Human capital externalities and rural–urban migration: Evidence from rural China

China Economic Review 19 (2008) 521–535

Contents lists available at ScienceDirect

China Economic Review

Human capital externalities and rural–urban migration: Evidence from rural China☆ Zhiqiang LIU Department of Economics, State University of New York at Buffalo, Buffalo, NY 14260, United States

a r t i c l e

i n f o

Article history: Accepted 10 April 2008 Keywords: Rural–urban migration Human capital externalities Rural development

JEL classification: J24 O15 O18

a b s t r a c t This study examines the determinants of rural–urban migration paying special attention to the role of human capital externalities in the rural sector. Using data from a well-known household survey in China, we find that in rural areas human capital externalities have a discouraging effect on rural–urban migration—everything else being the same, a rural resident from a county rich in human capital is less likely to migrate to the city than his counterpart from another county poor in human capital endowment. We also find some evidence that human capital exerts positive external effects on the likelihood for a rural resident to choose off-farm employment and on labor income in the rural sector. These results are robust to alternative model specifications and estimation methods. One important policy implication from this study is that expanding education opportunities in rural areas can help curtail rural–urban migration and therefore alleviate urban unemployment pressure. © 2008 Elsevier Inc. All rights reserved.

1. Introduction Rural–urban migration is an essential part of economic development, through which human resources move from the agricultural sector where the marginal product of labor is low or zero to the urban industrial sector where the marginal product of labor is high. While rural–urban migration helps improve the efficiency of sectoral allocation of resources, it also exacerbates, to a large extent, the widespread problem of urban unemployment in developing countries. This is especially true in China, as reforms of state-owned enterprises have resulted in massive layoffs of redundant workers in the city. According to the basic Harris–Todaro model,1 urban job creation programs are ineffective in solving the unemployment problem because they raise the expected payoffs to rural migrants and, as a result, lead to a higher, rather than lower, level of urban unemployment.2 One solution is to control rural–urban migration through administrative means. A policy tool that China relied upon is the hukou or household registration system coupled with a ration system on staple food in urban areas. Although that policy has successfully warded off unplanned migrations, it has been criticized as an unfair development policy that promotes industrial growth at the expense of the agricultural sector and urban development at the expense of the rural sector. It also is contrary to the principle of economic efficiency that requires the free movement of human resources across different regions and sectors. From a rural development

☆ An earlier version of the paper was presented at the Chinese Economists Society's International Conference on “Governing Rapid Growth in China: Efficiency, Equity and Institutions” held in Shanghai, China, July 2–4, 2006. I thank Belton Fleisher and two anonymous referees for comments and suggestions. The usual disclaimer applies. E-mail address: [email protected]. 1 See, Todaro (1969) and Harris and Todaro (1970). Lucas (2004) offers an alternative explanation on why rural–urban migration persists despite of poor employment prospects in the urban sector. He emphasizes the role of cities as places in which new migrants can accumulate the skills required by modern production technologies, which raise the life-time earnings of migrants. 2 Rural migrants are also blamed for being a strain on urban infrastructures, such as transportation, healthcare and education systems, and for rising crimes and other social problems in urban areas.

1043-951X/$ – see front matter © 2008 Elsevier Inc. All rights reserved. doi:10.1016/j.chieco.2008.04.001

522

Z. Liu / China Economic Review 19 (2008) 521–535

perspective, rural–urban migration may be harmful because it depletes the rural sector of valuable human capital as the young and better-educated rural residents are among the first to migrate to the city.3 There is a need to identify rural–urban migration policies that can strike a balance among three objectives: efficiency, urban unemployment, and rural development. A substantial amount of research has been devoted to the understanding of the determinants of the migration decision.4 Lall, Selod, and Shalizi (2006) contains a comprehensive survey of the empirical literature, which finds that rural–urban income gaps, education, gender, family composition are among the factors that affect the migration decision in the way predicted by various theories of migration. Numerous studies examining the Chinese experience also offer corroborating evidence. Notable studies include, among others, Zhao (1999a,b), Hare (1999), De Brauw, Huang, Rozelle, Zhang, and Zhang (2002), and Giles and Mu (2005). The present study makes a contribution to the vast literature on rural–urban migration. Our approach differs from the existing literature by focusing on human capital externalities in the rural sector. This allows us to gain new insights into the general role of human capital in influencing the migration decision and to derive policy recommendations that help curtail rural–urban migration and hence alleviate urban unemployment pressure. We argue that human capital at the aggregate level can act as an inhibitor to migration. The economic intuition is straightforward. The migration decision of an individual is made by weighing the expected income or utility of migration against that of no migration. Migration occurs when the net gain is positive, and any factor that reduces this net gain would have a dissuading effect on migration on the margin. Our argument is that if human capital can raise, through its external effect, the payoffs from rural activities, an increase in the overall level of educational attainment of the rural population would make staying in the rural sector a more attractive choice (at least for some potential migrants) relative to migrating to the city. This would be particularly true if there are plenty of local off-farm employment opportunities, in which human capital externalities are more important than in on-farm productions. Thus, higher local human capital endowment would make migration a less attractive option relative to staying in the rural sector, and on-farm activities less attractive than local off-farm employment. The main finding of the paper, based on an econometric analysis of a well-known household survey in China, is that in rural areas human capital externalities have a discouraging effect on rural–urban migration. That is, everything else being the same, a rural resident from a county rich in human capital is less likely to migrate to the city than his counterpart from another county poor in human capital endowment. This result is obtained from both binomial and multinomial logit regression analyses. While in the former rural residents are assumed to choose between migrating to the city and staying in the rural sector, in the multinomial logit model they are faced with three employment choices: migrate to and work in the city, work in the rural off-farm sector, and work in the rural on-farm sector. The multinomial logit regressions also suggest that human capital exerts a positive external effect on the likelihood for a rural resident to choose local off-farm employment. Corroborating with this result, we find a positive relationship between labor income in the rural sector and local human capital endowment. All of these results are robust to alternative model specifications and estimation methods. One important policy implication of our findings is that expanding education opportunities in rural areas can be a viable strategy to curtail rural–urban migration and hence help alleviate urban unemployment pressure—quite contrary to the inference drawn by some previous studies that improving education of the rural population would hasten rural–urban migration. This policy recommendation accords well with China's official policy as reflected in the motto “leaving the farmland without leaving the village.” The rest of the paper is organized as follows. We briefly discuss the data in Section 2 and econometric methodology in Section 3. Section 4 presents the estimation results. We offer some concluding remarks in the last section. 2. Data The data used in this study come from the Chinese Household Income Project 1995, also known as CHIP95. The survey contains two distinct samples of the urban and rural populations of China selected from substantially larger samples drawn by the National Bureau of Statistics of China and cover cities and towns of various sizes from different regions. In the rural sample, which we use in this study, eighteen provinces and one municipality were chosen to represent the whole country. These are Liaoning, Hebei, Jilin, and Shanxi in the north, Shandong, Jiangsu, Zhejiang, and Guangdong as eastern coastal provinces, Anhui, Henan, Hubei, and Hunan from the interior, Gansu, Shananxi, Sichuan, Guizhou, Jiangxi, and Yunan in the west, and Beijing as a representative of then three large province-level municipalities. The rural sample consists of two parts. The first contains information on the respondent's age, gender, education, employment status, ownership sector of employment, economic sector of employment, and annual labor income. It also contains information on whether an individual went to work or look for work in the city. The second part of the rural sample contains information on household characteristics, such as landholdings, household composition, incomes, expenditures, and assets. It also contains some information about the village that the household resides, such as whether the village is located in the suburb of a large- or medium-size city, the type of terrain and availability of telephone services. In this study, we focus on rural individuals who were between 16 and 60 years of age and reported complete information on schooling, age, gender, and employment. Full-time students, pre-school children, and individuals who were retired or disabled are excluded from the sample. This results in a sample of 21,451 individuals.

3 It is conceivable that rural–urban migration may benefit rural development if migration becomes a source of remittances to rural areas, especially if the remittance is used in education and productive investments. 4 There are several comprehensive reviews of the theoretical literature. See, e.g., Hare (1999) and Lall et al. (2006). To avoid unnecessary repetition, we do not attempt to provide one of our own in this paper.

Z. Liu / China Economic Review 19 (2008) 521–535

523

Table 1 presents definitions and summary statistics of the variables used. We classify the sample individuals into three groups: migrants, local off-farm workers, and local on-farm workers. Migrants are identified as people who left their households for at least one month to work or to look for work in the city. They consist of about 6.5% of our sample individuals. Local off-farm workers, accounting for about 18% of the sample, are those who engaged primarily in local off-farm activities. Local on-farm workers are those who engaged mainly in on-farm activities, representing over 75% of the sample individuals. Many of non-migrant workers earned labor income which includes regular wage, bonuses, subsidies, and non-regular income (such as contract income and sales commission) from work unit, income from other sources, such as the market value of income in kind received from the work unit and income and subsidies received for serving as village cadre. Pension and unemployment benefits are excluded. Labor income varies widely across individuals partly because of large variations in the number of days allocated to non-household business related activities. Those who received labor income devoted an average of 35.5 days during the sample year to working outside their households, with a standard deviation of 84 days. Measuring local human capital endowment is of central importance to our empirical investigation and, therefore, warrants a brief discussion. We measure the level of local human capital for each county in two ways. The first, denoted by Cedyr, is the average schooling in years of the sample individuals in a county. This measure is similar to the ones adopted by Rauch (1993) and Acemoglu and Angrist (2000) who estimate human capital externalities in U.S. cities and states, respectively, and Liu (2007) who estimates the external returns to education in Chinese cities. The second measure, denoted by Colsh, is the share of college graduates among the sample individuals in a county. This measure is similar to the one adopted by Moretti (2004) and Liu (2007). Cedyr and Colsh may capture different mechanisms through which human capital generates external effects. As Cedyr is the simple average of educational attainment of individuals in a local area, it indicates the general availability of a productive labor force, which is more important for off-farm than for on-farm activities. By contrast, Colsh indicates the availability and relative size of a well-educated, elite group. College-educated workers can help raise others' productivity due to the existence of a complementary

Table 1 Definition of variables and summary statistics Definition Key dependent variables Migrant Local off-farm worker Local farm worker Local labor income

=1 if left the household for at least a month to work or look for work in cities, =0 for all others =1 if engaged primarily in local non-farming activities, =0 for all others =1 if engaged primarily in farming activities, =0 for all others Total annual labor income, excluding pension and unemployment benefits (yuan)

Individual characteristics College Vocation school High school Middle school Primary school Age Male Married Workdays

