Three Essays on Regional Income Disparity DISSERTATION

Three Essays on Regional Income Disparity DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Yue Hua, B.A. Graduate Program in Agricultural, Environmental and Development Economics The Ohio State University 2014 Dissertation Committee: Mark Partridge, Advisor Alessandra Faggian Joyce Chen

Copyrighted by Yue Hua 2014

Abstract This dissertation is composed of three essays on regional income disparity. The first essay examines the decision between internal migration and home production for rural households and its impact on rural income distribution. By constructing counterfactual scenarios under which households are allowed to switch freely between internal migration and home production, this study finds that the migrant households in the studied region could have earned more had they choose not to migrate and work in local sectors, given the results that show remittances earned by the migrant households are less than their simulated home production earnings. The findings also illustrate that there would also be less income inequality in this area if migrants choose to work locally. These results are compatible with the fact that the internal migration in the study area is very likely to be involuntary, possibly due to the lack of arable land and insufficient local nonfarm job opportunities, usually provided by township and village enterprises. The second essay explains why U.S. workers earn more in urban areas than in rural areas. This spatial wage disparity can result from the agglomeration effect that makes workers and firms more productive in cities, or from sorting by observed or unmeasured individual skills. By employing a female panel dataset, this paper empirically tests the sorting and learning hypotheses that might drive the urban wage premium for female workers, using both the simple OLS and the fixed-effects estimation. We also appraise ii

the gender differences by comparing the male-sample and female-sample results. Findings suggest that sorting by observed ability has little impact on urban wage premium, while unmeasured fixed variables are the major driving forces. Agglomeration effect does exist for female workers, albeit not as evident as the effect on their male counterparts. The third essay focuses on the empirical tests of the existence, magnitude and driving forces of the intra-urban wage gradient, by employing the 2000 Census data for three large U.S. metropolitan areas and by estimating wage equations for urban workers. Findings suggest that within metro areas, commuting times and wages are not converging over time, which indicates that temporary spatial employment disequilibrium is not an underlying reason for the intra-urban wage gradient. Results also show that both the urbanization and the localization economy are linked to the wage level. This supports the hypothesis that the concentration of firms in the central urban area is sustained by an agglomeration effect, an effect that is becoming more and more important over time in the process of urban wage determination. iii

Acknowledgments I would never have been able to finish my dissertation without the guidance of my committee members, help from friends, and support from my family. I would like to express the deepest appreciation to my advisor and committee chair Professor Mark Partridge, for introducing me to urban and regional economics, for providing me with persistent and excellent research guidance, for funding my trips to academic conferences, and for giving me opportunities to review journal article and to share my ideas with the non-academic public. I would like to thank my committee members, Professor Alessandra Faggian, for helping me finalize my research topics and providing me with the experience of reviewing journal article, and Professor Joyce Chen, for a great introduction to development economics and for providing constructive and critical suggestions for my dissertation. I would like to thank Professor Abdoul Sam, who introduced me to econometrics and helped me on many technical problems. I would also like to thank Professor David Blau, a great teacher, who opened the door of labor economics for me. I would like to thank Brent Richard, Sanghyo Kim, Michael Farren, and Wendong Zhang, who as good friends, were always willing to help and give their best suggestions. Finally I would like to thank my parents and my girlfriend. They were always supporting me and encouraging me with their best wishes. iv

Vita June 2006...Wuhan Foreign Language School June 2010...B.A. Economics, Nankai University September 2010 to present...graduate Teaching Associate, Department of Agricultural, Environmental and Development Economics, The Ohio State University Fields of Study Major Field: Agricultural, Environmental and Development Economics Specialization: Urban and Regional Economics, Labor Economics, Chinese Economy v

Table of Contents Abstract... ii Acknowledgments... iv Vita... v List of Tables... viii List of Figures... x Chapter 1: Migration Decision and Rural Income Inequality in Northwestern China... 1 1.1 Introduction... 1 1.2 Remittances in Rural China and Literature Review...5 1.3 Data Description...8 1.4 Estimation Strategies...10 1.5 Estimation Results...15 1.6 Inequality Analysis and Policy Implications...22 1.7 Conclusion... 28 Chapter 2: What Drives the Urban Wage Premium? Evidences from a Female Panel.. 30 2.1 Introduction... 30 2.2 Data and Sampling Issues... 34 2.3 A Simple Model of Learning in Cities... 37 2.4 Testing for the Sorting Hypothesis...40 2.5 Testing for the Learning Hypothesis...45 vi

2.6 Gender Difference Analysis...48 2.7 Conclusion...51 Chapter 3: Spatial Disequilibrium or Agglomeration? New Evidence on Intra-Urban Wage Gradient...53 3.1 Introduction...53 3.2 Previous Empirical Works... 55 3.3 Conceptual Framework...58 3.4 Data Description and Intra-Urban Wage Variation...60 3.5 Estimation Strategies and Results...65 3.6 Testing for Agglomeration Effects...69 3.7 Conclusion...73 Bibliography...75 Appendix A.1: Poverty Map of China..81 Appendix A.2: Test of Overidentifying Restrictions 82 Appendix B: Descriptive Statistics for Low-Weeks-Worked-Per-Year Workers...83 Appendix C.1: Super-PWPUMA Definitions..84 Appendix C.2: Super-PWPUMA Maps of the Studied Metropolitan Areas... 86 Appendix C.3: Industries and Occupations... 88 Appendix C.4: Zonal Import Ratio for Private-Sector White Workers... 90 vii

List of Tables Table 1.1 Selected Descriptive Statistics...10 Table 1.2 Results of the Home Production Participation Equation...16 Table 1.3 Selected Results of the Nonmigrant Household Income Model...18 Table 1.4 Selected Results of the Migrant Household Income Model...20 Table 1.5 Average Household Income in Three Scenarios...21 Table 1.6 Decomposition of the Rural Household Income Inequality...24 Table 2.1 Selected Descriptive Statistics...41 Table 2.2 OLS and Fixed-Effects Results...44 Table 2.3 Results of the Mover's Model...47 Table 2.4 Summary of the Male-Sample Results...50 Table 3.1 Intra-Urban Wage Variation...62 Table 3.2 Intra-Urban Wage Variation for Public Sector Workers in the Chicago MSA... 63 Table 3.3 Intra-Urban Commuting Time Variation (in minutes)...64 Table 3.4 Travel Time Coefficients for Public and Private Sector Workers...67 Table 3.5 Travel Time Coefficients for Private Sector Workers by Race...68 Table 3.6 Travel Time Coefficients for Private Sector White Workers by Gender...69 viii

Table 3.7 Total Employment and Import Ratio (IR) Coefficients for Private-Sector White Workers...70 Table 3.8 Specialization Index Coefficients for Private Sector White Workers...71 ix

