ThE WORLD BANK ECONOMIC REVIEWI Vol C ii mCprcnrl br 1 99 i Nu111 1-3) T,he Soviet Economnic Decline 'William Easterly and Stanley Fischer - Contingent Valuation and Actual Behavior: Predicting Connections to New Water Systems in' the State of 'Kerala: ndia - Charles C. Griffin, John Briscoe, -Bhanwar Singh, Raidhika Ramasubban, and Ramesh Bhatia Political Influence on the Central Bank: International Evidence- Alex Cukierman and Steven B. Webb Pioneers for Profit: St. Petersburg Entrepreneurs in Services Martha de Melo, Gur Ofer, and Olga SandIer Apprenticeship Contracts, Small Enterprises, and Credit Markets in Ghana Ann D. Velenchik Inequality and Growth Reconsidered: Lessons from East Asia Nancy Birdsall, David Ross, and Richard Sabot Another Look at the East Asian'Miracle - Gustav Ranis Cumulative Indexes of Authors and Titles for Volume 9 THE WORLD BANK ECONOMIC REVIEW ID1)1 1 01o Mloshe Strqitlll ( 0 '1 1 1 I 1\(, 1 ) 10 Sandr a (G i I 1)1 ItRI I l 110 \RI) KaUshik Basu. (Cornell lniversirt and University oi Delhi (,regorv K. Ingraim Guillernmo (alvI, University of Nlaryland John Page a t1ahan Eaton, Biostion Universirt [ant H. Pritcherr Alberto (o ivalinint, Columbia University Jacques Van der (idag Mark R. Rosenz%veig. University of Pennsylvania Tbe World Bank Economici.ii RevieLt is a proifessio iil joairniaI for the dissemina tion of World Bank- spoiisored research That iniforms policy anialyses and choices. Ir is directed to an international readership amon0lg ecoiionoists and social scientists in government, business, and international agencies, as xvell as in universities and development research institutions. The Reviewc emphasizes policy relevance and opera- rional aspects if economics, rather thati primarily theorerical and merhodiological issues. It is intended for readers familiar wvith eciiniomic theorv and analsis but nor necessarilv proficient in advaniced mathemati- cal or econometric rechniques. Articles will illustrate how professional research can shed light on policy chhoices. Inconsistency with Bank policv will not be grouiids for rejection of an article. Articles will he drawn primarily from work conducted by World Batik staff and consultants. Before beiiig accepted for publication bv the Editorial Board, all articIls are reviexved by twvo reterees who are not mem- bers of rhe Bankk's staffi articles must also be recommeilded by at least onie external member of rhe Edirtirial Boa rd. The R(vici mnay on occasioti publish articles on specified topics by non-Bank conltributors. Anv reader interested in preparing such an arricle is InVited to sulbmit a proposatl of nior more than two pages in leligth to the Editor. The views and interpretrations expressed in this journal are those of the authors and do not iiecessarily represent the viexws and policies of the tVorld Bank or Of its Executire Directors or rhe countries they rep- resenr. The WXorld Bank does not guarantee the accur.cv of the dara iincluded in this publicatrion atid accepts 11o responstibilirv wharsoever tor an) consequences otf their use. When maps are used, the bound- aries, deito min at iotns, and other informationt do irot ini pl% vi the part of the \World Bank GroLip any i dg- inetilr On tihe legal status of any territory or the endclrsen,enit tr acceptance of isch hltuitidaries. C olnttienrs or brtef notes respoiniding tio Ret 'ite articles art rvelcomne t nd w% ill be ciosiLde red for puIlicaL- tiottitrht the extenlt rhat spice permits. Please direct all editorial torrespo ndence to thle Edritor Tlbe Wierld Baik, Fet iotmie Ret vwi. The World Btink Washinigroii 1D.( . 20433 LISA. TR kk il?l B.ink CLin,itt titte Ret ten is pit hINIshed tlhre riittes i Netr tJitLinary, Iay, and Septem betr) b l the World Kitnk. Siitgle copies nist be purclhased at $1') 9S Sub4s.ription rate are ,ts follows: Indvidi l11s1 bwIsSltttlswts I -vear subScriprion LS$25 USi4S 2-!y e.r tibscriptitn tUS$4h USS86 )-y eir tsubscription US$6 S( U.LSS I '15 Orders shOLild be setir to: Witrld Batik Publictriots, Box -24---9S6 Philadelphia.. PA 191-0-7956 U.S.A. SubsCrpriontI, re availaktble mtI1ht. charge To re.tdtrs wit m.tiling addresses in developinig utlunrries and in st cialist eLTotnties ilt rransirion. Writteni reqILICte is req iired eveer r vio ears to renew such sLihscripriotis. c 1995 Thbe InternatiotnalI Batnk for Reconistruction and Developmentr /t t IIF \Xti 1) u. N. K All rights reserved Mantuf.ctured in the United SraTes of Americ. IPsN 0)258-6 t) Matrerial in This journ.tl is copvrighted. The Witorld Bank enCO ur.tges dissenrixiiion of irs work and will noritrall give periiiissioti proiilptI .tnd, when rlte i itIded rep roduction is for noncoilnierciala pitrposes withoutr .tking . ftee. Periissioii rTO ittike phitiicopiei ie grallted throitigh the Co pright Clearanct ciretr 2- (tongress Street. Salei, NIA 19-l ! U.S.A. Thiis journia.l is indexed teglrl.ls ii (t trriit (CoiittntZs/So, iml C- Be/.avi'tr.l Si, ittnoc.. Index to I,tteritm- it,ul.S.7.,itmsti1S. /it 11r11 I/ bet (I'l mit Litere,tiiri Pitli/ic 4Aff tire Iitsir itratiion t.rt ice, andl Soi ieal SiitctLs ( itittot Index .r It is tv,il.ille in mcru(ct irTii through Ilni%tirsir% Nli icroiltiht. Ilic., (1) Nitrrth Zeeb Road. Ainth Arbor. \ lmclim-,in 4S 1 0,L US .A. THE WORLD BANK ECONOMIC REVIEW Volume 9 September 1995 Number 3 The Soviet Economic Decline 341 William Easterly and Stanley Fischer Contingent Valuation and Actual Behavior: 373 Predicting Connections to New Water Systems in the State of Kerala, India Charles C. Griffin, John Briscoe, Bhanwar Singh, Radhika Ramasubbani, and Ramesh Bhatia Political Influence on the Central Bank: 397 International Evidence Alex Cukiernman and Steven B. Webb Pioneers for Profit: St. Petersburg Entrepreneurs in Services 425 Martha de Melo, Gur Ofer, and Olga Sandler Apprenticeship Contracts, Small Enterprises, 451 and Credit Markets in Ghana Ann D. Velenchik Inequality and Growth Reconsidered: Lessons from East Asia 477 Nancy Birdsall, David Ross, and Richard Sabot Another Look at the East Asian Miracle 509 Gustaiv Ranis Cumulative Indexes of Authors and Titles for Volume 9 535 THE WORLD BANK Et.ONOMIC REVIEW. VOl 9, NO. 3: 341-371 The Soviet Economic Decline William Easterly and Stanley Fischer Soviet growth from 1960 to 1989 was the worst in the uworld after we control for investment and human capital; the relative performance worsens over time. There is some evidence that the burden of defense spending modestly contributed to the So- viet debacle. The declining Soviet growth rate from 1950 to 1987 can be accounted for by a declining marginal product of capital with a constant rate of growth of total factor productivity. The Soviet reliance on extensive growth (rising capital-to-out- put ratios) was no greater than that of market economies, such as Japan and the Republic of Korea, but a low elasticity of substitution between capital and labor implied especially acute diminishing returns to capital compared with the case in market economies. Why did the per capita economic growth of the former U.S.S.R. decline and then stop, contributing to the final collapse of the Soviet economic and political system? Accounts of the declining Soviet economic growth emphasize different causes: the Soviet reliance on extensive growth, which, given the slow growth of the labor force and the falling marginal productivity of capital, eventually ran out of payoff; the declining rate of productivity growth or technical progress associated with the difficulties of adopting and adapting to the sophisticated technologies being introduced in market economies; the defense burden; and a variety of special factors relating to the absence of appropriate incentives in the Soviet system, including corruption and demoralization (Banerjee and Spagat 1991; Bergson 1987b; Desai 1987; Ofer 1987; and Weitzman 1970). In this article, we examine alternative explanations with special care to place the Soviet growth performance in an international context. In section I, we start with an overview of the data and of the Soviet growth record and compare it to that of other countries by using a standard growth regression. We also examine the role of the defense burden in the performance of the Soviet economy and then compare the Soviet pattern of extensive growth-rising capital-to-output ratios-to that of other William Easterly is with the Policy Research Department at the World Bank. and Stanley Fischer is the First Deputy Managing Director at the International Monetary Fund. The authors are grateful to Karen Brooks. Padma Desai, Alan Gelb, Eugene Gravilenkov, Barry Ickes. Barry Kostinsky, Martha de Melo, Gur Ofer, Sergio Rebelo, Bryan Roberts, Marc Rubin, Randi Ryterman, Martin Schrenk, Marcelo Selowsky, Michael Walton, Alwyn Young, and anonymous referees for comments; to seminar participants at the Massachusetts Institute of Technology, the University of Michigan, the Federal Reserve Board, the University of California at Los Angeles, and the World Bank; to Professor Mark Schaffer of the London School of Economics for kindly making his historical Western, official, and Khanin's data available in machine-readable form; and to Elana Gold and Mary Hallward for diligent research assistance. C) 1995 The International Bank for Reconstruction and Development /THE WORLD BANK 341 342 THF WORDI) BANK F:ONOMI( REVIFW. VOL. 9, NO 3 economies. In section 11, we assess the role of extensive growth in the decline of growth by reexamining and updating estimates of the aggregate production func- tion and in particular the elasticity of substitution between labor and capital. We then discuss how Soviet elasticities of substitution compare with those in other econo- mies. In the conclusion, we offer an interpretation of our results. 1. THE SOVIET GROWTH RECORD The fundamental problem in evaluating Soviet growth is the quality of data (see Fischer 1992). As a result of both methodological problems-particularly deflating nominal data-and incentives to misreport output within the Soviet system, official data on Soviet output overstate growth. Western analysis of Soviet growth relies on classic studies, including Bergson (1961) and the U.S. Central Intelligence Agency (CIA 1982, various years). The CIA makes the work- ing assumption that physical quantities as presented in the official data were not systematically misreported. Thus the difference between the Western estimate that per capita Soviet gross national product (GNP) increased between 1928 and 1987 by 3.0 percent a year (4.3 percent a year for aggregate GNP) and the official (Soviet) estimate that net material product (NMP) per capita increased by 6 per- cent a year results mostly from pricing corrections and also from differences in the coverage of NMP and GNP.' The classic Western estimates generally assume that data on Soviet investment and capital are more accurate than data on out- put. This view is supported by Bergson (1987a) and disputed by Wiles (1982). The Western data through 1985 are conveniently summarized in Ofer (1987). The data set we use in this article is constructed from Western data on output, industrial production, employment, and the capital stock for the U.S.S.R. as a whole. It includes the series on value added and capital stocks in industry by Powell (1963), CIA (1982), and CIA (various years) and the series on GNP, labor input, and capital stock for the entire economy from Moorsteen and Powell (1966), Powell (1968), CIA (1982), CIA (various years), and Kellogg (1989). All these series use 1937 rubles for the period 1 928-60, 1 970 rubles for the period 1960- 80, and 1982 rubles for the 1980s. The direct source of our data sets is Gomulka and Schaffer (1991), who spliced together series from the sources described. We look at how our preferred data set differs from both the official U.S.S.R.-wide data on real output, industrial production, employment, and the capital stock in the material sector in 1973 rubles, and from Khanin's (1988) data, which are also at U.S.S.R.-wide level, for output, employment, and the capital stock in the material sector.2 (Note that both the official and 1. This section draws on Fischer ( 1992). The Soviet concept of NMP omitted from GNP those services not directly related to prodiuction, such as passenger transportation, housing, and the output of government employees not produLCilng material output. 2. Ericson (1990) argues that the Khanin data are preferahle to the Western series, hut Bergson (1987a, 1991a) criticizes the Khanin data for a poorly documented methodology and the use of unweighted averages of physical indicators. Easterlv and Fischer 343 Table 1. Growth Rates in the Soviet Economy, Based on Different Data Sources, 1928-87 (average annual percent) Western data f_ ficial data Khanin's data Total Material Material Indicator and period economy Industry sectors Industry sectors Output per u'orker 1928-39 2.9 5.0 1 1.4 12.5 0.9 1940-49 1.9 1.5 2.1 0.1 1.0 1950-59 5.8 6.2 8.3 8.9 5.3 1960-69 3.0 2.8 5.4 5.7 2.7 1970-79 ).1 3.4 4.1 5.2 1.2 1980-87 1.4 1.5 3.0) 3.4 0.2 Capital per w.-orker 1928-39 5.7 6.5 8.7 11.9 5.9 1940-49 1.5 0.1 2.7 1.5 1.3 1950-59 7.4 3.9 7.7 8.0 3.5 1960-69 5.4 3..4 7.1 6.1 3.8 1970-79 5.0 4.1 6.8 6.3 1.9 1980-87 4.0 4.0 i.3 5.6 0.1 So,urce: Gomnulka and Schaffer (1991). Khanin's data are presented for the material sectors only and do not include consumer services). Soviet Growth in International Comparison Growth rates of the Western series for different periods are presented in table 1, with the official data and Khanin's data given for comparison. The Western growth rates for output per worker and capital per worker are well below the official rates, and Khanin's data are in turn below the Western data. All series show growth declining sharply starting from the 1 950s. Although we have greater confidence in the Western data than in the likely exaggerated official estimates and the inadequately documented methodology of Khanin, the disparities sug- gest some caution in the use of all data sets on aggregate Soviet growth. How does the Soviet growth record compare with the rest of the world's? We use the Western gross domestic product (GDP) series to compare Soviet per capita growth from 1960 to 1989 with per capita growth rates for 102 economies. (We look here at per capita rather than per-worker growth to enlarge the sample of comparators and make it consistent with the cross-sectional growth literature.) The next-to-last column of table 2 shows that Soviet per capita growth was slightly above the global average in the period 1960-89. Soviet growth no longer looks respectable, however, once we control for the standard growth determinants from the empirical literature. The last column of table 2 shows the residual from inserting the U.S.S.R. into the core regression of Levine and Renelt (1992), which relates growth to initial income, population Table 2. The U.S.S.R. in the Levine-Renelt Growth Regression, 1960-89 Actual per Population 1'redicted capita grouth, Per capita growth, 1960-89 Secondaryschool Investment growth 1960-89 income., 1960, (average annual enroliment, 1960 ratio to (average annual (average annual Growth Variable (1985 U.S. dollars) percent) (percent) (,DP, 1960-89' percent" percent) residual Average for sample excluding the U.S.S.R. 1,792 2.07 21 21 2.00 2.00 0.00 Average for the U.S.S.R. 2,796 1.05 58 29 4.70 2.36 -2.34 Predicted effect on the $' growth differential hetween the world average and the U.S.S.R. 0.35 0.39 1.17 1.49 2.70 0.36 -2.34 Note: The sample excluding the U.S.S.R. has 102 coLutries. a. Per capita income in 1960 for the L.S.S.R. is calculated based on the Bergson (1 991h) purchasing power paritv value for 1985 and hackcastinig to 1960 using the actual per capita growth given in the sixth column. b. The CIA series for investmenit is available at five-year intervals for 1960-75; we iiterpolated to obtain values for the other years. Source: T he regression is from Levine and Renelt (1992). For all countiries except rhe U.S.S.R Levine and Renelt (1992); for the U.S.S.R.: actual per capita growth is from Goftilka and Schaffer (1991) and Marer afd others (1992); per capira income is from Bergsori (1991 bh1; population growth is from Feschbach (1983), IMF and others (1991), and N9arer and others (1 992); secondary school enrollment is from uNEsCo (197S) and Mater and others (1992); investment rate is from CIA (various years) and Marer and others (1992). Easterly and Fischer 345 growth, secondary school enrollment, and the ratio of investment to GDP. The Levine-Renelt regression including the U.S.S.R. is as follows: (1) Per capita growth 1960-89 = -0.83 + 17.49 investment 1960-89-0.35 C;DP per capita 1960 (0.85) (2.68) (0.14) + 3.16 secondary enrollment 1960 - 0.38 population growth 1960-89 - 2.34 dummy for U.S.S.R. (1.29) ((A.22) (1.43) 103 observations R2 = 0.46 standard errors in parentheses Except for population growth, Levine and Renelt showed these variables to be robust to alternative specifications in growth regressions (although concerns about endogeneity remain). The regression results are identical to the Levine- Renelt original, which excludes the U.S.S.R., because we are dummying out the Soviet observation. Excepting initial income, the values of the Soviet right-hand side variables should have implied very rapid growth.3 As shown in the third row of table 2, high Soviet human and physical capital accumulation should have increased Soviet growth by 2.7 percentage points above the world average. As it was, growth was only barely above average, hence the large negative residual of 2.3 percentage points for 1960-89. Soviet per capita income in 1989 was only half of what it would have been if the average relationship between growth and the right-hand-side variables had held for the period 1960-89. The Soviet residual in this ordinary least squares regression is not actually significant in a two-sided test at the 5 percent level. However, it is notable that the only economies with worse residuals are generally both small and poor: Guinea-Bissau, Jamaica, Liberia, Peru, Suriname, and Zambia. The presence of so many small and poor economies among the large outliers makes us suspect heteroskedasticity. The suspicion is justified. We split the 1960-89 sample into thirds on the basis of total real GDP (that is, population times per capita income valued at purchasing power parity) and reran the above regression for the top and bottom thirds ranked by total GDP. (The U.S.S.R. is included in the top third ranked by total GDP, and we continued to dummy it out.) The Goldfeld-Quandt test statistic for heteroskedasticity indicates that we can reject homoskedasticity.4 The test results are as follows: Sum of squared residuals in third of sample with lowest real GDP: 88.3 Sum of squared residuals in third of sample with highest real c.Op: 33.9 F (29, 28) = 2.61 (significant at I percent level) 3. Some would argue that even initial income was more favorable (that is, lower) for subsequent growth than we have depicted it here, because some argue that the Bergson (1991b) estimate of Soviet per capita income that we have used is too high. This would make the Soviet conditional growtl performance look even worse. 4. The Goldfeld-Quandt test statistic is equal to the ratio of the sum of squared residuals in these two subsample regressions and is distributed as an F-statistic with the number of degrees of freedom of the numerator and denominator corresponding to the degrees of freedom in the subsample regressions. 346 THE WV)RI D) BANK F[LONMI(C RFVIEW, Vl1. 9. NI. On the basis of test results, we performed weighted least squares using the log of total real GDI' as the weighting series. The results are as follows: (2) Per capita growth 1960-89 =-0.43 + 15.93 investmenit 1960-89 - 0.28 GDP per capita 1960 (0.73) (2.19 (0.(8) + 2.56 secondary e nrollimenit 1960 - 0.24 population growth 1960-89 - 2.28 dummy for U.S.S.R. (073! (0.16) (0.48) 102 observations R2 (weighted) = 0.84 standard errors in parentheses The Soviet dummy becomes highly significant with weighted least squares, with a t-statistic of 4.8. Taking into account that the only economies doing worse than the U.S.S.R. were small makes the Soviet performance look even worse. After correcting for heteroskedasticity, the Soviet economic performance conditional on investment and human capital accumulation was the worst in the world from 1960 to 1989. How does the comparative Soviet performance evolve over time? Because the World Bank data used by Levine and Renelt begin only in 1960, we compared the Soviet performance also with the cross-country Summers-Heston (1991) data set that extends back to 1950. We performed a pooled time-series, cross- sectional regression using decade averages for the same specification as before (except that unfortunately we had to omit the secondary school enrollment vari- able for lack of reliable Soviet data for the 1950s). We used the same Soviet data as in the previous regression, but now broken down by decade. For each decade, we used intercept dummies as well as a separate Soviet dummy. We continued to use weighted least squares with the weighting series being the log of total GDP, because the Goldfeld-Quandt statistic still indicated a significantly larger vari- ance for small economies.) The results are as follows: (3) Per capita growth by decade = 0.022 + 0.120 investmentitD,r by decade (0.005) (0.016i - I .5E-06 (,DP per capita, initial year ot each decade - 0.626 population growth by decade (3.6E-0 7 (0.143) + 0. 005 1960s dnimnv - 0.( )05 19-70s dummy - 0. 01 s 1 980s dummy (0.004) (0.003) ! 0. 00) +0. 024 duLImIy for l S. S. R. 1 950s - 0 0(8 dummI for U. S. S. R. 1 960s (0( 11 (().0°ll) 1 7.)17 dummy for I.S.S.R. 1 97(s-0.023 diummy for I.S.S. R. 1 980s (0.(09) 1(.009) 391 observations R2 (weighted) = 0.54 standard errors in parentheses Although economic growth decelerated worldwide in the 1970s and even more in the 1980s, growth deceleration in the U.S.S.R. is notable even by comparison 5. The F-statistic for the ratio of the sum of squiared residuals in the bottoni third to that in the top third of the sample ranked by total GDP (in purchasing power parity prices from Summers and Heston 1991) is F (124. 121) 2.03. which is significant at the I percent level. Easterly and Fischer 347 with the world pattern: Soviet economic growth was significantly above the world average in the 1950s and significantly below even the poor world growth of the 1980s. Note, especially, the good performance of the U.S.S.R. in the 1950s, even controlling for high investment: it suggests that whatever the weaknesses of Soviet central planning in hindsight, these weaknesses were unlikely to have been apparent before 1960. Possible Explanations for Poor and Declinzing Soviet Growtvt Before turning to the classic, extensive growth hypothesis, we first consider two other possible factors in the relative Soviet decline: the defense burden and Soviet disincentives for innovation. Could the poor and declining growth per- formance be explained by the burden of defense on the Soviet economy? Al- though measurement is problematic, the burden seems to have been high and rising. In table 3, we show some estimates of the Soviet defense burden as a share of GDP. Over the entire period since 1928, Soviet defense spending rose from 2 percent of GDP to the much higher levels of the mid- and late 1980s, around 15 to 16 percent of GDP. Over the period 1960-89 in which the Soviet growth decline occurred, the rise in the defense burden is more modest-from 10 to 13 percent in 1960 to 12 to 16 percent in the 1980s. The international evidence for adverse effects of defense spending on growth is ambiguous-see Landau (1993) for a recent survev. Landau (1993) himself finds an inverted U relationship: military spending below 9 percent of GDP has a positive effect on growth, but above 9 percent of GDP it has a negative effect on growth. To see whether this affects the Soviet dummy in the growth regressions, we inserted defense spending into the decade-average growth regressions per- formed earlier (equation 2). We also included a variable measuring war casual- ties per capita on national territory to ensure that the military spending variable was not simply proxying for wars. Because the military spending data are avail- able only for recent periods, we used data from the 1980s only. Landau covers only developing economies, so we used, instead, data from Hewitt (1993) that covers all economies (including the U.S.S.R. itself). The data for both Landau and Hewitt are mainly from the Stockholm Interniational Peace Research Insti- tute. The data on war casualties are from Easterly and others (1993). The re- gression including a quadratic function of military spending is as follows: (4) Per capita growth 1980-88 = -0.003 + 0.127 investment/kao 1980-88 - 2.7E-06 (.iui per capita 1980 10.(17) (0.038) 1. I F-06) -1.34 population growth 1980-88 +0.007 secondary enrolliioeitt 1970 ± 0.0081 inilitarv spending/.DP 1980-88 (0.38) c0.0171 (0.0024) -0.00041 (military spendinghc.ir)2 - 0.746 war casualric per iaplta 1980-88 - 0.0155 dummiy for U.S.S.R (0.00011 (0.343) (U.0268t least squarcs weighted hv log of total CtDP 77 observationis R- (weighted) = 0.59 stanidard errors in parentheses 348 THE WORLD BANK ECONOMIC REVIEW, VOL. 9, NO. Table 3. The Soviet Defense Burden as a Share of GDP, Based on Different Data Sources, 1928-87 (percentage of GDP) Brada and Graves Year Ofer Higb estimate Low estimate Steinberg 1928 2 1950 9 - - - 1960 12 13.34 9.90 - 1961 - 13.86 10.60 - 1962 - 14.93 11.39 - 1963 - 15.49 12.32 - 1964 - 15.03 12.17 - 1965 - 14.49 11.79 - 1966 - 14.11 11.54 - 1967 - 14.40 11.95 - 1968 - 14.45 12.14 - 1969 - 14.61 12.08 - 1970 13 13.83 11.48 13.28 1971 - 13.56 11.30 13.76 1972 - 13.80 11.34 13.61 1973 - 13.33 11.03 13.14 1974 - 13.71 11.28 13.15 1975 - 14.14 11.53 13.57 1976 - 14.32 11.62 13.30 1977 - 14.07 11.26 12.98 1978 - 14.00 11.09 13.08 1979 - 14.53 11.43 13.05 1980 16 15.06 11.82 13.91 1981 - 15.48 11.75 14.03 1982 - 15.36 11.70 14.58 1983 - 15.51 11.63 14.36 1984 - 15.55 11.57 14.37 1985 - - - 14.79 1986 - - 14.49 1987 - - 14.63 -Not available. Note: The Ofer growth rates are based on current rubles; those of Brada and Graves and Steinberg are based on constant rubles. Source: Brada and Graves (1988), Ofer (1987), and Steinberg (1987, 1990). We confirmed Landau's result of an inverted U-shaped relationship between growth and defense spending. Military spending reduces the magnitude and sig- nificance of the Soviet dummy. However, as Landau also notes, this result is not very robust; omitting Israel and Syria from our sample eliminates the signifi- cance of military spending. The defense explanation for the Soviet decline is plausible but not firmly established with cross-sectional data. We will test the defense hypothesis further with the Soviet time series in the production function estimates in the next section. Another possible explanation for poor and declining Soviet growth could be adverse incentives under central planning for technological innovation (Berliner 1976). Recent theoretical and empirical literature argues that endogenous tech- nological innovation, as measured by resources devoted to research and devel- Easteriv and Fischer 349 opment (R&D), significantly explains relative growth performance across econo- mies (Coe and Helpman 1993; Lichtenberg 1992; Romer 1989). Western estimates of the Soviet research effort, presented in figure 1, show R&D spending rising as a share of GDP. The R&D share is above the 2 to 3 percent of GDI' in the leading industrialized economies. In 1967, about 1.5 per- centage points of Soviet GDP was estimated to be for defense and space R&D (Bergson 1983). The share of defense and space R&D in total R&D is believed to have fallen in the period 1959-84 (Acland-Hood 1987), implying an even steeper rise in civilian R&D. It is well known that the lack of market incentives made the allocation of Soviet R&D inefficient and inhibited the diffusion of innovations. This would explain a low growth payoff for a given amount of R&D; it does not explain why that growth payoff would have fallen over time. The Extensive Growth Hypothesis As noted in the introduction, the conventional hypothesis for the Soviet de- cline in growth is the pattern of extensive growth, defined by Ofer (1987) as growth mainly through input accumulation rather than through productivity growth. Ofer notes that a key feature of extensive growth is a rising capital- output ratio. Figure 2 shows the evolution of the capital-output ratios implied by the alternative data series for 1950-87. We begin the graphs in 1950 because of the extreme volatility of all of the capital-output series before 1950, which no doubt is related to shocks such as collectivization and World War II as well as questionable data quality. The total economy (GDP) series in figure 2 shows the capital-output ratio increasing two and a half times between 1950 and 1987. Figure 1. Research and Developinent Expenditures as a. Percenltage Ot/GDP in the US.SXR., 1950-87 R&D expenditure (percenitage of GDP) 3.5- 3.() 2. 9 2.0 1.9 1 .A ) I I I I I I I I I I I 19S() 1993 1996 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 SuourLce joint Fc)nonmi Committee (1990). 350 THE WORLD BANK ECONONMI REVIEW, VOL. 9. NO. . Figure 2. Capital-Output Ratios in the U1.S.S.R., 1950-86 Capital-output ratio 5.0 - _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 4.5- Totail econoImy/ '4.0- 3.5- 3.0 -, '-_. < ~~~~~~~~~~~Industryj 2.5 ---- -- -------- -- - - - 2.0 - 195(0 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 Source: Authors' calculations based on Western itt.a. The capital-output ratios industry first decline in the 1950s and then rise sharply after 1960. The capital-output ratio in the official series also rises steadily beginning at the end of the 1950s, more than doubling between 1958 and 1987. Khanin's data, by contrast with the other two series, show only a small increase in the capital-output ratio between the early 1950s and 1987. In the neoclassical model, a rising capital-output ratio implies capital deepen- ing during the transition to a higher steady state, but this capital deepening will sooner or later run into diminishing returns that will cause growth to slow or stop. With technical change, steady-state growth is feasible with a constant capi- tal-output ratio. If the Soviet reliance on capital deepening is to explain its growth decline in relation to market economies, then market economies must have rela- tively constant capital-output ratios. Capital-output ratios in market economies have indeed long been thought to remain relatively stable, according to the fa- mous Kaldor stylized fact (see, for example, the recent statement by Romer 1 990). However, recent research on capital accumulation in market economies con- tradicts the Kaldor stylized fact, shedding a new perspective on the Soviet expe- rience of extensive growth. Appendix table A-1 lists the annual growth rates of Easterly and Fischer 351 the capital-output ratios in a selection of recent growth-accounting studies and a few older ones. All studies agree that the capital-output ratio in the United States has remained remarkably constant, which perhaps accounts for the con- ventional wisdom that Kaldor's stylized fact holds. But several recent studies point to capital-output ratios rising at Soviet-style rates in Japan and in some of the East Asian newly industrializing economies (NiCs), such as Korea (Young 1994b; Kim and Lau 1994; King and Levine 1994; Benhabib and Spiegel 1994; Nehru and Dhareshwar 1993).6 These studies present a new fact for the litera- ture on Soviet extensive growth to explain: why did the extensive growth strat- egy lead to eventual stagnation in the U.S.S.R., whereas the same strategy sus- tained rapid growth in Japan and Korea? Moreover, King and Levine (1994), Benhabib and Spiegel (1994), and Nehru and Dhareshwar (1993) show that rising capital-output ratios are a feature of growth for many economies. (See also Judson 1994, who shows the capital- output ratio rising systematically with income.) The three studies compute capi- tal stocks for a large sample of economies, using a variety of data sources (such as Summers and Heston 1991 and World Bank data) and a variety of assump- tions about initial capital stocks and depreciation rates. The three concur that rising capital-output ratios are by no means rare: the median annual rate of growth in the capital-output ratio in their respective samples is around 1 per- cent, and in a full quarter of the samples the annual rate of growth in the capi- tal-output ratio is greater than 1.7 percent.7 And it is not only developing econo- mies that are shown to have rapid capital deepening. For example, the studies concur that capital-output ratios in Austria and France increased at more than 1.5 percent a year. As either a cause or a consequence of falling output growth, the level of the Soviet capital-output (K-Y) ratio had become extreme by the 1980s. The K-Y ratio as measured by the Western GDP and total capital stock series was 4.9 in 1985, which is higher than any of the 1985 K-Y ratios in the Benhabib-Spiegel and King-Levine exercises. In the Nehru-Dhareshwar sample there are onlv four economies with a K- Y ratio above that of the U.S.S.R. in 1985, none of which seems especially relevant as a comparator-Guyana, Jamaica, Mozambique, and Zambia. One other implication of the extensive growth model-also pointed out by Ofer (1987)-is that investment ratios have to rise over time if growth is to be maintained while the capital-output ratio rises. The Soviet investment share doubled between 1 950 and 1975, as can be seen in the Western estimates pre- sented in figure 3. After 1975 the investment share continued to increase, but more slowly. 6. The literature has less of a consensus on whether the East Asian economies have high TFP growth. But even those who argue for high productivity growth acknowledge rising capital-output ratios in East Asia. See, for example, World Bank (1993) and Pack and Page (1994). 7. For the two studies that use Summers and Heston data )Benhabih and Spiegel 1 994 and King and Levine 1994), we omitted Africa from the sample because ratios of investment to GDA are implausibly extreme (both high and low) for the 1950s. 352 THE WORLD BANK ECONOMIC REVIEW, VOL. 9, NO 3 Figure 3. Share of Total Investment in GNP in1 the USSR., 1950-86 Total investment in GNP (percent) 35 - 30- 25 - 20- 15 - 20 -/ 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 Source. Joint Economic Committee (1990). How unusual is the doubling of the investment rate over a twenty-five-year period? In the Summers and Heston (1991) international data base for 1950-75, eight of fifty-two economies-most notably Japan and Taiwan (China)-had a doubling or more of investment rates.8 When the sample period is shifted for- ward by ten years to expand the sample, six of seventy-two economies had a more than doubling of investment during 1960-85, among which fast-growing Korea and Singapore are of particular interest. Again, Soviet extensive growth was not so different from East Asian extensive growth. What was unusual was the fact that Soviet growth was declining while its investment was rising. The standby of Soviet industrialization-investment in machinery and equipment- also increased sharply as growth declined. The importance of machinery to growth has been emphasized by de Long and Summers (1991, 1992, 1993), but the Soviet data suggest that a high ratio of machinery investment to GNP is not sufficient to generate growth. It. PRODUCTION FUNCTIONS AND EXTENSIVE GROWTH If extensive growth explains the Soviet growth slowdown, we should see it reflected in a falling marginal return to capital accumulation. The alternative explanation for falling growth would be a slowdown in total factor productivity (TFP) growth for some reason other than excessive reliance on capital accumula- 8. We continue to exclude African economies from this and the following sample. Easterly and Fischer 353 tion. In this section, we assess the consistency of the data with these alternative explanations. Productivity Growth In table 4 we show TFP growth for the total economy and the industrial sector in the U.S.S.R. TFP growth is calculated assuming a Cobb-Douglas production function with labor's share equal to 0.6 and the share of capital equal to 0.4 (slightly above that used by Bergson (1979, 1983) and the cIA (1982), but within the conventional range for developing economies).9 With the assumption of Cobb- Douglas production (unit elasticity of substitution between capital and labor), we see a strongly declining trend in TFP growth after the 1950s. The most interesting aspect of table 4 is that the 1950s once more stand out as an exceptional period in Soviet growth. It is especially striking that even con- servative Western estimates for the industrial sector imply productivity growth in that decade of more than 4 percent a year. The Western GNP data give a more pessimistic assessment of Soviet productivity performance, implying that pro- ductivity growth in the U.S.S.R. started being poor in 1960. By contrast, Khanin's data, which uniformly exhibit lower overall growth than Western GNP data, imply positive post- 1950 productivity growth, a result of the lower rates of growth of capital in the Khanin series. There is only a moderate difference between the official and Western data on factor input growth, but Khanin shows substan- tially lower rates of growth of capital (table 1). The lower capital growth rates reflect Khanini's view that hidden inflation is as serious in capital goods indus- tries as in consumer goods, a view shared by the "British school" of Hanson (1984), Nove (1981), and Wiles (1982). How do the Soviet TFI' growth rates look in a comparative context? We calcu- lated TFP growth rates for the East Asian extensive growers assuming the same Table 4. Total Factor Productivity Growth in the Total Economy and in Industry in the U.S.S.R., 1928-87 (average annual percenit) Period Total economiy Industry 1 928-39 0.6 2.4 1940-49 1.3 1.5 1950-59 2.8 4.6 1960-69 0.8 1.4 1970-79 0.1 1.8 1980-87 -0.2 -0. 1 Note: TFP calculated as (growth of output per worker) - 0.4 (growth of capital per worker). Growth rates are taken from table 1. Source: Authors' calculationis. 9. It has lonig been i stylized fact in the development literature that capital shares are higher in developing than in industrial economies (see, for example, De Gregorio's 1992 estimate that the capital share is between 0.4 and 0.55 for Larin America). Western estimates of Soviet per capita income suggest it was a developing rather than an industrial country. The Benhabib-Spiegel, Nehru-Dhareshwar, and King-Levine cross-economy data sets all assume a capital share of 0.4. 354 THE WORLD BANK ECONOMIC REVIEW, VOL.. 9, No. ; Cobb-Douglas production function with a capital share of 0.4 across economies, using the alternative estimates for East Asian accumulation and growth by the various authors. Table 5 shows how Soviet TFP growth rates compare with the East Asian cases in such a comparison. Soviet TFP growth, measured assuming a common production function, was worse than that of the East Asian extensive growers. There is a lively controversy about whether East Asian TFP growth in fact was above average or whether it was low, as argued by Young (1992, 1994a, 1994b). Young's estimates are lower than those shown in table 5 because he estimates a higher capital share, as discussed in section 1. The important point for this section is that East Asia seemed to get more output growth and less diminishing returns out of its extensive growth strategy than the U.S.S.R. did. How does Soviet TFP growth compare with TFP of other economies within the sample of all economies? To answer that question, we performed an exercise similar to our growth regression in section 1. We used TFP estimates from Nehru and Dhareshwar (1993), who also assume a Cobb-Douglas production function with a capital share of 0.4, for a sample of seventy-eight developing and indus- trial economies. As with per capita growth rates, TFP growth rates of small econo- mies seem to have much higher variance than those of large economies.'° Hence we use weighted least squares again with total GDP in 1960 as the weights for a regression of TFP growth rates in 1960-90 on a constant and a dummy for the Soviet observation. The results are as follows: (5) Aniial TFP growth 1960-90 = 1 321 - 1.12 Soviet dummy (0.065)(0.214) least squares weighted by total GDP '8 observations R2 (weighted) = 0.83 standard errors in parentheses Assuming the same Cobb-Douglas production function for all economies, Soviet TFP growth is a little more than 1 percentage point below the GDP- weighted average of productivity growth in all economies, and the difference is significant. The Low Elasticity of Substitution Hypotbesis Table 4 suggests that a slowdown in TFP growth corresponded to the overall growth slowdown. Such a productivity slowdown is puzzling when, as we have seen, R&D expenditures were increasing. However, the TFP calculations in table 4 imposed the Cobb-Douglas production function with a unit elasticity of sub- stitution between capital and labor. Following the pioneering work of Weitzman (1970, 1983) and later contributions (including Desai 1976 and 1987, and Bergson 1979), we investigated whether a constant elasticity of substitution (CES) func- 10. Perkins and Syrquin ( 1989) note some of the extreme negativeTFP observations for small economies; these extreme negative values give a smaller mean T FP growth for small economies than for large ones. We find only a larger variance for small economies, not a significantly smaller mean. Easterlv and Fischer 355 Table 5. Productivity Growth Assuming the .Same Production Function Structure in the U.S.S.R. and in East Asia, Based on Different Data Sources, 1950-88 (average annual percent) Change in capital-output Productiv ity Data source and economy Period ratio growth Western data U.S.S.R. 1950-87 2.5 0.8 1960-8' 2.6 0.2 1970-8' 3.0 -0.3 Young (1994b) Singapore 1970-90 2.8 0.8 Korea, Rep. of (excluding agriculture) 1966-90) 3.6 1.5 Taiwan, China (excluding agriculture) 1966-90 2.6 2.3 Kim and Lau (1994) Singapore 1964-90 1.4 2.1 Korea, Rep. of 1960-90 3.5 1.7 Taiwan, China 1953-90 3.1 2.2 Japan 1957-90 3.2 1.9 Benhabib and Spiegel (I1994) Japan 1965-85 2.6 1.9 Korea, Rep. of 1965-85 2.8 2.1 Singapore 1965-85 2.4 2.1 Taiwan (China) 1965-85 3.0 1.9 Nehru and Dhareshu'ar (1993) Japan 1950-90 2.7 1.7 Korea, Rep. of 1950-90 3.7 2.8 King and Levine (1994) Japan 1950-88 2.3 2.5 Korea, Rep. of 1950-88 3.0 2.0 Singapore 1950-88 2.9 2.8 Taiwan, China 1950-88 2.6 2.4 Note: Productivity growth is calculated assuming a conistanit Cobb-Douglas capital share of 0.4 across all economies, using data on output, capital stocks, and labor. tion provides a better representation of the data than the Cobb-Douglas produc- tion function does. Weitzman's basic finding was that a CES production function with a low elas- ticity of substitution of 0.4 fit the data better than the Cobb-Douglas and that the hypothesis that the elasticity of substitution was one could be rejected. With his specification, declining marginal product of capital, rather than decliningTFP growth, explained the growth slowdown. This result has been hotly debated. Bergson (1983) criticized this result on the grounds that it implied implausibly high estimates of the marginal product of capital in earlier years. Desai (1987) 356 rHE WORLD BANK ECONOMIC REVIEW, VoL. 9, NO. 3 concurred with Weitzman's finding for aggregate industry but argued that Cobb- Douglas was an adequate representation for some branches of industry. In general, there will be multiple specifications of production functions con- sistent with the data. A classic article by Diamond, McFadden, and Rodriguez (1967) showed that it is in general impossible to identify separately a time- varying elasticity of substitution parameter and the bias of technical change (neutral, labor-augmenting, or capital-augmenting). However, following Weitzman, we examined whether some plausible alternatives fit the data better than others. We first identified the substitution parameter by presuming it to be constant over time. We also assumed technical change to be neutral, but we did allow it to vary over time. We then tested the alternative of a low elasticity of substitution against a hypothesis of falling TFP growth. Estimation of production functions in industrial economies is the subject of a large literature. The usual method is to estimate parameters of factor demands derived from the cost function, the dual of the production function (see Jorgenson 1983 for a survey). This is obviously inappropriate for a nonmarket economy, such as the U.S.S.R. Direct estimation of production functions is usually thought to be tainted by endogeneity of the factor supplies, particularly capital; we be- lieved this would be much less of a problem in the nonmarket system of the U.S.S.R. Table 6 shows elasticities of substitution estimated by nonlinear least squares, and recalculated TFP growth rates for 1950-87 (assuming Hicks-neutral techni- cal progress) for subperiods with the CES form, where t denotes time: (6) ln(Y/L) = c It195(-i9 + C,t 1960-69 + c3t197(_ 9 + c4t1980_8' + [c /(l+c,)] lnIc6 (KIL)W' +C5.c5 + (1-C6)I + c.. We find indeed that the Western data on Soviet output and capital growth per worker lend themselves to the CES form with low elasticities of substitution be- tween capital and labor (significantly below one) and roughly constant rates of TFP growth. We regard these results as very suggestive, but we note again the caveats. First, these results are conditional on assumptions that we need tp identify the production function. Second, the results were much less sharp whenh we used the entire 1928-87 sample, where, as indicated earlier, the data before 1950 are volatile. Finally, when we tried similar estimates on Khanin's and official data, we found some support in the official data for the low-elasticity hypothesis but none in Khanin's data. The results with the Khanin data are intriguing because they support a story of unit elasticity of substitution, a high capital share (higher than was imposed for table 4), and poor (although not strongly declining) productivity growth. According to Khanin's data, growth declined mainly because capital growth slowed (see table 1 again). Given Bergson's (1987a, 1991a) criticisms and the Easterly and Fischer 357 Table 6. Elasticities of Substitution and Total Factor Productivity Growth with Estimated Constant Elasticity of Substitution Functions, U.S.S.R, 1950-87 Total Industrial Variable or statistic economv sector Elasticitv of substitution (c5) -0.37 -0.21 (-9.8) (-3.1) TFP growth 1950-59 (cl) 1.09 -0.19 (3.4) (-2.3) 1960-69 (c2) 1. 