploma or certificate 0.79 ().89 3.67 0.7 5 0.92 6.59 (percentage) 11.50) (20.66) Teachers' average primary reaching 13.23 13.58 (.97 12.84 14.20 3.78 experience (years) (1.69) (6.63) Stibrotal (physical and pedagogical inputs) n.a. n.a. 36.81 n.a. ri.a. 85.38 Pedagogical processes Average intensity oftestingstudents 2.44 2.51 5.00 2.22 2.75 38.()1 (O= never; 5 = almosteverv lesson) (2.49) (18.9.3) TFime spent on whole-class instructioni 30.98 32.50 -9.82 32.95 30.81 13.81 (percentage) (7.15) (10.05) Class time teacher copies notes onto 12.95 1 1.23 - 3.8.3 12.76 11.96 - 1.78 blackboard (percentage) (11.33) (5.27) Classtimespent instrticting simallgroups 15.36 14.71 0.53 15.04 14.78 0.21 (percentage) (4.65) (1.85) Individual written assignments done dtiring class time (O = never; 5 = almost every 377 3.66 7 2 4.04 3.36 47.34 lesson) (2.46) (16.17) Timespentprovidingindividualinstruction 12.19 12.72 1.49 12.13 13.71 4.43 in class (3.85) (11.44) Use of textbooks in instruction (O = never; 7.92 8.33 12.90 7.94 8.39 14.30 20 = almost every lesson) (6.90) (7.65) Lecturestowholeclass(O=never, 4.22 4.17 -0.88 4.22 4.05 -3.45 5 = almost everv lesson) (1.4.3) (5.59) In-class homevwork review (0 = never; 3.88 3.94 -0.32 .3.98 3.92 0.34 5 = almost every lesson) (2.90) (3.0)5) Sttidents copy notes from blackboard 3.95 3.97 0.43 3.9.3 4.02 1.54 (0 = never; .5 = almost everv lesson) (0.98) (3 5.3) Subtotal (pedagogical processes) n.a. n.a. 11.85 na. n.a. 111.67 fabZ!le -onstinutes onl the to/llou ing paIge. ) Table 5. (continued) A Il-age and primary sch.7ools Bottom and to) 25 percent of allprimary-level schools Difference in Mean for Dilference in total Mean for Mean tor total scores of bottom 25 Mean tfor top 25 scores ot bottom 2.5 all-age prinn1ary 111-age and perCent Of perc-enit of percent and top 2 5 \Variahle scho ols sc/ ),)ols pri,mary sc/ools Schools schools percc'tt of scl/ools School-level organization, climiiate, cnd conitrol School antotnomv Relative influence of the Ministrv of Educationcomparedtotheprincipal 0.18 0.14 0.57 0.19 0.16 0.44 on the school's organizatione (ratio) (4.68) (.3.59) Principal'sinflueliceon thecurriculum 1.62 1.48 4.72 1.54 1.39 - 5.SI (range: none = 0 to high = 6) (-.66) (6.59) Teachers' iifluence on cLrriculum 2.08 2.19 2.15 2.00 2.35 7.26 (range: none = 0 to highi = 6) (4.39) (14.82) 4 W irk -centered environmnent urriculum or pedagogy first or second mnostcommon theme in staffmeetinigs 0.20 0.17 2.4.3 0.16 0.29 10.64 (I =ves; 0=no) (1.47) (6.42) Instructional assistance and leadership 8.58 8.82 (.79 6.65 1 1.42 1 5.40 by principal (hours per week) (0.64) (12.59) Index of commitment to teaching basic skillsand/orcritical thinking(range: 1.73 1.78 0.58 1.84 1.71 -1.51 nonie = 0 to high = 3) (2.50) (6.54) Average frequency of receiving help from other teachers to improve teaching 1.32 1.48 2.06 0.90 1.23 4.20 skills (range: never = 0 to daily = .5) (3.39 (6.89) Community involvement Parents who attend PTA meetings 2.5.32 29..59 -.3.50 .30.71 25.61 4.17 (percentage) (4.11) (4.91) Index ofcommunity involvement(range: 0.98 0.96 0.37 0.90 1.03 -1.78 none = 0 to high = 2) (1.05) (5(.5) Order/v er 'tronment Uniform requiremeit strictly enforced 0.30 0.41 6.10 0.23 0.45 1.3.01 I yes; 0 = no) (4.21) (8.98) Class-timilespenton discipline 11.17 11.83 2.05 10.10 11.44 4.20 (percentage) (5.27) ( 0.8]) S,chol type School operates shifts(1 = yes;(0=no) ().8 o.80 -((.18 0.94 (.71 14.19 (0.17) (12.92) Studenits grouped by ability (1 = yes; 0.65 (0.62 0.74 0.61 O.5 5 - 1.78 0 = no) (1.14) (2.75) City school, Kingston excluded (I =yes; 0.13 0.16 -0.16 0.1i 0.29 - 1.45 0= no) (1.35) (11.85) Rural school (1 = yes; 0 = no) (.75 0.46 -1.(9 0.77 0.48 -1.()8 (17.4.5) (17.25) Subtotal (school-level organization, n.a. n.a. 1.85 n.a. na. 60.40 climate, and control) All-age school iitercepts (suitoral) I 0 88.20 0.81 0.23 51.21 (55.64) (32.3 1) loal difference in scores n.a. n.a. 138.66 naa. i.a. 308.66 ( 51.64) (46.39) Note: Thc sample si7e is 35 5i. O t the 508 children for w hoi test scores w ere a ailabl e. Ic0 w ere droppe c d heaIse th attended priv ate (prepar atory ) schools, 85 were dropped because they could not be matched wit h anN school for w hich c had school and reacher questionInaires, and 5 8 were d ropped bccause of Incomplete data from the school and teacher questioiniaires. Standard err-ors are in parentheses. The standard crior is calculated as the sum of the standard error on the math contribution aiid of the standard error on the readungcotitribution ,ind ignores the covariance hetxeen the two sets of estimates (thus it is an upper bound on the true standard error). a. The index measures the available equipimient In the school. including telephone. typewriter. telev ision, computer, radio, and copying or duplicating machine. b. Includes libraries, labs, and studios. c. Includes maps, charts, science kits, and dictionaries. d. Includes pens, pencils, paper, notebooks, complete set of required textbooks, and dictionaries. e. The influence of the Ministry of Educatioii on the organization of thc school is scaled on a range from none = 0 to high = 12. The influence of the principal is scaled on a raiige from none = 0 to high = 18. Sorrrce: Authors' calculations. 2,56 1 HIF Wt)oI 1) RANK I( )NONMI( RI VIEW, V()l. ". No. schools on only four dimensions-conducting more vision tests, increasing the amount of time spent testing students, increasing the intensity of textbook use, and decreasing the amount of time spent on doing written assignments in class- could reduce the performance difference between the two by more than 40 percent. Of course, given the caveats already laid out in our discussion, these areas should not he the focus of government education policy without further researcih, perhaps in the form of randonized experiments. VII. CONCLUSION Physical and pedagogical inputs, pedagogical practices, and school organiza- tion and climate variables all appear to influence achievement among primary school students in Jamaica. Indeed, it appears that physical and pedagogical inputs, usually the focus of economic studies, may not be the most important determinanits of student achievement, because pedagogical process variables are more often significant. Our policy simulations reveal that although primary schools are objectively better than all-age schools, there are much bigger differences withiln each type of school. A randomly selected child placed in the average "bad" school (those in the bottom quartile of our school quality index) would lag almost four grade- levels behind a child in the average "good" school (those in the top quartile in quality). Meanwhile, the difference in performanice for a student in an average all-age school compared with that for a student in an average primary school is about one and a half grade-levels. Finally, our analysis suggests that concentra- tion on just a few aspects of pedagogical process and school climate and organi- zation mav lead to substantial improvement in student achievement. Hlowever, we cautioll that, as with most studies of this genre, we are unable to explain with statistical precision most of the difference betweeni student achievement in good schools compared with bad schools, and we know that our results may suffer from selectivity bias because many students were not tested. There is undoubtedly much room for school improvement in many other developing countries. Before considering reforms, we believe that it is important to collect and analyze microdata and to conduct cost-benefit analyses wherever possible. Unfortunately, our attempts at cost-benefit analvsis were stvmied bv the difficultv of establishinig costs for the pedagogical practice and school man- agement factors that were found to be most important. Of course, nearly all analyses based on nonexperimental data are subject to biases of various sorts. Whether the school improvements suggested by such analyses ultimatelv work can be confirmed (or disproved) only by experimenting with them and rig- orously measuring their effects. REFERFNCES The word "processed" describes informally reprodticed works that may not be com- monly available throUgh library systems. Gleuruwe aind otbers 257 Andersoni, Lorin W., Doris W. Ryan, and B. J. Shapiro. 1989. The IEA Classroom Environment Study. Oxford: Pergamon Press. Chubb, John E., and Terry M. Moe. 1990. Politics. Markets, and America's Schools. Washingtoni, D.C.: Brookilngs Institution. Coleman James S., Thomas Hoffer, and Sally Kilgore. 1982. High School Achievement: Public, Catholic, and Private Schools Conipared. New York: Basic Books. Creemers, Bert, Tony Peters, and David Revnolds, eds. 1989. School Effectiveness and Sc-hool improvem1ient. 2d ed. Amsterdam: Lisse; Rockland, Mass.: Swets and Zeitlinger. Ednionds, Ronald. 1979. "Effective Schools for the Urhan Poor." Educationial Leadership 37:1S-24. Elley, Warwick B. 1992. Hon', in the World Do Students Read? The Hague: The Interna- tional Association for the Evaluationi of Educational Achievemenit. Fuller, Bruce. 1987. "What School Factors Raise Achievemelit in the Third W/orld?" Review of Educational Research 57(3):255-92. Fuller, Bruce, and Prema Clarke. 1994. "Raising School Effects While Ignoring Culture?" Review of Educational Research 64(1):119-57. Glewwe, Paul, and Hariai Jacobv. 1994. "Studenit Achievemenit and Schooling Choice in Low-Income ConLitries: Evidence from Ghana." louirnal of Humtan Resources 29(3):843-64. Grosh, Margaret. 1991. "The Household Survey as, a Tool for Policy Change: Lessons from the Jamaican Survey of Living Conditions." .SMS Working Paper 80. Washing- ton, D.C.: World Bank. Hanushek, Fric. 1986. "The Fcononi,cs of Schoolinig: Production and Efficiency in Public Schools." Journal of Economic Literature 243:1141-77. Harbison, Ralph, and Eric Hanushek. 1992. Educational Performance of the Poor: Les- sons from Rural Northeast BraZil. New York: Oxford University Press. Harris, Abigail. 1993. "The Validity of the California Achievement Test as a Measure of Educational Productivity in Jamaica." Paper presenited at the Annual Meeting of the Comparative and International Education Society, Klingston, Jamaica. March. Processed. Heyneman, Stephen P., and William A. Loxlev. 1983. "The Effect of Primary-School Quality on Acadeimic Achievement across Tvventy-nine High- and Low-Income Coun- tries." The American Journal of Sociology 88(6):1162-94. James-Reid, Olga. 1989. "Resources Management and Development in Some Caribbean Primary and All-Age Schiools." In The Major Project in the Field of Education in Latin American anzdi the Caribbean. Bulletin 20. Sa.ntiago, Chile. Jimenez, Emmanuel, and Marlaine Lockheed. 1994. "Private and Public Secondary Edu- cation in Developinig Countries: A Comparative Look." Draft. Johnson, D. W., Geoffrev iMaarLyama, Roger Johnson, Deborah Nelson, and Linda Skon. 1981. "Effects of Cooperative, Competitive, anid Individualistic Goal Structures on Achievement: A Meta-analysis." Psychological Bulletin 89(1):47-62. Levin, Henry, and )\larlaine E. [.ockheed, eds. 1993. Effective Schools in D)eveloping Countries. Washingtoni, D.C.: Failmer Press. Lockheed, Marlaine E., and Adriaan M. Verspoor. 1991. Improving Primarv Education in Dei'eloping Countries. Oxford: Oxford lJniversity Press. Maddala, G. S. 1983. Iinmited-Dependent and Qualitative Variables in Econoniietrics. Cambridge, U.K.; New York: Cambridge University Press. 258 il 1U WORI 1) BANk I ONO( N R l \II.\X. V()l 9, NO 2 Miller, Errol. 1988. "Bovs and the Educational System." People 1(2):4, 11-12. Newsletter of the Population Policy Coordinating Committee of the Government of Jamaica, Klingstoni. 1990. "Contemiiporary IssLies in Jamaican Education." In Colin Brock and Donald Clarkson, eds., Ediucation in Central Amlerica and the Caribbean. London: Routledge Press. Planning Institute of jamaica. 1992. lamaica: Survl!ey of Living Conditions Analytical Review 1990. Kingston. Postlethwaite, T. Neville, and Kenneth N. Ross. 1992. Effective Schools in Reading: 17nplications for Educational Planners. Hamburg: Interinationial Association for the Evaluation of Educational Achievemeint. Psacharopoulos, George, Carlos C. Rojas, and Eduardo Velez. 1993. "Achievement Eval- uation of Colombia's Escuela Nueva: Is Multigrade the Answer." Comparative Educa- tion Review 37(3):26.37-6. Purkey, Stewart C., and Marshall S. Smitil. 1983. Effective Schools: A Review. Ele- mentary School Journal 83(4):42,7-2. Raudenbtush, Stepheni WV., SUwanna EamIIsuLkkawat, IkechuLku Di-lbor, Molianed Kamali, and Wimol Taoklam. 1992. "On-the-Job Improvements in Teacher Competence: Pol- icv Options anid Their Effects on Teachinig and lIearilinig in Thailand." Policy Research Working Paper 889. World Bank, Populatioll and HuLimani Resources Department, Washington., ).C. Processed. Robitaille. David F., and Robert A. Garden. 1989. The IFA Studty of Mathematics I: Context and ()utcnmes of School Mathenmatics. Oxford: Pergamon Press. Rutter, Michael, Barbara Maughan, Peter Mortimore, and Janet Ouston, with Alan Smithi. 1979. Fifteen Thousand Hours: Secondary Schools and Their Efftects on Chil- drenz. Cambridge, Mass.: Harvard lJniversiry Press. Trevor, Hamilton. 1990. "Jamaicani Priniarv Educationi Policy Study: Final Report." Min- istry of Education, Kingstoni. Processed. Velez, Eduardo, Ernesto Schiefelbein, and Jorge "alenzuela. 1993. "Factors Affecting Achievement in Primary School: A Review of the Literature for Latin America and the Caribbean." IIRO Working Paper 2. World Bank, Humnani Resources Development and Operation Policy Department, Washington, D.C. Processed. Walberg, Herbert1. 1982. "What Makes Schooling Effective?" Contemporary Education Review 1:1-34. 1986. "Synthesis of Research on Teaching." In Merlin C. Wittrock, ed., Hand- 0ook of Research (on Teaching, 3rd ed. New York: Macmillan: 214-29. Wirtrock, Merlin C., ed. 1986. Handbook of Research on Teaching, 3rd ed. New York: Macmillan. World Bank. 1990. Primnary Education. A Wor-ld Bank Policy Paper. Washington, D.C. 1993. World Development Report 1993: Investing in Health. New York: Oxford Universitv Press. I11 A () R I 1) K A \ K I t O N () N1 I( R F V I I A.\\ S) I 9. N ( . I.' 2 S 9 - 2 K Natural Resource Management and Economywide Policies in Costa Rica: A Computable General Equilibrium (CGE) Modeling Approach Annika Persson and Mohan Munasinghe A computable general equilibrium (CGE) model is used to trace the effects of govern- ment policies on Costa Rican forests in the presence of incomplete markets. The results indzcate that correcting the market failure would, as expected, reduce deforestation. More interestingly, in the presence of the marketfailure, lowering the tax on unskilled labor reduces deforestation because people gain employment in other parts of the economy. Tixatron of other produced goods chaniges the incentives for deforestation. For example, a tax on agricultural products elevates the relative price of capital and shifts resources auayfrom the capital-intensive industrial sector toward the agricultural and forest sectors; aIs a result, such a tax increases deforestation. Currently, the economic analysis of environmental issues relies mainly on project-level studies, which use cost-benefit analyses and environmental assess- ments. However, economywide policies (both macroeconomic and sectoral) fre- quently have much more powerful environmental effects than mere project-level investments. Some progress has been made in identifying the environmental consequences of sectoral policies involving, for example, energy, water, or agri- cultural pricing. Nevertheless, the impacts of broad macroeconomic reforms (such as exchange rate devaluation, trade liberalization, privatization, and other fiscal and monetary stabilization policies) on natural resource and pollution management are far more difficult to trace. This difficulty hampers efforts to design sustainable development strategies that meet economic, social, and envi- ronmental criteria in a more balanced way (Munasinghe 1993). In the case of the World Bank, for example, the general lack of knowledge about links between economic policies and the environment has delayed at- tempts to gradually expand the application of environmental analysis to cover economywide or policy-based lending. Such lending is the second largest use of Bank resources (about $5.8 billion' annually, or 27 percent of total lending in 1. A billion is 1,(0( million. Annika Persson and Mohan Munasinghe are with the Environment Department at the World Bank. C- 1995 The International Bank for Reconstruction and Development/THE WORLD BANK 2 59 260 TilL 'WROk IDBANK [( ONO MI(: RFVI I`W VVOl ). NO .' 1993). Lack of knowledge about links has also hampered efforts to develop more effective National Environmental Action Plans. These plans are prepared by borrowing countries, with Bank assistance, to help determine priority activ- ities in order to address national environmental issues. A recent study argues that many instances of environmental damage are caused by market failures and policy distortions and exacerbated by unemploy- ment, landlessness, and poverty (Munasinghe, Cruz, and Warford 1993). Therefore, broad policy reforms, which usually promote efficiency or reduce poverty, also should be generally beneficial for the environment. Some of these reforms, however, may have negative environmental effects depending on preex- isting (and often localized) constraints, such as inadequately defined property or resource rights. It is important for decisionmakers to be able to trace the complicated paths by which policy changes at the macro level ultimately affect incentives to use re- sources efficiently at the micro level of firms or households. The objective is not necessarily to modify the original broader policies (which have conventional economic or poverty-related goals), but rather to design more specific or lo- calized complementary measures that remove economic distortions or con- straints (Munasinghe and Cruz 1994). These additional measures would help to mitigate the negative effects or enhance the positive effects of the original poli- cies on the environment. Such complementary actions would include both market-based approaches (such as Pigovian taxes on environmental externalities or the allocation of limited pollution rights coupled with marketable permits) and nonmarket methods (such as command-and-control techniques). The ideal approach is a general equilibrium analysis that traces both the economic and the environmental effects of economywide policy reforms. When such comprehensive methods are not possible, as in developing countries where data and skills are relatively scarce, partial equilibrium approaches that help to identify the most important impacts of economywide policies are frequently used. Because the full consequences of a policy are not traced, both quantitative and qualitative results of the partial equilibrium model may be wrong. For example, taxes that are not lump sum may carry over from the sector for which they were intended into other sectors of the economy and affect consumption and production decisions there as well. In this context, the main purpose of this article is to investigate the effects of economywide policies in Costa Rica on forest areas and the environment as described in Persson (1994). We also seek to determine whether new measures involving the allocation of property rights to these forests will yield different results when analyzed by using a general equilib- rium model instead of a more conventional, partial equilibrium approach. Deforestation and soil erosion are major environmental problems in Costa Rica. Some data on forest clearing over time are shown in table 1. To evaluate how sectoral and economywide policies can help to control deforestation, the computable general equilibrium (CGE) model used here highlights the economic activities and factors that specifically affect deforestation in Costa Rica. The Persson and Munasinghe 261 Table 1. Percentage of Total Land Area in Forests and Agriculture, Costa Rica, Selected Years Sector 1963 1973 1986 Forestry 67 57 40 Agriculture 30 40 57 Source: Sol6rzano and others (1991). model goes beyond standard approaches in two important respects. First, it can simulate the effect of introducing property rights on forest resources and thus encourages sustainable management of forests by private individuals who value future returns to forestry. Second, it includes markets for logs and cleared land: loggers deforest to sell timber to the forest industry and for exports, and squat- ters clear land for agricultural production and for sale to the agricultural sector as the latter expands and requires more land. The model retains features that are fairly standard in most CGE models. The sectors that produce tradable outputs-forestry, agriculture, and industry-are price takers in the world market, whereas infrastructure and services produce nontradable output. To focus on the natural resource sectors, the domestically mobile factors include-aside from capital and (skilled and unskilled) labor- cleared land and logs. The supplies of both labor and capital are exogenous. The demand for these factors arises from the producing sectors (agriculture and industry) and from the deforestation activity of loggers and squatters. In the model the supply of cleared land is initially based on Costa Rica's total land area that has been deforested. Additional cleared land is made available from in- creased deforestation. The rate of land clearing depends on the definition of property rights and on taxes (or subsidies) that affect forestry and agriculture, and the expansion of squatting activities augments the cleared-land factor. Agri- cultural production provides the demand for cleared land. Poorly defined property rights in Costa Rican forests play an important role in deforestation. The model indicates that correction of this market failure would reduce deforestation. When property rights are well defined and the interest rate is exogenous, the value that loggers assign to preserving the forests is crucial: to stop deforestation, the benefits from preserving the forests must be significantly higher than the value of the logs and the cleared land. In the model, tax policies may generate unexpected side effects, and substitu- tion effects between inputs in the producing sectors may be important. There- fore, when possible impacts of macroeconomic policies are investigated, the general equilibrium approach generates results that are different from those derived from a partial equilibrium analysis. Section I discusses the extent and causes of deforestation in Costa Rica. Sec- tion 11 summarizes other CGE models applied to environmental problems. The general features of our CGE model are discussed in section 111. The base case data and the assumptions and limitations of the model are presented in section IV. 262 ITHI :WRII) BANK F(:ONOMI(: RLVIFW, Vol.. ), No 2 Section V summarizes the chief results, and section VI presents the conclusions of the study. The appendix provides a summary of the structure of the model. 