List of Figures Figure 2.1 Urban Wage Premium for Female Workers...34 Figure 2.2 Detailed Urban Wage Premium for Female Workers...35 Figure 3.1 Intra-Urban Wage Gradient in the Boston, New York and Chicago MSA for Private Workers...64 Figure A.1 China Annual Per Capita Income by Province...81 Figure C.1 Super-PWPUMAs of the New York MSA...86 Figure C.2 Super-PWPUMAs of the Chicago MSA...87 Figure C.3 Super-PWPUMAs of the Boston MSA...87 x

Chapter 1 Migration Decision and Rural Income Inequality in Northwestern China 1.1 Introduction Internal rural-urban migration in China has become a highly noticeable social and demographical phenomenon since the beginning of the economic reform in 1978. Urban areas, especially large cities, have been developing disproportionately, which is the consequence of the unbalanced domestic development strategy focusing on urban industrialization since 1949. 1 This rural-urban divide and growth disparity motivates huge amounts of rural surplus laborers to seek jobs in urban areas for both pecuniary and nonpenuciary payments. A migrant worker in China can be defined as a rural laborer who works in urban area but does not have a formal urban resident identity, or urban hukou. 2 Thus, migrant workers are treated very differently from urban workers in terms of remuneration, both by the government and by the urban employers. They receive lower wage than their equally-productive urban colleagues, and are often excluded from the basic social security programs. 1 The adoption of this development strategy reflects the doctrine of learning from the Soviet Union and Stalinist Communism. Mao believed that the heavy machinery sector should always be the first priority in economic development to enhance the national defense. 2 China s modern hukou System (Household Registration System) started in 1950. It divides residents formally into rural residents and urban residents based on their residence and biological kinship. 1

As the world s largest floating population, the number of the Chinese migrant worker was 180 million at the end of the 1990s. In 2012, this number has exceeded 250 million, 3 and it has been predicted that the number of migrant workers will continue to increase swiftly, reaching approximately 300 million by 2015 (Yang, 2009). Most of the migrant workers are low-skilled young workers with limited educational attainment. These workers are employed mainly in mining, manufacturing, construction and low-skilled service sectors which often require intensive physical work. The characteristics of the Chinese migrant workers can be attributed to both rural labor supply and demand. On the supply side, China has 674 million rural residents, with more than 150 million registered rural households. Traditional rural Chinese household makes fertility decision that aims at maximizing the household s productivity, thus girls are often discriminated against and boys with more physical working abilities are preferred. On the other hand, the one-child policy is difficult to implement in many rural areas due to the lack of monitoring and judicial resources, so rural households often choose to have multiple children. Children within one household would naturally compete for limited parental cares and educational resources with their siblings. Poorly educated parents and undeveloped rural school system can also adversely affect the children s intellectual development and school performance, especially those children who are left behind. The result is high dropout rates and low average educational attainment. 3 Reported in China Statistic Yearbook by the National Bureau of Statistics of China. 2

On the demand side, shortage of arable land has always been a binding constraint on agricultural production in China (Zhu and Luo, 2009). As reported annually by the Ministry of Land and Resources, China s arable land per capita has been declining for more than 10 years. In 1997, arable land per capita was 1.58 mu, and this number declined to 1.38 mu in 2011. 4 This is not only because of the increasing rural population, but also because of the shrinking amount of total arable land due to urban sprawl and environmental deterioration. Adoption of advanced agricultural technologies and machineries in some area leads to a high marginal product of labor, which also decrease the demand for agricultural laborers. When both rural farm and nonfarm sectors are unable to absorb all these redundant laborers due to limited capacity, migrating to urban areas becomes the only option to seek better pecuniary payments for these rural surplus laborers. Urban firms have strong incentive to hire these migrant workers who would accept a much lower wage relative to their equally-productive urban counterparts. Internal migration is not barrier free and costless, especially for the relatively poor households. Successful migration often requires contacts, know-how and some capital or at least the ability to borrow (Davin, 1998). In China, transportation costs are high for the undeveloped areas that are remote and inaccessible. Living costs (rents and utility fees) in urban areas are also increasing due to the long-lasting domestic inflation. One of the most important social welfare issues in cotemporary China is the income inequality in the rural area. The Gini coefficient of the rural area is 0.39 in 2011, very 4 1 mu = 667 m 2 3

close to 0.4, which is the recognized security line of uneven distribution. 5 Large income inequality in China often induces severe hierarchical conflicts between the rich and the poor, resulting in discrimination against the poor, crimes against the rich and distrust of the government. While most of the minority population resides in rural areas, income inequality across different minorities can even create secessionism which may threaten national security. Thus, rural income inequality can significantly affect the social stability in China via different channels and should receive some serious inspections. Remittances, defined as money transmitted to rural households by migrant workers (Adams, 1989), are very likely to change the income distribution in the source communities. Hence, examining the possible effects of remittances on rural income inequality is critical to explore the welfare implication of internal migration, and can also provide useful guidance for policy makers who aim at constructing a more stable and harmonious society. Using data from a poverty monitoring survey of rural households in Gansu and Inner Mongolia province in northwestern China, this study contributes to the existing literature by performing quantitative analyses on remittances and income inequality in this area for the first time. This region is worth studying because it is geographically and economically different from other macroregions in China (Naughton, 2007), which can lead to significant differences in the household decision between migration and home production. Findings show that while initially less-endowed migrant households are better off by sending out migrant workers and receiving remittances in absolute term, 5 Reported by the National Bureau of Statistics of China. 4

they can actually be even more economically improved if they choose to work in local sectors, assuming such job opportunities are always available. Also, to reduce the overall income inequality in this region, local government should focus either on increasing migrant workers net earnings in the urban area, or on creating more local job opportunities. The rest of the paper is organized as follows. Section 1.2 illustrates the importance of remittances for rural households and provides a brief review of some representative empirical evidence. Section 1.3 presents the methodology used for estimation and simulation. Section 1.4 describes the data. Section 1.5 reports and interprets the estimation and simulation results. Section 1.6 focuses on inequality analysis and policy implications. Section 1.7 concludes. 1.2 Remittances in Rural China and Literature Review In China, it is conventional for migrant workers to send or bring part of their annual earnings back to their rural households as remittances. Prior studies showed that the remittances share of total household income is generally low in Asian countries. One study on 12 villages in northern India finds that the average remittance share in total household income is 6.9% (Tumbe, 2011). Other studies on Thailand, Malaysia and Indonesia find even lower remittance shares (Zhou, 2006). China, however, has a much higher remittance share than other Asian countries due to both economic and cultural reasons. Migrant workers, under the influence of collectivism culture, possess a strong sense of responsibility to earn both pecuniary and non-pecuniary resources for their 5