=1 for college graduates, =0 for all others =1 for vocational school graduates, = 0 for all others =1 for high school graduates, =0 for all others =1 for middle school graduates, =0 for all others =1 for primary school graduates, =0 for all others Age in years =1 for males, = 0 for females =1 for married people, =0 for all others Number of days allocated to non-household business related activities

Family characteristics Family size Family labor force Pre-school children Disable persons Landholdings Family wealth

Number of household members Number of household members in the labor force Number of pre-school children in the household Number of disabled persons in the household Land under the household control (mu) Total value of all financial asset at the end of 1995 (10,000 yuan)

Community characteristics Cedyr Colsh Colsh90 Suburb of a city Flatland Telephone services Share of household farming Share of TVE sector Share of other ownership sector Share of agriculture sector Share of industry sector Share of service sector Share of government sector Share of other economic sector

Average level of education in years in a county based on the sample College graduates as a share of county population based on the sample (%) College graduates as a share of county population based on the 1990 census (%) =1 if the village is a suburb of middle or large sized city, =0 otherwise Type of terrain: = 1 for flat land, =0 for hilly or mountainous land =1 if telephone service is available, =0 otherwise Share of the county labor force working on household farm (%) Share of the county labor force working in town and village enterprises (%) Share of the county labor force working in other ownership sector (%) Share of the county labor force working primarily in the agricultural sector Share of the county labor force working primarily in the industrial sector Share of the county labor force working primarily in the service sector Share of the county labor force working primarily in the government sector Share of the county labor force working primarily in other economic sectors

Mean (standard deviation) 0.065 (0.246) 0.178 (0.383) 0.757 (0.429) 4007 (7572)

0.00531 (0.0727) 0.0116 (0.107) 0.0863 (0.281) 0.417 (0.493) 0.311 (0.463) 35.73 (11.843) 0.502 (0.500) 0.767 (0.423) 35.5 (83.712)

4.35 (1.374) 3.17 (1.236) 0.28 (0.550) 0.063 (0.271) 7.99 (6.552) 4885 (9136)

6.01 (1.036) 0.531 (0.875) 0.140 (0.225) 0.0422 (0.201) 0.469 (0.499) 0.588 (0.492) 79.78 (16.843) 6.68 (9.621) 13.54 (10.426) 78.54 (15.597) 12.41 (11.390) 2.85 (2.819) 2.25 (2.097) 3.95 (2.992)

524

Z. Liu / China Economic Review 19 (2008) 521–535

relationship between skilled and unskilled workers or pure spillover effects. They may also raise the productivity of others by playing the roles of local business leaders or entrepreneurs or government officials. 3. Econometric methodology The econometric models we adopt in this study are similar to those widely used in the literature, which are based on the assumption that each individual makes the migration decision to maximize expected income or utility. In this framework, individual, family, and community characteristics affect the migration decision through their differential influences on the expected incomes or utilities from migrating to the city and staying in the rural sector. Instead of trying to isolate these influences, a reduced-form approach is commonly employed to estimate the net effects of these variables on observed migration outcomes. We extend this conventional model by including a measure of local human capital endowment to isolate the impact of human capital externalities on the migration decision of rural residents. The net utility gain in present value terms from migration is modeled as an unobservable variable, y⁎, which is determined by the following function: y4 ¼ b VX þ u;

ð1Þ

where X is a vector of individual, household, and community characteristics including local human capital endowment, and u has a standard logistic distribution with mean zero and variance one. However, the net gain is not observable. What is observable is a dummy variable, y, indicating whether an individual has migrated or not; that is, y = 1 if y⁎ N 0, y = 0 if otherwise. The logit model for the migration decision is: Prob ð y ¼ 1Þ ¼

eb VX 1 þ eb VX :

ð2Þ

To explore the role of human capital externalities in influencing employment choice of rural residents, we also estimate the following multinomial logit model: ebjVX Prob ðz ¼ jÞ ¼ P2 ; bkV X k¼0 e

ð3Þ

where z = 2 for migrant workers, z = 1 for local off-farm workers, and z = 0 for local on-farm workers. In Eqs. (2) and (3), the coefficient on local human capital endowment identifies the impact of human capital externalities on individual decisions regarding migration and employment choice, respectively. However, as Acemoglu and Angrist (2000) argued, the interpretation of this coefficient estimate derived from the standard regression method may be complicated if local human capital endowment is measured by the within-group average of individual educational attainment, which is also included as a separate independent variable in the regression. In this case, it is possible for the coefficient on average education to be nonzero even though the external effect of human capital is in fact absent. To avoid this problem, we measure individual educational attainment by a set of dummy variables indicating the highest grade level completed, instead of years of formal schooling. A major econometric challenge of the study is to obtain consistent estimates since county average education and unobserved county-specific characteristics are likely to be correlated. We break this correlation in three ways. First, we introduce proxy variables for unobserved, county-specific factors into our regressions and assume that the new disturbance is not correlated with county average education. Second, we use a lagged measure for county-level education to identify the external effect of human capital. The lagged variable, which is the share of college graduates in the county population five years prior to our sample period, is predetermined and, therefore, is uncorrelated with the error term. Third, we use the method of instrumental variables to obtain consistent estimates. The instrumental variables approach also offers a solution to biases in estimates that stem from the possibility that county average education is endogenous or measured with error. The key to this approach is to identify instruments that are correlated with county average education and orthogonal to unobserved county-specific characteristics. We propose two instruments for average county education. One is human capital endowment at the province level. The other is the share of county population below the age of 25, whose education was mostly likely affected by the Compulsory Education Law that China promulgated in 1986.5 4. Results 4.1. The effect of human capital externalities on the migration decision: binomial logit regressions We begin with the estimation of the migration decision using the logit model. Columns (1) and (2) of Table 2 present the estimates from two alternative specifications of the logit model that have been commonly used in the literature. In column (1) the migration decision depends on individual and family characteristics, while in column (2) community characteristics are also allowed to affect the decision. These estimates are very much consistent with the findings reported in previous studies. 5

More discussion about the instruments will be given in the next section.

Z. Liu / China Economic Review 19 (2008) 521–535

525

Table 2 Rural–urban migration decision: binomial logit regressions (1)

(2)

(3)

(4)

(5)

(6)

(7)

(8) a

− 0.671 (−0.909) 1.0506⁎⁎⁎ (3.740) 0.827⁎⁎⁎ (5.003) 0.682⁎⁎⁎ (4.744) 0.521⁎⁎⁎ (3.559) 0.0936⁎⁎⁎ (3.406) − 0.00218⁎⁎⁎ (−5.565) 0.969⁎⁎⁎ (14.690) − 0.856⁎⁎⁎ (−8.502)

−0.572 (−0.775) 1.0069⁎⁎⁎ (3.517) 0.796⁎⁎⁎ (4.742) 0.673⁎⁎⁎ (4.612) 0.478⁎⁎⁎ (3.229) 0.0882⁎⁎⁎ (3.159) −0.00213⁎⁎⁎ (−5.347) 0.982⁎⁎⁎ (14.625) −0.860⁎⁎⁎ (−8.343)

− 0.574 (− 0.777) 1.0237⁎⁎⁎ (3.573) 0.829⁎⁎⁎ (4.906) 0.701⁎⁎⁎ (4.775) 0.495⁎⁎⁎ (3.339) 0.0887⁎⁎⁎ (3.175) − 0.00213⁎⁎⁎ (− 5.353) 0.978⁎⁎⁎ (14.545) − 0.857⁎⁎⁎ (− 8.303)

−0.469 (−0.633) 0.994⁎⁎⁎ (3.467) 0.777⁎⁎⁎ (4.623) 0.650⁎⁎⁎ (4.442) 0.460⁎⁎⁎ (3.098) 0.0873⁎⁎⁎ (3.125 −0.00212⁎⁎⁎ (−5.318) 0.986⁎⁎⁎ (14.666 −0.861⁎⁎⁎ (−8.348)

–

–

1.0287⁎⁎⁎ (3.590) 0.835⁎⁎⁎ (4.944) 0.707⁎⁎⁎ (4.816) 0.499⁎⁎⁎ (3.367) 0.0887⁎⁎⁎ (3.175) −0.00212⁎⁎⁎ (−5.341) 0.975⁎⁎⁎ (14.493) −0.865⁎⁎⁎ (−8.369)

0.998⁎⁎⁎ (3.480) 0.782⁎⁎⁎ (4.650) 0.655⁎⁎⁎ (4.474) 0.463⁎⁎⁎ (3.121) 0.0875⁎⁎⁎ (3.129) −0.00211⁎⁎⁎ (−5.309) 0.984⁎⁎⁎ (14.620) −0.869⁎⁎⁎ (−8.416)

− 0.447 (− 0.604)) 1.0121⁎⁎⁎ (3.528) 0.815⁎⁎⁎ (4.823) 0.681⁎⁎⁎ (4.631) 0.477⁎⁎⁎ (3.211) 0.0877⁎⁎⁎ (3.137) − 0.00212⁎⁎⁎ (− 5.320) 0.982⁎⁎⁎ (14.581) − 0.857⁎⁎⁎ (− 8.296)

−1.218 (−1.190) 0.560 (1.593) 0.731⁎⁎⁎ (4.197) 0.650⁎⁎⁎ (4.361) 0.440⁎⁎⁎ (2.906) 0.0907⁎⁎⁎ (3.007) −0.00215⁎⁎⁎ (−4.999) 0.976⁎⁎⁎ (13.338) −0.831⁎⁎⁎ (−7.454)

0.162⁎⁎⁎ (4.977) 0.0826⁎⁎ (2.092) 0.0185 (0.303) − 0.169 (−1.434) − 0.00938 (−1.587) − 0.191⁎⁎⁎ (−4.159)

0.141⁎⁎⁎ (4.234) 0.0653⁎ (1.825) −0.0105 (−0.168) −0.0654 (−0.545) −0.0019 (−0.316) −0.0704 (−1.487)

0.136⁎⁎⁎ (4.068) 0.0704⁎⁎ (1.958) − 0.0125 (− 0.199) − 0.0631 (− 0.525) − 0.00284 (− 0.470) − 0.0706 (− 1.492)

0.143⁎⁎⁎ (4.298) 0.0614⁎ (1.713) −0.00969 (−0.155) −0.0657 (−0.548) −0.000784 (−0.129) −0.0710 (−1.497)

0.137⁎⁎⁎ (4.103) 0.0694⁎ (1.931) −0.0131 (−0.209) −0.0634 (−0.527) −0.00271 (−0.463) −0.0701 (−1.480)

0.144⁎⁎⁎ (4.336) 0.0603⁎ (1.682) −0.0104 (−0.166) −0.0660 (−0.549) −0.000720 (−0.118) −0.0704 (−1.485)

