复旦大学：《发展经济学》阅读材料与文献_The Coast–Noncoast Income Gap, Productivity, and Regional Economic Policy in China

NAL OF COMPARATIVE ECONOMICS 25, 220-236(1997) CLE NO. JE971462 The Coast-Noncoast Income Gap, Productivi and Regional Economic Policy in China Belton M. Fleisher The Ohio State University, Columbus, Ohio 43210 Jian Chen Kenyon College, Gambier. Ohio 43022 Received August 21. 1996: revised June 2. 1997 Fleisher, Belton M, and Chen, Jian-The Coast-Noncoast Income Gap, Productiv ity, and Regional Economic Policy in China We postulate that inferior factor productivity in Chinas noncoastal provinces is principal reason for their lower economic growth despite high investment rates relative to provincial GDP. We find that total factor productivity is roughly twice as high in the coastal provinces and estimate that investment in higher education and foreign direct investment helps explain the productivity gap. We speculate that despite its elatively modest estimated return, investment in infrastructure may be necessary to attract foreign direct investment and to retain university graduates in the interior Comp. Econom., October 1997, 25(2), pp. 220-236. The Ohio State University, Co- lumbus, Ohio, 43210; Kenyon College, Gambier, Ohio 43022. 0 1997 Academic Pres Journal of Economic Literature Classification Numbers: O15, 018, 047, 053 INTRODUCTION This paper is an attempt to understand the persistent and widening income gap between coastal and interior China and to suggest appropriate policies I This paper has benefited from the help of Dongwei Su and the comments of Mario Crucini, Pok-Sang Lam, Guang H, Wan, Shaowen Wu, Yong Yin, and two anonymous referees. We also thank Gary Jefferson and Barry Naughton, who offered extensive and valuable suggestions as scussants in the AEA session""Empirical Analysis of the Chinese Economy, New Orleans 1997, and participants in a seminar at the Center for Chinese Studies, University of Michigan, including Robert Dernberger, Junling Hu, David Li, Kenneth Lieberthal, and Albert Park. Xiaojun Wang provided excellent research assistance. Please send communications to B M. Fleisher, fleisher. l@osu. edu 0147-5967/97$2500 opyright e 1997 by Academic Press 220 All nights of reproduction in any form reserved

JOURNAL OF COMPARATIVE ECONOMICS 25, 220–236 (1997) ARTICLE NO. JE971462 The Coast–Noncoast Income Gap, Productivity, and Regional Economic Policy in China1 Belton M. Fleisher The Ohio State University, Columbus, Ohio 43210 and Jian Chen Kenyon College, Gambier, Ohio 43022 Received August 21, 1996; revised June 2, 1997 Fleisher, Belton M., and Chen, Jian—The Coast–Noncoast Income Gap, Productiv￾ity, and Regional Economic Policy in China We postulate that inferior factor productivity in China’s noncoastal provinces is a principal reason for their lower economic growth despite high investment rates relative to provincial GDP. We find that total factor productivity is roughly twice as high in the coastal provinces and estimate that investment in higher education and foreign direct investment helps explain the productivity gap. We speculate that despite its relatively modest estimated return, investment in infrastructure may be necessary to attract foreign direct investment and to retain university graduates in the interior. J. Comp. Econom., October 1997, 25(2), pp. 220–236. The Ohio State University, Co￾lumbus, Ohio, 43210; Kenyon College, Gambier, Ohio 43022. q 1997 Academic Press Journal of Economic Literature Classification Numbers: O15, O18, O47, O53. 1. INTRODUCTION This paper is an attempt to understand the persistent and widening income gap between coastal and interior China and to suggest appropriate policies 1 This paper has benefited from the help of Dongwei Su and the comments of Mario Crucini, Pok-Sang Lam, Guang H. Wan, Shaowen Wu, Yong Yin, and two anonymous referees. We also thank Gary Jefferson and Barry Naughton, who offered extensive and valuable suggestions as discussants in the AEA session ‘‘Empirical Analysis of the Chinese Economy,’’ New Orleans, 1997, and participants in a seminar at the Center for Chinese Studies, University of Michigan, including Robert Dernberger, Junling Hu, David Li, Kenneth Lieberthal, and Albert Park. Xiaojun Wang provided excellent research assistance. Please send communications to B. M. Fleisher, fleisher.1@osu.edu. 0147-5967/97 $25.00 Copyright q 1997 by Academic Press All rights of reproduction in any form reserved. 220 AID JCE 1462 / 6w10$$$121 09-30-97 14:16:24 cea