family members. The remittances also serve as an income smoothing tool by diversifying the source of income for rural households, especially for those households whose income depends solely on farming activities, which is very sensitive to natural conditions. As reported by the Consultative Group to Assist the Poor (CGAP), the overall remittances share in total migrant household income in China is 30%-50%. Remittances serve as a critical source of rural household expenditures on education, job training, medical care and daily life. Remittances, as a critical part of the total household income, can also affect the rural income distribution conditional on the position of the migrant households in the income spectrum. If the first households to make a migration decision are from the upper end of the income distribution, they are more likely to receive high returns from this high-risk investment (Taylor, 1986). However, if the households who choose to migrate are at the lower end of income distribution and are forced to do so because they have surplus laborers that cannot be absorbed by the local labor market, the remittances earned by these migrant workers may help the migrant households to catch up with their wealthier nonmigrant neighbors and the local income inequality could be reduced.. Stark (1986) takes remittances as an extra net income of the migrant households that has no substitutes if migrant households do not participate in migration. Adams (1989) issued a study on remittances and inequality in rural Egypt, in which he criticized and improved the Stark approach. The Adams approach no longer conceives remittances as purely exogenous, but treats remittances as a substitute for home production earnings, which highlights the existence of the opportunity cost of migration. Enlightened by the 6

Adams approach, Bradford and Boucher (1995) further contribute to the research of this topic by the improving the methodology. They formally introduce the migration participation equation, which enables them to control for the sample selection problem (incidental truncation) when estimating the impact of remittances on rural household income. The theoretical framework on this topic came to its maturity before studies on rural China started in the late 1990s. Zhao (1999) first conducted a case study on migration decision and earning differences in Sichuan province in southwestern China. He finds that earnings received by migrant workers are much higher than what they would receive as workers in local farm or nonfarm sectors. But internal migration in the studied area has high fixed costs. Du, Park and Wang (2005) restudied this issue using data from four western provinces. They observe that the rural policy reform initiated in the late 1990s benefits rural households disproportionately, in the sense that the poorest rural households still find migration extremely costly, while middle-class rural households benefit a lot from lower agricultural taxes, more transportation subsidies, better working conditions and higher labor insurance coverage. Thus, migration can enlarge income inequality under this scenario. Zhu and Luo (2009) study the impact of migration on income inequality in Hubei province, situated in central China. In their sample, nonmigrant households have richer land resources and higher initial household wealth, but remittances received by migrant households are more than home production earnings of nonmigrant households, thus participating in rural-urban migration is both individually and socially optimal. 7

1.3 Data Description The data used in this study comes from a China regional poverty monitoring survey, lasted consecutively from 1999 to 2004. The survey was supported by the World Bank and conducted jointly by the National Bureau of Statistics of China and the China Poverty Relief Office in Gansu and Inner Mongolia Province, which are two large and less-developed provinces in northwestern China (See Appendix A.1 for a poverty map). The dataset actually used is a two-period pooled cross sectional dataset composed of the 2001 and 2004 survey (Wave 3 and Wave 6). 6 The 2001 and 2004 survey covered 700 and 1500 rural households, respectively. The surveyed households were located in 150 administrative villages in 15 poverty-stricken counties, subject to independent random sampling with no time or spatial correlation between two years. This sample is representative of the low-income rural population in this region, which is the population of interest in this study. Among the 700 households surveyed in 2001, 295 households participated in rural-urban migration, while the other 405 households did not. Among the 1500 households surveyed in 2004, 633 households sent out at least one migrant worker, while the other 867 households opted to stay in rural area. The share of the migrant households in 2001 and 2004 was 42.1% and 42.2%, respectively. All the 928 migrant households in the sample received a strict positive amount of remittances, while the 1272 nonmigrant households received no remittances. 6 The data of this program is non-public and can only be used by authorized personnel. The 2001 and 2004 data were released to the graduate students in the department of agricultural economics at Renmin University for pedagogical purpose, other data remains undisclosed and is unavailable upon request. An introduction of the program in Chinese can be found at http://www.wl.cn/895306. 8

The survey provides quite comprehensive household level information, but does not collect detailed individual information for every member in the household. The numbers of household laborers in three age brackets (age 16-30, age 31-50 and older than 50) are used to control for household s age structure. We use the number of household members with an educational attainment above 9 th grade to control for household s education level and unobserved preference for education. In rural China, children receive governmental subsidies that cover part of their expenditures on education until 9 th grade. After 9 th grade, they are no longer eligible for the benefit and the households will face a different budget constraint if they make further investment in child education. So household s education preference can be revealed by the number of relatively well-educated members in the household. Table 1.1 reports the mean of the key variables for the full sample and for the two subsamples characterized by migration decisions. It is shown that migrant households in this sample are poorer than nonmigrant households with 28.5% less total household income on average. Migrant households have 68.1% more young laborers with age between 16 and 30, but own 103.3% less arable lands than nonmigrant households on average. Migrant households also receive less education and vocational training. These are important indicators that migrant households in this sample are initially less endowed and might be forced to participate in migration by sending out surplus laborers. The fertility decision of nonmigrant households to bear more children further reflects that they are located at the upper half of the income distribution. 9

Table 1.1 Selected Descriptive Statistics All Nonmigrant Migrant Households Households Households Total Household Income (CNY) *** 9648.75 10646.29 8284.62 Laborers of Age 16 to 30 *** 0.61 0.47 0.79 Laborers of Age 31 to 50 * 0.76 0.74 0.79 Laborers of Age above 50 *** 0.55 0.49 0.62 Members with Schooling above 9 *** 0.16 0.18 0.14 Members with Job Training ** 0.14 0.16 0.13 Arable Land Area (m 2 ) *** 1933.98 2458.99 1209.79 Number of Children under 6 ** Remittances (CNY) *** 0.34 1157.84 0.44 0 0.19 2744.87 Home Production Earnings (CNY) *** 7622.53 8638.29 4592.16 Number of Observations 2200 1272 928 Note: (1) 1 CNY = 0.16 USD (2) Home production earnings are defined as earnings from local units, including both the income from farming activities and the income from being employed by local enterprises. (3) *** Difference significant at 1%, ** significant at 5%, * significant at 10% 1.4 Estimation Strategies The substituting relationship between home production earnings and remittances is plausible for the rural households in China, which verifies the applicability of the Adams approach. Participating in migration always incurs a high fixed cost and is typically considered as a critical alternative for working at home, including both farm and nonfarm productive activities (Zhao, 1999). Because most of the migrant workers are young males 10