0.137⁎⁎⁎ (4.109) 0.0672⁎ (1.867) − 0.0121 (− 0.194) − 0.0628 (− 0.522) − 0.00181 (− 0.297) − 0.0710 (− 1.500)

0.150⁎⁎⁎ (4.298) 0.0543 (1.431) −0.0484 (−0.738) −0.0722 (−0.572) −0.00647 (−1.007) −0.0440 (−0.816)

Community characteristics Cedyr –

– –

Colsh90

–

–

–

–

−9.512⁎⁎ (−1.995) –

− 0.108⁎⁎ (− 2.159) − 11.693⁎⁎ (− 2.426) –

–

–

−0.0855⁎ (−1.776) –

–

Colsh

− 0.0811⁎ (− 1.686) –

Suburb of a city

–

Flatland

–

Telephone services

–

Share of household farming Share of TVE sector

–

Share of agricultural sector Share of industry sector Share of service sector Share of government sector Province dummy R-sq Sample size

–

0.177 (1.069) −0.442⁎⁎⁎ (−5.554) −0.202⁎⁎⁎ (−2.753) 4.241⁎⁎⁎ (2.652) −8.648⁎⁎⁎ (−9.182) −4.140⁎⁎ (−2.333) 8.189⁎⁎⁎ (5.707) 2.832 (1.251) −4.935⁎ (−1.612) Yes 0.222 21,451

0.178 (1.073) − 0.447⁎⁎⁎ (− 5.588) − 0.193⁎⁎⁎ (− 2.622) 4.908⁎⁎⁎ (2.952) − 8.427⁎⁎⁎ (− 8.825) − 4.589⁎⁎ (− 2.559) 8.651⁎⁎⁎ (5.937) 2.688 (1.189) − 1.401

0.181 (1.087) −0.456⁎⁎⁎ (−5.695) −0.196⁎⁎⁎ (−2.658) 4.918⁎⁎⁎ (2.956) −8.404⁎⁎⁎ (−8.791) −4.0278 (−1.308) 8.703⁎⁎⁎ (5.967) 2.830 (1.251) −4.0278 (−1.308) Yes 0.222 21,337

0.150 (0.904) −0.445⁎⁎⁎ (−5.586) −0.212⁎⁎⁎ (−2.885) 3.775⁎⁎ (2.329) −8.572⁎⁎⁎ (−9.093) −3.965⁎⁎ (−2.241) 7.773⁎⁎⁎ (5.355) 2.539 (1.123) −4.818 (−1.577) Yes 0.222 21,337

0.143 (0.859) − 0.441⁎⁎⁎ (− 5.513) − 0.198⁎⁎⁎ (− 2.689) 4.586⁎⁎⁎ (2.772) − 8.282⁎⁎⁎ (− 8.659) − 4.625⁎⁎⁎ (− 2.591) 8.244⁎⁎⁎ (5.635) 2.088 (0.926) − 4.317 (− 1.411) Yes 0.223 21,451

Individual characteristics College Vocation school High school Middle school Primary school Age Age-squared Male Married

Family characteristics Family size Family labor force Pre-school children Disable persons Landholdings (mu) Family wealth (10,000 yuan)

–

– – – Yes 0.192 21,451

Yes 0.222 21,451

−9.648⁎⁎ (−2.028) – 0.148 (0.888) −0.436⁎⁎⁎ (−5.482) −0.209⁎⁎⁎ (−2.842) 3.797⁎⁎⁎ (2.346) −8.580⁎⁎⁎ (−9.112) −4.043⁎⁎ (−2.287) 7.736⁎⁎⁎ (5.336) 2.384 (1.055) −5.0719⁎ (−1.662) Yes 0.222 21,451

– −180.542⁎⁎⁎ (−4.320) 1.0756⁎⁎⁎ (3.335) −0.309⁎⁎⁎ (−3.378) −0.227⁎⁎⁎ (−2.882) 1.894 (0.975) −7.904⁎⁎⁎ (−7.671) −2.332 (−1.165) 7.368⁎⁎⁎ (4.801) 0.907 (0.371) −10.022⁎⁎⁎ (−2.729) Yes 0.224 18,922

Notes: Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level. a Colsh90, college graduates as a share of county population in 1990, is used in place of Colsh.

The likelihood of migration is positively associated with educational attainment up to the level of vocational school. College graduates are less likely than those with below primary school education to migrate. At the first glance, the negative coefficient on College seems surprising. But this is expected. College graduates are automatically granted urban hukou (see Liu,

526

Z. Liu / China Economic Review 19 (2008) 521–535

2005) and, therefore, have the option of settling down permanently in the city. They are in the rural sector either by individual choice or by official job assignment. Young workers are more likely than older workers to seek work in the city. Males are significantly more likely than females to migrate, while married people have a lower propensity to migrate than singles. Whereas family size and household labor force raise significantly the likelihood for one to migrate to the city, the number of pre-school children, the number of disabled family members, family landholdings, and family wealth tend to reduce the propensity of migration. These results are consistent with those reported in numerous studies, such as Zhao (1999a,b) and De Brauw et al. (2002). Some of the community characteristics, which were not always considered in previous studies, are shown to play important roles in influencing the migration decision. People live in the suburb of middle- or large-size cities are more likely to migrate, albeit the estimated coefficient is not statistically significant. Consistent with the findings reported in Zhao (1999a), individuals reside in plains and villages having telephone services are less likely to migrate. The next two independent variables–the county employment shares of the household farming and TVE (town and village enterprise) sectors–are introduced to capture the availability of surplus labor and local off-farm employment opportunities, respectively. As the marginal product of labor is typically below the average product or possibly equal to zero in household farming, one or several members from a household can switch to off-farm jobs or look for work in the city without affecting substantially the output of household farming. Therefore, the proportion of people engaged in household farming indicates the availability of surplus labors. The estimated coefficient on this variable is positive and statistically significant, indicating indeed this is the case. Since TVE employment is an alternative to migration, a larger TVE sector reduces the net gain from migration and therefore works as a disincentive to migration. The estimate shows that the availability of employment opportunities in the TVE sector reduces the likelihood of migration. The last four independent variables are intended to capture the structural differences across rural counties. The proportions of people employed in the agriculture and government sectors are negatively associated with the likelihood of migration, while the employment shares of the industrial and service sectors are positively associated with rural–urban migration. However, the estimated effects of these variables are not all as expected, partly because of the presence of high correlation between the share of people in the household farming sector and the share of people in the agricultural sector, and among employment shares of the TVE, industry, and service sectors.6 The results of columns (1) and (2) also indicate that, while these county-specific factors affect the migration decision, they have virtually no impact on the estimates associated with the individual characteristics and most of the family characteristics.7 The county-specific characteristics also help mitigate potential omitted variable problems in regressions reported in columns (3) through (8), where we introduce county average education or college graduates as a share of the sample to investigate whether local human capital endowment exerts an external effect on the migration decision. Without controlling for the county-specific factors, the estimated effect of human capital externalities may be biased if local human capital endowment is correlated with factors such as the structure of the local economy, the availability of schools and wage jobs in local areas. For instance, rural counties in close proximity to cities may benefit from the development and expansion of the latter in terms of rural transformation and general education. By the same token, counties that are relatively more developed are likely to invest more in education and, as a result, would have a larger human capital endowment. It should be noted that introducing local human capital endowment as an additional explanatory variable into the logit regressions of columns (3) and (4) has little impact on the estimated coefficients associate with other variables. Note, particularly, that the educational attainment of an individual remains positively related to the propensity to migrate. By contrast, the average educational attainment (as well as the share of college graduates) at the county level has a negative and statistically significant effect on the likelihood of migration. This suggests that in the rural sector human capital externalities work to restrain rural–urban migration. Since the estimated coefficients on College remain negative, there is the question whether the negative external effect of human capital is an artifact resulting from the inclusion of college graduates in the sample. To address this concern, we rerun the logit regressions of columns (3) and (4) using the subsample that excludes college graduates. As the estimates reported in columns (5) and (6) indicate, the estimated effect of human capital externalities remains negative and statistically significant at the 5% level. Excluding college graduates from the sample also allows us to assess the extent to which the estimated coefficient on Colsh is subject to the type of bias discussed in Acemoglu and Angrist (2000), as in column (4) Colsh is the within-county average of another independent variable, namely College. Note that College is no longer an independent variable in the regression of column (6), and the estimated coefficient on Colsh is very comparable to that reported in column (4). This indicates that the estimated effect of human capital externalities on the migration decision is most likely free of the bias that arises when one explanatory variable is the with-group average of another independent variable. As discussed in the previous section, Cedyr and Colsh may capture different aspects of human capital endowment. To examine whether each of these two measures has an independent effect on the migration decision conditional on the other, we introduce them together in column (7). In this specification, not only are the estimated coefficients on Cedyr and Colsh negative and statistically significant at the 5% level but also larger than their counterparts reported in columns (3) through (6), where only one of

6 The six county-specific employment shares are constructed based on our sample individuals including migrants and non-migrants. We also experimented with an alternative definition of these share variables that excludes migrants and found that our main findings are not affected in any significant way, but the estimated coefficients for the employment shares of the household farming, agriculture, and service sectors switched signs. 7 The likelihood ratio test rejects the null hypothesis that the community characteristics have no effect on the migration decision at the 1% level of significance. The likelihood ratio statistic is 303.16, having a Chi-square distribution with 9 degrees of freedom.