1 -0.25 (3.1) (-2.9) 1970-79 (c.) 1.16 -0.21 (3.2) (-2.3) 1980-87 (C4) 1.09 -0.22 (3.2) (-2.3) Intercept (c-) -0.82 -0.95 (-9.8) (-16.7) Capital share parameter (cj) 0.96 0.47 (71.2) (6.2) Autocorrelation coefficient na.. 0.81 (15.8) R2 0.9987 ().9995 Durbin-Watson 1.9228 1.2968 n.a. Not applicable Note: Estimates are based on Western data. t-statistics are in parentheses. Source: Authors' calculations. limited information about the methodology behind Khanin's data, these differ- ing results can only point to the need for further research into Khanin's ap- proach to see whether his work represents a valid criticism of the Western esti- mates. For the moment, we are forced to regard conclusions based on Khanin's data as unproven. We would have liked to examine the implications of the "British school" of Hanson (1984), Nove (1981), and Wiles (1982), who made somewhat similar claims to Khanin's. However, we could not do so because those researchers did not provide alternative time series for output and capital. Note that a lower estimate for the growth rate of capital over the entire period, as implied by these authors' arguments, would imply higher TFP growth but does not necessarily imply a higher estimated elasticity of substitution. For our purposes here, the most striking feature of table 6 is that the implied rates of TFP growth show no significant decline between the 1950s and 1980s. Thus, once we free up the functional form of the production function, we find no evidence for a slowdown in TFP growth. In table 4, both extensive growth and declining productivity growth account for the overall fall in growth; in table 6, extensive growth-diminishing returns to capital-accounts entirely for the growth slowdown. Of course, as Ofer (1987) points out, extensive growth and low productivity growth are not necessarily independent of each other. The So- viets may have felt compelled to pursue extensive growth because productivity 358 THE WORLD BANK ECONOMN11C REVIEW, VOL. 9, NO. ; growth was low; there may have been a mechanism in the Soviet system by which rapid growth in capital leads to poor TFP performance. The Western GDP estimates yield a constant TFP growth rate of 1 percent a year, in contrast to the sharply falling TFP growth implied by the Cobb-Douglas estimates in table 4 for the 1970s and 1980s; the industrial TFP growth is slightly negative and also constant. Because the U.S.S.R. was spending increasing R&D resources and borrowing Western technology in the 1970s and 1980s, we find a constant rate of TFP growth more plausible than a falling rate. We cannot see a compelling story for the sort of worsening efficiency in resource use and tech- nology adaptation (in the sense of Nishimizu and Page 1982) that would be required to explain falling TFP growth. In figure 4 we examine a second implication of the estimates in table 6: these are estimates of the "share of capital" implied by the production function pa- rameters for the Western estimate of GDP for 1 950-87, under the hypothetical assumption of marginal productivity pricing. When the elasticity of substitution is less than one and the capital-output share is rising, the share of capital will fall over time. Diminishing returns will not be verv severe with a high capital share, but will be severe with a low capital share. In figure 4 the share of capital falls Figure 4. Share q/GCapital in outtpuit and Marginal Product of Capitalfor the Total Econzomy of the l17.S.S.R.. 1950-87 Percent 80 - 70- 60- Sihare of capital in output 50 40J 30 20- 1Marginal product of capital _ 10 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 Note: Calculated from eWestern estimaites of capital stocks, emiployment, anld vailue addled for industry andi for the whole economy. Source: Authors calculations. Easterly and Fischer 3S9 sharply for the economywide production function (a similar graph obtains for the industry estimates). Figure 4 also shows closely related data on the marginal product of capital implied by the CES estimates. The Western GNP data imply high rates of return to capital in the early 1950s, then the rates declined to about 3 percent in 1987. A CES function with a high capital share acts much like a linear function of capital, so that the marginal product of capital can stay high for as long as the capital share is high. With a very capital-intensive production of goods, includ- ing capital goods, the Soviets were close for a while to the model of growth through rapid reproduction of capital described by Feldman in the 1920s as using machines to make more machines (in Domar 1957). Rebelo (1991) shows formally that constant returns to reproducible factors in the capital goods sector are sufficient to generate a constant, sustained rate of growth even without TFP growth. However, as the capital share begins to fall, the marginal product will begin to decline. The decline can be precipitous when the elasticity of substitution is particularly low. Although we find the extreme values of the marginal product of capital and capital's share in figure 4 surprising, they do not logically rule out the CES form; the capital-labor ratio in a nonmarket economy could be driven to levels that would not be observed in a market economy. Indeed, the data presented in figure 4 suggest that a market economy could not have gone through the growth process of the Soviet economy between 1950 and 1987. The low wage shares in the early period would probably have pre- vented any but a subsistence wage equilibrium in those periods. The low mar- ginal product of capital by the mid-1980s would have been inconsistent with equilibrium and would have meant that investment in industry and the capital- labor ratio would have been lower. What would have happened in the early years if there had been a market economy? One possibility is that different tech- nologies would have been adopted. Similarly, in the later period there may well have been other technologies available-but less amenable to central planning- that yielded a positive return to capital. It is also possible that if the extensive growth route had been closed off in a market economy, there would have been more incentive for Soviet entrepreneurs to attempt to improve technology. International Comparisons of Elasticities of Substitution As we did with the earlier regression estimates, we put the results on the Soviet elasticity of substitution and TFP growth in an international context. We have fragmentary evidence that the low elasticity for the U.S.S.R. is not unique among planned economies. Rusek (1989) reports the remarkably low elasticity of 0.10 for Czechoslovak industry. Sapir (1980) estimates an elasticity of 0.13 for Yugoslav manufacturing. The characterization of the Soviet (and Soviet-type) data by a low elasticity of substitution and constant TFP growth provides a natural way to reconcile the lack of success of the Soviet extensive growth strategy with the high payoffs 360 THE WORLD BANK ECONOMIC REVIE.W, V(L. 9, NO. X from capital deepening in Japan, Korea, and other market economies. If East Asian economies have higher elasticities of substitution, then this would explain the much weaker force of diminishing returns to capital deepening. De la Grandville (1989: 479) insightfully speculates "whether part of the explanation of miracle growth in Japan or South Asian countries lies not only in a high saving rate but also in a high elasticity of substitution between factors in their industrial sectors." Yuhn (1991) tested De la Grandville's conjecture for Korean manufacturing and found an elasticity of substitution near unity. Bairam (1989) presents evi- dence that Japan's pre-World War II development was characterized by an elas- ticity of substitution above unity. Wang (1995) argues for an elasticity of substi- tution above two in manufacturing in Taiwan (China). High elasticities of substitution would have made it easier for East Asian economies to absorb rapid capital accumulation without severe diminishing re- turns. Young (1994b) assumes a unit elasticity of substitution in the East Asian tigers he studies and estimates relatively high capital shares (particularly in Singapore). He thus reconciles high and constant output growth in East Asia with rapid increases in capital-output ratios and less-than-spectacular produc- tivity growth (that becomes even negative TFP growth in Singapore). Similarly, Kim and Lau (1994) cannot reject a zero rate of TFP growth in the East Asian NICS, but have an output elasticity of capital beginning at about 0.6 (and then slowly declining to about 0.4). For market economies in general, a recent study estimating the elasticitv pa- rameter from the convergence behavior of the cross-sectional national per capita income data argues that the elasticity of substitution is slightly above one (Chua 1993). The classic studies by Berndt and Wood (1975) and Hudson and Jorgenson (1978) report elasticities of substitution slightly above one for the United States (an estimate that is somewhat higher than in earlier studies). There has also been a large literature on elasticities of substitution in develop- ing economies. The early focus of this literature was the hypothesis that less- developed economies had fixed-proportions technology (zero elasticity of sub- stitution). This hypothesis has been soundly rejected by the literature. There is little consensus on whether the elasticities of substitution are at or significantly below unity. White's (1978: 33) summary of this literature was that "the esti- mates tend to clump between 0.5 and 1.2." See also the survey by Morawetz (1976), which reports a similarly wide range of estimates. Elasticities of substi- tution around or above unity have been found in places as disparate as Brazilian manufacturing (Tyler 1974), Pakistani manufacturing (Mahmood 1992), Ven- ezuelan manufacturing (Sines 1979), Colombian agriculture (Thirsk 1974), Rho- desian GDP (Muzondo 1978), Egyptian manufacturing (Girgis 1974), Puerto Rican industry (Reynolds and Gregory 1965), Philippine manufacturing (Williamson 1971), and Ghanaian industry (Roemer 1975). Our results in a comparative context suggest the plausible story that dimin- ishing returns to extensive growth were much sharper in the U.S.S.R. than in Easterly and Fischer 361 market economies because the substitutability of capital for labor was abnor- mally low. In the concluding section, we will speculate why substitutability may have been low in a planned economy. Combining Regression Evidence with Production Function Evidence As a final exercise, we inserted the other apparent correlate of declining growth-defense spending-into our production function estimates (we took the midpoint of the Brada and Graves (1988) estimates in table 3 spliced to- gether with the Steinberg (1987, 1990) estimate for 1985-87). Specifically, we allowed the Hicks-neutral rate of technological progress to depend linearly on the share of defense spending in GDP in the production function estimated with Western GDP and capital stock data: (7) ln(Y/L) = co + c1t + c2t(defense spending) + [c5/(l + C5Ol In [c,(KIL)(l+cs)1c5 + (1 - c6)I. The results for the nondefense parameters are virtually identical to the estimates shown in table 6 and are not reported here. We find that defense spending does indeed have a significant and negative effect on the rate of increase in the total productivity term in the production function. However, the effect is not very quantitatively important: every additional 1 percent of GDP spent on defense lowered productivity growth by 0.07 percent. The increase during 1960-87 of 2.2 percentage points in the defense share thus would have lowered growth by 0.15 percentage points. Moreover, the parameters of the CES function are virtu- ally unchanged from our earlier regression so the low substitutability, diminish- ing returns story still holds. We also tried equipment investment and R&D spend- ing as independent influences on the technical progress term, but both gave insignificant results. How do we reconcile our production function estimates with our earlier cross- sectional growth regression evidence using the Levine-Renelt specification? The Soviets' high capital-output ratio and low substitutability of capital for labor implies a lower derivative of growth with respect to the investment rate than in other economies with lower K-Y ratios and more substitutable capital for labor. To see this, assume zero depreciation and labor growth for simplicity and define y and k as output and capital, respectively, per worker. Assume a CES function y = A(ykP + I - #)`P). Growth will be given as a function of the invest- ment ratio (i/y = Ak/y) as follows: (8) Ay'/Y = y (ily) [ (k/y) (r + ( 1- y)k- P) ]] As is well known, a higher k/y implies a lower marginal effect of the investment ratio on growth simply because a given investment rate translates into lower capital growth. With a unit elasticity of substitution (p = 0), this is the only way 362 THE WORLD BANK E(ONOMIC REVIEW, VOL. 9, NO. 3 that the level of capital influences the marginal effect of investment. With a less- than-unit elasticity of substitution (p < 0), higher capital has an even stronger negative effect on the coefficient on investment in a growth equation. Although obviously not the only explanation, this is consistent with the large negative residual for the U.S.S.R.-and increasingly negative residuals over time-in the cross-sectional regressions. In the same vein, why were the Soviet "TFP" growth rates, calculated as (growth per worker) - 0.4(capital growth per worker), so poor compared with the TFP growth rates of other economies? If the true production function is CES with a low elasticity of substitution, then a Cobb-Douglas calculation of TFP will yield the following: (9) Av-yAk/k = y [(y + (1 - I)k-) '-1 Ak/k + AA/A. If the elasticity of substitution is indeed unity (p = 0), then the Cobb-Douglas calculation gives the "true" technical change AA/A. However, if the elasticity of substitution is less than unity (p < 0), measured Cobb-Douglas "TFP growth" will be lower, the higher the stock of capital per worker is. The Soviets' high capital stock and low elasticity of substitution would explain the poor Cobb- Douglas productivity growth compared with that of other economies. We conclude from our reexamination of the aggregate data that-subject to admittedly a priori restrictions on the form of the production function-the original Weitzman (1970) story of Soviet growth deceleration provides a de- scription of the final decades of the U.S.S.R. as well. With neutral technical change and a constant-and low-elasticitv of substitution, the decline in So- viet growth can be explained by diminishing returns to capital accumulation alone. This helps explain why the average growth performance was poor when we take into account the rapid capital growth and high education levels. Our comparative approach illuminates further the extensive growth hypothesis of the literature on Soviet growth, and the low substitutability of capital for labor in the U.S.S.R. explains why the outcome of Soviet extensive growth was so different from that of East Asian extensive growth. III. CONCLUSIONS AND DIRECTIONS FOR FURTHER RESEARCH One natural question to ask is why Soviet capital-labor substitution would have been more difficult than in market economies such as Japan and Korea and whether this difficulty was related to the Soviets' planned economic system. Recent work on models of endogenous economic growth stresses the notion of a broad concept of capital, including human capital, organizational capital, and the stock of knowledge, which can substitute easily for raw labor and perhaps replace it altogether (Rebelo 1991; Jones and Manuelli 1990; Parente and Prescott 1991). Other recent models of endogenous growth stress the incredible diversity of capital goods and intermediate inputs that are successively explored by mar- Easterlv and Fischer 363 ket economies (see the presentation of such models in Barro and Sala-i-Martin 1995, chapter 6). A broad variety of capital goods would be more substitutable for labor than a narrow range of capital goods. Could one explanation for the Soviets' substitution problems be that, under an autocratically directed economic system, they accumulated a narrow rather than a broad range of capital goods because only certain types of capital were amenable to central allocation? Was substitutability low because they were missing such market-oriented types of physical and human capital as entrepreneurial skills, marketing and distribu- tional skills, and information-intensive physical and human capital (because of the restrictions on information flows)? Do other economies-especially devel- oping ones-with highly state-controlled economic systems also have low sub- stitutability between capital and labor? Although diminishing returns are weakened if the elasticity of substitution is high, they still apply. Could Soviet-style growth declines await other economies that rely too much on extensive growth? After all, the U.S.S.R. had its period of rapid growth from the 1930s through 1950s, when it appeared to be following a linear output-capital production function, as we have shown. The Soviet experi- ence can be read as a particularly extreme dramatization of the long-run conse- quences of extensive growth. Weitzman (1990) describes Soviet growth (as ana- lyzed by Ofer 1990) as the best application of the Solow (1956) neoclassical model ever seen. Krugman (1994) has an entertaining description of the unan- ticipated Soviet growth decline as a metaphor for the alleged future growth slowdown in East Asia. Of course, if the story of low elasticity in the U.S.S.R. is correct, then the metaphor is highly exaggerated. Or are there other factors that stave off diminishing returns to capital accu- mulation in East Asia and other market-oriented extensive growers? Rapid growth in human capital would mitigate diminishing returns to physical capital in East Asia. However, we have seen that educational investments were also high in the U.S.S.R. Also, the fact that capital-output ratios rise implies that capital is growing faster than other factors, assuming constant returns to scale. Note that an elas- ticity of substitution above unity between capital and labor limits diminishing returns asymptotically such that sustained growth through capital accumula- tion is possible indefinitely. This is an important area for further research. Our results with the U.S.S.R. in the international cross-sectional growth and productivity regressions suggest that the planned economic system itself was disastrous for long-run economic growth in the U.S.S.R. Although this point may now seem obvious, it was not so apparent in the halcyon days of the 1950s, when the Soviet example was often cited as support for the neoclassical model's prediction that distortions do not have steady-state growth effects. Economic systems with low substitutability may deceptively generate rapid growth with high investment, only to stagnate after some time. Because a heavy degree of planning and government intervention still exist in many developing economies, the eventual fate of the Soviet economic system continues to be of interest. 364 THE WORID BANK I( ONON11( REVIEW, VOL. 9, NO. Table A-1. Trends int Capital-Output Ratios in Growth-Accounting Studies Annual percentage change Source and economny Period in capital-output ratio Western data (this article) U.S.S.R. 1950-87 2.5.3 Maddison ( 1989) Argentina 1950-84 0.61 Brazil 1950-84 0.86 Chile 1950-84 -0.22 China 1950-84 2.48 France 1950-84 -0.45 Germany, Fed. Rep. of 1950-84 0.16 India 1950-84 1.54 Japan 1950-84 -0.91 Korea, Rep. of 1950-84 0.03 Mexico 19.()-84 0.50 Taiwan, China 1950-84 -0.34 United Kingdom 1950-84 0.62 United States 1950-84 -0.07 U.S.S.R. 1950-84 3.75 Young (1 994b) Hong Kong 1966-91 0.84 Singapore 1970-90 2.79 Korea, Rep. of (excluding agriculture) 1966-90 3.62 Taiwan, China (excluding agriculture) 1966-90 2.55 Kim and Lau (1994) France 195.-9() 0.68 Germany, Fed. Rep. of 1960-90 1.16 Hong Kong 1966-90 1.11 Japan 1957-90 3.19 Singapore 1964-90 1.38 Korea, Rep. of 1960-90 3.50 Taiwan, China 1953-90 3.13 United Kingdom 1957-90 0.68 Ulnited States 1948-90 -0.19 Elias (1992) Argentina 1950-80 0.39 Brazil 1950-80 -0.54 Chile 1950-80 -0.39 Colombia 1950-80 -0.79 Mexico 1950-80 0.44 Peru 19(0-80 1.22 Venezuela 1950-8() 0.75 Easterly and Fischer 365 Annual percentage change Source and economy Period in capital-output ratio Chenery, Robinson, and Svrquin (1 986) Canada 1947-73 0.66 France 1950-73 0.13 Germany, Fed. Rep. of 1950-73 0.33 Italy 1952-73 -0.63 Netherlands 1951-73 0.17 United Kingdom 1949-73 0.69 United States 1949-73 0.00 Benhabib and Spiegel (1 994) Hong Kong 196.5-85 -0.20 Japan 1965-85 2.56 Korea, Rep. of 1965-85 2.78 Singapore 1965-85 2.41 Taiwan. China 1965-85 2.97 United States 1965-85 0.63 Percentile of sample, 75th 1965-85 1.72 50th 1965-85 0.80 25th 1965-85 0.21 Nehru and Dhareshwar (1993) Japan 1950-9(0 2.70 Korea, Rep. of 1950-90 3.70 United States 1950)-90 0.20 Percentile of sampleb 75th 1950-90 1.84 50th 1950-90 1.06 25th 1950-9f0 0.38 King and Levine (I 994) Hong Kong 1950-88 -0.80 Japan 1950-88 2.33 Korea, Rep. of 1950-88 3.05 Singapore 1950-88 2.94 Taiwan, China 1950-8X 2.63 United States 1950-88 0.40 Percentile of sample, 75th 1950-88 1.69 50th 1950-88 0.95 25th 1950-88 0.23 a. The sample includes seventy-seven countries, excliding Africa. b. The sample includes seventv-two countries. c. The sample includes seventv-four countries, excluding Africa. 366 THE W'ORL1) BANK ECONOMIC REVIF'V Vl>l . 9, NO. I REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Acland-Hood, Mary. 1987. "Estimating Soviet Military R&D Spending." In Carl G. Jacobsen, ed., The Soviet Defence Enignma. Oxford: Oxford University Press. Bairam, Erkin. 1989. " Learning-by-Doing: Variable Elasticity of Substitution and Eco- nomic Growth in Japan, 1878-1939." journal of Development Studies 25(3):344- 53. Banerjee, Abhijit, and Michael Spagat. 1991. "Productivity Paralysis and the Complex- ity Problem: Why Do Centrally Planned Economies Become Prematurely Gray?" Journal of Comparative Economics 15(December):646-60. Barro, Robert J., and Xavier Sala-i-Martin. 1995. Econo)mic Growth. New York: McGraw-Hill. Benhabib, Jess, and Mark M. Spiegel. 1994. "The Role of Human Capital in Economic Development: Evidence from Aggregate Cross-Country Data." Journal of Monetary Economics 34(October):143-73. Bergson, Abram. 1961. The Real National Income of Soviet Russia since 1928. Cam- bridge, Mass.: Harvard University Press. .1979. "Notes on the Production Function in Soviet Postwar Industrial Growth." Journal of Comparative Economics 3(2, June): 116-26. . 1983. "Technological Progress." In Abram Bergson and Herbert Levine, eds., The Soviet Economy: Toward the Year 2000. London: Allen & Unwin. 1987a. "On Soviet Real Investment Growth." Soviet Studies 39 (3, July):406- 24. 1987b. "Comparative Produ(ctivity: The U.S.S.R., Eastern Europe, and the West." American Economic Review 77(3):342-57. .1991a. "Real National Income Measurement in Soviet Perspective." Paper pre- pared for the Bureau of Labor Statistics (BLS) and Eurostat Conference on Economic Statistics for Economies in Transition. Washington D.C. February. Bureau of Labor Statistics, Washington, D.C. Processed. . 1991b. "The U.S.S.R. before the Fall: How Poor and Why?" Journal of Eco- nomic Perspectives 5(4, Fall):29-44. Berliner, Joseph S. 1976. The Innovationi Decisiont in Soviet Industry. Cambridge, Mass.: MIT Press. Berndt, Ernst, and David 0. Wood. 1975. "Technology, Prices, and the Derived De- mand for Energy." The Review of Economics and Statistics 57(3):259-68. Brada, Josef C., and Ronald L. Graves. 1988. "Slowdown in Soviet Defense Expendi- tures." Southern Economic Journal 54:969-84. Chenery, Hollis, Sherman Robinson, and Moises Syrquin. 1986. Industrialization and Growth. New York: Oxford University Press. Chua, Hak. 1993. "Factor Substitution and Slow Convergence." Economic Growth Center Discussion Paper. Yale University, Economic Growth Center, New Haven, Conn. Processed. CIA (U.S. Central Intelligence Agencv). 1982. U.S.S.R.: Measures of Economic Growth and Development. Washington, D.C.: U.S. Government Printing Office. Easterly and Fischer 367 Various years. Handbook of Economic Statistics. Washington, D.C.: U.S. Gov- ernment Printing Office. Coe, David T., and Elhanan Helpman. 1993. "International R&D Spillovers." NBER Working Paper 4444. National Bureau of Economic Research, Cambridge, Mass. Processed. De Gregorio, Jose. 1992. "Economic Growth in Latin America." Journal of Develop- ment Economics 39(1):59-85. De la Grandville, Olivier. 1989. "In Quest of the Slutsky Diamond." American Eco- nomic Reviiew 79(June):468-81. de Long, J. Bradford, and Lawrence Summers. 1991. "Equipment, Investment, Rela- tive Prices, and Economic Growth." Quarterly Journal of Economics 106(2, May):445-502. . 1992. "Equipment Investment and Economic Growth: How Robust is the Nexus?" Brookings Papers on Economnic Activity no. 2:157-211. . 1993. "How Strongly Do Developing Countries Benefit from Equipment In- vestment?" Journal of Monetary Economiics 32(3):395-415. Desai, Padma. 1976. "The Production Function and Technical Change in Postwar So- viet Industry: A Reexamination." Americani Economic Review 66(3, June):372-81. 1987. The Soviet Economty: Problemzs and Prospects. Oxford: Basil Blackwell. Diamond, Peter. Daniel McFadden, and Miguel Rodriguez. 1978. "Measurement of the Elasticity of Substitution and Bias of Technical Change." In Melvyn A. Fuss and Daniel McFadden, eds., Productionz Economics: A Dual Approach to Theory and Applications. Amsterdam: North-Holland. Domar, Evsev. 1957. "A Soviet Model of Growthi." In Essavs in the Theory of Eco- nomic Growth. New York: Oxford University Press. Easterly, William, Michael Kremer, Lant Pritchett, and Lawrence H. Summers. 1993. "Good Policy or Good Luck? Country Growth Performance and Temporarv Shocks." Journal of Monetary Economics 32:459-83. Elias, Victor J. 1992. Souirces of Growth: A Study of Seven Latin Anmerican Economies. San Francisco: i(cS Press. Ericson, Richard E. 1990. "The Soviet Statistical Debate: Khanin vs. Tssu." In H. S. Rowen and Charles Wolf, Jr., eds., The Impovierished Superpower. San Francisco: ics Press. Feschbach, Murray. 1983. "Population and Labor Force." In Abram Bergson and H. S. Levine, eds., The Soviet Economny: Toward the Year 2000. Boston: Allen & Unwin. Fischer, Stanley. 1992. "Russia and the Soviet Union: Then and Now." In Olivier J. Blanchard, Kenneth A. Froot, and Jeffrey D. Sachs, eds., The Transition in Eastern Europe. Vol. 1. Chicago: The University of Chicago Press. Girgis, Maurice. 1974. "Aggregation and Misspecification Biases in Estimates of Fac- tor Elasticity of Substitution: The Case of Egypt." Weltuvirtschaftliches Archiv 110(1):114-47. Gomulka, Stanislaw, and Mark Schaffer. 1991. "A New Method of Long-Run Growth Accounting, with Applications to the Soviet Economy, 1928-87, and the U.S. Economy, 1949-78. Centre for Economic Performance Discussion Paper 14. Lon- don School of Economics and Political Science, London. Processed. Hanson, Phillip. 1984. "New Evidence about Soviet Investment." Radio Liberty Re- search 13 September. 368 THE WORLD BANK FCONOMI(: REVIFW, VOL. 9, No. ; Hewitt, Daniel. 1993. "Military Expenditures, 1972-1990: The Reasons behind the Post-1985 Fall in World Military Spending." IMv1F WP/93/18. International Monetary Fund, Fiscal Affairs Department, Washington, D.C. Processed. Hudson, Edward A., and Dale W. Jorgenson. 1978. "Energy Policy and U.S. Economic Growth." American Economic Review 68(2):118-23. IMF (International Monetary Fund), World Bank, Organisation for Economic Co- operation and Development, and European Bank for Reconstruction and Develop- ment. 1991. A Study of the Soviet Economy. 3 vols. Washington, D.C. Joint Economic Committee. 1990. Measures of Soviet GNP. Washington, D.C.: U.S. Government Printing Office. Jones, Larry E., and Rodolfo Manuelli. 1990. "A Convex Model of Equilibrium Growth: Theory and Policy Implications." Journal of Political Economy 98(October):1008- 38. Jorgenson, Dale W. 1983. "Econometric Methods for Modeling Producer Behavior." In Zvi Griliches and Michael Intriligator, eds., Handbook of Econometrics. Vol. 3. Amsterdam: North-Holland. Judson, Ruth. 1994. "Do Low Human Capital Coefficients Make Sense? A Puzzle and Some Answers." Massachusetts Institute of Technology, Department of Economics, Cambridge, Mass. Processed. Kellogg, Robert L. 1989. "Inflation in Soviet Investment and Capital Stock Data and Its Impact on Measurement of Total Factor Productivity." Central Intelligence Agency, Office of Soviet Analysis, Washington, D.C. Processed. Khanin, G. 1. 1988. "Ekonomicheskii Rost: Al'ternativnaia Otsenka." Kommunist no. 17. Kim, Jong-l1, and Lawrence J. Lau. 1994. "I'he Sources of Economic Growth of the East Asian Newly Industrialized Countries." Journal of the Japanese and Interna- tional Economies 8(September):235-7/1. King, Robert, and Ross Levine. 1994. "Capital Fundamentalism, Economic Develop- ment, and Economic Growth." Carnegie-Rochester Series on Public Policy 40(June):259-92. Krugman, Paul. 1994. "The Myth of Asia's Miracle." Foreign Affairs 73(6):62-78. Landau, Daniel. 1993. "The Economic Impact of Militarv Expenditures." Policy Re- search Working Paper 1138. World Bank, Policy Research Department, Washing- ton, D.C. Processed. Levine, Ross, and David Renelt. 1992. 'A Sensitivity Analysis of Cross-Country Growth Regressions." American Economic Review 82(4, September): 942-63. Lichtenberg, Frank R. 1992. "R&D Investment and International Productivity Differ- ences." NBER Working Paper 4161. National Bureau of Economic Research, Cam- bridge, Mass. Processed. Maddison, Angus. 1989. The World Economy in the 20th Century. Paris: Development Center of the Organisation for Economic Co-operation and Development. Mahmood, Zafar. 1992. "Factor Price Shocks, Factor Substitution, and Their Implica- tions for Policy." International Economic Journal 6(4):63-73. Marer, Paul, Janos Avray, John O'Connor, Martin Schrenk, and Daniel Swanson. 1992. Historicallv Planned Economies: A Guide to the Data. World Bank: Washington, D.C. Easterly and Fischer 369 Moorsteen, Richard H., and Raymond P. Powell. 1966. The Soviet Capital Stock, 1928- 1962. Homewood, Ill.: Irwin. Morawetz, David. 1976. "Elasticities of Substitution in Industry: What Do We Learn from Econometric Estimates?" World Development 4(1): 11-15. Muzondo, Timothy R. 1978. "Sectoral Elasticities of Substitution between Capital and Labour: Rhodesia, 1954-75." South African Journal of Economics 46(1):36-47. Nehru, Vikram, and Ashok Dhareshwar. 1993. "A New Database on Physical Capital Stock: Sources, Methodology, and Results." Revista de Analisis Econ6mico 8(1):37- 59. Nishimizu, Mieko, and John M. Page, Jr. 1982. "Total Factor Productivity Growth, Technological Growth, and Technical Efficiency Change: Dimensions of Productiv- ity Change in Yugoslavia, 1965-78." Economizic Journal 92(368):920-36. Nove, Alec. 1981. "A Note on Growth, Investment, and Price Indices." Soviet Studies 23(1, January): 143-4 S. Ofer, Gur. 1987. "Soviet Economic Growth: 1928-85.-" Journal of Economic Litera- ture 25 (4, December): 1767-1833. 1990. "Macroecononic Issues of Soviet Reforms." In Olivier J. Blanchard and Stanley Fischer, eds., NBER Macroeconomiics Annual 1990. Cambridge, Mass.: MIT Press. Pack, Howard, and Johni M. Page, Jr. 1994. "Accumulation, Exports, and Growth in the High-Performing Asian Economies." Carnegie-Rochester Series on Public Policv. 40(June): 199-236. Parente, Stephen L., and Edward C. Prescott. 1991. "Techniology Adoption and Growth." NBER Working Paper 3733. National Bureau of Economic Research, Cambridge, Mass. Processed. Perkins, Dwight W., and Moshe Svrquin. 1989. "Large Countries: The Influence of Size." In Hollis Chenery and T. N. Srinivasan, eds., Handbook of Development Eco- nomics. Vol. 2. Amsterdam: North Holland. Powell, Raymond P. 1963. "Industrial Production." In Abram Bergson and Simon Kuznets, eds., Economic Trends in the Soviet Union. Cambridge, Mass.: Harvard University Press. . 1968. "The Soviet Capital Stock and Related Statistical Series for the War Years." Yale University, Economic Growth Center, New Haven, Conn. Processed. Rebelo, Sergio. 1991. "Long-Run Policy Analysis and Lonig-Run Growth." Journal of Political Economy 99(3, June):500-2 1. Reynolds, L. G., and Peter Gregory. 1965. "Wages, Productivity, and Industrialization in Puerto Rico." Yale University, Economic Growth Center, New Haven, Conn. Processed. Roemer, Michael. 1975. "The Neoclassical Employment Model applied to Ghanian Manufacturing." Journal of Development Studies 11(2):75-92. Romer, Paul. 1989. "What Determines the Rate of Growth and Technical Change?" PRE Working Paper 279. World Bank, Country Econiomics Department, Washing- ton, D.C. Processed. 1990. "Capital Accumulation in the Theory of lIong-Runi Growth." In Robert Barro, ed., Modern Business Cycle 7heory. Cambridge, Mass.: Harvard University Press. 370 THE WORLD RANK ECONOMIK REVIEW, VOL. , NO. 3 Rusek, Antonin. 1989. "Industrial Growth in Czechoslovakia, 1971-85: Total Factor Productivity and Capital-Labor Substitution." journal of Comparative Economics 13(2):301-13. Sapir, Andre. 1980. "Economic Growth and Factor Substitution: What Happened to the Yugoslav Miracle?" The Economic Journal 90:294-313. Sines, Richard H. 1979. "Sectoral Elasticities of Substitution between Labor and Capi- tal in Venezuelan Manufacturing: A Cross-Sectional Micro Analysis." World Devel- opment 7(1):79-82. Solow, Robert M. 1956. "A Contribution to the Theory of Economic Growth." Quar- terly journal of Economics 70(1, February):65-94. Steinberg, Dmitri. 1987. "Estimating Total Soviet Military Expenditures: An Alterna- tive Approach Based on Reconstructed Soviet National Accounts." In Carl G. Jacobsen, ed., The Soviet Defence Enigma: Estimating Costs and Burden. New York: Oxford University lPress. . 1990. "Trends in Soviet Military Expenditure." Sov.iet Studies 42(4):675-99. Summers, Robert, and Alan Heston. 1991. "The Penn World Table (Mark 5): An Ex- panded Set of International Comparisons, 1950-88." Quarterly Journal of Econlom- ics 106(May):327-68. Thirsk, Wayne. 1974. "Factor Substitution in C.olombian Agriculture." American Journal of Agricultural Economics 56(1):73-84. Tyler, William G. 1974. "Labour Absorption with Import-Substituting Industrializa- tion: An Examination of Elasticities of Substitution in the Brazilian Manufacturing Sector." Oxford Ecotionoic Papers 26(1):93-103. UNESCO (United Nations Educational, Scientific, and Cultural Organization). 1975. UNESCO Statistical Yearbook. New York: United Nations. Wang, Eric C. 1995. "Factor Substitution Approach to Testing the Time Variations in Production: The Case of Taiwan's Manufacturing." Applied Economics 27(Janu- ary):107- 16. Weitzman, Martin L. 1970. "Soviet Postwar Economic Growth and Capital-Labor Sub- stitution." American Economic Review 60(5, December):676-92. -. 1983. "Industrial Production." In Abram Bergson and H. S. Levine, eds., The Soviet Economy: Toward the Year 2000. Boston: Allen & Unwin. 1990). "Comment." In Olivier J. Blanchard and Stanley Fischer, eds., NBER Macroeconomics Annual: 1990. Cambridge, Mass.: MIT Press. White, Lawrence J. 1 978. "The Evidence on Appropriate Factor Proportions for Manu- facturing in Less Developed Countries: A Survey." Economic Development and Cul- tural Change 27(1):27-59. Wiles, Peter. 1982. "Soviet Consumption and Investment Prices and the Meaningful- ness of Real Investment." Souiet Studies 24(2, April):289-95. Williamson, J. G. 1971. "Relative Price Changes. Adjustment l)ynamics and Productiv- ity Growth: The Case of Philippine Manufacturing." Economic Development and Cultural Change 19(4):507-26. World Bank. 1993. The East Asian Miracle. New York: Oxford University Press. Young, Alwyn. 1992. "A Tale of Two Cities: Factor Accumulation and Technical Change in Hong Kong and Singapore." In Olivier J. Blanchard and Stanley Fischer, eds., NBER Macroeconomics Annual 1992. C.amlbridge, Mass.: MIT Press. Easterlv and Fischer 371 . 1994a. "Lessons from the East Asian NKls: A Contrarian View." European Economic Revieuw 38(4)964-73. . 1994b. "The Tyranny of Numbers: Confronting the Statistical Realities of the East Asian Growth Experience." NBER Working Paper 4680. National Bureau of Economic Research, Cambridge, Mass. Processed. Yuhn, Ky Hyang. 1991. "Economic Growth, Technical Change Biases, and the Elastic- ity of Substitution: A Test of the De la Grandville Hypothesis." Review of Eco- nomics and Statistics 73(2):340-46. THE WORLD BANK FCONOMil REVIE W VOl. 9, NO. 3 371-395 Contingent Valuation and Actual Behavior: Predicting Connections to New Water Systems in the State of Kerala, India Charles C. Griffin, John Briscoe, Bhanwar Singh, Radhika Ramasubban, and Ramesh Bhatia In 1988, families in Kerala State in India were surveyed to ascertain their willingness to pay for household connections to a piped water supply system. In 1991 the fami- lies in these communities were surveyed again and their actual decisions recorded. This article explores the validity of the findings of the 1988 study on the basis of actual behavior. It looks at the question of benefit revelation: did people behave as they said they would? And it looks at the question of benefit transfer: did people in one site behave as they were predicted to behave, on the basis of the predictions of a behavioral model for a different site? The data were also used to analyze the policy relevance of behavioral modeling. The ability to put a value on environmental resources is a core problem in envi- ronmentally sustainable development in industrial countries (Carson and Mitchell 1993 and Mitchell and Carson 1989) and developing countries alike (Serageldin and Steer 1994). During the past twenty years there has been a vigorous and contentious debate about the relative merits of various approaches (Brookshire and others 1982; Arrow and others 1993). The "indirect approach" draws con- clusions from actual behavior; the "direct approach," or the contingent valua- tion method, draws conclusions from responses to hypothetical questions. The "benefit-transfer" issue in environmental economics, which is concerned with transferring valuations from one population to estimate how a second popula- tion would value the same resource, further complicates the debate (Pearce 1993). Charles C. Griffin and John Briscoe are with the Eastern Africa Department, Population and Human Resources Division, anid with the Transport. Water, and Urban Development Department, Water and Sanitation Division, at the World Bank; Bhanwar Singh and Radhika Ramasubban are with the Centre for Social and Technological Change, Bombay, India; and Ramesh Bhatia is with the Institute of Economic Growth, Delhi University, India, and the International Institute for the Management of Irrigation (i[Mi), Colombo, Sri Lanka. Financial support for this work was provided by the Danish International Development Agency, the United Nations Development Programme, the Swiss Development Corporation, the Norwegian Agency for Development, and the World Bank. The study benefited from the assistance and cooperation of the Kerala Water Authority. David Guilkey of the University of North Carolina assisted with the econometric analyses. The data are available from John Briscoe. The authors wish to thank Dale Whittingtoni for helpful advice in revising the manUscript and three reviewers for their excellent comments. (© 1995 The International Bank for Reconstructioni and l)evelopment / TiHk WORLI) BANK 373 374 THE WORLD BANK ECONOMIC: REVIEW. VOL. 9, N). 3 These approaches have been applied to other areas, including water supply policy in developing countries (World Bank 1992). The fundamental issue in water supply policy is predicting the response of consumers to a service to which they have not previously had access or to characteristics that they have not previously experienced. Piped water supply, higher prices, and improved reli- ability are examples (see Briscoe and others 1990; World Bank Water Demand Research Team 1993). Mirroring the debate in the environmental economics literature, two funda- mental approaches are used to analyze this fundamental issue. The indirect ap- proach involves observing actual behavior, modeling this behavior, and then deriving the willingness to pay for water connections from the value of time spent fetching water or from housing values (see North and Griffin 1993, for example). The direct (contingent valuation) approach, by contrast, involves tak- ing a survey (through a carefully designed and administered questionnaire) of households' willingness to pay specified prices for hypothetical services. Our first concern is with the validity of the direct approach in predicting actual behavior. In environmental economics, because contingent valuation methods are used to value public goods or environmental amenities, it is not possible to validate the hypothetical responses of the interviewed population through actual market behavior. For a water system, however, it is possible to test the results of a contingent valuation survey by comparing the responses given when the water system was hypothetical to the actual behavior once the water system becomes available. Our second concern is to test the benefit-transfer hypothesis as it applies to direct valuation approaches. Contingent valuation methods are subject to hypo- thetical bias, strategic bias, and compliance bias. Hypothetical bias can be re- duced if the sample is well aware of the nature of the good; strategic bias can be reduced if the sample has little or nothing to gain by undervaluing the good. Compliance bias can be reduced through careful development of the survey, training, and supervision of the fieldwork. These problems can be reduced, at least hypothetically, by using the benefit-transfer approach, in which the behavior of a group that already has the service is projected onto a second group to predict the second group's willingness to pay for the good or service in question. This approach may be of interest for three reasons: the second group has little knowl- edge of the good, the second group may behave strategically, and the technique requires little additional data collection. The contingent valuation survey in Kerala was carried out in pairs of communities that differed primarily in their having or not having a water system available. We can therefore test the accuracy of benefit-transfer predictions from the villages where water systems already existed in relation to what we will call "benefit-revelation" predictions based on the contingent valuation responses of the villages where no water system existed. Our third concern is to test the relative accuracy of simple tabulations of willingness to pay from a contingent valuation survey and predictions from a behavioral model using the same data. Simple tabulations may provide all the Griffzn and others 37S information we need to decide whether people will hook up to a new water system under a planned tariff structure. Behavioral models are also important, however, because they allow us to simulate the impacts of changes in policy variables, estimate price elasticities, quantify changes in welfare, and control for nonpolicy variables that also affect behavior. Yet such models may be misspecified, the functional form may be inappropriate, or the assumed distri- bution of the errors may be wrong. We are able in this analysis to test actual behavior against both simple tabulated responses and predicted responses based on a behavioral model. In this article we analyze two surveys of willingness to pay for improved water services in Kerala State in India. The first survey was done in 1988; a follow-up survey of the same households was conducted in 1991. Section I recaps the findings of the 1988 study, and section II describes the 1991 follow-up survey. The article then explores the validity of the findings of the 1988 study on the basis of actual behavior. Section III looks at results for the scarce-water environment and section IV for the saline-water environment. Section V presents conclusions. 1. RECAP OF THE 1988 SmLIDY The 1988 study of the willingness to pay for water in villages in northern areas of the Indian state of Kerala was part of a multicountry study of willing- ness to pay for water in rural areas of developing countries (World Bank Water Demand Research Team 1993). Singh and others (1993) reported the results of the 1988 contingent valuation study. The original study conducted contingent valuation surveys among families living in three types of traditional drinking water environments-abundant water, scarce water, and salt-water intrusion (the latter is referred to as saline water in text). Pairs of communities were se- lected within each of the three water environments. Each pair included a site-A community and a site-B community. Households in the site-A communities had piped water service already avail- able. Within the site-A communities, two types of households were surveyed: those who had already decided to connect at the existing connection costs and tariffs and those who had decided not to connect. C onnectors were asked whether they would continue to connect, for a range of hypothetical tariffs higher than the current tariff. Nonconnectors were asked whether they would connect, for a range of hypothetical tariffs and connection costs. Households in the site-B communities had no piped system but could expect to have one installed soon. Families were asked whether they would connect for a variety of hypothetical connection costs and monthly charges. All households in both sites in each type of water environment were asked about their willing- ness to pay if the reliability of the water system were improved.' Thus, the im- 1. Reliability was characterized as follows: "Now I would like you to tell me what you would do if the service through the piped water system was greatly improved. Imagine that the water supply was available every day for most of the day, that the flow in the taps w.as alwavs good, and that the water was clean." 376 1HF WORLI) BANK ECONOMIC RFVIFW, VOL. 9, NO. pact on the choice of connecting of three characteristics of water systems- connection charge, tariff, and reliability-was tested through the contingent valuation method. In the original survey there were 1,150 households distributed approximately evenly across the three types of water environments, including the entire popu- lation of connectors in the three A sites, a sample of nonconnectors in the A sites, and a sample of potential connectors in the B sites. Table 1 provides some basic information about the survey sites and shows how willingness to pay var- ied across connectors and nonconnectors in the A sites and overall for the B sites. The exact questions posed during the interviews and the econometric issues that had to be resolved are discussed in detail in Singh and others (1993: 1,932- 35) and will not be reproduced here. Table 2 suLmmarizes how well the 1988 survey followed best practice in designing contingent valuation surveys, or what has become known as the "Seven Pillars of NOAA" (National Oceanic and At- mospheric Administration) (Portney 1994). These rules are designed to over- come some of the known problems with the technique. Generally, the survey Table 1. Location and Types of Survey Sites, with Sample Size and Maximum Willingness to Pay, Kerala, India A sites: B sites: improved water source available no improved water Site characteristic Connectors Nonconnectors source available Water abundant Location Ezhuvathuruthv Ezhuvathuruthy Nannamukku All households 66 819 1.497 Household sample 66 100 200 Water scarce Location Elapully Elapully Elapully All households 86 723 876 Household sample 86 100 200 Water abundant but witb. saline intrusion Location Ezhuvathuruthv Ezhuvathuruthy Vallikkunnu All households 98 768 1,313 Household sample 98 100 200 Total household sample 250 300 600 Average maximum monthly tariff bid (rupees) 19.3 8., 5.5 Average maximum connectioni charge bid (rupees) n.a. 355 267 Average maximum bid for improved service (rupees) 25.0 9.7 n.a. n.a. Not applicable because the bidding game was n1ot conducted in that site. Note: The exchange rate in 1988 was about 14 rupees per U.S. dollar. Source: Authors' calculations. Griffin and others 377 Table 2. Comparison of National Oceanic and Atmospheric Administration (NOAA) Guidelines for Contingent Valuation Surveys with Guidelines Used in the Kerala 1988 Surv'ey NOAA guidelines Procedu res used in the Kerala survey 1. Interview in person rather than Interviewed household head personally. on the telephone. 2. Question about a future, Survey asked about willingness to pay for a new or hypothetical occurrence rather improved water system or a change in tariff policy, than a historical event. not for an existing service. 3. Referendum format in which Bidding game format used: interviewer suggested respondent "votes" on a benefit prices for monthly tariff or connection cost; with a known price (as opposed respondenit aniswered yes or no to each quote. to open-ended questions). 4. Interviewer begins with a Interviewer described the exact nature of the good or scenario accurately describing service to be provided: existing quality of piped the benefit or program. water services into the home at various monthly tariffs and connection charges, plus improved quality of service at various monthly tariffs. 5. Survey reminds that payment for Survey had no reminders, hut questions about other the new benefit reduces other consumption and assets preceded the contingent consumption. valuation questions. 