with relatively high physical working ability, the household would have to face a large decline in the household productivity if it chooses to send out a migrant worker (Li, 2011). Thus, there are obvious opportunity costs of getting remittances. In this study we admit the fact that households must sacrifice some positive amount of home production earnings to participate in migration and receive remittances, and vice versa. Two economic counterfactual groups are created. In the first group, the 928 migrant households are assumed to be nonmigrant households and receive home production earnings instead of remittances. In the second group, the 1272 nonmigrant households are assumed to be migrant households and receive remittances at the expense of some home production earnings. Thus remittances and home production earnings are treated as substitutes, which is the core of the Adams approach. These constructions are based on the assumption that rural households would always make working decisions between migration and home production, so there are no idle laborers, which is not implausible given the working motivation and ability of the rural young laborers in China. In the first counterfactual group, we know the income of the actual 1272 nonmigrant households, but not the potential home production earnings for the 928 migrant households. Therefore this part of (counterfactual) income for the migrant households needs to be simulated. By the same argument, we need to simulate the potential remittances for the 1272 nonmigrant households in the second counterfactual group. By comparing the income level and income distribution of the actual group with those of the two counterfactual groups, we can determine whether participating in migration is the 11

first-best choice for migrant households, and symmetrically, whether staying in rural areas is optimal for nonmigrant households. 7 To simulate the home production earnings under the first counterfactual scenario, the unbiased estimators of the coefficients of the nonmigrant household income model are required. The nonmigrant household income model takes a linear form and is written as log(y nr i ) = ρd i +βx i + δd i G i + e i (1.1) where Y i nr is the total household income for household i, the superscript nr indicates that all households are assumed to be nonmigrant households and receive no remittances. D i represents the year dummy set to one if the household was surveyed in 2004, to allow for different intercepts for two periods. X i is a vector of control variables, including household s information on initial endowments and its demographics. G i is a subset of X i and contains variables that may affect household income differently across the two periods, such as educational level and job training. e i is the disturbance term. Because we are now in the first counterfactual scenario, the actual Y nr i is unknown for the 928 migrant households. We first estimate the actual nonmigrant subsample with 1272 observations to get the unbiased coefficients for the income equation, and then apply these coefficients to predict the home production earnings for the 928 migrant households, giving their characteristics. Since the nonmigrant households might be self-selected into home 7 An alternative approach is to use propensity score matching, which requires many details to be filled in such as how to model the score and how to do inference. However, the implementation of matching is not yet standardized and different researchers might reach very different conclusions, even when using the same data and covariates (Angrist and Pischke, 2009). We thus believe regression should still be the starting point for this work. 12

production activities, this subsample estimation may incur the sample selection problem which may prevent us from obtaining unbiased estimates. Following the Heckman s twostep method for selection bias correction, a participation equation that incorporates the exclusion restrictions is estimated, to test for the existence and magnitude of potential sample selection bias. The participation equation is written as: R i = ρd i +αz i + δd i K i +v i (1.2) R i is the home production dummy which equals unity if the household engages in home production and receives no remittances, and equals 0 otherwise. Z i is a vector of exogenous variables that contains all the control variables in X i and some variables that affect household s home production decision but not their income directly. K i contains all the variables in G i plus the exclusive restrictions. The Inverse Mills Ratio (IMR) λ i is then calculated from the participation equation and added into Equation (1.1) as a regressor, which yields: log(y i nr ) = ρ c D i +β c X i + δ c D i G i +λ i + u i (1.3) The selection-bias-corrected coefficients ρ c,β c and δ c can now be used to simulate the deterministic part of potential home production earnings for migrant households. For the unobserved residuals, a random value is generated based on the observed error term of nonmigrant households: μ i = φ i Φ -1 (r) (1.4) 13

where φ i is the estimated disturbance of nonmigrant households, r stands for a random number between 0 and 1, and Φ -1 is the inverse of the cumulative probability function of the standard normal distribution (Zhu and Luo, 2009). The estimating process is much the same for the second counterfactual group in which all households are assumed to be migrant households. We acquire unbiased coefficients of the wage equation from the actual migrant subsample and apply those coefficients to simulate the unknown household income for nonmigrant households, assuming they received some remittances. The migrant household income model takes the form: log(y i r ) = ρd i +βx i + δd i G i + e i (1.5) Y i r now stands for the total household income with superscript r indicating all the households are now assumed to be migrant households. Using the actual migrant household subsample with 928 observations, a first stage migration participation equation is estimated, the inverse mills ratio is then calculated and added into Equation (1.5), which yields: log(y r i ) = ρ w D i +β w X i + δ w D i G i +λ i + u i (1.6) Subsequently, the selection-bias-corrected coefficients ρ w, β w and δ w are used to simulate the potential remittances for nonmigrant households. Unobserved residuals are constructed using the observed residuals of migrant households and following the same procedures in the first counterfactual. 14

1.5 Estimation Results We first examine the counterfactual scenario in which all households are assumed to be nonmigrant households, and receive some positive amount of home production earnings. This counterfactual scenario is designed to explore whether migrant households would earn more had they chose to and work locally and not to migrate. Table 1.2 presents the selected estimation results for Equation (1.2). Recall that the econometric purpose of estimating this equation is to obtain the inverse mills ratio for the subsequent selection-bias correction. Furthermore, the estimators can be used to verify the demographic patterns of migrant households (and symmetrically nonmigrant households), complementary to the sample statistics in Table 1.1. The number of adult laborers is shown to be negatively related to the probability of engaging in home production, regardless of the laborers age. This relationship can be well-explained by the large amount of surplus laborers in rural China area at all age levels. Since rural industrial sectors are undeveloped and have very limited capacity to absorb surplus laborers, rural-urban migration is the only option for surplus laborers to get a job with better payment. Hence, for an average rural household, more laborers often imply a higher probability of sending out migrant workers. The positive relationship between the number of relatively well-educated members and the probability of home production is consistent with the fact that migrant households are less educated on average in the studied area, which can make them disadvantageous in competing with nonmigrant households on local resources. Rural job training programs 15

serve as a less costly substitute for formal education for these migrant households. These vocational training programs mainly focus on low-skill service jobs, such as truck driving, mail sorting and haircutting (Li, 2011). The trainings are fairly job-oriented and most of the trainees can be employed in urban servicing sectors after they graduate, resulting in a larger proneness of participating in rural-urban migration. Table 1.2 Results of the Home Production Participation Equation Coefficient Std. Error Laborers of Age 16 to 30-0.46 *** 0.05 Laborers of Age 31 to 50-0.82 *** 0.07 Laborers of Age above 50-0.32 *** 0.07 Members with Schooling above 9 0.64 *** 0.11 Members with Job Training -0.99 *** 0.29 Marriage (=1 if married) 1.12 *** 0.25 Number of Children under 6-0.08 0.06 Arable Land Area (m 2 ) 0.0005 *** 0.0001 Distance to the nearest bus station 0.26 *** 0.06 Distance to the nearest town center 0.12 *** 0.05 Year (=1 if surveyed in 2004) -3.06 *** 1.31 Constant 4.06 0.52 Observation 2200 Log-likelihood -662.61 Pseudo R 2 0.25 Note: (1) Other control variables include interactions of year dummy with education, training, arable land, and two distance variables. (2) Standard errors in parentheses; * significant at 10%, ** significant at 5%, *** significant at 1%. (3) The two exclusion restrictions passed the F-test for weak instruments with a F statistic of 13.26. 16