Z. Liu / China Economic Review 19 (2008) 521–535

527

these two variables is included in each regression. We also run a regression that includes the interaction term between Cedyr and Colsh as an additional explanatory variable. In this specification (not reported), the coefficient on the interaction term is negative albeit insignificant, indicating that the negative external effect of human capital as measured by Cedyr (i.e., county average educational attainment in years) is reinforced by Colsh (i.e., the share of college graduates) and vice versa. Recall, one reason for us to include the community-specific characteristics is to alleviate potential omitted variable bias in the estimates for the external effect of human capital. However, this strategy may not be very effective if there are other county-specific and unobserved factors that are correlated with the county-specific human capital endowment, but are not included in the regressions. Furthermore, the negative relationship between the average education at the county level and the likelihood of migration may be due to reverse causality—increased opportunities to migrate reduce the likelihood that local residents will complete high school as argued by De Brauw and Giles (2008).8 To address these concerns, we replace Cedyr and Colsh by a measure of county-level human capital in a past period. Under the assumption that the error term is not serially correlated, the external effect of human capital can be consistently estimated using this lagged variable. We obtain from the 1990 census college graduates as a share of county population (denoted as Colsh90) and use it as the lagged variable for both Cedyr and Colsh. The estimates are reported in column (8). In this specification, the estimated coefficient on Colsh90 is negative and statistically significant at the 1% level, similar to those on Cedyr and Colsh in columns (3) through (7).9 4.2. The effect of human capital externalities on migration, local off-farm and on- farm employment: multinomial logit regressions In the preceding empirical analyses, we assume that rural residents have only two choices: migrating to the city or staying in the rural sector. In reality, people who decide to stay in the rural sector can choose between on-farm and off-farm employment. If local human capital endowment has a restraining effect on migration, does it generate a positive external effect on off-farm employment and hence makes the latter a more attractive choice than on-farm employment or migrating to the city? To answer this question, we estimate a series of multinomial logit regressions to isolate the effect of human capital externalities on the employment choice of rural residents. Table 3 presents the results of multinomial logit regressions with specifications similar to the logit regressions discussed in the previous subsection. Since the coefficient estimates from a multinomial logit regression do not bear any relationship to either the sign or the magnitude of the marginal-effect estimates for the corresponding independent variables, we also report the marginal-effect estimates (in square brackets) in addition to the coefficient estimates and asymptotic t-statistics. The results from the migration decision equation are very similar to those obtained from the binomial logit regressions. There is no new finding to report. However, it is worth reiterating that the estimated external effect of human capital on the migration decision is negative and statistically significant at the 10% level or higher across all four specifications. Now turn to the estimates from the rural off-farm employment equation. Similar to its effect on the migration decision, more education makes it more likely for an individual to take local off-farm jobs as oppose to on-farm jobs. However, people with primary school education are less likely to seek off-farm employment than those did not complete primary school. Older people and males are more likely to opt for off-farm jobs, while married individuals are less likely to do so. Five of the six family characteristics are shown to be significant predictors for whether one chooses to engage in rural off-farm activities. Whereas family size and wealth enhance one's odds of working off the farm, household labor force and the number of pre-school children lower one's chance of holding an off-farm job. Everything else being equal, people from families with larger landholdings are less likely to find themselves employed in the off-farm sector. Some of the county-specific characteristics are shown to have significant effects on the decision to choose off-farm employment in the rural sector. The availability of telephone services and the relative size of the TVE sector are, as expected, positively associated with the probability for one to work in the off-farm sector. People from counties that have a larger share of their population working on household farms or being employed in the agriculture sector are less likely to obtain off-farm employment. The effects of employment shares in the industrial and service sectors are somewhat puzzling—people from counties that have a larger share of their population employed in these sectors are less likely to find themselves working in the off-farm sector. The partial effect of human capital endowment at the county level on individual decisions to work off the farm seems to depend on the specific measure used for the former. When we use county average education to proxy human capital endowment in the specification of model (1), human capital is found to reduce the likelihood for one to switch from the on-farm to off-farm sector. But the estimate is not statistically significant at even the 10% level. When human capital is measured by college graduates as a share of county population in the specification of model (2), its coefficient estimate is positive and statistically significant at the 5% level. This indicates that people from counties with a larger share of their population being college graduates are more likely to opt for off-farm employment. This result holds in model (3), where both Cedyr and Colsh are included as explanatory variables, and in model (4), where Colsh90 is used. 4.3. The effect of human capital externalities on rural off-farm labor income and off-farm employment opportunity The analyses above suggest that there is a negative association between local human capital endowment and the likelihood of migration, and a generally positive relationship between local human capital endowment and the likelihood of moving from the 8

Another way of dealing with the reverse causality issue is to use the instrumental variables method. See our IV estimates in Section 4.5. The estimated coefficient on Colsh90 is substantially larger than that on Colsh. This difference is mainly due to the fact that Colsh90 is college graduates as a share of the county population whereas Colsh is college graduates as a share of the sample which consists of individuals between 16 and 60 years of age. 9

528

Z. Liu / China Economic Review 19 (2008) 521–535

Table 3 Migration, local off-farm, and on-farm employment: multinomial logit regressions (1)

(2)

(3)

(4)

Migration

Local off-farm

Migration

Local off-farm

Migration

Local off-farm

Migration

Local off-farm

−0.152 (−0.204) [−0.0106] 1.852⁎⁎⁎ (6.217) [0.0436] 1.0475⁎⁎⁎ (6.150) [0.0255] 0.738⁎⁎⁎ (5.006) [0.0190] 0.423⁎⁎⁎ (2.830) [0.0126] 0.0500⁎ (1.771) [0.00196] −0.00166⁎⁎⁎ (−4.143) [−0.0000533] 1.136⁎⁎⁎ (16.630) [0.0291] −0.927⁎⁎⁎ (−8.838) [−0.0245]

1.648⁎⁎⁎ (6.880) [0.192] 2.0818⁎⁎⁎ (11.900) [0.235] 0.949⁎⁎⁎ (10.013) [0.106] 0.395⁎⁎⁎ (5.082) [0.0430] −0.225⁎⁎⁎ (−2.900) [−0.0277] −0.149⁎⁎⁎ (9.061) [−0.0175] 0.00184⁎⁎⁎ (8.886) [0.000220] 0.655⁎⁎⁎ (14.500) [0.0717] −0.336⁎⁎⁎ (−4.119) [−0.0354]

− 0.0433 (− 0.058) [−0.00733] 1.818⁎⁎⁎ (6.101) [0.0428] 0.990⁎⁎⁎ (5.845) [0.0240] 0.684⁎⁎⁎ (4.651) [0.0176] 0.385⁎⁎⁎ (2.581) [0.0116] 0.0487⁎ (1.723) [0.00193] − 0.00165⁎⁎⁎ (− 4.110) [−0.0000531] 1.144⁎⁎⁎ (16.755) [0.0295] − 0.932⁎⁎⁎ (− 8.886) [−0.0247]

4.588⁎⁎⁎ (6.607) [0.184] 2.0757⁎⁎⁎ (11.887) [0.234] 0.944⁎⁎⁎ (10.058) [0.106] 0.393⁎⁎⁎ (5.115) [0.0429] −0.226⁎⁎⁎ (−2.923) [− 0.0277] −0.149⁎⁎⁎ (−9.042) [− 0.0174] 0.00184⁎⁎⁎ (8.865) [0.000220] 0.656⁎⁎⁎ (14.539) [0.0716] −0.338⁎⁎⁎ (−4.146) [− 0.0356]

−0.0226 (−0.030) [−0.00675] 1.840⁎⁎⁎ (6.166) [0.0433] 1.0332⁎⁎⁎ (6.064) [0.0252] 0.719⁎⁎⁎ (4.866) [0.0185] 0.405⁎⁎⁎ (2.709) [0.0121] 0.0491⁎ (1.736) [0.00194] −0.00165⁎⁎⁎ (−4.111) [−0.0000531] 1.139⁎⁎⁎ (16.662) [0.0292] −0.927⁎⁎⁎ (−8.825) [−0.0245]

1.592⁎⁎⁎ (6.621) [0.185] 2.0797⁎⁎⁎ (11.896) [0.234] 0.950⁎⁎⁎ (10.015) [0.106] 0.398⁎⁎⁎ (5.124) [0.0434] − 0.222⁎⁎⁎ (− 2.868) [− 0.0274] − 0.148⁎⁎⁎ (− 9.036) − 0.0174] 0.00184⁎⁎⁎ (8.863) [0.000220] 0.654⁎⁎⁎ (14.490) [0.0715] − 0.338⁎⁎⁎ (− 4.141) [− 0.0356]

−0.778 (−0.758) [−0.0264] 1.350⁎⁎⁎ (3.733) [0.0280] 0.949⁎⁎⁎ (5.400) [0.0214] 0.685⁎⁎⁎ (4.573) [0.0165] 0.367⁎⁎⁎ (2.407) [0.0104] 0.0513⁎ (1.683) [0.00190] −0.00167⁎⁎⁎ (−3.859) [−0.0000507] 1.135⁎⁎⁎ (15.300) [0.0274] −0.893⁎⁎⁎ (−7.899) [−0.0224]

1.702⁎⁎⁎ (6.978) [0.195] 2.128⁎⁎⁎ (11.403) [0.236] 1.000⁎⁎⁎ (10.016) [0.110] 0.415⁎⁎⁎ (5.107) [0.0445] −0.211⁎⁎⁎ (−2.568) [−0.0251] −0.157⁎⁎⁎ (−8.922) [−0.0179] 0.00195⁎⁎⁎ (8.762) [0.000226] 0.678⁎⁎⁎ (13.900) [0.0726] −0.307⁎⁎⁎ (−3.521) [−0.0316]

0.151⁎⁎⁎ (4.452) [0.00390] 0.04777 (1.309) [0.00162] −0.0448 (−0.706) [−0.000631] −0.0869 (−0.715) [−0.00203] −0.0120⁎ (−1.937) [−0.000144] −0.0174 (−0.354) [−0.000902]

0.0786⁎⁎⁎ (3.335) [0.00855] −0.0757⁎⁎⁎ (−2.928) [−0.00897] −0.161⁎⁎⁎ (−3.51) [−0.0185] −0.102 (−1.262) [−0.0115] −0.0498⁎⁎⁎ (−9.467) [−0.00573] 0.109⁎⁎⁎ (4.456) [0.0127]

0.158⁎⁎⁎ (4.695) [0.00412] 0.0385 (1.057) [0.00136] − 0.0428 (− 0.674) [−0.000577] − 0.0913 (− 0.751) [−0.00216] − 0.00999⁎ (− 1.605) [−0.0000852] 0.0195 (− 0.395) [−0.000959]

0.0793⁎⁎⁎ (3.371) [0.00858] −0.0747⁎⁎⁎ (−2.892) [− 0.00882] −0.161⁎⁎⁎ (−3.536) [− 0.0185] −0.103 (−1.274) [− 0.0115] −0.0503⁎⁎⁎ (−9.574) [− 0.00580] 0.108⁎⁎⁎ (4.435) [0.0126]

0.152⁎⁎⁎ (4.485) [0.00393] 0.0451 (1.233) [0.00154] −0.0450 (−0.709) [−0.000638] −0.0874 (−0.717) [−0.00205] −0.0112⁎ (−1.786) [−0.000117] −0.0182 (−0.369) [−0.000924]

0.0784⁎⁎⁎ (3.327) [0.00852] − 0.741⁎⁎⁎ (− 2.862) [− 0.00877] − 0.161⁎⁎⁎ (− 3.534) [− 0.0185] − 0.102 (− 1.263) [− 0.0115] − 0.0504⁎⁎⁎ (− 9.572) [− 0.00580] 0.109⁎⁎⁎ (4.461) [0.0127]