1. THE STATUS OF FORESTS IN COSTA RICA Deforestation in Costa Rica is proceeding at a rapid pace, and concern is growing both inside the country and among environmental organizations in the rest of the world. Quesada Mateo (1990) mentions the following economic and ecological benefits that Costa Rica may lose if deforestation continues: access to construction materials and other wood products; species of plants and animals that have potential uses for consumption and industrial production; recreation and ecotourism; control of erosion and sedimentation; and education and re- search possibilities (see, for example, Hartshorn and others 1982). The green- house effect and concerns about the rich biological diversity in Costa Rica may be important to other countries and environmental organizations. Deforestation and erosion are the main environmental problems in the coun- try (Blomstrom and Lundahl 1989; Foy and Daly 1989). According to Blomstrom and Lundahl 1989, originally most of Costa Rica was forested, but in 1977 only 31 percent (16,000 kM2) remained covered with forests. They estimate that in 1983, 14 percent of the area was still covered with forests. Sol6rzano and others (1991) give the more conservative estimate that about 40 percent of the land is still covered with forests. This difference in estimates is probably caused by differences in what types of forests were investigated. The lower estimates probably concern only primary forests, whereas the higher esti- mates include secondary forests and intervened forests (Sader and Joyce 1988). Most of the deforestation has occurred since 1950. If deforestation continues at the current rate, the commercial forests of Costa Rica will be exhausted within the next five years. The life zones with the highest rates of deforestation are the tropical wet forests, which are also the life zones with the highest levels of biodiversity (Sol6rzano and others 1991). Carriere (1991a, 1991b) describes the process of deforestation as taking place in several stages. First, a logging company involved in high-grading (that is, extracting only the most commercially valuable species) clears a vehicle tract to extract lumber. Thereafter, the road is improved by the government under pres- sure from lobbying groups. This in turn enables local peasant families to clear and use the remaining forest for subsistence agriculture until the decreasing yields force them to sell the land, if it is titled, or abandon it. However, the land is still suitable for pasture and is therefore packaged commercially by urban- based real estate companies and sold to cattle ranchers. After a few years, the land is almost completely degraded and unsuitable for any kind of economic use. This view is shared by Keogh (1984). The Costa Rican government is taking steps to preserve the forests. More than 13,000 km2 of forests have been designated as national parks, although de- forestation in the past had been encouraged to diversify the country's production away from coffee and banana crops (Biesanz, Biesanz, and Biesanz 1987). Perssont and Munasinghe 26.3 Four groups are responsible for deforestation in Costa Rica (Lutz and Daly 1990). First, the timber industry may be responsible for deforesting as much as 20 hectares annually. Logging requires a special permit from the government, but about half of the trees are cut illegally. Domestically cut logs are processed locally and are typically used in construction. Exports of wood and wood prod- ucts are small, and imports are negligible. The current import tariff on logs is 5 percent (Lutz and Daly 1990). Efficiency in forestry is low, and only a few species are commercially utilized. About 54 percent of the logs are processed, and of these about half finally reach the market (Quesada Mateo 1990). The main part of the logs used in the timber industry are bought from sources other than the industry itself. Second, banana firms and other companies are expanding their plantations rapidly. The main products cultivated in Costa Rica are rice, coffee, fruits, sugar cane, beans, maize, and sorghum (Hugo and others 1983). Lutz and Daly (1990, p. 6) state that erosion is visible in some areas but that farmers "do not produce in obviously unsuitable ways to destroy the environment. . . . For example, living fences are widely used, which reduce erosion, and protective forest cover is left intact next to creeks, on contours or steep slopes, etc." Third, in recent decades cattle ranchers have expanded their activities rapidly at the expense of forested areas. This type of land conversion, however, may now be limited because most of the land that can be sustainably used for pasture has already been cleared. In the 1950s and 1960s, investment in cattle increased significantly, encour- aged by foreign aid and investment and also by government aid as credit and provision of infrastructure. The increase in cattle ranching caused a rapid in- crease in deforestation. The pasture trend boomed in the 1970s, but since then profits have decreased. More than 70 percent of the farmland is in pasture, only 2.5 percent is in coffee, and 1.1 percent in bananas (Biesanz, Biesanz, and Biesanz 1987). Fourth, squatting is taking place on both privately owned and government land. Some of the squatters produce agricultural outputs, but others sell the cleared land to cattle ranchers or other landowners. Those who buy "in good faith" from squatters are not prosecuted. About twice as much is paid for cleared land as is paid for forests. Squatting is an important cause of deforestation in Costa Rica. By clearing the land, it is possible to get formal owvnership to the land (Blomstrom and Lundahl 1989) or in some cases at least to the "land improvements." Squatting by small- holders nowadays appears to constitute a less significant part of deforestation in an overall context, although it may be locally important (Lutz and Daly 1990). If ownership may be obtained at no other cost than that of clearing the land, the forests can be seen as a type of common property and the cleared land as private property in the traditional sense. However, Costa Rica is not a typical case of undefined property rights in which there is an open-access resource (see, for example, Dasgupta 1982); rather, Costa Rica is a case of insecure land 264 THF WORLD IBANK F( ONOMII RFVIEW. VO1.9. No. 2 tenure. This implies that there is no crowding effect on the stock of the resource, which is what occurs when each agent maximizes personal profit without taking into consideration the effect on the stock of the resource. Instead of the traditional open-access problem, there is a form of short-term property rights when deforestation occurs, but the property rights to the stand- ing forests are not protected. Because of the structure of property rights, the logger or squatter will continue deforestation only until the marginal cost of deforestation equals the marginal revenue. The social cost of deforestation will then be higher than the private cost because "the world's" willingness to pay for the preservation of the Costa Rican forests will not be included in the private cost. Chichilnisky (1993) shows that the difference in cost functions is a major cause behind deforestation in the developing world. Chichilnisky develops a north-south trade model in which the difference in trade patterns between north and south is explained by a difference in property rights. Thus, deforestation would be driven by the difference in private and social objectives. For example, the loggers' main interest may be the profitability of the logging operation itself without much consideration about future, alternative uses of the land. Another critical economic factor may be the existence of higher private dis- count rates; that is, deforestation may be caused by discounting the future value of the forests. A high discount factor implies that future gains from the forests have much less value than the gains from deforestation today. The benefits of tropical forests are often more significant in the long term than in the short term. However, the regenerative capacity of tropical forests is low, and the discount- ing of future environmental benefits may often make it more profitable to har- vest forest resources as quickly as possible. Forest investments, such as replant- ing, take a long time to yield returns, and individuals therefore find little attraction in conservation and reforestation activities. In many developing coun- tries, private market rates are very high and often exceed the rate that would be socially justifiable (Barbier, Burgess, and Markandya 1991). Poor people often face even higher discount rates because of credit constraints. A final factor to be considered is that the Costa Rican tax structure for income and property taxes is regressive. Sales taxes and other indirect taxes constituted 70 percent of the total tax revenues in 1970, and there are indications that this figure may still be high. Although property taxes are low (in some cases about 1 percent of the actual market value), property and income tax evasion is a prob- lem that costs the country approximately 100 billion colones a year. Remedies may include raising the price of land by increasing land taxes, increasing tax collection rates, and prosecuting tax evaders more effectively. II. THE MODELING APPROACH TO ENVIRONMENTAL PROBLEMS As may be concluded from the previous section, the main reasons for de- forestation and thereby erosion are as follows: Persson adtd Munasingbe 26.5 * The price of land is too low because the total social opportunity value of the rain forests is not included. * Undefined property rights make the private cost of deforestation lower than the social cost of deforestation. * Discount rates are too high. This implies that the value of future gains from the forests is deemed to be lower than that of the gains from deforestation today. * Economywide policies, for example, those defining the tax system, may cause deforestation. CGE models have been applied before to environmental problems, mainly to evaluate issues involving air pollution and pollution taxes. Bergman's (1990a) model is designed to simulate the effects of environmental regulation and energy policy on the Swedish economy. The environmental market failure is in this case corrected by the creation of a market for emission permits. The cost of emission permits for carbon dioxide, nitrogen, and sulphur is incorporated in the cost functions. Jorgenson and Wilcoxen (1990) analyze the economic impact of environmen- tal regulations on the U.S. economy by simulating long-term growth with and without environmental regulations. The share of abatement costs in total costs is estimated for each industry, as is the share of investment in pollution control equipment and the cost of pollution control devices in motor vehicles. The model is run with and without these costs to estimate the economic impacts. There are not many examples of CGE models dealing with the impact on the economy of overexploitation of natural resources. The London Environmental and Economics Centre (1992) constructs a model against the background of environmental problems in Thailand. The sources of the environmental issues are economic growth, exchange rate problems, and government policies (on such matters as agriculture, taxation, and land tenure) that promote deforesta- tion. The approach implies that each sector produces, for example, a fixed amount of air pollution or deforestation. The environmental impacts are not part of the model per se; that is, the environmental degradation or improvement is not fed back into the model so as to affect future production and consumption decisions. Some of the findings point out that export taxes on rice and rubber increase investment in soil conservation, increase the use of agrochemicals, and shift land cultivation from rubber to rice. Not much work has previously been done on the modeling of undefined property rights in a general equilibrium context, where the results may differ from those of a partial equilibrium model. Devarajan (1990) suggests incor- porating a partial equilibrium model in the general equilibrium framework. He suggests removing the first-order condition that labor be paid the value of its marginal product in some sectors and replacing that condition with one that reflects the suboptimal behavior of the sector. This new condition will make possible an analysis of how policy interventions in the system affect deforesta- 266 THF.WORI I) iANKF :ONO.MI1(REVIEW, V(.I .,No 2 tion, but the model has to be dynamic in order to take account of both the stock and the flow effects of deforestation. Unemo (1993) models the suboptimal use of land in Botswana, caused by overgrazing of cattle, as a result of undefined property rights. Land is seen as an open-access resource, and the effects of overgrazing are incorporated as crowd- ing effects in the cattle owner's production function. The quantity of output is determined not only by the number of cattle the individual owns but also by the whole population of cattle grazing on the land. One of the findings shows that a fall in the price of diamonds considerably increases pressure on land because mining becomes less profitable than cattle ranching. In order to model property rights-related behavior in Costa Rican forests, it is assumed that the private cost of deforestation is lower than the social oppor- tunity value of the forests when property rights are undefined. When property rights are defined, the social value of the rain forests is incorporated in the utility functions of the squatters and therefore in the private cost of deforestation. This approach facilitates the analysis of the role of undefined property rights and follows the approach used by Chichilnisky (1993). 111. GENERAL FEATURES OF THE MODEL This model is a static CGE model of an open economy, even though it has certain implicit dynamic features, such as the discount rate included in the future valuation of forested land. The model differs from standard CGE modeling be- cause of the inclusion of undefined property rights and of the modification of the way in which the markets for logs and cleared land function. The model has two types of sectors. The tradables-producing sectors (for- estry, agriculture, and industry) are assumed to be price takers on the world market in the standard Heckscher-Ohlin fashion. The nontradables-producing sectors are infrastructure and services. In addition, there are two sectors that clear land: logging and squatting. Loggers clear land to obtain logs for the forest industry and for exports, and squatters clear land to sell it to the agricultural sector. The domestic intersectorally mobile production factors are unskilled labor, skilled labor, and capital. Logs and cleared land are specific to forestry and agriculture, respectively, although logs can be traded on the world market. No reafforestation is possible in the model. The key elements of the CGE model are introduced next; a more detailed mathematical description is given in the appendix. The main linkages in the model are shown in figure 1. Factor Market Equilibrium and the Stock of Land The supplies of both labor and capital are assumed to be exogenously given, and for factor markets to clear, these supplies must equal the demands for labor and capital, respectively. Demands arise from the producing sectors in addition Persson and Munasinghe 267 Figure 1. Main Linkages of tbe CGE Model Land l lo)gs r Lab,or Factor demand capital Factor supply Fac tor C i ~~~~~~i markets Factor Wages, ecosts rents Dernandl for f _ ~~~~intermediate goods oducers F Households tStales ConsLumer , revenuies expenditures < C ~~~~~~Pro(let 'iupply of C.onsumption of procluce(l goodIs Imports Exports final goods ) ~~~~) to the amounts used for deforestation by squatters and loggers. The demand for each production factor (such as capital or labor) within both the tradables- producing and the nontradables-producing sectors, and in the deforestation sectors, is given by the partial derivative of the cost function for the relevant sector with respect to the price of the same production factor. Both loggers and squatters generate demand for unskilled labor for deforestation, but only loggers generate demand for capital. Costa Rica's total area has been divided into two types of land: cleared and forested. Cleared land is produced through deforestation. The amount of cleared land produced depends on the definition of property rights, on taxes and subsidies on the factors of production, and on profits in forestry and agriculture. Logs are assumed to be tradable. Therefore, the demand for forest land by the logging sector and the world market price determine the rate of deforestation. This demand is equal to the partial derivative of the logging cost function with respect to the user cost of logs plus the net export of logs. The supply of cleared land is composed of the stock of cleared land plus the land deforested by squatters. The demand for cleared land is the demand by the agricultural sector, which is set equal to the partial derivative of the cost function of the agricultural sector with respect to the user cost of cleared land. 268 YHF WORI 1) BANK F.( ONO:MI RFVILW. Vol 9, NO.2 The combination of production factors can be influenced by taxes and sub- sidies. Thus, a given user price will be greater by a percentage tax than the corresponding supply price or smaller by a percentage subsidy. Technology, Costs, and Producer Behavior The production factors have been aggregated into a composite input. Cleared land is combined with capital to yield an aggregate, which in turn is combined with logs to generate an aggregate for land, capital, and logs. Skilled labor is added to that combination and is combined with unskilled labor to yield the composite factor input. This aggregation is accomplished through the use of constant-elasticity-of-substitution (CES) production functions. The technology is specified to exhibit constant returns to scale. The relation between inputs and output is given by typical Leontief production functions for each sector. Because the technology exhibits constant returns to scale, the marginal cost and the average cost of production in a given sector can be expressed as a linear function of prices, relevant input/output coefficients, and indirect tax rates. Producers are assumed to maximize profits. The producer output prices in the tradables-producing sectors are given by the world market prices. Assuming perfect competition, pure profits are nonpositive and output is nonnegative and positive only if pure profits are equal to zero. In the nontradables-producing sectors, the sector-specific capital is endogenously adjusted so that price equals marginal cost. Prices, Domestic Demand, Foreign Trade, and Market Clearing For a good produced in the tradables-producing sectors, the domestic pro- ducer price is equal to the world market price of the identical good. In the nontradables-producing sectors, the domestic user price is equal to the producer price times the tax rate. The intermediate demand of a good is given by the technology assumptions. Domestic final demand is given by a linear expenditure system, derived from the consumers' utility maximization. To equilibrate the market for a good, the net export for that good is defined as the difference between domestic supply and demand. Deforestation Sectors The two sectors responsible for deforestation, logging and squatting, interact with the rest of the economy through their demands for capital or labor, or both; through their supplying forest products and cleared land to the rest of the economy; and through changes in the relative prices of factor inputs and sectoral outputs. Logging. Logging is assumed to have a capital-intensive technology (Repetto 1988). In addition, the technology is assumed to exhibit decreasing returns to scale in order to reflect the diminishing amount of available forests and the fact Persson and Munasinghe 269 that much of the logging is done illegally. The production of logs is assumed to depend only on two factors of production: labor and capital. A log-linear pro- duction function is used. The technology used to model the diminishing yields in deforestation exhibits decreasing returns to scale; this implies that the returns to the production factors fall with increased deforestation. Deforestation for land and deforestation for logs are assumed to be independent of each other, and therefore the increased deforestation for logs does not affect the returns to deforestation for land. Similarly, returns to deforestation for land do not affect increased deforestation for logs. However, increased deforestation for logs im- plies decreasing yields for loggers and increased deforestation for land implies diminishing returns for squatters. In the case of undefined property rights, loggers take only the private cost of deforestation into account. When property rights are well defined, the oppor- tunity value of saving the forests is included in the loggers' cost function. Squatting.2 The forested land cleared by squatters is seen as common prop- erty, but because the stock of forested land is not included in the squatters' production function, there is no crowding effect. The base case assumes unde- fined property rights. The squatters' production function for cleared land increases monotonically with labor inputs. The squatters' total revenue from land clearing is the price paid for the cleared land. Part of the land cleared by squatters is sold to the agricultural sector, and the rest is used for subsistence agriculture by the squat- ters themselves. However, because both activities (selling of cleared land and using it for subsistence agriculture) occur, the returns at the margin must be the same in each case. The squatters are assumed not to sell the timber from their deforestation. Other uses of the timber, such as for firewood, are assumed to be negligible. When property rights are undefined, the squatters' total private cost for land clearing depends only on the amount of labor needed to clear the land. This private cost does not include the future value of the forests and the cost of environmental damage. Therefore, the total social cost of deforestation is this private cost plus the future benefits from cleared forests that are foregone by clearing the land today. The future value of the forests is assumed to be greater than the value of the forests today. The analysis of the definition of property rights can then be accomplished through the simulation of two regimes. In the case of undefined property rights, the present-day squatters do not take the future value of the forests into account. When property rights are well defined, the squatters own their land and do take the future value of the forests into account. The owners of forested land (that is, squatters who are aware of the future), decide whether to preserve the forests or clear land. 2. The section on squatting is inspired mainly by Johan%son and Lofgren (1985). 270 TII U. WORID) IANK E(:>NOMI(I Rl I IVW, V' )1. NO. 2) When property rights are undefined, the squatters have no market avail- able for their forests. A simple partial model of land clearing by squatters (in which each squatter receives an equal share of the private profits) is used to show that land will be cleared until marginal cost equals marginal revenue. This result corresponds to a maximization of private profit, given the insecure land tenure. When property rights are well defined, the forests do have a market. The squatters take the future value of the forests into account, and they can choose either to clear forested land or to preserve the forests. This is consistent with the condition for socially optimal forestry-that a tree should be harvested when the market value is equal to the shadow value (Hellsten 1988). This result corre- sponds to the optimization of net social benefits. It can be deduced from the foregoing that more land is cleared when property rights are undefined than when property rights are well defined. This is because the squatters' marginal cost of deforestation is lower when property rights are undefined than when property rights are well defined and the cost includes the future value of the forests. A more detailed analysis of the supply function indicates that when property rights are well defined, deforestation is increased by a change of technology toward more efficient use of labor in the production of cleared land, by an increase in the time preference rate, or by an increase in the supply price of cleared land. Conversely, deforestation is reduced by increases in the future value of the forests or the price of labor. When property rights are undefined, land clearing is not affected by the future value of forests and the rate of time preference. The effects of other variables are the same as when property rights are well defined. The profit maximization condition for the squatters in the general equilibrium model includes a term reflecting the opportunity value of saving the forests for alternative uses or for deforestation at a later time. When property rights are undefined, the weight given to this term is zero because the future tenure of the forest is uncertain. When property rights are well defined, this term is included in the profit maximization. Macroeconomic Closure and Measures of Welfare The current account is assumed to be constant, and the current account surplus is defined as the sum of net exports. There are three welfare measures in the model: the disposable income (which is implicitly determined from the cur- rent account), the green gross domestic product (green GDP), which is deter- mined as the sum of factor incomes plus a term that diminishes with increased deforestation to reflect the negative welfare effects of deforestation), and utility (which is determined from the consumers' utility function). Utility maximization results in a linear expenditure system for goods, based on a transformed Cobb- Douglas utility function. Persson a7nd Mintasinghe 2 71 IV. BASE CASE DATA, ASSUMPTIONS, AND LIMITATIONS OF THE MODEL The data used in this version of the model were originally drawn from Bricefio (1986) and the Costa Rica National Accounts (Banco Central de Costa Rica 1990). However, because the sectors of production were not consistent between the two studies, these data were adjusted in Ravent6s (1990). The input-output matrix in table 2 was calculated by Sol6rzano and others (1991) from the disaggregated data used in Ravent6s (1990). The remaining differences were added to the net exports column by the authors. Land-use data are shown in table 3. The economic rent to timber shown in table 2 was calculated from data from Sol6rzano and others (1991). Deforestation in 1986 (see table 2) was assumed to equal average deforestation between 1973 and 1989. The value in 1986 prices was calculated by using the increase in the consumer price index between January 1985 and December 1986.1 The rent to the production factor "cleared land" was subtracted from the rents to capital in the agricultural sector, and the labor used for land clearing by squatters was subtracted from the labor used in the same sector. The labor and capital used for logging were subtracted from the payments to those factors in the forest sector. Logging was assumed to be responsible for half of total deforestation, and land clearing by squatters was assumed to be respon- sible for the other half. No estimates of elasticities of substitution between production factors were available. It is reasonable to assume that they are imperfect substitutes, and all substitution elasticities were therefore assumed to be less than one. As a base case, the substitution elasticity between land and capital in agriculture was set at 0.5. The substitution elasticity between the "cleared land and capital" aggregate and logs was assumed to be 0.8 in forestry, and the substitution elasticity be- tween the "cleared land, capital, and logs" aggregate and labor was set at 0.8 for each producing sector. Compared with other studies, such as Bergman (1990a, 1990b), these values appear reasonable. The remaining elasticities concern ag- gregates involving land and logs in sectors that cannot use those factors as inputs in production; therefore, the shares of those inputs in production always have to be zero. Those elasticities were set to zero, which is consistent with a fixed- coefficient (Leontief) technology. The parameters in the production functions for squatters and loggers are judgment-based estimates, assuming a labor-intensive technology for the squat- ters and a capital-intensive technology for loggers. In concluding this section, we note several limitations in the data and model formulation. First, because of the various data adjustments, the results of the simulations are mainly indicative and not necessarily precise quantitative measures. 