Households who possess more arable land tend to input more human capital resources in home production, especially in farming activities. In China, lands are nationalized and non-tradable. The only way for rural households to obtain arable land is to sign the land leasing contract with local government through the Housing Responsibility System. Thus, households cannot purchase or sell arable lands to change their initial land endowment in the short run. Lastly, marriage tend to keep the workers stay in rural area, while number of young dependents does not seem to affect one s working decision significantly. The exclusion restrictions used in the participation equations (the first stage of the Heckit Method) are the distance to the nearest bus station and the distance to the nearest town center. In the two surveyed provinces, infrastructure, including the road system, is extremely undeveloped, due to historical and geographical reasons. Thus the transportation cost is fairly high for migrant workers, considering their circulatory pattern. This indicates that the distance to the nearest bus station and town center can substantially affect household s migration decision, which is proven by the positive and statistically significant results in Table 1.2. It should be admitted that, in theory, these two distance variables could be expected to affect rural household income. But the surveyed area is remote and home production is largely characterized by traditional and autarky agriculture, so it is economically valid to ignore the feeble influences that the distance variables might have on household incomes through channels other than internal migration. Test of overidentifying restrictions is shown in Appendix A.2. 17

Table 1.3 reports the selected estimation results of Equation (1.1) and Equation (1.3) for the subsample composed of 1272 nonmigrant households, with and without the IMR. Table 1.3 Selected Results of the Nonmigrant Household Income Model OLS without IMR OLS with IMR Equation (1.1) Equation (1.3) Laborers of Age 16 to 30 0.02 (0.02) 0.07 *** (0.02) Laborers of Age 31 to 50 0.17 *** (0.02) 0.15 *** (0.03) Laborers of Age above 50 0.14 *** (0.03) 0.12 *** (0.03) Members with Schooling above 9 0.11 *** (0.03) 0.11 *** (0.03) Members with Job Training -0.05 (0.31) -0.09 (0.31) Marriage (=1 if married) -0.21 *** (0.06) -0.14 (0.09) Number of Children under 6-0.13 *** (0.03) -0.13 *** (0.03) Arable Land Area (m 2 ) 0.0003 *** (0.00004) 0.0003 *** (0.0004) Year (=1 if surveyed in 2004) -0.17 ** (0.08) -0.26 ** (0.12) Inverse Mills Ratio N/A N/A 0.25 *** (0.07) Constant 8.95 *** (0.87) 9.05 ** (0.15) Observation 1272 1272 Adjusted R 2 0.34 0.34 Note: (1) Other control variables include interactions of year dummy with education and training, (2) Standard errors in parentheses; * significant at 10%, ** significant at 5%, *** significant at 1%. The statistical significance of the inverse mills ratio indicates the existence of the sample selection bias, albeit the fact that the selection-bias corrected coefficients are not substantially different from the basic OLS coefficients. The positive IMR coefficient indicates that the nonmigrant households are above average in their unobservable 18

characteristics with regard to the migrant households in the sample. Most of the statistically significant results are consistent with the economic reality in rural China. More laborers lead to more household income, while the middle-aged laborers (often the household heads) tend to contribute the most to the household income. The number of young dependents has negatively affects the household income. More arable lands are indicators of higher yields from home production, which naturally results in higher household income. Job training and marital status show no significant impacts on household income in this sample. The selection-bias-corrected coefficients from Equation (1.3) can now be applied to predict the potential total household income for the 928 migrant households had they received some home production earnings instead of remittances. For the second counterfactual group in which all nonmigrant households are assumed to be migrant households, the estimation results for the first-stage migration participation equation are exactly the same as the results from Equation (1.2) in magnitudes, but with opposite signs, because the only change is the dependent dummy variable, which now equals 1 if household participated in internal migration. Analogously, the inverse mills ratio is obtained and added into the second-stage household income regression as an independent variable. Table 1.4 presents the selected estimation results of Equation (1.5) and Equation (1.6) for the 928 observed migrant households, with and without the IMR. 19

Table 1.4 Selected Results of the Migrant Household Income Model OLS without IMR OLS with IMR Equation (1.5) Equation (1.6) Laborers of Age 16 to 30 0.09 *** (0.02) 0.11 *** (0.02) Laborers of Age 31 to 50 0.11 *** (0.03) 0.15 *** (0.03) Laborers of Age above 50 0.11 *** (0.03) 0.14 *** (0.03) Members with Schooling above 9 0.19 *** (0.04) 0.28 *** (0.03) Members with Job Training 0.03 (0.05) -0.09 (0.31) Marriage (=1 if married) -0.09 (0.19) 0.05 (0.09) Number of Children under 6 0.003 (0.14) -0.13 *** (0.03) Arable Land Area (m 2 ) 0.0002 *** (0.00004) 0.0003 *** (0.00004) Year (=1 if surveyed in 2004) -0.24 ** (0.14) -0.27 * (0.19) Inverse Mills Ratio N/A N/A -0.42 *** (0.09) Constant 10.45 *** (0.31) 11.68 *** (0.46) Observation 928 928 Adjusted R 2 0.32 0.32 Note: (1) Other control variables include interactions of year dummy with education and training, (2) Standard errors in parentheses; * significant at 10%, ** significant at 5%, *** significant at 1%. Sample selection bias is similarly obvious for migrant households given the statistical significance of the inverse mills ratio. By comparing the magnitude of the coefficients in Table 1.3 and Table 1.4, it is clear that the effect of the number of laborers on household income is largely the same for both migrant and nonmigrant households. Same argument applies for the effect of arable land areas on household income. This confirms that migrant households are not less productive than the nonmigrant households on the margin or facing a structurally different production function. For them, the most 20

plausible reason of a lower household income is they could not make the surplus laborers more productive by allocating them to local jobs. The selection-bias-corrected coefficients for the migrant household income equation can now be applied to predict the potential household income for nonmigrant households had they received remittances at the expense of less home production earnings. Table 1.5 reports the household income for the actual sample and the simulated household income for the two counterfactual samples. 8 Table 1.5 Average Household Income in Three Scenarios Total Nonmigrant Migrant Actual Sample 9648.75 10646.29 8284.62 Counterfactual 1 10228.31 10646.29 9655.37 Counterfactual 2 8931.42 9403.31 8284.62 The reported numbers suggest that migrant households could earn 16.5% more on average by switching to local jobs. In contrast, the nonmigrant household income would drop by 11.7% if they chose to participate in internal migration. Overall, the entire region is better off with higher total household income if counterfactual (1) is realized. It should 8 The simulations are based on the assumption that migrant households would have the same marginal productivity in home production as nonmigrant households, and vice versa. This assumption is arguable and the simulated income of nonmigrant households could be underestimated, because in the sample, they are better endowed and more likely to be well-informed and risk-resilient, which could make them earn more from migration than the actual migrant households. Symmetrically, the simulated income of migrant households could be overestimated. 21