6. Survey reminds that substitutes In-depth questioning about cost, distance, and other exist for the hypothetical benefit characteristics of the household's sources of water in question. (and volume of consumption) preceded the contingent valuation questions. No specific reminder was given during the contingent valuation questions, but it was clear to the respondent that the context of the survey was general use of various sources of water. 7. Follow-up questions to make No follow-lip questions, but interviewer evaluated the sure respondent understands the quality of response. There was a follow-up survey choices made. to ascertaii actual behavior after the water svstem was put in place, and respondents were asked to explain divergence between hypothetical (1988) and actual behavior ( 1991). Source: Portney ( 1994) and authors' review of the Kerala survey and procedures. meets or exceeds the desired specifications, but the true test is whether the hy- pothetical responses adequately predict subsequent behavior. In the earlier analysis (Singh and others 1993), we estimated an econometric model that controlled for individual, household, and community characteristics that could affect the responses to the three policy variables in which we were interested: monthly tariff, connection cost, and reliability. The purpose of esti- mating the full model was to isolate the impact of the policy variables on the choice to connect and then to perform simulations of how changes in these 37S THE WORL[) BANK Ft ONOMIC REVIFW. VOL. 9, NO I policy variables would affect the number of water connections demanded and, consequently, consumer welfare. Table 3 contains the full model estimated in Singh and others (1993), along with coefficient estimates and t-statistics. Monthly tariff was an important de- terminant of whether respondents were willing to pay for yard connections. For each 1 percent increase in the monthly tariff, the probability of choosing a yard tap fell by almost 1.5 percent. However, connection cost was an even greater impediment limiting connections, probably because of the very high implicit interest rates prevailing. Whereas a ten-rupee increase in the mean monthly tar- iff quoted in the bidding game would cause connection probabilities to fall by approximately 27 percent, a ten-rupee increase in the implicit monthly connec- tion cost would cause connection probabilities to fall by approximately 82 percent.2 Another finding is that household characteristics matter. The decision to con- nect was positively and strongly affected by nonpolicy variables: higher levels of income, assets, and schooling. Water scarcity also matters. People in scarce- water areas were much more likely to connect, everything else being equal, than those in areas where water is abundant. Improved reliability turns out to be important only for current connectors. Those who were already connected indi- cated that they would be willing to pay substantially more for more-reliable service. More-reliable service did not, however, affect the probability that nonconnectors would decide to connect to the system. The findings indicate that large potential welfare gains would be generated by a more liberal connection and pricing policy. Because of the problem of con- nection costs, we estimated that consumer surplus could be increased by at least 450 percent by amortizing the connection cost and folding it into the monthly tariff. We estimated that connections in the population would rise from the current 250 to about 2,500 under this policy regime. 11. THE 1991 FOLLOW-UP SURVEY In the period following the 1988 contingent valuation study, improved water services were made available in the scarce- and saline-water areas. Site-B house- holds in these two areas had to decide whether to connect to the improved sys- tem. The original site-B families in these two water environments were resur- veyed in 1991 to determine whether they had connected. 2. lUnder reasonable assumptions about the interest rate and amortization schedule, the impact of the tariff and connection charges on the probability of connecting can be equalized, suggesting that respondents treated the tariff as a recurrent cost and the connection charge as a capital cost. The monthly charge for the connection cost in this example is calculated as if the connection cost were financed for six years at an annual real interest rate of 5 percent. A ten-rupee increase in the average monthly cost would thus correspond to an increase in the total coniection charge from the bidding game average of Rs219 to Rs842. In fact, rural Indians are likely to face much stiffer credit market conditions, so this example is for most people probably a suibstantial understatement otf the full connection costs when high interest rates are taken into account. Griffin and otbers 379 Table 3. Probit Estimates of the Probability of Choosinig a Yard Tap (1988 Sunrey) Variable Coefficient Constant -0.3009 (- 1.28) Tariff -0.0605 (-20.31) Connection charge -0.0010 (-13.21) Improved service -0.0582 (-1.11) DListance to Current source (meters) 0.00002 (0.06) Queue at current source (minutes) (0.0028 (1.54) Per capita inicome (rupees) 0.00002 (1.93) Household has electricity 0.3345 (3.83) Number of rooms in dwelling 0.0861 (3.50) Household has femiiales in government service -0.0997 (-0.57) Household has males in government service 0.1664 11.78) Hindu household -0.1908 (-2.07) Fenmale-headed household 0.0569 (0.66) Female respondenit -0.2749 (- 3.70) Some primary school 0.5092 (3.39) Primary school complete 0.6293 (4.45) Middle school complete 0.9608 (6.51) Secondarv school complete 1.1325 (8.09) More than secondary school 1.2898 (7.99) Scarce-water area 0.3474 (3.54) Saline-water area -0.2315 (-2.19) A-site noncoiiltecting household -0.3070 (-2.45) B-site village household -0.4921 (-3.87) Pseudo R2 0.28 Sarnple Size, 9,720 Households 1,150 X' (22) 3,272 Note: Dependent variable: whether respondent would choose to coinnect at each price quored. Estimates are weighted by the population of the sampling unit. Standard errors are corrected using a method explained in Singh and others ( 1993). The estimiating equation is signiificant at better than the 0.00001 Ilevel for a likelihood ratio test (X') The (lmitted site dumniv is A-site coinecting household. t-staristics are in parentheses. a. The sample size of 9,720 results from multiple observationis for each household in the sample. See Singh and others (1993 for a complete explanation. Source: Authors' calculations. .380 I Ii-F. WORLD) BANK F( ONOWIR RFVIFI', VOl. 49 No.; Data Issues In the 1991 follow-up survey, an attempt was made to contact all of the respondents in the original sample in the two villages. The first issue to address in any follow-up survey is sample attrition. We lost 25 of 200 households (12.5 percent) from the sample in the scarce-water site and 59 of 200 households in the saline-water site (29.5 percent) hetween 1988 and 1991. The loss of house- holds appears to have been random in the scarce-water site in that it left the original income distribution virtually intact. In the saline-water site, attrition was concentrated among the poorer households.l "Connectors" in this analysis are defined as those who either were already connected to the new system at the time of the follow-up survey, had applied for a connection, or had made a decision to connect once applications were called for by the water authority. Both connectors and noniconinectors were asked a short series of questions convcerning their decision and concerining exogenous changes that had come about since the original survey. We predicted that a household would coninect if the maximum bid for con- nection in 1988 was higher than the actual connection cost in 1991. If the house- hold was connected at the time of the interview in 1 991, the respondent was asked how much the connection actually cost. If the household was not con- nected, the interviewer estimated the cost of coninecting for that household on the basis of its distance from the distribution line.4 iludgiiig tbh.e Vahli(ditv oft Predictions What constitutes a success or failure in niaking a prediction based on the 1988 survey? In this section we lay out the criteria to he used in judging the success of the experiment. The first and most important criterionl, as in anyv sampling procedure, is whether we predict the correct proportion of coniectors, notwithstanding whether we predict the exact behavior of each household. We would like to get as close as possible to the correct proportion. If we err, we prefer to err on the side of underpredictling the proportion of coninectors because the concern with the di- rect approach is that its inherent biases tend toward overvaluationi. Furthier- more, we would not want to recommend installation of a water system on the basis of estimated demand and revenuc forecasts that are too optimistic. The second criterion, aimed directly at the heart of evaluating the technique, is whether specific households behave as they said they would. There are three elements here. The first is gross accuracy, or the proportioni of the surviving sample for which actual behavior was correctly predicted. The second is the I. The exact c omparison of the two samples is asailahle fritni the authors. 4. The coninecrioi co st in 1991 was deflated to I 9S8 ripees so that the nililmhers Could he compared in real terms. We used the average inflationi rate fronI 1980(-9() as reported in World Batik ( 1992 7'.9 perceit. This lonig-rermii average rate Was ised to aa olid Leat -to-vear flutUiationis for this investnenst and in recol)gitioii of the illirrecisioll lt o tfinttal estimatecs Griffin and others 381 specificity of connector predictions or the proportion of connecting households predicted correctly. The third is the sensitivity of predictions for nonconnectors or the proportion of nonconnecting households predicted correctly. The statistical epidemiological literature distinguishes between specificity and sensitivity (see Kleinbaum, Kupper, and Morgenstern 1982). Specificity has to do with the problem of false positives, that is, the number of families that were predicted to connect but that actually did not conniect. Sensitivity has to do with the problem of false negatives, that is, the number of families that were pre- dicted not to connect hut that actually did conniiect. Comparing Benzefit Revelation, Benefit Transt er, a7nd a Behavioral Model We apply the above procedures to assess whether the behavior of families at the B sites could best be predicted by benefit-revelation or benefit-transfer meth- ods. There are, accordingly, two research questions. First, did people at the B sites behave as they said they would? We call this the benefit-revelation ques- tion. Second, did people at the B sites behave as we had predicted on the basis of the predictions of a behavioral model for the A sites (after substituting the char- acteristics of households in the B sites)? We call this the benefit-transfer ques- tion. Prima facie, it is not obvious which of the two strategies implicit in the above questions would be most promisilng. To partially assess the validity of the inferences drawn from the behavioral model (Table 3) estimated using the 1988 data (Singh and others 1993), we also compare predictions from it with actual behavior. Is the actual behavior of the families in the B sites in 1 99 1 predicted accurately by a behavioral model based on the responses of those families in 1988? We call this the behavioral modeling qLiestion. Ill. RESUlTS FOR THF SC.ARCE-WATER SITE This section analyzes the results for the scarce-water site. The results are also robust for the saline-water site, but we do not place equivalent weight on these findings, for reasons to he explained in the next section. Benefit Revelatio7n The fundamental quiestion that the follow-up survey was designed to address is the simplest one: did people behave as they said they would? It was necessary to exclude those families for whom predictions could not be made because they bid at the maximum (Rs700) in 1988, and their actual connection costs in 1991 (which varied from house to house, depending on the distance to the public water line) exceeded this maximum bid.i This situation applied to 13 of the 161 5. For this group, we do nor really observe a maxinmumi bid because we know onily that it was at least Rs7 O{. T hus we cannotr iiake i comparisonl with the coninlectioll w.ost faced in 1991. 382 THI. WORI D BANK l-.(:()NOMI(: REVIFW, VOL.. 9, NI. I Table 4. Comparisoni of Actual and Predicted Behavior of B-Site Households in Water-Scarce Areas (number of households) Actual bebavior Predicted behavior Did connect Did not connect Total Benefit rei elation, Will connect I * 6 21 Will not coninect 7 120 127 Total 22 126 148 Bene/ft tralnsfer, all A sites' Will connect 27 76 103 Will not connect 1 65 66 Total 28 141 169 Benefit transfer, scarce-water A site Will connect 28 10N 128 Will not connect () 41 41 Total 28 141 169 Behavioral modelingAI Will connect 10 13 23 Will not connect 18 128 146 Total 28 141 169 a. Prediction based on 1988 survey of B-site households in the scarce-water area. b. Prediction based Oin 1988 survey of A-site households in all three water areas. c. Prediction based oIi 1988 survey of A-site households in the scarce-water area. d. Predictioin based on probit model of 1988 B-site households in the scarce-water area. Source: Author-s' calculations. families (7 of which actually connected).' The results for the remaining 148 respondents are presented in table 4 (under "benefit revelation"). The results indicate that 14.9 percent of the respondents (22/148) did con- nect. This is not statistically different (at the 5 percent significance level) from our prediction that 14.2 percent (21/148) of the families would connect. The gross accuracy of the predictions based on benefit revelation was 91 percent, as shown in table 5. That is, the behavior of 91 percent [(15 + 120)/1481 of the families was conisistent with the intentions they declared in the 1988 contingent valuation survey. The specificity of connector predictions, the percentage of those predicted to connect who actually did connect, was 71 percent (15/21). The sensitivity of nonconnector predictions, the percentage of those predicted not to connect who actually did not, was 94 percent (120/127). Thus, simple tabulations of respondents' answers to the contingent valuation survey in 1988 were remarkably accurate in predicting both the overall proportion of the 6. After losinig 25 households because ofz attritioni, we lost another 12 households that had missing values for connection cost in 1991 and 2 with missing values for the 1988 contingent valuation questionis, which left us with 161 households for analysis in the 1991 sample. (;riffin and otbers .3 83 Table 5. Sutinmarv Statistics on1 the Accuracy ot Predictions of Behavior of B-Site Households in Water-Scarce Areas (percentage) Sample statistic Valuse Benefit revelation' Gross accuracy 91 Specificity of coniiector prediction 71 Sensitivity of predictioni for nonconinectors 94 Benefit transter, all A sites' Gross accuracy 54 Specificity of coninector prediction 26 Sensitivity of prediction for nonconnectors 98 Benefit transfer, scaree-w'ater A site, Gross accuracv 41 Specificity of connector prediction 22 Sensitivity of prediction for nonconinectors 10() Behavioral nodLleli ng'" Gross accuracy 82 Specificity of coninector prediction 43 Sensitivity of predictioni for nonconnectors 88 Note: Gross accuracy is the percentage of the surviving sample for which actual behavior was correctly predicted. Specificity of connector prediction is the percenitage of those predicted to connect and who actually dlid. The sensitivity of prediction for nonconnectors is the percentage of those predicted not to connect who actually did [lot. a. Predictioin based oin 1988 survev of B-site households ill the scarce-vater area. b. Predictioll based on 1988 survey of A-site households in all three water areas. c. Prediction based oni 1988 survey of A-site households in the scarce-water area. d. Predictionu based on probit model of 1 988 B-site households in the scarce-water area. Soutrce: Authors' calculations. sample connecting and the household-specific choices once the households were given the opportullity of connecting to the piped water supply system in 1991. In addition to determining whether families connected or not, the follow-up survey asked why. A variety of responses were given during these open-ended discussions. Only one emerged with any consistency-three of the thirteen re- spondents whose behavior was inconsistent with their 1988 response cited "changed economic circumstances." Two of the seven "unpredicted connec- tors" fell into this group, as did one of the six "unpredicted nonconnectors." For the nonconnectors, more than 75 percent indicated that "inability to pay the connection cost" was the primary reason for not conniecting to the system, just as we had predicted from the analvsis of the 1988 data. Did reliability affect connection decisions in the sample in 1991? We pre- dicted, on the basis of the 1988 results, that households already connected were concerned with reliability, but it was not an importanit consideration for those who were not already coninected to a water system. The results of the ex post .384 THiF WORI 1) IANK F(:(ONt)MIC RFVIFW, V(l. 9. NO. 1 investigations confirm this finding. Only a small proportion of nonconnectors (13 percent) replied that an inadequate quantity of water from the system was the main reason for not connecting, and no respondents mentioned service qual- ity as a decisive reason for not connecting. However, those who had connected in 1991 unanimously expressed dissatisfaction with the reliability of the system. Every connector complained about the quantity of water available from the system during the dry season, and all but one of the connectors found the quan- tity inadequate even in the monsoon season. Benefit Transfer Many reasonable people are justifiably suspicious of answers to hypothetical questions by subjects who have a strong interest in the outcome of the research. Extrapolation based on information gathered from a similar group that is not subject to these problems may reduce these biases and thus be a more reason- able basis for predicting actual behavior. The strategy of benefit transfer de- pends on the validity of models that can allow Us to extrapolate from behavior or valuation of benefits in one area to populations of known characteristics in other areas. This method is still in its infancy (Pearce 1994; Desvousges, Naughton, and Parsons 1992; and Boyle and Bergstrom 1992). In the United States there has been only one legal proceeding in which contingent valuation estimates from one study were used to estimate values in another site, and the court refused to accept the "transferred values" as legitimate evidence (Brookshire and Neill 1992). Three guidelines have been developed for use in research on benefit transfer. First, the study site should be very similar to the policy site. Second, the policy change or project at the study site should he very similar to that proposed at the policy site. And third, the valuationi procedures used at the study site should be analytically sound and carefully coniducted (Pearce 1994). The Kerala study fol- lows these guidelines exactly. To address the benefit-transfer research question, the contingent valua- tion survey of households at the A sites was used to estimate a model ex- plaining the probability of connection for them. As described in Singh and others (1993), respondents' choices were modeled using a random utility framework, in which the probability that a family chooses to conlnect is de- termined by household characteristics, characteristics of the irnproved sup- ply system, and characteristics of alternative water supplies. Table 3 con- tains estimates for the full model using all observations for all A and B sites; appendix table A-I contains analogous estimates for the three subsamples used in this section: all A sites, the scarce-water A site, and the scarce-water B site. Under reasonable assumptions on the distribution of errors, probit estimates are unbiased. They are not efficient, however, because each house- hold recorded repeated bids. The standard errors have been corrected for within-household correlations among groups of observations (described in detail in Singh and others 1 993). Griffln and others 385 Before proceeding, it is important to explain what these estimating equations do and do not do. The coefficients are consistent estimates of the true effect in the population of each variable on the decision to hook up. Given the high t-statistics for the important variables, we can be fairly certain that if other samples were drawn from these populations, we would get roughly the same coefficient estimates. However, the pseudo R2 statistics at the bottom of tables 3 and A-I indicate that only about one-third of the variation in the dependent variable-whether to connect-is explained by the model. Although this statis- tic is relatively high for cross-sectional models, it suggests that the specified economic and social variables determine only a limited fraction of behavior. Using this model for one site to predict responses at another site might be problematic for several reasons. For example, some variables might be out of range. The coefficients are valid only for the range of variables occurring in the sample observed in the first site. Another possibility is that the predictable com- ponent of behavior may be overwhelmed by unobservable and site-specific ef- fects and random effects. Still another possibility is that comparison across time periods without new information might be problematic, in tlhat some of the observed variables in the equationi may have changed, but the changes cannot be observed. One example is income. In predicting behavior in 1991, we bumped up every houselhold's income by the real growth rate of the economy, but doing so fails to capture relative chianges in real income across the households. Using models of behavior at the A sites to predict behavior at the scarce- water B site gives completely inaccurate predictions. This is true whether house- holds in all the A sites or only the households in the scarce-water A site are used (table 4). We assunme that the results using the scarce-water A site would be the most accurate, as this site is the closest sample to the scarce-water B site. Using the benefit-transfer approach, we predicted that a very high proportion of the B- site households would connect-76 percent (I 28/169). From table 4, only 16.6 percent (28/169) of the families connected. The difference is statistically signifi- cant. Gross accuracy is 41 percenit, specificity of connectol predictions is 22 percent, and sensitivity of noniconinlector predictions is 100 percent. In short, the performance of the benefit-revelation method (based on the responses to coIn- tingenit valuation questions at the B site) is vastly superior to that of the benefit- transfer method (based on modelling behavior at the A site and transferring the results to the B site). We do not have to look far for reasonis why the benefit-transfer approach works so poorly in this case. The basic reason is not model specification (in the next subsection we demolIstrate that Lising data from the B site to estimate the coefficients in the model tendis to overpredict conntectors slightly, but otherwise performs well in explaininig behavior), but the mucil lower bids in the B site in 1988. The descriptive statistics at the bottom of table I show that the average connection bid in the B sites in 1988 (Rs267) WalS o111n 75 percenit of the average for the nonconinlectilng A sites (Rs35S 5). The aver-age maximlulIL bid for the monthly tariff in the B sites was Rs._S, onily 28 percenit of the average for the connecting )86 1111i WO)Ri I) lANK Is (I)NOMIt( RI VII W, Vol I., N() 1 A sites (Rs19.3) and 63 percent of the average for the nonconnecting A sites (Rs8.7). The coefficients for the dummy variable for the villages in the B sites in the full model (table 3) are negative and highily significant. In other words, the households in the B sites placed a relatively low value on the new water system in 1988, and they were telling the truth about it: their responses in 1988 pre- dicted their own behavior very well. The respondents in the A sites turned out to be very pgor substitutes for the respondents in the B sites, apparently because of unobservable, site-specific factors. Behav,ioral Modielinig As described above, the benefit-revelation method gave reliable predictions of actual behavior at the B site. Although this is an important finding, in prac- tice it provides an answer to only part of the question of interest to researchers and policymakers, who are also concerned with likely responses to changes in a cluster of related policy variables (such as tariffs, connection charges, and reli- ability) while controlling for other confounding variables (such as education and income). The simple tabular analysis of the contingent valuation data from the B site cannot, of course, answer such questions. The numbers in table 4 were generated by predicting behavior in the scarce- water B site using only the data from that sample and the model appearing in the third column of appendix table A- 1 In table 4, the sample size is increased to 169.' We had predicted that 13.6 percent (23/169) of families would connect, whereas in fact 16.6 percent (28/169) of the larger sample of families did con- nect. Thus, the model meets our criterion of being conservative, that is, it under- estimates connectionis. The difference is not statistically significant at the 5 per- cent level. The gross accuracy was 82 percent, compared with 91 percent for the simple analysis. The specificity of coniector predictions fell from 69 percent in the simple case to 43 percent. The sensitivitv of nonconnector predictions was 88 percent, down from 95 percent in the case of the simple statistics. The behavioral model estimated using the contingent valuation data from the B site does not predict behavior quite as well as the simple descriptive statistics do. Because the behavioral model should give us more information rather than less, this result suggests that fulictionlal form and distributional assumptions 7. We also performed the anralsis summnarized in table 4 (unider "behavioral modeling ) using the whole-sarnple estimates presented in table 3. The results are almost identical. The counterpart to table 4 is available fromii the aLuthors. 8. The change in samilple size fromii 148 to 169 cai be explainied as follows, we started with 148 in table 4 (uilnder benefit revelation , and then we gainied I, that had been dropped because their bids were at the maximnum. For these households, we could predict the probahilitv that thev would connect because of the properties of the estimatinig equation. We gained another 12 houselholds, for which the willinginess-to- pay data were inissiiig in 1988, but for which we knew whether thev coniniected in 1991. For these households we could predict whether thex would cionnect usin1g the mTodel. We lost 4 households with a missing value for ot c or more indepenidenit variables Ill the equ1ationis, even though we knew the outconme variables for thern. Griffin and others 387 may be more restrictive than we expected. But we doubt that conclusion and attribute the lower accuracy to the fact that the model allowed us to include those observations at the maximum bid that had to be dropped for table 4 (un- der "benefit revelation"). These households introduced virtually all the addi- tional error. If we use behavioral modeling to predict for the same sample that appears in table 4 (under "benefit revelation"), we get similar results to those in that table (results not reported here). Thus, inferences from the econometric model probably provide good, conser- vative guidance for water policy. Inferences from the model about elasticities (with respect to price, income, and reliability, for instance) and changes in wel- fare are likely to be reliable. This result suggests that the strong policy conclu- sions of the initial study (Singh and others 1993) should be taken quite seriously. That article suggested that substantial improvements in welfare would be pos- sible with drastic reform of water sector policies in Kerala. Our results do sug- gest, however, that considerable care should be taken to make sure that the range of responses to the contingent valuation questions do not artificially cen- sor those who might bid very high or very low. In the simple benefit-revelation case, the censoring requires us to throw out observations; in the behavioral model case, it reduces the predictive accuracy of the model, even though we can pre- dict values for those observations. IV. RESUITS FOR THE SALINE-WATER SITE We will only briefly present the results for the saline-water site because the results buttress the findings from the scarce-water site but are less interesting and less reliable.' As already noted, sample attrition was much greater in the saline-water site and was concentrated among the poorer households (we lost 30 percent of the sample, compared with only 13 percent for the scarce-water site). In the saline-water site, we lost twenty-one households for the benefit- revelation exercise because they were at the maximum bid for connection cost. Nineteen of these did not connect, and two did. We lost an additional three households with a missing value for connectioni cost in 1991. The results for the saline-water site are shown in table 6. We predicted that 0 percent (0/117) of families would connect, whereas in fact 15.4 percent (18/117) of families did connect. So, although we were infinitely wrong about the proportion connecting, we were at least wrong in the desired direction, underpredicting connectors. Our failure to predict connectors appears to have been caused by exogenous changes that took place at the saline-water site between surveys. During the 9. We do not include here thie additional analyses performed for the scarce-water site because they do not add new information. The results are basically the same as those for the scarce-water site, with the benefit-transfer model performing slightly better than for the scarce-water site (but still badly overpredicting connections) and the behavioral model basically replicating the benefit-revelation results with less accuracy. These results are available fromi the authors on request. 388 THE WORLD BANK ECONOMIC REVIFW, VOL. 9, NO. .3 Table 6. Comparison of Predicted Behavior of B-Site, Saline Water Area Households in 1988 and Actual Behavior in 1991 (number of households) Actual behavior5 Predicted behav'ior' Did connect Did not connect Total Will connect 0 0 0 Will not connect 18 99 117 Total 18 99 117 a. Based on the maximum bid that households were willing to pay for a connection in 1988. b. Based on the actual number of connections after true cost was known. Source: Authors' calculations. follow-up survey, 78 percent of the households in this area that did connect, despite our predicting that they would not, cited "unanticipated improved eco- nomic circumstances" as the reason for connecting. Another 22 percent said it was because they could borrow money for the connection (in contrast, no one in the scarce-water site borrowed to finance the connection). The gross accuracy of predictions based on the intentions households declared in the 1988 contingent valuation survey was 85 percent. Thus, even though we predicted no connectors, because the proportion connecting was small, we did very well in predicting behavior for the whole sample. The specificity of connec- tor predictions was 100 percent. We were by definition perfectly correct in get- ting our predicted connectors right, because we predicted none. The sensitivity of nonconnector predictions was 85 percent. The results for the saline water site show the fragility of our endeavor to compare bids with behavior three years later, with almost no information about what happened in between. Even so, the results are fairly robust, despite our misgivings even about the 1988 data in the saline-water site. For example, analysis of the 1988 data suggested that a family in a saline-water area was willing to pay less for an improved water supply than a similar family in a water-rich area (see table 3), which was counterintuitive. But more important, between the first and second rounds of the surveys, there were major exogenous changes in the saline-water villages. V. CONCLUSIONS Contingent valuation studies suffer from three potential sources of bias (Cummings, Brookshire, and Schultze 1986). First, strategic bias might arise because respondents perceive it to be in their interest to respond inaccurately to the questions. Second, hypothetical bias might arise because respondents are not fully acquainted with the good or service in question. And third, compliance bias might arise because respondents give replies they believe the questioners would find most satisfactory. A priori, in this particular setting, it might be predicted that strategic biases would be relatively high, with respondents under- estimating their willingness to pay in the hope that this might lower the tariffs Griffin and others 389 that would be charged. Hypothetical biases would be relatively low because the good in question-a piped water supply-is familiar to all. Compliance biases might be relatively high, because this has happened with surveys in the Indian subcontinent in the past (for example, Mamdani 1972). The comparison of actual behavior with that emanating from the contingent valuation survey amounts to a resounding confirmation that, in this particular study, the net effect of these biases was small. The caveat regarding "this par- ticular study" cannot be overstressed for several reasons. First is the issue of strategic and compliance bias. Recognizing the potential for these sources of error, the study made strenuous efforts to reduce to a minimum the effects of these biases: by explaining procedures in the survey; by being-and being seen as-independent of the supply agency; and by training interviewers rigorously. There is considerable evidence that "quick and dirty" willingness-to-pay sur- veys of a similar nature in the past have yielded nonsensical results (Saunders and Warford 1972). Accordingly, the survey instruments were developed with great care in the course of a multicountry study (World Bank Water Demand Research Team 1993) and were alreadv "tried and tested" by the time of the 1988 survey in Kerala. As shown in table 2, the instrument met the NOAA stan- dards five years before the standards were developed. Furthermore, the instru- ment was carefully pretested in Kerala, and modifications were made so that it was appropriate for the local setting. Second is the issue of hypothetical bias. Proponents of the contingent valua- tion methodology have understood for some time that the greatest problem for contingent valuation studies arises not from strategic but from hypothetical bias (see Cummings, Brookshire, and Schultze 1986). Stimulated by the controversy on the damage caused by the Exxon Valdez oil spill in Alaska, a rigorous and heated debate on the contingent valuation methodology has taken place, with much of the attention focused on the issue of hypothetical bias (see Portney 1994; Hanneman 1994; and Diamond and Hausman 1994). Diamond and Hausman (1994: 62) have made a forceful denunciation of the method, arguing that "contingent valuation is a deeply flawed methodology for measuring nonuse values." With respect to this controversy, the Kerala water study simultaneously (a) provides clear evidence that carefully conducted contingent valuation stud- ies can provide reliable information on how people value well-defined goods and services and (apparently paradoxicallv) (b) does not contradict the concerns that underlie the Diamond-Hausman argument. The key issue here is that, even though the good at stake in the Kerala study was tangible and simple, there was still a problem with hypothetical bias in parts of the study. This problem emerged in the finding that an important ser- vice characteristic-reliability-was of major importance to those who were already connected, but was not perceived as being important by those who had not directly experienced the service. Well-conducted contingent valuation stud- ies can provide reliable and valuable information on behavioral responses to well-defined and well-understood goods such as a household water supply. But 390 THE WORLD BANK F(.ONOMIC REVIEW. VOL. 9, NO. 3 this finding in no way vitiates the very serious problems arising when this method is used to assess such abstract concepts as "existence values." In the words of Diamond and Hausman (1994: 62), "it is precisely the lack of experience both in markets for environmental commodities and in the consequences of such de- cisions that makes contingent valuation questions so hard to answer and the responses so suspect." Benefit-Transfer Literature The results are equally striking for the prospects of using the estimates based on a behavioral model for one population to predict the behavior of another population. Virtually all the characteristics of the study population and the al- ternative water sources are apparently quite similar at the A and B sites. When behavior at the scarce-water B site was estimated from a well-specified and carefully estimated model of actual behavior at the scarce-water A site, how- ever, predictions were wrong for about half of the sample. The number of con- nections was overestimated by a factor of four. This finding is not surprising, for two reasons. First, even when the determi- nants of behavior are easy to specify (as in this case), detailed models of this behavior are formulated, the full required set of data is collected, and sophisti- cated statistical tools are applied, less than one-third of variance can be ex- plained (see table 3). Second, the results of a multicountry study of willingness to pay for water (World Bank Water Demand Research Team 1993) shows that both appropriate specifications and parameter estimates vary considerably in different locations. Attempts to estimate behavior (and thus benefits) in a particular community on the basis of results of studies in other communities in other settings can reach conclusions that are seriously erroneous. This can occur even when the commu- nities, the natural conditions, and the service to be offered are apparently quite similar. Substantial additional information is collected when the expected ben- eficiaries are interviewed directly. Assessing the Demand for Services througb Bebavioral Models Benefit revelation through direct methods has great potential for assessing the demand for services, especially for capital-intensive and costly services such as water supply in developing countries. Carefully designed and conducted con- tingent valuation studies can produce reliable estimates of the demand for water and sanitation and are appropriately becoming widely used for this vital func- tion (Whittington and others 1992; Altaf, Jamal, and Whittington 1992). The sample size requirements are modest (a couple hundred families sufficed in this case). As experience with these studies accumulates, it has been possible to substantially improve quality, increase speed, and reduce cost by the judi- cious use of off-the-shelf survey components. If the policy interest is limited (in the present case, to the number of families who would connect to a new supply), then simple tabular analyses may suffice. If the policy interest is more complex Griffin and others 391 (for example, elasticities with respect to price and service reliability; simulations of policy changes; or welfare analysis), then behavioral models using economet- ric techniques need to be estimated. Caveats The contingent valuation method is validated under a very specific set of circumstances. Of particular importance (as stressed earlier) is the fact that hy- pothetical biases for the service evaluated in this study are much more limited than for many of the environmental resources that the technique is typically used to value. Of equal importance is the careful design and conduct of these studies. In all cases this meant several weeks of pretesting and adaptation to local circumstances, meticulous training and supervision of interviewers, and careful cleaning of data. More specifically, it is noted with concern that the relative success of the set of studies of which the Kerala one is part (World Bank Water Demand Research Team 1993) has inspired some investigators to con- duct two-day studies of hundreds of households to determine willingness to pay. In the past such studies gave willingness-to-pay surveys a (well-deserved) bad name. The old adage of "garbage in, garbage out" is certainly applicable to such poorly conducted studies. In some circumstances, researchers and policy analysts will have access to rich data sets on populations that (a) already have access to the service of inter- est and (b) are very similar to the population for which the service is to be introduced. In such circumstances, it would appear that carefully specified and estimated models of behavior could be used to predict behavior by the unserved population. The results of this study, however, show that site-specific factors are of major importance and that predictions based on extrapolations may be far off the mark, even when many conditions in the population and environment are apparently similar. What We Learned That Can Improve Survey Design Much more care should be exercised in defining the range and the increments for bidding games or referenda. They should be connected as closely as possible to actual costs, or the ranges should be pretested for validity. Often researchers pick very high willingness-to-pay bids for the top of the range, which they think are higher than anvone would pay, vet when the data come back they show that large proportions of the sample have chosen the higlhest bid. That does not give us good information about willingness to pay. In our case, despite efforts to design the best possible survey in 1988, the top of the ranges for tariffs and connection charges was too low. For example, in the scarce-water site, one con- nector who bid at the maximum of Rs700 in 1988 actually paid Rs2,547 for the connection. In this case, we could have avoided the problem by reviewing the range of actual connection costs in 1988 before finalizing the survey. Quick analysis of the survey data immediately after the fieldwork, with a plan to return to the field for additional work, would resolve many questions 392 THE WORLD BANK FC(ONO.11( RFVIF:W, VOL. 9. No. 3 that we had in the analysis. In the saline-water site, for instance, we should have caught right away (in 1988) that the bidding behavior was not as we expected, and households should have been reinterviewed. After the 1991 survey, we should have immediately gone back to understand better why six households that had bid RslO0 in 1988 for connections actually hooked up at an average cost of Rsl,380. Doing the analysis quickly in the field, preparing a qualitative ques- tionnaire to complement the quantitative work, and aggressively addressing is- sues of data quality could raise the validity of benefit-revelation techniques con- siderably and increase their value for assessing demand. If other researchers conduct the same type of test reported in this article, we suggest collecting additional information about the households over the elapsed time between the hypothetical bids and the actual choice. Repeated observa- tions on income, assets, family size, community characteristics, precipitation, and traditional water quality would help control for endogenous and exogenous changes over the period. Implications for Kerala To the extent that contingent valuation, or benefit revelation, has been vali- dated by this exercise, it has been driven by the nonconnectors. We predicted that very few would connect to the water systems, given the policies in place. We were right. Standard water systems as designed in India and in much of the developing world do not make people better off. People respond by letting the systems fall apart or by telling us, as they did in the 1988 survey, that they would not connect even at the fairly low prices that were quoted. Their behavior in 1991 just turned the hypothetical rejection of the system by the vast majority of the sample into an actual rejection. We also predicted, on the basis of the 1988 data, that once people connected, they would become concerned about the poor quality and low reliability of the system. They would not care much about these problems unless they connected. We were right again. Although we would not have been able to make these predictions without a survey and a technique that worked as expected, we still do not know if we were right about the prescriptions we offered for solving these problems. On the basis of simulations, we recommended that the water authority raise tariffs, fold the connection cost into the tariff, and vastly expand access to private connections. We predicted that both connections and revenue would explode. People would consider themselves much better off. They would begin agitating to pay more to create a more reliable system. Making these policy changes, and tracking the results, would be the true test of the work. We know the technique worked. We do not know yet if the economic analysis was accurate, because the changes in policies that we recommended as a result of the analysis have not been tested. Griffin and others 393 Table A-1. 1988 Survey: Probit Estimates of Probability of Choosing a Yard Tap for Subsamples Variable All A sites Scarce-water A sites Scarce-water B sites Constant -0.3302 0.6083 -0.1693 (-0.877) (1.00) (-0.47) Tariff -0.0520 -0.0461 -0.0947 (-15.57) (-10.54) (-6.08) Connection charge -0.0010 -0.0008 -0.0017 (-8.32) (-3.76) (-8.49) Improved service -0.1394 0.0678 (-3.15) (0.97) Distance to current source 0.0103 0.0120 0.0013 (meters) (0.75) (1.31) (2.83) Queue at current source 0.0013 0.0067 -0.0076 (minutes) (0.64) (0.82) (-1.33) Per capita income (rupees) 0.00004 0.00005 -0.000001 (1.99) (1.05) (-0.04) Household has electricitv 0.0024 0.3689 0.2937 (0.02) (1.29) (1.35) Number of rooms in dwelling 0.0441 0.0721 0.2677 (1.53) (1.33) (2.75) Household has females in -0.1894 0.1652 -0.3037 government service (-0.67) (0.48) (-0.41) Household has males in -0.1495 -0.0694 0.0741 government service (-1.15) (-0.30) (0.32) Hindu household -0.1138 -(.9599 -0.3984 (-0.89) (-3.24) (-2.17) Female-headed household 0.0153 0.5330 -0.1017 (0.12) (2.26) (-0.51) Female respondent -0.3183 -0.3582 0.1192 (-2.68) (-1.66) (0.71) Some primary school 0.7960 1.2941 0.3312 (2.30) (2.47) (1.47) Primary school complete 1.1315 1.3099 0.2859 (3.74) (3.44) (1.22) Middle school complete 1.0451 0.3247 0.3709 (3.13) (0.73) (1.41) Secondarv school complete 1.4341 1.7492 0.6238 (4.77) (4.60) (2.63) More than secondary school 1.7274 2.0476 0.4956 (5.30) (4.53) (1.36) Scarce-water area 0.6983 n.a. na. (4.41) Saline-water area -0.1446 n.a. n.a. (-0.91) A-site noniconnecting household -0.4557 -1.1346 n.a. (-3.14) (-5.0()) B-site village household n.a. n.a. n.a. Pseudo R2 0.26 0.30 0.34 Sample size u,228 1,700 1,416 X2 (degrees of freedom) 1,704(21) 697(19) 554(17) n.a. Variable not appropriate to the sample. Note: Dependent variable: choice to connect in the bidding game. Estimates are weighted by the population of the sampling unit. Standard errors are corrected using a method explained in Singh and others (1993). The estimating equations are significant at better than the 0.00001 level for a likelihood ratio test (X). The omitted site dummy is A-site connecting households. t-statistics are in parentheses. Source: Authors' calculations. 394 THF WORIL) B'ANK F) oNoMl( REVbFW, Vol. 9, No. 3 REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Altaf, Mir Anjum, Haroon Jamal, and Dale Whittington. 1992. "Willingness to Pay for Water in Rural Punjab, Pakistan." Water and Sanitation Report 4. World Bank, Transport, Water, and Urban Development Department, UNDP-World Bank Water and Sanitation Program, Washington, D.C. Processed. Arrow, Kennetlh, Robert Solow, Paul Portney, Edward Leamer, Roy Radner, and Howard Schuman. 