3. Personal commUnication with Pedro Ravent6s, INCAE, Costa Rica, Julv 1991. Table 2. Base Case Data for the CGE Model, Costa Rica, 1986 (billions of colones) Sector and item Forestry Agriculture Industry In/rastructure Services Consumption Net exports Total Sector Forestry 0.003 0.022 0.391 0.002 0.000 0.124 0.011 0.552 Agriculture 0.004 4.033 2.488 0.000 0.000 3.535 3.924 13.984 Induistry 0.137 1.405 7.390 3.418 1.343 14.426 -9.643 18.477 Infrastructure 0.004 0.293 0.826 0.647 0.684 8.366 0.000 10.821 Services 0.038 0.602 1.546 1.487 2.160 11.230 0.000 17.062 Land 0.000 2.070 0.000 0.000 0.000 0.000 0.000 2.070 v4 Logs 0.022 0.000 0.000 0.000 0.000 0.000 (.000 0.022 r2 ] Capital 0.176 2.052 2.168 1.633 5.256 0.000 0.000 11.285 Unskilled labor 0.141 2.714 1.525 1.754 1.439 0.000 0.000 7.573 Skilled labor 0.002 0.045 0.999 1.299 4.904 0.000 0.000 7.250 Indirect tax 0.025 0.747 1.145 0.580 1.276 0.000 0.000 3.773 Total 0.552 13.984 18.477 10.821 17.062 .37.681 -5.708 92.870 Note: The subsectors of the input-output table are aggregated into the five production sectors as follows: forestry (forestry and fishing); agriculture (bananas, unprocessed coffee, sugar cane, cacao, basic grains, cotton, tobacco, livestock, other agricultural products, coffee processing, grains milling, sugar refining); industry (meat and milk, fish rinning, edible oils, bakeries, other manufactured goods, drink, tobacco products, textiles and clothing, leather and shoes, timber and furniture, paper and printing, chemical products, oil refining, tire products, plastic and rubber, glass and ceramic, construction materials, metal products, electric products, transport equipment, other manufacturing): infrastructure (construction, transport, electricity, gas and water); and services (banking and finance, commerce, ow% nership of dwellings, general government). The GNP is 26.148 billions of colones. Source: Authors' calculations froni Ravenr6s (1990) using Briceno (1986), Solorzano and others (1991), and Banco Central de Costa Rica (1990). The adjustments have been calculated from Solorzano and others (1991). Table 3. Land Use Data, Costa Rica, Selected Years 1963 1973 1986 Area (sqluare Perc-entage of Area (square Percentage ot Area (square Percentage of Land use kilometers) total land kilometers) total land kilometers) total land Agriculture 1,544.796 30.09 2,048.512 39.90 2,944.616 57.36 _4 Primary forest 3.154.280 61.44 2,666.005 51.93 1,760.622 34.30 Seconidarv forest 299.011 5.82 283.571 5.52 292.850 5.70 Other 135.593 2.64 135.593 2.64 135.593 2.64 Total 5.133.680 100.00 5.133.681 100.00 5,133.681 100.00 Nwte: Percentages may not add up to 100 because of rounding. Source: Authors' calculations based on dara from Sol6r7an0 and others (1991). 274 THF WORII) hANK F:(ONO%MI( RFVIEW, VOl .9, NO.2 Second, because the model developed here is essentially static, the results are comparative snapshots of different policy experiments. A more dynamic version of the model is being developed, in which the stock and flow effects are taken into account and valid results are derived for a longer-term planning horizon. Third, the approach developed here does not include some other possible linkages with deforestation. Migration and population growth are two causal factors that may be important (Harrison 1991), but they are not investigated. Furthermore, the model neither allows for reafforestation nor includes erosion and other external effects of deforestation. To make an economic valuation of such environmental effects (that is, to incorporate them in the conventional economic analysis) would be a formidable task (for examples of valuation in developing countries, see Munasinghe 1993). V. RESlULTS Simulations of different policy experiments and a condition of well-defined property rights generated some results that are different from what could be expected from the partial equilibrium framework discussed earlier. These differ- ences are caused by substitution effects in the producing sectors. Today's situation, with undefined property rights, was taken as a base case. As a first step, property rights were defined, and the opportunity value of the forests (the H-value) was set 28 percent higher than the value derived from deforestation (Sol6rzano and others 1991). The discount rate was set at 10 percent. The results are displayed in table 4. A comparison of the first and second columns shows that defining property rights results in a dramatic de- crease in deforestation and an increase in the net import of logs (not shown in the table). Activity in the forest sector increases significantly because of the increase in imports of logs. Because deforestation by squatters ceases, activity in the agricultural sector declines. The decline, however, is smaller than in forestry because the existing stock of land remains. The welfare measures remain con- stant because consumption of different goods is unchanged. Sensitivity analysis (the third and fourth columns in table 4) shows that even when forests have a relatively small opportunity value (for example, an H-value of 0.0792), deforestation decreases dramatically. However, for deforestation to cease completely, a high opportunity value (an H-value of 0.4792) is required. Both the opportunity value of forests and the interest rate are exogenous to the model. Varying the interest rate while keeping the opportunity value fixed shows that high interest rates promote deforestation and that deforestation promotes high interest rates. The results of varying the interest rate may be deduced from table 4 because a decrease in the interest rate is equivalent to an increase in the opportunity value. The equivalent of the results in table 4 in terms of interest rates is shown in figure 2. We can conclude that although deforestation increases with the interest rate, the relationship is not linear. Persson and Munasinghe 275 Table 4. Effect of Future Valuation on Value of Production (billions of colones) Property rights Defined (H-value)- Item Undefined 0.4792 0.2792 0.0792 Deforestation Logging 0.020 0.000 0.002 0.010 Squatting 0.020 0.000 0.000 0.000 Total 0.040 0.000 0.002 0.011 Production Forestry 0.552 0.713 0.711 0.691 Agriculture 13.984 13.876 13.876 13.879 Industry 18.477 18.416 18.417 18.424 Utility 0.232 0.232 0.232 0.232 Green GDP 31.962 31.972 31.971 31.971 Disposable income 37.681 37.679 37.679 37.679 a. H-value is the future value per unit of forest. Source: Authors' calculations. Figure 2. The Effects of Changes in Interest Rates on7 Deforestation When Propertv Rigbts Are Well Defjined Value of deforestation (billions of colones) 0.012 - 0.010 - 0.008 - Loggr / 0.006 - 0.004- 0.002 2 Squatters 0- / ~ ~ ~~ ~~ ~~~~~ IIII 0) 1 20 30 40 S0 60 Interest rate (percent) Source Authors' calculktions. Table 5. The Effects of Taxes and Subsidies on Production Factors (billions of colones) Logs Land Unskilled labor Capital With With With With With With With With Item Base case tax subsidy tax subsidy tax subsidy tax subsidy Deforestation Logging 0.020 0.000 0.018 0.020 0.019 0.029 0.013 0.036 0.011 Squatting 0.020 0.044 0.004 0.000 0.255 0.210 0.000 0.020 0.020 Total 0.040 0.044 0.023 0.020 0.273 0.239 0.013 0.056 0.031 Production Forestry 0.552 0.000 1.562 0.672 0.000 0.000 0.689 0.515 0.574 Agricultuire 13.984 14.204 13.748 13.877 15.122 14.922 13.877 1.3.987 13.983 Industry 18.477 19.248 17.017 18.430 17.656 17.946 18.424 18.487 18.471 Utility 0.232 0.243 0.214 0.231 0.226 0.229 0.232 0.232 0.233 Green GDP 31.962 32.017 31.848 31.960 31.757 31.797 31.965 31.943 31.958 Disposable income 37.681 37.748 37.561 37.678 37.592 37.620 37.679 37.680 37.682 Source: Authors' calculations. Persson and Munasinghe 277 The effects of taxes on logs, land, unskilled labor, and capital are summarized in table 5. A tax increase of 10 percent on logs generates predictable results, with no deforestation from loggers and no production in the forest sector. Re- sources are shifted to the agricultural sector, with an increase in deforestation for land and an increase in total deforestation. The increase in total deforestation can be explained by lowered prices of capital and unskilled labor as a result of discontinued production in the forest sector. The tax increase actually results in a higher level of utility, as well as an increase in green GDP. A tax decrease of 10 percent on logs generates somewhat surprising results, however, because total deforestation, as well as deforestation from both squat- ters and loggers, decreases. The increased prices of unskilled labor and capital reduce domestic deforestation, but because logs are imported at the world mar- ket price, production in forestry increases. The increase in production in this sector is offset by reduced production in the agricultural and industrial sectors. The reduction in the size of the agricultural sector reduces deforestation by squatters. Utility, income, and green GDP are reduced. Taxes and subsidies on land generate expected results: a corresponding change in deforestation by squatters and roughly constant deforestation by loggers. Both the tax and the subsidy are distortionary and reduce utility, income, and green GDP. A tax increase of 10 percent on unskilled labor adversely affects the forest sector, but logging continues and logs are exported. The price of unskilled labor in the sectors that cause deforestation (that is, logging and squatting) is actually reduced, because those sectors are considered to be "informal" in the sense that their activities are to a large extent illegal and remain unaffected by government tax policies. Resources are shifted to agriculture, and as a result there is a large increase in land clearing by squatters. Agriculture gains an advantage over in- dustry, and industrial production is reduced. A tax reduction of 10 percent generates largely the opposite results. These results also hold for the experiments with capital tax policies. Substitution effects prove to be important for policy experiments involving tax changes on goods produced in tradables sectors, as summarized in table 6. The effects of tax changes on goods from the forest sector generate small econ- omywide effects, because this sector is small compared with the others. The industrial sector (which uses forest products relatively intensely as an intermedi- ate input) gains from the tax reduction and grows, but the forest sector itself suffers. The effects are reversed for a doubling of the tax on forest products. Deforestation remains largely unaffected in both cases. When the tax on agricultural products is reduced to half, the agricultural sector actually decreases. The industrial sector benefits because of its extensive use of agricultural products as intermediate inputs, and the forest sector is reduced. Deforestation for logs remains constant, but deforestation for land is somewhat reduced. Utility, income, and green GDP measures are reduced. A double tax on agricultural products generates the opposite effects. A tax on 278 THE WORLDI ANKIA.ONOMR NRFVI1FW.VVOL 9.1NO.! Table 6. The Effects of Changes in Taxes on Final Products (billions of colones) Sector (change relative to original tax rate) Base Forestry Agriculture Industry Item case Half Double Half Double Half Double Deforestation Logging 0.020 0.020 0.020 0.020 0.020 0.020 0.020 Squatting 0.020 0.020 0.020 0.019 0.022 0.018 0.025 Total 0.040 0.040 0.040 0.039 0.042 0.038 0.045 Production Forestry 0.552 0.528 0.609 0.405 0.902 0.160 1.363 Agriculture 13.984 13.984 13.985 13.258 15.706 13.964 14.030 Industry 18.477 18.503 18.418 19.243 16.663 18.268 18.909 Utility 0.232 0.229 0.240 0.135 0.459 0.064 0.572 Green GDP 31.962 31.941 31.980 31.589 32.815 31.324 33.253 Disposable income 37.681 37.669 37.709 37.317 38.544 37.051 38.984 Source: Authors' calculations. products from the industrial sector generates the same effects as a tax on agri- cultural products, although the magnitude of the changes is larger. VI. CONCLUSIONS The results of the CGE modeling study support the more conventional partial equilibrium approach that establishing property rights tends to decrease de- forestation (see, for example, Southgate 1990). The reason is that such rights allow forest users to capture the future benefits of reduced logging damage today. On the basis of a recent environmental accounting study (Sol6rzano and others 1991), this potentially avoidable loss is initially presumed to be 28 per- cent of the value of the residual stand. Using an interest rate of 10 percent, the simulation indicates that deforestation is dramatically reduced to 5 percent of the base level as both loggers and squatters internalize the losses associated with deforestation and reduce corresponding activities. Significant reductions in de- forestation occur even when the estimate of logging damage is substantially reduced. The CGE results concerning the effects of changes in the discount rate also parallel the predictions of partial equilibrium models: higher interest rates promote deforestation, and lower interest rates contribute to conservation. Beyond confirming the direct results of partial equilibrium analyses, the CGE approach also makes an important contribution by clearly identifying the indi- rect effects arising from intersectoral linkages. To determine the total impact, these indirect effects must be combined with the direct effects attributable to policies that are specific to the forest sector. For example, partial equilibrium analysis predicts that stumpage price increases will act directly to reduce log- ging. The model shows, however, that although deforestation from logging will indeed decline, total deforestation will nevertheless increase. This phenomenon arises from indirect linkages captured by the general equilibrium analysis. The Perssonz anrd Munasinghe 279 contraction of the logging and forest industry sectors causes a shift of resources toward agriculture, and as agriculture expands, deforestation increases. The importance of such indirect effects is also demonstrated in the analysis of economywide policy changes, such as an increase in the wage rate. Because of intersectoral resource flows, the general equilibrium model captures effects of changes in wages that are different from partial equilibrium results. If the wages of unskilled labor are increased (for example, by minimum wage legislation), the model predicts that deforestation could worsen. Although logging declines be- cause of the increased direct costs of higher wages, this decline is more than offset by the indirect effect of intersectoral flows, because the industrial sector (where minimum wage legislation is more binding) is much more adversely affected by the higher labor costs. Labor and capital thus tend to flow to agricul- ture, leading to the conversion of even more forest land for farming. These examples underline the importance of pursuing sectoral reforms in the context of growth. Without alternative employment opportunities, reducing logging activities will tend to direct labor and capital resources toward agricul- ture, industry, and other sectors. Expansion of some of these sectors could lead to a second round of effects on forestry and ultimately to more severe deforestation. APPENDIX. SUMMARY OF THE CGE MODEL STRUCTURE This appendix provides the key equations in the model and explains in mathe- matical terms the general features of the model described in section Ill. Factor Market Equilibrium and the Stock of Land The supplies of labor and capital are assumed to be exogenously given. The market equilibrium conditions for capital, unskilled labor, and skilled labor, respectively, are given by equations A-1, A-2, and A-3: (A-1) K =ZE 3 kg where K is capital, T is tradable goods, N is nontradable goods, C is the cost function, Pk is the user price of capital, and kg is the capital used in deforestation by loggers; (A-2) M = ' TTN - + Lq + Lg - 'aPM where M is unskilled labor, PM is the user price of unskilled labor, and Lq and Lg are the labor used in deforestation by squatters and loggers, respectively; and (A-3) E =Z,T,N aP apE where E is skilled labor and PE Is the user price of skilled labor. 280 1 I:<) WOR II BANK F.(:(ON where QF is the production of logs. When property rights are well defined, the logging companies take the oppor- tunity value of the forests into account. The opportunity value is represented by a function H(d), and is exogenous to the model. Demand for unskilled labor and demand for capital, respectively, are (A-13) ~~~~~~~P p + aH(d)j_ OPF(kg)', Pk + aHl(d)]_ kg k+a(Lg)] oePF ( kg)"3 Squatting. Squatters have a monotonically decreasing production function for cleared land with unskilled labor as the only factor of production: (A-14) dq = d(Lq) = (Lq)Y; -y < 1. The squatters' total revenue from land clearing is the price paid for the cleared land. When property rights are undefined, the squatters' total private cost for clearing the land will depend only on the amount of labor needed: (A-15) C(dq) = Pq' d- I (dq) where PqL is the price of labor in squatting. When property rights are undefined, the land is seen as common property. No market for the forests is available to squatters. The squatters' total private profit from land clearing is (A-16) rj(dq) = R(dq) - C(dq) 282 l HF WORI IDBANK F.( ONOM I Rl:VII W. Vol q. NO.' where RI is the profit function, R is the revenue function, and C is the cost function. However, for the squatting sector as a whole, land will be cleared until margi- nal cost equals marginal revenue. From the profit function, the squatters' de- mand for unskilled labor is (A- 17) Lq (P, PM and the supply of land cleared by squatters is -yp (A-1 8) dq-(fAt W When property rights are well defined, there is a market for the forests. Assuming that all squatters are identical, and that every squatter owns an equal share of the land that was previously available for squatting, the total private profit from clearing the forested land is now (Johansson and Lofgren 1985) (A-19) II(dq) = Psdq - C(d?)_ H(dq) where i is the interest rate. Deforestation will occur until marginal cost equals marginal revenue, and from the profit function the following are the equations for the labor used in deforestation by squatters and the total deforestation by squatters, where H is the future value of the forests. (A-20) LI La+[(L qyi/(1 + p + at H[(Lq)Y,j/(j + i)I (A-21) dq a [ Lq ] Macroeconomic Closure The current account is assumed to be constant. This implies that (A-22) ZiCY,N Pi Xi = B where B is the current account surplus. Equation (A-22) indirectly determines the disposable income. The gross domestic product (GDP) is determined as the sum of factor incomes, plus a term reflecting the value of diminished deforestation: (A-23) GDP = PKK + Pr E + P, M + PS S + S e T, N 1 Qi - A(dq + dg)H(l) Persson and Mninasinghe 283 where H(1) is the future value per unit of forests, and arj is the indirect tax rate. The model is solved by maximizing consumer utility, U, subject to a budget constraint: (A-24) Max U = In |II A)bHl, where E b, = 1 for ieN, T subject to I - E P, D, 0 where bi is the expenditure share devoted to good i, D, is the domestic final demand of good i, and I is disposable income. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Banco Central de Costa Rica, Secci6n de Cuentas Nacionales. 1990. Cuentas Nacionales de Costa Rica 1980-1989. San Jose, Costa Rica. Barbier, Edward B., Joanne C. Burgess, and Anil Markandya. 1991. "The Economics of Tropical Deforestation." Ambio 20(2, April):55-58. Bergman, Lars. 1990a. "Energy and Environmental Constraints on Growth: A CG;E Modeling Approach." Journal of Policy Modeling 12(4, Winter):671-91. - . 1990b. "General Equilibrium Effects of Environmental Policy: A cGE-Modeling Approach." Research Paper 6415. Economic Research Institute, Stockholm. Processed. Biesanz, Richard, Karen Zubris Biesanz, and Mavis Biesanz. 1987. The Costa Ricans. Englewood-Cliffs, N.J.: Prentice-Hall. Blomstrom, Magnus, and Mats Lundahl. 1989. Costa Rica en Landstudie. Blickfang Centralamerika, Latinamerika-Institutet, Stockholm. Briceino, F. 1986. "10 Table for Costa Rica." University of Costa Rica, Department of Economics, Costa Rica. Processed. Carriere, Jean. 1991a. "The Crisis in Costa Rica: An Ecological Perspective." In David Goodman and Michael Redclift, eds., Environment and Development in Latin Amer- ica: The Politics of Sustainability. Manchester, U.K.: Manchester University Press. . 1991b. "The Political Economy of Land Degradation in Costa Rica." Interna- tional Journal of Political Economy 21(1, Spring): 10-31. Chichilnisky, Graciela. 1993. "North-South Trade and the Dynamics of Renewable Re- sources." Structural Change and Economic Dynamics 4(2, December):219-48. Dasgupta, Partha. 1982. The Control of Resources. Oxford: Basil Blackwell. Devarajan, Shantayanan. 1990. "Can Computable General Equilibrium Models Shed Light on the Environmental Problems of Developing Countries?" Paper prepared for WIDER conference on "The Environment and Emerging Development Issues." Helsinki, September 3-7. Processed. Foy, George, and Herman Daly. 1989. "Allocation, Distribution, and Scale as Determi- nants of Environmental Degradation: Case Studies of Haiti, El Salvador, and Costa Rica." Working Paper 19. World Bank, Environment Department, Washington, D.C. Processed. 284 THE ''ORIA) KANK V tO()NOMI( RPI\ [liWV. 9lt 9, NO. ' Harrison, Susan. 1991. "Population Growth, Land Use, and Deforestation in Costa Rica, 1950-1984." Interciencia 16(2, March-April):83-93. Hartshorn, Gary, Lynne Hartshorn, Agustin Atmella, Luis Diego G6mez, Alfonso Mata, Leonardo Mata, R6ger Morales, Rafael Ocampo, Douglas Pool, Carlos Quesada Mateo, Carlos Solera, Ra[l Sol6rzano, Gary Stiles, Joseph Tosi, Alvaro Umana, Carlos Villalobos, and Robert Wells. 1982. Costa Rica: Country Environmental Pro- file: A Field Study. Tropical Science Center, San Jose, Costa Rica. Hellsten, Martin. 1988. "Socially Optimal Forestry." Journal of Environmental Eco- nomics and Management 15(4, December):387-94. Hugo, V., V. Gonzalez, R. Jimenez, and T. Vargas. 1983. Problemas Econ6micos en la Decada de los 80. San Jose, Costa Rica: Agencia para el Desarrollo International. Johansson, Per-Olov, and Karl-Gustaf Lofgren. 1985. The Economics of Forestry and Natural Resources. Oxford: Basil Blackwell. Jorgenson, Dale W., and PeterJ. Wilcoxen. 1990. "Intertemporal General Equilibrium Modeling of U.S. Environmental Regulation." Journal of Policy Modeling 12(4, Winter):715-44. Keogh, R. M. 1984. "Changes in the Forest Cover in Costa Rica through History." Turrialba 34(3):325-31. London Environmental and Economics Centre. 1992. "Case Study for Thailand." In David Reed, ed., Structural Adjustment and the Environment. Boulder, Colo.: West- view Press. Lutz, Ernst, and Herman E. Daly. 1990. "Incentives, Regulations, and Sustainable Land Use in Costa Rica." Working Paper 34. World Bank, Environment Department, Wash- ington, D.C. Processed. Munasinghe, Mohan. 1993. Environmental Economics and Sustainable Development. World Bank Environment Paper 3. Washington, D.C.: World Bank. Munasinghe, Mohan, and Wilfrido Cruz. 1994. Economywide Policies and the Environ- ment. World Bank Environment Paper 10. Washington, D.C.: World Bank. Munasinghe, Mohan, Wilfrido Cruz, and Jeremy Warford. 1993. "Are Economywide Policies Good for the Environment?" Finance and Development 30(3, September):40- 44. Persson, Annika. 1994. "Deforestation in Costa Rica: Investigating the Impact of Mar- ket Failures and Unwanted Side Effects of Macro Policies Using Three Different Modeling Approaches." Beijer Discussion Paper Series 48. Beijer Institute, Stockholm, Sweden. Processed. Quesada Mateo, Carlos A. 1990. Estrategia de Conservaci6n para el Desarrollo Sosteni- ble de Costa Rica. San Jose, Costa Rica: Ministerio de Recursos Naturales, Energia y Minas, Republica de Costa Rica. Ravent6s, Pedro. 