be noted that the reported results ignore the human capital accumulation of the migrant workers and the long-run effect that migration might have on the development of the source areas. It has been documented that migrants may bring back to the home country increased skills and knowledge that could only be picked up in cities but are transferable to the home environment (Stark et al. 1997). But it is more likely that it will be high-skill individuals working in creative and dynamic sectors of the economy that will contribute, upon return, to the development of the home areas (Rosenzweig, 2005). Hence, the results should still be reliable given the fact that most of the migrant workers in northwestern China are low-skilled individuals. We next present evidence that shows the switching from internal migration to home production is not only individually optimal, but also socially desirable in terms of income distribution. 1.6 Inequality Analysis and Policy Implications Income inequality is studied by decomposing the Gini coefficient into three easilyinterpretable terms, following the seminal works by Lerman(1985) and Stark(1986). The differences between these inequality indicators of the observed group and those of the two counterfactual groups can be used to illustrate the impact of different job choices on rural income distribution, while also provide evidences for its underlying reasons. We write the Gini coefficient for rural household income as a function of the covariance between income and its cumulative distribution, which is: 22

(1.7) where is the Gini coefficient of total household incomes for all the 2200 surveyed households. is the cumulative distribution of total incomes, and denotes the average income for 2200 households. Utilizing the properties of the covariance, Equation (1.7) can be written as G K k 1 K = k 1 S k G k R k (1.8) Where S k is the share of income component k in total household income, i.e. S k =y k /y; G k is the Gini index corresponding to income component k; and R k is the Gini correlation of component k with total income. Equation (1.8) enables us to decompose the role of remittances and home production earnings in inequality into three terms: (a) S k : the magnitude of remittances/home production earnings relative to total income; (b) G k : the inequality of remittances/home production earnings, and (c) R k : the correlation of remittances/home production earnings with total income. Table 1.6 shows the results of Gini decomposition for the three groups. The first panel of Table 1.6 shows the result of the traditional Stark (1985) approach where we can obtain a measure of the overall impact of remittances upon village income inequality. Total Gini coefficient is at the level of 0.33. The remittances are distributed quite unevenly across migrant households, with a Gini coefficient of 0.74. Referring to the third panel, the simulated remittances are even more unequally distributed, with a Gini 23

coefficient of 0.9, and the total income distribution with Gini coefficient of 0.4 is significantly less desirable than the actual income distribution with a Gini coefficient of 0.33. Table 1.6 Decomposition of the Rural Household Income Inequality Observed Group Share in Gini for Gini correlation Contribution to total income income component with total income Gini coefficient (S) (G) (R) (SGR) Remittances 0.12 0.74 0.02 0.04 Home Production Earnings 0.79 0.41 0.90 0.28 Other Income 0.09 0.41 0.42 0.01 Total Income 1.00 0.33 1.00 0.33 Counterfactual Group 1 Remittances N/A N/A N/A N/A Home Production Earnings 0.91 0.34 0.90 0.28 Other Income 0.09 0.41 0.49 0.03 Total Income 1.00 0.31 1.00 0.31 Counterfactual Group 2 Remittances 0.72 0.90 0.32 0.21 Home Production Earnings 0.19 0.51 0.88 0.17 Other Income 0.09 0.41 0.45 0.02 Total Income 1.00 0.40 1.00 0.40 Note: Other income is calculated by subtracting remittances and home production earnings from total household income. 24

With a dominating share of 72% in the total income, remittances now contribute 0.21, which is over 50%, to the total income inequality. The second panel shows that without the income component of remittances, the Gini coefficient would drop to a more desirable level of 0.31. A 0.02 decrease in the Gini might not be of strong economical importance, but an increase of the Gini from 0.33 to 0.4 can surely be devastating. The results confirm that by switching from migration to home production, the entire region would be better off (or at least remain stable) in terms of income distribution, and more importantly, we should expect to see a sizable increase in the income inequality if more households choose to participate in rural-urban migration. Home production earnings are highly correlated with the total household income in all the three scenarios, with the Gini correlation around 0.9, indicating that these earning opportunities are not well-diffused across the entire region and are only taken advantage of by those at the upper end of income distribution. Internal migration, instead, is widely accessible in the studied region with a Gini correlation of 0.02 in the observed sample. This further suggests that households are taking migration as an alternative when local jobs are unavailable. Given the above findings, two types of policies should be designed to increase the total average income and to decrease the income inequality in the studied area. The first is to aim at achieving the first-best outcome by encouraging households to participate in home production. However, this outcome can only be obtained when rural laborers are allowed to switch freely between home production and migration. Since the agricultural land endowments of rural households are fixed in the short run due to institutional 25

reasons, as discussed above, many rural laborers cannot participate in agricultural production due to the lack of land. Moreover, the rural low-skill service industries have very limited capacity to absorb surplus laborers. Thus, surplus laborers of migrant households can hardly switch to home production due to the binding constraint of limited agricultural resources and local job opportunities. This constraint also tends to result in a low marginal return to labor input on home production for migrant households, by the law of diminishing marginal returns. Policies of this type should focus on creating equal opportunities for rural households to participate in home production, mainly by developing township and village enterprises (TVEs). TVEs have been proved to be efficient in absorbing rural surplus laborers since most of them are very labor intensive (Naughton, 2007). TVEs often cluster to focus on the entire processing industry for agricultural products in the target area, which can add sizable extra values to agricultural products. Rural household can increase their incomes by selling fresh agricultural products directly to the TVEs at the contracted prices, which also reduce the market risk they may encounter. 9 The second set of policies, which is more enforceable in the short run, should focus on increasing remittances for the migrant households, mainly by increasing migrant workers nominal earnings and reducing their migration costs. There are two general ways to increase migrant worker s earnings in China: increasing their wage rate and preventing them from being cheated by the employer. The latter case is by no means a 9 The Chinese central government issued the Decision on accelerating the development of the TVEs in middle and western region in 1993 to officially confirm the benefits and encourage the development of the TVEs. See http://www.mofcom.gov.cn/article/b/bf/200207/20020700031377.shtml for the complete policy. 26