0.168⁎⁎⁎ (4.731) [0.00412] 0.0301 (0.780) [0.00109] −0.0846 (−1.272) [−0.00159] −0.0909 (−0.712) [−0.00206] −0.0145⁎⁎ (−2.208) [−0.000225] −0.0133 (−0.237) [−0.000641]

0.0839⁎⁎⁎ (3.352) [0.00888] −0.0851⁎⁎⁎ (−3.056) [−0.00972] −0.181⁎⁎⁎ (−3.775) [−0.0201] −0.0940 (−1.115) [−0.0103] −0.0446⁎⁎⁎ (−8.217) [−0.00499] 0.0835⁎⁎⁎ (3.009) [0.00947]

−0.0246 (−0.762) [−0.00248] –

–

–

–

–

Colsh90

–

–

6.400⁎⁎ (1.916) [0.774] –

− 0.0146 (− 0.446) [− 0.00122] 6.0368⁎ (1.767) [0.742] –

–

− 8.334⁎ (− 1.730) [−0.258] –

−0.123⁎⁎⁎ (−2.407) [−0.00336] −10.821⁎⁎ (−2.209) [−0.325] –

–

Colsh

−0.0977⁎⁎ (−2.000) [−0.00263] –

Suburb of a city

0.116 (0.691) [0.00396] −0.431⁎⁎⁎ (−5.336) [−0.0119] −0.158⁎⁎ (−2.113) [−0.00527]

−0.188⁎ (−1.723) [−0.0223] −0.0313 (−0.617) [−0.00197] 0.225⁎⁎⁎ (4.093) [0.0267]

0.0889 (0.528) [0.00320] − 0.421⁎⁎⁎ (− 5.235) [−0.0116] − 0.173⁎⁎ (− 2.326) [−0.00575]

−0.185⁎ (−1.701) [− 0.0218] −0.0342 (−0.676) [− 0.00234] 0.236⁎⁎⁎ (4.260) [0.0280]

0.0822 (0.487) [0.00300] −0.424⁎⁎⁎ (−5.251) [−0.0117] −0.162⁎⁎ (−2.169) [−0.00543]

− 0.185⁎ (− 1.693) [− 0.0218] − 0.0349 (− 0.687) [− 0.00241] 0.236⁎⁎⁎ (4.271) [0.0281]

−165.919⁎⁎⁎ (−4.194) [−4.497] 1.0357⁎⁎⁎ (3.165) [0.291] −0.281⁎⁎⁎ (−3.047) [−0.00754] −0.201⁎⁎⁎ (−2.511) [−0.00597]

39.901⁎⁎⁎ (2.482) [5.0860] −0.553⁎⁎⁎ (−2.650) [−0.0660] 0.0445 (0.751) [0.00601] 0.199⁎⁎⁎ (3.406) [0.0232]

Individual characteristics College

Vocation school

High school

Middle school

Primary school

Age

Age-squared

Male

Married

Family characteristics Family size

Family labor force

Pre-school children

Disable persons

Landholdings (mu)

Family wealth (10,000 yuan)

Community characteristics Cedyr

Flatland

Telephone services

Z. Liu / China Economic Review 19 (2008) 521–535

529

Table 3 (continued) (1)

Community characteristics Share of household farming Share of TVE sector

Share of agricultural sector Share of industry sector Share of service sector Share of government sector Province dummy R-sq Sample size

(2)

(3)

(4)

Migration

Local off-farm

Migration

Local off-farm

Migration

Local off-farm

Migration

Local off-farm

3.897⁎⁎ (2.300) [0.116] −7.413⁎⁎⁎ (−7.557) [−0.212] −5.849⁎⁎⁎ (−3.218) [−0.139] 7.539⁎⁎⁎ (5.067) 0.225] 2.628 (1.153) [0.0803] −4.651 (−1.495) [−0.128]

−1.972⁎⁎ (−2.230) [−0.244] 1.406⁎⁎⁎ (2.771) [0.192] −6.326⁎⁎⁎ (−5.167) [−0.712] −3.920⁎⁎⁎ (−4.203) [−0.484] −1.861 (−1.221) [−0.006] −0.383 (−0.193) [−0.0266] Yes 0.232 21,451

2.709⁎ (1.633) [0.0839] −7.613⁎⁎⁎ (− 7.883) [−0.218] −5.192⁎⁎⁎ (− 2.893) [−0.121] 6.616⁎⁎⁎ (4.453) [0.200] 2.454 (1.077) [0.0752] −5.461⁎ (− 1.767) [−0.150]

−2.131⁎⁎⁎ (−2.486) [− 0.258] 1.358⁎⁎⁎ (2.703) [0.187] −6.214⁎⁎⁎ (−5.071) [− 0.701] −4.0249⁎⁎⁎ (−4.446) [− 0.492] −1.712 (−1.120) [− 0.208] −0.565 (−0.284) [−0.0445] Yes 0.232 21,451

3.531⁎⁎ (2.092) [0.106] −7.235⁎⁎⁎ (−7.348) [−0.207] −5.824⁎⁎⁎ (−3.218) [−0.138] 7.132⁎⁎⁎ (4.770) [0.214] 2.089 (0.909) [0.0645] −4.674 (−1.511) [−0.128]

− 2.0175⁎⁎ (− 2.278) [− 0.248] 1.390⁎⁎⁎ (2.739) [0.189] − 6.300⁎⁎⁎ (− 5.129) [− 0.709] − 3.958⁎⁎⁎ (− 4.236) [− 0.487] − 1.776 (− 1.160) [− 0.214] − 0.584 (− 0.293) [− 0.0498] Yes 0.232 21,337

0.700 (0.350) [0.0216] −6.898⁎⁎⁎ (−6.518) [−0.186] −3.585⁎ (−1.757) [−0.0659] 6.0261⁎⁎⁎ (3.824) [0.175] 0.884 (0.359) [0.0270] −10.688⁎⁎⁎ (−2.885) [−0.280]

−0.909 (−0.921) [−0.105] 1.422⁎⁎⁎ (2.589) [0.185] −8.0663⁎⁎⁎ (−5.752) [−0.898] −4.891⁎⁎⁎ (−4.907) [−0.573] −1.0741 (−0.668) [−0.124] −0.0634 (−0.029) [0.0303] Yes 0.238 18,922

Notes: Numbers in parentheses are asymptotic t-statistics; numbers in square brackets are marginal-effect estimates. ⁎⁎⁎ statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level.

on-farm to off-farm sector in rural areas. Can these results be attributed to human capital externalities that increase the relative payoffs of jobs in the rural off-farm sector? One way of ascertaining this question is to look for evidence that, ceteris paribus, human capital endowment in rural areas help raise the income of those engaged in off-farm activities. This is what we do next using a subsample of individuals who reported labor income but did not migrate to the city. To establish a point of reference, we begin with the estimation of a basic earnings equation that specifies the log of labor income as a function of individual characteristics, including educational attainment, age, age-squared, and gender. Also included in the regression is the number of days worked for labor income. It is important to control for days worked because some, especially the less educated, may have worked outside their households only for part of the year.10 The estimates reported in column (1) of Table 4 are in conformity with the human capital theory. Earnings increase with educational attainment and age. On average, males' earnings are about 14% higher than females'. And as expected, earnings are positively associated with days worked. Also note that the estimates associated with these variables are remarkably stable both quantitatively and in terms of statistical significance across the models reported in Table 4. In columns (2) and (3), we expand the basic earnings equation to include our measures of human capital endowment, namely Cedyr and Colsh, respectively. First, the coefficients on these two human capital measures are both positive, indicating a positive external effect on labor income emanating from human capital at the county level. The estimated coefficient on Cedyr, 0.0379, shows that a one-year increase in county average educational attainment will increase the labor income by 3.79%. Similarly, the estimated coefficient on Colsh, 6.310, suggests that a one-percentage point increase in the share of college educated in the county population will increase the labor income by about 6.3%. Second, the estimated coefficients on Cedyr and Colsh are statistically significant at the 15% and 10% levels, respectively. To check whether the estimated external effect of human capital is subject to omitted variable bias, we expand the basic earnings equation to include, as in the previous subsections, a set of community-specific characteristics. As the estimates reported in columns (4) and (5) indicate, while most of the community characteristics are shown to have significant effects on the labor income, they cause no discernible change in the estimated external effect of human capital. As Colsh is the within-county average of another independent variable, namely College, the coefficient estimate for Colsh may suffer from the type of bias discussed in Acemoglu and Angrist (2000). To check the extent to which the estimated external effect of human capital as given in column (5) is biased, we rerun (in column (6)) the regression of column (5) with college graduates excluded from the sample and College removed from the regression. As the estimated coefficients on Colsh in columns (5) and (6) are very comparable in terms of both magnitude and statistical significance, they do not seem to be an artifact of the regression specification that includes, as a separate regressor, the within-group average of an independent variable. In column (7), we introduce both measures of human capital

10 One seemingly obvious solution is to run the regressions using a sample consisting of those who had full-time off-farm jobs. But to do so would undermine in no small way the objective of the analysis—whether human capital has raised the payoffs of staying in the rural sector (both off- and on-farm activities). It is conceivable that having access to local labor market even on a part-time or seasonal basis would help reduce the incentive to migrate, everything else being the same. For examples of earnings regressions using Chinese data, see Liu (1998), Liu (2003), and Fleisher and Wang (2005).