PROVINCIAL PRODUCTIVITY IN CHINA to help the lagging interior provinces catch up to their more prosperous counterparts Aware of the political danger and perhaps also sensitive to the inequity of favoring coastal development, the central government has taken steps to pro- mote the growth of enterprises in the interior, focusing particular attention on steps to encourage investment in rural enterprises(Yang and Wei, 1996) Evidently this strategy has yet to produce the desired results. We hypothesize that a use of the persistent and widening income gap between the coast and interior is lower factor productivity in the noncoastal provinces We report tests of hypotheses that total factor productivity(TFP)and TFP growth vary across provinces. We identify factors contributing to the produc tivity gap and derive implications for policies that may help the interior provinces approach parity with their coastal counterparts The rest of the paper proceeds as follows. In Section 2, we deal with methodological issues and outline the basic theoretical and econometric proce dure. Section 3 contains our econometric results. The last section summarizes and draws policy implications MODELING TFP AND TFP GROWTH 2.1. Methodological Issues The first methodological issue addressed is frontier- versus standard produc tion-function estimation Lau and Brada(1995) point out that an advantage of using the frontier approach is obtaining the relative contributions of techno- logical growth and improvements in technical efficiency to TFP growth, which is important in forecasting how long current growth trends will continue. We have chosen not to use a frontier estimation approach for two reasons. First, accuracy in allocating the" residual" of the production relationship between technical efficiency and technological progress depends critically on the accu- racy with which inputs have been measured Because we focus on all sectors of the Chinese economy at the provincial level we do not have access to accurate capital-stock data. 3 See, for example, Chen and Fleisher(1996)and Yang and Wei(1996). Chen and Fleisher contains references to earlier studies on the provincial distribution of income and production. In particular, rising per capita income in 10 coastal provinces, which we define to include Beijing because of its location and to exclude Guangxi and Hainan because of inadequate data, has outstripped growth in the interior, so that between 1978 and 1993 the coast/noncost ratio of mean GDP per capita grew from 2.53 to 2.82, or 11%. s As described below, we are able to estimate the desired production-function parameters without data on the capital stock because we estimate a growth model, which requires data investment. Discussion of difficulties in using capital stock data in China to estimate aggregate production functions can be found in Chen et al. (1988)and Chow(1984), especially pp. 202 205. We also attempt to correct for inclusion of nonproductive investment in the data described below

PROVINCIAL PRODUCTIVITY IN CHINA 221 to help the lagging interior provinces catch up to their more prosperous counterparts. Aware of the political danger and perhaps also sensitive to the inequity of favoring coastal development, the central government has taken steps to pro￾mote the growth of enterprises in the interior, focusing particular attention on steps to encourage investment in rural enterprises (Yang and Wei, 1996). Evidently this strategy has yet to produce the desired results.2 We hypothesize that a major cause of the persistent and widening income gap between the coast and interior is lower factor productivity in the noncoastal provinces. We report tests of hypotheses that total factor productivity (TFP) and TFP growth vary across provinces. We identify factors contributing to the produc￾tivity gap and derive implications for policies that may help the interior provinces approach parity with their coastal counterparts. The rest of the paper proceeds as follows. In Section 2, we deal with methodological issues and outline the basic theoretical and econometric proce￾dure. Section 3 contains our econometric results. The last section summarizes and draws policy implications. 2. MODELING TFP AND TFP GROWTH 2.1. Methodological Issues The first methodological issue addressed is frontier- versus standard produc￾tion-function estimation. Lau and Brada (1995) point out that an advantage of using the frontier approach is obtaining the relative contributions of techno￾logical growth and improvements in technical efficiency to TFP growth, which is important in forecasting how long current growth trends will continue. We have chosen not to use a frontier estimation approach for two reasons. First, accuracy in allocating the ‘‘residual’’ of the production relationship between technical efficiency and technological progress depends critically on the accu￾racy with which inputs have been measured. Because we focus on all sectors of the Chinese economy at the provincial level we do not have access to accurate capital-stock data.3 2 See, for example, Chen and Fleisher (1996) and Yang and Wei (1996). Chen and Fleisher contains references to earlier studies on the provincial distribution of income and production. In particular, rising per capita income in 10 coastal provinces, which we define to include Beijing because of its location and to exclude Guangxi and Hainan because of inadequate data, has outstripped growth in the interior, so that between 1978 and 1993 the coast/noncoast ratio of mean GDP per capita grew from 2.53 to 2.82, or 11%. 3 As described below, we are able to estimate the desired production-function parameters without data on the capital stock because we estimate a growth model, which requires data on investment. Discussion of difficulties in using capital stock data in China to estimate aggregate production functions can be found in Chen et al. (1988) and Chow (1984), especially pp. 202– 205. We also attempt to correct for inclusion of ‘‘nonproductive’’ investment in the data as described below. AID JCE 1462 / 6w10$$$121 09-30-97 14:16:24 cea

FLEISHER AND CHEN Our second reason is, in a sense, philosophical and rests on the belief that there is an inherent arbitrariness in distinguishing between the levels of technology and technical efficiency. One source of this arbitrariness is the need to specify the mathematical form of the time paths of technical progress and technical efficiency. The allocation of TFP change between technical progress and changes in technical efficiency depends on the time paths as- umed. Arbitrariness also arises in attempting to allocate the causes of failure to adopt""best'available technology, which may arise from: (i) failure to invest in physical capital in which the technology is embodied; (ii) lack of human capital, or knowledge of the best available technology; and (iii)adverse incentives due to market institutions. government controls. etc. Economic reforms since 1979 are designed to take care of item(iii) and are evidently reflected in the increased efficiency identified by Lau and Brada in the early years of the reform era. If TFP is below its maximum due to(i) or(ii),is this necessarily"inefficient?" The answer depends in part on one's view of capital markets, available resources, and capital constraints. This study fo- cuses on(i)and (ii)as possible explanations of provincial differences in TFP The second methodological issue is specification of the form of the produc tion function In our empirical work, we assume a Cobb-Douglas production function with Hicks-neutral technology. G. S. Maddala(1979) points out that within the class of functions... Cobb-Douglas, generalized Leontief, homogeneous translog, and homogeneous quadratic, differences in the func tional form produce negligible differences in measures of multi-factor produc- tivity. ' Imposing the Cobb-Douglas specification in the context of the Solow growth model is analogous to the standard growth-accounting technique of using hypothetical factor elasticities to compute TFP or TFP change as a esidual. We, however, estimate our(constant)factor elasticities simultane- ously with our estimates of TFP and TFP growth 2.2. An Empirical Model The Cobb-Douglas production function with Hicks-neutral technology is Y,,=A,KhLiI'e'u where i and t index the provinces and time, respectively. We specify Ais A, oe,+2. as the systematic component of TFP at time t, which includes all factors contributing to output other than labor L and physical capital K at Another potentially serious problem, however, is pointed out by Guang H. Wan(1995), who argues that alternative specifications, e.g, Hicks-neutral, Harrod-neutral, can influence estimates of the degree of technical change