1993. "Report of the NOAA Panel on Contingent Valuation." Department of the Interior, Washington, D.C. Processed. Boyle, K. J., and J. C. Bergstrom. 1992. "Benefit Transfer Studies: Myths, Pragmatism, and Idealism." Water Resources Research 28(3):657-63. Briscoe, John, Paulo F. de Castro, Charles Griffin, James North, and Orjan Olsen. 1990. "Towards Equitable and Sustainable Rural Water Supplies: A Contingent Valuation Study in Brazil." The World Bank Economic Review 4(2):115-34. Brookshire, David S., Mark A. Thayer, William D. Schultze, and Ralph D'Arge. 1982. "Valuing Public Goods: A Comparison of Survey and Hedonic Approaches." Ameri- can Economic Revieuw 72(1):165-77. Brookshire, D)avid S., and H. R. Neill. 1992. "Benefit Transfers: Conceptual Issues and Empirical Issues." Water Resources Research 28(3):651-56. Carson, Richard T., and Robert Cameron Mitchell. 1993. "The Value of Clean Water: The Public's Willingness to Pay for Boatable, Fishable, and Swimmable Quality Water." Water Resources Research 29()7):2445-54. Cummings, Ronald G., David S. Brookshire, and William D. Schultze. 1986. Valuing Environmental Goods: An Assessment of the Contingent Valuation Method. Totowa, N.J.: Rowman and Allanheld. Desvousges, W. H., M. C. Naughton, and G. R. Parsons. 1992. "Benefit Transfer: Conceptual Problems in Estimatinig Water Qualitv Benefits Using Existing Studies." Water Resources Research 28(3):675-83. Diamond, Peter A., and Jerry A. Hausman. 1994. "Contingent Valuation: Is Some Number Better than No Number?" Journal ol Economic Perspectives 8(4):45-64. Hanneman, W. Michael. 1994. "Valuilng the Environment through Contingent Valua- tion." Journal of Economic Perspectives 8(4):19-43. Kleinbaum, D. G., L. L. Kupper, and Hal Morgenstern. 1982. Epidemiological Meth- ods: Principles and Quantitative Methods. Belmont, Calif.: Wadsworth. Mamdani, Mahmood. 1972. The Myth of Population Control: Family, Caste, and Class in an Indian Village. New York: Monthly Review Press. Mitchell, Robert Cameron, and Richard T. Carson. 1989. Using Surveys to Value Pub- lic Goods: The Contingent Valuation Method. Washington, D.C.: Resources for the Future. North, James H., and Charles C. Griffin. 1993. "Water Source as a Housing Character- istic: Hedonic Property Valuation and Willingness to Pay for Water." Water Re- sources Research 29(7):1923-29. Pearce, David W. 1993. Economic Values and the Natural World. London: Earthscan Publications. Griffin and others 395 1994. Project and Policy Appraisal: Integrating Economics and Environment. Paris: Organisation for Economic Co-operation and Development (OECD). Portney, Paul R. 1994. "The Contingent Valuation Debate: Why Economists Should Care." Journal of Economic Perspectives 8(4):3-1 7. Saunders, Robert J., and Jeremy J. Warford. 1972. Village Water Supply: Economics and Policy in the Developing World. Baltimore: Johns Hopkins University Press. Serageldin, Ismael, and Andrew Steer, eds. 1994. Valuing the Environment: Proceed- ings of the first Annual International Conference on Environmentally Sustainable Development. Washington, D.C.: World Bank. Singh, Bhanwar, Radhika Ramasubban, Ramesh Bhatia, John Briscoe, Charles C. Griffin, and Chongchuni Kim. 1993. "Rural Water Supply in Kerala, India: How to Emerge from a Low-Level Equilibriunm Trap." Water Resources Research 29(7, July): 1931-42. Whittington, Dale, Donald T. Lauria, Albert M. Wright, Kyeongae Choe, Jeffrey Hughes, and Venkateswarlu Swarna. 1992. "Household Demand for Improved Sanitation Services: A Case Study of Kumasi, Ghana." Water and Sanitation Report 3. UNDP- World Bank Sanitation Program, Washington, D.C. Processed. World Bank. 1992. Worldt Development Report 1992. New York: Oxford University Press. World Bank Water Demand Research Team. 1993. "The Demand for Water in Rural Areas: Determinants and Policy Implications." The World Bank Research Observer 8(1, January): 47-70. TiiE WORLI) BANK I-( ONOMIC REVIEW V(l . 9, NO 397 -423 Political Influence on the Central Bank: International Evidence Alex Cukierman and Steven B. Webb Political influence on the central bank is measured here by looking at the probability that a central bank governor will be replaced shortly after a political change of government. The governor changes about half the time within six months of a nonconstitutional or other radical change of government-a military coup or a res- toration of democracy. The governor is much less likely to change within six months following a routine change in the head of government-about one-fourth of the time in developing countries and one-tenth in industrial countries. These indicators vary across countrics and correlate statistically with inflation and its variability and with real growth and real in2terest rates. Differences in the vulnerability of the central bank to political instability, in political instability itsellf and in central bank turn- over in nonpolitical periods seem to be a major part of the explanation for why developing countries have, on average, higher and more variable inflation than in- dustrial countries dio. Economists and policymakers generally feel that the degree of autonomy of the central bank from political authorities is an important determinant of policy choices and of economic performance. Empirical verification of these presumptions has been difficult, however, because the autonomy of the central bank is not easily quantified. Most previous studies have used legal indexes from central bank char- ters to quantify the autonomy of the central bank (Parkin 1986; Grilli, Masciandaro, and Tabellini 1991; Alesina and Summers 1993). These measures help account for cross-country inflation differentials within industrial economies but not within de- veloping economies. Low inflation is not associated with the legal independence of central banks in developing countries because of the small degree of association between actual and legal independence among those countries. In industrial coun- tries, the frequency with which a central bank governor is replaced within three months of the time designated by law is more than ten times higher than in other periods. In developing countries it is only 2.2 times higher. Alex Cukierman is with rhe Department of Economics at Tel-Aviv University and with the Center for Economic Research at Tilburg liniversity, the Netherlands. and Steven B. Webh is with the Latin America and the Caribbean Country Department III at the World Bank. The authors acknowledge impnrtant research assistance by Charles Guo, Pantelis Kalaitzidakis, Bilin Nevapti, Pedro Rodriguez, and Kenneth Xu and assistance by Stephan Haggard and Yasuhito Asami in reviewing the political data in detail. The authors are thankful for the useful comments of several individuals and groups, including Thomas Havrilesky, Suzanne l ohmanm, and Sylvia Maxfield. Research for this article was funded by World Bank research grant RpO 677-77. (O 199 The International Bank for Reconstruction and Development /THE WORLI) BANK 397 398 THF WORLD IKANK F ONOMIC: RFVIEW, VOl.. 9, NO. x Hence, more behavioral indexes of independence are called upon. Cukierman, Webb, and Neyapti (1992) and Cukierman (1992) made an initial step in this direction by using the actual average term in office of the governor as a proxy for central bank independence in developing countries. They found a significant negative association between this proxy and inflation across developing coun- tries. The use of this proxy is based on the presumption that below some thresh- old a shorter term in office of the chief executive officer (to whom we refer as "governor," although the actual title may be president or chairman) of the bank is associated with lower central bank independence. It should be noted, in this context, that actual terms in office in developing countries are substantially lower than in industrial countries. Cukierman (1994) shows theoretically how the ex- pected length of tenure of the central bank governor relates inversely to infla- tion because of the governor's effective rate of time preference. This article presses the inquiry further by taking account of how turnover of the governor relates to political events. Although considerable evidence now establishes the negative association be- tween inflation and central bank autonomy, the reduction of political influence on the central bank is not the only institutional device for assuring price stabil- ity. Regardless of whether a country has a dependent central bank, the country may still enjoy price stability if it consistenily pegs its currency to that of a country with stable prices. Examples include Argentina in recent years, Belgium, the Netherlands, and some francophone African countries. An even stronger commitment to maintain a fixed exchange rate is to use a foreign currency for legal tender, as in Panama. But it should be emphasized that central bank au- tonomy and various degrees of commitment to a fixed exchange rate are com- plementary institutional arrangements and are not mutually exclusive. This ar- ticle focuses on the documentation and the effects of the varying degrees of political influence on the central bank. We and others have investigated the effects of exchange rate commitments elsewhere. Empirical investigations of the effect of exchange rate pegs on inflation appear in Cukierman, Rodriguez, and Webb (forthcoming) and in Anayadike-Danes (1995). Cukierman, Kiguel, and Liviatan (1992) offers a general discussion of dollarization. A variety of research-theory, case studies, and statistical analysis-shows that political instability worsens macroeconomic policy and outcomes (Edwards and Tabellini 1991; Haggard, Kaufman, and Webb 1992; Haggard and Webb 1994). The effects are strongest on inflation, but other variables such as growth and the real interest rate are also affected. Some macroeconomic deterioration results from the direct effects of political change on policy, particularly fiscal policy. Some also come from the effect on institutions, such as the central bank, and on their ability and incentive to follow policies for the long-run benefit of the economy. So the research agenda includes the question of the relative impor- tance of different channels through which political instability influences eco- nomic outcomes, as well as the underlying questions of whether and how politi- cal instabilitv affects institutionis such as the central bank. Cukierman and Webb 399 It seems likely that different kinds of political instability would have quite different effects. If political changes reflected changes in basic attitudes toward economic policy or if they were traumatic and irreversible for the politicians involved, then the instability would motivate politicians to control the central bank tightly and keep it at their disposal to help them stay in power. If, however, the political changes were alternations in power of two or three parties that shared a consensus on many basic tenets of economic policy, then the parties might agree to grant the central bank considerable autonomy to pursue price stability, so that the aspect of economic policy on which they agreed would not suffer from the political contests over other issues. These questions motivated the research for this article, although the article does not answer all of them. This article presents new behavioral indexes of political influence on the cen- tral bank. The indexes focus on the propensity of the governor of the bank to lose office following a political transition. The data base and the indexes derived from it were designed to address four specific questions derived from the broader questions in the previous paragraphs. First, is turnover at the central bank significantly different shortly after politi- cal transitions than in other periods? To answer this question, we compare the frequency with which governors are changed in periods shortly after a political transition with that in other periods. We refer to these other periods as "nonpo- litical," although recognizing that even then some turnover may result from political influence. Nevertheless, in the presence of heavy political influence we should expect central bank governors to be changed with significantly higher frequency in periods immediately after political transitions, which we call "po- litical periods," than in the nonpolitical periods. The tests presented in the ar- ticle suggest that this is indeed the case for the entire sample of sixtv-seven economies, for the subgroup of developing economies, and, surprisingly, for the industrial economies as well. Given this result, the frequency with which central bank governors are changed in political periods can serve as an index of the extent of political influence over the central bank and thus lead to more refined indicators of central bank autonomy. Second, for how long does a political transition, after its occurrence, increase the probability of a change of the central bank governor? To answer this ques- tion we examine a number of two- or three-month successive intervals follow- ing political transitions and calculate the frequency of changes at the central bank within each interval. The frequency of changes is high on average immedi- ately after a political transition and usually decreases as more months go by. The cutoff between political and nonpolitical periods is obtained by considering the last interval within which the frequency of changes at the central bank is significantly higher than the average frequency at the ten-month-or-more lag. The demarcation line between political and nonpolitical periods is then defined as the upper bound of this critical interval. With this procedure, periods within six months of the latest political transition are defined as political in the overall sample, and any period beyond that is nonpolitical. 400 THE WORLD BANK ECONOMIC REVIEW, VOL. 9. NO. 3 Third, is the political vulnerability of the central bank systematically related to the level of economic development and the type of political regime? We cal- culate a measure of the political vulnerability of the central bank and examine its relation with a country's level of development and the nature of its regime (whether it is always democratic, always authoritarian, or mixed). Vulnerability of the central bank is defined as the fraction of political transitions that are followed, within the subsequent political periods, by a replacement of the gover- nor of the bank. Frequent removal from office of the bank's governor following political tran- sitions probably reflects gross political influence, because the governor's term in office is not shielded by law or custom from political changes. This type of influence has been common in developing economies, such as Botswana, China, Costa Rica, and Indonesia, and has been particularly high in countries that switch between democratic and authoritarian regimes, such as Brazil, Chile, India, the Republic of Korea, Nicaragua, Peru, and Uruguay. In Argentina before the 1990s, even though the law specified a four-year term, the governor was always re- placed when the government changed. Among industrial economies gross politi- cal influence on the central bank is rare-Sweden appears to be an exception- but more subtle and mild influences are common. For instance, in the United States, the index of central bank vulnerability is zero, implying the absence of gross political influence. But, as documented in Havrilesky (1992), there are several other, milder, channels of political influence on the Federal Reserve. Fourth, is the political vulnerability of the central bank systematically related to measures of economic performance such as inflation, growth, and interest rates? Cross-sectional and panel-data regressions provide evidence on this question. The article is organized as follows. Section I presents the data set and dis- cusses the conventions used to organize it. The procedure for distinguishing be- tween political and nonpolitical changes of central bank governors is developed and applied in section II to the entire sample as well as to broad subgroups of countries. Section III discusses indexes of central bank political vulnerability for broad subgroups of countries and investigates the effect of different types of political transitions on central bank vulnerability. The effects of vulnerability and of nonpolitical turnover on inflation, growth, and real rates are briefly ex- amined in section IV. Section V estimates empirically how much of the strong cross-sectional association between inflation and its variance is due to their com- mon association with central bank vulnerability and nonpolitical turnover. Sec- tion VI follows with concluding remarks. I. THE DATA To assess and quantify the degree of political influence on the central bank, we have put together a multicountry data set on political and central bank insta- bility. Political instability is measured by counting political transitions of vari- ous types. Instability at the central bank is measured with data on the frequency Cukierman and Webb 401 and timing of replacement of central bank governors. The research focuses on variables that were available for a large number of countries on a uniform basis. The sample consists of sixty-seven economies with matched data on political and central bank instability. The economies are listed in table A-1. The sample includes all the major industrial and developing economies, but excludes most Eastern European economies. The data cover 1950-89, where possible, but start later for economies that achieved political independence or established a central bank after 1950. The data are divided into two subperiods: 1950-71, corre- sponding to the Bretton Woods era, and 1972-89. In each subperiod as well as in the total period, an economy is included only if data are available for at least ten years because data for shorter periods would often be unrepresentative. Political Transitions Instability of the executive branch of government seems most relevant for central bank autonomy. Consequently, we selected four types of political change as indicators of political instability: change of the head of government, change of the party in government, change of the fundamental rules of government as embodied in switches from authoritarian to democratic regimes or vice versa, and irregular changes of government from one authoritarian ruler to another. These types of changes form a hierarchy, so that each political event is coded as a level of instability: a change of the head of government without a change of party (low instability); a nonviolent change of party without a change of type of government (medium instability); an irregular change of authoritarian govern- ment without a change of form of government-a coup-(type 2 authoritarian); or a change of the form of government between democratic and authoritarian- a coup against a democratic government or a restoration of democracy-(high instability).' Our previous research indicated that the autonomy of the central bank dif- fered markedly between industrial and developing countries (Cukierman, Webb, and Neyapti 1992). Here we examine whether this distinction is also important for the effects of political instability on the central bank. Within each economic group, countries are divided into subgroups with democratic regimes for the whole period, with authoritarian regimes throughout, and with regimes that alternate between democratic and authoritarian. We classified as democratic only the countries and periods in which the head of government was chosen as a result of competitive elections. All others were classified as authoritarian, covering a variety of systems in which the government did not face serious pressure from electoral politics. The classifications thus differ somewhat from what one would 1. Data come from and were checked against several sources. The most comprehensive single source, and usually the initial one, is The Europa World Yearbook. Changes of only the economic team (not necessarily incLiding the central bank) are also events that may be relevant for assessing central bank autonomy. To investigate the relevance of these events, we have begun collecting data on changes of the Minister of Finance or Treasury. These data are less readily available than information on changes of the head of government, and we do not vet have information on a broad enough sample to report the results. 402 THE WORLL) BANK FCONOMIC RFVIFW, VOL. 9, NO. code as the degree of political liberty, such as with Gastil's indexes. We chose to exclude from our data political events that challenged and perhaps threatened to change a government but did not do so-strikes, riots, attempted coups, and elections in which the previous government was reelected. To assess the impli- cations for an institution such as the central bank, it seemed better to focus on institutionally well-defined events that actually changed the government to which the bank reported. Changes at the Central Bank The raw data on instability at the central bank consist of the actual dates of changes of the governors of the central banks in our sample of countries. The turnover of governors is only an imperfect indicator of actual central bank in- dependence in developing countries (for more discussion of this proxy, see Cukierman, Webb, and Neyapti 1992; Cukierman 1992; and Cukierman and others 1993). Low turnover does not always imply a high level of central bank independence-relatively subservient governors may stay in office longer pre- cisely because they do not stand up to the executive branch. This may be true for countries with exceptionally low turnover rates such as Denmark, Iceland, and the United Kingdom and for countries with stable authoritarian governments. In such countries, low turnover is probably unrelated to independence. Above some threshold turnover rate, however, higher turnover of the governor reflects lower central bank independence. Sufficiently high turnover rates make the tenure of the central bank governor shorter than that of the executive branch and thus make the governor more susceptible to influence by the executive branch and more discouraged from trying to implement longer-term policies. Because the electoral cycle is at least four years in most countries, the threshold turnover rate is probably between 0.2 and 0.25 (implying a governor's tenure of four to five years). In addition, governors with very short terms of office, such as three years or less, generally have more difficulty in implementing long-term policies (for example, the main- tenance of price stability) for any electoral cycle. One limitation of the turnover variable is that all the industrial countries have turnover rates at or below the threshold rate in the years we examine. Turnover rates in developing countries, however, span a range that goes well above the threshold point. The measures of political influence on the central bank developed here consider the links between political instability and subsequent turnover at the central bank and hence appear to be relevant for both industrial and developing economies. They also make it possible to distinguish between the frequencies of changes at the central bank in political periods, as defined earlier, and in nonpolitical periods. A relatively high frequency of turnover in the political periods indicates substantial influence of political instability on the central bank. A relatively high level of turnover the rest of the time most likely indicates that the central bank is more dependent even in politi- cally tranquil times. Cukierman and Webb 403 II. POLITICAL CHANGE AND THE TIMING OF CENTRAL BANK TURNOVER To what extent does political instability translate into instability at the cen- tral bank? Are there systematic differences in central bank turnover between periods immediately following a political change and nonpolitical periods? If there are differences, how does their magnitude vary with country characteris- tics? To answer these questions, this section measures political instability by the frequency of political transitions, as defined in section 1, and instability at the central bank by the turnover of the governors. Before attempting to give precise answers to these questions, it is instructive to take a broad look at the data. Table I presents central bank turnover figures (measured as the number of changes of central bank governor per month) for various intervals following a political transition. Thus, the average frequency of changes of governor within one month following a political transition is 0.063 per month; within two to three months it is 0.026, and it is only 0.015 at ten months or more after a political transition.2 These numbers correspond to gov- ernors' average terms of office of 1.3, 3.2, and 5.6 years, respectively.3 The numbers in table 1 can be interpreted as estimates of the probability per month of a change in central bank governor conditional on being within a time interval, i, that follows a political transition by i to i + 2 (or by i to i + 3) months.4 Table 1. Frequency of Change of Central Bank Governor at Various Intervals by Economy Group, 1950-89 (number of changes per month) Interval Number of (number of months since a political transition) Economy group economies 0-1 2-3 4-6 7-9 IO or more All 67 0.063 0.026 0.025 0.013 0.015 Industrial All 20 24 0.008 0.013 0.013 0.003 0.009 Democratic only 18 0.020 0.009 0.015 0.002 0.009 Mixed 2 0.060 0.0(0 0.000 0.024 0.013 Developing All 47 0.096 0.041 0.034 0.022 0.018 Democratic only 9 0.046 0.016 0.065 0.011 0.025 Authoritarianonly 16 (.089 0.017 0.006 0.025 0.015 Mixed 22 0.105 0.053 0.038 0.023 0.019 Source: Authors' calculations. 2. When a change in central bank governor occurs within a short time after two or more previous political transitions, the central bank change is attributed only to the most recent political transition. 3. In other words, let x be a turnover number from table I; then, the corresponding average term in office, in years, is given by I1/2x. 4. This interpretation requires the probability of two or imnre changes within a subperiod to be negligible. Because the time intervals considered are only two or three months, this assumption is supported by the data. 404 THE WORLD BANK ECONOMIC REVIEW. VOL 9, NO. 3 Table 1 reveals that, for the entire sample of economies, this probability decreases monotonically with the number of months that have elapsed since the last politi- cal transition. It is almost three times larger in the zero-to-one-month interval than in the two-to-six-month intervals. For seven-month intervals and beyond, this probability drops further-to about 60 percent of its value in the two-to- six-month intervals. Central bank changes within one month of a political transition are highly likely to result directly from the political change, but central bank changes more than nine months after a political transition are unlikely to be the re- sult of a change in government and are more likely to be largely nonpolitical. The challenge is to find the average elapsed time across countries, presum- ably between one and nine months, at which to set the cutoff between politi- cal and nonpolitical periods. We take the probability of turnover long after a political change (ten months or more) as a background rate against which to compare months in the intermediate range. The political period would then be defined to include months at intervals (after a political change) for which the average rate of central bank turnover is significantly higher than the background rate. The appendix describes the tests for differences between the probability of a change of central bank governor ten or more months after a political change and the probability in each of the intervals with shorter lags, using the normal approximation to the binomial distribution. It shows that for the whole sample the probability of a turnover at the central bank is significantly larger in the zero-to-one-month, two-to-three-month, and four-to-six-month intervals than in the ten-month-and-more interval. The average probability of a change of central bank governor during the seven-to-nine-month interval following a po- litical change does not differ significantly from the probability in the ten-or- more lag period. Hence, the evidence in the appendix supports a choice of six months following a political transition as the cutoff between political and non- political periods in the overall sample. In summary, for the entire sample of economies, the evidence supports the view that instability at the central bank rises in periods following political tran- sitions and that this increase is significant. The estimated probability of a change of governor at the central bank is more than two times larger in periods within six months after a political transition than in periods that are more removed from political change. Thus, political instability increases instability at the cen- tral bank and weakens its independence from political authorities. The profile of estimated probabilities of a change at the central bank varies between industrial and developing economies and between economies with dif- ferent political regimes. We distinguish three types of political regimes: stable democracies, stable authoritarian regimes, and mixed regimes (those alternat- ing between democratic and authoritarian regimes). Most industrial economies are stable democracies, none are authoritarian-only for the period under in- vestigation, and only two (Spain and France) are mixed regimes. Developing Cukierman and Webb 405 economies include all three political regimes-democratic only, authoritarian only, and mixed. Table 1 also presents the estimated probability of a change of central bank governor at various intervals following a political transition for the different subgroups of economies. The probability of central bank turnover is higher in developing economies than in industrial countries at all intervals . The frequency declines monotonically for developing economies but has two peaks for indus- trial economies-one in the zero-to-one-month interval and a lower peak in the four-to-six-month interval. The tests reported in the appendix reveal that for developing economies the appropriate cutoff between political and nonpolitical periods is at six months as in the overall sample, and there is clear evidence of a link between political instability and instability at the central bank. The average frequencies of central bank turnover within the three political subgroups of developing economies are summarized in the lower part of table 1. In the authoritarian and mixed subgroups, the two largest, the frequency in the zero-to-one-month interval is substantially higher than in the subsequent inter- vals. Statistical tests reveal that in all three cases the difference in frequencies between the zero-to-one-month interval and the ten-month-or-more interval is significant. Comparing the three subgroups, the average frequency of central bank instability in the zero-to-one-month and the two-to-three-month intervals is highest on average in the countries with mixed regimes, second highest in authoritarian-only countries, and lowest in democratic-only countries. The evidence is less dramatic for industrial economies, but still supports the view that in those countries the average frequency of central bank changes in the zero-to-one-month interval after a political change is significantly higher than the background frequency in the ten-month-or-more interval. The frequen- cies in the two-to-three-month, four-to-six-month, and seven-to-nine-month intervals are not significantly different from the frequency for the ten-month- or-more lag. Thus, for industrial economies as a subgroup we locate the cutoff between political and nonpolitical periods after the zero-to-one-month interval. We present most of the results in the rest of the article for both the zero-to-one- month and zero-to-six-month cutoffs. The tendency of mixed regimes to have higher frequencies of central bank changes than democracies in the zero-to-one-month interval also appears within the industrial countries, although this finding is based only on the experience of France and Spain. Also, as was the case for industrial democracies, there is a second peak in the four-to-six-month interval for democratic-only developing economies. 111. POLITICAL VULNERABILITY OF CENTRAL BANKS Computing the frequency of central bank turnover in intervals at various lags from political changes was important for assessing the size and duration of the effect of political changes on the propensity for central bank turnover. For a 406 THE WORLD BANK ECONOMIC REVIEW, VOL. 9. NO. 3 summary measure of political influence, however, we compute an index of the political vulnerability of the central bank, defined for each country as the frac- tion of political transitions that are followed promptly by a replacement of the central bank governor: Number of replacements of the central bank governor within i months following a political transition /(i)_ ,i= 1,6. Number of political transitions Table 2 reports average values of the index of central bank political vulnera- bility for industrial and developing economies as well as for democratic, au- thoritarian, and mixed regimes within each group. The overall average value of vulnerability is 0.24. That is, almost a quarter of all political transitions are followed by a replacement of the central bank governor within six months. As with the frequency of turnover, the vulnerability varies widely across country groups and subgroups. Political vulnerability is more than three times larger in developing economies than in industrial ones. A similar picture (not shown) emerges when the sample is broken into two subperiods (1950-71 and 1972- 89). Central banks of developing economies with mixed regimes are the most vulnerable on average. Table A-1 in the appendix presents the vulnerability of central banks to all types of political transitions for individual economies by subgroups. For econ- omies with a small number of political transitions in a subperiod, the vulnerabil- ity ratios are highly sensitive to the effect of a truly nonpolitical central bank Table 2. Political Vulnerability of Central Banks by Economy Group, 1950-89 Vulnerability, Frequency of political Economnygroup Within six months Within one nmonth change (per year) All 0.24 0.12 0.27 Industrial (.10 0.05 0.32 Developing 0.35 (.18 0.24 Industrial Democratic only 0.10 0.04 0.32 Mixed 0.12 0.12 0.33 Developing Authoritarian only 0.22 0.17 0.14 Democratic only 0.30 0.09 0.18 Mixed 0.39 0.20 0.30 a. Average share of political transitions followed by a replacement of the central bank governor within the noted period. Source: Authors' calcuiations. Cukierman and Webb 407 change accidentally happening after a political change.5 Although vulnerability numbers for individual economies in such cases should be viewed cautiously, they are still useful for the overall statistical analysis. Within each subgroup, central banks are arranged in descending order of their political vulnerability in 1972-89. By this measure, Sweden has the most politically vulnerable central bank among industrial democracies in that subperiod. Among democratic-only developing economies, the central bank of Botswana is the most vulnerable. Among developing economies with mixed regimes, the Argentine central bank is the most vulnerable, with sometimes more than one change of governor in the first six months after a political change.6 The last column of table 2 reports the average frequency of political transi- tions per year for each country group. Industrial economies have political tran- sitions more frequently than developing economies, because they have much more frequent democratic changes of government, with and without party changes. Within developing economies, the subgroup with mixed regimes has a frequency of political transitions that is about twice that of either democratic- only or authoritarian-only regimes. Authoritarian-only regimes have the lowest number of political changes of any type. We now have four indicators of central bank autonomy: the two developed in Cukierman, Webb, and Neyapti (1992)-the index of legal autonomy and the (total) turnover rate of the governor-plus vulnerability and its complement, the frequency of nonpolitical turnover of the governor. Of course the latter two can be reaggregated into the total turnover rate by making appropriate adjust- ments of units and multiplying vulnerability by the frequency of political transi- tions in each country. How different are these four measures? Table 3 shows the matrix of correlation coefficients for the four measures and the frequency of political turnover. Most of the various indicators are not closely correlated to one another, although total turnover is closely correlated with its two deriva- tives. The frequency of political transitions, an indicator of vibrant democracy or of more-fundamental political instability, is not significantly correlated with the various central bank variables. How is the political vulnerability of the central bank affected by the type of political instability in a country? To answer this question, we regressed the in- dexes of central bank vulnerability at lags of zero to one month and zero to six months on various types of political transitions, on a dummy for developing countries, and on a dummy for purely authoritarian regimes. Table 4 summa- rizes the results. With the vulnerability index for the lag of zero to six months, 5. This has been the case at least once in Jamaica and once in the Netherlands. In the Netherlands the decision to replace the governor in 1967 was made prior to the adjoining political transition, but was inplemented after it (de Haan 1995). Thus, vulnerahility is a noisy measure of political influence whose quality rises with the number of political transitions in a country. 6. In the 1990s, Argentina has vastly increased its commitment to price stability by upgrading the autonomy of its central monetary institution and by endowing it with the authority to function as a currency board. 408 THE WORLD BANK FCONOMIC REVIEW, VOL. 9, NO. i Table 3. Correlation between Various Indicators of Central Bank Autonomy, 1950-89 Index of legal Frequency of central bank Total central political Indicator autonomy bank turnover Vulnerability change Total central bank turnover -0.05 Vulnerability (six months) -0.11 0.78 Frequency of political transitions -0.05 0.06 -0.11 Nonpolitical central bank tuLrnover -0.02 0.88 0.60 -0.21 Source: Authors' calculations. the developing-country dummy has a significant positive impact on central bank vulnerability. But none of the indicators of political change had a statistically significant effect on vulnerability. Low- and medium-level changes clearly have no effect, which is consistent with the ambiguous theoretical priors. For high- level political change, the coefficient has a positive sign, as expected, but is not quite significant, although it was with some earlier versions of the sample. The vulnerability of central banks to political influence depends on the type of political transition. Table 5 reports the indexes of political vulnerability by Table 4. The Impact of Political Change on Central Bank Vulnerability, 1 950-89 Estimates Explanatory For lag of For lag of variable 0-I month 0-6 months Constant 0.09 0.16* (1.6 1) (2.16) High-level political change 0.73 0.96 (1.46) (1.44) Type-2 authoritarian transitions -0.16 -0.47 (-0.29) (-0.65) Medium-level political change -0.15 -0.22 (-0.92) (-1.01) Low-level political change -(.09 -0.19 (-(1.52) (-0.78) Dummy for purely authoritarian regimes 0.04 -0.08 (0.55) (-0.79) Dummy for developing countries 0.09 0.24-* (1.48) (2.96) Number of observations 110 110 Adjusted R2 0.059 0.143 Significant at the 5 percent level Significant at the I percent level. Note: The estimated equation is a pooled cross-sectional time-series regression in which there are two observations for each country, one for 1950-71 (where available) and one for 1972-89. The dependent variable is central bank vulnerability. t-statistics are in parelitheses. Source: Authors' calculations. Cuekierman and Webb 409 type of political transition for different economy groupings. The fraction 0.61 at the intersection of the mixed developing-economy row with the "high-level" column means that 61 percent of high-level political transitions in developing economies are followed within six months by a replacement of the central bank governor. The other numbers in the table are defined in a similar manner. There is generally little difference between the vulnerability to medium-level and to low-level political transitions within each country group. The vulnerability to medium- and low-level political transitions is more than twice as high in devel- oping economies as it is in industrial ones. Furthermore, this difference seems to be due to a difference in the level of development, rather than to differences in regimes, because in democratic developing economies vulnerability to medium- and low-level political transitions is more than twice as large as that of in- dustrial democracies. The vulnerability to these types of political transitions in mixed-regime developing economies is of the same order of magnitude as the vulnerability in democratic developing economies. The highest level of central bank vulnerability occurs in the face of high-level political transitions, which are all in developing economies (with mixed regimes, by definition). For developing economies as a whole, central bank vulnerability to type-2 authoritarian transitions (0.46) is larger than vulnerability to medium- and low-level transitions by a factor of almost two. The 0.46 figure, however, masks quite different tendencies in authoritarian-only regimes, which have an average vulnerability of 0.20, and in mixed regimes, where type-2 authoritarian transitions lead to an average vulnerability of 0.55. Table 5. Central Bank Vulnerability (within Six Months) by Type of Political Transition and Economy Group Type of political transition Economy group High lev'el Type-2 authoritarian' Medium let-el, Low letveld All 0.58 0.46 0.15 0.19 Industrial 0.00 n.a. 0.11 0.08 Developing 0.61 0.46 0.24 0.26 Industrial Democratic n.a. n.a. 0.11 0.08 Mixed 0.00 n.a. 0.12 0.17 Developing Authoritarian only n.a. 0.20 n.a. 0.23 Democratic only n.a. n.a. 0.24 0.25 Mixed 0.61 0.55 0.24 0.28 n.a. Not applicable, because there were no political transitions of that type for that category of country. a. Change of regime, from democratic to authoritarian or vice versa. b. Replacement of one authoritarian regime by another one. c. Change of party without a change in regime. d. Change of head of government without a change in regime or party. Source: Authors' calculations. Table 6. Estimates for Central Bank Vulnerability and Inflation, 1950-89 Estimiation with six-month vulnerability index Estimation with one-month vulnerability index Dependent variable is Dependent variable is Dependent variable is Dependent variable is transformed standard devliation transformed standard deviation Explanatory variable inflation, I) of D inflation, D of D Constant 0.046* 0.019 0.054** 0.020 (1.73) (1.28) (2.16) (1.53) Vulnerability - 0.l64t* 0.106'* (lag 0-1 months) {3.95) (4.79) Vulnerability 0.092- ` 0.070* (lag 0-6 months) (2.89) (4.02) Nonpolitical turnover 0.158** 0.108*-* (lag over I month) (2.12) (2.73) Nonpolitical turnover 0.239*** 0.108** (lag over 6 months) (2.60) (2.14) High-level 0.312 0.237** 0.208 0.193* political change (1.50) (2.08) (1.04) (1.83) Type-2 authoritarian 0.308 0.190 0.268 0.167 transitions (1.38) (1.56) (1.24) (1.45) Medium-level 0.044 0.010 0.029 0.003 political change (0.65) (0.27) (0.44) (0.09) Low-level 0.126* 0.041 0.109 0.028 political change (1.72) (1.02) (1.53) (0.74) Dummy 0.030 0.023 0.016 0.017 authoritarian only (0.98) (1.37) (0.52) (1.08) Dummy first period -0.087** -0.021** -0.083** -0.019* (1950-1971) (-4.53) (-2.02) (-4.44) (-1.95) Dummy -0.007 -0.005 0.006 -0.001 developing countries (-0.26) (-0.32) (0.22) (-0.10) Numberof observations 110 110 110 110 Adjusted R2 0.34 0.34 0.37 0.41 -Not available. * Significant at 10 percent level. Significant at 5 percent level. * * t Significant at I percent level. Note: t-statistics are in parentheses. Source: Authors' calculations. Cukierman and Webb 411 IV. EFFECTS OF CENTRAL BANK VULNERABILITY ON THE ECONOMY Besides being of independent interest, measures of central bank autonomy are useful for testing the effect of this autonomy, or its absence, on the economy. Such a project is largely beyond the scope of this article. Nevertheless, this sec- tion and the next briefly report evidence on the relation between some of our measures and the performance of the economy as reflected by inflation, real growth, and real interest rates. Each of these variables is regressed in a pooled cross-sectional time series on central bank vulnerability, on nonpolitical turnover, and on other control vari- ables. Nonpolitical turnover is measured as the average number of nonpolitical changes of central bank governor (more than one month or more than six months after the latest political transition) per year. Inflation Previous work has shown a significant positive relationship between inflation or the rate of depreciation in the real value of money (D) and the (total) turn- over of central bank governors for developing economies (Cukierman, Webb, and Neyapti 1992: tab. 7; Cukierman 1992: tab. 20.2). Table 6 here extends the effort by relating D and its standard deviation to the index of vulnerability, nonpolitical turnover, various types of political instability, and three dummies: one for the Bretton Woods era, one for having only authoritarian regimes in the period, and one for being a developing country. This formulation makes it pos- sible to evaluate the relative importance of lack of central bank autonomy and of political instability for the creation of inflation. Table 6 shows results with the zero-to-one-month and zero-to-six-month measures of vulnerability, and the results are very similar for the two different measures. The regressions in table 6 back the view that the first channel is more important. The political vulnerability of the central bank as well as nonpolitical turnover at the bank have a positive and significant impact on inflation (D) and its vari- ability. Political instability, particularly of the high-level sort, contributes to in- creased inflation variability. Low-level political instability has a marginally sig- nificant effect in raising the level of inflation. Medium-level political instability-alternation between parties (in a democracy)-has essentially zero effect on inflation, probably because frequent transitions of this sort require fundamental consensus on policy issues. Also, worldwide inflation is lower dur- ing the Bretton Woods period, even after allowing for the effect of other vari- ables. This is consistent with the view that fixed exchange rates have a stronger disciplinary effect on policy than flexible rates do. But the higher worldwide inflation after 1972 may also be the result of the larger shocks that affected the world economy in the post-Bretton Woods period. The insignificant authoritar- ian-only dummy means that having an authoritarian rather than a democratic 7. D = 7t/( I + it), where 7i is the annual inflation rate. 412 THE WORLD BANK E(:ONOMIC RFVI1W, VOL. 9, NO. 3 regime throughout does not help to explain differences in average inflation, once the central bank and political variables are taken into account. A dummy is not needed for mixed-regime countries, because the variable for high-level political changes is nonzero only in mixed-regime countries. The most important result in table 6 is the insignificance of the dummy vari- able for developing economies. Inflation is significantly higher on average in developing economies, and this could not have been accounted for by just look- ing at the overall frequency of turnover at the central bank (Cukierman, Webb, and Neyapti 1992). Once turnover is disaggregated into its constituent parts, however, and account is taken of the different types of political instability, the distinction between developing and industrial economies no longer contributes to explaining differences in inflation outcomes. In other words, the differences in vulnerability of the central bank to political instability, in central bank turn- over in nonpolitical periods, and in political instability can fully account for the developing economies having higher average inflation than industrial economies. Growth Recent empirical literature on growth has identified several variables, such as initial gross domestic product (GDP) and education, that are significantly related to real growth (see Barro 1991, for example). Does the degree of anti-inflation commitment by the monetary authorities, for which our indexes of central bank autonomy are a proxy, have any significant impact on growth after controlling for these variables? Table 7 answers this question by presenting growth equa- tions that take account of initial GDP, initial primary and secondary education enrollment ratios, a decade-by-decade change in terms of trade, and our indexes of central bank autonomy. With a full sample of countries, nonpolitical turn- over of central bank governors has a marginally significant positive sign, con- trary to priors. Brazil, Korea, and Botswana are outliers, however, because they achieved high growth despite high rates of central bank turnover and high vul- nerability. With those countries excluded, the six-month vulnerability indicator has a significantly negative sign.8 This finding supports the view that, other things being equal, higher political dependence of the central bank tends to re- tard growth in most countries. It is possible that political vulnerability of the central bank is a proxy for general economic and political instability, both of which deter growth, possibly by slowing down investment and innovation. There is some evidence that in developing countries, higher central bank vulnerability is associated with lower levels of investment (Cukierman and others 1993). Sorting out the channels through which political vulnerability of the central bank affects growth clearly deserves further work. 8. Botswana could legitimately be discarded from the sample, because its growth was primarily due to the discoverv of diamonds, not to good macroeconiomic policy. Brazil and Korea seem to be simply countries where lack of central bank autonomy was not a hindrance to growth, at least not in the period covered here. Fuirther iustification for the exclusion of these countries appears in Cukierman and others (1993). Cukierman and Webb 413 Table 7. The Impact of Central Bank Vulnerability and Nonpolitical Turnover on Economic Growth, 1960-88 Estinate Explanatory variable Full sample Sample minus three' Constant -0.15 0.73 (-O.I5) (0.77) Initial GDP -0.22*' -0.25**' (-2.46) (-2.83) Change in terms of trade 28.9** 28.1-* (4.8 7) (5.01) Initial primary education 2.03' 2.53*** enrollment ratio (1.98) (2.55) Initial secondary education 1.59 1.34 enrollment ratio (1.22) (1.04) Nonpolitical turnover 5.8WX -2.39 of central bank governors (1.80) (-0.66) Political vulnerability -0.78 -1.51* of central bank (-1.1$) (-2.30) Dummy for the 1 960s 1.9'; 1.42* (3.01) (2.61) Dummy for the 1970s 1. 