530

Z. Liu / China Economic Review 19 (2008) 521–535

Table 4 Rural labor income regressions (1)

(2)

(3)

(4)

(5)

(6) a

(7)

(8) b

Cedyr

–

–

–

Colsh90

–

–

6.310⁎ (1.809) –

–

6.150⁎ (1.749) –

5.885⁎ (1.651) –

0.0441 (1.527) 7.0188⁎⁎ (1.971) –

–

–

0.0350 (1.227) –

–

Colsh

0.0379 (1.539) –

College

1.282⁎⁎⁎ (6.082) 1.0975⁎⁎⁎ (8.336) 0.746⁎⁎⁎ (8.540) 0.539⁎⁎⁎ (7.128) 0.292⁎⁎⁎ (3764) 0.00940 (0.812) −0.000138 (−0.879) 0.137⁎⁎⁎ (2.925) 0.00338⁎⁎⁎ (20.426)

1.269⁎⁎⁎ (6.015) 1.0818⁎⁎⁎ (8.194) 0.728⁎⁎⁎ (8.260) 0.525⁎⁎⁎ (6.894) 0.284⁎⁎⁎ (3.652) 0.00994 (0.859) −0.000146 (−0.929) 0.139⁎⁎⁎ (2.965) 0.00337⁎⁎⁎ (20.281)

1.238⁎⁎⁎ (5.838) 1.0965⁎⁎⁎ (8.331) 0.743⁎⁎⁎ (8.509) 0.539⁎⁎⁎ (7.120) 0.291⁎⁎⁎ (3.752) 0.00884 (0.764) −0.000131 (−0.829) 0.138⁎⁎⁎ (2.936) 0.00339⁎⁎⁎ (20.443)

Yes 4239 0.353

Yes 4239 0.353

1.150⁎⁎⁎ (5.612) 1.0579⁎⁎⁎ (8.307) 0.738⁎⁎⁎ (8.708) 0.497⁎⁎⁎ (6.774) 0.280⁎⁎⁎ (3.731) 0.00862 (0.769) −0.000116 (−0.763) 0.173⁎⁎⁎ (3.794) 0.00293⁎⁎⁎ (17.388) −0.557⁎⁎⁎ (−6.593) −0.00664 (−0.149) 0.370⁎⁎⁎ (7.229) 4.0584⁎⁎⁎ (5.923) 4.130⁎⁎⁎ (9.976) −3.751⁎⁎⁎ (−3.486) −0.980 (−1.295) 5.105⁎⁎⁎ (3.773) 5.748⁎⁎⁎ (3.267) Yes 4239 0.397

–

Yes 4239 0.353

1.181⁎⁎⁎ (5.793) 1.0499⁎⁎⁎ (8.229) 0.730⁎⁎⁎ (8.574) 0.491⁎⁎⁎ (6.656) 0.277⁎⁎⁎ (3.692) 0.00992 (0.885) −0.000135 (−0.882) 0.174⁎⁎⁎ (3.808) 0.00291⁎⁎⁎ (17.258) −0.547⁎⁎⁎ (−6.451) 0.00526 (0.117) 0.353⁎⁎⁎ (6.964) 3.770⁎⁎⁎ (5.144) 4.0879⁎⁎⁎ (9.848) −3.482⁎⁎⁎ (−3.187) −1.118 (−1.445) 5.387⁎⁎⁎ (3.944) 6.576⁎⁎⁎ (3.774) Yes 4239 0.397

Vocation school High school Middle school Primary school Age Age-sq Male Days of work Suburb of a city Flatland Telephone services Share of household farming Share of TVE sector Share of agricultural sector Share of industry sector Share of service sector Share of government sector Province dummies Sample Adj-R-sq

1.0526⁎⁎⁎ (8.256) 0.735⁎⁎⁎ (8.670) 0.495⁎⁎⁎ (6.729) 0.279⁎⁎⁎ (3.710) 0.00863 (0.768) − 0.000118 (− 0.768) 0.173⁎⁎⁎ (3.789) 0.00296⁎⁎⁎ (17.504) − 0.549⁎⁎⁎ (− 6.489) − 0.00169 (− 0.038) 0.361⁎⁎⁎ (7.021) 4.139⁎⁎⁎ (6.004) 4.133⁎⁎⁎ (9.945) − 3.881⁎⁎⁎ (− 3.594) − 1.0618 (− 1.397) 5.228⁎⁎⁎ (3.847) 5.643⁎⁎⁎ (3.193) Yes 4201 0.397

1.134⁎⁎⁎ (5.528) 1.0468⁎⁎⁎ (8.207) 0.724⁎⁎⁎ (8.500) 0.487⁎⁎⁎ (6.613) 0.275⁎⁎⁎ (3.657) 0.00930 (0.930) −0.000127 (−0.829) 0.175⁎⁎⁎ (3.841) 0.00291⁎⁎⁎ (17.237) −0.543⁎⁎⁎ (−6.397) 0.00367 (0.081) 0.368⁎⁎⁎ (7.179) 3.652⁎⁎⁎ (4.969) 4.0852⁎⁎⁎ (9.845) −3.466⁎⁎⁎ (−3.174) −1.241 (−1.599) 5.388⁎⁎⁎ (3.946) 5.996⁎⁎⁎ (3.394) Yes 4239 0.398

– 71.764⁎⁎⁎ (5.899) 1.119⁎⁎⁎ (5.403) 1.0984⁎⁎⁎ (8.225) 0.719⁎⁎⁎ (8.269) 0.466⁎⁎⁎ (6.159) 0.237⁎⁎⁎ (3.048) 0.0172 (1.478) −0.000242 (−1.524) 0.158⁎⁎⁎ (3.292) 0.00296⁎⁎⁎ (16.416) −0.6340⁎⁎⁎ (−4.615) 0.0747 (1.496) 0.339⁎⁎⁎ (6.478) 6.0628⁎⁎⁎ (7.817) 3.434⁎⁎⁎ (7.675) −8.430⁎⁎⁎ (−7.044) −3.513⁎⁎⁎ (−4.283) 3.697⁎⁎⁎ (2.637) 1.591 (0.852) Yes 3708 0.417

Notes: Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ Statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level. a Based on the subsample excluding college graduates. b Colsh90, college graduates as a share of county population in 1990, is used in place of Colsh.

endowment at the county level. In this specification, their coefficients are positive and somewhat larger than those obtained when they are introduced separately. The most noticeable change is that the estimated coefficient on Colsh becomes statistically significant at the 5% level, increasing from the 10% level in the previous column. As an alternative way of addressing the potential omitted variable problem, we rerun column (5) replacing Colsh by its lagged value, i.e., Colsh90, and present the results in column (8). In this specification, the estimated external effect of human capital remains positive and is statistically significant at the 1% level. Another possible channel through which human capital externalities make off-farm employment a more attractive choice than migrating to the city or on-farm employment is the creation of employment opportunities in the off-farm sector, especially in the TVE sector. Data limitation precludes a comprehensive investigation into the job-creation role of human capital. Nevertheless, we find some suggestive evidence for it. Table 5 presents two sets of simple regressions that explain the cross-county variation in employment shares by the variation in human capital endowment. In panel A, the dependent variable is the share of non-migrant individuals who engaged in off-farm activities, while in panel B the dependent variable is the share of non-migrant individuals who are employed by TVEs. With the sole exception in row 2 of panel A, we find positive relations between the share of off-farm employment and local human capital endowment, and between the share of TVE employment and local human capital endowment. These relations are particularly strong in terms of statistical significance when Cedyr or Colsh90 is used to measure local human capital endowment.

Z. Liu / China Economic Review 19 (2008) 521–535

531

Table 5 Simple regressions of off-farm employment on human capital endowment (1)

(2)

(3)

Panel A: Off-farm employment share Cedyr Colsh Colsh90 Constant Adj-R-sq Sample size

0.0606⁎⁎⁎ (4.309) – – −0.160⁎ (− 1.872) 0.137 112

– −0.100 (−0.053) – 0.203⁎⁎⁎ (10.093) 0.00 112

– – 32.0618⁎⁎⁎ (4.323) 0.150⁎⁎⁎ (7.573) 0.150 101

Panel B: TVE employment share Cedyr Colsh Colsh90 Constant Adj-R-sq Sample size

0.0313⁎⁎⁎ (4.006) – – −0.122⁎⁎⁎ (0.0474) 0.119 112

– 0.346 (0.331) – 0.0631⁎⁎⁎ (5.724) 0.00 112

– – 19.743⁎⁎⁎ (4.826) 0.0337⁎⁎⁎ (3.097) 0.182 101

Notes: Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ statistically significant at the 1% level; ⁎ statistically significant at the 10% level.

4.4. The effect of human capital externalities by education group The external effect of human capital can come from two sources. The first is complementarity between skilled and unskilled workers. As long as skilled and unskilled workers are not perfect substitutes for each other, an increase in average education (hence the supply of skilled workers) will raise the marginal products and wages of unskilled workers. The second source is the pure spillover. As unskilled workers gain access to a larger pool of human capital, their marginal products and equilibrium wages increase still further. Since the complementarity and pure spillover effects reinforce each other for unskilled workers, the external effect of human capital is expected to raise the labor income of the less less-educated rural residents in the rural sector and hence lower their likelihood of migration. However, for skilled workers an increase in county average education may generate two opposing effects on earnings. On the one hand, an increase in the supply of skilled workers depresses wages of skilled workers. On the other hand, by raising the marginal products of skilled workers, the pure spillover effect raises the demand for and wages of skilled workers. Therefore, for skilled workers the external effect of human capital can be either positive or negative, depending on the relative magnitudes of these two opposing effects. Since the estimates reported so far are obtained by pooling individuals with different educational backgrounds, they indicate the average external effect of human capital externalities across different education groups. It is important, however, to assess to what extent our estimates are driven by the external effect of human capital on the migration decision and labor income of the less-educated individuals. To accomplish this, we divide our sample into two subsamples–one consisting of individuals with middle school or less education and the other consisting of individuals with at least high school education–and rerun some of the regressions reported in Tables 2 and 4. To conserve on space, we report in Table 6 only the estimates associated with our measures of human capital endowment.11 By and large, these estimates are consistent with those obtained based on the full sample: human capital exerts a negative external effect on the likelihood of migration and a positive external effect on labor income in the rural sector. The only two anomalies occur when Cedyr is used in the subsample of individuals with at least high school education in column (4). The estimates associated with Colsh and Colsh90 also suggest that, while human capital externalities affect the migration decision and labor income of people in both education groups, the less less-educated population seem to be somewhat more responsive to the external effect of human capital than the more educated rural residents. 4.5. Instrumental variables estimates In this section, we use the instrumental variables approach to further address the issue of omitted variable bias that may arise because of unobserved factors being correlated with county-level human capital. The IV approach also allows us to obtain consistent estimates if county-level human capital is endogenous or measured with errors. The key to this approach is to identify instruments that are highly correlated with the educational attainment at the county level but uncorrelated with the error term. Based on this principle, we choose the following instruments for Cedyr: province average education (Pedyr) defined as the average education in years of our sample individuals by province,12 the share of county population below 25 years of age (Cshage25), and the interaction term between

11

The full table is available from the author upon request. The instruments for Colsh are college graduates as a share of province population (Pcolsh), Cshage25, and the interaction term between Pcolsh and Cshage25. While the resulting IV estimates for Colsh have expected signs and statistically significant at the 1% level, the estimates associated with all three instruments in the first stage regression all have the wrong sign. For this reason, the IV estimates for Colsh should be considered with a grain of salt. 12

532

Z. Liu / China Economic Review 19 (2008) 521–535

Table 6 Estimating the external effects of human capital by education group Middle school education and below (1)

(2)