trivial one in reality. Every year, large number of migrant workers suffers from delayed payments and even no payments due to the dishonesty of their employers and little legal protections they can rely on. Most of the migrant workers have no knowledge on law, so it is the government s responsibility to guarantee a legal employment contract to be signed before any de facto employment takes place. Employers should be legally convicted and punished for their dishonest and unlawful behaviors. The government is also responsible for reducing the costs of migration, chiefly by reducing the living cost and the transportation cost. In China, the largest part of the living cost for migrant workers comes from the housing rents in urban areas. In the metropolitan areas such as Beijing, Shanghai and Guangzhou, the monthly rent of a common apartment is very likely to exceed a migrant worker s monthly earnings, resulting in the formation of urban slums with poor and unhygienic living conditions for the occupying migrant workers. Thus, government should invest more in constructing the economically affordable house which is specifically designed for poor urban households and migrant workers. Migrant workers should also receive rent subsidies and transportation subsidies as received by less-endowed urban residents. 10 It is important to note that these two types of policies are not mutually exclusive. They can function together to provide both the equality of opportunity and the equality of outcome for current migrant households in the surveyed area. 10 The suggested policies are concerned by the three-nong problem since 2001 and are included explicitly in the 2004-2009 No.1 Central Document issued annually by the Chinese central government. 27

1.7 Conclusion Internal migration in China after the economic reform has played a critical role in increasing household income level and changing income distribution in rural areas. Remittances are the most direct pecuniary rewards from migration, and serve as an important source of income for migrant households. However, there is no consensus on the impact of internal migration, since nonmigrant households and migrant households have very different initial endowments and face different job choice constraints in different areas, while the remittances and home production earnings they receive varies significantly across different regions. This is why most of the studies on this topic are place-based studies. Using data from two large and undeveloped provinces in northwestern China, this empirical study finds that migrant households in this area are initially less endowed, and the remittances they received by participating in migration are lower than potential home production income they could have earned if they switched to home production. Thus, rural-urban migration is not the optimal decision for migrant households if they can switch costlessly to local jobs. The decomposition of the Gini coefficient also shows that encouraging internal migration is also not the first-best choice for a social planner if she aims to achieve a higher average income and lower income inequality in the studied region. Policies should be implemented to increase migrant worker s remittances by increasing their nominal wages and reducing their migration costs, and protect their rights to be fully and punctually paid. More importantly, government should spend more 28

budgets on protecting existing arable lands and encouraging the development of the TVEs to absorb the surplus laborers more adequately in the studied area. 29

Chapter 2 What Drives the Urban Wage Premium? Evidences from a Female Panel 2.1 Introduction U.S workers earn higher hourly wages in urban areas than in rural areas. 11 The positive relationship between city size (often measured by population) and wages has received ample discussion since Marshall (1890) and Weber (1899). Glaeser and Maré (2001) report the urban wage premium to be 18% to 34% for male workers, based on four different datasets. The simple model in Roback(1982) s seminal paper suggests regional real wage differences can be explained largely by differences in local amenities. Variation in housing prices and income should equalize firm profits across regions while also ensuring that workers are equally well off everywhere. However, this equilibrium state has never been reached in reality, and this general framework does not impede the analyses of the driving forces of the nominal urban wage premium. Certain firms can benefit substantially from locating in cities through the agglomeration effect even if the cities are high-wage areas. 12 11 In the NLSY79 data, "urban area", "city" and "metropolitan statistical area" are used interchangeably. Also, "non-urban area","rural" and " non-metropolitan statistical area are used interchangeably. Urban area is defined as central core or city and its adjacent, closely settled territory which has a combined total population of 50,000 or more. 12 In this study we are interested in firm agglomeration measured by nominal wages as that is what affects firm competitiveness for traded goods. Thus we ignore the household cost-of-living that only affects real wage and focus on firm effects on nominal costs. 30

Analysis of wage determination starts logically from labor supply and labor demand. On the labor supply side, one of the suggested driving forces of the urban wage premium is that wages are higher because more productive workers choose to live and work in cities. Combes et al (2003) provide strong evidence on sorting by skills using a large panel of French workers. Their findings suggest that worker's individual ability accounts for a large fraction of the existing urban wage premium in France. This sorting argument naturally raises another critical question, why cities are more preferable for skilled workers, if they are paid the same wage based on their ability wherever they work? Glaeser and Maré (2001) argue speculatively that urban density creates more opportunities for interactions between peers, which are more important to highly educated individuals. Also, cities are centers of consumption which are more valuable for high-skilled workers. Thus high ability workers have potential incentives to sort themselves into cities, and models controlling for measured abilities or unmeasured skills should reveal these incentives. Another driving force of the urban wage premium could be that workers become more productive in cities through interaction and human capital accumulation. Alfred Marshall (1890) and Jacobs (1968) suggest that cities are "intellectual furnaces" where knowledge and skill are flowing between individuals as if they are traded mysteriously in the air. Urban density can speed up the rate of interactions between low-skilled and high-skilled workers and increase the probability that low-skilled workers learn from their high-skilled co-workers. Cities can also allow tacit information to be shared 31

between participants in business meetings and informal contacts (McCann, 2013). 13 While cities provide a superior environment for interaction and learning, workers will choose to locate in cities under the expectation that this learning process will generate higher economic returns over time. Urban wage premium as the result of sorting by observed/unobserved abilities implies a wage level effect. According to this hypothesis, explicitly controlling for individual fixed abilities should eliminate most of the urban wage premium. Furthermore, workers who move to cities should immediately receive a nontrivial wage gain and workers who leave cities should immediately experience a sizable wage loss. On the other hand, urban wage premium as the result of interaction/learning process implies a wage growth effect that suggests the existence of the agglomeration effect. For those who move to cities, wages should grow smoothly over time, and for those who leave cities, wages should not decline instantly. Moreover, there should not be a considerable change in the urban wage premium after controlling for fixed abilities. In the existing male-sample studies, wage level effect and wage growth effect are constantly found to be functioning simultaneously, each explaining part of the existing urban wage premium. On the labor demand side, if wages are higher in urban areas, why firms choose to locate in cities instead of fleeing high wage areas? The prevailing explanation is that firms enjoy lower production costs in urban areas. The reduction of production costs comes not only from better labor market matching in cities (Wheaton and Lewis, 2001), where firms (especially high-tech firms) have a much lower cost to select specialized 13 These "tacit information" are knowledge that are partially shared on a non-market basis, including information on new products, personnel, technology and market trends. 32