1990. "Commercial and Tax Reform in Costa Rica in the Mid-1980's." Preliminary version. INCAE, Alajuela, Costa Rica. Processed. Repetto, Robert C. 1988. "Economic Policy Reform for Natural Resource Conserva- tion." Environment Paper 4. World Bank, Environment Department, Washington, D.C. Processed. Sader, Steven, and Armond Joyce. 1988. "Deforestation Rates and Trends in Costa Rica, 1940 to 1983." Biotropica 20(1):11-19. Sol6rzano, Ra6l, Ronnie de Camino, Richard Woodward, Joseph Tosi, Vincente Watson, Alexis Vasquez, Carlos Villalobos, Jorge Jiminez, Robert Repetto, and Wilfredo Cruz. Persson and Muntasinghe 285 1991. Accounts Overdue: Natural Resource Depreciation in Costa Rica. San Jose, Costa Rica: Tropical Science Center. Southgate, Douglas. 1990. "The Causes of Land Degradation Along 'Spontaneously' Expanding Agricultural Frontiers in the Third World." Land Economics 66 (1, February):93-101. Unemo, Lena. 1993. "Environmental Impact of Government Policies and External Shocks in Botswana: A CGE-Model Approach." Beijer Discussion Paper Series 26. Beijer Institut, Stockholm, Sweden. Processed. T H I WX O 10 i D1 1 A N, k I t ( I N ( I Mi I ( R I X I f W'. XI O I . q , N l) 2 . 2 .> --Io 4 The Role of Infrastructure in Mexican Economic Reform Andrew Feltenstein and Jiming Ha This article estimates the relationship betwt!eien the prol ision ot public infrastructure and private outpuit in sixteeni sec-tors in Mexico. The sector-specific cost functions depenzd on wt-ages, the cost of capital, ind the nominal values of the stocks of three types of infrastructure: e -cctricity: trcinsport, and conimmiinications. The article concludes that iuifrastriucture in electricity a17d conMMUi 7Ciuiton1s gene7rrallr reduces the cost of sectoral produliction., huit tran7sportatiZon infrastructure tends to increase costs of sectoral proidic- tio?i. It appears that MNexican public expedliture ot7 CIcCt riclty and! commun11s1lications has enbaincedl the prodluctiv'itv ot priivatc prodchtion, but expendituire on trcinsport may actually have had a detrimental eflfect on priv'ate ouitpuit. In addition, cilthough in general labor and finrastructurc are substitutes, in the case of electricity and communi- cations inzfrastructuire, capital and( infrastrutctutre arc comiplements. in the case of tranis- port inifrastructure these comnclusioins are rei, crsed. During the period 1985-91 the government of Mexico carried out policies that were considered models of econiomic stabilization. Starting in 1985, Mexico embarked on an ambitious program that combined fiscal incentives and trade liberalization. The average tariff rate was reduced from 25 percent in 1985 to approximately 12.5 percent in 1990; at the same time, the number of tariff categories was reduced from ten to three. The results of the trade policies were very positive: there was a steady improvement in the productivity of the export sector, and, until 1990 the trade balance remained positive even while the vol- ume of imports increased. The Mexican trade balance turned slightly negative in 1990-91, and today the country suffers from well-publicized problems in its external economic situation. While these changes in the foreign trade regimiie were being undertaken, tax policy officials in Mexico experimented with a number of tax policy tools to promote private capital formation. Major such initiatives were general and industry-specific tax credits, employment tax credits, and corporate tax rate Andrew Feltetnstein is with the Department of Ficonomics at the University of Kansas. and Jiming Ha is with the Europe 11 Departmetit ofthe Internat ional Monerary FuLid i I MF). The research for this article was supportedi by a WYorld Banik granit for the Infrastructure InadLquacies Project in Mexico. The authors would like to thank Frank Lvy% and Daniel Ocks for thei- comments and assistance andtdNeil Roger for origilalls proposing the topic and offering continuLIous gidance. ©, 1995 Tiet Interniationial Baiik tor Recoiistruuito(nl and l)eVclop1m1en1t / TrlE WORI D BANK 2S7 28S TII 't)RI I ) BANK I (O)N OII( I:IVIF W, II " \z) 2 reductiois. To avoid the negative effects of large budget deficits, the government has reduced both currenit and capital spendiig. As a result of these domestic and foreign sector reforms, the Mexican inflation rate in the wholesale price index was reduced from 88.4 percent in 1986 to 20.5 percent in 1991. The hoped-for outcomiie of these changes wa1s an increase in domestic invest- ment and output. Despite having carried out all the apparently correct macro- economic policies, Mexico has not, however, observed these desired outcomes. Indeed, real incomlie growthi has remained relatively flat: from 1985 to 1991 the average rate of growthi of real gross domestic product (GDP) was only 1.8 percent. Why has a successful stabilization program apparently failed? In this article we will estimate the relationship between the provision of Mexican public infra- structure and private output. As part of the program of fiscal austerity intro- duced in 1986, the authorities have reduced capital expenditure, as well as expenditure on maintenance. Thus a positive estimated relationship between public capital and private output would offer a partial explanation for the slow growth in real income. In estimating these output elasticities, we will offer a first step toward answer- ing a broad question. How can we judge whether there is a shortage of infra- structure in Mexico that has been caused bv the government's cuts in capital expenditure? This question cannot be addressed without first estimating the optimal level of infrastructure. Such a c*alculation will depend on the type of elasticity estimates we will carry out here. If, for example, our estimated output elasticities turn out to be zero, then it would clearly be desirable for the govern- ment to reduce its expenditure on public infrastructure. Positive elasticities might offer initial evidence that the cuts in capital expeniditure maya have been counterproductive. A simple analysis of output elasticities may, of course, be insufficient to deter- mine whether the current stock of public capital is optimal. If, for example, nonefficient prices are currentiv being charged, then it may well be that the stock of infrastructure is too large. Thus Winston (1990, 1991) argues that public spending on road infrastructure in the United States should be curtailed. We should note that even if the stock of a certain type of public infrastructure is too large, this does not necessarily imply that all sectoral elasticities would be nega- tive. Feltenstein and Ha (1994) use the results of this article to calculate marginal benefits and social rates of return to infrastructure, giving some further insight into the optimalitv of the current stock of Mexican infrastructure. Table 1 offers an indication of the extenit to which the rate of growth of the national stocks of Mexican infrastructur-e has slowed. We also note that, as infrastructure growth has stagnated, the rate of growth in real GDP has also slowed, Of course table I should be viewed as offering only casual evidence of a connection between infrastructure and real output. Appendix table A-1 gives a full listing of the stocks of infrastructure from 1970 to 1990. Figures I to 3 show the real national stocks of infrastructure for, respectively, the electricity, transport, and communications sectors. We see, for example, that Feltenstemn and Ha 289 Table 1. Average Ainntal Rzates of Grotvth of Kel Real Indicators in Mexico, 1970-9() (percent) Indicator I 97()- 75 197 7- I98()-s i 1 98 5-90( GDP 6.i . 1.9 1.3 Total eniployvimenlt 3. S 4.1 3.6 0.4 National intrastrulcture steloS Electricity 14.8 1(I.9 1.8 - 1.2 Transport 6.2 11. 3 3.8 3.1 com nltllicat tIO s 56.9 49.2 8.1 (). I Source: im (varioLIs IsLues, 199 1) mnd Ban.co c1e Mexico (1992). electricity infrastructure has declined since 1981. Transport infrastructure rose sharply froni 1985 to 1988, then slowly declined (see figure 2), although over the whole 1985-90 period it rose more rapidly than did real GDP. Communications infrastructure began to decline in 1984, rose slightly in 1986-87, and has de- clined since then. We will develop anl econometric analysis of the sectoral elasticities of Mexican industries with respect to the national stocks of public infrastructure. We con- struct a micro-data set containing both national infrastructure series and firm- specific inputs required for estimating both sectoral cost functions as well as factor demand equations. We then estimate these equations, thereby deriving the elasticities of sectoral production with respect to public infrastlructure. The parameter estiniates we derive may he used to examine a number of issues. In particular, they may be used to draw conclusions about whiether the Figure 1. Real Stock of Electric,it-V Infrastructture, 19 70-90 index numbers based on constarit 1970 prices 350 - 300- 250- 200 - 150 - S O -- 1970 1974 1978 1982 1986 1990 Source These inclex numbers are (lerivedI from data given in table A-i. 290 THE WORI 1) BANK F(:()NO.,11( RKI- EW. Vol. 1, NO. 2 Figure 2. Real Stock of Transportation Inftrastructure. 1970-90 Index numbers basecl on constant 1970 prices 250 - 200 - 150- 100- 1970 1974 1978 1982 1986 1990 Source: Thesc index numbers are derivt d from datai given in table A-1. failure of the Mexican economy to respond to various incentives and stabiliza- tion programs may partly stem from lack of infrastructure. 1. ESTIMATING THE EFFECTS OF PUBLIC CAPITAL ON PRIVATE OUTPUT Several recent studies have examined whether, or to what extent, public sector infrastructure capital contributes to private sector productivity growth and changes in the structure of production costs. Among such studies are Auerbach (1990), Aschauer (1989, 1990), Hulten (1992), Munnell (1990b), and Shah (1988). A further direction of research has been to examine the effectiveness of fiscal incentives in generating private investment that may act as a substitute for public infrastructure. Rajagopal and Shah (1990a, 1990b), Shah (1992), and Shah and Baffes (1990) are among sorne of the recent articles in this area. The results reported in the literature are generally obtained by using an aggre- gate production function framework to estimate the relationship between out- put, or total factor productivity (TFP) growth, and public sector capital. The reported elasticities of output and labor productivity with respect to changes in public infrastructure capital formation are very diverse. Using cross-sectional data of nine countries in the Organization for Economic Cooperation and Devel- opment (OECD), Ford and Poret (1991) estimate the average elasticity of TFP with respect to changes in infrastructure to be about 0.45. Hulten and Schwab (1991) note an interesting pattern: that the size of the estimated output elasticity of public capital varies directly with the degree of aggregation of infrastructure; both they and Tatom (1991), however, find no statistically significanit relation- ship between the growth of infrastructure and TFP growth. Thus Aschauer (1989) Feltenstein and Ha ' 91 Figure 3. Real Stock of Communications Infrastructure. 1970-90 Index numbers hased on constant 1970 prices 60- 50- 40- 30- 20- 0T- I IT IF I lITI 1970 1974 1978 1982 1986 1990 Source: These index numbers .ire dlerived from clato given in t;ible A-I. and Munnell (1990b), using aggregate U.S. data, find output elasticities of 0.39 and 0.34, respectively. Munnell (1990a), which is based on regional data, esti- mates a smaller elasticity of 0.15, but Duffv-Deno and Eberts (1991), using metropolitan data, find a statistically insignificant elasticity of personal income with respect to local public infrastructure of 0.08. Econometric inference regarding the effects of public infrastructure capital de- pends critically on specification of the production or cost functions, or both. Using plant-level cross-sectional and panel data, Tvhout and Westbrook (1991) and Ha and van Wijnbergctn (1992) examine the returns to scale of Mexican manufactur- ing industries by estimating translog production and cost functiolls. Their results show that the translog functions fit the data verv well. Ha and van Wijnnbergen (1992) reject the specification of the constant elasticity of substitution (CES) pro- duction function and the corresponding cost function. Berndt and Hansson (1991) and Nadiri and Mamuneas (1991) study the effects of public infrastructure capital on the cost structure an(d the performance of Swedish and U.S. economies. In this study we use an approaclh similar to that of Nadiri and Mamuneas (1991). 11. MODEI. SPECIFICATION AND ESTIMATION If the cost of production in the private sector is affected by the types and quantities of public sector capital services, the traditional cost functions can be modified to include the "externality" associated with these capital services. We write the cost function for an industry as (1) C = O(uv, v.g, t) 292 1111 WORI I) KANK F ONO( \l( Rl\ 11[> VO1 '! I where C is a cost functioni that is twice continuously differentiable, iv is an N-dimensional vector of prices of private inputs (including labor, intermediates, and capital), y is the output quantity of the industry, g is an M-dimensional vector of the nominal values of the stocks of general infrastructures of the economy (includinig both public and private infrastructure and capital for re- search and developilmeit), and t is an index of time representing disembodied technical change. We are supposling that enterprises pay a service charge for the use of public infrastructure, such as electricitv, commuLnicationis, and transport. This service charge is not, however, market determined, and the government generaliv underprices these public services. Thus both the available stock, whici is gener-ally in short supply, and the regulated price have an effect on enterprises' costs. Our specificationi of the cost functioni in equaltion I is derived from a simple production function augiiiented by public infrastructure, Y = F (K, L, g), which supposes that public infrastructure may enih.anice private production. This is similar to the specificationi in Feltenstein and Morris (1990), as well as to the endogenous growth model of Barro (1990). Infrastructure affects the cost structure of an industry in two ways. First, a larger qualintity (and hetter quality) of infrastrLucture will shift the cost per unit of output dowinward in an industl-y if it receives anly benefit from improved or larger services provided by that infrastructure. This can be called the "produc- tivity effect." This infrastructure need not he costless. Thus, for example, enter- prises would genierally pay governiment-determined prices for the services pro- vided by electricity infr-astructure. Secoid, firms will adjust their production decisionis with respect to their own labor, intermediates, and capital stock if services provided by infrastructure are substitutes for or comirplemenits of their owIn factors of productioll. That is, the effects of infrastructure may not be neutl-rll with respect to private sector factor demands. To estimate the effects of public capital on the productivity and production structure of the industries, we specify a translog cost functioni augmented by the nominal stocks of infrastructure: (2) In C = fi0 + Xfi In it, + I In y + Bt + EX,XiJf, In wiv In tv, + ,f,.,In wit In v +• Xjf,, In wit + fl, In yt + f,B, In vIn y + X4fl), In g_ + X_,X4v%j In wi Iln g5 + X,P( In v In g, + e. To account for industry differences, we have introduced dummies in the intercept and slope coefficients of input price and output variables. That is, fl = fl0 + LX a%,D1,, g' = B, + El, (cY,D,, fv, = , + 4- a,,iD,,, and o= + 4 a_,i,D,,,D whiere Di, refers to industry duLiinies taking values Feltenstein and Ha 293 1 and 0, 17 is an identification industry index, t is a trend variable, and E is the error term. One niight also introduce industry dummies for the intercept and slope coefficients of the types of capital stocks. Given the relative scarcity of data, we have chosen not to do so. It should be noted that the effects of infra- structure are not constrainied to be the same across sectors. In particular, the coefficients T' are different for different sectors because of the use of the dummy variables. We have avoided using separate time and industry subscripts so as to avoid unreadable notation. Our specification of the intercept terms implies that we are emiiploving a fixed-effect model. We should note that our estimates may therefore be sensitive to spurious correlationl. Using Shepherd's lemma, we obtaini equation 3i, representing factor demands: (3) s,= + EX ,, In w, + In + /3,,t + X,(1) In g, where s, is the share of the ith private input in the total cost. Input shares in each industry depend not only on relative factor prices, output, and techniology change, but also on the infrastructure services. The parameters (ls determine the magnitude of the factor-bias effects associated with these types of infrastructure. I We estimate equations 2 and 3 jointly. This simultaneous estimation is supe- rior to estimating equation 2 only, because joint estimation employs more infor- mation. We will also impose regularity restrictions on the parameters required by the properties of the cost functions. Once the cost and factor share functions are estimated, we will carry out certaini other studies associated with the value of infrastructure. In particular, we will examine the impact of public capital on private production costs and factor demands. The spillover effects of infrastructure on cost and input shares are captured by the magnitudes and signs of the parameters (1), and ,),. The cost elasticities with respect to infrastructure cani be computed using (4) D) In C,, = + X, ±),s In i,1,h + I,, In 1sh - a ,l gy. where s stands for types of publicly financed capital, and / is an industry identi- fication index. The elasticity estimates lTh4 measure the productivity effect of infrastructure. Previous studies in the literatuL-e show a negative sign for these elasticities. Firms beniefitiig from the availability of infrastructure will adjust the structure of their factor demanids. The factor adjustmenit effect is measured by the elasticity of factor shar-es with respect to infrastructure, as,&,la In g,. If the factor cost share incr-eases, decreases, or does not change, the types of infrastruc- ture are factor-using, factor-saving, or neutr-al. Combining the two effects, we 1. Ain anotinmolis referee has noted that otlnly under certain restrictive conditioins can we omit the user charge for infrastruCture services. Bh uIsing iii intl lid - infrastruLctrLr st(ocks. we dlo IncorpiorpiC ir ser prices. WVe justif' the incorporationi of, for example. the real stock of electricity itifrastrticttire hy otur a ssnrilptioin of full capacity Urilization1 anid hetnce Yxcess dema.nd. Thus the available stock, as well as the price, of electricity infrastriuctiru hais ai 111iipact 011 enterpris costs. 294 Tl I W'ORI 1) IANk IOt()N)MAI C RIV\IFX'.\%0 I No.' obtain the total effects of public sector capital services on input demand as (5) r1 lnx, = d . it = 11, b + 0 In g. ~~~Si'., The sum of the productivity effect and the factor adjustment effect can offset each other. A positive, negative, or zero sign in equation 5 implies that the particular publicly financed or privately provided capital service and the ith private input are complements, substitutes, or neutral, respectively. Suppose there are two inputs: capital and labor. If we use equation 5 to compute the effects of infrastructure on capital demand, then the effects on labor demand are given as (6) Innxi, n5" = l''g =71bst I- l _ .Wl In section IV we estimate the equation of capital share and the cost function simultaneously. We also use equations 5 and 6 to compute the effects of infra- structure on capital and labor, respectively. 111. DESCRIP-TION OF THE DATA AND ESTIMATION METHOD We have complete data on both real and nominal capital stocks for sixty-three sectors during 1970-90. As numbered and identified in the Mexican national accounts (Governmenit of Mexico, various issues), these sectors are 5, 7-32, 35- 67, and 70-72. To reduce the size of the model to be estimated, we aggregate these sectors into sixteen groups (see table 2). The aggregation corresponds to that used in the Mexican national accounlts. For the estimationi of equations 2 and 3 the following data are used. For each industry, the qualtity of output is meastired as the sectoral gross domestic prod- uct (GDP) in constant prices with base vear 1980. Costs are given by sectoral value added, normalized so that the residuals from equations 2 and 3 are compa- rable. The real wage rates are calculated by dividing the mean annual remunera- tion (at current prices) by the consumer price index (normalized at 1980 = 1). The quantity of labor input is given by sectoral employment data. We use the net stock of sector-specific capital, valued at 1980 constant prices, to represent sectoral iniputs of capital. W/e use the three-month Mexican Trea- sury bill interest rate as a proxy for the cost of borrowing and hence of the cost of capital. Clearly this is a crude measure, because it assumes a uniform cost of capital across sectors. Given the limitations of this study, we have been unable to carrv out a more detailed analysis that would permit disaggregated costs. In particular, data limitations make it impossible to construct measures of capital costs for each sector. The capital stocks of sectors 61 (electricity), 64 (transport), and 65 (communications) are used as three types of infrastructure, measured at the national level. These data (except the interest rate) are the same as those used by Jarque (1988), whose sample was from 19)70 to 1984. Feltenstein and Ha 295 Table 2. Aggregation of Sectors and Estimates of Coefficients for Input Price and Output Variables Estimates Coefficient on Sector numbers Coefficient on output Iffo,t the Mexican 1Iztercept relative factor qiantttv' of Sector national accounts term], A<, prices, 0-r the ndustry, 03 Mining 5. 7-10 56.791 -0.987 9.769 (3.61) (-2.08) (-3.39) Food products and 11-23 72.588 -1.098 -1 1.208 tobacco (3.68) (-2.09) (- 3.46) Textiles 24-2X 67.200 --1.122 - 10.727 (3.79) (-2.25) (-3.49) Wood products 29-30 52.067 --1.112 -9.265 (3.65( (-2.46) (-3.39) Paper and print 31-32 59.904 -0.972 10.037 (3.98) (-2.09) (-3.56) Chemicals and petroleum 35-42 63.949 -0.992 - 10.448 (3.711 (-1.99) (-3.45) Nonmetallic minerals 43-45 60.144 -0.980 -10.100 (3.82) (-2.0(7) (-3.50) Basic metals 46-47 60.636 -0.698 - 10.049 (3.97) (-1.49) -3.5 4) Machinery and equipment 48-58 70.584 - 1.023 - 11.(08 (3.77) (-1.99) (-3.48) Other manufacturing 59 53.362 - 1.037 -9.557 (4.0.5) -2.38) (-3.6.5) Constructioni 60 /2.179 - 1.386 11.230 (3.65) (-2.6.3) (-3.45) Electricitv 61 n.. n.a. n.a. Commerce and hotels 62-63 99.712 - 1.465 - 13.377 (4.02) (-2 50) ( 3.66) Transport and 64-65 nal. nla. n.a. communications Financial services 66-67 86.404 - 1.195 - 12.459 (4.11) (-2.18) (-3.72) Medicine 70-72 74.493 - 1.329 - 11.316 (3.56) (-2.45) (-3.38) n.a. Not applicable. Note: The sector numilbers correspond to those in the Mexican nationial accounts. Estimates are for equation 2, a translog cost finctioni augmented by the nominal stocks of infrastructure. The numbers in parentheses are the values ot t-statistics. Soirce: AutIlors' calculations, Banico de Mexico (1992), and Government of Mexico (various issues). We suppose that users of infrastructure pay a charge for services provided. In the absence of capacity utilization data, we also suppose that there is 100 percent utilization of infrastructure. Accordingly, we use the nominal stock of infrastructure as an explanatory variable in equation 2. We are thus assuming that the availability of infrastructure, as well as its price, affects costs. It should be noted that this assumption does not imply that we assume that the public goods in question provide service at zero marginal cost up to the capacity con- straint and then provide no more. Rather, we are simply assuming that the use of infrastructure, which benefits the economy, is equal to the provision of infra- 296 1111: WORI 1' hANK I( O)NO)MI( 1K1 VI I W, V(.1'. NV l structure, which our data reflect. The appendix gives a detailed description of our estimation techniques, as well as a listing of regularitv conditions. IV. EMPIRICAL RESULTS The translog functions seem to fit the data well, because the adjusted R2 of the cost functioni is 0.99',, and that of the capital share equation is 0.713. The estimates of the parameters of equations 2 and 3' are given in tables 2 and 3. We may note that the time parameters used to represent disembodied technical change are not significant. We should also note that the translog functions in- clhde one variable in more than one term, so that the t-statistics should not be judged as with a conventional regression. We thus observe the response of marginal cost to changes in infrastructure capitals. These are the estimated values of IV' for each sector. The results in table 4 show that infrastructure in commun1ications and electricity is marginal cost- reducing, and infrastructUre in the transport sector is marginal cost-increasing. Table 3. Estimates of Coefficients for the Plroduc-tioni Struc-tuire Variables P[a rawn eer Estimate Direc t impact (,/ inlrastructo ire st cks ElectricitN, (1)1 1.852 (().82) Transport, (1)1 - 1.897 (-0.97) Communications, 1)i (0.647 (1.12) hPtro rblis Electricity () ,,() 0.028 (0.74) irransporr. '), r, -0( 02)6 (-0.74) Co11111mc1,1Ation1s 1 -0.00744 (-0.98) Ii metri jd, -() 0.00371 DliisZez ' ,riabl)le cross terns Relative factor prices and output, /3,() I 1.3 (2. 