9,** 1.11 (2.74) (2.26) Number of observations 129 120 Adjusted R2 0.23 0.26 Significant at the 10 percent level. Significant at the 5 percent level. *** Significant at the 1 percent level. Note: The sample consists of, at most, three observations (one for each decade) for each country. The estimated equation is a pooled cross-sectional time-series regression. The dependent variable is real per capita GDP growth. Central bank vulnerability is characterized in terms of changes in central bank governors that occur within six months of a political transition. t-statistics are in parentheses. a. Botswana, Brazil, and the Republic of Korea are exciluded. Source: Authors' calculationis. Real Interest Rates on Deposits Previous evidence (Alesina and Summers 1993; Cukierman and others 1993) suggests there is a negative relation between the variability of ex post real inter- est rates and central bank independence. In our work, we have used as proxies for central bank independence the bank's legal independence for industrial econo- mies and the bank's turnover of governors for developing economies. We now reexamine the relation between central bank independence and real interest rates by using central bank vulnerability and nonpolitical turnover as proxies for a lack of central bank independence. The second column of table 8 presents the effect on the variability of real rates. It appears that both central bank vulner- ability and nonpolitical turnover significantly increase the variability of ex post real deposit rates. This further supports and amplifies the conclusion that the variability of ex post real deposit rates is lower in countries with more indepen- dent central banks. Although it has the expected negative sign, the coefficient on legal independence is not significant. 414 THE WORL[) BANK ECONOMIC REVIEW, VOL. 9, No I Table 8. The Impact of Alternative Measures of Central Bank Autonomy on the Ex Post Real Deposit Rate Estimate Dependent variable is Dependent variable is real ex post standard deviation Explanatory variable deposit rate, R of R Constant 1.98 1.04 (0.64) (0.3) Vulnerability -5.5 6* 8.68* (lag 0-6 months) (-1.89) (2.1) Nonpolitical turnover -1.98 36.24** (lag more than 6 months) (-0.15) (3.0) Legal independence -3.26 -5.60 of central bank (-0.48) (-0.7) Number of observations 34 34 Adjusted R2 0.()5 0.52 Significant at the 10 percent level. Significanlt at the 5 percent level. * Significant at the 1 percent level. Note: t-statistics are in parentheses. Source: Authors' calculations based on data from Easterly, Rodriguez, and Schmidt-Hebbel (1992), Cukierman,Webb,and Neyapti (1992), and Cukierman (1992:ch. 19). The first column in table 8 relates the average level of the ex post real deposit rate to measures of central bank independence. The higher the political vulner- ability of the central bank, the lower the average real deposit rate, which cor- roborates a similar finding in Cukierman and others (1993). The effect of low- ering the real deposit rate probably reflects the higher implicit taxation of financial savings in countries with more politically dependent central banks. V. CENTRAL BANK INDEPENDENCE AND THE MEAN AND THE VARIABILITY OF INFLATION The strong cross-country association between the mean and the variability of inflation is a well-known empirical regularity. When legal independence and the total turnover of central bank governors are used as proxies for central bank independence, up to a quarter of this association is accounted for by their com- mon association with central bank independence (Cukierman 1992: ch. 18, 22). For this article, the same experiment was repeated with central bank vulnerabil- ity and nonpolitical turnover as indexes of central bank autonomy. Legal in- dependence was added as a regressor in industrial economies. To calculate the fraction of the correlation between the mean and the standard deviation of in- flation that results from their common association with central bank indepen- dence, we proceeded as follows. First, the cross-sectional covariance between D (the rate of depreciation in the real value of money) and its standard deviation was calculated. Second, both D and its standard deviation were regressed on central bank vulnerability, nonpolitical turniover, and (for industrial economies) Cukierman and Webb 415 legal independence. The predicted values of D and of its standard deviation were also calculated. Third, the covariance between those predicted values was calculated and compared with the overall covariance between D and its stan- dard deviation. The experiment was done for the entire 1950-89 period and for the subperiods 1950-71 and 1972-89. The fractions of the covariability between D and its standard deviation that are the result of their mutual link to central bank in- dependence are 0.30 for the whole period and 0.40 and 0.31 for the two subperiods. Thus, with the more refined measures of central bank autonomy presented in this article, about one-third of the association between inflation and its variability is attributable to their common association with central bank independence. VI. CONCLUDING REMARKS Stability and other characteristics of government institutions have always been recognized in economic history and in country studies as crucial determinants of macroeconomic stability. Recently this recognition has spread to theoretical work on macroeconomics and has been the focus of some cross-country statistical analysis (Fischer 1991; Edwards and Tabellini 1991; Cukierman, Edwards, and Tabellini 1992; Alesina and others 1995; Haggard, Kaufman, and Webb 1992). A survey appears in Alesina and Perotti (1994). As one would expect, political instability is positively related to inflation and negatively related to growth. There are various theories for how political instability causes macroeconomic instability, most of them not mutually exclusive. One contender is that political instability shortens time horizons of policymakers and that it decreases the abil- ity of the political system to efficiently resolve disputes over real incomes. This article has demonstrated that political instability causes instability at the central bank as well. But the spillover from political instability to instability at the central bank varies across country groups and types of political transitions. It is particularly large when the political regime changes from democratic to authoritarian or vice versa. This finding supports the view that when political change is deep enough to involve fundamental rules of the game and in other circumstances where political change would probably mean the party in power would not be back soon, if ever, then the expected effect of greater political instability would be shortened time horizons, as discussed above, and therefore less autonomy and stability for the central bank. Each new government would want to use the central bank to try to stay in power as long as possible and would have little concern for the associated detrimental longer-run effects. How- ever, greater frequency of low- and medium-level political change, such as just changing the head of government or the party, does not on average lead to reduced central bank autonomy as proxied by its political vulnerability. Actu- ally, in democratic governments in which the party changes frequently, the rul- ing parry might tvpically lack the strength to impose its will unilaterally and 416 THE WORLD BANK ECONOMIC REVIEW, VOL.. 9. NO 3 might thus agree to a compromise that would endure changes of party. This was explicit in the setup of the reformed Chilean central bank in 1989 (Arriagada and Graham 1994). An analysis of the effect of political instability on central bank independence in countries with democratic and stable rules of the political process appears in Cukierman (1994). An important issue for future work is the possibility of reverse causality be- tween the performance of the economy and our measures of gross political in- fluence on the central bank. The largely cross-sectional nature of the variables in this article precludes the use of Granger-Sims and other methods to test the exogeneity of our measures of political instability with respect to the perfor- mance of the economy. Hence, strictly speaking, it is possible that some of the significant relationships between these two groups of variables are due to cau- sality running from the economy to political influence. This is probably less likely to be the case for real interest rates than for real growth. There is evidence that real growth is affected by political instability (Barro 1991; Alesina and others 1995) which may, in turn, cause instability at the central bank. The measure of central bank vulnerability, however, is more likely to be ex- ogenous with respect to the economy than other behavioral indexes of political influence on the central bank. Even if political instability responds to the perfor- mance of the economy, vulnerability-defined as the ratio between instability at the central bank and political instability-depends on slowly changing institu- tions and sociopolitical norms, and thus may be largely exogenous to the cur- rent economic performance. Because the main contribution of this article is in the extraction of empirical regularities by matching two new data sets, our approach has been inductive rather than deductive. We deliberately avoided committing to and testing a par- ticular model because we believe that at this early stage the broad regularities in the data can be uncovered more efficiently without positing a particular model. Our hope is that the regularities uncovered here will encourage the construction of more precise models and further empirical testing of hypotheses. Table A-1. Political Vulnerability of Central Banks for Individual Economies by Subgroup, 1950-71 and 1972-89 1 950-71 1972-89 Vulnerability Number of Number of central Vulnerability Number of Number of central (within six political bank turnovers (within six political bank turnovers Economy months) transitions Total Nonpolitical, months) transitions Total Nonpolitical' Industrial, democratic Sweden 0.000 1 2 2 0.500 4 4 2 Ireland 0.000 5 3 3 0.286 7 3 1 Japan 0.200 5 4 3 0.222 9 4 2 Belgium 0.000 9 2 2 0.167 6 3 2 Switzerland 0.059 17 2 1 0.167 18 3 0 Finland 0.105 19 3 1 0.125 8 2 1 Italy 0.071 14 1 0 0.077 13 2 1 Australia 0.333 3 1 0 0.000 3 3 3 Austria 0.000 4 3 3 0.000 2 3 3 Canada 0.000 3 2 2 0.000 4 2 2 a Denmark 0.143 7 2 1 0.000 4 0 0 9 ;Germany, Fed. Rep. 0.333 3 2 1 0.000 2 2 2 United Kingdom (.000 6 2 2 (.000 3 2 2 Iceland 0.000 .3 1 I 0.000 0 0 Netherlands 0.143 7 1 0 0.00() 3 1 1 New Zealand 0.250 4 2 1 0.000 5 4 4 Norway 0.000 6 2 2 0.000 7 1 1 United States 0.000 4 2 2 0.000 4 3 3 Industrial, mixed France 0.056 18 2 1 0.500 2 4 3 Spain n.a. 0 5 5 0.167 6 3 2 Developing, authoritarian Egypt 0.(00 3 8 8 1.000 1 4 3 Mexico 0.250 4 2 1 0.667 3 4 2 China - - - - 0.50() 2 4 3 (Table continues on the following page.) Table A-1. (contiEnued) 19.50-71 1972-89 Vulnerability Number of Number of central Vulnerability Numtber ot Number of central (uithin six political bank turnovers (uwithin six political bank turnovers Economty months) transitions Total Nonpoliticael months) transitions Total Nonpolitical' South Africa 0.00( 4 2 2 0.500 2 3 2 Taiwan (China) 0.000 3 2 2 (.500 4 2 0 Fthiopia n.a. 0 0 0 0.400 5 5 3 Hungary - - - 0.3.33 3 2 1 Yugoslavia (former) n.a. 0 5 5 (.100 10 4 3 Kenya n.a. ( 2 0.000 1 2 2 Morocco 0.167 6 4 .3 0.(00 4 2 2 Romania - - 0 0 0.000 1 3 3 Tanzania ll.a. 0 I 1 0.00( 1 2 2 Uganda - - 0 0 0.000 4 2 2 Iidoniesia 1.(0(0 1 5 4 - 0 3 3 Qatar _ - 0 0 _ 0 0 0 Singapore n.a. 0 1 1 n.a. 0 6 6 Zaire ().000 1 2 I- 4 4 Developing, democratic Botswana - - - - 1.000 1 6 .5 Costa Rica 0.400 5 10 8 0.750 4 11 8 Malta ().000 1 1 1 0.500 2 5 4 Jamaica 1.000 1 3 2 0.333 3 7 6 West Samoa - - - - 0.333 3 3 2 Barbados - - - - 0.000 4 2 2 Israel 0.000 3 2 2 0.000 6 3 3 Bahamas - - - - n.a. 0 .3 3 Developing, mixed Argentina 1.714 7 20 8 1.111 9 16 6 Brazil 0.500 8 9 5 1.000 5 11 6 Chile 0.750 4 6 3 1.000 1 12 11 India 0.000 2 6 6 0.667 6 7 3 Korea, Rep. of n.a. 0 2 2 0.667 3 6 4 Peru 0.333 6 11 9 0.667 3 6 4 Uruguay 1.000 2 6 4 0.600 5 6 3 Colombia 0.167 6 5 4 0.500 4 3 1 Poland 0.000 1 0 0 0.500 6 6 3 Venezuela 0.167 6 5 4 0.500 4 7 5 Honduras 0.200 5 2 1 0.400 5 3 1 Portugal 0.000 1 4 4 0.385 13 7 2 Turkey 0.600 5 7 4 0.273 11 7 4 Greece 0.053 19 2 1 0.250 8 5 3 Ghana 0.000 4 2 2 0.200 5 4 3 Nigeria 0.000 2 2 2 0.200 5 3 2 Nepal 0.250 4 3 2 0.143 3 2 Panama 1.500 2 4 1 0.125 8 2 1 4 Thailand 0.250 4 4 3 0.111 9 3 2 Xc Malaysia 0.000 3 0 0 0.000 2 2 2 Pakistan 0.167 6 4 3 (.000 3 5 5 Philippines 0.01i0 4 3 3 0.000 2 2 2 -Not available. Vulnerability is not available when data on either the number of political transitions or the numilber of central bank turnovers are not available. ni.a. Not applicable. Note: Within each subgroup, central baniks are arranged in desceiidinig order of vulilerability during the 1971-89 subperiod. a. The nonpolitical turnover is defined as the umiber of clanges in the governor of the central bank that occurred more than six months after a political transition. Sources: Authors' calculations. 420 THE WORLD BANK F(ONOMIC REVIEW. VOL. 9, No). APPENDIX. TESTS FOR THE DETERMINATION OF THE CUTOFF BETWEEN POLITICAL AND NONPOLITICAL PERIODS To test for possible differences between the probability of a change of central bank governor in the ten-month-or-more intervals and in each of the other in- tervals in table 1, we use the normal approximation to the binomial distribu- tion. More specifically, let c1 be the number of changes of central bank governor in interval i following a political transition. As in table 1, i may assume the interval values (0-1), (2-3), (4-6), (7-9), and (10 or more). We denote each of these intervals by its lower bound. Thus i assumes the values 0, 2, 4, 7, and 10. Let n, be the number of intervals of type i in the sample. This number is de- termined by the total number of political transitions in the sample. The fre- quency of central bank governor changes within interval i is given by (A-1) fi=c/ni fori=0,2,4,7, 10. Let P1 be the true conditional probability of a change of central bank gover- nor in interval i following a political transition. The null hypotheses to be tested are (A-2) H,,:P,=PI() fori= 0, 2, 4, 7 against each of the alternative hypotheses that P, is significantly larger than P10. The specification of the alternative hypotheses accommodates the possibility that the probability of a turnover at the central bank is larger at shorter lags following a political transition. Under each of the null hypotheses the values (A-3a) Zi fi - , i = 0, 2, 4, 7 I+ where (A-3b) f, + 'I(' ,i=0,2,4,7 ni + n,( have approximately a standard normal distribution, provided both n, and n10 are larger than 30.9 (See, for example, Hunltsberger, Croft, and Billingsley 1980: 302.) The null hypotheses should be accepted for small values of the z,'s and rejected for large positive z.'s. Table A-2 displays these statistics. 9. These conditions are always satisfied for the entire sample of economies and the two main subsamples. For i = 0, 2, the terms n, and n are the numbers of two-month periods in the appropriate intervals. For i = 4, 7, they refer to the number of three-month periods. This variation in the length of the basic tine unit is introduced to make its length identical to the number of months within each interval (two months for i = 0, 2 and three months for i = 4, 7). In either case, the value ofn is sufficiently large to make the normal approximation to the binomial valid. Cukierman and Webb 421 Table A-2. Values of the Test Statistic Z1 for the Null Ho: P, = plo Interval (number of montbs) Economy group 0-1 2-3 4-6 7-9 All economies 12.0 2.8 2.9 -0.6 Industrial 3.2 -0.2 1.1 -1.6 Developing 13.0 4.0 3.3 -0.7 Note: Sample size is sixty-seven economies. Source: Authors, calculations. Because a z, of 1.96 implies that the null is rejected at the 0.05 level, we conclude that the probabilities of a turnover at the central bank are significantly larger for the 0-1, 2-3, and 4-6 intervals than for the 10-or-more intervals. However, there is no significant difference in the probability of a change of central bank governor between the seven-to-nine-month and the ten-or-more-month inter- vals. The evidence in table A-2 therefore supports the conclusion that for the whole sample the cutoff between political and nonpolitical periods occurs at a lag of six months following a political transition. Accordingly, we define periods of up to six months following a political transition as political and periods seven or more months after the latest political transition as nonpolitical. The respec- tive frequencies are 0.037 and 0.015, respectively. The z statistic testing the significance of the difference between these two estimated probabilities is a huge 9.6. A replication of these tests separately for each of the two economic categories of countries reveals that for developing countries the appropriate cutoff between political and nonpolitical periods is still at the upper end of the 4-6 interval. For developing economies the z statistic also very significantly rejects the null hypothesis that the frequency within the entire period of zero to six months is no different from the frequency in periods that are seven or more months after a political transition. In the industrial economies, the probability of a change at the central bank in the 0-1 interval is significantly higher than in the 10-or-more interval. The frequencies in the two-to-three-month, four-to-six-month, and seven-to-nine- month intervals, however, are not significantly different from frequencies in the 10-or-more interval. Thus, for industrial economies we should locate the cutoff between political and nonpolitical periods after the 0-1 interval. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Alesina, Alberto, Sule Ozler, Nouriel Roubini, and Phillip Swagel. 1995. "Political In- stability and Economic Growth." Draft. Alesina, Alberto, and Lawrence H. Summers. 1993. "Central Bank Independence and Macroeconomic Performance: Some Comparative Evidence." Journal of Money. Credit and Banking 25(May):I51-62. 422 THE WORLD BANK FCONOMIC RFVIEW. VOl. 9, NO. 3 Alesina, Alberto, and Roberto Perotti. 1994. "The Political Economy of Growth: A Critical Survey of the Recent Literature." The World Bank Economic Review 8(3):351-71. Anayadike-Danes, M. K. 1995. "Comment on 'Measuring the Independence of Central Banks and Its Effect on Policy Outcomes' by Cukierman, Webb, and Neyapti." The World Bank Economic Review 9(2):335-40. Arriagada, H. G., and Carol Graham. 1994. "Chile: The Maintenance of Adjustment and Macroeconomic Stability during Democratic Transition." In Stephan Haggard and Steven B. Webb, eds., Voting for Reform: Democracy, Political Liberalization and Economic Adjustment. New York: Oxford University Press. Barro, R. J. 1991. "Economic Growth in a Cross Section of Countries." Quarterly Journal of Economics 106(May):407-43. Cukierman, Alex. 1992. Central Bank Strategy, Credibility, and Independence: Theory and Evidence. Cambridge, Mass.: MIT Press. . 1994. "Commitment through Delegation, Political Influence, and Central Bank Independence." In J. 0. de Beaufort Wijnholds, S. C. W. Eijffinger, and L. H. Hoogduin, eds., A Framework for Monetary Stability. Boston: Ktuwer Academic Publishers. Cukierman, Alex, Sebastian Edwards, and Guido Tabellini. 1992. "Seigniorage and Political Instability." American Economic Review 82(June):537-55. Cukierman Alex, Miguel Kiguel, and Nissan Liviatan. 1992. "How Much to Commit to an Exchange Rate Rule? Balancing Credibility and Flexibility." Revista de Analisis Econ6mico 7(1, June):73-89. Cukierman, Alex, Steven B. Webb, and Bilin Neyapti. 1992. "Measuring the Indepen- dence of Central Banks and Its Effect on Policy Outcomes." The World Bank Eco- nomic Review 6(3):353-98. Cukierman, Alex, Pantelis Kalaitzidakis, Lawrence H. Summers, and Steven B. Webb. 1993. "Central Bank Independence, Growth, Investment, and Real Rates." Carnegie- Rochester Conference Series on Public Policy 29(Autumn):1-46. Cukierman, Alex, Pedro Rodriguez, and Steven B. Webb. Forthcoming. "Central Bank Autonomy and Exchange Rate Regimes-Their Effects on Monetary Accommoda- tion and Activism." In Harry Huizinga and Sylvester Eijffinger, eds., Positive Politi- cal Economy: Theory and Evidence. de Haan, Jakob, 1995. Personal communication. University of Groningen, the Nether- lands. August. Easterly, William, C. A. Rodriguez, and Klaus Schmidt-Hebbel, eds. 1992. "Public Sec- tor Deficits and Macroeconomic Performance." Draft. Edwards, Sebastian, and Guido Tabellini. 1991. "Explaining Fiscal Policies and Infla- tion in Developing Countries." Journal of International Money and Finance 10 (March, Supp.):16-48. The Europa World Yearbook. Various years. London: Europa Publications. Fischer, Stanley. 1991. "Growth, Macroeconomics, and Development." In Olivier J. Blanchard and Stanley Fischer, eds., NBER Macroeconomics Annual 1991. Cambridge, Mass.: National Bureau of Economic Research. Grilli, Vittorio, Donato Masciandaro, and Guido Tabellini. 1991. "Political and Mon- etary Institutions and Public Financial Policies in the Industrial Countries." Eco- nomic Policy: A European Forum 6(October):341-92. Cukierman and Webb 423 Haggard, Stephan, Robert Kaufman, and Steven B. Webb. 1992. "Democracy, Dicta- torship, and Inflation in Middle-income Countries." World Bank, Country Econom- ics Department, Washington, D.C. Processed. Haggard, Stephan, and Steven B. Webb. 1994. Voting for Reform:-Democracy, Politi- cal Liberalization and Economic Adjustment. New York: Oxford University Press. Havrilesky, Thomas. 1992. The Pressures on American Monetary Policv. Norwell, Mass.: Kluwer Academic Publishers. Huntsberger, D. V., D. J. Croft, and Patrick Billingsley. 1980. Statistical Inference for Management and Economics. 2d ed. Boston: Allyn and Bacon. Parkin, Michael. 1986. "Domestic Monetary Institutions and Deficits." In J. M. Buchanan, C. K. Rowley, and Robert Tollison, eds., Deficits. New York: Basil Blackwell. THE WORLD BANK ECONOMIC REVIEW, VOL. 9, NO. 3: 425-450 Pioneers for Profit: St. Petersburg Entrepreneurs in Services Martha de Melo, Gur Ofer, and Olga Sandler Russian private entrepreneurs in services are truly pioneering, because many services- especially trade, financial services, and most business anzd consumer services-are poorly developed. This article uses 1992 and 1993 survey data from St. Petersburg to assess the characteristics of these entrepreneurs, their firms, and the markets in which they deal. Evaluation of the firms' performance establishes how well they are doing and provides insight into the determinants of success. Their performance was surprisingly good and can be attributed to several factors, including the existence of a substantial gap between the desired and the actual levels of many services and the high level of education and motivation of the entrepreneurs themselves. Policy priorities are to achieve macroeconomic stability, a transparent legal framework and simplified regulations, a well-designed tax code, further privatization of real estate, and better access to fi- nance. Direct assistance programs would be useful in providing information, counsel- ing, and financing to small and medium-size firms. In McKay (1970), "pioneers for profit" were foreign industrial entrepreneurs in Russia prior to World War 1. Today, pioneers for profit are the Russian Federation's new domestic entrepreneurs. Under communism, private owner- ship of the means of production was not allowed, and markets as we know them did not exist. Prices were fixed and goods were distributed according to a central plan. Thus Russian entrepreneurs, who are creating and managing private busi- nesses in today's unstable macroeconomic and regulatory environment, are en- tering new territory. And entrepreneurs in services are truly pioneering, because many services-especially trade, financial services, and most business and con- sumer services-are poorly developed. Such activities were previously suppressed for a number of reasons, including the Marxist doctrine that held them to be speculative and unproductive. As a result, Russians have had relatively little experience in services, as customers or providers. Our interest in private entrepreneurs in services grew out of a study of the gap between the actual levels of services in the Russian Federation and other former Martha de Melo is with the Policy Research Department at the World Bank; Gur Ofer is with the Department of Economics at the Hebrew University in Jerusalem; and Olga Sandler is an independent consultant. The authors thank Leila Webster, Alan Gelb, and three anonymous referees for advice and the Leontief Center in St. Petersburg for logistical support in the conduct of the surveys. 0 1995 The International Bank for Reconstruction and Development / THE WORLD BANK 425 426 THE WORLI) BANK EC:ONOMIC REVIEW, VOL. 9, NO. 3 Soviet states and those to be expected on the basis of other countries' experi- ences.' Our interest was heightened by early evidence that the source of private sector growth in Central and Eastern European countries, in transition from a centrally planned to a market economy, was largely services. Because service firms are needed in most transactions, in some sense they are the market, provid- ing information, communication, and distribution. The market infrastructure that service firms provide facilitates the development of new agricultural and manufacturing activities as well as other services. Private business activity was officially allowed in the Soviet Union starting in 1987, although only limited private sector development took place before 1991. Since 1991 a new legal framework has been developed. Property rights, contract law, and company law have been woven into the fabric of society, although some laws and decrees are unstable or ambiguous and institutions are weak. Privatization of both small and large state enterprises has proceeded relatively rapidly in the Russian Federation, but its pace has varied considerably by re- gion. Privatization of commercial real estate has proceeded much more slowly for several reasons, including ambiguities in ownership and the absence of pri- vate ownership rights for land. In a typical market economy, the majority of all private firms are small and medium-size, and the majority of these are in services (on services, see, for ex- ample, Giarini 1987; on the importance of small firms, see Storey 1983, 1989 and Acs and Audretsch 1993). However, in the late 1980s only 22 percent of Russian firms had under 100 employees, compared with 94 percent in the United States, and a minority of the Russian small and medium-size firms were in ser- vices (Brown, Ickes, and Ryterman 1994). The share of services in gross domes- tic product (GDP) was an estimated 33 percent in Russia in 1990, compared with 50 percent or more in upper-middle-income countries and 67 percent in the United States (Goskomstat, various issues). Thus, the simultaneous growth of services and small firms is an important feature of restructuring for economies in transition to a market economy. Such growth is relatively easy because privatization of small firms is simpler than privatization of large ones and many service firms require only modest investment, simple technology, and limited market infrastructure. Even so, it is surprising how quickly the emerging market economy in the Russian Federation is being transformed. By 1993, although the number of stores per thousand inhabitants was still only 2 to 2.5, compared with 7 to 10 in most Western European countries, the share of services in GDI' had increased to 43 percent (Earle and others 1994; Goskomstat, various issues). Estimates for the first half of 1994 put services at around 50 percent of GDP. Both new firms and privatized firms are contributing to a partly unrecorded but growing private sector, characterized by smaller firms and a strong concentration in services. 1. See Easterly, de Melo, and Ofer (1994) for estimates of the gap in different types of services between the former Soviet Union and other countries. Services are defined there and in this article as all sectors of the economy other than agriculture, manufacturing, construction, and utilities. de Melo, Ofer, and Sandler 427 This article draws on survey data from 1992 and 1993 to look at the charac- teristics and performance of entrepreneurs in the services sector in St. Peters- burg. Section I describes the surveys and the firms that were included. Section II discusses the characteristics of the owners, their firms, and the markets in which they deal. Section III discusses firm performance and its relation to a variety of internal and external factors. The questions of why performance has been so good and whether it can be expected to continue are addressed here. And section IV draws some policy implications from the survey findings. Judged in the con- text of the Russian Federation's transition-with its turbulent environment, declining output, and lack of market and private sector tradition-firm perfor- mance looks remarkably strong. 1. THE ST. PETERSBURG SURVEYS Although the successful development of service firms is vital to the Russian Federation and other economies in transition, in earlv 1992 little was known about the emergence of service sector entrepreneurs and firms there or about the problems they faced. The lack of information prompted us to undertake our own surveys, which are unique in that they focus on service firms and provide observations of the same firms at two points in time-albeit close together. Our purpose was to understand the entrepreneurs, their firms, and the external busi- ness environment; to identify factors that encourage or impede growth; and to recommend policies to accelerate development. The surveys were undertaken in St. Petersburg-a city with historically close economic and cultural links with Europe and a large number of professionals who have worked in the city's military industrial complex. Although St. Peters- burg is not typical of the Russian Federation, the development of a market economy there is likely to precede such a development in many other parts of the Russian Federation and in other former Soviet states. A better understanding of this experience can contribute to some timely policy lessons. The first survey was undertaken in January 1 993 and included eighty-six en- trepreneurs (see de Melo and Ofer 1994). Firms were randomly chosen from 10,000 registered firms and spanned the full range of services, including trade, business services, consumer services, transport, finance, and health and educa- tion. An estimated one out of three firms responded to an initial inquiry and expressed willingness to be interviewed; a randomly chosen subsample was drawn from this population. Prior to the interview, firms were screened to ensure that each was majority privately owned and majoritv Russian-owned and that ser- vices accounted for 50 percent or more of revenues. No restrictions were placed on a firm's legal form or the number of its employees. An interviewer then per- sonally visited each firm and administered a broadly conceived questionnaire, which took an average of three hours. Most questions were qualitative in na- ture, although quantitative data were collected for major aggregates such as employment and gross revenues. 428 THE WORLD BANK ECONOMIC RFVIFW, VOl. 9, NO. Table 1. Comparison of 1992 Survey and Statistical Office Data on Service Firms, Including Cooperatives, in St. Petersburg by Service (percentage of total firms) Statistical Su rvey Service Office data dalta Trade 40 37 Business services 37 31 Consumer services 12 1 5 Health and education 9 8 Transportation 2 5 Finance 0 4 Total 1(0 100 Note: The Statistical Office data include some cooperatives and small enterprises with majority public ownership; services account for 52 percent of all activities. The survey data are based on eighty-six firms. Source: Authors' survey data and St. Petersburg Statistical Office ( 1993). Ten months later, in November 1993, a shorter, telephone follow-up survey was undertaken. All eighty-six original firms were located, although only eighty- two entrepreneurs could be interviewed. The four firms that could not be inter- viewed in the follow-up survey were two trading firms and one firm each in business services and house repair. The most surprising finding was that all eighty- six firms interviewed in January 1993 were still functioning at the end of the year. This success is unusual in the face of the high failure rate of small and medium-size enterprises in market economies. For simplicity, data from the first survey, which reflect end-1992 or the period of operation prior to end-1992, are labeled 1992; data from the second survey, which reflect end-1993 or calendar year 1993, are labeled 1993. Where data from only the first survey are used, the total number of firm observations is eighty-six; where panel data are used, the total number of observations is eighty- two. The sectoral distribution of survey firms, shown in table 1, is similar to that of all small and medium-size firms and cooperatives in St. Petersburg in 1992.2 And the survey firms are believed to be typical of the wider population in size, financing, market orientation, and major complaints. Although biases may ex- ist, their experience is interesting and provides insight into the Russian Federation's new private sector. II. SURVEY FINDINGS ON OWNERS, FIRMS, AND MARKETS St. Petersburg pioneer entrepreneurs are part of a spontaneous and wide-rang- ing human conversion to free enterprise. Talented people in state industries and government institutes are opting for a life of risk and opportunity in the newly 2. The data from the St. Petersburg Statistical Office were received only in lune 1993: they were not available at the time of the original survey. de Melo, Ofer, and Sandler 429 emerging private sector. In table 1 their firms are classified by their primary revenue activity into six categories of service activity. According to the survey data, the largest is trade (thirty-two firms or 37 percent), which includes whole- sale and mixed-wholesale-and-retail companies, as well as five retail-only com- panies (including one kiosk owner) and three restaurants. The second- largest category is business services (twenty-seven firms or 31 percent), which includes firms providing computer services, research and development, and legal and management consulting. The third-largest category is consumer services (thir- teen firms or 15 percent), which includes travel, repair, custom tailoring, hair- dressing, and publishing companies. The other three groups-health and educa- tion (public-type services that cater to both businesses and households), transportation, and financial services-include only a few firms each. Owners Who are these pioneers for profit and why did they become entrepreneurs and go into services? They are on the whole highly educated, energetic, ex- perienced, and resourceful people. Almost 85 percent have a universitv educa- tion, and 20 percent have graduate-level education. As also found in Ukraine, a strong educational background has provided them with both marketable special- ized skills and general analytical abilities useful in setting up and maintaining a business successfully in a complex, difficult environment. (For a comparison with private entrepreneurs in Ukraine, see Stone and Novitskaya 1993.) Entrepreneurs' explanations of why they started a business were, in order of frequency, to make more money, to become independent, to pursue their profes- sion more freely, and to pursue a hobby (for example, computer programming, commercial design, or the arts). Many entrepreneurs went into services related to their previous public sector activity, such as computer support or educational training. Others established service firms because of the relatively low capital investment required or because they saw the gap in services, such as tourism and restaurants, as a chance to establish a niche. Some 10 percent started out in manufacturing but switched to services, often trade, to replenish their cash flow. One entrepreneur began manufacturing security equipmenit but, faced with in- sufficient new orders, switched to repairing and servicing existing security systems. Virtually all the entrepreneurs interviewed were once employed in the public sector, where they were managers, academics or technical specialists, or work- ers. Table 2 gives a profile of the owners by background. The former managers represent insiders, the privileged elite of the communist regime. They neverthel- ess encompass more than the "nomenklatura," a very restricted group of direc- tors, high-level bureaucrats, and party officials. As a group, former managers were slightly older than the survey average, and most had university or technical education. Many of the former academics and technical specialists had com- pleted postuniversity degrees, and they employed a high proportion of profes- sionals in their firms. The former "workers" were either white-collar (clerical 430 THE WORiLD BANK E(IONOMI(: REVIEW. VOL. 9, No. I Table 2. Profile of Owners of Service Firms in St. Petersburg, 1992 Owner's background Academic or technical Indicator Manager specialist Worker Total Number of owners 41 34 11 86 Average age 45 41 35 42 Women (percent) 17 6 18 13 Higher education (percent) 8( 97 55 84 Previous employment (percent) State enterprise 85 19 55 57 Academic institute 5 68 9 29 Same as present 4. 31 55 42 Note: The data are based on eighty-one to eighty-six firmis, depending on the characteristic. X2 values for cross-tabIulations: age = 0.1)6; gender = 0.30; education = 0.00; previous work places = 0.00: and activitv = 0.26. (The XI statistic shows the probabiliry that the data are distributed randomlv across the cells. A low X2 indicares a meaningful difference in owner characteristics.) Source: Authors' calculations. and administrative) or blue-collar workers; they were typically younger, and had a lower level of education than the other two groups. The characteristics of firms run by entrepreneurs with different backgrounds are shown in table 3. Firms owned by former managers were more likely to be privatized and to engage in trade. The four former government officials in this group all ran (largelv privatized) trading firms whose goods were provided pri- marily by state enterprises-supporting the popular view that bureaucrats were in a position to arrange good deals for themselves. Their trading firms were large, averaging more than 200 employees, and more likely to have state enter- prises as customers. Firms owned by academics or technical specialists were more likely to be new firms, to engage in business services, and to receive foreign aid in some form. Firms owned by white- or blue-collar workers were more likely to be small, and they received no foreign assistance. In many ways, entrepreneurs of small firms in services behave like entrepre- neurs of small firms in other sectors. They rely heavily-and significantly more than in market economies-on personal contacts to accomplish business objec- tives, substituting such contacts for the information networks and standard gov- ernment and business procedures that exist elsewhere. For example, in Poland, where private sector development was more advanced, survey data in late 1992 showed that more than 90 percent of entrepreneurs there also found personal contacts to be the most frequent and efficient way of finding customers, even customers abroad (Wyznikiewicz, Pinto, and Grabowski 1993). These personal contacts originate in family relationships, ethnic or regional ties, work relation- ships, or military service. They are used to hire workers; to sell products, includ- ing exports to former Soviet republics; and to deal with the government, banks, de Melo, Ofer, and Sandler 431 Table 3. Profile of Service Firms in St. Petersburg by Owners' Backgrounds, 1992 Owner's background Academic or technical Indicator Manager specialist Worker Total Number of firms 41 34 11 86 Average number of full-time workers 59 62 20 55 Privatized (percent) 51 12 27 33 Received foreign aid (percent) 14 29 0 18 Service (percent) Trade 51 21 36 37 Financial and business services 22 49 27 35 Consumer services 20 12 9 15 Health and education 2 15 9 8 Transportation 5 0 18 5 Note: The data are based on seventy-seven to eighty-six firmis, depending on the characteristic. X' values: privatized = 0.01; activity = 0.02; and foreign aid = 0.52. Souirce: Authors' calculatioins. and state enterprises. Relying on personal contacts, however, is less efficient than using established, impersonal market institutions, and it can limit the abil- ity to expand business activity. In other ways, entrepreneurs in services differ from entrepreneurs in other sectors. A comparison with St. Petersburg entrepreneurs in private manufactur- ing (see Webster and Charap 1993) reveals that service sector entrepreneurs are more diverse in age, gender, and work experience; they operate newer, more independent firms; they face more competition in their output markets; and they are more active in seeking new markets through diversification. Firms Survey firms were quite young: as of January 1 993 their average age since start-up or their becoming majority private was less than two years, compared with eight years for Polish firms at that time (Wyznikiewicz, Pinto, and Grabowski 1993). Most survey firms were new firms and less than one-third had been privatized. Trading firms were particularly likely to have been privatized and accounted for two-thirds of all privatizations in the survey. Several consumer services had also been privatized. Most firms providing business services were new, suggesting that conditions of entry are particularly important for such firms. The distinction between new and privatized firms, however, is not as sharp as might be expected. Of the new firms, most originated within a public organiza- tion, often with no formal arrangements but having important economic links. For example, the owner of a new firm might moonlight on the job in the host organization, rent space from it (in theory or in fact), use its equipment, or 432 THE WORID BANK ECONOMIC RFVIFW, VOL. ), NO. I provide it with services. In many cases such symbiotic relations gradually erode, leading to full separation, but twenty-two firms classified as new in the survey still maintained one or more of these relations with their host organization. (This is also true in other transition countries. In the former Czechoslovakia, for ex- ample, new entrepreneurs' personal contacts with former state enterprise man- agers were important in maintaining access to materials, second-hand equip- ment, and market information, according to McDermott and Mejstrik 1992.) For privatized firms in the St. Petersburg sample, the minority public ownership was relatively high in a few firms, but it accounted for only 3 percent of enter- prise equity on average. Some 80 percent of equity was held by individuals; the remaining shares were held by other firms. Most survey firms became private-either as new or privatized companies- in 1990 and 1991. As of January 1993 their average age was twenty-one months. Only two were founded in 1987, the year Michail Gorbachev officially allowed private businesses to open. About half the firms were registered as joint-stock closed or its equivalent-limited liability. The rest were registered as private individual, small enterprise, leasehold, or cooperative. Introduced in 1990, the joint-stock designation is popular because owners are not personally respon- sible for their debts. An additional seven firms had converted to this form by end- 1993. Firm size. Firms in the survey varied widely in size, from a health care company with one person and monthly revenues of $100 (R1O,000) to two firms with 1,000 employees or more and a trading company with monthly revenues of $20 million (R2 billion).3 Survey data indicate that 87 percent of private firms have fewer than 100 employees, compared with 22 percent under the old regime. However, this figure understates the extent of change in the industrial structure, because only registered firms are included in the survey. Registration entails responsibilities, such as bank accounts and tax payments, that discourage some businesses. Thus, many new businesses are uLndeclared and part of Russia's "shadow economy," recently officially estimated at more than 20 percent of GDP. Unregistered businesses include self-employed crafts workers and service providers who operate with or without a license. Monthly revenues for different activities are shown in table 4, although Russia's high inflation makes such estimates difficult. The median monthly revenue for survey firms during the last quarter of 1992 was $20,000 (R2 million), suggest- ing an annual turnover of about a quarter of a million dollars. As expected, gross revenues of trading firms were high, reflecting the high value of goods purchased for a given value added. But revenues of the few survey firms in fi- nance and transport were higher, reflecting the former's large profits and the latter's need to recover capital costs. In general, premises reflected revenues, 3. Rubles are converted to U.S. dollars using