Panel A: Rural–urban migration decision: logit regressions Cedyr −0.0571 (−1.089) – Colsh – −8.984⁎ (−1.794) Colsh90 – – Panel B: Rural Cedyr Colsh Colsh90

labor income regressions: 0.0686⁎⁎ (2.119) – –

At least high school education (3)

(4)

(5)

(6)

– – − 184.305⁎⁎⁎ (−4.282)

0.0357 (0.349) – –

– −19.450 (−1.415) –

– – −125.447 (−1.189)

−0.0881 (−1.395) – –

– 5.125 (0.659) –

– – 62.761⁎⁎ (2.298)

dependent variable is log of labor income – – 6.729⁎ (1.695) – – 74.821⁎⁎⁎ (5.415)

Notes: For panel A, columns (1) and (4) have the same specification as column (3) of Table 2; columns (2) and (5) have the same specification as column (4) of Table 2; and columns (3) and (6) have the same specification as column (8) of Table 2. For panel B, columns (1) and (4) have the same specification as column (4) of Table 4; columns (2) and (5) have the same specification as column (5) of Table 4; and columns (3) and (6) have the same specification as column (8) of Table 4. Coefficient estimates for other independent variables not reported here are available from the author upon request. Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ Statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level.

these two variables.13 Province average education is likely to be correlated with county average education because both are influenced to a large extent by the same provincial government through its budgetary expenditures on rural education. Everything else being the same, we would expect that the higher the average education among all rural counties in a province the higher the average education of each individual county. By contrast, there is no compelling reason for us to believe that province average education plays a role in the migration decision or labor income of a rural individual. It is conceivable that human capital externalities are less likely to operate at the province level over a geographically much dispersed area, which makes it more difficult and costlier for individuals from different counties to interact. The relative size of the county population in the age group below 25 is expected to be positively correlated with county average education because these individuals' education was most likely and positively affected by the Compulsory Education Law enacted in China in 1986 that requires all school-aged children to complete at least 9 years of formal schooling.14 The interaction term is used to capture the diminishing effect of the Compulsory Education Law on county average education—ceteris paribus, we would expect counties in provinces with high educational achievement to experience somewhat smaller increases in average education because of the Compulsory Education Law than counties in provinces with low educational achievement. Statistical tests involving these instrumental variables are encouraging. While the instruments are important determinants of county average education—in the reduced reduced-form regression their coefficient estimates all have the expected signs and are highly significant, they do not seem to belong in any of the structural regression equations.15 For example, when they are included in the migration decision regression, none of them obtains a coefficient estimate that is statistically significant.16 Panel A of Table 7 presents the IV estimates for three sets of regressions–binomial logit, multinomial logit, and rural labor income regressions–with two alternative measures for local human capital endowment. Overall, the IV estimates associated with county-level human capital are qualitatively consistent with, but quantitatively larger than, those reported in Tables 2 through 4. Everything else being the same, human capital externalities tend to reduce rural residents' propensity to migrate to the city, increase their propensity to choose rural off-farm relative to on-farm employment, and raise their labor income in the local offfarm sector. These results are further corroborated by the IV estimates (in Panel B of Table 7) from an alternative specification in which individual educational attainment is measured in years of schooling and treated as an endogenous variable as well. The instrument for individual educational attainment is a dummy variable which assumes the value of one for those whose schooling was potentially affected by the Compulsory Education Law (i.e., those below the age of 25 in the year our sample was collected) and zero for those whose schooling was not affected by the law. 4.6. Addressing the issue of potential sample selection bias Recall, in this study migrants are identified as people who left their households for at least one month to work or look for work in the city. Many previous studies adopted similar definitions for migrants. In Rozelle, Gou, Shin, Hughart, and Giles (1999) for example, migrants are those who leave the village for at least one month per year for wage-earning jobs, but retain direct ties to the village by, at the very least, returning during the Chinese New Year or annual peak season farm operations. Since most studies are

13 This identification strategy is very similar to the ones adopted by a growing line of empirical research that tries to measure the effect of peer group influences on individual performance and decision making. See, for example, Evans, Oates, and Schwab (1992). 14 Although it might be argued that the share of individuals less than 25 years of age should be included in the migration decision regression to capture the peer influence among the young who have a greater propensity to migrate, the statistical test shows that this variable plays no role in the migration decision. 15 The coefficients (asymptotic t-statistics) in the first stage regression are 1.356 (11.808) for Pedyr, 17.922 (19.167) for Cshage25, and −3.237 (−20.195) for the interaction term Pedyr ⁎ Cshage25. The first stage regression also includes as explanatory variables all exogenous variables in the migration decision regression. Also see footnote 12. 16 The estimated coefficients (asymptotic t-statistics) are 1.098 (1.139) for Pedyr, 4.322 (0.580) for Cshage25, and − 0.125 (−0.098) for Pedyr ⁎ Cshage25.

Z. Liu / China Economic Review 19 (2008) 521–535

533

Table 7 Instrumental variables estimates of human capital externalities Binomial logit regressions

Multinomial logit regressions

(1)

(2)

(3)

Migration

Migration

Migration

Local off-farm

Migration

−0.532 (−1.488) –

1.0266⁎⁎⁎ (4.494) –

−0.104 (−0.140) 1.990⁎⁎⁎ (6.257) 0.903⁎⁎⁎ (4.548) 0.903⁎⁎⁎ (4.548) 0.533⁎⁎⁎ (3.081) 0.233 21,451

PANEL A: Cedyr and Colsh are endogenous County-level education Cedyr −0.724⁎⁎ – (−2.048) Colsh – −99.905⁎⁎⁎ (−2.599) Individual-level education College −0.524 0.363 (−0.710) (0.442) Vocation school 1.230⁎⁎⁎ 0.955⁎⁎ (4.012) (3.324) High school 1.111⁎⁎⁎ 0.682⁎⁎⁎ (4.881) (3.935) Middle school 0.945⁎⁎⁎ 0.511⁎⁎⁎ (4.789) (3.223) Primary school 0.656⁎⁎⁎ 0.363⁎⁎ (3.821) (2.353) R2 a 0.222 0.222 Sample size 21,451 21,451 PANEL B: Cedyr, Colsh, and Edyr are endogenous County-level education Cedyr −0.705⁎⁎ – (−2.070) Colsh – −127.13⁎⁎⁎ (−2.933) Individual-level education Edyr 0.980 1.025 (0.893) (0.934) 2a R 0.219 0.219 Sample size 21,451 21,451

Rural labor income regressions (4)

(5)

(6)

Local off-farm

Local income

Local income

–

–

–

−71.923⁎ (−1.849)

77.476⁎⁎⁎ (3.050)

0.727⁎⁎⁎ (5.561) –

1.599⁎⁎⁎ (6.646) 1.755⁎⁎⁎ (9.299) 0.496⁎⁎⁎ (3.641) 0.00465 (0.041) −0.473⁎⁎⁎ (−5.014)

0.531 (0.641) 1.790⁎⁎⁎ (5.997) 0.926⁎⁎⁎ (5.298) 0.588⁎⁎⁎ (3.686) 0.319⁎⁎ (2.052) 0.232 21,451

0.921⁎⁎⁎ (2.729) 2.112⁎⁎⁎ (12.059) 1.028⁎⁎⁎ (10.454) 0.511⁎⁎⁎ (5.866) − 0.143⁎ (− 1.735)

1.024⁎⁎⁎ (4.989) 0.874⁎⁎⁎ (6.659) 0.518⁎⁎⁎ (5.550) 0.334⁎⁎⁎ (4.230) 0.194⁎⁎ (2.532) 0.401 4239

0.948⁎⁎⁎ (4.153) 1.054⁎⁎⁎ (8.279) 0.723⁎⁎⁎ (8.500) 0.492⁎⁎⁎ (6.697) 0.273⁎⁎⁎ (3.628) 0.398 4239

−0.500 (−1.451) –

1.058⁎⁎⁎ (4.890) –

–

–

–

−108.78⁎⁎ (−2.484)

54.128⁎ (− 1.670)

0.615⁎⁎⁎ (4.288) –

0.610 (0.544) 0.218 21,451

−1.357 (−1.500)

0.62 (0.580) 0.218 21,451

− 1.367 (− 1.513)

0.192⁎⁎⁎ (2.663) 0.379 4013

0.274⁎⁎⁎ (3.941) 0.376 4013

37.875⁎⁎ (2.345)

2.203 (0.162)

Notes: Besides county-level and individual-level education variables, these regressions also include all exogenous independent variables contained in their counterparts reported in Tables 2, 3, and 4. Specifically, columns (1) and (2) use the same specifications as columns (3) and (4) of Table 3, respectively; columns (3) and (4) use the same specifications as columns (1) and (2) of Table 4, respectively; columns (5) and (6) use the same specifications as columns (4) and (5) of Table 4, respectively. Coefficient estimates for other independent variables not reported here are available from the author upon request. The instruments for Cedyr are Pedyr (province average education), Cshage25 (share of individuals less than 25 years of age in a county), and the interaction term between Pedyr and Cshage25. The instruments for Colsh are Pcolsh (province share of college graduates), Cshage25, and the interaction term between Pcolsh and Cshage25. The instrument for Edyr (individual educational attainment in years of schooling) is a dummy variable which assumes the value of one for those whose schooling was affected by the compulsory education law implemented in 1986 and zero for those whose schooling was not affected by the law. Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level. a Pseudo R-squares are reported in columns (1) through (4).

based on surveys conducted in rural areas, they may exclude migrants who were not present at their rural households at the time of the survey. This raises the question that whether or not these surveys comprise random samples of migrants and non-migrants. If they do not, the resulting estimates would be susceptible to sample selection bias. As this issue applies to our sample as well, we address it from two perspectives. First, suppose that the migrants included in our sample are somewhat different from migrants at large in that they have relatively smaller gains from migrating to the city. This perhaps is one of the reasons why they were present at their rural households at the time of the survey.17 On this account, it is reasonable for us to assume that most (certainly not all) of them are marginal migrants and our sample missed out on some non-marginal migrants. Can we learn anything meaningful about the migration decision from an econometric analysis of a sample consisting of non-migrants and marginal migrants? Our answer is affirmative. Recall, the economic theory of migration suggests that all those who reap positive net gains from migration would choose to migrate. Since the net gains vary across individuals depending on personal and environmental characteristics, a change in any factors that affect the benefits and costs of migration (hence the net gains) may alter the migration decision of marginal migrants, who by definition are indifferent or nearly indifferent between migrating to the city and staying in the rural sector, but not necessarily change the decision of non-marginal migrants who have relatively large net gains from migration. Empirically, the negative external effect of human capital is expected to depend in large part on the 17 A referee pointed out that the migrants included in our sample are likely to be those who failed to migrate to the city and had to return to their rural households. While there is some truth in this contention, a rational individual would decide to return to home if he finds ex post that he is better off staying where he was, even though he could have successfully moved to the city and be worse off. In this sense, it is not whether or not one makes it in the city, rather, as economic theory suggests, it is whether or not it pays for one to migrate.