222 FLEISHER AND CHEN Our second reason is, in a sense, philosophical and rests on the belief that there is an inherent arbitrariness in distinguishing between the levels of technology and technical efficiency. One source of this arbitrariness is the need to specify the mathematical form of the time paths of technical progress and technical efficiency. The allocation of TFP change between technical progress and changes in technical efficiency depends on the time paths as￾sumed. Arbitrariness also arises in attempting to allocate the causes of failure to adopt ‘‘best’’ available technology, which may arise from: (i) failure to invest in physical capital in which the technology is embodied; (ii) lack of human capital, or knowledge of the best available technology; and (iii) adverse incentives due to market institutions, government controls, etc. Economic reforms since 1979 are designed to take care of item (iii) and are evidently reflected in the increased efficiency identified by Lau and Brada in the early years of the reform era. If TFP is below its maximum due to (i) or (ii), is this necessarily ‘‘inefficient?’’ The answer depends in part on one’s view of capital markets, available resources, and capital constraints. This study fo￾cuses on (i) and (ii) as possible explanations of provincial differences in TFP. The second methodological issue is specification of the form of the produc￾tion function. In our empirical work, we assume a Cobb–Douglas production function with Hicks-neutral technology. G. S. Maddala (1979) points out that ‘‘within the class of functions . . . Cobb–Douglas, generalized Leontief, homogeneous translog, and homogeneous quadratic, differences in the func￾tional form produce negligible differences in measures of multi-factor produc￾tivity.’’ Imposing the Cobb–Douglas specification in the context of the Solow growth model is analogous to the standard growth-accounting technique of using hypothetical factor elasticities to compute TFP or TFP change as a residual. We, however, estimate our (constant) factor elasticities simultane￾ously with our estimates of TFP and TFP growth.4 2.2. An Empirical Model The Cobb–Douglas production function with Hicks-neutral technology is given by Yi,t Å Ai,tKb i,tL10b i,t eei,t , (1) where i and t index the provinces and time, respectively. We specify Ai,t Å Ai,0eg1i t/g2i t 2 as the systematic component of TFP at time t, which includes all factors contributing to output other than labor L and physical capital K at 4 Another potentially serious problem, however, is pointed out by Guang H. Wan (1995), who argues that alternative specifications, e.g., Hicks-neutral, Harrod-neutral, can influence estimates of the degree of technical change. AID JCE 1462 / 6w10$$$122 09-30-97 14:16:24 cea

PROVINCIAL PRODUCTIVITY IN CHINA time t, g, as the rate of technological change, and Eu as an error term with the usual properties, which may also be viewed as random productivity shocks. The labor force evolves as Lio e"! where n, is the rate of labor force growth. Output per worker, a close correlate of income per capita, is y,=A.Ai- -b where y= Y/L and k= K/L From Eq (1), we specify a Solow growth equation B B which, on the basis of the assumption that convergence to the steady state occurs at the rate x(0<x< 1), leads to In In (In A1, 0+g1,+ g2/) (1-e-)ny/-1+l(3) Our TFP estimates are based on Eq. (3), which allows us to obtain all produc- tion-function parameters directly and simultaneously and does not require data on the capital stock 2. 3. Explaining Technological Change Making the standard growth accounting assumption that the error term in e above equations, uil, represents provincial productivity shocks, we define TFP in year t as TH Ai0 +814+ g2F + ui, and specify the following regression equation to explain provincial TFP differentials In amAmis-1+ asC+ ant agf+aoCI aioN Ti-I +U,(4) where the right-hand variables and hypothesized qualitative relationship with TFP are The quadratic trend term was suggested by an anonymous referee to capture a possible slowdown in TFP growth that may have occurred around 1985. This may be inferred by comparing the empirical results of Lau and Brada(1990)with those of wu(1995)