an estimated purchasinlg povwer parity exchange rate of R (10 (=$1. This compares to the nominal exchange rate of R400 = SI in the fourthli qarter of 1992 on the Moscow International Currency Exchange Market. de Melo, 01er. and Sandler 433 Table 4. Size of Service Firms in St. Petersburg by Revenue and Employees, 1992 Monthly revenue Number of full-time (R 1,000) employees Main activity Average Median Average Median Trade 98,711 4,000 95 25 Business services 9,200 1,50() 22 1 1 Consumer services 6,522 600 40 17 Health and education 985 50)( 101 Transportation 9,925 9,750 5 6 Finance 38,750 38,750 146 41 Total 40,263 2,(00 55 13 Note: Dara on monthly revenue, or gross sales, are based o1n sixty-four firms: data on full-time employees are based on eighty-five firms. The estimated purchasing power exchange rare tor rubles is R100 = $1. Source: Authors' calulations. with high-revenue firms occupyinlg large, well-equipped, and well-guarded offices. The median number of full-time employees in 1992, also shown in table 4, was thirteen, and the average, fifty-five. The median is quite typical of small and medium-size enterprises in Western Europe; the average is strongly affected by a few large firms (a transportation company and two trade companies) that had more than 800 employees. Trading firms tended to be larger than average; busi- ness services, smaller than average; and consumer services, about average. Diversification. Diversification has both static and dynamic dimensions. In 1992 many firms-especially those in trade or business services-were engaged in several activities, basically hedging their bets. These multiple activities were generally small-scale and sometimes were carried out as subsidiaries, under separate legal structures. This early enmphasis on diversification is consistent with the contemporary experience of Ukrainian private enterprises that were found to "grow by diversification, not expansion" (Stone and Novitskaya 1993). During 1993 St. Petersburg entrepreneurs in services focused more on lowering their costs or upgrading the quality of their product than on undertaking additional activities. This strategy is consistent with the Johnson and Loveman (1995) view that firms are more likely to survive when the attention of management is concentrated on a set of core activities. In addition to engaging in multiple activities at a given point in time, firms also diversified over time. This was particularly common before 1992, when many firms were using trial and error to explore a range of activities. During 1 993 the rate of diversification moderated, with most firms maintaining the same primary activity but some still looking for the right niche. Eighteen firms- most of them quite small-changed their main activity between 1992 and 1993. Patterns of diversification over time can be interesting to policyinakers; the popularity of a given activity may indicate which firms should be followed up by tax authorities, which activities should receive priority attention for simplifica- 434 THE WORLD) HANK lC;ONOMIC RIlVIEW, VOL. 9, NO. X tion and clarification of regulations, and which entrepreneurs might require as- sistance during the transition. For example, the fact that no firms diversified into consumer services but several firms diversified out suggests that consumer services should not be high on the tax collector's list. Also, the fact that research and development firms continued to diversify out of business services and into trading or manufacturing and construction suggests that they might require as- sistance during the transition. The majority of entrepreneurs engaged at some point in trading, suggesting that substantial gains in efficiency could be achieved by simplifying and clarify- ing trade regulations. In 1992 about 33 percent of survey firms moved into trading, or new areas of trading, and only 5 percent curtailed overall trade op- erations. Within trading, entrepreneurs moved away from wholesale trading in foodstuffs, for which margins had narrowed and supplies-especially from Ukraine-had been disrupted. In 1993, 25 percent of firms initiated or expanded trading activities, but 10 percent curtailed or stopped such activities. Trading vas still profitable but was no longer the only lucrative pursuit. The diminishing attraction of trading is consistent with developments elsewhere. In Poland, for example, the speculative gains in trading in 1990 and 1991 were disappearing by late 1992. Financial services continued to be profitable over the two-year period, and interest in real estate and related construction and renovation activities was grow- ing strongly in 1993, again suggesting areas of focus for authorities. About 10 percent of survey firms started or expanded some real estate activity: real estate services, building repair, building construction, or architectural design. Only one firm cut back on its real estate activities. About 10 percent of firms also diversified out of services into manufacturing or construction, but this trend is difficult to interpret without indications of diversification rates in the opposite direction. Financing. All over the world, small and medium-size enterprises obtain their financing primarily from the personal resources of their founders and reinvested profits, and only secondarily from bank credit. This is true for St. Petersburg entrepreneurs, wlho as a group rely very little on bank credit. On average, firms were established mainly with the owner's personal financial resources, supplemented in part by early profits and loans from family and friends. Working capital and investments have been financed largely from internal profits. Bank credits contributed only 7 to 11 percent of financing-compared with an estimated one-third to a half in the United States (Storey 1982). Government programs contributed only 1 to 2 percent. The median nominal interest rate paid by survey entrepreneurs in 1992 was 120 percent, with inflation running at an annual rate of over 2,000 percent. In reality, real interest rates were much higher than nominal rates would imply because of discounting up front, extra fees, and bribes, but they were still prob- ably strongly negative in real terms. Not one entrepreneur had a loan with a de Melo, Ofer, and Sandler 435 maturity beyond twelve months, although several said their short-term credit lines could be rolled over. A number of entrepreneurs said that available loans were too short-term to interest them. In 1992 only 22 percent of St. Petersburg service firms had bank credits, and this share dropped to 18 percent in 1993. These figures can be compared with data from other surveys in late 1992 and early 1993 showing 40 percent of private firms receiving bank credit in Poland (Wyznikiewicz, Pinto, and Grabowski 1993) and 50 percent in Ukraine (Stone and Novitskaya 1993). In both cases, bank credit covered only a small share of total financing needs, but credit access was clearly higher than for the St. Petersburg entrepreneurs in ser- vices. Access to bank credit may have been lower in the Russian Federation than elsewhere for the new private sector at this time, but it appears also that banks prefer to lend to manufacturing firms-because collateral is available, because the owners have better personal connections with bIanks, and because manufac- turing firms are still viewed as socially more useful. The reason for the drop in the share of firms with bank credit from 1992 to 1993 was not stricter lending criteria. Rather, it seems to have been the entrepre- neurs' reluctance to borrow, given higher-albeit still negative-interest rates and uncertainty about future input and output prices. According to the 1993 survey data, nominal interest rates had risen to 250 percent or more on an an- nual basis, with inflation running at about 1,000 percent, and entrepreneurs said the maximum annual interest rate they would pay for a ruble loan was from 100 to 300 percent. In other surveys undertaken around this time, Stone and Novitskaya (1993) and Johnson and Loveman (1995) found that private entre- preneurs in Ukraine and Poland were also reluctant to borrow at high nominal interest rates, even when real interest rates were low or negative. In Ukraine, the main explanation was uncertainty, as here; in Poland, it was cash flow problems arising from early repayment of principle under high nominal rates. Main problems. Entrepreneurs were asked about the main problems faced by their firms. Table 5 shows that in 1992 the most frequently mentioned problem- especially by firms in trade and financial services-was the instability and uncertainty in the macroeconomic and legal environment, including crime and corruption. (This was also true for private entrepreneurs in Ukraine, accord- ing to Stone and Novitskaya 1993. In Poland, where inflation had moderated by late 1992, entrepreneurs' biggest complaint was lack of financing.) Almost half the trading firms in the 1992 St. Petersburg survey were subject to pressure from the mafia in deciding from whom to buy, where to sell, and how to price their products. Problems of financing and taxes ranked next in importance. In 1993 the same three problems were dominant. Most firms said the power of the mafia was as strong or stronger than before, and taxes were an increasing problem for firms remaining in trade and consumer services. Trading firms were subject to higher tax rates than were firms engaged in other activities, and con- sumer services were still suffering from low demand. Financing was the main 436 FHF. WORLD BANK ECONOMIC RFVIEW, VOl. 9. No. 1 Table 5. Main Problems of Service Firms in St. Petersburg, 1992 and 1993 (percentage of firms) Workspace Main activity Instability Finance Taxes Demand and utilities Other 1992 Trade 38 28 14 3 0 17 Business services 24 28 28 12 4 4 Consumer services 25 8 25 8 8 24 Health and education 17 50 0 17 0 17 Transportation 0 50 25 25 0 0 Finance 67 0 0 0 33 0 Total 29 27 19 9 4 13 1993 Trade 19 19 33 0 11 18 Business services 29 19 19 5 5 25 Consumer services 20 20 40 0 20 0 Health and education 33 33 0 0 17 17 Transportation 0 75 0 25 0 0 Finance 25 ( 25 0 0 . 0 Manufacturing and construction 29 43 0 0 14 14 Total 23 24 23 3 10 19 Note: Data are based on seventv-nine firts in each year. Percentages may not add to 100 because of rounding. Source: Authors' calculations. problem for half the firms changing activity between 1992 and 1993, and was a prime reason for the change. For most firms, however, financing was needed for expansion, not survival. Other problems mentioned were low demand (caused in part by a shortage of cash in state organizations, the closing of state enterprises, increased competi- tion, the firm's own low-quality products, and market domination by more es- tablished firms); lack of affordable workspace with adequate utilities; a scarcity of good workers; delays in banking transactions, especially for transfers outside the city limits; and unsatisfactory business relations-especially difficulties in finding honest partners. These last two problems particularly affected entrepre- neurs in trade and business services. Markets In fully functioning market economies, there are numerous examples of mar- ket failure. Until recently, however, markets did not exist at all in the Russian Federation. For new private firms there, the problem is not market failure but market creation and development. As shown in table 6, the firms' market orientation varied according to the back- ground of owners. Former managers were more likely to deal with the state sector and households. Former academics and technical specialists were more likely to deal with other private firms and to import current inputs, and they were the only own- de Melo, Ofer, and Sandler 437 Table 6. Market Orientation of Service Firms in St. Petersburg by Type of Ownzer, 1992 (average percentage share) Owner's background Academic or technical Market orientation Manager specialist Worker Total Customers Public administration 1 8 4 State enterprises 27 27 X 24 Private companlies 20 43 31 31 Households and individuials 51 19 52 39 Foreigners ( 5 0 2 Suppliers State enterprises 71 49 36 58 Private companies 13 28 49 23 Individuals 9 14 14 12 Foreigners 6 9 1 7 Note: Data tor customers are based on seventy-eight firms, and data for suppliers are based on fifty- six firms. Source: Authors' CalCLIlarltois. ers to export their services. Former workers were the least likely to deal with state enterprises, presumably because they had few useful contacts; they relied on private companies as suppliers and households as customers. For all the survey firms, state enterprises were clearly the most important input suppliers, although they were less important than for private manufactur- ing firms in Russia and Czechoslovakia in the early 1990s (see Webster and Charap 1993 for Russia and McDermott and Mejstrik 1992 for Czechoslova- kia). Customers were more diverse than suppliers; and sales concentration was moderate, with less than 15 percent of firms receiving more than half their rev- enues from a single customer, typically a state enterprise. Most entrepreneurs sold their services mainly locally and in surrounding towns and looked to their competitors for pricing guidelines (for further findings on wages and prices in private service firms, see de Melo and Ofer 1994). Business services were the most dependent on a single large customer, with 70 percent receiving a quarter or more of their revenues from one source. In 1993, survey firms-especially those whose main customer in 1992 was the state sector-made a clear shift away from state sector customers and toward private firms and households. One explanation was the failure of public enterprises to pay their bills; another was the drop in state sector output. The low reported average export share in sales reflects the reluctance of sev- eral trading firms engaged in the export of raw materials (oil, timber, and met- als) to talk about these operations, which may not have been legal. In response to more general questions, 25 percent of the firms said they were engaged in some form of export, and 5 percent identified exports as their primary sales 438 THF WORLD BANK F(jONOMIC REVIFW, VOL. 9, NO. . market. Of exporters who spoke freely, half were traders, acting as intermediaries; the others provided exports of tourism, transportation, design, and information services. Most firms exported to the former Soviet states; only a quarter ex- ported elsewhere. Constraints to exporting were said to include payment delays, government bureaucracy, and marketing, but exporting was also discouraged by foreign exchange surrender requirements, unfavorable foreign exchange con- version rates by the central bank, capital gains tax on increases in the domestic currency value of foreign currency deposits, high bank fees on foreign exchange transactions, and visibility with mafia groups. Despite these impediments, 80 percent of exporting firms increased exports in 1992; in 1993, 40 percent in- creased exports. Competition was moderate and increasing. In 1992 about 50 percent of the entrepreneurs said there were ten or more companies-typically small, private firms-competing in their output market. Less than 10 percent said they had no competition. Most entrepreneurs felt the pressure of competition and reacted by reducing prices in real terms, improving quality, and diversifying. Other strate- gies included cutting costs, increasing market share by expansion, and agreeing on market shares with competitors. In 1993, some 70 percent of firms dealing in all markets-real estate, capital equipment, material inputs, service inputs, and service sales-found them more competitive than a year earlier. Others found conditions the same. Only a few found markets less competitive. 111. FIRM PERFORMANCE The firms' performance was evaluated to establish how well the new pioneers for profit are doing and to gain some insight into the determinants of success. Performance Indicators The interviewer asked the entrepreneur for a qualitative assessment of five performance indicators: current profits, growth of profits, growth of real sales, expansion plans, and capacity utilization.4 A sixth performance indicator was developed from quantitative data on employment, on the assumption that grow- ing employment signals success. In the original survey, an overall grade of I to 4 was assigned to each firm by the interviewer. It was based partly on the performance indicators and partly on a qualitative impression of the firm, taking into account the qualities of the entre- preneur and a personal inspection of the facilities. Three-quarters of firms were classified as moderately successful (grade 3) or highly successful (grade 4). A comparable interviewer grade was not possible in 1993 because the follow-up survey was conducted by telephone. Instead, the owner was asked to compare the overall well-being of the firm in 1993 with that for the period ending in 1992. 4. The original survey also asked entrepreneurs about their share of new equipment and found that one-third had 100 percent new equipment and over half had more than 50 percent new equipment. de Melo, Ofer, and Sandier 439 By this evaluation, only 15 percent noted a deterioration in their firm's general well-being during 1993 (grade 1), and more than 50 percent cited an improve- ment (grades 3 and 4). A composite grade was calculated each year from the available overall grade and five of the performance indicators (see table 7). Ca- pacity utilization, because it did not show positive and significant intercorrelation with the other indicators, was excluded from the composite grade. Table 7. Performance of Service Firms in St. Petersburg, 1992 and 1993 (percentage of firms) Indicator 1992 1993 Cuirrent profits Negative 4 5 Break even 12 10 Small profits 49 60 Large profits 35 25 Grouwth of profits Less profitable 26 22 About the same 21 29 More profitable 53 49 Growth of real sales Decreased 19 12 No significant change 18 13 Increased 63 75 Expansion plans Will contract 0 0 No change 14 13 Will expand 86 87 Capacity utilization Capacity underutilized 19 20 Near full capacity 64 72 Capacity constrained 16 8 Growth in employment Decreased 8 35 Stayed the same 27 4 Increased up to and including the median 15 10 Increased above the median 50 51 Composite grade 1 = lowest 0 0 2 15 10 3 44 54 4 = highest 41 36 Note: Data are based on seventy-one to eighty-one firmis, except for those on employment growth in 1992, which are based on fifty-two firms. Performance indicators are qualitative except for employment growth, which is based on number of employees. Source: Authors calculations. 440 THE WORLD BANK ECONOMIC REVIEW, VOL. 9, NO. i The most significant correlations among the five intercorrelated performance indicators are between expansion plans on the one hand and current profits and growth of sales on the other, and between growth of sales and growth of profits. In 1992 the five intercorrelated performance indicators are also strongly corre- lated with the overall grade; in 1993, equivalent correlations are positive but somewhat weaker. All performance indicators are significantly correlated with the composite grade, and most of the correlations are quite strong, suggesting that the composite grade is a good indicator of performance. Although the cal- culated composite grade remained at the same level in 1993 as in 1992, the inclusion of dynamic performance indicators such as growth in profits, sales, and employment means that the same rating is a sign of continued improvement. Most of the firms that responded that there was no change in 1993 were already successful in 1992. Together, the performance indicators and overall grades give the impression of good and improving performance. This impression is reinforced by the oral testimonies of many owners, who spoke of improvement and expansion in a variety of directions: internal organizational changes, stabilization of working patterns, greater specialization, improvements in quality, diversification into promising new areas, investment in equipment and real estate, improvements in the skill levels and work ethic of employees, development of good client rela- tions, and expansion of markets and business connections-including those out- side the Russian Federation. Some even cited improvement in relations with vari- ous government officials, including tax authorities. An interesting question is whether particular indicators or overall grades can predict subsequent performance. If so, surveys could be used to anticipate prob- lems. With this in mind, we examined cross-period correlations. Most cross- period correlations were found to be either significantly lower or weaker than correlations within periods, but a few of the indicators were significantly corre- lated across periods. Expansion plans in 1 992 appeared to be a particularlv good predictor of 1993 performance, especially as measured by growth of sales. In fact, 80 percent of firms with concrete expansion plans in 1992 experienced an increase in real sales in 1993, and 90 percent of firms experiencing an in- crease in real sales had concrete expansion plans in the previous period. Several factors might explain the lack of strong intertemporal correlation among other variables. The unstable environment might result in radical changes in the firms' financial positions in the short run, and short-term cycles in firm performance might lead to inevitable corrections of short-term deviations from longer-term trends. Employment growth-the only performance indicator that was quantified- exhibited the lowest and least significant levels of correlation with other indica- tors. This is not surprising. Although expanding employment can be a clear sign of success, some firms may have raised profits and productivity by reducing employment. Also, employment growth may cut into current profits, even if it leads to higher profits over time. de Melo, Ofer, and Sandier 441 Table 8. Enmployment Growth in Service Firms in St. Petersburg by Various Indicators, 1992 and 1993 (annualized percent change) Full-time workers All workers, Indicator 1992 1993 All firms 28 65 Main activity in 1992 Trade 140 36 Business services 1 56 26 Consumer services 0 37 Health and education 116 104 Transportation -3 107 Finance 5(07 35 Firms that changed activity in 1993 18 -2 Firms that did not change activity in 1993 28 71 Main activity rin 1993 Trade 139 14 Business services 209 96 Consumer services 0 0 Health and education 124 93 Transportation -3 107 Finance 232 117 Manufacturing and construction I I -41 Numnber o zworkers in the firm in 1992 Less than 8 42 162 8 to 39 34 -20 40 to 99 79 3 100) or niore 22 91 Firrm history New 103 40 Completely privatized 110 60 Privatized with residual state ownership 3 81 Note: Data are hased on forty-nine firms. a. Includes the full-time equivalent of part-time workers. Source: Authors' calculations. Table 8 shows that employment of full-time workers in survey firms increased at an annual rate of 28 percent in 1992, and employment of all workers in- creased at an annual rate of 65 percent in 1993.5 This rapid growth indicates that private services are making an important social contribution by providing employment for state sector employees who lose their jobs. Analysis of the compo- nents of employment growth in 1993 shows that reliance on part-time employ- ees declined. Firms hired a higher share of full-time workers, reflecting perhaps 5. Discussion is based on forty-nine firms that provided answers for both 1992 and 1993. The definition of employment changes because growth for all employees is not available for the earlier period, yet this measure is preferred. The full-time equivalent of parr-time workers is calculated as half-time where informatio>n on hours per week is not available. 442 THF WORLD BANK ECONOMIC REVIFW, VOL. 9, No. 3 a growing appreciation of their own investment in the human capital of their employees but also the convenience of hiring on long-term contract rather than permanent hire. Contract hiring, which provides substantial flexibility to the employer but reduces workers' job security, increased significantly between 1992 and 1993. Based on the categorization of firms by main activity in 1992, activity- specific growth rates varied widely in 1992 and in 1993. In 1992, employ- ment growth was concentrated in business services, trade, and finance-that is, those activities that were the most repressed under communism. A period of consolidation with lower growth followed in 1993. Firms providing pub- lic-type services (health and education) increased employment in both peri- ods. Based on1 the categorization of firms by main activity in 1993, employ- ment in trade slowed noticeably, but in business services and finance it remained high, reflecting the diversification trends discussed earlier. Firms engaged in consumer services in both 1992 and 1993 experienced zero growth, reflecting low demand; however, firms that diversified out of consumer ser- vices increased their employment. Employment growth in 1992 was concentrated in medium-size firms (with forty to ninety-nine workers). In 1993 these firms stabilized, and growth was generated by either very small, new firms or large firms with 100 or more employees. From the perspective of ownership, employment growth was more rapid over the two periods in new firms; this is understandable because such firms often started small. Employ- ment growth in completely privatized firms was much higher in 1992 than in privatized firms with residual state ownership, probably because these latter firms were ten times larger and hence more likely to have surplus labor. Why Such Good Performance? There are several possible explanations for the good overall performance of the firms surveyed. Two possible explanations are particularly difficult to gauge but could be important. First, there may be biases in the sample selection. Suc- cessful firms may have been more likely to respond to the interview request letter. Also, a positive performance bias may prevail among registered firms, from which the sample was drawn, with much of the failure and instability oc- curring among the undocumented population of unregistered firms. Both sources of selection bias are possible but difficult to verify. Second, despite the general negative effects of inflation, some firms, especially in trade and financial ser- vices, may have profited from it. Other possible explanations that may be important are listed below: * There is a substantial gap between desired and actual levels of many services-such as trade, financial services, and business services-and new private service firms have an opportunity to fill this gap. Despite declines in real incomes, this gap exists because such services were repressed under the previous regime and also because the centralized trading system collapsed. de Melo, Ofer, and Sandler 443 Service firms in St. Petersburg are performing somewhat better than comparable manufacturing firms, giving support to this explanation (see de Melo and Ofer 1994). The private sector is able to attract high-quality, quick-learning entre- peneurs, given the disarray in public administration, budget-supported government agencies, and academic institutions. These entrepreneurs have formed a critical mass that permits some firms to operate successfully by dealing primarily with other private businesses. Their capabilities do not imply that the initial quality of services was necessarily high but rather that their learning curve is steep. In fact, several entrepreneurs claimed improvements in the quality of their product and the work habits of employees in the short period between surveys. * New, smaller, private firms have an advantage over large, established firms, which are less flexible, may have managers that are less motivated, and may be burdened with social expenditures, including expenditures on housing. A recent survey of large, privatized firms in Moscow and Vladimir oblasts supports this explanation (Webster and others 1994). * The environment is one in which competition, although increasing, has not yet reduced margins to the minimum, and so companies are able to change activities and to absorb bad debts, loss of customers, and other hazards without incurring bankruptcy. Wyznikiewicz, Pinto, and Grabowski (1993) provides survey evidence on the decline in trading margins in Poland, which in 1992 and 1993 was further along in the transition process than the Russian Federation was. An interesting comparison can be made with registered private firms in Ukraine, where 20 percent of firms surveyed in April 1992 were found to have stopped operation (no reason given) eight months later. The Ukraine survey covered manufacturing firms and included a higher percentage of small firms (defined as having less than twenty-five employees) and a much higher percentage of firms with bank credit. It also had less scope for selection bias. Otherwise, the charac- teristics of firms, the entrepreneurs, and their environment appear to be similar (see Stone and Novitskaya 1993). Internal and External Factors Affecting Performance Internal and external factors associated with firm performance include activ- ity, owner's background, firm history, firm size, output market, bank financing, and main problems imposed by the business environment. These are discussed below. Other factors-such as age of firms and entrepreneurs, gender of owner, owner's previous workplace, size of entrepreneur's ownership share, and share of professionals in the labor force-were uncorrelated with performance indica- tors in both periods and are therefore not discussed. It is also of interest to note that service entrepreneurs in St. Petersburg have had very little contact, not to mention assistance, from government agencies or Western firms. 444 THE WORLD BANK EC:ONOMI( REVIEW, VOL. 9. No. 1 An analysis of the relation between performance and main activity shows that by far the best performance in both periods was demonstrated by financial firms, followed by trade and business service firms. The weakest showing was by con- sumer service firms, which suffered from declining demand for such services as hairdressing, tailoring, and watch repair. An analysis of performance by type of entrepreneur suggests that general char- acteristics of entrepreneurship and capabilities, rather than specific managerial experience or skills, are important. Thus, two characteristics showing signifi- cant correlation with firm performance during 1992 were the number of other firms owned by the entrepreneur and the entrepreneur's level of education. In- terestingly, the entrepreneur's level of satisfaction with the firm's employees was negatively correlated with performance. One interpretation is that an entrepreneur's ability to recognize labor problems was a sign of professionalism. In several successful firms, including a private hospital and an investment firm, the owners were concerned with worker performance and instituted very strict hiring policies. In addition, there was an association between firm performance levels and the entrepreneur's background, as shown in table 9. Firms run by former academics or technical specialists performed better than the other two groups by most criteria during both periods. This was true even though firms run by former state managers were more concentrated in trade, a more profitable activity, and even though many state managers had previous experience in Table 9. Performance of Service Firms in St. Petersburg bv Owner's Background, 1992 and 1993 (average composite grade) Owner's background Academic or technical Indicator Manager specialist Worker Total 1992 Current profits 3.1 3.3 2.9 3.2 Growth of profits 2.8 .3.1 2.8 2.9 Growth of sales .3.0 3.4 2.8 3.2 Growth in emplovment 2.7 3.5 3.3 3.1 Expansion plans 3.8 3.8 3.7 3.8 1993 Current profits 3.1 3.0 3.0 3.1 Growth of profits 2.7 3.1 3.2 2.9 Growth of sales 3.4 3.7 2.9 3.5 Growth in employment 2.3 3.2 3.3 2.8 Expansion plans 3.8 3.9 3.7 2.8 Note: The performance indicator values are scaled from a low of I to a high of 4. Data are based on seventy-three to eighty-one firms in 1992 and seventy-one to eighty-one firms in 1993, except those for employment growth, which are based on fifty-two firms. Source: Authors' calculations. de Melo. Ofer, and Sandier 445 service activities. This finding suggests that in services the insider status may not have the same importance it has for private firms in manufacturing (Webster and Charap 1993). Also, continuity may reflect the conservative nature of traditional managers, but services, which are strongly customer- oriented, require a radical departure from previous practices. Another pos- sible explanation is that the analytical skills acquired by former specialists give them a comparative advantage in confronting the uncharted waters of the private service sector. There was also some variation in performance by firm ownership. In 1992 privatized firms ranked lower than new firms. These differences narrowed dur- ing 1993 as some privatized firms cut their large, inherited work forces. Also, within the category of privatized firms, those with residual state ownership im- proved their rankings in 1993, which suggests that they may have benefited from special privileges given to state enterprises. According to almost all indicators, large firms, as defined by revenue, per- formed consistently better than small firms did, and this was true for both sur- vey periods (table 10). The better performance of large firms may be because higher revenues are associated with financial and other resources that help firms to deal with the bureaucracy, including tax authorities, and to obtain credit. Large firms, as defined by number of employees, also performed better than small firms, although the relationship is somewlhat weaker. Table 10. Performance of Service Firms in St. Petersburg by Revienue, Customers, and Credit History, 1992 and 1993 (composite grade) Indicator 1992 1993 Monthly revenue (1,000 rubles) Less than 100 2.5 2.5 100 to 999 2.9 3.1 1,000 to 9.999 .3.3 3.2 10,000 or more 3.5 3.3 Total 3.2 3.2 Main customer State sector 3.1 3.3 Private firms 3.5 3.3 Households 2.8 3.2 Total 3.1 3.2 Credit htstorv No credit 3.1 3.2 Received credit 3.5 3.3 Note: Data for monthly revenue are based on eighty-one firms in both vears; data for main customer on fifty-six firms: and data for credit history on eighty firms. Monthlv revenue categories are adjusted for inflation in 1993: see table 4 for ruble and dollar values. The composite grade runs from I (the lowest) to 4 (the highest); grades shown are averages for the group. Source: Authors' calculations. 446 TiiE WORI) BANK F(IONOMIC RFVIEW, VOL. 9, N). ; Data from the original survey indicated that firms selling to other private firms performed better than those selling to the state (table 10). Firms selling to households performed less well than the other two groups and were typically privatized, had a low share of professional workers, had low employment growth, and were run by former state enterprise managers who were continuing in the same activity, had less education, and owned no other companies. The conven- tional wisdom is that private companies working with the state are in a position to take advantage of rents created by distortions and to make lucrative profits. But the evidence on this is mixed. In 1992, firms selling mainly to the state sector had lower overall perfor- mance than those selling mainly to other sectors. But in 1993, firms reducing their share of sales to the state showed weaker performance, and firms main- taining or increasing their sales to the state performed equally to or better than firms selling mainily to other sectors. The association between exports and over- all performance is clearer. The seven firms that increased their export sales share over time performed better than average in both 1992 and 1993, and the four firms that reduced export shares were consistently weak overall performers (not shown). In 1992, firms that received bank credit were more successful, but not in 1993 (table 10). One reason may be that performance moderated in trading firms, which are often able to obtain credit by using traded goods as collateral. Another may be that real interest rates were significantly higher in 1993 than in 1992, with a resulting reduction in implicit subsidies to borrowers. To the ex- tent that a positive correlation between performanice and bank credit does exist, we would like to think that the causality runs from the quality of the firm and the soundness of its expansion plans to its ability to receive credit. However, reverse causality, where the independent ability to receive cheap bank credits helps to improve performance, cannot be ruled out-particularly if real interest rates were negative, as they appear to have been. In some cases, obtaining credit may simply reflect the cleverness of the owner who perceives a substantial inter- est rate subsidy. Alternatively, some entrepreneurs may have received bank credit through corruption or personal connections. Receiving credit was correlated in the survey with the use of hribes and with owners who previously worked in the state sector. Table 11 shows how the main problems varied with the firms' composite grades. In both 1992 and 1993, macroeconomic and legal instability was the most frequently mentioned problem for strong firms. Financing was an important con- straint for both moderate and strong firms, suggesting that improved access to credit could foster more rapid private sector growth. Taxes were mentioned with higher frequency in 1993 than in 1992 by all categories of firms, and the share of weak firms that mentioned taxes as their main constraint increased from one- third to a half. The importance of taxes to weak firms may be explained by their severe shortage of funds, a shortage that could be eased by lower taxes but not by credit, because the firms were ineligible for credit. Because strong firms were de Melo, Ofer, and Sandier 447 Table 11. Performance of Service Firms in St. Petersburg by Main Problems, 1992 and 1993 (percentage of firms) Performance Workspace and level Instability Finance Taxes Demand utilities Other 1992 Weak 8.3 25. 0 33. 2*5.0 0.0 8.3 Moderate 33.3 19.4 19.4 8.3 5.6 14.0 Strong 32.3 35.5 12.9 3.2 3.2 12.8 Total 29.1 26.6 19.( 8.9 3.8 16.5 1993 Weak 12.5 12.5 50.0 0.0 0.0 25.0 Moderate 19.1 26.2 21.4 4.8 9.5 19.1 Strong 31.0 24.1 17.2 0.0 13.8 13.9 Total 22.8 24.1 22.8 2.5 10.1 17.7 Note: Data are based on seventy-nine firms for both years. x values: 1993 =0.39; 1992= 0.54. Weak firms have composite grades less than 2.5; moderate firms, between 2.5 and 3.5: and strong firms, over 3.5. Source: Authors' calculationis. relatively liquid-and possibly because they were able to avoid or evade taxes- they complained less about high taxes than weak firms did. IV. CONCLUSIONS AND POLICY IMPLICATIONS The newly emerging private sector in St. Petersburg is developing despite an uncertain and unstable political and economic environment. It is coping with both the short-term shock effect of disrupted trade, declining output, and social disorder and the longer-term transition effect of changing ownership, laws, and institutions associated with moving from a centrally planned to a market economy. In 1992, entrepreneurs were exploring their options and looking for a niche. In 1993 the pace of diversification moderated, and output markets be- came increasingly dominated by the private sector. Firms received little finan- cial support in either year and complained about taxes and lack of financing. Firm performance was on average quite good from the time firms were estab- lished until the end of 1992, and this performance improved in 1993. Explana- tions for the good performance cannot be found in any dramatic improvement in the Russian Federation's economic, legal, administrative, or political envi- ronment during 1 993. Inflation continued at an annual rate of close to 1,000 percent, production continued to deteriorate, and political instability persisted. Rather, the explanation for good performance seems to lie partly in the gap between demand and the previous dearth of services-especially trade, business services, and financial services-and partly in the personal qualities of the St. Petersburg pioneers for profit-their high level of education, technical experience, and entrepreneurial attributes. 448 THE WORLD BANK ECONOMIC REVIEW, VOL. 9, NO. x A reasonable prognosis for services would be continuing growth with lower profit margins. On the one hand, the lack of a well-developed service sector has an advantage; the development of small and medium-size firms has not been blocked by large, preexisting firms whose managers collude with politicians and the bureaucracy to preserve their economic monopolies. On the other hand, profit margins will be squeezed as competition continues to increase; private service entrepreneurs may find it more difficult to survive. (See McDermott and Mejstrik (1992) for an account of difficulties faced by small and medium- size manufacturing enterprises in Czechoslovakia during the early stages of tran- sition.) The survey findings have several clear policy implications. First is the need to establish a stable macroeconomic environment-an objective frequently men- tioned by the entrepreneurs surveyed. Students of private sector development in Eastern Europe single out relative stability as the key element in private sector development in Poland, Hungary, and the Czech Republic and instability as the key element in delaying such growth in Romania, Bulgaria, and the Russian Federation. See Earle and others (1994) and Johnson and Loveman (1995). En- trepreneurs' aversion to high interest rates is one reason why rapid and broadly based stabilization is necessary. Second, as part of the general business environment, a stable legal framework is also important, together with transparent and evenhanded implementation of government regulations. Complex or changing laws and procedures, confusing or conflicting instructions, favoritism, and bribe taking all work against small and medium-size firms that have less financial and human resources to deal with these obstacles than large firms do. Aside from regulations covering the entry and operation of businesses, those pertaining to the rent and renovation of real estate are important. Third, given the relatively low access of small and medium-size firms to bank credit, and the possible conflict between credit expansion and macroeconomic stabilization, the next priority is arguably the creation of a well-designed tax code that permits and encourages self-financing. Private sector activities are expanding and clearly must form an important part of the tax base. But taxes need to be rationalized (for example, discriminatory treatment of trade and other services should be eliminated) and set at levels that are moderate enough to permit substantial reinvestment of profits. Specific tax relief for small and medium-size enterprises-perhaps on reinvestment of a limited amount of nomi- nal profits-would allow these firms to grow even more quickly, generating more jobs and providing the distribution and business service infrastructure re- quired to stimulate and facilitate private sector development in agriculture and manufacturing. Fourth, continued privatization of real estate is an extremely important fac- tor in private sector development. The experience in Eastern Europe demon- strates that the transfer of ownership in real estate assets to small private busi- nesses may be the most important aspect of the privatization process-more de Melo, Ofer, and Sandler 449 important than the transfer of the business itself (Earle and others 1994). The slowness of real estate privatization in the Russian Federation is no doubt a major barrier to faster growth of private firms. Fifth, private firms need better access to financing in order to grow rapidly. This requires actions on both the demand and supply sides. Banks need to develop procedures and personnel to evaluate credit applications, and firms need to develop good business proposals. In regard to the latter, direct assistance would be helpful, although important issues arise on how such programs are designed and delivered. See de Melo and Ofer (1994), Kondratowicz and Maciejewski (1994), Piasecki (1992), and de Koning and Snijders (1992) for a variety of specific policy recommendations for direct assistance programs. At the time of the St. Petersburg surveys, the private sector in the Russian Federation was growing without the support of the government or of Western business. Direct assistance may not have priority over macroeconomic stabiliza- tion, but if such assistance is warranted in established market economies, it is clearly justified in the Russian Federation, where the market infrastructure is still weak and private growth so important. Given the high education level of the St. Petersburg entrepreneurs, information, counseling, and training should have a good payoff there. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Acs, Zoltan J., and David B. Audretsch. 1993. Small Firms and Entrepreneurship: An East-West Perspective. New York: Cambridge University Press. Brown, Annette, Barry Ickes, and Randi Ryterman. 1994. "The Myth of Monopoly: A New View of Industrial Structure." World Bank Policy Research Working Paper 1331. World Bank, Policy Research Department, Washington, D.C. Processed. de Koning, Aad, and Jacqueline Snijders. 1992. "Policy on Small- and Medium-sized Enterprises in Countries of the European Community." International Small Busi- ness /ournal 10(3):25-39. de Melo, Martha, and Gur Ofer. 1994. Private Service Firms in a Transitional Economy: Findings of a Survey in St. Petersburg. Studies of Economies in Transformation 11. Washington, D.C.: World Bank. Earle, John S., Roman Frydman, Andrzej Rapaczynski, and Joel Turkewitz. 1994. Small Privatization: The Transformation of Retail Trade and Consumer Services in the Czech Republic, Hungary, and Poland. Prague: Central European University Press. Easterly, William, Martha de Melo, and Gur Ofer. 1994. "Services as a Major Source of Growth in Russia and Other Former Soviet States." Policy Research Working Paper 1292. World Bank, Policy Research Department, Washington, D.C. Processed. Giarini, Orio, ed. 1987. The Emerging Service Economy. Oxford: Pergamon Press. Goskomstat. Various issues. Social/Economic Situation of Russia. Moscow: Goskomstat. Johnson, Simon, and Gary L.oveman. 1995. Starting Over in Eastern Europe: Entrepre- neurship and Economic Renewal. Boston: Harvard Business School Press. 450 THE WORLD BANK FE(CONOMIC REVIEW. V(Ol. 9, NO. I Kondratowicz, Andrzej, and Wo ciech Maciejewski. 1994. Small and Medium Private Enterprises in Poland. Warsaw: Adam Smith Research Center. McDermott, Gerald A., and Michal Meistrik. 1992. "The Role of Small Firms in the Industrial D)evelopment and Transformation of Czechoslovakia." Small Business Economics (4): 179-200. McKay, John. 1970. Pioneers for Profit: Foreign Entrepreneurship and Russian Indus- trialization, 188.5-191.3. Chicago: University of Chicago Press. Piasecki, Bogdan, ed. 1992. Policy on Small and Medium-sized Enterprises in Central and Eastern European Countries: Survey and Comparatiue Study. Organizing Com- mittee, 19th International Small Business Congress, Warsaw. Stone, Andrew, and Irina Novitskaya. 1993. "Ukrainian Private Enterprise: Profiting against the Odds." World Bank, Private Sector Development Department, Washing- ton, D.C. Processed. Storey, David. 1982. Entrepeneurship and the New Firm. New York: Praeger Publishers. 1983. The Small Firm: An International Survey. New York: St. Martin's Press. 1989. The Performzance of Small Firms: Profits, Jobs, and Failures. London and New York: Routledge. St. Petersburg Statistical Office. 1993. Main Activity Indicators for Small Enterprises and Cooperatives in St. Petersburg and Leningrad Regiotz in 1992. St. Petersburg. Webster, Leila, and Joshua Charap. 1993. A Survey of Private Manufacturers in St. Petersburg. World Bank Technical Paper 228. Washington, D.C. Webster, Leila, with Juergen Franz, Igor Artimiev, and Harold Wackman. 1994. Newly Privatized Russian Enterprises. World Bank Technical Paper 241. Washington, D.C. Wyznikiewicz, Bogdan, Brian Pinto, and Maciej Grabowski. 