82) Relative factor prices and time trentl. l's,, - 0.001 58 (- 0. 3) OLItput and time trelld /3, -0.00196 t( .32) Rel.tive factor prices own elasticity , 0. 05) 12.4. 1 Output oWIt elasticity. /3,l (0.524 (3.86) Note: Estimaers are for equatimi 2, a traliislOg Colt fu uCt on a tIgnicnred by the noniiial I sotocks of infrastructure. Il-.tatisrics are in parenileses. Source: AutbHI S' CalCUlatiC0iis. Feltensteio and Ha 2)7 Table 4. The Impact of Infrastructuire onl Marginial Cost Selc tor grorip Electricity Transport C(omnmnmicaltists Mining 0.16X 0.169 - 0.061 - 0.83) .97) (- 1.14) Food products and tobacco -0. I5 0. 1 161 - (0.59 (-()X71 (1.04) (-1.26) Textiles 0.165 0. 171 -0.062 (- .87) 1.0)4) -1.26) Wood products -0.208 0.212 0.067 (-0.98) 11.14) (- 1. 19) I'aper and print -0.194 ().19 -0 0.063 (-0.941 (1.0)91 (-1.16) Chemicals and petroleuIm 0.162 0.168 -0.064 1-0.85) I .08) (-1.27) Nonimiletalliic miner,als )1' 77 0.18 -1 0.066 (-0.88) (1.(6) (-1.25) Basic meetals ().i159 0. 5(1 - 0.(55 (-(1.78i () 10.850 i- 1.03) Machiniery andl cquipnient 0.146 (1 147 -0.057 - 0.80() (() 94) 1. - 8IX Other mnanufacturing - 0.192 0.2 1) - (0.069 (.871 (1.06)1 1-1 19) Constructioni -( (.147 11.159 (-0.58 - (.XI 1 1.0 1) (1.23) Commerce anid hotels -1 114 1 (1.14) - ().(5) ( (! X7) (m 'I () (- 1.19) Financial services (. 15(0 1.1 54 -(1(.5) (-11.86.) 11(31 .()( 1.13)1 Medicine ().1 9 1.1(64 i.057 (.9(1) L(.118) (-1.24) Note: Values are the Jh.ingeC. In scCotral m.arginal cost with respect tu tht e change In infrastructure, 1'". t-statistics are in parentheses. Source: A ithors' c.alc lations. After the model is estimated, we use equations 4, 5, and 6 to calculate the effects of the different types of infrastructure on production cost and factor demands for each group. The cost elasticities withi respect to each of the infra- structure capital services are shown in table 5. The numbers in table 5 are the means of the estimates of the average imppact of the infrastructures on each of the sixteen industries during the sample period. The first column in table 5 shows that increascs in electricitv infrastructure reduce productioni costs in all sectors except basic metals, machinery and con- struction. Transport infrastructure, however, has a cost-increasing effect in most cases. Infrastructure in communicationis reduces produLCtioll costs in all sectors except the basic metals sector, but to a lesser extent than infrastructure in elec- tricity does. The results are similar to those found by Jarque (1988). Jarque regressed the Solow residuals on three types of infrastructures and found posi- tive coefficients for the infrastructures of electricity' and comImunications and negative coefficienits for transport infrastructure in most sectors. Jarque, how- ever, for eachi sector assumed a Cobb-Douglas productioll function with con- stant returns to scale, but we do not impose that restriction. Indeed, our meth- 298 1 III1 WR) IANK FK ONONI R IM IEF%', (01) . . Ni ) 2 Table 5. The Cost Elasticitv with Respect to lIifrastructure Sect(or grozip ElectricitY Tralisport Communications Mv1ining -(.)025 -0 0(2 - 0.036 Food products and tobacco - 1 19 0. 142 - (0.08 Textiles -0. 78 0.113 -(0(9( Wood produCts -0.3 19 0.3 13 - (0.72 Paper and print -). 25 3 0. 229 -( 0_ 7 Chemicals and petroleumlI -0. 104 () 127 -0.10 7 Nonmetallic minierals - 127 (.1 72 - 0.097 Basic metals (.105 - 0.254 0.031 Machinery and equipimient 0.046 - 0.074 -0.043 Other manitiuaciirinig - 0.054 0.147 -(.07-1 ConstruCtion 000)6 (1.0)42 -0.093 (Commerce and( hotels -0.096 ). 12 - 0.(74 Financial services -0.0)7 (.11 -0.1)32 Mediciuie -(.178 0.201 -0.089 Note: These niLiiritrs representr the ilmeanis for 1980-90 of the corresponding sectoral elasticities, II,,/, =) In C,l/ In g'. Sonrce: Au thors calcLIa3ti ins. odology has a number of additional elements of value added beyond those that are in jarque's. Not only do we avoid imposing the restriction that elasticities be time-invariant, but our specifications also permit the identification of a number of secondary effects. Thus, for example, we are able to estimate the interaction between the scale of the economy and the level of infrastructure. We are also able to determinie whiether infrastructule types and private inputs are comple- ments, substitutes, or neutral. Decomposing the estimates of 17,.,5A using equation 4, we find a large secondary effect with respect to the interaction between the scale of the economy and the level of infrastructure.2 Thus the direct effect of electricity, which is measured by the estimate of (k, in table 3, is positive, ais is the interactioni with relative factor prices, measured by the estimate of q)Io/I.F However, the overall elasticities, as shown in table 5, are ailmost all negative. This is because the estimates of the effect of scale, wlichi are captured by the last term in equation 4, are all negative and more than offset the direct effect. It is difficult to explain whv electricity and communications generally reduce sectoral productionl costs but transport infrastructure tends to raise them. We have, in fact, no good explanation for the couliterintuitive result for transport. It is possible that our assumption of full-capacitV utilization overestimates the use of transport infrastructure, thereby giving the wrong" sign. An anonymous referee has suggested an alternative explanation. If there is, indeed, full-capacity utilization, then infrastructure can no longer be properly considered a public good; rather, it should be treated as a congested public good, and there should be governmenit intervention in the form of congestioni tolls. Differential pricing 2. We a1-e grateful to one of the referees for poinriog out rhis featurt-e of the estimates. Ole miglt also wish to address a related topiC, nlaiely, the degree of scale econiomiies lper se. This is the subjeCt of Ha anid van Wijihergen 11992), which estimiaztes the retinrs to scale of Mlexican manuifaCturiigiindustries. Feltenstein aznd Ha 299 Table 6. Elasticitv of Labor Demand with Respect to Intrastructure Sec-tor group Electricity Transport Co,nnitzincations Mining -0.064 0.03 5 -0.026 Food products and tobacco -0.149 0 170 - 0.08() Textiles -0(.109 (.142 -0.081 Wood products - 0. 352 0. 44 -0.063 Paper and print - 0.299 0_2-72 - 0.045 Chemicals and petioleum -0.138 0. 159 -(.098 Nonmetallic minerals -( .166 0.209 - 0.087 Basic metals 0.027 -() 18 0.052 Machinery and equipment 0.01 5 -0.045 -0.034 Other manufactUrIng -0.1()1 (.19( - 0.(059 Construction 0.046 0)A'( - 0.086 Commerce and hotels - 0.130 0.19 - (.065 Finanicial services - 0.136 0. 156 -0.019 Mledicine -()0206 0.227 - (.082 Note: These nLumbers represent the means tor 1980-901 (i the corresponding sectoral elasticities, 11.' 0 In X, ,fid InI g'. Source: Authors' CJlcuI.tlonsS- would result, with a rise in peak prices and a decrease in off-peak prices. The large positive externality associated with infrastructure may thus largely disap- pear. A further possibility is that the measured stocks of transport infrastructure may have a broader coverage than is actually correct. An analysis of data sets is, however, beyond the scope of this study. Using equations 5 and 6, combined with estimated parameters, we calculate the elasticities of capital and labor demands with respect to infrastructure. Table 6 shows the elasticities of sectoral demand for labor with respect to infrastructure. We see that electricity and communications infrastructures are generally sub- stitutes for labor. The exceptions to this rule are the same sectors for whicih these types of infrastructure were cost-increasing (see table 5). Transport infrastruc- ture, however, is generally a complement of labor. These results seem plausible if we suppose that transport is largely comiiposed of roads. Thus an increase in roads would lead to a corresponding increase in demand for labor, possibly because the increased provision of roads causes industries to use more labor as, for example, drivers and loaders. Table 7 shows the sectoral elasticities of capital demiland with respect to infra- structure. Our results are similar to those of Berndt and Wood (1975, 1979), where demand elasticities are estimated using rranslog cost functions applied to U.S. data. They found that, in the United States there was substitutability be- tween energy and labor, and complementarity between energy and capital. We find that communications infrastructure is a substitute for private capital in all sectors other than in the basic metals sector, wlhere it complements both labor and private capital. Transport infrastructure, however, is a substitute for private capital in most sectors. As a final exercise, we estimate the cost function associated with a Cobb- Douglas production functioni, using aggregate data. We do this to see if our 300 i HF. WORI 1) BANK F(CONOMIC REVIEW, VO. 9 NO. 2 Table 7. Elasticity of Capital Demand with Respect to Infrastructure Sector grouip Electric-ity Transport Communications Mining 0.108 -0.126 -0.071 Food products and tobacco 0.50( -0.436 -0.252 Textiles 0.310 -0.249 - 0.192 Wood products -0.069 0.080 -0.138 Paper and print -0.161 0.143 -0.081 Chemicals and petroleum 0.107 - 0.069 - 0.163 Nonmetallic minerals 0.002 0.052 - 0.132 Basic metals 0. 155 -0.301 0.018 Machinery and equipment 0.509 -0.507 -0.165 Other manufacturing 0.034 0.064 -0.094 C onstruction 0.759 - 0.596 - 0.274 Commerce and hotels 0.110 -0.066 -0.129 Financial services 0.016 0.014 -0.060 Medicinie 2.592 -2.388 -0.822 Note: These numbers represent the means for 1980-90 of the corresponding sectoral elasticities. 7?K,I, a In x,1,/a In g, Sousrce: Auth ors' calculation.c result of a positive cost elasticity of output in relation to infrastructure is robust with respect to specification of the structure of the cost function, as well as to aggregation of data. We therefore estimate the aggregate cost of production in the Mexican economy as a function of total output, relative interest and wage rates, and stocks of infrastructure. Of course, the translog specification could be used to estimate the aggregate cost function. We have not done so, however, because we wish to show that even a highly simplified version of our methodol- ogy retains the same general conclusions. Estimates using a translog specifica- tion yield results broadly similar to those reported in tables 2 and 3 for individ- ual industries. The Cobb-Douglas specification permits a conventional interpretation of t-statistics and also indicates how a policymaker may use a simple methodology to make an initial judgment as to the benefits of public infrastructure. Because a cost function must be homogeneous of degree one in factor prices, the relative factor price enters as an explanatory variable. The estimates are log (C) = 9.176 + 0.172 log (RIW) + 0.494 log (Y) - 0.120 log (GE) (7.49) (10.37) (6.27) (-1.39) + 0.128 log (GT) - 0.019 log (GC) ( 1.37) (-0.72) R2 = 0.99, DW Stat = 1.84 where C denotes total cost; R, rate of interest; W, wage rate; Y, GDP; GE, electricity infrastructure; GT, transport infrastructure; and GC, communica- tions infrastructure. Several aspects of the estimates are interesting. First, the cost elasticities of infrastructure have the same signs as those in our estimation of a translog cost function, that is, negative for electricity and communications and positive for Feltenstein and Ha 301 transport. Second, these elasticities are not statistically significant. This is be- cause the relationship between cost and infrastructure is probably more compli- cated than a log-linear functional form. It suggests, therefore, that a more gen- eral specification of the cost function, for example, a translog function, should be used. Recall, however, that the translog function includes variables in more than one term, so that its t-statistics should not be judged as with a conventional regression. Thus we should not make direct comparisons between statistical significance in this estimation and the estimates reported in tables 2 and 3. As a matter of fact, the log-linear function imposes an ad hoc assumption that the elasticities are time invariant. V. SUMMARY AND CONCLUSION We have constructed a data base for Mexico that includes time series for an aggregation of the national accounts into sixteen sectors. In particular, we have developed a series for sectoral capital stocks, employment levels, outputs, and corresponding prices and wages. We have considered three types of infrastructure-electricity, transport, and communications-and have derived annual stocks of each type of infrastructure. We have then estimated sector- specific cost functions in which the cost of output depended on wages; the cost of capital, represented by the interest rate; and the nominal values of the stocks of the three types of infrastructure. We conclude that infrastructure in electricity and communications generally reduces the cost of sectoral production but transport infrastructure tends to increase costs of sectoral production. These results are similar to those in studies carried out by Jarque (1988) and Shah (1992) in Mexico. Jarque, in particular, explains the counterilItuitive result for transport by noting that the sector has often been used by goverinments as an employer of last resort. Thus increases in spending on infrastructure are largely nonproductive and cause increased tax burdens, leading to the observed outcome. It should be noted, however, that Shah's (1992) results differ in a number of ways from ours. In addition, we find that, although in general, labor and infrastructure are substitutes, in the case of electricity and communications infrastructure, capital and infrastructure are complements. In the case of transport infrastructure these conclusions are re- versed. Here our results are similar to those of Berndt and Wood (1975, 1979) for the United States. It would therefore appear that public expenditure on electricity and communi- cations has enhanced the productivity of private production, but expenditure on transport may' actually have had a detrimental effect on private output. These results should be treated with caution, however, given the uncertaini nature of data on infrastructure stocks. APPENDIX. ESTIMATION TECHNIQUE We estimate equations 2 and 3 simultaneously because it is likely that the disturbances from the two are correlated. We use a seemingly unrelated estimate, 302 I Ii. W0RI I) BANK [I()NOMI R FVII- W. VO[. N(o _' which gives more efficient estimates than regressions applied separately to each equation. Furthermore, because the two equations share the same parameters, they must be estimated jointly to impose cross-equation constraints. We assume that the disturbances of the two equations are independent across observations but have free covariance across equations. A consistent estimate of the covariance matrix is formed and used to weigh the observations when the equations are reestimated. The objective function can be written as (A-I) Q(b) = e(b)' (S-1 - IT) e(b) where e(b) is the vector of stacked residuals (a function of the parameter vector b), S is an estimated covariance matrix of the disturbances, and IT is the identity matrix with an order equal to the number of observations. If S is recomputed from b(i) at each iteration, this estimator converges to the maximum likelihood estimator if the disturbances are assumlled to be multivariate normal. For a de- mand system, this method yields estimates which are invariant with respect to which equation is dropped. The regularity conditions on the parameters required by the properties of the cost function are considered in the estimation. The cost function should be hlomogenieous of degree one in factor prices. There are two inputs in the model: labor and capital. We therefore divide total cost bv the rental price of capital (the interest rate) and use the relative price, wir, as one of the explanatory variables in equations 2 and 3. For the cost function to be concave in input prices, the Hessian matrix [a2C/1aivjrtvIj, of the cost function should be negative semi- Table A-1. Stocks of Infrastruicture, 197/0-90 (millions of pesos at 197( constant prices) Year Elet tricitv Trans prl c(oll))fllifliCcltioils 1 97( 1 9,) 1 3.1 9,06 '.5 147.8 149 1 2 2,8.4 9,562.9 153.6 1972 26,09 1.1 9,86.7 180.7 197.3 30,14 1.8 10,066.8 218.2 1 974 33,986.5 I o,045. 8 248.3 1975 37,846.9 12,244.8 1 398.7 1976 44,798.6 1 7,0(5. 9 3,658.4 1977 46,669.3 I 8,495.8 5,524.5 1978 51,148.1 19,536.6 7,458.9 1979 55,761.8 2)0,368.6 9,137.7 198( 6.3,419.0 20,92 1.4 8,436.3 1981 70,175.4 22,809.5 10,342.3 1982 73,962.6 23,875.1 1 ,124.0 1983 73,587.9 24,5 11.5 1 1,276.7 1 984 71,613.8 24,630.4 11,818.7 1 985 69,216.9 25,21 3. 3 12,471.3 1986 66,555.8 25.,096.4 11,689.6 1987 64,195.5 26.58.3.) 1 1,975.3 1988 65,630.1 .317 ,46.4 1.3,409.7 1 989 65,340.5 3(0.763.4 12,769.1 1990 6.5,1 1(.6 29,372.1 12,564.5 So/urce: Banco dc Mexico (1992!. c1-lteustemz and Ha 3() '3 definite. Following previous empirical studies of translog functions, we do not impose inequality constraints in the estimation. We assume that the errors associated with equations 2 and 3 are optimizing errors and are jointly distributed with zero expected values and with a positive definite covariance matrix. The rationale for the stochastic specification is that firms may make ranidomii errors in clhoosing their- cost-minimizing input bundles. REFERENCES The word "processed" describes informally reproduced works that mayi not be com- monk' available through library systems. Aschauer, D. A. 1989. "Is Public Expeildituire Produictivc?" journal oJ Monetary Eco- nomics 23(2, March):7- -200. 1990. "Whl Is IntrastRIctUre Important?' In Alicia H. Munnell, ed., Is Tlerc a Shortfall in Public Capital lnvestment?' Proceedings of a conferenice held at Harwich Port, Mass. Junie. Conference Series 34. Federal Reserve Bank of Bostoin. ALierbach, Alan J. 1990. "The Cost of C'apital and Investimienit in De[veloping Countries." wps 410. World Bank, Country Fconomilics Department. Wlashington, D.C. Processed. Banco de Mexico. 1992. La iKncufesta dc Accrv'os. D)cpreciacin V Formaci(on tic Capital a'el Banco de Mexico. Direcci6in de Investigaci6n Fconomilica. Barro, Robert. 1990. "Government SpendinIg in a Si inpIc Model of Endogenous Growth." journal ot-IPolitical Econiomiy 98(SUPPIC11em t): 103-25. Berndt, Ernst R., and Bengt Haisson. 1991. "Measuirinig the Contribution of Public Infrastructure Capital in Swedeni." NBER Workmig Paper 3842. National Btirealu of Economilc Researchi, Ca mbridge, Mass. Processed. Berndt. Ernst R., and David 0. Wiood. 1975. "Techlinology, Prices and the Derived De- mand for Energy." Rev'iew' of lE-oniomiics and Statistics 57(3, AugLust):259-68. 19 1979. "Enginieerinig anid Econiometric Interpretationis of Enaergy Capital Comple- mentarity." American bconomic Rev'iei 69(3, .1J1ic:342-54. Duffy-Deno, K. T., and R. W. Eberts. 1991. "Public InfrastrIuCtiure and Regional Eco- nomic Developimicint: A SimulltanleoLis Equations Approach." Journal o° Urban Eco- nzo;nics 39(3, November): 39-43. Feltensteini, Andrew, and liming Ha. 1994. "The Social and Economic Benefits of Mexi- can InfrastructurC." World Bankl, Latin America and the Caribbean Departmelit, Washington, D.C. Processed. Feltensteini, Andr-cw, anid Stephen Morris. 1990. "Fiscal Stabilization and Exchaiige Rate Instability: A Thcoretical Approach and Somni Policy Conclusion1s Using Mexican Data." Journal of Public Economiics 42(3. August):329-56. Ford, Robert, and P'ierie Poret. 1991. "Infrastruictiure and Private Sector Productivitv." Workinig Paper 91. oE(:i) Ecoiiolilic Studies 17. Organization for Ecoiionoic Coopera- tion and Developimienit. Paris. Processed. Governmeint of Me\ico. Various issLues. .Siste'niia de (ite7tas Nacionales (1c Mexico. Of- fice of Programiniig and BuLdget, Mexico CitN. Ha, jiming, and Sweder \an W\iinhel-gen. 199. D)raft. "Estimatinig Returns to Scale of MCXiCan ManUfaCtUring IndUStrieS." International M\1onetary FUnId (IMF), EUrope 11 I)epartment, Wash i ngto in D.C. Processed. 304 1il H WoRi 1) hANK I( ONONI Rl Vl \W. VOl ., N() ' Hulten, Charles R. 1992. "Growth AccoulitilIg Wheln Techinical Change Is Embodied in Capital." Americaln Economic Review 82(4):964-80. Hulten, C. R., and Robert M. Schwalb. 1991. "Is There Too Little Public Capital? Infra- structure and Fconomiiic Growvth." Paper presented at the America.n Enterprise Institute (AEI) Conlferenlce on1 InfrastrIlcture Needs and Policy Options for the 1990s. February. Washlilngton, D.C. Processed. IMF (International Mlonctary Fulid). 1993. M'exicO: Recent Econ(om11ic Developments. Washingtoni, D.C. Processed. . Internationlal Financiall Statistics. Various issues. Washinlgtol, D.C. Jarque, Carlos NM. 1988. "An Empirical Study of the Deterniinants of Production in Mexico." Secretariat of Budget and Planning, Mexico City. Processed. Munlinell, A. H. 1990a. "tiow Does Public InfrastruLCture Affect Regional Economic Performance." Federal Reserve Bank of Boston. Newz England Fconiomic Reviewu', no. (September/October): 11-32. 1 990b. "'Whv Has Productivity Declinied? Productivity and Public Investimienlt." New England Econo mnic Review, no. ( ua rv/Februarv(:3-22. Nadiri, M. I., and T. 1. Man. liueas. 1991. "Thc Effects Of Public Infrastructure and R&D Capital on the Cost Stricture and Performance of U.S. Manufacturing Industries." NBER Working Paper X3887. National Bureau of ECOn1OI11;C Researci, Cambridge, Mass. Rajagopal, Dagmnar and Anwar Shah. 1990a. T'The Effectivenless of Tax Incenitives in a Rational Expectations Model of Investmenit Behaviour." NYorld Bank, Public Eco- nomi.cs I)ivision, Wi.ashington, D).C. Processed. . 1990)b. "Testing for the Effectiveniess of Investnieiit Incenitives in Physical and Knowledge Capital." World Bank, Public Economics Division, Washington. D.C. Processed. Shah, Anwar. 1988. "Public Infrastructure. Industrial Productivity and Profitability in Mexico." Policv; Planiiiig, and Research Workling Paper 100. World Bank, Policy Research Deliartment, Washington, D.C. Processed. 1992. "Dyn1am;ics of P'ublic InfrastruLCture, Industrial Productivitv and Prof- itablltv." Rev'iewc oi Fconiomliic-s and.Statistics 74(Fehruarv):28-36. Shah, Anwar, and J. Raffes. 1990. "Tax Policies and Investillent in Phvsical and R&D Capital: An Enmpirical Applicationi of a Dvinamic Factor Demand Model with Endo- genous (Capacitv Uitilizatioil." World Baiik, Public F.cononiics Division, Washington. D.C. Processed. Tatom, J. A. 1991. "Public Capital and Private Sector Performilance." World Bank, Wash- ington. D.C. Processed. Tybiout, James R. alnd M. Daniel Westbrook. 1991. "Traide l.iberalization and the Struc- ture of Productioll in Mexicani Mallufacturing Industries." CGeorgetowni ULniversitvy Departmenit of Econoniics, Washiiigton, D.C. Processed. Winstonz. C. NI. 1990. "How Efficienit Is CurrenIt InfrastruLcture Spetiditig and Pricinig?" In Alicia H. uInn111CII, ed., Is Tbere a1 Sbortbl ll in Public Capital Inlnestm7hLnt? Proceed- ings of a coniferenice held at Harwicli Port, Mass. lune. C'oniferenice Series 34. Federal Rescrve Banik of iBoston. Processed. 