PROVINCIAL PRODUCTIVITY IN CHINA 223 time t, gi as the rate of technological change, and ei,t as an error term with the usual properties, which may also be viewed as random productivity shocks.5 The labor force evolves as Li,0 Å eni t , where ni is the rate of labor￾force growth. Output per worker, a close correlate of income per capita, is yi,t Å Ai,tk10b i,t where y å Y/L and k å K/L. From Eq. (1), we specify a Solow growth equation ln yi,t Å 1 1 0 b (ln Ai,0 / g1i / g2i t 2 ) / b 1 0 b ln si,t01 0 b 1 0 b ln ni,t / wi,t , (2) which, on the basis of the assumption that convergence to the steady state occurs at the rate l(0 õ l õ 1), leads to ln yi,t 0 ln yi,t01 Å (1 0 e0lt )F 1 1 0 b (ln Ai,0 / g1i / g2i t 2 ) / b 1 0 b ln si,t01 0 b 1 0 b ln ni,tG 0 (1 0 e0lt )ln yi,t01 / ui,t . (3) Our TFP estimates are based on Eq. (3), which allows us to obtain all produc￾tion-function parameters directly and simultaneously and does not require data on the capital stock. 2.3. Explaining Technological Change Making the standard growth accounting assumption that the error term in the above equations, ui,t , represents provincial productivity shocks, we define TFP in year t as ti,t Å Ai,0 / g1i t / g2i t 2 / ui,t and specify the following regression equation to explain provincial TFP differentials. ln ti,t Å a0 / ∑ 5 mÅ1 amxm,i,t01 / a6C / a7t / a8t 2 / a9Ct / a10ln ti,t01 / £i,t , (4) where the right-hand variables and hypothesized qualitative relationship with TFP are 5 The quadratic trend term was suggested by an anonymous referee to capture a possible slowdown in TFP growth that may have occurred around 1985. This may be inferred by comparing the empirical results of Lau and Brada (1990) with those of Wu (1995). AID JCE 1462 / 6w10$$$122 09-30-97 14:16:24 cea

24 FLEISHER AND CHEN x, is a measure of investment in housing(to correct for the inclusion of expenditure on new housing in total investment ), and is negative, x2 is a measure of the vintage of the physical capital stock, and egative, xs is a measure of investment in human capital, and is positive xa is a measure of infrastructure(highways, railways, and waterways) and is positive, xs is foreign direct investment(FDI) as a share of total investment, an Is positive, C is a dummy variable I for coastal provinces and Beijing t is the year of observation(1979=1..1993= 15), and U is an i.i.d. error term. We follow Wolff (1991)in including lagged TFP, Ti-l, in Eq. (4). Full definitions and sources of the variables are included in the Appendix 7 The rationale for including vintage as a contributing factor to TFP and TFP growth is neatly summarized by Wolff (1991). Although Wolff uses the rate of change of the capital stock as a proxy for vintage, we have chosen to define variable x2 as a weighted average of the age of existing capital, specifi I(y2i-I Iuxt-j+ 1)l, where I is real accumulation of fixed assets 8 The contribution of human capital to production is by now part of received knowledge. It would be appropriate to include investment in human capital in conjunction with physical-capital investment in Eq (3) We do not do this because data on the actual magnitude of human-capital (1988) stock data have been purged of housing and other nonproductive capital. Jefferson et al.(1992) use corrected data for state and collective industry. It would be ideal for such net capital stock data for each province, but constructing such data is a task that is far beyond our urrent resources. In order to solve the problem that annual data for this variable are not available 1978 to 1993, we use an instrument for housing in estimating Eq.(4). The instrument is obtained by regressing a measure of housing area(square meters) per capita on per capita real income. The predicted level of per capita housing is then used as the measure of variable x Unfortunately, variables x(infrastructure)and xs(FDI) are not available annually from 1978 to 1993. Therefore in our empirical work we have treated them nvironmental variables Details are contained in the notes to table 2 Data for accumulation of fixed assets is available after 1952 for all provinces in our sample. We deflate using a price index obtained from series on construction both in nominal prices and in]fixed Chen et al. (June, 1988)assert that the data on construction in fixed prices are unreliable. However, our alternative is to use the provincial national income deflator that can be obtained by comparing national income and national income at fixed prices. We chose to use the construction deflator on the assumption that it provides an index closer to the correct one for accumulation than any alternative. We also used the same data to construct a variable(AK/ K)=(n2-o Iy), which is conceptually similar to the variable used by Wolff. The empirical results are not very sensitive to which of these variables is used to estimate Eq (4)

224 FLEISHER AND CHEN x1 is a measure of investment in housing (to correct for the inclusion of expenditure on new housing in total investment), and is negative,6 x2 is a measure of the vintage of the physical capital stock, and is negative, x3 is a measure of investment in human capital, and is positive, x4 is a measure of infrastructure (highways, railways, and waterways), and is positive, x5 is foreign direct investment (FDI) as a share of total investment, and is positive, C is a dummy variable Å 1 for coastal provinces and Beijing, t is the year of observation (1979 Å 1 rrr 1993 Å 15), and £i,t is an i.i.d. error term. We follow Wolff (1991) in including lagged TFP, tt01 , in Eq. (4). Full definitions and sources of the variables are included in the Appendix.7 The rationale for including vintage as a contributing factor to TFP and TFP growth is neatly summarized by Wolff (1991). Although Wolff uses the rate of change of the capital stock as a proxy for vintage, we have chosen to define variable x2 as a weighted average of the age of existing capital, specifi- cally, Vi,t Å (t jÅ1 [(Ii,j/(t jÅ1 Ii,j)(t 0 j / 1)], where Ij is real accumulation of fixed assets.8 The contribution of human capital to production is by now part of received knowledge. It would be appropriate to include investment in human capital in conjunction with physical-capital investment in Eq. (3). We do not do this because data on the actual magnitude of human-capital 6 Chen et al. (1988) report estimates of production functions for state industry in which capital￾stock data have been purged of housing and other nonproductive capital. Jefferson et al. (1992) use corrected data for state and collective industry. It would be ideal for us to use such net capital stock data for each province, but constructing such data is a task that is far beyond our current resources. In order to solve the problem that annual data for this variable are not available for 1978 to 1993, we use an instrument for housing in estimating Eq. (4). The instrument is obtained by regressing a measure of housing area (square meters) per capita on per capita real income. The predicted level of per capita housing is then used as the measure of variable x1 . 7 Unfortunately, variables x4 (infrastructure) and x5 (FDI) are not available annually from 1978 to 1993. Therefore in our empirical work we have treated them as ‘‘environmental’’ variables. Details are contained in the notes to Table 2. 8 Data for accumulation of fixed assets is available after 1952 for all provinces in our sample. We deflate using a price index obtained from series on construction both in nominal prices and in fixed prices. Chen et al. (June, 1988) assert that the data on construction in fixed prices are unreliable. However, our alternative is to use the provincial national income deflator that can be obtained by comparing national income and national income at fixed prices. We chose to use the construction deflator on the assumption that it provides an index closer to the correct one for accumulation than any alternative. We also used the same data to construct a variable (DK/ K)i,t Å (Ii,t/(t jÅ0 Ii,j), which is conceptually similar to the variable used by Wolff. The empirical results are not very sensitive to which of these variables is used to estimate Eq. (4). AID JCE 1462 / 6w10$$$122 09-30-97 14:16:24 cea