1993. "Coping with Capi- talism: The New Polish Entrepreneurs." World Bank, Private Sector Development Department, Washingron, D.C. Processed. THE WORL D BANK FCO(NO)MI( REVIF'W. VOl . 9- N 0. O 45I-475 Apprenticeship Contracts, Small Enterprises, and Credit Markets in Ghana Ann D. Velenchik This article uses data from a 1992 survey of manufactutring enterprises in Ghana to describe the importance of apprenticeship in the manufacturing sector and analyze the structure of the contracts in wthich apprenticeship takes place. The article pre- sents three major findinigs. First, the training of apprentices is both a widespread activity and a part of the training of a large fraction of entrepreneurs and maanufac- turing workers. Second, two primary types of contracts are apparent in the data: those in which apprentices pay fees for their training and those in which they do not. Third, for those firms training apprentices, the choice of contract type is strongly correlated with other characteristics of the firm, particularly its usc of credit. Ap- prenticeship fees are one among many informal sources of firm finance. Many studies of manufacturing enterprises in Africa, particularly those focused on small firms, have noted the importance of apprenticeship in the backgrounds of individual entrepreneurs (House, Ikiara, and McCormick 1993; Harris 1971), as a component of current labor forces (Liedholm and Mead 1986), and as a mechanism for providing training (Squire 1981; Berry 1985). In most of these studies, however, discussions of apprenticeship are incidental to the main sub- ject, and attention has rarely been focused on the institution of apprenticeship itself. Although there is a small literature in education and anthropology focus- ing on the structure and educational content of individual firms' apprenticeship programs, the economics literature includes no systematic portrait of either the prevalence and scope of apprenticeship in manLfacturing or the microeconomic details of the contracts involved. This article begins the process of filling in both those gaps by providing a quantitative assessment of the importance of appren- ticeship in Ghanaian manufacturing and presenting an analysis of the contrac- tual relationships between apprentices and employers that characterize this institution. My analysis of the determinants of the structure of apprenticeship contracts places the institution of apprenticeship at the intersection of labor and financial Ann D. Velenchik is with the Department of Economics at Wellesley College. She would like to thank Tyler Biggs, Paul Collier, Tom Downies, Chris Udry, participants in the North East Universities Development Consortium Conference and the African Studies Association Meetings, and three anonymous referees for their comments and their insights on this issue. The research was conduIcted, in part, while the author was Visiting Scholar in the Department of Economics at Northwestern University. C 1995 The International Bank for Reconstruction and Development /THE WORI ni BANK 451 452 FHF W(ORI D BANK W(ONO1(1( RFVIEW, VOL. 9, No. I markets and highlights the relationship between firms' behavior in credit mar- kets and contract choice in the apprenticeship arena. The discussion extends the examination of contracting behavior in manufacturing and of African economic institutions more generally. The growing literature on economic institutions in developing countries provides a strong source of motivation for this work. Another motivation comes from the role of apprenticeship as a training insti- tution in Africa, where skilled labor is generally considered to be quite scarce. The past two decades have seen a dramatic expansion of educational systems throughout the continent (see Psacharopoulos and Woodhall 1985 and Knight and Sabot 1990), many of which have included the establishment of govern- ment vocational and technical schools geared toward providing skilled workers to the manufacturing sector. In many countries these schools exist alongside the traditional apprenticeship system. A great deal has been written about the ex- pansion of formal education in Africa, with particular attention on evaluating the impact of education on productivity, earnings, inequality, and growth (see Collier and La! 1 986; Hazlewood and others 1989; and Knight and Sabot 1990). Little is knowT1, however, about the role of apprenticeship in generating human capital for the manufacturing sector. Understanding the structure of apprentice- ship contracts is a first step toward analyzing how useful apprenticeship is in skill formatioln, productivity increase, and growth in the manufacturing sector. This article presents three major findings. First, the training of apprentices is both a widespread activity and a part of the training of a large fraction of entre- preneurs and manufacturing workers. Second, two primary types of contracts are apparent in the data: those in which apprentices pay fees for their training and those in which they do not. The third finding is that, for those firms training apprentices, the choice of contract type appears to be strongly correlated with other characteristics of the firm, particularly its use of credit. Because firms charging fees are more likely to use informal rather than formal sources of work- ing capital, apprenticeship fees are one among many informal sources of firm finance. As background for the discussions of the prevalence and structure of appren- ticeship contracts, section I presents a description of the data and the study from which it is drawn. A broad quantitative discussion of the incidence of apprentice- ship is presented in section II. A description of contracting forms, their impor- tance, and so 0, this type of financing arrangement is consistent with contract types in which no fees are paid and zero or positive wages are received during training. If MP() - aC < 0, the contract is one in which wages are zero and fees are paid on a continuous basis during training. No contracts of this type were observed in these data. 460 THE WORDI) BANK FCONOMIC REVIEW. VOL. 9, No. . With apprentice financing, the apprentice pays for the training before it be- gins. To finance the full amount, the apprentice pays a fee of aC at the beginning of period 0, and receives a wage of W0 = MP,) during the training period. The second-period wages are the same as under the pay-as-you-go strategy. The ap- prentice-financing strategy would be manifested as the contract type in which both fees and wages during training are positive. Apprentices may also choose to finance only part of their training in advance, paying for the rest through deductions from wages during the training period. In that case the apprentice pays a fee of F < aC before training begins and receives wages of W1) = MPo - aC + F during training. The post-training wages will be the same as in the strategies described above. This strategy is consistent with both contract types in which fees are paid, where the case with no wage payments is a special case in which MPo - aC + F = 0. Employer-financing strategies allow apprentices to pay their share of training costs by deductions from wages in the second period. Under full employer finan- cing, the apprentice bears none of the costs during training and receives a wage of WO = MPW. At the end of training the apprentice owes the firm aC(1 + r), which is paid as a deduction from the wage in the post-training period, which is then WI = MPo. If the apprentice bears part of the cost during training, the apprentice receives W1! > MPo - aC but less than MPo during training, at the end of which time the apprentice owes the firm laC - (MPO - W0)I(1 + r). Repay- ment through wage deductions in the post-training period yields W, = MPo + (1 + r)(MP,) - W9,). The second term reflects the fact that the worker receives the return on that fraction of the cost of training actually paid through wage deduction during the training period. This wage is less than the wage the worker would receive in alternative employment. Both apprentice- and employer-financing arrangements contain credit compo- nents and therefore give rise to default risk and moral hazard issues. These is- sues are not the same as those surrounding the sharing of investments in specific human capital. Those arrangements are made to avoid inefficient separations, but not failure to repay debts or to provide services that have already been paid for, as is the case regarding the financing of the apprentice's portion of the investment, which is the issue under consideration here. Mechanisms for deal- ing with default risk and moral hazard issues can be incorporated into contracts by reducing the incentives to default and providing contract enforcement mechan- isms to deter individually desirable breaches of contract. When apprentices finance their training by paying fees in advance, firms then have an incentive to fail to provide that training or to dismiss the apprentice before training is completed. One mechanism for reducing default incentives is for the apprentice to defer part of the payment of the fee to the end of the training period. In 40 percent of the firms that charge fees an average of 20 percent of the fees are deferred to the end of training. Older firms are less likely to have deferred fees, indicating that reputation factors may provide another disincentive for breach of contract. Although firms do not engage in repeated Velenchik 461 contracts with any individual apprentice, they do need to continue to attract apprentices and hence must worry about reputation effects. Firms for which apprenticeship is more important also may find reputation more important, which is verified by the finding that firms in which apprentices form a larger fraction of the labor force are less likely to receive deferred fees. Under employer financing, the apprentice has an incentive to borrow the train- ing costs from the firm and then leave for a higher-paying job without paying those costs back. It should be noted that this situation is different from that of the firm's having financed investment in specific human capital and wishing to retain the worker in order to appropriate the returns. In that case, the worker has not borrowed from the firm, and the worker's earnings would not be en- hanced by leaving, so there is no real problem of moral hazard. Under employer financing of general human capital, the worker can actually benefit from leav- ing before the debt is repaid. To prevent this occurrence, firms might require apprentices to post a performance bond at the beginning of the training period, to be refunded once the worker's debt has been paid off. Although such a scheme is possible, it seems unlikely, because an apprentice who could post a bond large enough to cover the debt would also be able to pay that amount as a fee, thereby financing that part of the training directly. An alternative type of bonding scheme is through deferred compensation, in which workers are paid lower wages dur- ing the training period and higher wages during the post-training period. By deferring some of the compensation to the later period, the firm induces the apprentice to stay. This mechanism would be indistinguishable in practice from shared investments in specific human capital, as both would generate upward- sloping wage profiles. Moral hazard problems may be diminished by external contract enforcement mechanisms. Agreements between apprentices and masters that generate a bond- ing of the apprentice to the master while also defining the master's obligations may be made verbally or in writing. External enforcement of these agreements may be provided by the legal system, trade unions, crafts associations, and other social networks. Nearly half of the firms in the sample have unionized labor forces or belong to trade associations that set explicit rules governing appren- ticeship. Among the firms in which fees are charged, it is common for either the beginning or the end of the apprenticeship, or both, to be marked by a ceremony hosted by the apprentice's family. This ceremony is generally the source of the in-kind fees recorded in the data and may be the reason for some of the deferred fees as well. The ceremony usually includes the provision of drinks by the ap- prentice and the pouring of libation to the ancestors. These ceremonies are at- tended by relatives, friends, and business associates of both the master and the apprentice. Masters and apprentices often come from the same social network, with 85 percent of the apprentices in the sample having found their masters through the recommendations of friends and relatives. The guests at the cere- mony, including the ancestors, are witnesses to the contract itself, and the social network and kinship ties provide a strong external enforcement mechanism. 462 IHE WORLF) BANK F( ONOMIC: RFVIEW, VOL. 4, No. . Apprenticeship contracts can be written to combine several of these financing strategies: the apprentice may finance part of the training through initial fees, part through deductions from wages, and part through lower wages in the post- training period. Although the characteristics of contracts can be observed, the combination of strategies underlying each contract cannot be. However, the circumstances under which each strategy might be expected to be chosen can be described and related to observed contract types. V. THE DETERMINANTS OF CONTRACT CHOICE The choice of financing strategies, and hence the choice of contract type, is influenced by the characteristics of the training itself as well as by at- tributes of the firms and individuals participating in the apprenticeship. The foregoing discussion indicated that a contract type that does not involve pay- ments of fees by apprentices can be generated by both pay-as-you-go and employer-financing strategies. Therefore, the only clear distinction between strategies that can be made through an examination of contract types is be- tween those that involve some apprentice financing through fee payments and those that do not. The following analysis consequently focuses on what determines the use of apprentice financinig as evidenced by the charging of fees. Before turning to a discussion of contract choice it is useful to present some evidence about the validity of interpreting apprenticeship fees as apprentice fi- nancing. The cost of training includes both time and equipment, and these data indicate that firms charging fees are more than three times as likely to require their apprentices to provide their own tools than are firms that do not charge fees. In providing their own tools, these apprentices are financing the capital costs of their training, so it seems reasonable to interpret the fees as payment for the labor and materials costs of the training. Contract choice should be influenced by the degree of specificity of the train- ing. The more specific the training, the lower the cost of training to be borne by the apprentice. This would make it easier for the apprentice to finance the train- ing through deductions from wages, as in the contract types without fees. More specific training also gives rise to steep wage profiles designed to induce the worker to stay with the firm. Such profiles would also be used when employers finance the acquisition of general training, which is the second strategy giving rise to contract types without fees. Thus, contracts without fees should, on av- erage, involve a greater fraction of specific training. The data set does not include information about the content of training; I use information about the post-training behavior of apprentices as a proxy. The more specific the training, the greater the probability that the apprentice will continue working for the firm after traininig is completed. If the training offered by firms that do not charge fees is indeed more specific, a larger fraction of those apprentices would continue in the firm. The data set provides three means Velenchik 46 3 Table 4. Contract Type and Workers Remaininig in Their Apprenticeship Firms, Ghana, 1992 (percent) Fee- No-fee- charging charging Item lirms firms Mean share of surveyed apprentices who finished training the previous year and stayed on at the firm 20 76 Mean share of current apprentices surveyed who plan to stay on after completing their training 12 So Share of surveyed workers who are working in the firm in which they did an apprenticeship and paid a fee for it 28 n.a. Share of surveyed wnrkers who are working in firms other than that in which they did an apprenticeship and paid a fee for it 55 na. n.a. Not applicable. Source: Author's caICUlatian1,. for examining this relationship, and the relevant data are presented in table 4. These data support the hypothesis that firms not charging fees offer more spe- cific training than do firms charging fees. It would also he expected that firms offering specific training or employer financing of general training would offer post-training earnings profiles designed to discourage quits. That is, more specific training should be associated with steeper earnings profiles providing higher returns to seniority. This is indeed the case for these data, in which the returns to an additional year of seniority among workers in firms in which apprentices are not charged fees is more than twice as high (4.6 percent compared with 2.0 percent) as in firms that charge fees.' This behavior is consistent with Hashimoto's (1981 ) argument about the sharing of investments in specific human capital and is consistent as well with the forego- ing discussion of employer financing. It should be noted that decisions about the type of training to offer and the financing strategy involved may be made simultanieously. A firm offering more specific training will bear a larger fraction of the costs, and the ability to do so may be driven by the same factors determining the firm's use of employer finan- cing of the worker's share of costs. Total training costs might also be relevant to contract choice. In particular, lower training costs make it more likely that the apprentice can pay through deductions from wages during the training period. There is n1o clear association, however, between the level of training costs and the choice between apprentice and employer financing. Once costs are too high to be borne easily as wage deductions, there is no obvious reason to expect 2. These estimates were obtained by estimating an earnings tunctioni for the 191 individuals in the sample of workers who are employed by firms currently training apprentices. The retirnis to experience were allowed to differ according to the fee-charginig practice of rth individtals' employers. t-rests indicate that the coefficienits arL different at a i percent significance level 464 THF WORL D RANK I CONOMI RFVIFW, VOL . , No). apprentice or employer financing to be more likely. Similarly, the duration of training, which might affect costs, is also influenced by financing decisions. Employer financing requires that the apprentice stay on after training is com- pleted and would give rise to longer measured apprenticeships for any given training costs. Although table 3 indicates that apprenticeships in which fees are charged are of longer average duration, this could be due to a number of factors, including the more general nature of the training, and cannot be clearly related to costs. In addition to the characteristics of training, contract choice is also influ- enced by the characteristics of firms and apprentices. The evidence in table 3 indicates that firms charging fees are significantly smaller on average than those not charging fees. Small size in itself, however, is not a characteristic that should exert independent influence on contract choice. Firm size is endogenous and determined bv a number of other factors that may also influence the choice of contract type. A principal task of the rest of this analysis will be to identify factors that, although thev may be correlated with size, are more reasonable sources of differences in firm behavior. One set of possibilities is related to the general labor market behavior of larger firms. There is a large literature (see Brown and Medoff 1989) document- ing the widespread existence of large earnings differentials associated with em- ployment in larger firms. The factors that cause large firms to "overpay" their workers may also cause them to "undercharge" their apprentices, in which case contract choice is simply an extension of a broader class of uncompetitive labor market behavior. If larger firms do not charge fees as part of a general pattern of paying rents to their employees, then explanations of contract choice would rely on explanations of the employer-size wage premium. Although this article does not discuss all of the possible sources of the employer-size wage premium, sev- eral of these explanations are also relevanit here. Do firms that do not charge fees provide rents to their workers and appren- tices? The evidence in table 3 indicates that apprentices who are not charged fees do indeed earn more during their apprenticeships than their fee-paying coun- terparts do. But these differences could he generated by differences in the degree of specificity of the training and overall training costs, and thus cannot be easily interpreted as reiits. Apprentices who are not charged fees may well be advantaged, but we cannot measure this advantage using current earnings with- out having data oni training costs. There is strong evidence, however, that nonapprenticed employees are advantaged by employment in larger firms. Jones and Teal (1994) find evidence of large-em- ployer size effects in estimates for the whole sample of workers. They estimated that the premiums for small, medium, and large firms in relation to microenterprises are 21, 38, and 55 percent, respectively. Estimates for the subsample of workers em- ployed in firms offering apprenticeship also show a substantial advantage to em- ployment in larger firms. Medium-size firms pay 40 percent more than firms with fewer than thirtv workers do, and large firms pay 73 percent more. Velenchik 465 The fact that employees of larger firms earn rents does not necessarily imply that employees of all firms in which no fees are charged also earn rents, because the correlation between contract choice and size is not perfect. Among the twenty- four firms with more than thirty workers, five charge fees, and there is no statis- tically significant difference in wages between large firms that charge fees and those that do not. Among the seventy-three firms with fewer than thirty work- ers, twenty-one do not charge their apprentices. Interestingly, among the smaller firms, those that do charge their apprentices actually pay their workers more. Overall, there is no clear evidence that all firms that do not charge fees also pay rents to their employees. To the extent that the decision to charge fees is strongly correlated with firm size, however, broader explanations of large-firm behavior, particularly in regard to wage premiums, may provide insight into contract choice. A source of these wage premiums that ma' also explain contract choice is unionization. Larger firms are more likely to be unionized, and wages in union firms are 9 percent higher than wages in non-union firms (Jones and Teal 1994). If unions also cause firms to undercharge for training, or if unions forbid the charging of fees, unionization would also drive contract choice. Seventeen of the firms in this sample are unionized, and none of them charge fees. Although it is possible that unionization and contract choice are both caused by a third factor (perhaps the profitability of the firm), it seems reasonable to attribute at least part of the decision not to charge fees to union influence. But there are also twenty-three non-union firms that do not charge fees, so the union factor is not a complete explanation of contract choice. Another common explanation of employer-size wage premiums is rent shar- ing. If larger employers are able to earn economic profits or rents, they may share these rents with their workers and apprentices. There are several potential sources of rents. Market power in output markets would allow these firms to charge higher prices. These data include no obvious measure of market power, so we cannot directly explore the relation between market power, profits, and contract choice. Quantity discounts on raw materials are another source of rents. Differences in the nature of raw materials used across firm sizes make direct comparisons of raw materials costs impossible, although the subsequent discus- sion will explore differences in value added that may capture variation in input costs in relation to output values. If firms face differential costs of financing their working capital needs, this would also give rise to rents for those firms having cheaper or easier access to funds. Differential costs of working capital would also give rise to variation in contract choice if, as described above, choice of contract type is based on a decision about financing of training. Employer financing of training is a form of informal credit from firms to apprentices, and it is likely that firms finance training when they have access to funds at a lower interest rate than that avail- able to apprentices or when firms are able to access more funds than apprentices can. This is the idea that Berry (1985) reflects in her discussion of the expense of working capital to firms using apprentices. Table 5. Conttract Type anld Firmi Credit Market Behavior, Ghaana, 1992 (percent unless noted otherwise) No-fre-charging firms Fee-charging firms t-test for With unpaid With paid With unpaid With paid difference Vlariable apprentices apprentices apprentices apprentices in Pmeansa Value added per worker per year" (cedis) 414,852 830,324 322,103 173,183 Relations with b1anking sector Share of firms receiving bank loanis 16 6 6 0 Share of firms receiving overdraft facilities 33 17 6 7 Share of firms holding bank accounts 100 79 75 63 a Relations wtith suppliers Share of firms receiving credit from suppliers 33 35 19 32 Mean length of supplier credit for those receiving it (days) 45 51 9 29 Mean implied interest oni supplier credit tor those receiving it 7 4 10 13 Informal credit markets Share of firms receiving informal credit 33 26 44 51 Share of firms using informal safekeeping services or mobile bankers 0 4 19 18 Share of firms participating in informal savings groups 0 3 0 7 Relations with customers Share of firms making sales on cash basis 50 38 75 66 Share of firms making sales on credit basis 67 62 44 27 Share of firms receiving advance payment from customers 50 44 56 68 Note: The sample size is ninety-seveni firms. a. Difference in means between fee-charging and no-fee-charging firms (for firnis paying their apprenrices a wage and those that do nor). and in dicate the rest is significant at 10 and 5 percent, respecrively. b. Data were available for only ninery-one firms for this variable. Source: Author's calculations. Velenchik 467 Although a firm's access to or cost of financing cannot be measured, its behav- ior in other arenas involving financing can be used as an indicator. If the hy- pothesis above is correct, firms that charge apprenticeship fees will differ from firms that do not in terms of their use of alternative sources of finance, their relationships with other business partners, and their behavior and experience in credit markets generally. In particular, firms that charge fees to their appren- tices would be expected to differ from other firms in their mechanisms for ob- taining working capital. Normally, businesses meet their working capital needs through some com- bination of their own liquidity and formal sources of credit such as banks and input suppliers.3 Firms with weak cash flow and lacking access to bank and supplier credit will use informal sources of finance, such as borrowing from money lenders, or holding funds with informal savings collectors or mobile bank- ers in order to have access to funds at month's end, or requiring advance pay- ments from customers. This use of informal finance can be seen as a spillover from formal financing markets, in which agents who are rationed out of formal sources turn to informal markets. From the perspective of contract choice, firms with good cash flows or access to formal finance at reasonable costs would be expected to be able to finance their apprentices' training and would be less likely to charge fees. These firms would also face lower costs for working capital and could share this advantage with workers by not charging fees. The charging of fees to apprentices, which is similar to the practice of requiring advance payments from customers, is itself a form of informal finance, because the fees are generally paid before the training is received. Therefore, firms charging apprentices fees would also probably use other informal sources of credit. To the extent that these sources are more ex- pensive than formal sources, these firms would not be earning rents that could be shared with workers. Data about several features of firms' credit and financial market behavior are presented in table 5. Although it would be desirable to include a measure of the strength of the firms' cash flows, the data set does not include information that would allow such a measure to be constructed. The average value added (value of sales minus cost of raw materials) per worker per year is a blunt measure of the availability of rents to the firm and is relevant for examination of the rent- sharing hypothesis. These data were available for only ninety-one of the ninety- seven firms. Firms that charge for apprenticeship have significantly lower value added per worker, indicating less access to rents than firms that do not charge. A firm is considered to have received bank credit if it had either a loan or an overdraft facility at a bank or other formal financial institution in the past three years. Bank accounts include both demand and time deposits at all formal finan- cial institutions. From table 5, firms that charge for apprenticeship are less in- 3. Supplier credit lies at the border between formal and informal credit. For this analysis it is treated as formal because suipplier credit is a common feature of business practices in developed manufacturing sectors throughout the world and is part of their formal credit practices. 468 THE WORLL) BANK ECONOMIt RFVIF.W, V(. 9. NO. 3 volved with the banking sector, both as borrowers and as depositors, than are firms offering the other form of contract, and this indicates limited use of the banking system as a source of working capital. Firms were asked to describe the form of payment (cash, credit, or advance payment) they used with the primary suppliers of their three major raw materi- als and then to give the details of the last transaction of each type they con- ducted. Each firm can engage in all three types of transactions. A firm is said to receive supplier credit if any of its principal raw materials suppliers extend credit. Table 5 presents two characteristics of supplier credit for those firms receiving it. The first is the mean value of the elapsed time between delivery of the goods and final payment, and the second is the implicit interest rate paid for the credit, measured in terms of the cash discount forgone. Although firms that charged their apprentices were no less likely to have received credit from suppliers than were firms that did not charge, they did get shorter repayment periods and pay higher interest rates, in terms of forgone cash discounts, than the other firms did. This evidence supports the hypothesis that firms without strong cash flows or access to credit from the formal sector are more likely to charge fees. Firms were asked if they had borrowed money on the informal credit market (this includes loans from friends, relatives, suppliers, clients, other enterprises, and informal moneylenders) in the past three years. They were also asked if they had participated in any informal savings groups or had used the services of a mobile banker or savings collector-an individual who holds cash for safekeeping in exchange for a fee. As shown in table 5, firms that charge fees are more likely to be involved in the informal credit market, both as borrowers and as users of savings collectors than are firms that do not. Firms were asked to describe the form of payment they used with their most important customers or the form they used most frequently when they had a large number of infrequent customers. They were also asked to describe the details of their most recent transaction of each type. From table 5, firms charg- ing apprenticeship fees were more likely to deal in cash or advance payment with their clients and less likely to extend them credit. Because the elapsed time between order and delivery of goods is far less than the mean duration of ap- prenticeship, it seems reasonable to conclude that firms that do not finance their production costs out of working capital would be unlikely to finance training. Worker financing of training is itself a source of informal credit. The evidence indicates that firms that are more likely to use informal credit use several forms of informal credit, including borrowing from their customers in the form of advance payment and borrowing from their apprentices in the form of advance fees. Contract choice in apprenticeship markets, then, seems to mimic financing choices in other arenas and to be part of an overall financing strategy emphasiz- ing informal credit arrangements. The relation between these credit market behaviors and the choice of appren- ticeship contract is summarized in table 6, which presents the results of a probit estimation of the influence of such capital market features on the probability Velenchik 469 Table 6. Probit Estimates of the Influence of Capital Market Features on the Probability That a Firm Charges a Fee to Apprentices, Ghana, 1992 Estimation excluding Estimation including value added per worker v'alue added per worker Mean effect on Mean effect on Variable Coefficient probabilitvy Coefficient probability' Sector dummy Food 0.156 0.058 0.557 0.1 5 (0.214) (0.697) Woodworking and furniture -0.587 -0.181 -0.596 -0.071 (1.188) (1.107) Metalworking -1.170*' -0.285 -1.058" -0.1(00 (2.309) (1.892) Firm size -0.0005 -0.0001 -0.0017 -0.0002 (0.262) (0.744) Value added per worker -1.44e-6'"' -3.20e-7 (2.865) Capital market bebavior dummy5 Firm received bank credit -0.475 -0.148 -0.193 -0.034 (0.931) (0.307) Firm holds bank account -0.641 * -0.194 -0.374 -0.049 (1.633) (0.906( Firm received informal 0.462 0.18( 0.270 0.049 credit (1.326) (0.697) Firm used mobile banker 1.275< 0.465 1.149 0.289 (1.805) (1.509) Firm made sales on cash basis 1.01 I 0.384 1.002** 0.242 (2.771) (2.595) Firm made sales on credit basis -1.084 > -0.273 -1.002"* -0.297 (2.863) (2.391) Firm made sales with 0.927t " 0.356 0.732* 0.161 advance payment (2.305) (1.639) Constant 0.542 n.a. 1.022 n.a. (0.829) (1.435) Number of observations 97 91 X2 45.13 48.36 Pseudo R2 0.3432 0.4020 Log likelihood -43.175 -35.959 Significant at 10 percent. e* Significant at 5 percent. n.a. Not applicable. Note: The dependent variable takes the value I if the firm charges a fee. t-statistics are in parentheses. a. The mean effect on probability is the effect of each variable on the probability a firm chiarges fees, calculated as the mean of the effect across all observations. b. The value is 1 if the variable holds, () otherwise. Source: Author's calculations. 470 THE WORLD IIANK X:(OlNOMIC REVIEW, VOL. 9, NO. I that a firm charges a fee to apprentices. The dependent variable takes the value 1 if the firm charges a fee. The independent variables include sectoral dummies, firm size, and value added per worker, followed by a series of dummy variables that take the value I if the firm made sales on a cash basis, made sales on a credit basis, made sales with advance payment, received bank credit, held bank accounts, received informal credit, and used informal mobile banking services.4 The effect of each variable on the probability that a firm charges fees, calculated as the mean of the effect across all observations, is reported in the second col- umn for each set of estimates. For the continuous variables, this represents the change in probability in response to a one-unit change in the value of the vari- able. For the dummy variables, it represents the difference in the probability of charging a fee generated by having rather than not having that characteristic. The table reports estimates of two versions of the probit, one that excludes and one that includes the value-added measure. There is substanitial correlation among the independent variables, particu- larly those related to banking and informal credit, and this correlation results in Imprecise estimates. One mechanism for dealing with the imprecision is to com- press some of the measures. The results in table 7 were constructed by replacing the two banking measures with a single dummy that takes the value 1 if the firm had any relationship (borrowing or deposits) with formal banking institutions and replacing the two informal credit markets with a single dummy that takes the value I if the firm had any relation (borrowing or using mobile bankers) with informal credit institutions. These results indicate that a relation between credit market behavior and the use of apprenticeship fees exists and that firms with more access to rents, as measured by value added, are indeed less likely to charge fees. The magnitudes of the credit market effects are smaller in the estimates where value added is included. This is not unreasonable, given that profitability, as proxied by value added, would be a determinant of firms' access to formal credit, generating some substantial correlation among the variables in these regressions. The credit market results show that firms that do not provide financing to their customers are less likely to provide financing to apprentices. Use of the formal banking sector is negatively correlated, but use of informal credit mar- kets is positively correlated, with the charging of apprenticeship fees. It must be emphasized, however, that this shows only correlation, not causation. Both credit market behavior and contract choice would be influenced by factors that con- trol the firm's access to and cost of funds. The probit provides results consistent with the notion that worker financing of apprenticeship results from poor firm access to formal sources of working capital. The probit analysis is not a test of that hypothesis, however, nor is it a direct test of the hypothesis that firms which do not charge apprenticeship fees are engaging in rent sharing with their workers. 4. Because there is no variation across contract types in the use of supplier credit, supplier credit indicators were niot included in the prohit. Velenchik 471 Table 7. Probit Estimates with Formal and Informal Banking Dummy Variables, Ghana, 1992 Estimation excluiding Estimation including value added per worker value added per worker Mean effect on Mean effect on Variable Coefficient probability, Coefficient probability' Sector dummy Food 0.125 0.048 0.399 0.136 (0.175) (0.510) Woodworkinig and furniture -().531 -0.208 -0.479 -0.147 (". 09) (0.927) Metalworking -1.120* -0.397 -0.948* -0.256 (2.280) (1.78.) Firm size -0.0004 -0.(0(01 -0.0010 -0.0002 (0.211) (0.440) Value added per worker -1.32e-6'* -2.95e-7 (2.663) Capital market bebai nbr dummyb Firm had relationship -1.024* -0.372 -0.646 -0.190 with formal bank (2.366) (1.450) Firm had relationship with 0.701 " * 0.323 0.488 0.167 informal financial sector (2.011) (1.278) Firm made sales on cash basis 1.068W 0.323 0.968* * 0.318 (3.094) (2.626) Firm made sales on credit basis -0.962" -0.354 -0.9(0* -0.247 (2.675) (2.252) Firm niade sales with 0.893* 0.2 87 0.7 1 0.237 advance paymilent (2.309) (1.625) Constant 0.689 n.a. 1.121 n.a. (1.030) (1.589) Number of observations 97 91 X2 44.74 47.56 Pseudo R2 0.3403 0.3954 Log likelihood -43.368 -36.359 Significanit at 10 percent. * Significant at S percent. n.a. Not applicahle. Note: t-statistics are in parentheses. a. The effect of each variahle on the prohabilitv that a firm charges fees, calculated as the mean of the effect across all observationis. h. The value is I if the variable holds, 0 otherwise. Sou2rce: AUthor's calculations. It would also be expected that characteristics of the apprentices influence the type of contracts in which they are found. The evidence in table 3 shows little difference in the educational characteristics or ages of apprentices, although it does indicate that women tend to be concentrated in contract types that do not involve wage payments but do involve fees. In addition to these personal charac- 472 THE WORI ID BANK FCt:ONMI(: RlVIW, VOL. 9, Ni. I Table 8. Probit Estimates of the Probability an Apprentice Paid a Fee, Ghana, 1992 Meain effect on Variable Coefficient probability Age (years) 0.0389 .0(79 (0.197) Age squared -0.0006 -.00(01 (0.1337) Highest schooling-level dummy Primary -0.1014 -.0403 (0.210) Middle school 0.4422 .1703 ( .15 1. ) Secondary schlool -0.6054 -.2278 i 1.44 1) Apprentice is male -0.6909:- -.2556 (2.196) Apprentice sends remittances to parents -O.7003" -.2586 (2.28 l) Apprentice receives cash from parents 0.8X68X .3083 (2.7S5) Constant 0.2086 n.a. (0.088) Significanit at 5 percent. n.a. Not applicable. Note: For the dummy variables, the valuLe is I if the variable holds, 0 otherwise. The sample size is 209, the log-likelihood is -78.502864, the X (8) is 49.93. and the pseudo R2 is (0.241 3. The absolute values of t-statistics are in pareiitheses. Source: Author's calcuilations. teristics, it would also be expected that the apprentices' ability to finance their apprenticeship influences contract choice. Most apprenticeship fees are paid by parents, and many apprentices have little knowledge of the specific arrange- ments that were made. If parents are the main source of funds for apprentices, then it would be expected that the apprentices for whom fees are paid are more able to draw on parental resources. The survey included questions to the ap- prentices about their receipt of funds from and remittance of funds to their parents. These variables, along with more standard measures of human capital, are included in probit estimates of the probability an individual apprentice paid a fee. These estimates are presented in table 8. As expected, male apprentices are less likely to have paid fees.' In the total sample, access to parental financial resources increases the likelihood of being in a paying apprenticeship by 31 percent; in contrast, apprentices who send cash home are 5. Given the differences between men and women in this sample, an attempt was made to analyze their behavior separately. Small sample size for the women (sixty-four) and lack of variation across them in personal characteristics such as education and age made estimation of the same probit impossible, but the available results for women do indicate the same positive relatioiiship between receivitng cash from parents and paying fees as is seen for the sample as a wholc. Velent7hik 473 26 percent less likely to have paid fees. Other human capital variables, including education and age (a proxy for experience) do not exert statistically significant in- fluence on contract choice, although it seems that the most educated individuals- those who have completed secondary school-are less likely to be in apprentice- ships in which fees are charged. If the qualifications and abilities of the two groups-those paying and those not paying fees-of apprentices do differ, the differences are not captured in standard measures of human capital. Overall, the evidence indicates that contract choice in apprenticeship, insofar as it is based on decisions about the financing of training, is influenced by the relative access to funds of firms and apprentices. In contexts such as manufac- turing in Ghana, where firms are small and informal and not necessarily clearly distinguishable from households, it seems reasonable that some firms would be unable to finance training as easily as their apprentices can. VI. SUMMARY AND DIRECTIONS FOR FtUTtlRE RESEARCH A great deal of training takes place on the job in the manufacturing sector in Ghana, particularly in small firms. Although this article does not explore the content of this training and therefore cannot assess its quality, it seems clear that the training of a new generation of craft workers is an important part of economic activity in the small-scale manufacturing sector. Equally important, the evidence indicates that these firms provide relatively general training, generat- ing skills that are useful in other firms and that seem to provide apprentices with preparation for entrepreneurship. Both King (I 977) and Berry (1985) note that apprenticeship in a trade is a route to self-employment, and the importance of apprenticeship training in the backgrounds of the entrepreneurs in this sample supports this idea. The important implication of this fact is that apprenticeship should not be excluded from general analyses of vocational training in Ghana and that the human capital being generated in small-scale manufacturing should be considered when policy in this area is made. The human capital approach taken here provides insights into the character- istics of traditional apprenticeship programs. This article has expanded the ideas of King and Berry by looking more closelv at the financing arrangements in- volved in apprenticeship contracts. King's contention that apprentices pay for their training is correct, because all apprentices ultimately bear the cost of the general portion of their training. But what is truly interesting about fee-paid apprenticeships is that they involve apprentice financing as well. Berry's conten- tion that the usefulness of apprentices to firms is that they do not have to be paid is insightful because it focuses on the importance of constraints on working capital in motivating the structure of apprenticeship contracts. Although these contract forms do alleviate constraints on working capital, the fact that appren- tices do not receive wages is unlikely to actually increase firm profits. The small payments received by apprentices would be exactly offset by their low produc- tivity. Only if apprentices received less than their productivity minus training 474 THE WORLD BANK V(ONOMC RFVIFW, VOL. 9, NO. l costs would this practice actually enhance firm profits. This scenario would be possible only if the firm charged the apprentices more than their training cost, which the firm could do only if it had some market power. Given the large number of small firms engaged in the training of apprentices in any trade, this kind of market power seems unlikely. The correlation between apprenticeship contract choice and other credit market behavior neither shows causation nor identifies the root cause of differences in these behaviors across firms. It does indicate, however, that the practice of firm financing of training is related to the way in which firms finance other activities. This implies that changes in the functioning of urban capital markets could in- fluence patterns of human capital formation as well as cause the more well- known effects on technological change, employment, and growth. This analysis has provided a first step to understanding the role of apprenticeship in manufacturing in Ghana. It does, however, neglect some important questions, including the determinants of the firm's decision to train apprentices and to offer general or specific training and the influence of apprenticeship training on the em- ployment opportunities, career paths, and wages of apprentices. Future work must address these issues. In an era of budgetary crisis and structural adjustment that are forcing many African countries to reduce their education budgets, the role of alter- native training institutions such as apprenticeship must be explored. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Baah-Nuakoh, Amoah, and Francis Teal. 1993. "Economic Reform and the Manufac- turing Sector in Ghana: Final Report." World Bank, Regional Program on Enter- prise Development (RPED)), Washington, D.C. Processed. Bas, Daniel. 1989. Oin-the-Job Training in Africa." International Labour Review 128(4):485-96. Becker, Gary. 1975. Human Capital. 