1991. "Efficcint Transportation Infrastruicture Policy." Journal o lEconomic Per- spectives 5(1, Winter):I I 3-2- T H E (OR LI) R A N K I ( iNoMIi( R I V I L W, \ OI 9, N O . ': 3 0 5 - 3 3 3 The Current Account in Developing Countries: A Perspective from the Consumption-Smoothing Approach Atish R. Ghosh and Jonathan D. Ostry According to the consumption-smoothing view, a high degree of capital mobility im- plies that agents are able to fully smooth their consumption in the face of shocks. This article develops a framework to test whether, indeed, the current account in developing countries acts as a buffer to smooth consumption in the face of shocks to national cash flow, which is defined as output less investment less government expenditure. Using vector autoregression analysis, we estimate the optimal consumption-smoothing cur- rent account with data Jrom a sample of forty-five developing countries. We find that for a majority of the countries, the hypothesis of full consumption smoothing cannot be rejected, suggesting that capital mobility may alter all be quite high in this group of countries. Several recent studies (for example, Mathieson and Rojas-Suarez 1992 and Montiel 1994) have suggested that the effective degree of capital mobility in developing countries has been increasing in recent years. Although the vast majority of developing countries continue to maintain some form of restriction-such as exchange controls and quantitative restrictions, for example-on capital movements, these restrictions were not very successful in stemming the large capital outflows ("capital flight") that took place in the 1970s and 1980s. When developing countries have resorted to capital controls to stem occasional surges of inward capital flows, these controls have been largely ineffective, and evasion has been widespread. Recent evidence has shown that, even in developing countries with extensive capital controls, domestic interest rates have tended to move quite closely with international interest rates adjusted for expected exchange rate changes. Again, this suggests that the effec- tive degree of capital mobility in these countries may be quite high. Although these studies are suggestive, there are a number of problems with trying to gauge the degree of capital mobility by simply looking at either the Atish R. Ghosh is in the Woodrow Wilson School of Public and International Affairs at Princeton University. and Jonathan D. Ostry is in the Research Department of the International Monetary Fund. Work on this article was completed while Atish R. Chosh was a consultant with the World Bank. The authors thank Peter Montiel and Carmen Reinhart for helpful comments on a previous draft and Ava Ayrton-Lilaoonwala for assistance with the data. © 1995 The International Bank for Reconstruction and Development/THE WORLD BANK 30 5 306 IHF X;R.I) DANK0K N()N()I( REA V 1, the country is tilting con- sumption toward the future. For 0 = 1, the consumption-tilting component is identically zero, and consumption is equal to the country's permanent cash flow. The primary focus of our analysis is on the consumption-smoothing component of the current account. We abstract from long-term trends in foreign saving and focus instead on the short-run dynamics of the current account around its trend. The model ignores liquidity constraints, imposing only the intertemporal solvency constraint. Because absence of liquidity constraints are part of the null, the empiri- cal findings below should shed light on the extent to which liquidity constraints may have been important. Because there is no reason to assume that the consumption-tilting parameter will be unity in all cases (see table 1 for the em- pirical estimates), it is necessary to detrend the current account data by first removing the consumption-tilting component. Removing this component will en- able us to identify correctly the consumption-smoothing component of the cur- rent account, with which the model is concerned. As shown below, the consump- tion-smoothing component will be a stationary time series, which has a number of econometric advantages because standard statistical tests can be applied.5 3. The separability between the consumption and investment decisions lets us write output as exog- enous to the consumption decision, but not (necessarily) to the investment decision. Specifically, if output is given by q, = ca, f(k,), where a is a productivity shock, f(k) is the production function, and k is the capital stock, then a shock to a, affects current output but not investment, whereas a future shock to productivity, a,+, j > 0, affects investment and future output (both by changing productivity and by changing the capital stock) but leaves current output unaffected. 4. Under quadratic utility, f = ,Sr(1 + r)/[0(1 + r)2 - 1]. Clearly, when S = 1/(1 + r). 0 = I and there are no consumption-tilting dynamics. Whether the consumption-tilting parameter is above or below unity depends on the relation between the interest rate and the time preference rate. S. Our model identifies the stationary component of the current account with consumption-smoothing behavior; more generally, it could include other transitory factors. 310 THF. W OR I) BANK B ONO MI( REV Il+', V( I.. NO. N Table 1. Unit Root Tests and Cointegrating Parameter Cointegrating parameter Augmented Dickey-Fuller between consumption and statistic national cash flow inclusive of Current National cash interest payments Country account flow; Estimate Standard error Africa Bostwana -0.54 -3.47 1.61 0.08 Ethiopia -2.32 -4.21 0.69 0.04 Ghana -1.67 -4.25 0.93 0.04 Kenya -2.81 -4.23 0.80 0.03 Liberia - 3.06 - 2.94 1.13 0.18 Malawi -3.00 -3.90 0.83 0.04 Mauritius -4.03 -4.66 0.86 0.07 Morocco -2.89 -3.52 0.81 0.04 Nigeria -3.02 -3.45 0.31 0.10 Senegal -3.74 -4.08 0.86 0.03 Tanzania -0.86 -4.76 0.36 0.06 Tunisia -2.71 -3.22 0.95 0.02 Zambia -5.72 -5.11 0.77 0.10 Asia Hong Kong -3.13 -2.45 1.10 0.02 India -3.56 -3.30 0.95 0.01 Indonesia -3.42 -3.37 0.75 0.04 Korea, Rep. of -3.52 -2.11 1.00 0.02 Malaysia -4.31 -3.21 0.93 0.08 Pakistan -3.17 -1.93 1.04 0.02 Papua New Guinea -2.53 - 2.99 0.49 0.15 Philippines -4.00 -3.30 0.98 0.01 Singapore -1.44 -2.10 1.20 0.04 Sri Lanka - 5.16 - 3.43 0.74 0.02 Thailand -5.08 -1.64 0.93 0.01 Middle East Egypt -1.91 -3.35 0.37 0.04 Iran, Islamic Rep. of -3.37 -4.43 1.22 0.11 Israel -2.27 -3.52 0.89 0.03 Jordan -2.70 -3.17 0.44 0.05 Saudi Arabia -2.47 -4.17 0.45 0.20 Latin America and the Caribbean Argentina -2.26 -3.75 0.92 0.04 Bolivia - 2.56 -3.17 1.16 0.04 Brazil -3.31 -2.32 0.97 0.03 Chile - 1.99 -3.64 0.66 0.09 Colombia -3.24 -3.28 0.96 0.02 Ecuador -4.45 -3.74 0.88 0.03 El Salvador -3.05 -4.51 0.82 0.02 Guatemala -3.83 -3.00 0.91 0.01 Honduras -4.08 -4.09 0.80 0.01 Jamaica -1.58 -3.701 0.65 0.09 Mexico -4.23 -3.07 0.99 0.02 Panama -2.31 -3.25 1.02 0.04 Paraguay -3.40 -2.93 0.87 0.01 Peru -3.92 -3.49 0.98 0.04 Uruguay -2.00 -3.78 0.93 0.04 Venezuela -6.00 -3.97 1.03 0.04 a. In first differences, A(q, - i, - g,). Source: Authors' calculations based on World Bank and IMF data. Chosh and Ostrv AIl The consumption-smoothing component is given by (4) ca*= y, - it - g, - Oc, where ca, is the consumption-smoothing current account, y, is the gross national product or GDP plus interest income on existing foreign assets, q, + rb,. Substi- tuting equation 3 into equation 4 yields, after some manipulation, x 1 (5) ca, = -L(+ ,[E,A(q,+, - itg,+ , 1+)] where A is the (backward) difference operator, Ax, x= - x,_. Equation 5 shows that the consumption-smoothing current account is identically equal to minus the present discounted value of expected changes in national cash flow.6 Equation 5 thus embodies the intertemporal approach to the current account in a clear and simple way. Shocks to national cash flow (or to any of its components-output, investment, government spending) that are expected to be permanent have no effect on the current account, because their expected change is zero. By contrast, favorable transitory shocks lead to improvements in the current account, and the extent of the movement in the current account is a decreasing function of the persistence of the shock. Equation 5 shows that creating the consumption-smoothing current account series requires estimating the present value of expected changes in national cash flow, where the expectation is conditional on the information set used by eco- nomic agents. This is a daunting task because, in general, the information set used by agents to forecast future values of these variables is unknown to the researcher. It turns out, however, that precise knowledge of what information is employed by the agent is not needed. This is because, as shown by Campbell and Shiller (1987) in a somewhat different context, the current account itself reflects all the information available to agents for the purpose of forecasting these variables. Therefore, by including the current account in the conditioning infor- mation set, we can fully capture agents' expectations of shocks to output, invest- ment, and government expenditure. Following Campbell and Shiller (1987), we estimate an unrestricted vector autoregression (VAR) in [A(q, - it - g,), caj], where cat is the actual (detrended) current account, defined (in analogy to equation 4) as cat = y, - it - g, - Oc,. Some estimate of 0 is required for the VAR estimation. We explain below how an estimate of this parameter may be obtained. The VAR may then be written as I ca, * t1221 c a,_l I or more compactly as (7) xt =t Ix + e,. 6. It is similar to Campbell's ( 1987) expression for household saving as the expected present value of future declines in household labor income. 312 THF WORIDB) IANK ((flN(OMI RFVWW. \ Nl .9. From equation 7, the k-step-ahead expectation is simply (8) E,(x,,k) = 4kx, so that E,A(q,±k - -k = [1 x],kX. Therefore, it is possible to write the consumption-smoothing current account defined by equation 5 in terms of the VAR given by equation 6. Specifically, the expression for ca* (in equation 5) is given by 1 (9) ca,= - = ( 0]ZX (1 + ry[1]I.x = -[1 0]I/(1 + r) + X j~0O (1 r) i = -[1 0][*/(1 + r)][I - *I/(1 + r)] x, rx,. This expression is valid as long as the infinite sum in equation 5 converges, which it will if the variables in the VAR are stationary.7 Assuming that (q, - it - g,) is 1(1), its first difference will be stationary. Because under the null the current account is a discounted sum of A(q, - i, - gt), it too will be stationary. An important implication of the intertemporal smoothing model is that the current account should, in general, Granger-cause changes in national cash flow. That is, in a regression of the change in national cash flow on the lagged current account and the lagged change in cash flow, the coefficient on the lagged current account should be statistically significant (Sargent 1979). From equation 5, cat is equal to (minus) the expected present discounted value of A(q-i -g), where the expectation is conditional on the agents' entire information set. If agents have more information about the evolution of national cash flow than is contained in past values of the national cash flow, then the current account ought to Granger-cause changes in cash flow. If, for example, a change in administrations portends higher future government spending, then the country should run a current account surplus. The surplus would then Granger-cause the subsequent decline in national cash flow. This is completely analogous to the notion, presented by Campbell (1987) in his study of the permanent income hypothesis, of "saving for a rainy day." It remains only to describe how to calculate the consumption-tilting parame- ter, 0, so that the actual data on the current account can be purged of its consumption-tilting component. As argued previously, the optimal current ac- count series, ca*, will be an 1(0) process. Under the null hypothesis that the actual consumption-smoothing component of the current account is equal to car', the actual (detrended) current account is also 1(0). This means that the left- hand side of equation 4 is 1(0) and therefore that 0 may be obtained as the cointegrating parameter between consumption, c,, and national cash flow inclu- sive of interest payments, y, - i, - g,. 7. Therestrictionsonthe r coefficientsimplythat 'I = -421 and22 - 412=(1 + r). GIbosb and COstry 313 If the theoretical model-with its infinitely lived representative agent who has a constant subjective discount rate-is taken literally, then 0 would be constant over the entire sample. Moreover, values of 0 that differed from unity, although not at all inconsistent with the theoretical model, would have the troubling implication that the most patient country would eventually own the entire world. We do not believe that such an extreme conclusion is necessarily war- ranted. Instead, we view the use of the infinite-horizon, constant-discount-rate model as a simple abstraction. The model provides a practical means of remov- ing the trend in the current account that results from, among other things, shifts in demographic and other factors not captured here and allows us to focus on the consumption-smoothing aspect of the current account, which is our primary interest. Once the optimal current account series, ca;:, has been calculated, a number of tests may be performed. First, an implication of the intertemporal model is that the current account should Granger-cause subsequent movements in na- tional cash flow. This is easy to test using the results of the VAR estimation. Second, again using the VAR estimates, the model implies two restrictions on the parameter values. Specifically, from equation 9, the restrictions on the parame- ter vector r- = [r, F] are (10) [Il ',, -, -[I1 0]w(1 + r)] [Il- ifI + r)]- = [0]' that is, F, =0, and r,, = 1. The requirement that the coefficient on national cash flow, Fr, be close to zero and that on the current account, r,a, be close to unity can easily be tested. Third, under the null, the variances of the actual consumption-smoothing current account and the optimal consumption- smoothing current account should be equal. This equality-of-variances restric- tion can also be tested. 11. DATA AND ESTIMATION A large cross-sectional sample of developing counltries for which the necessary national accounts data were available over a sufficienitly long period was chosen for the analysis. The main source for all data was INMF (various issues). Details of data sources and sample periods are provided in appendix table A-1. In the case of a few countries, for which the IMF data either contained errors or were unavailable over a sufficiently long period, data from World Bank (various years) were used instead. Thirteen African countries were employed for the analysis: Botswana, Ethiopia, Ghana, Kenya, Liberia, Malawi, Mauritius, Mo- rocco, Nigeria, Senegal, Tanzania, Tunisia, and Zambia. Eleven Asian countries were used: Hong Kong, India, Indonesia, the Republic of Korea, Malaysia, Pakistan, Papua New Guinea, the Philippines, Singapore, Sri Lanka, and Thai- land. Five Middle Eastern countries were included: Egypt, the Islamic Republic of Iran, Israel, Jordan, and Saudi Arabia. Finally, sixteen countries from Latin America and the Caribbean were used: Argentinia, Bolivia, Brazil, Chile, Co- 314 iIF I'5OR I 1) iHANK I ('NO)N[( RIViI IW VOI . 'N. ) lombia, Ecuador, El Salvador, Guatemala, Honduras, Jamaica, Mexico, Panama, Paraguay, Peru, Uruguay, and Venezuela. All data are at an annual frequency and cover in most cases about thirty observations ending in 1990. Granger-Causality Tests As described in the previous section, the empirical analysis first ensures that the variables entering the VAR are rendered stationary. This is done by working with the national cash flow in first differences and the current account series with the stochastic trend removed, as described above. The Augmented Dickey- Fuller (ADF) test statistics for the current account and national cash flow (in first differences) are given in table 1. The test statistics are based on regressions in which the dependent variable is the first difference either of the consumption- smoothing current account or of the first difference of national cash flow. The regressors include a constant, the lagged level of the corresponding variable, and up to ten lags of the first difference, with insignificant lags dropped. The ADF statistic is the t-statistic on the lagged-level variable. From the critical values reported in, for example, Engel and Yoo (1987), it can be seen that in the majority of cases the variables in question are indeed stationary. These results should, however, be interpreted in light of the observation that the ADF test has limited power to reject the null of a unit root in favor of near alternatives. Table 1 also reports the cointegrating parameter between consumption and national income net of investment and government consumption. In a number of cases, the cointegrating parameter differs significantly from unity, indicating the presence of consumption-tilting dynamics. The results for the cointegrating pa- rameter are on the whole sensible, being for the most part below unity, as would be expected for a group of mostly borrowing developing countries and as the literature on "stages in the balance of payments" would suggest. Notable excep- tions include a number of middle-income developing countries, for example, Botswana, Hong Kong, Iran, Korea, Singapore, and Venezuela. Finally, it may be noted that, because 0 is estimated by means of a cointegrating regression, its use in the subsequent statistical tests is legitimate. As mentioned above, the first implication of the intertemporal model is that the current account should Granger-cause subsequent movements in national cash flow. We use a standard t-test to gauge whether the data satisfy this impli- cation. The results are reported in table 2 for the different regions. In Africa (table 2), the current account Granger-causes national cash flow in about half the countries: Kenya, Malawi, Mauritius, Senegal, Tunisia, and Zambia. Interestingly, in virtually all the countries the sign of the coefficient estimate is negative, implying that the country increases its current account surplus (reduces its deficit) when national cash flow is expected to decline in the future, as theory requires. For Asia (table 2), the results suggest Granger causality in five cases: Hong Kong, India, Korea, Papua New Guinea, and the Philippines. In four additional cases (Indonesia, Malaysia, Singapore, and Thailand) the results are quite close Gbosh and Ostry 31.5 Table 2. Granger-Causality Test Statistics Test Test Country statistica Country statistic, Africa Middle East Botswana -0.84 Egypt -1.24 Ethiopia -0.98 Iran, Islamic Rep. of -2.59* Ghana -0.85 Israel - 1.08 Kenya -1.98* Jordan -2.99# Liberia -0.51 Saudi Arabia -3.50* Malawi - 2.55 S Mauritius -3.61 * Morocco -0.74 Nigeria - 1.48 Senegal - 3.72* Tanzania - 1.82 Latin Arnerica and the Caribbean Tunisia -2.57* Argentina 1.82 Zambia -4.57* Bolivia -3.07* Brazil -2.74* Chile -2.41 * Asia Colombia -1.85 Hong Kong -2.29* Ecuador - 3.09 * India -2.89* El Salvador -4.07* Indonesia -1.55 Guatemala -5.15* Korea, Rep. of -2.67* Honduras -3.40 Malaysia -1.67 Jamaica -1.78 Pakistan 0.58 Mexico -3.88* Papua New Guinea -2.89* Panama -2.94* Philippines -3.63* Paraguay - 2.89 * Singapore - 1.41 Peru - 1.62 Sri Lanka -0.66 Uruguay -0.64 Thailand - 1.48 Venezuela -4.73* * Indicates statistical significance at the 95 percent level. a. t-statistic on the lagged (detrended) current account in a regression of national cash flow (in first differences) on lagged national cash flow (in first differences) and lagged current account. Source: Authors' calculations based on World Bank and IMF data. to passing the test at standard significance levels. The coefficient estimates are negative (as required by theory) in all but one of the eleven Asian cases. Only for Pakistan and Sri Lanka is the Granger-causality restriction clearly not satisfied. Apart from India, where previous research has found evidence of relatively effective capital controls over much of this period (Montiel 1994), the results reported in table 2 are broadly consistent with those found in previous literature that uses different analytical approaches. The results for the Middle East are also encouraging: three of the five coun- tries pass the test at standard significance levels, and all the point estimates are negative, as predicted by theory. Finally, the Latin America and the Caribbean region, which contains some of the relatively more developed countries of our sample, performs best among the regions: eleven of the sixteen countries exhibit Granger-causality from the current account to national cash flow, and a further three are borderline. Again, in virtually all cases, the point estimates are negative as required. On the whole, therefore, this first implication of the consumption- 316 THF F.'WORIID IUANK F(O )N0 MI REVIFEW, V\)L.. 9. NO 2 smoothing approach is satisfied for about 60 percent of the countries in the sample. This result compares favorably with estimates of the model for the large industrial countries (see Ghosh 1990). Formal Tests of the Model Before turning to the formal tests of the model, it may be useful to show visually how well the model performs in terms of accounting for actual current account movements in developing countries. To this end, we use the VAR esti- mates for each country to generate a predicted current account series. This benchmark time series can then be compared visually with that of the actual (detrended) data to determine how well the consumption-smoothing model tracks current account developments. Such a visual comparison, although less rigorous than the formal statistical tests presented below, has the advantage of being less subject to the objection that the low power of statistical tests may be driving our results. In addition, the visual impression conveyed by the figures can be related to the parameters of the VAR given in appendix table A-2. Specifi- cally, using the definitions of the actual and optimal (detrended) current account, ca,>= -[1 0][*/(1 + r)][I - 'I'/(1 + r)]- 1x, is the optimal current account, and ca, = [0 1]x, is the actual current account, it is clear that the two will be equal if -[1 0][I/(1 + r)][I - */(1 + r)]-l = [0 1]. Postmultiplying by [I - * /(1 + r)] and adding [0 1I[I/(1 + r)] yields -[1 0][*/(1 + r)] + [0 1][I/(1 + r)] = [-1 1][*/(1 + r)] = [0 1] if actual and optimal (detrended) current accounts are equal. Thus, the model will fit well whenever the elements of the first column of the VAR matrix are approximately equal, and the difference between the elements of the second column is equal to (1 + r). Figures 1 to 4 plot actual and predicted (detrended) current account balances for the forty-five countries in the sample.8 In the vast majority of countries, the actual and predicted variables move closely together, suggesting that the inter- temporal optimizing model indeed can account for developments in the current 8. As mentioned previously, the variables in the figures are detrended. However, because the consumption-tilting parameter is in most cases close to unity (see table I ), the variance of the actual and detrended current accounts (in first differences) should be similar. In fact, on average over the sample, the variance of the change in the detrended current account is equal to about 92 percent of the variance of the change in the raw (undetrended) current account. Ghoslb and Ostrv -317 accounts of a significant proportion of developing countries.9 It may also be noted that the visual impression conveyed by the figures when actual and pre- dicted current account movements diverge is in line with our priors. For exam- ple, for a number of countries affected by the debt crisis-including to some extent Argentina, Chile, the Philippines, Uruguay, and Venezuela-actual cur- rent account balances (surpluses) exceed optimal (predicted) current account balances for most of the period since 1982, as would be expected if capital inflows had been less than desired. Turning now to the formal statistical tests, it is clear from equation 9 that if the consumption-smoothing model is valid, then the coefficient on A(q - i - g), ry, should be zero and that the coefficient on ca, r;,, should be equal to unity. In other words, the actual (detrended) current account should be precisely equal to the optimal consumption-smoothing current account defined by equations 4, 5, or 9, which will be the case if the parameter restrictions FP, = 0 and F,a = 1 are satisfied. These two parameter restrictions can be tested individually by the use of a standard t-test and jointly by a chi-squared test. 1" The results are once again quite encouraging (see table 3). As far as the coefficient on cash flow (J,) is concerned, the vast majority of the point esti- mates are extremely small (between minus one-quarter and plus one-quarter) as predicted by the theory. In about one-quarter of the cases, the estimates of ry are statistically different from zero, but even within this group of countries the point estimates are frequently quite small. For r,,,, again the vast majority of the estimates are not statistically different from unity, as the consumption-smoothing model requires. The point estimates are also close to unity in all but a few cases, implying, as noted with reference to the VAR parameters, that the point estimate for ',' is close to that for I21 and that 'I22 - '12 is close to (1 + r). As far as regional differences are concerned, the parameter r, differs most frequently from its theoretically predicted value in the African countries, where the null hypothesis is rejected in three of the thir- teen countries. This pattern of rejection can also be seen in appendix table A-2, which reports the underlying VAR parameters. The incidence of rejection is quite small in the other regions, with two rejections among the eleven Asian countries, zero rejections among the Middle Eastern countries (the sample of which is admittedly quite small and where standard errors were quite large in two of the 9. With respect to the raw (undetrended) series, the model on average accounts for about 74 percent of the variance of the change in the current account. 