PROVINCIAL PRODUCTIVITY IN CHINA TABLE ESTIMATES OF TFP AND TFP GROWTH Dependent variable log difference real GDP per capita(1978-1993) Regression coefficients Independent variable (absolute /-statisticsp In(n) 0.152(9.62) Employment growth rate 0.37(2.70 0.74(154) Jilin, Guangxi, Hainan, Qinghai, and Tibet are excluded because of insufficient data. The TFP and linear trend coefficients(not shown here, but illustrated in Fig. I)are all highly significant, as are most of the quadratic trend coefficients investment are very difficult to construct. Doing so for China would be an extremely time- and resource-expensive project(see Jorgenson and Fraumeni, 1992). We therefore have elected to estimate the impact of human capital by the flow of graduates in the second stage of our re search. We measure infrastructure by the aggregate length of waterways paved highway, and trunk railway per square kilometer of area. Foreign agement technology. 4 ably embodies the latest in production and man- 3. ECONOMETRIC RESULTS The estimates of the coefficients of investment share, employment growth and lagged percapita GDP for Eq. (3)are shown in Table 1, and the estimates Despite lack of data on human-capital investment as such, Mankiw et al. (1992)do include a proxy in their well-known study io A referee and others who have commented on earlier drafts of this paper correctly point out that the flow of graduates from universities in a province is only a proxy for the change in the province s population or labor force with university degrees, as there is a significant migration of university graduates toward the"bright lights" in coastal provinces, especially the major cities. We would, of course, have used information on the population of educated workers had annual data been available. Commentators have also noted that any correlation between university education and TFP may reflect the impact of lower levels of educational attainment or even the attainment of literacy. Our attempts to deal with these comments are indicated below I It has also been pointed out to us that our measure of transportation infrastructure is onl a crude approximation and may well be poorly correlated with an interior province's access to the coast, which is critical for export-oriented industries 12See Shang-Jin Wei(1993)for a similar view

PROVINCIAL PRODUCTIVITY IN CHINA 225 TABLE 1 ESTIMATES OF TFP AND TFP GROWTHa Dependent variable log difference real GDP per capita (1978–1993) Regression coefficients Independent variable (absolute t-statistics)b ln(I/Y) 0.152 (9.62) Employment growth rate 00.37 (2.70) ln yt01 00.74 (15.4) a Jilin, Guangxi, Hainan, Qinghai, and Tibet are excluded because of insufficient data. b The TFP and linear trend coefficients (not shown here, but illustrated in Fig. 1) are all highly significant, as are most of the quadratic trend coefficients. investment are very difficult to construct. Doing so for China would be an extremely time- and resource-expensive project (see Jorgenson and Fraumeni, 1992).9 We therefore have elected to estimate the impact of human capital by the flow of graduates in the second stage of our re￾search.10 We measure infrastructure by the aggregate length of waterways, paved highway, and trunk railway per square kilometer of area.11 Foreign direct investment presumably embodies the latest in production and man￾agement technology.12 3. ECONOMETRIC RESULTS The estimates of the coefficients of investment share, employment growth, and lagged percapita GDP for Eq. (3) are shown in Table 1, and the estimates 9 Despite lack of data on human-capital investment as such, Mankiw et al. (1992) do include a proxy in their well-known study. 10 A referee and others who have commented on earlier drafts of this paper correctly point out that the flow of graduates from universities in a province is only a proxy for the change in the province’s population or labor force with university degrees, as there is a significant migration of university graduates toward the ‘‘bright lights’’ in coastal provinces, especially the major cities. We would, of course, have used information on the population of educated workers had annual data been available. Commentators have also noted that any correlation between university education and TFP may reflect the impact of lower levels of educational attainment or even the attainment of literacy. Our attempts to deal with these comments are indicated below. 11 It has also been pointed out to us that our measure of transportation infrastructure is only a crude approximation and may well be poorly correlated with an interior province’s access to the coast, which is critical for export-oriented industries. 12 See Shang-Jin Wei (1993) for a similar view. AID JCE 1462 / 6w10$$$123 09-30-97 14:16:24 cea