2d ed. Chicago: University of Chicago Press. Berry, Sara. 1985. Fathers WY'ork for Their Sons. Berkeley: University of California Press. Brown, Charles, and James Medoff. 1989. "The Emplover Size-Wage Effect." Journal of Political Econo7ny 97(October): 10'7_59. Collier, Paul, and Deepak Lal. 1986. Labour and Povert in Kenva. New York: Oxford University Press. Coy, Michael, ed. 1989. Apprenticeship: From Theory to Method and Back Again. Albany: State University of New York Press. Harris, John. 1971. "Nigerian Entrepreneurship in Industry." In Peter Kilby, ed., En- trepreneirship and Economic Develop?nent. New York: The Free Press. Hashimoto, Masanori. 1981. "Firm-Specific Human Capital as a Shared Investment." Americani Economic Review 71(3):475-82. Hashimoto, Masanori, and Ben Yu. 1980. "Specific Capital, Employment Contracts, and Wage Rigidity." BellJournal of Economics 11(2):536-49. Velenchik 475 Hazlewood, Arthur, Jane Armitage, Albert Berry, John Knight, and Richard Sabot. 1989. Education, Work, and Pay in East Africa. Oxford, U.K.: Clarendon Press. House, William, Gerrishon Ikiara, and Dorothy McCormick. 1993. "Urban Self-Employ- ment in Kenya: Panacea or Viable Strategy?" World Development 21(July):1205-23. Jones, Tracy, and Francis Teal. 1994. "Earnings and Wages in the Ghanian Manufac- turing Sector: A Preliminary Report on a Panel Data Survey." University of Oxford, Centre for the Study of African Economies, Oxford, UJ.K. Processed. King, Kenneth. 1977. The African Artisan. London: Heinemann. Knight, John, and Richard Sabot. 1990. Education, Productivity, and Inequality: The East African Natural Experiment. New York: Oxford University Press. Liedholm, Carl, and Enyinna Chuta. 1976. "The Ecoonomics of Rural and Urban Small Scale Industries in Sierra Leone." African Rural Economy Paper 14. Michigan State University, East Lansing. Processed. Liedholm, Carl, and Donald Mead. 1986. "Small Scale Industry." In Robert Berg and Jennifer Whitaker, eds., Strategies for African Development. Berkeley: University of California Press. Psacharopoulos, George, and Maureen Woodhall. 1985. Education for Development: An Analysis of Investment Choices. New York: Oxford lJniversity Press. Squire, Lyn. 1981. Employment Policy in Developing Countries: A Survev of Issues and Evidence. New York: Oxford University Press. Steel, William, and Leila Webster. 1992. "How Small Enterprises in Ghana Have Re- sponded to Adjustment." The World Bank Economic Revieuw 6(.3):423-38. University of Ghana. 1992. "Country Background Paper: Economic Reform and the Manufacturing Sector in Ghana." Department of Economics, Legon, and University of Oxford, Centre for the Study of African Economies, Oxford, U.K. Processed. THF,- WORLI ) BANK lKON4MI( REVIEW, V\tI 9, NO. 1: 477-508 Inequality and Growth Reconsidered: Lessons from East Asia Nancy Birdsall, David Ross, and Richard Sabot East Asian economies have experienced rapid growth over three decades, with rela- tively lou! levels of income inequality, and appear to have also achieved reductions in income inequality. We argue that policies that reduced poverty and income inequality, such as emphasizing high-quality basic education and augmenting labor demand, also stimulated grou'th. Closing two virtuous circles, rapid grouwth and reduced inequality led to higher demand for, andi supply of, educatiopn. Moreover, low levels of income inequality may have directly stimulated growth. We present cross-economy regression results that are consistent uwith a positive causal effect of lowv inequality on economic growth and with loU! inequality of- income as an independent contributing factor to East Asia's rapid growth. We conclude that policies for sharing growth can also stimu- late grouth. In particular, investment in education is a key to sustained growth, both because it contributes dlirectly through productivity effects and because it reduces in- come inequality. It has long been the conventional wisdom that there is a tradeoff between aug- menting growth and reducing inequality. Two explanations are given for the view that an unequal distribution of income is necessary for, or the likely conse- quence of, rapid economic growth. The first, following Kaldor (1978), is that because a high level of savings is a prerequisite of rapid growth, income must be concentrated in the hands of the rich, whose marginal propensity to save is relatively high. The second, following Kuznets (1955), is that as labor shifts from sectors with low productivity to sectors with high productivity, aggregate inequality must initially increase substantially and only later decrease; Robinson (1976: 437) observed that this pattern had "acquired the force of economic law." The conventional wisdom has in fact been repeatedly questioned. Twenty years ago, prior to the fulfillment of East Asia's remarkable potential, Ahluwalia Nancy Birdsall Is ExCcutive Vice-President at the Inter-American Development Bank, David Ross is with the Department of Economics at Bryn Mawr College, and Richard Sabot is with the Center for Development Economics at Williams College and the Policy Research Department at the World Bank. The authors are grateful to Kristin Forbes, Rebec.a Foster, and Jennifer Keller for their able assistance. The authors also wish to thank for helpful comments the participants in the authors' session of the 1994 meeting of the American Economics Association and in seminars at St. Antony's College, Oxford; Williams College; Kennedy School of Government, Harvard University; and the Inter-American Development Bank, where this article was presented. ©D 1995 The International Bank for Reconstructioni and Development / THE 'WORLI) BANK 477 478 THE WORLI) BANK FCONOMIC REVIFW, VOL. 9, NO.3 (1974: 13) concluded, on the basis of the cross-country data then available, that "[T]here is no strong pattern relating changes in the distribution of income to the rate of growth of GNP. . . . This suggests that there is little firm empirical basis for the view that higher rates of growth inevitably generate greater in- equality." Anand and Kanbur (1993) demonstrate the sensitivity of the U-shaped relation between inequality and per capita income to changes in functional form and choice of observations. Alesina and Perotti (1994), Alesina and Rodrik (1994), Fields (1989), and Persson and Tabellini (1994) provide supporting evidence for why high inequality might constrain growth. In this article we show how the recent experience of East Asia also casts doubt on the conventional wisdom. East Asian economies have had rapid growth over three decades with relatively low levels of income inequality, and although evidence on changes in income inequality is fragile, these economies appear to have also achieved reductions in income inequality. Our argument is straight- forward: policies that reduced poverty and income inequality, such as empha- sizing basic education and augmenting labor demand, also stimulated growth. Moreover, low levels of income inequality may have stimulated growth. The rapid growth in the East Asian region has been unparalleled. Eight "high- performing Asian economies"-labeled HPAES in World Bank (1993)-are Hong Kong, Indonesia, Japan, the Republic of Korea, Malaysia, Singapore, Taiwan (China), and Thailand. Not only have the HPAEs outperformed the industrial economies since 1960, but they have grown at higher rates than other develop- ing economies (see table 1). They have also grown at higher rates than the cur- rent industrial economies did during their earlier periods of most rapid growth: the annual average compound growth rate of GDP per capita of industrial econo- mies was 1.1 percent for 1820-70 and 1.4 percent for 1870-1913 (Maddison 1982). The HPAEs also stand out for persistence of rapid growth: although other developing economies have experienced rapid growth for several years, few have managed to sustain such growth over three decades (Easterly and others 1993). The decline in poverty in East Asia has been equally remarkable. The per- centage of people below the poverty line has fallen far more rapidly in East Asia than in other regions. For example, in the period from 1970 to 1987, the percentage of people below the poverty line in Indonesia fell from 58 percent (among the highest of all developing economies for which measures are available) to 17 percent (among the lowest of all developing economies)(World Bank, various years b). As would be expected, the abso- lute number of poor people has also declined significantly in the HPAEs (World Bank, various years b). Although declines in income inequality are more dif- ficult to document, measures such as the Gini coefficient show improvements in each of the HPAEs from 1965 to 1990 (World Bank 1993). Findlay and Wellisz (1993); Kuo, Ranis, and Fei (1981); and Leipziger and others (1992) also cite evidence suggesting that, unlike in many other developing econo- mies during this period, income distribution in East Asian economies either improved or, at the very least, did not worsen. (However, for some years Birdsall, Ross, and Sabot 479 Table 1. Average Annual Growth in Per Capita GNP, Selected Economies, 1960-80 and 1965-90 Country 1960-80 1965-90 High-performing East Asian economies Hong Kong 6.8 6.2 Indonesia 4.0 4.5 Japan 7.1 4.1 Korea, Rep. of 7.0 7.1 Malaysia 4.3 4.0 Singapore 7.5 6.5 Taiwan (China) - 7 Thailand 4.7 4.4 Average 5.9 5.5 Selected developing economies Argentina 2.2 -0.3 Bangladesh - 0.7 Botswana - 8.4 Brazil 5.1 3.3 Cameroon 2.6 3.0 Chile 1.6 0.4 El Salvador 1.6 -0.4 Guatemala 2.8 0.7 Ghana -1.0 -1.4 India 1.4 1.9 Kenya 2.7 1.9 Mexico 2.6 2.8 Morocco 2.5 2.3 Nigeria 4.1 0.1 Pakistan 2.8 2.5 Panama 3.3 1.4 Peru 1.1 -0.2 Senegal -0.3 -0.6 Zimbabwe 0.7 0.7 Source: World Bank (various years b) and Taiwan, China (1994). within this period, some East Asian economies show evidence of worsening inequality. See, for example, Bruton 1992 and figure 1.) Figure 1 shows the relationship across economies between percentage growth in GDP (1965-89) and average income inequality (1965-90) as measured by the ratio of the income share of the richest 20 percent to that of the poorest 40 percent. The East Asian economies stand alone in the northwest corner; they achieved a combination of rapid growth and low inequality over the twenty- five-year period. In part, the association of rapid growth and low inequality is explained by policies and programs adopted by governments in East Asia that helped ensure widespread sharing in the benefits of economic growth. In addition to the post- war land reforms in Korea and Taiwan (China), a variety of other policies and programs had the effect of ensuring shared growth: public housing in Hong 480 THE WORLI1 BANK ECONOMIC: REVIEW, VOL. 9, NO. 3 Figure 1. Income InequalitV and Growth of GI)P, 1965-89 Growth of Gt1P per capita (percent) 8 RepuhlI it of Kore ('76) b/ Bl!otswana - lionI Kong ('70) U Repuhlic_io 0(/ Korea (7( * 4- Singapore ('83) Gabon Singapore ('9) 09 6 Hong Kong ('80) Indonesia ('71) - Japan,, * Mauritius A/ zA Thailand ('88) Indonesia ( 9(i) m a Malaysia ('65) Thailand ('65) Malaysia ('90) *BrjziI - ~ ~~~~~~~~~~~~~ ~ Brazill Spain Belgium * *France Colomhia Australi a 2 I Inited? ' Sri Lanka NMexico 2 Kingdomi ' MF AXswitzerlandaKevi Pak istan Philippines I n0 ThailandI / [-Si|gapore (4Mlaysia 6( [2 ~~~31 [ 18.71 18 SI Pakistaln 12') , / SIndonesiL (46) Pais i 1 119.1 Brazil (39) 40 T n Hong Kong (29) d (28) / ~~~~~~~~~-1.41 [-8.31 BangladeSh Malaysia (28) / (0 713) / / 1 19 Pakistan (191 20 -14.21 / -.91 4 Brazil ( 16) Bngladesh (18) * ndonesm (12) j iI 1-3.61 0 I , . , , ..I , I , _ __ 1 , , , , -: A IN 0 250 900) 9(1 1 .5( ( 2.0(00 ) 5(00 1.00() 1,5(0 3,500 Per capita income (1988 UI S dollars) AVole. Figure.s in parentheses are enrollment rates; bracketed numbers show residuals. a. Per capita income substantially exceeds tJS$3.50(0. Source. Sing;ipore and -loing Kong from Vo(rkli Bank ( 19941; all others from Behrmian and Scthneidler ( 1992). parable tests than children from other developing economies (Birdsall and Sabot 1994). In the first two sections of this article, we focus on two "virtuous circles" in East Asia. Section I describes a process of cumulative causation, in which educa- tion contributed to economic growth, which, in turn, stimulated investment in education. Section II shows how education contributed to low levels of income inequality which, in turn, stimulated investment in education. In section III we discuss how low inequality of income can have a direct, positive effect on eco- nomic growth, over and above its effect through education. We present cross- economy regression results that, contrary to the predictions of Kaldor and Kuznets, are consistent with a positive causal effect of low inequality on eco- nomic growth and with low inequality as an independent contributing factor to East Asia's rapid growth. We conclude that investment in education is a key to sustained growth, both because it contributes directly through productivity ef- fects and because it reduces income inequality. Throughout this article we invoke a variety of different results-from theory, microeconomic studies, and cross-economy studies-to piece together and illus- trate our story. None of these alone would constitute an airtight case. In par- ticular, the results of cross-economy statistical work must be interpreted with Birdsall, Ross, and Sabot 483 caution;' the statistical difficulties of working with cross-economy regressions- a high level of aggregation, substantial measurement error, inconsistencies in data collection and definition, and specification and omitted-variable bias-are well documented (for example, Levine and Renelt 1992). However, only by looking across economies is it possible to generate the level of variation in key indicators of growth, inequality, and policy needed to capture the effects hy- pothesized here. The cross-economy regression results reported below comple- ment, and are consistent with, theory and with the microeconometric evidence we present. l. EDUCATION STIMLILATES GROWTH, AND GROWTH STIMULATES EDUCATION The accumulation of human capital, as measured by the educational attain- ment of the population, has consistently emerged as an essential feature of eco- nomic growth and development. But the direction of causality implied by the positive correlation between educational attainment and per capita output in a cross-section of economies is unclear: it could simply indicate that education is a luxury consumer good that is increasingly demanded as incomes rise. This concern has been eased by cross-economy regressions in which the characteris- tics of economies decades ago are used as predictors of subsequent rates of growth. A widely cited contribution to this burgeoning growth-regression literature is Barro (1991). The core regression from that paper, which we use as our base regression, appears in the first column of table 2. (Descriptive statistics for the variables used in table 2 appear in appendix table A- 1.) The important contribu- tion of education is among the most robust findings of these growth regressions, proving to be relatively insensitive to changes in either specification or sample composition (Levine and Renelt 1992). These results are consistent with what human capital theory, the theory of investment in people, predicts. Education augments cognitive and other skills, which, in turn, augment the productivity of labor (Schultz 1961; Becker 1964). Moreover, human capital theory has been extensively tested in the microeconomic literature; Schultz (1988) and Strauss and Thomas (1995) review the evidence from dozens of studies, based on household- and firm-level data, showing that more-educated individuals receive higher wages and are more efficient in man- aging the health and nutrition of their households. Endogenous growth theory also predicts that educational investments will enhance growth. A larger stock of human capital facilitates the production of new ideas and technological progress or, for an economy that is not on the 1. Cross-economy comparisons have been a prominent feature of the now fifty-year empirical search for uniformities in the process of the transformation of low-income economies into high-income economies. Clark (1940) and Kuznets (1966) were pioneers in this effort. Chenery and Syrquin (1975) provided a comprehensive description of the structural changes that accompany the growth of developing economies and analyzed their relations. More recently, Romer (1990), Barro (1991), and King and Levine (1993) have used cross-country regressions to test endogenous growth models. 484 THE WORLI) BANK ECONOMI( REVIEW. VOL. 9, NO. 1 Table 2. Determinants of GDP Growth in a Cross-Economy Sample, 1960-85 Basic Basic regression regression Basic with income with regression share, excluding Basic manufactured with income education Variable regression exports share variables Per capita GDP, 1960 (thousands of 1980 dollars) -0.0075 -0.0069 -0.0075 -0.0020 (-6.25) (-5.135) (-4.730) (-1.801) Primary school enrollment rate, 0.0250 0.0271 0.0243 1960 (4.464) (4.532) (3.024) Secondary school enrollment rate, 0.0305 0.0262 0.0366 1960 (3.861) (1.723) (2.427) Ratio of manufactured exports 0.0007 to GDP, 1965 (1.539) Secondary school enrollment- 0.0005 export interaction, 1 960' (0.324) Income share ratiob -0.0013 -0.0018 (-1.897) (-2.406) Government consumption share of GDP, 1970-8 5 -0.1190 -0.0566 -0.1229 -0.1495 (-4.250) (-2.419) (-4.380) (-4.760) Average annual number of revolutions, 1960-85 -0.0195 -0.0168 -0.0176 -0.0268 (-3.095) (-1.987) (-1.867) (-2.493) Average annual number of assassinations, 1960-85 -0.0333 -0.0024 -0.0055 -0.0067 (-2.148) (-0.738) (-1.526) (-1.626) Absolute deviation in investment deflator, 1960(d -0.0143 -0.0139 -0.0)086 -0.0176 (-2.698) (-2.136) (-1.093) (-2.012) Constant 0.0302 0.0202 0.0)418 0.0706 R' 0.56 0.57 0.54 0.36 Number of observations 98 100 74 74 Note: The first column presents Barro's (1991) results; the other columns present new results. t-statistics are in parentheses. a. Ratio of manufactured exports to GDP multiplied hv 1 960 secondary school enrollment rate. b. Ratio of the total income of the top 20 percent to the total income of the bottom 40 percent for the first year for which data became available in each countrv. c. Average aniual ratio of real government consumption (exclusive of defense and education) to real C.DP. d. Magnitude of the deviation of the purchasinig power paritv value for the investment deflator (lJ.S. = 1.0) from the sample mean. Source: Barro (1991) and authors' calculations based on data from Barro (1991); World Baik (various years a): Clarke ( 1995); and World Bank data. technological frontier, relatively rapid adaptation of new ideas and acquisition of technological capability (Nelson and Phelps 1966; Romer 1990, 1993). More- over, rates of return to human capital may actually be increasing over some range because of spillover benefits, that is, when one more-educated worker makes an entire group of workers more productive. Birdsall, Ross, and Sabot 485 The magnitude of the contribution of investment in human capital to rates of economic growth can be assessed with counterfactual simulations derived from the basic regression model in table 2. Consider an economy characterized by the average value of each of the variables in the sample of ninety-eight economies in table 2. We simulated the changes in growth in per capita GDP for that economy during the period from 1960 to 1985 by assuming that the country had achieved primary and secondary school enrollment rates each half a standard deviation above or below the means of those variables for all countries. The one-standard- deviation difference in enrollment rates translates into a nearly 1.5-percentage- point difference in the annual per capita growth rate. The cumulative effect of this annual difference in growth rates over twenty-five years is large. The simu- lation indicates that a country with primary and secondary school enrollments half a standard deviation above the average in 1960 would have had a GDP per capita 40 percent higher than a country with 1960 enrollments half a standard deviation below the average. Similarly dramatic results emerge from simula- tions pairing actual countries. With Pakistan's 1960 enrollment rates, Korea's predicted growth rate falls, resulting in a per capita CDP for 1985 that is 39.6 percent lower than the one Korea actually attained. Allowing the impact on growth of primary school enrollment rates to vary by gender (not shown) reveals no significant difference between the coefficient val- ues for men and women, suggesting that increasing primary school enrollments for girls will be just as effective in stimulating growth as increasing enrollments for boys. This conclusion from cross-economv data is consistent with the microeconomic evidence that the private rates of return to education among wage earners are roughly the same for women as for men (Strauss and Thomas 1995; Summers 1992). Of course, the rate of participation in the wage-labor market is much lower for women than for men. However, the economic payoff to educating girls is given not only by the increases in the productivity of wage labor but also by nonwage social benefits derived from changes in behavior within households. Fertility is a notable example: the evidence is clear that educated mothers have fewer children (Behrman 1993). Educated mothers also are more efficient users of health services for them- selves and their children, send children to school who are better prepared to benefit from society's schooling investment, and are more likely to send their own daugh- ters to school (see Birdsall, Ross, and Sabot 1993). In Asia the fertility rate for women with more than seven years of schooling is 54 percent of the rate for unedu- cated women (Summers 1992). Closing still another virtuous circle, the fertility decline in East Asia beginning in the mid-1960s resulted in a marked slowing of the growth of the school-age population in the 1970s. This contributed to a growth of public expenditures on basic education per eligible child that was more rapid than elsewhere, permitting more rapid increases in the quantity and improvements in the quality of schooling provided. Differences among countries in the demand for skills, which have been ne- glected in regression studies of growth that test education's contribution, are 486 THF. WORLI) BANK ElONOMR: RFVI'IW, VOL. 9, NO. 3 also important. The resulting omitted-variable bias in these regressions may be a factor in the substantial overprediction of rates of growth for some countries with higher-than-predicted rates of enrollment in primary and secondary schools in the 1 960s. For example, countries such as Egypt, the Philippines, Sri Lanka, and the former Soviet Union, like East Asian economies, had greater human capital endowments than predicted for their initial levels of income. Weak de- mand for educated labor may help explain why these countries nevertheless have tended to underperform with respect to growth. Figure 4 illustrates the link between the demand for skills in the labor market and education's contribution to growth. The horizontal axis measures the mag- nitude of investment in human capital, as proxied, for example, by enrollment rates and basic education. The vertical axis measures the rate of return to in- vestment in human capital (schooling) and, implicitly, the contribution of in- vestment to growth for a given level of investment in human capital. S and D Figure 4. Demand, Shifts and the Retunis to Human Capital, Last Asian and Developin.g Fconolmt(tie( Rate of return to investment in I)' I hun c lTypical \ / \ ~~~~~~~East A.sian / Typical D' 7 leveloping - R1economy S( s S2 Quantity of investment in human capital Vote: S is the skill supply~ D is the demand ftnttion; S'and l'represent the same variables for a typic;al Fast Asian .eonotn. Birdsall, Ross, and Sabot 487 are, respectively, the skill supply and demand functions of the typical develop- ing economy; S' and D' are the skill supply and demand functions of the typical East Asian economy. We show S' as shifted to the right because, for example, there is greater public commitment to basic education in East Asia. (The supply function is more elastic because the distribution of income is more equal. There is less absolute poverty than in the typical developing economy, and families near the bottom of the distribution of income can better respond to increases in returns to investments in human capital.) D' is shifted to the right on the premise that, for any given rate of return to skill, skilled workers have been in greater demand in East Asia than in the typical developing economy. As drawn, in East Asia high levels of demand for skilled workers have offset the tendency for educational expansion to induce diminishing returns to invest- ment in human capital. By contrast, although enrollment rates are higher in some developing economies than others, for example economv X, the returns to education will be lower. Educational expansion (from So to S;) reduces the rate of return to education (from R. to R,) because economy X has the human capi- tal supply function of East Asia but the skill demand function of the typical developing economy. In contrast, in East Asia the net result of high demand for human capital as well as high supply is that, although skilled labor is more abundant than in the typical developing economy, the rate of return to invest- ment in human capital (at R,) is at least as high. It is widely acknowledged that macroeconomic, agricultural, and, in particu- lar, export-push policies contributed to a high demand for labor in East Asia (World Bank 1993).2 Did these policies, by ensuring adequate demand for skilled as well as unskilled labor, also help ensure a high economic return to education? We assess this question using the regression shown in the second column of table 2. We control for the supply of education and test the relevance of the demand for skills using variables that measure the degree to which an economy is oriented toward manufactured exports." Our prediction is that, in the stan- dard growth rate function, there will he a positive effect on growth of the inter- action between the degree of orientation of an economv toward manufactured exports and the level of educational endowments in 1960. That is, the more oriented toward manufactured exports the economy is, the greater the demand for skilled labor and the greater the impact of a given educational endowment on growth will be. 2. East Asia's emphasis on exports contrasted with the import-substituting, capital-intensive strategies that primarily served the elite and labor insiders in Latin America and elsewhere and did not generate a strong demand for labor (Banerli. Campos, and Sabot, forthcominig). 3. See, for example, Pack and Page (I 994) and Wood (1 994). Of course, a high rate of manufactured exports may also reflect the supply of skills-a tuniction in part of human capital accumulation (Balassa 1984). In the second columine of table 2, we corntrol for the supply-side channel by direct inclusion of school enrollment rates. And just as we mitigate simultaneity between growth and human capital accumulation by using 1960 enrollment rates, we mitigate simnltalneiry between growth and export intensity by using the 1 965 ratio of maniufactured exports to GDP. Table 3. Public Expenditure on Basic Education in Kenya, the Republic of Korea, Mexico, and Pakistan, 1970-89 Percentage change, Country and statistic 1970 1975 1985 1989 1970-89 Kenya Expenditure per eligible child (1987 dollars) 38.6 - 46.6 53.4 38.3 Expenditure as a percentage of GNP 4.0 - 4.9 4.9 22.5 Absolute expenditure index (1970 = 100), 100 - 220 286 186 Number of children eligible for basic education (millions)b 3.8 4.6 7.0 7.9 107.1 Korea, Rep. of Expenditure per eligible child (1987 dollars) 95.3 81.6 357.1 433.4 354.7 Expenditure as a percentage of GNP 3.1 1.9 3.8 2.7 -12.9 Absolute expenditure index (1970 = 100)a 100 91 388 444 344 Number of children eligible for basic education (millions)' 10.1 10.7 10.4 9.8 -3.0 4^ t Mexico Expenditure per eligible child (1987 dollars) 68.4 124.9 113.5 111.9 63.6 Expenditure as a percentage of GNP 1.6 2.6 2.0 2.0 25.0 Absolute expenditure index (1970 = 100), 100 222 255 259 159 Number of children eligible for basic education (millions)b 16.2 19.7 24.9 25.6 .58.6 Pakistan Expenditure per eligible child (1987 dollars) 7.9 9.4 13.4 - Expendirure as a percent of GNP 1.1 1.6 1.6 Absoluteexpenditureindex(1970= 100)1 100 150 277 - Number of children eligible for basic education (millions)b 21.0 26.6 34.4 42.2 101.4 - Not available. a. Absolute expenditures on basic education in real 1987 U.S. dollars used to calculate indexes for absolute expenditures on education. b. Calculated using enrollment rates and number of students in primary and secondary school. Source: UNESCO (various years) and World Bank (various years c). Birdsall, Ross, and Sabot 489 Because of substantial multicollinearity, the two variables we add in the sec- ond column of table 2-the ratio of manufacturing exports to GDP and the inter- action of the secondary school enrollment rate and that measure of exports- are not significant at the 5 percent level. However, joint tests reject the null hypothesis that the coefficients are zero.4 The results support the contention that the contribution of education to economic growth tends to be greater in countries in which manufactured exports are a higher proportion of GDP. The results are consistent with the view that in East Asia the stimulus that the greater supply of human capital gave to economic growth was augmented by the export orientation of those economies and the resulting skill-demanding growth paths they followed. Our finding is also consistent with the shift of East Asian export- ers into more technologically sophisticated and more capital- and skill-intensive goods, as rapidly rising wages of unskilled labor eroded international competi- tiveness in labor-intensive manufactured goods. Thus the combination of a greater supply of education and a greater demand for educated workers contributed to faster economic growth in East Asia than in other developing regions. In the other half of the virtuous circle of education and growth, rapid eco- nomic growth and altered household behavior gave positive feedback to greater investment in education in East Asia, and resources for education expanded rapidly. In the ratio of public expenditures on basic education to the number of school-age children, in East Asia rapid economic growth increased the numera- tor while declining fertility (a result of earlier investments in education) reduced the denominator. Table 3 indicates that in 1970, public expenditure on basic education per eligible child was not much higher in Korea ($95 in 1987 dollars) than in Mexico ($68). However, from 1970 to 1989 it more than quadrupled in Korea, to $433, whereas in Mexico it did not even double. As a consequence, in 1989, Mexican public expenditure on basic education per eligible child was only 26 percent of Korean expenditure per eligible child. What accounted for this divergence? It was not government commitment: public expenditure as a percentage of GNP over this period was actually declin- ing in Korea and rising in Mexico. For other East Asian economies as well, there is no evidence that greater government commitment to education produced the extraordinary performance in the provision of education.' In 1960 and in 1989, public expenditure on education as a percentage of CGNP was not much higher in East Asia than in other regions. In 1960 the share of education in public expen- 4. The F-statistic for the restriction that secondary school enrollments, export share, and interaction coefficients are each zero is 6.42, which exceeds the critical value for 1 percent significance with three and ninety degrees of freedom. More important, the corresponding F-statistic for the export share and interaction coefficients (6.37) also exceeds the critical value for I percent significance. 5. Government expenditure on education, expressed as a percentage of GNP, was not significant in explaining expected years of schooling in a cross-economy regression covering fifteen Asian and Latin American economies. See Tan arid Mingat (1992). 490 1-HE WORLD BANK E(:(ONOMI REVIEW. VOI Y, No) diture was 2.5 percent for East Asia, 2.4 percent for Sub-Saharan Africa, and 2.2 percent for all developing economies (Birdsall and Sabot 1994). Over the next three decades, all regions markedly increased the share of national output they invested in formal education, but in 1989, the share in Africa, 4.1 percent, was higher than the East Asian share of 3.7 percent, which barely exceeded the 3.6 percent average share for all developing economies (Birdsall and Sabot 1994). The initial conditions, the colonial legacy, are also not decisive in explaining why enrollment rates have been so much higher in East Asia than elsewhere. Although Korea had much higher enrollment rates in 1950 than did most other developing economies, the roughly 50- and 70-percentage-point increases since then in, respectively, primary school and secondary school enrollment rates ac- count for much of the current gap between Korea and other middle-income economies. Similar claims can be made for other East Asian economies. Two factors permitted a faster expansion of educational systems and enroll- ment rates in East Asia: faster economic growth and a more rapid decline of fertility. In Korea the absolute level of expenditure on basic education rose faster than in Mexico in part because GDP growth was faster. For example, from 1965 to 1975 GDP growth averaged 9.5 percent in Korea and 6.5 percent in Mexico (World Bank, various years b). This implies that over that period, given a con- stant share of GDP allocated to education, the resources available to the educa- tion sector in Korea rose by almost 400 percent; they rose also in Mexico but less, by about 250 percent. (Rapid growth also raises the demand for labor, hence wages, and in particular, the wages of teachers. Because the pay of teach- ers accounts for a large proportion of recurrent expenditure on education in low-income economies, the tendency for rising costs to reduce the benefits of rapid growth would be strong except for an important mitigating factor. Rapid accumulation of human capital in one period increases the potential supply of teachers in the next, thereby reducing the relative earnings premium that teach- ers command. Although growth induces increases in average wages, the wages of more-educated workers, including teachers, tend to rise at a slower rate. For example, in part because of low relative salaries, the per-pupil operating costs of primary schools are about 13 percent of GNP per capita in Indonesia and Malaysia, whereas they are almost 30 percent in francophone Africa; Tan and Mingat, 1992.) Over the two decades from 1970 to 1989 there was also a divergence be- tween Mexico and Korea in the number of school-age children. Although in Mexico the number of children eligible for basic education increased by nearly 60 percent, in Korea the number actually declined by 2 percent (Birdsall and Sabot 1994). The difference in fertility rates, of which these diverging trends are a reflection, was in part caused by earlier differences in educational attainment, in particular, in the educational attainment of women. Figure 4 illustrates another feedback mechanism. In the absence of the skill intensification of labor demand, educational expansion (from S0 to Si) reduces the rate of return to education (from R.) to R,) in the typical low-income economy. Birdsall, Ross, and Sabot 491 This lower return then slows further investment in education. Investment in education by households is greater in East Asia because the demand for edu- cated workers is greater, and consequently the returns to investment in school- ing are higher. S, - St is the difference between the East Asian economy and the typical developing economy in the level of investment in human capital induced by the greater demand for educated labor in the East Asian economy. A stronger demand for educated workers elicits a greater supply. II. EDUCATION LOWERS INEQUALITY, WHICH STIMULATES INVESTMENT IN EDUCATION In a cross section of more than eighty economies, there is a clear correlation between basic-education enrollment rates and lower levels of income inequality (Birdsall and Sabot 1994). There is evidence that causality runs in both direc- tions in this relationship-in a virtuous cycle of greater education as a cause and a consequence of lower inequality. Because educational expansion increases the number of workers holding high- wage jobs, income inequality can increase rather than decline; the change in the educational composition of the labor force causes a disequalizing effect analo- gous to that identified by Kuznets. With educational expansion, however, the relative abundance of educated workers increases, and another factor kicks in: the scarcity rents that the educated earn are eroded. The result is a compression of the educational structure of wages. This compression effect can offset the composition effect, leading to a reduction in the inequality of pay and hence in total income inequality (Knight and Sabot 1983, 1991). Differences between Brazil and Korea in educational opportunities in the 1970s and 1980s illustrate how, if education expands rapidly to all segments of a popu- lation, the inequality of pay can be reduced by the dominance of the compres- sion effect. In 1980, Brazil and Korea had similar levels of per capita income. Both countries had achieved universal primary education, and in both countries secondary and tertiary enrollment had increased. However, there was a large and widening gap between Brazil and Korea in the extent to which enrollment rates for secondary and tertiary education grew. In Korea between the mid- 1970s and mid-1980s the proportion of high school and postsecondary gradu- ates in a large random sample of the wage-labor force increased so sharply that the proportion of workers with an elementary education or less declined from nearly 20 percent to under 8 percent. By contrast, for the Brazil sample, the comparable proportion fell from over 70 percent to just over 60 percent (Park, Ross, and Sabot, forthcoming). Table 4 presents the results of estimating a standard Mincerian log wage regression on the Korean and Brazilian samples using 1976 and then 1985 data. In Korea, with the spread of high school and postsecondary education, the extra earnings of graduates at these levels fell. In 1976, workers with a high school education earned 47 percent more than primary school graduates; by 1 986 that 492 1TIF WS)RI 1) BANK F.( ONONMIC RFVIFW, VOL. 9, NI). Table 4. Male Wage Structure in Brazil and the Republic of Korea, 1976 and 1985 Brazil Korea, Rep. of Variable 1976 1985 1976 1986 Premium to 0.488 0.449 - - primary schooling (55.68) (67.23) Premium to 0.958 0.886 0.176 0.092 secondary schooliog (85.70) (110.53) (19.66) (7.54) Premium Lto - - 0.473 0.296 high school (48.19) (23.40) Premium to 1.593 1.508 0.969 0.655 higher edutcationi (100.22) (127.40) (71.48) (42.06) Experience 0.045 0.(48 0.067 0.078 (64.97) (8.3.91) (61.90) (69.61) Experience2 -0.0006 -0.0007 -0.001 -0.001 (61.41) (79.27) (39.13) (50.27) Constant 1.149 7.(43 10.231 11.779 R` 0.546 0.562 0.532 0.449 Number of observations 85.106 118,000 23,838 24,486 Mean log of wages 1.864 8.095 11.363 12.895 - Not available. Note: Dummy variables were included to control for region, occupation, industry, and head of household (Brazil only); t-statistics are in parentheses. Source: Authors' calculations. premium had declined to 30 percent. Similarly, the extra earnings of workers with higher education fell from 97 percent to 66 percent. In Brazil, where the expansion of enrollment rates in secondary and tertiary education was less marked, the premia to workers with secondary and tertiary schooling barely changed. In Korea the compression effect shown in table 4 offset the composition ef- fect of increased education; the net effect of educational expansion was to re- duce the log variance in wages by 22 percent. By contrast, in Brazil the compo- sition effect dominated; the net effect of the smaller educational expansion over the decade was to increase that measure of wage inequality by 4 percent. By 1985, 25 percent of the gap between Brazil and Korea in the inequality of pay could be explained by differences in the educational attainment of the labor force (Park, Ross, and Sabot, forthcoming). The reduction in the wage premium to educated workers in Korea between the 1970s and 1980s and the lack of any such decline in Brazil suggest that the more rapid spread of education in Korea, and elsewhere in East Asia, contributed over time to the below-average levels of income inequality there compared with those in Latin America. Closing the second virtuous circle, there has been a feedback effect from low inequality of income to high enrollment rates and thus greater education. Low inequality of income increases household demand for education and probably increases the public supply. On the demand side, it is likely that budgetary constraints combined with poor access to capital markets mean that poor households do not invest in their Birdsall. Ross, and Sabot 493 children's human capital even when the returns are high. The pressing need to use income simply to subsist crowds out high-return investments and constrains the demand for education. Table 5 pairs East Asian economies with other econo- mies having similar levels of average per capita income but with considerably higher levels of income inequality and hence lower absolute incomes of the poor. For example, although the per capita income of Brazil (in 1983) slightly ex- ceeded average income in Malaysia (in 1987), the bottom quintile received only 2.4 percent of total income in Brazil compared with 4.6 percent of total income in Malaysia. The per capita income of the bottom quintile in Brazil was thus only 54 percent of the per capita income of the bottom quintile in Malaysia. Given an income elasticity of demand for basic education of 0.50 (a conserva- tive figure), if the distribution of income were as equal in Brazil as in Malaysia, enrollments among poor Brazilian children would be more than 40 percent higher. Low levels of income inequality may have an influence on the supply side as well as the demand side of the market for education. For the government to provide subsidized basic educational opportunities for a large segment of the school-age population when the distribution of income is highly unequal, the tax burden on the rich has to be heavy. High-income families are likely to resist, for example, by trying to channel public spending for education into subsidies to higher education where their own children will be the beneficiaries (Birdsall and James 1993). If incomes are more equally distributed, the incidence of taxes to finance mass education need not be as concentrated, and resistance to such pro- grams by high-income families is likely to be weaker. Although public expenditure on education as a share of GNP is not higher in East Asia than in other developing regions, the share of public expenditure on education allocated to primary and secondary, as opposed to higher, education has been consistently larger in East Asia. Korea and Venezuela are extreme examples. In 1985, Korea allocated just 10 percent of its public education bud- get to higher education, but Venezuela allocated 43 percent (Birdsall and Sabot 1 994). As a result, although both countries spent a similar percentage of GDP on education, Korea spent more than twice as much as Venezuela on primary and secondary education, a fact that surely helps to explain Korea's more abundant and higher-quality educational opportunities. Birdsall and Sabot (1994) present similar data for other East Asian economies. Findings of analyses across economies are also consistent with the likelihood that more equality in the distribution of income leads to more investment in education. In a study of the determinants of secondary school enrollment rates, Williamson (1993) found that more egalitarian societies (measured using the ratio of the share of total income of the bottom 40 percent to the share of the top 20 percent) had higher secondary school enrollment rates. Using his esti- mated equation to decompose the difference in enrollment rates between Brazil and Korea, we found that none of the 27-percentage-point difference in enroll- ment rates could be explained by the difference in G(NP per adult. Nearly all of that portion of the gap that could be explained was caused by the greater Table 5. Absolute Income Share of LIowest Quintile in Selected Economies and Years Poorest 20 percent of households Total GNP GNP per capita Population (millions of Income share Absolute income Per capita income Economy (U.S. dollars) (millions) U.S. dollars) (percent) (U.S. dollars) (U.S. dollars) Botswana, 1986 840 1.1 924 2.5 23 115 Brazil, 1983 1,880 129.7 243,836 2.4 5,852 226 IC Costa Rica, 1986 1,480 2.6 3,848 3.3 127 254 4 Indonesia, 1976 240 135.2 32,448 6.6 2,141 79 Indonesia, 1987 450 171.4 77,130 8.8 6,787 251 Kenya,1976 240 13.8 3,312 2.6 86 31 Korea, Rep. of, 1976 670 36.0 24,120 5.7 1,375 191 Malaysia, 1987 1,810 16.5 29,865 4.6 1,374 416 Philippines, 1985 580 54.7 31,726 5.5 1,745 160 Source: World Bank (various years b). Birdsall, Ross, and Sabot 495 inequality in the distribution of income in Brazil. Were income distributed as equally in Brazil as in Korea, Korea's secondary school enrollment rate, instead of being 27 percentage points higher, would be only 6 percentage points higher. III. Low INEQUALITY OF INCOME STIMULATES GROWTH Levels of educational attainment above those predicted for economies having similar income levels help explain why, contrary to the conventional wisdom, we observe in East Asia both rapid growth and low levels of income inequality. Investment in education augments growth and reduces inequality and, closing a virtuous circle, rapid growth and low inequality induce higher investment in education. The adoption of a labor-demanding growth strategy that was in ac- cord with factor endowments and comparative advantage also contributed to both rapid growth and low inequality (World Bank 1993; Baner'i, Campos, and Sabot, forthcoming). Banerji, Campos, and Sabot (forthcoming) argue that in East Asia the incomes of those in the bottom half of the distribution of income were pulled up by the rapid growth of demand for wage labor. They show that in Korea from 1970 to 1990 wage employment in manufacturing grew at an annual rate of 19 percent and real wages in manufacturing increased at an an- nual rate of 9 percent. By contrast, in India over the same period both wage employment and real wages in manufacturing grew at less than 2 percent a year. But does low inequality stimulate growth independently of its effect on edu- cation? In this section we first present some cross-economy evidence that it does. We then suggest some reasons why. We point out how public policy in East Asia counteracted the Kaldor and Kuznets effects, thereby weakening the tendency for rapid growth to be associated with high inequality. Econometric Results Recent cross-economy studies of growth suggest a negative relationship be- tween income inequality and average annual growth in per capita GDP (Alesina and Rodrik 1994; Persson and Tabellini 1994). For the period 1970-88, Clarke (forthcoming) found this relationship to be robust to the choice of (five) in- equality measures and alternative specifications of the explanatory variables. We modified the Clarke regression because we wanted estimates for the period 1960-85 for comparability with our other growth regressions and because in Clarke's data set inequality observations for some countries are as recent as 1980. Current inequality, arguably, is simultaneously determined with growth. We assembled economy-by-economy observations from a variety of sources on the ratio of the income shares of the top 20 percent of population and the bot- tom 40 percent. We chose the earliest available observation and dropped obser- vations where the measure postdated 1970. This procedure yielded a data set with seventy-four observations.6 6. The list of econo