534

Z. Liu / China Economic Review 19 (2008) 521–535

Table 8 Migration decision: household household-level analysis (1)

(2)

(3)

(4)

County-level education Cedyr Colsh Colsh90

−0.347 (−0.511) – –

– −3.579 (−0.566) –

– – −71.628⁎ (− 1.837)

−0.782 (−1.537) – –

Household-level education Average Edyr Pseudo R2 Sample size

0.0487⁎⁎ (20160) 0.132 7873

0.0441⁎⁎ (2.064) 0.132 7873

0.0298 (1.273) 0.137 6958

0.119⁎⁎ (2.259) 0.133 7873

First-stage instruments Pedyr Cshage25 Pedyr ⁎ Cshage25

– – –

– – –

– – –

1.315⁎⁎⁎ (6.760) 15.829⁎⁎⁎ (10.400) −2.909⁎⁎⁎ (−11.173)

Notes: Besides county-level and household-level education variables, these regressions also include all the household-level variables contained in the regressions reported in previous tables. In column (4) Cedyr is treated as an endogenous variable. The instruments for Cedyr are Pedyr (province average education), Cshage25 (share of individuals less than 25 years of age in a county), and the interaction term between Pedyr and Cshage25. Numbers in parentheses are asymptotic t-statistics. ⁎⁎⁎ statistically significant at the 1% level; ⁎⁎ statistically significant at the 5% level; ⁎ statistically significant at the 10% level.

impact of human capital on the migration decision of marginal migrants regardless the proportion of marginal migrants in our sample. As our sample may contain more marginal migrants than non-marginal migrants, a negative estimated coefficient on county-level education does suggest that human capital externality has a restraining effect at least on the migration decision of marginal migrants, who, according to the theory of migration, are more responsive to changes in the costs and benefits of migration. We also conduct additional regressions of migration decision using the household as the unit of analysis to check the robustness of our results. We define a household as a migrant household if it has at least one member reported to have worked in the city. While this definition may still misclassify households with only one migrant who was not present at home at the time of the survey, it correctly identifies households with multiple migrants and having at least one migrant member included in the survey. In a sense, at the household level our sample represents more accurately the actual distribution of migrant and non-migrant households, and therefore sample selection bias is not as serious as at the individual level. Table 8 presents the estimates of logit models that explain the likelihood for a household to have at least one migrant member. Besides the measure of county-level educational endowment and average years of schooling of household members, these models also contain as covariates all the household- and county-specific variables (not reported in the table) that are included in the previous regressions. In columns (1) through (3), we use the three alternative measures of county-level education, respectively. In column (4), we treat county-level education, namely Cedyr, as an endogenous variable and use the three instruments discussed in the previous section. Overall, these estimates suggest a negative relationship between the likelihood of migration and county-level education, consistent with the results based on the individual-level data. The IV estimate, which accounts for omitted variable bias as well as reverse causality, is nearly significant at the 10% level. Second, there is some evidence suggesting that our data comprise a sample representative of the rural population. The survey used by Zhao (1999a) was conducted during the Chinese New Year, when most migrants return to celebrate the holiday with their families in rural areas. The timing of the survey ensures that the resulting sample represents well the distribution of migrants and non-migrants in the rural population. Her sample consists of a total of 1820 households from 18 counties in Sichuan province, while 798 of 7998 households in our sample were drawn from 8 counties in Sichuan province. As both samples were randomly selected in 1995 from the rural household survey network maintained by the National Bureau of Statistics of China, we can use Zhao's sample as a benchmark to gauge whether migrants are underrepresented in our sample. About 9% of the individuals in our subsample of Sichuan province identified themselves as migrants, very much in line with the proportion (8.4%) of migrants in Zhao's carefully-constructed sample. Another way of assessing whether our sample missed out on a significant number of migrants is to compare the average family size of our sample households with that reported in the 1990 census. Since the census was based on the location of one's hukou rather than the location of work or temporary residence, it should count all migrants as members of their households as long as they have not obtained urban hukou. Therefore, for the rural counties included in our sample if the average family size is significantly smaller than that reported by the census, it would indicate that our sample may have indeed excluded many migrants. However, the comparison suggests quite the opposite—the average family size of our sample is 4.35, larger than the census's figure of 4.1.18 Evidently, sample selection bias is not a serious issue in our study.

18 Assuming there is no change in the number of rural households and rural population grew by 6% between 1990 and 1995, the predicted average size of rural households in 1995 would be the same as our sample average of 4.35 (which is calculated based on the rural individual sample, rather than the household sample). According to China Statistical Yearbook (1997), over the same period China's overall and rural population grew by 6% and 2.1%, respectively. The low population growth in rural areas is mostly due to urbanization—reclassifying some former rural counties as urban towns.

Z. Liu / China Economic Review 19 (2008) 521–535

535

5. Concluding remarks In this paper, we investigate the determinants of rural–urban migration paying special attention to the role of human capital externalities in the rural sector. Our approach is in conformity with the exiting literature that views rural–urban migration as being driven by the difference in the expected incomes or utilities between migrating to the city and staying in the rural sector. We depart from the literature, however, by explicitly recognizing that the presence of human capital externalities in rural areas can restrain rural–urban migration. This is quite intuitive because any factor that has the effect of reducing the relative payoffs of migration will make migration a less attractive choice. Apart from the fact that it has not been considered by the vast literature, accounting the role of human capital externalities in the analysis allows us to shed new light on the general role of human capital in influencing individual decisions regarding rural–urban migration and employment choice in the rural sector. While for a rural resident the likelihood of migration increases with his own educational attainment, as shown by the existing literature and confirmed by the present study as well, our estimates show that, ceteris paribus, an increase in the average educational attainment or college graduates as a share of the rural population reduces a rural resident's propensity to migrate. We also find some evidence that human capital exerts a positive external effect on the likelihood for a rural resident to choose rural off-farm employment over on-farm employment. This result is corroborated by the finding of a positive relation between labor income and human capital endowment in the rural sector. Evidently, human capital externalities increase the payoffs of off-farm employment relative to migration and on-farm employment by raising the level of labor income in the rural off-farm sector. These results are by and large corroborated by the IV estimates. Our empirical analyses further suggest that, while human capital externalities play an important role in the migration decision of individuals of all education backgrounds, the less educated seem to be somewhat more susceptible to the influence of local human capital. A unique and important policy implication from our analyses is that expanding education opportunities in rural areas can be a viable strategy for curtailing rural–urban migration and alleviating urban unemployment pressure—quite contrary to the implication presented in some studies that promoting rural education would hasten rural–urban migration and hence exacerbate unemployment problem in the urban sector. While we find some suggestive evidence that human capital endowment and off-farm employment opportunities are positively correlated, it is desirable to look for direct evidence on how the concentration of human capital helps create jobs. Do areas rich in human capital have more entrepreneurs than areas poor in human capital? Do areas rich in human capital attract more public or private investments than areas poor in human capital? The answer to these questions will provide further insights into the role of human capital in rural–urban migration as well as rural development. We will look into these issues in future research. References Acemoglu, Daron, & Angist, Joshua (2000). How large are human capital externalities? Evidence from compulsory schooling laws. In Ben S. Bernanke, & Kenneth Rogoff (Eds.), NBER macroeconomics annual, Vol. 15. (pp. 9−59)Cambridge, MA: MIT Press. De Brauw, Alan, Huang, Jikun, Rozelle, Scott, Zhang, Linxiu, & Zhang, Yigang (2002). The evolution of China's rural labor markets during the reforms. Journal of Comparative Economics, 30, 329−353. De Brauw, Alan, & Giles, John (2008). Migrant opportunity and the educational attainment of youth in rural China. World Bank policy research working paper No. 4526. Evans, William N., Oates, Wallace E., & Schwab, Robert M. (1992). Measuring peer group effects: A study of teenage behavior. Journal of Political Economy, 100(5), 966−991. Fleisher, Belton M., & Wang, Xiaojun (2005). Returns to schooling in China under planning and reform. Journal of Comparative Economics, 33(2), 265−277. Giles, John and Mu, Ren (2005). Elder parent health and the migration decision of adult children: Evidence from rural China. Working paper. Hare, Denis (1999). ‘Push’ versus ‘pull’ factors in migration outflows and returns: Determinants of migration status and spell duration among China's rural population. Journal of Development Studies, 35(3), 45−72. Harris, John R., & Todaro, Michael P. (1970). Migration, unemployment, and development: A two-sector analysis. American Economic Review, 60(1), 126−142. Lall, Somik V., Selod, Harris, & Shalizi, Zmarak (2006). Rural–urban migration in developing countries: A survey of theoretical predictions and empirical findings. World Bank Policy Research Working Paper 3915, May 2006. Liu, Zhiqiang (1998). Earnings, education, and economic reforms in urban China. Economic Development and Cultural Change, 46(4), 697−725. Liu, Zhiqiang (2003). The economic impact and determinants of investments in human and political capital in China. Economic Development and Cultural Change, 51(4), 821−849. Liu, Zhiqiang (2005). Institution and inequality: The hukou system in China. Journal of Comparative Economics, 33, 133−157. Liu, Zhiqiang (2007). The external returns to education: Evidence from Chinese cities. Journal of Urban Economics, 61, 542−564. Lucas, Robert E., Jr. (2004). Life earnings and rural–urban migration. Journal of Political Economy, 112(1). S29–59 pt. 2. Moretti, Enrico (2004). Estimating the social return to higher education: Evidence from longitudinal and repeated cross-sectional data. Journal of Econometrics, 121, 175−212. Rauch, James E. (1993). Productivity gains from geographic concentration of human capital: Evidence from the cities. Journal of Urban Economics, 34, 380−400. Rozelle, Scott, Guo, Li, Shen, Minggao, Hughart, Amelia, & Giles, John (1999). Leaving China's farms: Survey results of new paths and remaining hurdles to rural migration. The China Quarterly, 158, 367−393. Todaro, Michael P. (1969). A model of migration and urban unemployment in less-developed countries. American Economic Review, 59(1), 138−148. Zhao, Yaohui (1999). Labor migration and earnings differences: The case of rural China. Economic Development and Cultural Change, 767−782. Zhao, Yaohui (1999). Leaving the countryside: Rural–urban migration decision in China. American Economic Review, 89(2), 281−286 Papers and Proceedings, May.