FLEISHER AND CHEN YN◆HEc)◆Js[C) -P 1978 Yuan FIG. 1. Provincial TFP and TFP growth, 1988 of TFP and TFP growth for the year 1988 are depicted in Fig. l, with coastal provinces indicated by the( C)notation. Our estimates of the determinants of TFP and TFP growth are containe in Table 2. As can be seen by comparing the second and fourth columns with the first and third columns, respectively, the variables other than trend and the coast-noncoast dummy can account for virtually all of the coast-noncoast productivity gap. The coefficient of capital vintage, while of the hypothesized The empirical formulation of Eq. (3)uses the arithmetic form, rather than the log form, of the employment-change variable, n, because annual employment growth in some provinces is occasionally negative. This specification is approximately equivalent to using the log of n+ I Thus, it is impossible to impose the constraint on the estimated factor elasticities implied by the constant-returns-to-scale assumption implicit in Eqs. (1)-(3). It is apparent I that the three city provinces, Beijing, Tianjin, and Shanghai, appear as"in the sense that they exhibit much higher than average TFP, One of our referees suggested that inclusion of these urban outliers" may have had a substantial effect on our econometric results. However, when Eqs. 3)and (4)are estimated without Beijing, Tianjin, and Shanghai, the estimated coefficients and their significance are changed very little I4 Based on the estimated coefficient of In(/n), the elasticity of capital is approximately 0. 2, implying a labor elasticity of approximately 0.8. This is at the low end of estimates of the Chen and Fleisher(1996), Chow (1994), and Chen et al. (1988). We suspect that one reason for this relatively low estimate is omission of a human-capital variable from Eq ( 3)

226 FLEISHER AND CHEN FIG. 1. Provincial TFP and TFP growth, 1988. of TFP and TFP growth for the year 1988 are depicted in Fig. 1, with coastal provinces indicated by the (C) notation.13,14 Our estimates of the determinants of TFP and TFP growth are contained in Table 2. As can be seen by comparing the second and fourth columns with the first and third columns, respectively, the variables other than trend and the coast–noncoast dummy can account for virtually all of the coast–noncoast productivity gap. The coefficient of capital vintage, while of the hypothesized 13 The empirical formulation of Eq. (3) uses the arithmetic form, rather than the log form, of the employment-change variable, n, because annual employment growth in some provinces is occasionally negative. This specification is approximately equivalent to using the log of n / 1. Thus, it is impossible to impose the constraint on the estimated factor elasticities implied by the constant-returns-to-scale assumption implicit in Eqs. (1)–(3). It is apparent in Fig. 1 that the three city provinces, Beijing, Tianjin, and Shanghai, appear as ‘‘outliers’’ in the sense that they exhibit much higher than average TFP. One of our referees suggested that inclusion of these ‘‘urban outliers’’ may have had a substantial effect on our econometric results. However, when Eqs. (3) and (4) are estimated without Beijing, Tianjin, and Shanghai, the estimated coefficients and their significance are changed very little. 14 Based on the estimated coefficient of ln(I/Y), the elasticity of capital is approximately 0.2, implying a labor elasticity of approximately 0.8. This is at the low end of estimates of the elasticity of production with respect to physical capital reported in the literature. See, for example, Chen and Fleisher (1996), Chow (1994), and Chen et al. (1988). We suspect that one reason for this relatively low estimate is omission of a human-capital variable from Eq. (3). AID JCE 1462 / 6w10$$$123 09-30-97 14:16:24 cea

PROVINCIAL PRODUCTIVITY IN CHINA TABLE 2 DETERMINANTS OF TFP Independent variables" Constant 0.46 4.35) 0.008 0.028 (5.08) (2.68 (2.92) 0.084 0.002 (3.08) 0.07) Coastal dummy X trend -0.028 -0.015 (0.66) (242 Trend- -0.00 (0.84) Trend× coastal dummy (0.73) In vintage - -0.0l1 -0.021 -0.4 (2.52) Transportation Route Length/Square Km 0.040 0.072 FDI/ 3.32 (0.8 (69.62) (4.78) Adj. R2 0.56 -0.001 For full variable definitions, see the Appendix. The infrastructure variable, transportatin route length, is measured as of 1986. In the ATFP formulation, In housing is used in differeon vintage and In( university grads/pop) are differenced. The change in transportation route expressed as 1/8 the change between 1986 and 1994, and FDI/ is the sum of FDI 3 divided by mean accumulation (D), as in the level equation. Further diagnostic tests are reported in Tables Al and A2. sign, is insignificant in both the level and the change regressions. The coefficient of the housing variable is indistinguishable from zero in the level regression, but negative and significant as hypothesized in the change regres- S One of our referees suggests that this lack of significance is because, in the pre-reform period and continuing into the early 1980s, investment, especially that undertaken by soe did not always embody the best available technology

PROVINCIAL PRODUCTIVITY IN CHINA 227 TABLE 2 DETERMINANTS OF TFP Dependent variables Independent variables a ln ti ln ti 0 ln tt01 Constant 4.97 0.52 0.061 0.46 (55.95) (4.35) (9.13) (4.40) Coastal dummy 0.70 .056 0.008 0.028 (5.08) (2.68) (1.34) (2.92) Trend 0.084 0.002 0.00 0.002 (3.08) (0.49) (0.07) (2.04) Coastal dummy 1 trend 00.028 00.015 (0.66) (2.42) Trend2 00.001 00.00 (0.84) (0.21) Trend2 1 coastal dummy 0.002 .001 (0.73) (2.94) ln vintaget01 00.011 00.021 (0.45) (0.91) ln housingt01 00.00 00.44 (0.00) (1.94) ln(university grads/pop)t01 0.014 0.011 (3.20) (2.52) Transportation Route Length/Square Km 0.040 0.072 (1.27) (0.70) FDI/I 1.94 3.32 (0.83) (1.60) ln tt01 0.94 00.049 (69.62) (4.78) Adj. R2 0.56 0.99 00.001 0.082 a For full variable definitions, see the Appendix. The infrastructure variable, transportation route length, is measured as of 1986. In the DTFP formulation, ln housing is used in difference form. Ln vintage and ln(university grads/pop) are differenced. The change in transportation route length is expressed as 1/8 the change between 1986 and 1994, and FDI/I is the sum of FDI 1984–1993 divided by mean accumulation (I), as in the level equation. Further diagnostic tests are reported in Tables A1 and A2. sign, is insignificant in both the level and the change regressions.15 The coefficient of the housing variable is indistinguishable from zero in the level regression, but negative and significant as hypothesized in the change regres- 15 One of our referees suggests that this lack of significance is because, in the pre-reform period and continuing into the early 1980’s, investment, especially that undertaken by SOE’s, did not always embody the best available technology. AID JCE 1462 / 6w10$$$123 09-30-97 14:16:24 cea