10. The standard errors need to be computed numerically as v(r)'7V(F), where E is the variance- covariance matrix of the parameters of the VAR, and v(r) is the gradient of [r,, r,,,j with respect to the VAR parameters. The standard errors used in table 3 are White heteroskedastic-consistent standard errors calculated as =(xx) '(X E,e,X)(X'X)- where e, and e, are the residuals from the ith and jth equation of the VAR. The x2 statistics reported in table 3 follow the chi-squared distribution with degrees of freedom equal to the number of parameter restrictions, in this case two. 318 THE WORLD BANK[O:ONOMIC1( RFVIEW.,V(OL..q,No ' Figure 1. Predicted Detrended and Actuial Detrended Current Account. Selected Countries in Afrnca Botswana Ethiopia 400 - 800 7 200- 400 0 200O -200- . -200 -400- -600 -40(1 1961 1966 1971 1976 1981 1986 1964 1969 1974 1979 1984 1989 Ghana Kenya 40,000 - 6(00 4,000- 2(0,000 - 2,0(o -2,0(0 -2(0,000 - 4,01(- -6 .0oo) -40,000 - -800( 1956 1966 1976 1986 1(69 1970 1975 1980 1985 Liberia Malawi 200- - 3(10 190- 2()() 100- 1(0 90 -50 - 10 0 -100 - 200 -150-1- - 300 1966 1971 1976 1981 1986 1965 1970 1979 1980 1985 1990 Mauritius Morocco 4,000 - 15,000 - 2,000 - 10,00( - 4 0ooo 5,11)0(5 0 - - 2.0(00 - 0- -4,000 --0,0( - 6,000 - -20,000 18 00 59 - 69 I, 9 M -20 10(I - 7 1 1954 99 64 (,9 74 79 84 89 1998 1963 1968 1973 1978 1983 1988 Ghosh and Ostrv 319 Nigeria Senegal 40.000 - 10(- _I°00 '- ' 8 20(9 -20,000 - 60 -30,00)0 - -0 -, 0 A 0 - -10,000 -2 -40 -20,000 - V '601 -30,000 -80 1961 1971 1981 1961 1966 1971 1976 1981 1986 Tanzania Tunisia 20,000 - 600 - 15000 - 400 - 10,000 - 20() - _ 9.000 0* -5,0(00 -20( - -10,00( - - 400 -1,000 -600 - 1966 1971 1976 1981 1986 1961 1971 1981 1991 Za0bia 2,000 - 1,500H 1.000H 900H - 500- -1,()() -1,500 1958 1963 1968 1973 1978 1983 Actual -------- Predicted Aote: Before being detrenided and demeaned, the origindl data were in billions of local currency. Source: Authors calculations based on data fromii IMF (various issues) and World Bank (various years). five cases), and three rejections among the sixteen countries in Latin America and the Caribbean. It may be noted that a value of IcF, > 0 (even if significantly different from unity) will result in the actual and optimal current accounts being positively correlated. That this is indeed the case is reflected by the fact that for nearly four-fifths of the countries in the sample, the correlation is above 90 percent. As for the joint test of the parameter restrictions implied by the model, two factors are obviously important. The first is how close the estimates are to their 320 THE WORLD BANK ECONOMIC REVIEW, VOI. 4, NO. 2 Figure 2. Predicted Detrended and Actual Detrended Current Account, Selected Countries in Asia Hong Kong India 20,000 20( - 10,000 . 10()- 0 Pk. -10,000( . - 10- -20,000 .3001, . -30,000 -20 -40,000 - -300- 1962 1967 1972 1977 1982 1987 19(11 1966 1971 1976 1981 1986 Indonesia Republic of Koiea 10,000 - 10,00(,0(JO 5.000 - 5,000.0)0- -5,000- '5,000.0O - -10,000 - 10.(00O,O()() 1961 1971 1981 1954 59 64 69 74 79 84 89 Malaysia Pakistan 10,000- 30.000- i,OOO - W (,000~ - o - -;i;ee ['\ - ^-o*ooo V -5,000 -- 30,00000 -10,000- - 5(,000 1971 1976 1981 1986 1963 1968 1973 1978 1983 1988 Papua New Guinea Philippines 300 - 40,000 - 200 - -n 0000\a -200 - . 20s,000 -', 300 1740 -4()19000 4-6 6 7 9 4 1974 1979 1984 1989 19/49 54 59 64 69 74 79 84 89 C;hosh and Ostrv 321 Singapore Sri Lanka 4,0X)0 - 15,0 0 3,000 - 10,00( 0 2,000(- 5,(00 A 1,000 - I - - --- U - 1,00( 10(l - 2,000 - - I5 (0 -3,0I0 -00,000 1969 1979 1989 195 1 56 61 66 71 76 81 86 91 Thailanci 100,000 _ _ I 50,(X)( - - ( 0,(( ....~ ...... - 50,000 - 100.0(10 - 150,00 _., - 20(,0001) --l-l-l-l-l-- 1951 561 66 7) 76 81 86 Actual -------- Predictedi AVote: Before being (letren(led and demeaned, the original data were in billions of local currency. Source Author,' CalCulations hased on data froITI IMF (I:arious issues) and World Bank (various years). theoretical values under the null hypothesis, and the second is how precisely the coefficients are estimated. In the African countries, the model is rejected in five cases; however, in the cases in which the data do not reject the model (eight of thirteen), parameters tend to be imprecisely estimated. For the Asian countries, where the incidence of rejection is slightly higher (six out of eleven), the parame- ters of the model are estimated much more precisely than in the African coun- tries, making rejections that much more likely despite point estimates that are reasonably close to the values predicted by the model. For the remaining coun- tries, the results are very favorable, with zero rejections out of five countries in the Middle Eastern region and with five rejections out of sixteen countries in Latin America and the Caribbean. In the latter group, the model is rejected in the important cases of Argentina, Uruguay, and Venezuela, where capital in- flows dried up for a part of the estimation period. In the case of Brazil, which was also adversely affected by the debt crisis, the model is not rejected, but in this case it is clearly because standard errors are extremely large. One factor that affects our confidence in the above results relates to the power of the statistical tests-that is the probability of a correct rejection. In this 322 I Hi woRi i) BANK vt ON)MI tt Rvviv.W. vo0.1, \(. - Figure 3. Predlicte l Detrended anid Actuical Detrentded Current Account, Selectedl Countrles irl the Middlle E.ast Fgypt Israel 4,0()0 -_ _4,00v)( - 2)1000 2,000 4,000 - -4 0wg) 6(00t 1975 198( 1985 1990 1961 1966 1971 1976 1981 1986 Islam1;ic RepubliC of Iran Jordan 4,00('0,( - 300 - 2 - 100 - 2,00o,O(Y) -~~~~~~~~~~~~0 200t 2.000,000 - 30- 0(1 I I I ~~~~~~- 400 I I I 1965 1975 1985 1970 1975 1980 1985 1990 SauLli Arabia 1()0,00(} 1 50,00() - -50(,(,(, Xr: - I()(),000 1969 1974 1979 1984 1989 ACttual -------- Predicted Aote: Before being detrended anid demeaned, the original data were in billions of local currency. Source. Author,' CalCulations based on data from ISIF (various issues) and World Bank (various years). context, the power of the test may be affected by the possibility of endogenous government behavior. For example, the government may act to smooth current account movements in the face of shocks to the economy. However, if this were the main reason for our strong results, the actual current account movements would be uniformly less volatile than movements in the benchmark series. In- stead, as we shall see below, for some countries the actual current account series is more volatile than the benchmark series, which suggests that endogenous government behavior is unlikely to provide a general explanation for the failure to reject the model in a majority of cases. Gbosh and Ostry 323 Figure 4. Predicted Detrended and Actual Detrended Currrent Account. Selected Countries in Latin America and the Catibhbean Argentina Bolivia 1.5 - 40( - 200 - -200 1.5 ~~~~~~~~~~-400 1961 1966 1971 19-6 1981 1986 195l1 56 61 66 71 76 81 86 91 Brazil Chile 2,000 800,0(0 - 1,(000- ,,,.,4(( ,()00 - ' ._ ,,,,--,,, 200,000 o- 0 A -oX\/' - 1.000 - .V . - 200,00- - 2,000) I - q0.() 1961 1971 1981 1950 61 66 71 76 81 86 91 Colombia Ecuador 600- 6,00 1 400- 40,000 - 200 200 20 .* - 20,000- v v -200- -4000 - -400- -60.000 - - 60o I - 80 .00( -1 I 1961 1966 1971 19:761981 1986 196F6 1971 1976 1981 1986 El Salvador Guatemala 1.500- LOW - 1,000 - ,00 ) 500 -1000~~~~~~o-2 0 - 500 F500- 1952 57 62 67 72 77 82 87 1951 56 61 66 71 76 81 86 91 - Actual -------- Predicted (Figure 4 continues on thefollowing page.) 324 THk.WORI )BANK F(:ONO(MI(:RFVII W,VOL.9,NO 2 Figure 4. (continued) Honduras Jamaica 400 1,500 - 300 - 1,000 200 90 100- -400 -1,9~~~~~~~~~~~~~~~~~~~~~~~~~00- -100 -900- -200 -1 -300 ,( 19S1 1901 1971 1981 1991 1961 66 71 76 81 86 Mexico Panama 4,000,000 - 800 - 3,000,000- 600 -A 2,000,000- 400 - 1,000(000- 20( - 0- -, o-k -1 000,0ooo- .; - 200 -2.000,000 400 -3,000,000- I -600- 1993 98 63 68 73 78 83 1991 16 61 66 71 76 81 86 Paraguay Peru 100000 0.06 - 50,000 - 1 °'°'' ' t ' i0.04 0.02- - S0,00( - - lOO,OOU~ ~ ~ ~~~~~% - 0-0 -) 1961 66 71 76 81 86 1961 1966 1971 1976 1981 1986 Uruguay Venezuela 50,000 - 0,00() - 40.000- 30,000 - (0,000- 20,000- 10,000- 0 - - A \ (1 ,. I - I.I A - 0.°000- - 90,000 - - 20,000- - 30,000 - 100,000 1996 1966 1976 1986 1991 96 61 66 71 76 81 86 91 A-tual -------- Predicted Note: tBefore being dcetrended and demeaned, the original data were in billions of local currency. Source: Authors' calculations hased on data from INIF (vrioLus issues) and World Bank (various years). Ghosh and Ostrv 325 In addition, as mentioned previously, our confidence in the model does not just depend on the results from the statistical tests; the time-series plots presented earlier certainly suggest that the model indeed captures economically significant movements in the current account for most of the countries in the sample. The consistency of the results reported here with those obtained by researchers using very different analytical approaches also strengthens our confidence in this ap- proach. For example, using a variety of different tests, Montiel (1994) classifies a large sample of developing countries into three categories: low, intermediate, or high capital mobility. Using his results for the overlap countries in Latin America and the Caribbean (the largest grouping in our sample), Montiel (1994) classifies Bolivia, Chile, Colombia, Ecuador, Guatemala, Jamaica, Mexico, Panama, and Uruguay as having either high or intermediate capital mobility. As can be seen from table 3 for the Latin America and Caribbean region, in seven of these nine countries, the consumption-smoothing approach would also lead one to conclude that capital flows have been sufficient to enable agents to fully smooth consump- tion, given the shocks they face. This conclusion should lend additional support to the view that the results presented here may be capturing some important aspects of capital mobility in developing countries. Variance of the Actual and Predicted Current Account As a further test of the consumption-smoothing model, we examine whether the current account in developing countries has been as volatile as would be expected, given the shocks experienced by these countries. The benchmark cur- rent account series generated by the consumption-smoothing model directly addresses this issue. If the variance of the actual current account is not statis- tically different from the variance of the benchmark current account, then we cannot reject the null hypothesis that agents indeed have been able to fully smooth consumption in the face of shocks. Table 4 provides an estimate of the ratio of the variance of the predicted current account to the variance of the actual current account. Both variables have been detrended in the manner described in section 1. The variance ratio has a standard error (not reported in the tables), but the x2 statistics reported in the table test whether the variance ratio is significantly different from unity.' 1 The results do not immediately suggest that the variability of actual current accounts has been too small (suggesting effective barriers to capital movements) or too large (suggesting excessive speculative flows) in light of the shocks hitting these economies. For example, the null hypothesis of equal variances is rejected in only three of the thirteen African countries, whereas there are only two rejec- tions among the eleven Asian countries. The results are as good or better for the remaining countries, with only one rejection among the five Middle Eastern countries, and three rejections among the sixteen countries in the Latin America 11. The X2 statistics follow a chi-squared distribution with degrees of freedom equal to the number of restrictions, in this case one. 326 THF.WORII)BANKF(oNt)MI(:RFVIEW',V 9O.qNt) 2 Table 3. Wald Tests of the Model National cashflow Current account Coefficient, t-statistica Coefficient, t-statistica Wald test Country P. rI = 0 r,a r-> = 1 statistic, x2b Africa Botswana -1.56 -2.23* 0.75 1.31 36.49* Ethiopia -0.05 -0.36 1.05 0.07 0.19 Ghana -0.04 -0.80 0.43 -3.56* 16.16* Kenya 0.20 2.22* 0.58 -1.62 5.72 Liberia 0.08 1.60 0.23 -7.00* 45.70* Malawi 0.12 1.09 0.74 -0.93 1.39 Mauritius -0.03 -0.17 0.99 -0.02 0.26 Morocco -0.04 -0.80 1.60 0.97 1.28 Nigeria -0.10 -0.56 0.88 -0.24 0.45 Senegal 0.25 2.78 1.04 0.13 8.43* Tanzania 0.21 1.24 0.80 -0.38 1.90 Tunisia 0.07 0.58 1.03 0.08 0.37 Zambia 0.15 1.50 0.57 -3.07* 9.32* Asia Hong Kong -0.60 -2.61i 1.47 0.87 7.26* India -0.18 - 1.13 2.16 1.71 3.12 Indonesia -0.12 -1.20 0.73 -1.08 3.48 Korea, Rep. of -0.83 -2.44* 1.50 1.39 9.42* Malaysia -0.21 -1.40 0.64 -1.64 5.00 Pakistan - 1.13 -3.90* -0.36 -4.12* 27.45* PapuaNew Guinea -0.19 -0.76 1.12 0.26 0.89 Philippines -0.32 -2.00' 1.13 0.35 6.05 Singapore -0.32 -2.91* 0.77 -1.05 12.26* Sri Lanka 0.26 6.50* 0.17 -6.92* 81.43* Thailand -2.89 -3.04* 2.52 1.21 25.86* Middle East Egypt 0.17 2.13 t 0.43 -1.68 5.67 Iran, Islamic Rep. of -0.10 -0.56 1.00 0.00 0.74 Israel 0.11 1.22 0.58 -1.14 1.54 Jordan -0.05 -0.25 0.92 -0.32 0.22 Saudi Arabia 0.17 0.71 1.00 0.00 0.50 Latin America and the Caribbean Argentina 0.15 1.59 -0.48 -7.79* 61.36* Bolivia -0.34 - 1.26 1.34 0.59 2.76 Brazil 0.14 0.45 2.63 0.84 0.73 Chile -0.14 -1.00 0.89 -0.37 2.13 Colombia -0.27 -1.69 1.04 0.18 4.43 Ecuador 0.05 0.50 0.93 -0.29 0.24 El Salvador -0.08 -0.67 0.80 - 1.25 2.97 Guatemala -0.85 -3.27* 1.69 1.73 13.88* Honduras -0.00 -0.01 1.15 0.45 0.25 Jamaica -0.13 -0.48 0.93 -0.12 1.95 Mexico 0.19 1.90 0.90 -0.36 4.29 Panama -0.08 -0.42 0.91 -0.22 0.74 Paraguay -0.44 -3.67t 1.29 1.53 13.89* Peru 0.21 - 1.62 0.80 -0.80 3.35 Uruguay -0.02 -0.40 0.35 -4.64* 36.75* Venezuela 0.05 0.63 0.47 -6.63* 50.97* a. Based on White (heteroskedastic-consistent) standard errors. * Indicates that the null hypothesis (rP = 0 for national cash flow and r,a = 1 for current account) is rejected at the 5 percent level. b. Tests the overall fit of the model. * Indicates rejection at the 5 percent level. Source: Authors' calculations based on World Bank and IMF data. Chosh and Ostry 327 Table 4. Variance Ratiofor the Predicted and Actual CurrentAccount Country Variance ratio, X-7 b Africa Botswana 1.19 0.03 Ethiopia 1.07 0.00 Ghana 0.19 35.54* Kenya 0.49 2.84 Liberia 0.07 297.60* Malawi 0.63 0.65 Mauritius 0.93 0.02 Morocco 2.49 0.60 Nigeria 0.75 0.09 Senegal 1.51 0.54 Tanzania 0.66 0.15 Tunisia 1.10 0.02 Zambia 0.45 13.99* Asia Hong Kong 2.33 0.70 India 4.24 1.50 Indonesia 0.48 2.39 Korea, Rep. of 2.05 1.26 Malaysia 0.37 5.77" Pakistan 1.59 0.51 Papua New Guinea 1.(9 0.09 Philippines 1.27 0.11 Singapore 0.52 2.77 Sri Lanka 0.17 145.76* Thailand 8.65 1.47 Middle East Egypt 0.20 7.35* Iran, Islamic Rep. of 0.93 0).01 Israel 0.38 2.18 Jordan 0.83 (.15 Saudi Arabia 1.01 0.00 Latin America and the Caribbean Argentina 0.24 16.41* Bolivia 1.78 0.25 Brazil 6.72 0.33 Chile 0.71 0.39 Colombia 0.92 0.03 Ecuador 0.94 0.02 El Salvador 0.60 2.82 Guatemala 2.72 1.73 Honduras 1.32 (.20 Jamaica 0.80 0.05 Mexico 0.78 0.18 Panama 0.84 1.0)4 Paraguay 1.39 0.98 Peru 0.73 0.40 Uruguay 0.12 88.51* Venezuela 0.24 117.22* a. Ratio of variance of optimal current account to variance of actual current account. b. x2 is the test statistic for the null hypothesis that the variance ratio is equal to unity. *Indicates rejected at the 5 percent level. Source: Authors' calculations based on World Bank and IMF data. 328 THF '()ORI1) BANK 1(ONOM IC REVIFW. VWl.. No 2 and the Caribbean region. It is also worth noting, as mentioned in the previous subsection, that for several countries, the point estimates of the variance ratio are below unity. That is, actual current account balances have been more vola- tile than optimal ones. This observation, in addition to the simple time-series plots presented previously (which do not depend on the power of statistical tests), should help to dispel the notion that endogenous government behavior- designed to smooth current account fluctuations in relation to the optimum- may be driving our results. 111. CONCLUSION A growing literature is suggesting that developing countries, far from being financially closed economies, may be better characterized as having a high de- gree of effective capital mobility. This article has presented an alternative meth- odology for gauging the extent of capital mobility in developing countries that avoids some of the difficulties associated with previous tests based either on examining the magnitude of gross capital flows or on establishing the extent to which financial market parity conditions hold. The basis of the tests in this article is the simple notion that, in a world of high capital mobility, agents should be able to fully smooth their consumption in the face of shocks. For the economy as a whole, this implies that the current account should act as a buffer to smooth aggregate consumption in the presence of shocks to national cash flow, defined as output (GDP) less investment less government expenditure. This article has shown that the consumption-smoothing model provides a natural benchmark against which to judge actual current account movements. If the level and volatility of such movements differ systematically from the move- ments predicted under the assumption of full consumption smoothing, then either there are effective barriers to capital mobility (less volatility of actual movements than predicted movements) or speculative factors drive capital movements (actual movements more volatile than optimal movements). Using data from a sample of forty-five developing countries, we empirically tested the consumption-smoothing model. In about two-thirds of our sample, the data were found to be consistent with the restrictions imposed by the model, a result that compares favorably with previous findings for the industrial coun- tries. For this sample of developing countries, therefore, the null hypothesis that agents have indeed been able to fully smooth consumption in the face of shocks could not be rejected. Less formally, we found that both the level and the volatility of current account movements predicted on the basis of the consumption-smoothing model are very close to the actual level and volatility of such movements observed in the data, again suggesting a relatively high degree of effective capital mobility in developing countries. In particular, we did not find any systematic tendency for actual current account movements to be smaller than optimal movements, as would be the case if there were generalized effective barriers to international capital movements. Gbosh and Ostry 329 These conclusions were supported for the vast majority of the countries in the sample, both by using formal statistical testing of the model's restrictions and through simple time-series plots of the predicted and actual data. In addition, the results obtained using the consumption-smoothing approach to assessing capital mobility in developing countries coincided in a number of cases with those re- ported in previous studies using very different analytical approaches. (Appendix tables start on the following page.) 3.30 THF. WORI I) BANK EC ONOMIC REVIEW. VOI. 9, No. 2 Table A-1. Sources and Sample Periods for National Accounts Countrv Source Sample period Africa Botswana World Bank 1960-88 Erhiopia IMF 1963-91 Ghana IMF 1955-90 Kenya IMF 1964-89 Liberia IMF 1965-88 Malawi IMF 1964-91 Mauritius IMF 1953-90 Morocco IMF 1957-91 Nigeria World Bank 1960-90 Senegal World Bank 1960-90 Tanzania IMF 1965-90 Tunisia IMF 1960-91 Zambia IMF 1957-87 Asia Hong Kong IMF 1961-91 India World Bank 1960-90 Indonesia World Bank 1960-90 Korea, Rep. of IMF 1953-91 Malaysia IMF 1970-90 Pakistan IMF 1960-91 Papua New Guinea IMF 1973-90 Philippines IMF 1948-91 Singapore IMF 1968-91 Sri Lanka IMF 1950-91 Thailand IMF 1950-90 Middle East Egypt World Bank 1974-90 Iran, Islamic Rep. of IMF 1964-90 Israel World Bank 1960-9(1 Jordan IMF 1969-91 Saudi Arabia IMF 1968-89 Latin Armerica and thei Caribbean Argentina World Bank 1960-90 Bolivia IMF 1950-91 Brazil World Bank 1960-90 Chile IMF 1955-91 Colombia World Bank 1960-90 Ecuador IMF 1965-90 El Salvador IMF 1951-91 GLuatemiiala IMF 1950-91 Honduras IMF 1950-91 Jamaicai IMF 1960-89 Mexico IMI' 1950-86 Panama IMF 1950-90 Paraguay IMF 1960-89 Peru World Bank 1960-90 Uruguay IMF 1955-91 Venezuela IMF 195()-91 Note: Annual data were collected for private c(oinsumption. investment, government consuTmption, GDP. and GNP in hillions of local currency. The ciw) deflator was used to convert noiminal into real magnitLides. Source: IMF (various issues) and World Bank (various years). Table A-2. VAR Coefficienzts and Associated Stanizdard Errors Stand7ard Standard Stanidard1 Standard Country Coefficient error Coefficient error Coefficient error Coetticient error Africa Botsxsana 0.51 0.20 - 0.12 0.14 (0.37 0.27 0.60 0.18 Ethiopia -0.04 0.20 -0.21 0.21 -0.09 0.12 0.80 0.13 Ghana 0.03 0.17 -0.19 0.22 -(.01 ().1 3 0.55 0.16 Keniya - 0.27 0.19 -0.47 0.23 -0.02 0.17 0.37 0.21 Liberia -0.09 0.23 -0.11 (1.22 -0.01 0.20 0.54 0.19 Malawi -0.05 0.20 -0.62 0.24 (.(9 0.17 0.28 0.2I Mauritiuis -(0.12 0.16 - 0.80 0.22 -0.16 (0.16 0.18 0.22 Morocco 0.06 0.18 -0.49 0.67 0.02 0.0.3 0.69 0.13 Nigeria 0.21 0.20 -0.19 0.13 0.15 0.23 (.77 0.IS Senegal -0.10 0.15 -1.06 0.28 0.16 0.10 0.24 (.18 Tanzania -0.04 0.19 -0.19 0.10 (1.22 0.23 0.82 0.12 Tunisia (.(01 0.18 -0.50 0.19 0.08 (.1.5 0.56 0.16 Zambia -0.0( 0.16 -0(.77 0.16 0.26 0.20 o .13 0.21 Asia HongKong 0.47 0.15 -0.27 0.12 (.(9) 0.17 0.71 0.13 India 1. 25 0.18 0.98 0.34 0.05 0.09 0.48 0.18 Indoncsia (0.2 i 0.2( 0.23 0.15 0.15 (1.21) 0.65 0.15 Korea,Rep.of 0.74 0.12 -(1.24 0.08 0.34 0.16 0.72 0.11 Malavsia 0.34 0.24 -0.23 0.13 0.32 0.35 0.58 0.20 Pakistan 0.54 0.18 0.10 0.l7 -0.00 0.18 0.40 (1.17 Papua New Guinea 0.31 0.24 -0.70 0.24 0.16 0.27 0.27 0.26 Philippines (.15 0.13 0.36 0.09 -0(.11 (1.18 0.59 0.12 Singapore 0.25 0.22 -0.26 0.18 0.02 0.20 0.56 0.17 SriLanika -0.35 (.15 -(0.1.3 0.20 0.04 0.12 (0.45 0.15 Thailarid 0.66 0.15 -0.30 (1.20 -0.14 0.15 0.54 0.20 (Table continues on the follou'ing page.) Table A-2. (conztinuiied) I I 1 12 '12 Il'22 .Standar, d Stacl 1nrd Standard Standard Counztry Coefficient eTroT Coe'tficient error Coeffzcient error Coefficieint error Middle East Egypt -0.22 0.19 -(.1.5 0.12 - 0.04 0.35 0.72 0.22 Iran, Islamiic Rep. of o.05 0.20 - 0.62 0.24 -0.04 0.18 0.33 0.22 Israel -(0.22 0.19 -0.18 0.17 -0.14 0.15 0.73 0.14 Jordan 0.49 ().19 -0.46 0.1.5 0.50 0.21 0.48 0.17 Sa diA,rabi.a 0.46 0.16 0.34 0.09 0.55 (.1 7 0.7.3 0.10 Latin A incrica and the Caribbean Argentin.a -0.01 0.18 0.22 0.12 -0.28 0.24 0.62 0.16 Bolivia -0.04 0.14 -0.37 0.12 -0.30 0.14 0.64 (0.12 t- Brazil 0.02 (0.21 - .30 01 0.06 (0.19 0.92 0.09 Clhile 0.21 (.18 0.39 0.16 (0.12 (.18 0.51 0.17 Colombia 0.51 (.19 -o.3(1 0.16 0.36 0.I 0.65 0.15 Ecuador 0. 14 0.20 -0.98 (0.31 (0.20 0.16 0.0l 0.24 El Salvador 0.29 (0.13 -0.52 0.13 0.30 0.16 0.28 0.15 Guatemala 0.40 0.13 - 0.87 (0.16 -0.06 0.14 0.06 0.17 Honduras 0.12 0.15 - 0.70 0.20( 0.10 0.12 0.41 0.16 Jamaica -0.11 0.20 -(.30 0.16 -0.28 0.21 0.65 0.17 Mexico -().10 0.14 -0.43 0.11 0.18 0.18 0.63 0.14 Panama -0.13 0.18 -0.29 0.09 -0.25 0.22 0.67 0.12 Paraguay 0.51 0.17 -0.74 0.25 ().22 0.10 0.18 0.15. Peru 0.25 (1.19 -0.35 0.22 0.11 0.15 0.48 0.17 UrugLIaV -0.01 0.18 -0.13 0(.21 -0.10 0.13 0.63 0.1.5 Venezuela 0.35 (.1.5 -0.48 0.10 0.82 0.23 0.03 0.15 Source: Authors' calculations based on World Bank and IMF data. Ghb slb and Ostrv 333 REFERENCES The word "processed" describes informally reproduced works that may not be com- moniv available through librarv systems. Campbell, Johi. 1987. "Does Saving Anticipate Declining Labor Income? An Alternative Test of the Permanent Income Hypothesis." Economletrica 55:1249-73. Campbell, John, and Robert Shiller. 1987. "Cointegration and Tests of Present Value Models." Journal of Politica/l Economy 95(October): 1062-88. Cooper, Richard N., and Jeffrey D. Sachs. 1985. "Borrowing Abroad: The Debtor's Perspective." In John T. Cuddington and GordonV Whitford Smith, eds., International Debt and tbe Developing Countries. Washington. D.C..: World Bank. Dooley, Michael, Jeffrey Fralikel, and Donald J. Mathieson. 1987. "International Capital Mobilitv: What Do Saving-Investmcint Correlations Tell Us?" IMF Staff Paper 34:503-30. Engel, Robert F., and Byunig Sam Yoo. 1987. "Forecastling and Testing in Cointegrated Systems." University of California Discussion Paper. University of California, Depart- ment of Economics, San Diego. Processed. Frenkel, Jacob A., and Assaf Razin. 1987. Fiscal l'olicies andi the World Economy: An Intertemporal Approach. Cambridge, Mass.: MIT Press. Chosh, Atish R. 1990. "Interinationlal Capital Mobility and Optimal Current AccouLit Behavior: An Empirical Investigationi." John M. Olin Discussion Paper 50. Princeton University, Woodrow Wilson School of PubliC and Interinational Affairs, Princeton, N.J. Processed. Forthcomiiing in EconomnicJournal. IMF (Interinational NMonetarv Fund). Various issues. International Financial Statistics. Mathieson, Donald, and Liliana Rojas-Suarez. t99'. "Liberalizing the Capital Account." INMF Working Paper WP/92/46. Processed. Also published in Finance and Developnment 29(December):41-43. Montiel, Peter J. 1994. "Capital Mobility in Developing Countries: Some Measurement Issues and Empirical Estimates." Thle WVorldt Bank Lcoomzic Rev'iee 8(3):311-50. Ostry, Jonathani D., anid Carmen Reinhart. 1992. "Private Saving and Terms of Trade Shocks: Evidencc from [Developing CouLntries." IMF Staff Papers 39 (3, September):495-5 17. Otto, Glenn. 1992. "Testing a Present-Value Model of the Current Account: Evidence from U.S. and Canadianl Time Series." Journal of International Money and Finance 11:414-30. Sachs, Jeffrey 1). 1982. "The Current Account in the Macroeconomic Adjustment Pro- cess. Scanndinavian Journal of Economfics 84(2):147-64. Sargent, Thomas J. 1979. Macroeconomic Theory. San Diego: Academic Press. Sheffrin, Steveni M., arid Wing Thye Woo. 1990. "Prcsent Value Tests of an Intertemporal Model of the (Currcnt Accouit.'' Journal ofl International Econoniics 29(November):237-53. World Bank. Various years. Wk'orld Tables. Baltimore: Johns Iliopkiis University Press. I H F W () K I I> B A \ K I t () X, () M I t R 1 k I I WX \ (, I q . S () 2 . . 