FLEISHER AND CHEN sion. The regression coefficients of the variables representing human capital transportation infrastructure, and foreign direct investment are all of the pre dicted sign, with I-statistics in the level regression of 3.20, 1.27, and 0.83, respectively. The coefficient of the natural log of university graduates/popula- tion is significant and slightly smaller in the change than in the level regres- sion. The regression coefficient of the infrastructure variable is not signifi cant by conventional standards in the level regression and it is insignificant by any reasonable standard in the change regression, although the point esti- mate of its magnitude is much larger in the change regression. Despite the low level of statistical significance, we take the estimated coefficients at face value and explore their implications for economic policy The estimated regression coefficient of FDI in the level regression is statisti cally insignificant, but it is marginally significant in the change regression and implies that raising the FDI/ ratio from the bottom of the distribution to the sample mean would increase TFP growth by about five percentage points per year. This seems implausibly large, but it nonetheless suggests that FDI may be an important source of TFP growth through the embodiment of new technology and managerial skills 16 One possible explanation of the increase in significance is that there is more collinearity in the TFP equation between lagged TFP and the housing variable than in the ATFP equation, where both are expressed in difference form In order to test for the sensitivity of our estimates to possible lack of correla the annual flow of university graduates and the presence of university graduates in provincial labor forces, we have done the following. First, using data available in various issues of the Statistical Yearbook of China from the population censuses of 1982 and 1990, we regress the log of the proportion of ty graduates in provincial populations on the log of new graduates from provincial universities. The regression coefficients(elasticities), which are highly significant, e 0.97 and 1.13 for 1982 and 1990, respectively. While it is apparently true that the flow of niversity-educated workers toward the"bright lights"has strengthened in recent years, it is also apparent that in the years covered by our sample, there is a strong relationship between annual flow of university graduates in a province and the proportion of university graduates in its population. Second, we regress the change in the proportion of university graduates in provincial opulations between 1982 and 1990 on the mean fiow proportion of newly graduated university students in the population in 1982 and 1990(multiplied by 7), obtaining an estimated regress coefficient of 0.75. This suggests that interprovincial differences in the flow proportion of gradu- tes underestimates the effect of interprovincial differences in the proportion of university. educated workers in the labor force. In the estimates of provincial rates of return to further education presented below we use our most" pessimistic" estimates of the relationship between the population proportion and the fiow proportion to obtain our final result. To test for the possibility that data on university graduates reflects the impact of literacy and/or lower levels of education on productivity, we have added the 1982 pre of illiterate and semiliterate persons in the provincial populations and the 1982 proportions of persons whose highest year of schooling is lower middle-school to the variables included in Table 2. The results are follows:(a) the coefficient and significance of the variables reported in Table 2 are virtually unaffected; and(b) the magnitudes and levels of significance of the middle-school and literacy variables are both extremely low

228 FLEISHER AND CHEN sion.16 The regression coefficients of the variables representing human capital, transportation infrastructure, and foreign direct investment are all of the pre￾dicted sign, with t-statistics in the level regression of 3.20, 1.27, and 0.83, respectively. The coefficient of the natural log of university graduates/popula￾tion is significant and slightly smaller in the change than in the level regres￾sion.17 The regression coefficient of the infrastructure variable is not signifi- cant by conventional standards in the level regression and it is insignificant by any reasonable standard in the change regression, although the point esti￾mate of its magnitude is much larger in the change regression. Despite the low level of statistical significance, we take the estimated coefficients at face value and explore their implications for economic policy. The estimated regression coefficient of FDI in the level regression is statisti￾cally insignificant, but it is marginally significant in the change regression and implies that raising the FDI/I ratio from the bottom of the distribution to the sample mean would increase TFP growth by about five percentage points per year. This seems implausibly large, but it nonetheless suggests that FDI may be an important source of TFP growth through the embodiment of new technology and managerial skills. 16 One possible explanation of the increase in significance is that there is more collinearity in the TFP equation between lagged TFP and the housing variable than in the DTFP equation, where both are expressed in difference form. 17 In order to test for the sensitivity of our estimates to possible lack of correlation between the annual flow of university graduates and the presence of university graduates in provincial labor forces, we have done the following. First, using data available in various issues of the Statistical Yearbook of China from the population censuses of 1982 and 1990, we regress the log of the proportion of university graduates in provincial populations on the log of new graduates from provincial universities. The regression coefficients (elasticities), which are highly significant, are 0.97 and 1.13 for 1982 and 1990, respectively. While it is apparently true that the flow of university-educated workers toward the ‘‘bright lights’’ has strengthened in recent years, it is also apparent that in the years covered by our sample, there is a strong relationship between the annual flow of university graduates in a province and the proportion of university graduates in its population. Second, we regress the change in the proportion of university graduates in provincial populations between 1982 and 1990 on the mean flow proportion of newly graduated university students in the population in 1982 and 1990 (multiplied by 7), obtaining an estimated regression coefficient of 0.75. This suggests that interprovincial differences in the flow proportion of gradu￾ates underestimates the effect of interprovincial differences in the proportion of university￾educated workers in the labor force. In the estimates of provincial rates of return to further education presented below we use our most ‘‘pessimistic’’ estimates of the relationship between the population proportion and the flow proportion to obtain our final result. To test for the possibility that data on university graduates reflects the impact of literacy and/or lower levels of education on productivity, we have added the 1982 proportion of illiterate and semiliterate persons in the provincial populations and the 1982 proportions of persons whose highest year of schooling is lower middle-school to the variables included in Table 2. The results are as follows: (a) the coefficient and significance of the variables reported in Table 2 are virtually unaffected; and (b) the magnitudes and levels of significance of the middle-school and literacy variables are both extremely low. AID JCE 1462 / 6w10$$$123 09-30-97 14:16:24 cea

PROVINCIAL PRODUCTIVITY IN CHINA en the regression coefficient of lagged TFP is highly significant and implies elasticity of TFP growth with respect to TFP level of just over 0.9. In the ATFP regression, lagged TFP remains highly significant; it is almost exactly equal to one minus the coefficient of lagged TFP in the TFP regression. As discussed in Wolff (1991), this evidence of TFP convergence may be due to disembodied technology transfer from across provinces 4. EVALUATION AND POLICY IMPLICATIONS The three variables that are amenable to policy control by both the central and the provincial governments are investment in human capital, investment in transportation infrastructure, and foreign direct investment. It is instructive to view the net social pecuniary return to additional higher education in terms of a standard human-capital formulation in which the flow return of increasing the number of university graduates per year by the proportion AE/E in province j IS a3(AE/EDY. Here, a3 is the estimated elasticity of TFP with respect to university graduates, E is the annual number of college graduates in province j, and Y is a measure of aggregate provincial output. The one- year cost of such an investment would be(AE/EDE(B(Y/N)+ D)where B is the elasticity of output with respect to labor, N, is a measure of the labor force in province j, and D is the direct cost in terms of physical capital, instructional staff, and support staff of one year of university education for one person. The expression A(Y /N) is the indirect cost, or foregone output, for one typical individual who leaves the labor force for one year to attend college. By setting the return and cost expressions equal to each other and assuming that the direct cost of one year of college is equal to the foregone- production cost, we can solve for the implicit rate of return to higher educa tion in the jth province, p. We obtain +1=(1+ 2B(E/N) where N is the number of years required to graduate from college. Equation (5)illustrates that the payoff to investment in human capital is greater the greater is the elasticity of TFP with respect to adding new university graduates, the smaller is the elasticity of production with respect to labor and the ratio of the current flow of new college graduates relative to the labor force, and We assume no depreciation and infinite lifetimes This is only a simplification. The nature of the solution is basically unaffected by this assumption

PROVINCIAL PRODUCTIVITY IN CHINA 229 The regression coefficient of lagged TFP is highly significant and implies an elasticity of TFP growth with respect to TFP level of just over 0.9. In the DTFP regression, lagged TFP remains highly significant; it is almost exactly equal to one minus the coefficient of lagged TFP in the TFP regression. As discussed in Wolff (1991), this evidence of TFP convergence may be due to disembodied technology transfer from across provinces. 4. EVALUATION AND POLICY IMPLICATIONS The three variables that are amenable to policy control by both the central and the provincial governments are investment in human capital, investment in transportation infrastructure, and foreign direct investment. It is instructive to view the net social pecuniary return to additional higher education in terms of a standard human-capital formulation in which the flow return of increasing the number of university graduates per year by the proportion DE/Ej in province j is a3(DE/Ej)Yj . 18 Here, a3 is the estimated elasticity of TFP with respect to university graduates, Ej is the annual number of college graduates in province j, and Yj is a measure of aggregate provincial output. The one￾year cost of such an investment would be (DE/Ej)Ej(b(Yj/Nj) / D) where b is the elasticity of output with respect to labor, Nj is a measure of the labor force in province j, and D is the direct cost in terms of physical capital, instructional staff, and support staff of one year of university education for one person. The expression b(Yj/Nj) is the indirect cost, or foregone output, for one typical individual who leaves the labor force for one year to attend college. By setting the return and cost expressions equal to each other and assuming that the direct cost of one year of college is equal to the foregone￾production cost,19 we can solve for the implicit rate of return to higher educa￾tion in the jth province, rs j . We obtain a3 2b(Ej/Nj) / 1 Å (1 / rs j) N , (5) where N is the number of years required to graduate from college. Equation (5) illustrates that the payoff to investment in human capital is greater the greater is the elasticity of TFP with respect to adding new university graduates, the smaller is the elasticity of production with respect to labor and the ratio of the current flow of new college graduates relative to the labor force, and 18 We assume no depreciation and infinite lifetimes. 19 This is only a simplification. The nature of the solution is basically unaffected by this assumption. AID JCE 1462 / 6w10$$$123 09-30-97 14:16:24 cea