3 * - 3 4 Comment on "Measuring the Independence of Central Banks and Its Effect on Policy Outcomes" by Cukierman, Webb, and Neyapti M. K. Anyadike-Danes Cukiermani, Webb, and Neyapti (1992) significantly broadened and deepened the empirical stuidy of the link between central bank independence and inflation per- formance. They conlsidered evidence for a group of countries much bigger than any examined previousl, most of the additions being developing countries, and they constructed a range of new indicators of indepenidetice based on central bank law, the results of a questionnaire, and the turniover rate of central bank gover- nors. In interpretilng their findings, however, they appeared to add a further, new dimension to this subject but without examininig the dimension's wider implica- tions. Almost as an afterthought, CLukierman, Webb, and Nevapti suggested that a counltry s choice of exclhanige rate regime might have some bearing on their results. Here I extend the analysis to show that a systematic examinationi of exchange rate-fixing arrangeinenits can play a role in both accoUtnting for the pattern of inflationl rates across their samnple of develop ing couLntries and sharpening the as- sociation between central bank independenice and inflation performance. L. CLKIERMAN, WEBB, AND NEYAPrTI ON I IE EXCHANGE RATE REGIME In the last paragraph before the concluding sectionl of their article, Cukierman, Webb, and Neyapti discuss the performance of their overall measure of central bank indepenidelce. This measure is a regression-based conistruct that combines information on legal indepenldence with the turniover rate of central bank gover- nors. The authors commilenit that "Austria, The Bahamias, Belgium, Luxembourg, Netherlands, and Pananma have lower inflationi in the 1980s than their central hank iiidependence would indicate, because their monietary policy is dominated by a policy rule fixing their exchanige rate to a relatively stable currelncy" (p. 382). In this statement the authors are apparently suggesting that the adoption of a fixed policy rule for the exchanige rate will alter the relation betweell central bank inde- Nl. K. Anvyadike-[)Dari is with the Depa.rtimienit of Economics vat the ULniversity of West Indies, Cave Hill CampLIs. Thte .utiior acknowledges the helpful COmInTITitS b Alex (&Likierman, Biln Nevapti, a.nd Steven R. We1l9. ; 1995S1 I'lieterna.tion(al Bankl Ifor Rec()i1%trL1Ctl( 1011an De\C1op1)F L-111t /TIIE WCR1R D BANKX ? 3 3.3 6 I IkW0(l II RANK [- ()NO)NI(RI- RIW, Vol.. 9. No. 2 pendence and inflation performance. I In fact, more than one-third of the develop- ing countries in their sample have fixed exchange rate rules. Explicit recognition of differences in exchange rate-fixing arrangenments increases the empirical sup- port for their central hypothesis that connects central hank independence and inflation. 2 11. TiiE EXCHANGE RATE REGIME AND INFLATION PERFORMANCE IN THE 1980s To bring the choice of exchanige rate regime into the picture, I classified the countries according to their exclhange rate-fixing arrangements by making use of the sumiimary table in IMF (various years). The IMF data identify three categories of exchange arrangements: a peg, limited flexibility, and more flexible arrangements. It turns out that only one country in the sample of developing countries is in the limited flexibility category-Qatar. Because Qatar's exchange arrangement is a peg, albeit witlh a slightly wider band, I treat it as a member of the pegged category. I then separate out the countries that have been in the same I MF category through- out the decade. That exercise yields two groups of countries: those with a pegged exchange arrangement and those with a flexible one. A third group, whose cate- gorization clhatnged durinig the decade, I refer to as having a combined exchalnge arrangement. These groups are showIl in table 1. Of the forty-seven developing countries in Cukierman, Webb, and Nevapti (1992, table 11), only forty-four are included here. I have omitted Tlaiwani (China) because it is not a member of the Interna- tional Monetary Ftlnd (IMF), and Hungarv and Poland becaulse they joined during the 1980s. Of the forty-four countries, sixteeni are In the pegged category, fifteen are in the flexible category, and thirteen are in the combined category. Within each category the countries are arranged in the table in order of increasinig average annual inflationi for the 1980s. It is immediately obvious that the classification by categorv has some connec- tion with inflation performance. This can be seen even more clearly by taking an initial cut at an inflation rate of 10 percent or less. Twelve of the sixteen pegged counitries had inflation rates of 10 percent or less. Onlv four of the fifteen flexible countries had inflation rates of 10 percent or less. The combined category seems to resemble the flexible category: four of the thirteen countries had inflation rates of 10 percent or less. One very significant outlier in the pegged category is Nicaragua. This observation actually provides an important insiglht into the IMF classifica- tion, which, clearly, refers to the exchange rate-fixinig arrangement, not the behavior of the exchange rate itself. For example, in the case of Nicaragua-a pegged exchange rate country-the exchange rate changed dramatically in the 1980s. t. Indeed, a similiir point is made hy CLikierimian (199' p. 4 7). 2. The authonm have brought to my attentiin that C(ukierm.l, Ro)driglez, and Webh IfortheOminiu have produced a paper that examitnj tes the IssLie for the indUistrial counttries, and t hev have also looked at the developing COuntr-IeS. Anvadike-Danes 337 Table 1. Inflation and Central Bank Independence, 1980-89 Auerage Turnover Index Transformed lExchange rate an,111al rate rate of i/ legal rate o/i categJry andi (i/flZ/7atomn central bank centralblank inflation countrl (percent) governor indepen11dence (pertenit) Pegged Malta 3 0.20) 0.44 0.02 Panama 3 0.20 0. 1) 0.0)2 Ethiopia 4 o. I o 0.40 0.03 Malaysia 4 0.2( 0.36 0.03 Qatar 4 0.00 (.2l 0.03 Romania 4 0.20) 0.30 0.03 Bahamas. The 6 0.2( 0.41 (.05 Barbados 0.10 0.38 o.05 Honduras 7 0)) 3 .43 0.05 Botswana 10 0.40 0.3 3 (.(9 Kenvy I 0.2() 20.44 0.09 Nepal 0 I ()0 I 8 0.08 Zimbabwe 12 0). 0 ( 0.2i) (. 11 Venezuela 19 (.5() 0.43 0.16 Tanzania 27 0.10 0.44 0.21 Nicaragua 128 0.40) 0.45 0.7 Flexible Morocco 7 0.2)) 0. 14 (0.06 Korea, Rep. of 8 (05() 0.27 0.07 India 9 0.3) 0.34 0.07 Indonesia 9 (0.20 0.27 (0(7 Philippinies I 3 0.20 0.43 0.1 South Africa 14 ((.2() ((.25 (.12 Nigeria 18 0. 1)) ((37 0.16 Colombia 21 0.21) n.27 0. 17 Costa Rica 23 0.41) o.47 ((.19 Turkey 41 0.4)) 0.46 0.28 Uruguay 45 (.30 ( 024 o34 Mexico 5( ((.30) 0.34 0.38 Yugoslavia 7.3 ().20 . 17 (l.5 I Brazil 119 0(.8( 0.2 1 0.68 Argentina 14i3 1 .00 o(4() .7 4 Combined Singapore 3 0.60 0.29 0A)2 Thailand 6 (1. I ) (.27 0.04 Pakistan 7 0.30 (0.21 0.06 China 8 (0.30) ((.29 0.((7 Western Samoa 12 l(.56 0.30) 0.11 Egypt 16 (.3() (0.49 ((.13 Chile 19 0.80 ).46 0.16 Zambia 25 (.5S() 03.33 0.20 GChania .37 0.2(0 0.31 0.28 Zaire 45 0.20 0(.43 0.34 Israel (2 0.2( 0.39 (.47 Uganda 72 0.2) 0.38 0.47 Peru l)08 0.(30) 0.4.3 0.64 Source: Author's calcLlations a.i(i Cukierniani, %ebb, and Nevapti (1992m tibles 2 anid 1). ]3 38 I HF H a (!RI D)bANK (:IN()NOIN Ri\'[FE' \f. N ) ' Ill. CUKIERMAN, WEBB, AND NEYAPTI'S RESULTS ON CENTRAL BANK INDEPENDENCE AND INFLATION PERFORMANCE Cukierman, Webb, and Nevapti construct a coinbiied measure of central bank independence derived from a regression of their index of legal central hank inde- pendence and the rate of turnover of central bank governors on a transformed value for the rate of inflation. Their values for the rate of turnover of central bank governor, central bank legal indepenidence index, and transforimied rate of infla- tion are given in table 1. A simple and effective way of summariziig these relations is to follow Cukier- man, Webb, anid Nevapti's procedure and regress the legal independence index and the turnover rate oin the transformed rate of inflationi. The resulting predicted value of the inflation rate can then be interpreted as a measuL-e of overall indepen- dence (see Cukierman, WVebb, and Nevapti 1992, p. 379 for a discussion). Unfortunately it is not possible to compare these results directly with those ob- tained by Cukiermani, Webb, and Neyapti, because their regressions were run over a much longer sample period. To provide a benchmark for comparisoni, I report resLilts for the regression of the independenlce nleastLres on inflation for the whole sample of forty-four coulitries. These are recorded in table 2. The coefficient on the legal independenice index does not have the predicted negative sign and is in- significanit, but that on the turnover variable has the predicted positive sigil and is significant. This pattern of findings does in fact tLiril out to resemble the most closely comparable estimates reported bIy CuLkieriliai, Webb, and Neyapti (see their table 8, p. 372). The results for separate regressionis for countries in rile pegged, flexible, and combibed categories are also reported in table 2. Neither of the coefficients for the pegged categorv is statistically well deternmined, and the coefficient on the legal independence index has ain inlappropriate sign. For the flexible category, hoth co- efficients are signed in accordance with the maintained hypothesis about the rela- tion between central hank iildependence aild inflation. The coefficient on the legal index is negative (althouIghi not well deternmined) aind that on the tuirnover rate is both positive and highly significanlt. For the combilled group, the legal indepen- dence index is significant and has a large coefficicit but with the wrong (positive) sign. The coefficient on the tLirilover variable is only just significant at the 10 per- cent level, but it, too, has the wroing (in this case negative) sign. To summarize, for the combiined category the results suggest a direct relation between inidependence aind iiiflation (thie more iindependence, the higher infla- tion), but for the pegged category the relation is negative but very weak and in the flexible category the relation is appropriately signed and rather more significanit. Thus, the pattern of evidence is broadly consistenit witi my interpretation of Cu- kierman, Webb , and Neyapti's hypothesis on central bank independence and ex- change rate policy rules. That is, some countries might not exhibit any systematic connection between measures of central hank indepeildenice and inflation be- cause, as Cukierman, Webb, and Nevapti put it, 'ttheir monietary policy is domi- Anvadike-Danes .339 Table 2. Estimation Reszilts for the Samnple ofLDe!eloping Counztries, 1980-89 C effilcu'nt Pegged Flexible Comblized Explanatory All 44 exchange rate egxchange rate excbange rate variabl)le coH17tries caltegory category category Colnst.ant 0.(1 0(.(9 0.1 (.10 ((.08) (. 59) (0.92) 0 ().47) Index of legal (.2( 0.28 (0.36 1.32 central haik (0.(A7) (0.te2) -0.85) (2 34) indepenldence Turinover rate of 0.41 (.52 0.69 -(0.39 the centraal hank (29 )7) (1.58) (3.97) ) 1.67) governor R2 (0.2)) ().25 0.57 0.43 Note: Estimation results ire from ordinlarv least st IL.ores. The depeildent variahle is the tra.nsforilmed rate of inflationi. The sample size is flrty-lour. t-r.itiois are in parenitheusi. Source: Author's alCulatiolls. nated by a policy rule fixing their exchange rate to a relatively stable currency" (p. 382). Table 3 presents the actual and fitted values for the transformed inflation rate, which are useful to investigate the properties of the combined measure of indepen- dence more deeply for the flexible category. The ordering of the independence measure, the fitted value from the table, does not match the actual inflation order- ing exactly. The two significant discrepancies are the Republic of Korea and Yugoslavia. Korea has a much lower inflation rate than expected, given its low legal independence rating and its relatively high turnover of governors, but the opposite is true of Yugoslavia. Nevertheless there is some degree of consistencv between the two series, and the rank correlation coefficient is 0.55, which is signif- icant at the 5 percent level. IV. CONCLUDING REMARKS Cukierman, Webb, and Neyapti's (1992) discussion of their evidence on the connection between inflation and central bank independence has exposed an im- portant issue: the role of the exchange rate regime. But they took the matter no further. It has been demonstrated here, using their data, that there were indeed significant differences in the 1980s in the inflation performance of developing countries with different types of exchange rate regimes. Moreover, those differ- ences were associated with systematic differences in the relation between inflation and Cukierman, Webb, and Nevapti's measures of central bank independence. Specifically, it has been slhowl (as Cukiermani, Webb, and Neyapti anticipated) that for counltries with pegged exchange rates (in Cukiernman, Webb, and Neyapti's terms, countries following an exchange rate rule), the connection between central bank independence and the countryvs inflation performance was much weaker than in countries where no such rule was in place. We also found 3401 1111 VWORI I DANK I( ONIINC RFVII , .VU N(1 ' Table 3. Actuczland Fitted Values of tbe TranisformTted Inflation Rate for Counitries with Flexilble Exchanzge Rates, 1980-89 Couintry A tual Fitted NAorocco O)06 0).2 Korea, Rep. ot t)07 (.38 India 0.07 0. __ InTdoTiesia 0)) (0.17 PIhilippilnes 0.1 I 011 Southi Africa 0.( 2 0. I8 Nigeria 0. 16 ((.07 (oloinhia (. 1 ( 1'7 Costa Rica (). 1 ( .24 Turkey '1.28 0. 24 LUrugLav V. 4 0)..5 Mexico (.3 8 (0). Yugoslavia .5 1 (.2 1 Brazil 10.68 0.61 Argentina (0.74 0.68 Source: ALthor'scalcularionssod (ukierman, %(chh-i and \NUIpat (1992, tahle I1). that for countries whose exchange rate reg,ime had changed over the decade, the relation between central bank independence and inflation was entirelv perverse. These findings raise several qLuestions that merit further investigationi. Amotng them are the interrelated issues of causality-does it run from the choice of ex- change rate regime to inflation, or the othier way around?-and credibilitv-is the simple choice of reginie sufficient? Future r esearch on the core question of central bank independence ma'v need to be more sharplv focused, because such indepen- dence seems likely to have very different mi.anifestationis and to critically depend on the nature of the exclhanige rate regime with which it is conmbined. REFERENCES Cukiermani, Alex. 1992. Central Banzk Strategy, Credibility, and Independence: Theory and Evidence. Cambridge. Mass.: MIT Press. Cukierman, Alex, Pedro Rodriguez, and Stevcn Webb. Forthcominig. "Central Bank Au- tonomy and Exchanige Rate Regimes: Their Effects on Monetary Accommodation and Activism." In Harry Huizinga and Svlvester Eijffinger, eds., Positive Political Economzly: Theor Yand Evidence. Tilburg, the Netherlands: C.enter for Fconomic Research. Cukierman, Alex, Steven B. WVebb, and Bilin Nevapti. 1992. "Measuring the Indepelidenice of Central Banks and Its Effect on Policy Outcomlies." The YYWorldl Bank Eco077mic Re- view} 6(3):353-98. INIF (International Monetar-y Fund). Various yea rs. Exchainzge Arrangements anid Ex- change Restrictions. Washington, D.C. World Bank New Publications Global Economic Prospects and the Developing Countries 1994 Globalization is transforming the world World an Projections 1994e95: economy, as reflected by the widening and intensifying international linkages in trade and Demographic Statistics finance. This new edition examines these timely Contains the most recent demographic and important issues along with GATT, the statistics on the world's population, with esti- outlook for private capital flows in light of the mates for 206 countries and territories projected Mexican crises, and the implications for as far as the year 2150. This edition incorporates industrial countries of the increasing global census and fertility-mortality survey results that economic role of developing countries, were unavailable for the previous edition. It including gains or losses from increased trade contains a new subregion, Central Asia, to and labor issues. 100 pages. Stock No. 12917. categorize the southernmost former Soviet $12.95. republics. 532 pages. Stock No. 44947. $34.95. World Debt Tables 1994-95: External Finance for Developing Countries Social Indicators of Development 1995 A complete and up-to-date presentation of Poverty alleviation, the main issue facing the data and a review of development on the World Bank and its member countries, is the external debt of and financial flows to develop- primary focus of this latest survey of social ing countries. Volume I-Analysis and Sum- mary Tables. 248 pages. Stock No. 12912. progress around the globe. The format highlights $19.Volume I1-Country Tables. 560 pages. poverty statistics for each of 190 countries. $21.95. Volume II-Country Tables. 560 pages. Indicators are given for GNP per capita, urban Not available separately. Two-volume set- and rural wages for unskilled workers, the Stock No. 12914. $150.00. adrrlwgsfrusildwres h consumer price index, and a range of other social Also available: checkpoints. 412 pages. Stock No. 45021. World Debt Tables 1994-95: Data on Dis- $26.95. kette. Stock No. 12915. $95.00. Also available: Social Indicators of Development 1995: Data World Bank Atlas 1995 on Diskette. Stock No. 12573. $45.00 An easy-to-use collection of color maps, charts, and graphs representing key social, economic, and environmental indicators for 209 countries and territories. Organized under three development themes: The People, The Stock No. 12922. $7.95. World Bank Publications Order Coupon CUSTOMERS IN THE UNITED STATES: CUSTOMERS OUTSIDE THE UNITED STATES: Complete this coupon and return to Contact your local World Bank Publications distributor The World Bank for information on prices in local currency and payment Box 7247-8619 terms. (See next page for a complete list of distributors.) Philadelphia, PA 19170-8619 If no distributor is listed for your country, use this order U.S.A. form and return it to the U.S. address. Orders that are Charge by credit card by calling (202) 473-1155 or sent to the U.S. address from countries with distribu- fax this completed order coupon to (202) 676-0581. tors will be returned to the customer. Quantity Title Stock Number Price Total Subtotal US$ * If a purchase order is used, actual postage will Postage and handling* US$ be charged. If payment is by check or credit card, postage and handling charges are Total US$ US$5.00 per order. For air mail delivery outside the U.S., include US$8.00 for the first item and US$6.00 for each additional item. CHECK METHOD OF PAYM1ENT O Enclosed is my check payable to The World Bank. O Charge my O VISA O MasterCard LI American Express Credit card account number Expiration date Signature O Bill me. (Institutional customers only. Purchase order must be included.) PLEASE PRINT CLEARLY Name Firm Address City State Postal code Country Telephone Distributors of World Bank Publications Prices and terms vary by country ARGENTINA COLOMBIA GREECE Carlos Hirsch, SRL Infoenlace l.tda. Papasotiriou S.A. Galeria Guemes Apartado Aereo 34270 35, Stournara Str. Florida 165, 4th Floor-Ofc. 453/ Bogota D.E. 106 82 Athens 465 1333 Buenos Aires COTE D'IVOIRE HONG KONG, MACAO Centre d'Edition et die Asia 2000 Ltd. Oficina del Libro Internat ional Diffusion 46-48 Wvndham Street Alberti 40 Africaines (CEDA) WiLning Centre 1082 Buenos Aires 04 B.P. 541 7th Floor Abidjan 04 Plateau Central Hong Kong AUSTRALIA, PAPUA NEW GUINEA, CYPRUS HUNGARY FIJI, SOLOMON ISLANDS, Center of Applied Resear h Foundation for Market VANUATU, AND Cyprus College Economy WESTERN SAMOA 6, Diogenes Street, Engomi Dombovari Ut 17-19 D.A. Information Services '.O. Box 2006 H-1117 Budapest 648 Whitehorse Road Nicosia Mitchani 3132 INDIA Victoria CZECH REPUBLIC Allied Publishers Private Ltd. National Information Center 751 Mount Road AUSTRIA PO Box 668 Madras- 600002 Gerold and Co. CS-1 13 57 Prague I Graben 31 INDONESIA A-101l Wien DENMARK Pt. Indira Limited SamfundsLitteratur Jalan Borobudur 20 BANGLADESH Rosenoems All11 I P.O. Box 1181 Micro Industries Development DK-1970 Frederiksberg C Jakarta 10320 Assistance Society (MIDAS) House 5, Road 16 DOMINICAN REPUBLIC IRAN Dhanmondi R/Area Editora Taller, C. por A. Kowkab Publishers Dhaka 1209 Restauracidn e Isabel la P.O. Box 19575-511 Cat6lica 309 Tehran BELGIUM Apartado de Correos 2190 Z-1 Jean De Lannoy Santo Domingo IRELAND Av. du Roi 202 Government Supplies Agency 1060 Brussels EGYPT, ARAB REPUBLIC OF 4-5 Harcourt Road Al Ahram Dublin 2 BRAZIL Al Galaa Street Publicacoes Tecnical Cairo ISRAEL Internacionais Ltda. Yozmot Literature Ltd. Rua Pexoto Gomide, 209 The Middle East Observer P.O. Box 56055 01409 Sao Paulo, SP 41, Sherif Street Tel Aviv 61560 Cairo CANADA R.O.Y. International Le Diffuseur FINLAND P.O. Box 56055 151 A Boul. de Mortagne Akateeminen Kirjakauppa Tel Aviv 61560 Boucherville, Qu6bec P.O. Box 128 J4B 5E6 SF-00101 Helsinki 10 ITALY Licosa Conimissionaria Renouf Publishing Co. FRANCE Sansoni SPA 1294 Algoma Road World Bank Publications Via Duca Di Calabria, 1/1 Ottawa, Ontario KIB 3W8 66, avenue d'lena Casella Postale 552 75116 Paris 50125 Firenze CHINA China Financial & Economic GERMANY JAMAICA Publishing House UNO-Verlag Ian Randle Publishers Ltd. 8, Da Fo Si Dong Jie Poppelsdorfer Allee 55 206 Old Hope Road Beijing 53115 Bonn Kingston 6 JAPAN PERU SRI LANKA & THE MALDIVES Eastern Book Service Editorial Desarrollo SA Lake House Bookshop Hongo 3-Chome, Bunkyo-ku 113 Apartado 3824 P.O. Box 244 Tokyo Lima 1 100, Sir Chittampalam A. Gardiner Mawatha KENYA PHILIPPINES Colombo 2 Africa Book Service (E.A.) Ltd. Intemational Book Center Quaran House, Mfangano St. Suite 1703, Cityland 10 SWEDEN P.O. Box 45245 Condominium Tower I FritzesCustomer Service Nairobi Ayala Avenue, H.V. dela S-106 47 Stockholm Costa Extension KOREA, REPUBLIC OF Makati, Metro Manila SWITZERLAND Pan Korea Book Corporation Librairie Payot P.O. Box 101, Kwangwhamun POLAND Case postale 3212 Seoul Intemational Publishing Service CH 1002 Lausanne Ul. Piekna 31/37 Korean Stock Book Centre 00-677 Warszawa Van Dierman Editions Techniques P.O. Box 34 P.O. Box 465 Yeoeida PORTUGAL CH 1211 Geneva 16 Seoul Livraria Portugal Rua Do Carmo 70-74 TANZANIA MALAYSIA 1200 Lisbon Oxford University Press Universitv of Malaya Cooperative Maktaba Street Bookshop, Limited SAUDI ARABIA, QATAR P.O. Box 5299 P.O. Box 1127, Jalan PantaiBaru Jarir Book Store Dar es Salaam 59700 Kuala Lumpur P.O. Box 3196 Riyadh 11471 THAILAND MEXICO Central Department Store INFOTEC SINGAPORE, TAIWAN, 306 Silom Road Apartado Postal 22-860 MYANMAR,BRUNEI Bangkok 14060 Tlalpan, Mexico D.F. Gower Asia Pacific Pte Ltd. Golden Wheel Building TRINIDAD & TOBAGO, NETHERLANDS 41, Kallang Pudding, #04-03 Systematics Studies Unit De Lindeboom/InOr-Publikaties Singapore 1334 #9 Watts Street P.O. Box 202 Curepe 7480 AE Haaksbergen SLOVAK REPUBLIC Trinidad, West Indies Slovart G.T.G. Ltd. NEW ZEALAND Krupinska 4 UGANDA EBSCO NZ Ltd. P.O. Box 152 Gustro Ltd. Private Mail Bag 99914 852 99 Bratislava 5 1 st Floor, Room 4, Geogiadis New Market Chambers Auckland SOUTH AFRICA, BOTSWANA P.O. Box 9997 For single titles: Plot (69) Kampala Road Oxford University Press Kampala NIGERIA Southem Africa University Press Limited P.O. Box 1141 UNITED KINGDOM Three Crowns Building Jericho Cape Town 8000 Microinfo Ltd. Private Mail Bag 5095 P.O. Box 3 Ibadan For suibscription orders: Alton, Hampshire GU34 2PG Intemational Subscription Service England NORWAY P.O. Box 41095 Narvesen Information Center Craighall ZAMBIA Book Department Johannesburg 2024 University Bookshop P.O. Box 6125 Etterstad Great East Road Campus N-0602 Oslo 6 SPAIN P.O. Box 32379 Mundi-Prensa Libros, S.A. Lusaka PAKISTAN Castello 37 Mirza Book Agency 28001 Madrid ZIMBABWE 65, Shahrah-e-Quaid-e-Azam Longman Zimbabwe (Pte.) Ltd. P.O. Box No. 729 Libreria Intemacional AEDOS P.O. Box ST 125 Lahore54000 Consell de Cent, 391 Southerton 08009 Barcelona Harare The World Bank Research Observer Concise and readable, the World Bank Research Observer examnines key issues in development economics, written in a style accessible to non- specialist readers. Articles cover the broad range of development issues, survey the latest World Bank research,and debate current development policy. Samples of recent articles: AIDS and African Development (July 1994) The Costs and Benefits of Soil Conservation (July 1994) Privatization: Lessons from Market Economies (July 1994) Parallel Exchange Rates in Developing Countries (February 1995) Incentives and the Resolution of Bank Distress (February 1995) Readership is made up of a wide audience of academics, researchers and others with a professional interest in development. The World Bank Research Observer is published twice a year. Subscription rates: 1-year: individuals US$20, institutions US$35 2-year: individuals US$36, institutions US$66 3-year: individuals US$40, institutions US$95 Individuals from developing countries and socialist economies in transition are eligible to receive complimentary subscriptions on request. 8' 9:) 4 9 _ | ~~~~~~~~~~~~~~~~~~~EC3N RE'V !E.. V'C . 2 :111llfilllil11!11;~~~~~~~~~~~~~0 9