workers from a large pool of applicants, but also from reduced transportation costs, technology spillovers (Marshall-Arrow-Romer externality), input sharing and proximity to consumers (Matano and Naticchioni, 2012). It is not this paper's major task to test for these possibilities, but any remaining urban wage premium unexplained by the laborsupply analyses should be reasonably attributed to these alternative explanations on the labor demand side. While existing literature has provided abundant evidences on the driving forces of the urban wage premium for male workers, there has been little consideration on whether these factors affect female workers differentially, and what the underlying implications are. This paper contributes to current literature by first providing evidences on the existence and magnitude of the urban wage premium for female workers, a cohort that has been intentionally circumvented by previous studies (Yankow, 2006). This study also captures possible gender differences on this topic by comparing the female-sample results with the male-sample results, using the most updated dataset. Basic findings suggest that learning in cities and sorting by observed education or ability explains only a marginal part of the urban wage premium, while sorting by unmeasured personal characteristics drives the bulk of the urban wage premium. Agglomeration effect only weakly affects women's wage growth in cities. The rest of the paper is organized as follows. Section 2.2 introduces the data and shows the urban wage premium for female workers, and addresses some sample selection issues for using the female sample. Section 2.3 presents a simple urban learning model to provide one theoretical framework for the learning hypothesis. Section 2.4 shows 33

descriptive statistics, explains the empirical strategy and ensuing results. Section 2.5 concludes. 2.2 Data and Sampling Issues The data used in this study comes from the National Longitudinal Survey of Youth 1979 (NLSY79). This survey covers a nationally representative sample of 12686 young men and women who were 14 to 22 years old when first surveyed in 1979. This paper uses a panel of female surveyed biennially from 1994 to 2010, which forms a longitudinal dataset with observations over 9 years. This dataset is ideal for both the wage level and the wage growth analyses. It uniquely provides the Armed Forces Qualification Test (AFQT) score, which are widely accepted in the existing literature as a standard measurement for individual ability. It also appears to be a favorable dataset for tracking workers' locational choices and wage change patterns over time, given the maturity of the sample (age 29-37 in 1994, age 45-53 in 2010). 14 We exclude observations who are separated, divorced or widowed, and those who are married but with spouse absent. 15 We omit any observations with hourly wage greater than 200 dollars or less than 2 dollars. Part-time workers who usually worked less than 35 hours per week are excluded, but we keep individuals with zero work hours. 16 In spirit of Juhn and Murphy (1992), nonworkers receive predicted wages based on the regression 14 Although an apparent urban wage premium is also observed using the NLSY97 data (2000-2010 panel), the sample is too young to provide any interesting results for the wage growth analysis. 15 We do not observe any urban wage premia for a sample composed of these cohorts. 16 We omitted the part-time workers because they are more likely to have changed jobs in a certain year than full-time workers (Timothy and Wheaton, 2001). We keep nonworkers to avoid the sample selection problem and make the sample more representative of the entire female labor force. 34

using the under-5-weeks-per-year sample. This imputation is reasonable giving the fact that these two samples are similar in terms of age, education, AFQT score, as well as number of children (see Appendix B), suggesting that those with short work weeks may be an appropriate base for which to estimate the wages of nonworkers. 17 Figure2.1 and Figure2.2 show the existence of an evident urban wage premium for female workers, while Figure2.2 provides a more detailed picture by further dividing urban areas into central city areas and non-central city areas 18 16 14 12 10 8 6 Urban Non-Urban 4 2 0 1994 1996 1998 2000 2002 2004 2006 2008 2010 Figure 2.1 Urban Wage Premium for Female Workers 17 The self-selection of women into the labor force can be quite different between rural and urban area. While women in rural areas have lower educational attainment on average, they have lower labor force participation rate and are more likely to be self-selected into the labor market, and are more likely to be self-selected into urban areas under the expectation of higher wages. This could cause the agglomeration effect to be overestimated for rural-urban movers. 35

18 16 14 12 10 8 6 4 SMSA- CC SMSA- NCC Non- SMSA 2 0 1994 1996 1998 2000 2002 2004 2006 2008 2010 Figure2.2 Urban Wage Premium for Female Workers For a simple rural-urban divide, the first figure shows that a constant and stable wage gap persists between non-urban and urban female workers, at the average level of 17%. The second figure identifies a 19.5% wage premium for non-central city workers and a 12% wage premium for central city workers. This pattern is consistent with the observed reality that high-income households tend to reside in the suburbs of U.S. cities, and poorer households tend to live near the center. 18 Based on a transport mode-choice perspective, Glaeser, Kahn and Rappaport (2008) argue that the poor relies more on public transit, and that central cities are the only parts of urban areas that are dense 18 The NLSY79 dataset does not provide information on place of work but rather location of residence. This is one limitation that has been clearly stated in many previous studies. 36

enough to support convenient public transit. Brueckner and Rosenthal (2009), in contrast, show that high-income households prefer new housing, whereas the poor tolerate old housing. Since a city's newest dwellings are often built in the suburbs, far from the CBD, the rich are therefore attracted to suburban locations, whereas the poor occupy older dwellings near the center. 2.3 A Simple Model of Learning in Cities While the process of ability sorting is more self-evident, the hypothesis of learning is vaguer and requires some conceptual formalization.we now present a simple model that depicts the learning process that purports individuals acquire skills by interacting with and studying from one another, while dense urban areas themselves increase the speed and efficiency of interactions. This model modifies Glaeser (1999)'s original model by explicitly showing that low-skilled workers can benefit from both localization and urbanization economies through the learning process. Workers are assumed to live for two periods and choose the location of their workplace (urban or rural) in each period. Localization economy and urbanization economy can both facilitate learning and wage growth. Within industries, low-skilled individuals can learn from high-skilled individuals through meeting and communicating. 19 Let s ek denotes the share of the high-skilled workers in industry e in city k 20. Meeting or communicating does not necessarily lead to learning, so a probability c ek (a i, a j ) 19 Although we allow skill levels to vary continuously within low-skilled group and high-skilled group, learning is assumed not happening within these two groups in this model. We also ignore the minor possibility that high-skilled workers learn from low-skilled workers. 20 We drop the assumption made by Glaeser (1999) that skilled and unskilled individuals are spread equally across all industries, which is unconvincingly strong. 37

is assigned to denote the likelihood that a low-skilled individual i learn some knowledge from a high-skilled individual j in one meeting in industry e in city k, a i and a j are the skill levels (individual ability) of the student and the teacher, respectively, which are strictly positive, and a j > a i. c ek is increasing in both a i, and a j, which allows different learning or teaching effects for individuals with different skill levels. Number of meetings per period is assumed to be equal and denoted as M(L ek ), which is assumed to be increasing in the scale of localization economy L ek, measured by the fraction of national employment in industry e in a certain city (Wheaton and Lewis, 2001). In each meeting, for a low-skilled worker, the probability of actual learning from a high-skilled individual is sc(a i, a j ), which is the probability of meeting a high-skilled worker times the probability of learning from her. Given the number of meetings, the probability of learning (at least once) in one period is 1-(1- sc(a i, a j )) M(L). Urban density per se can offer a unique environment for informal learning from social and business contacts. So we consider this urbanization-induced learning separately. Let h k denote the share of the high-skilled workers in city k. By the same argument above, we assign a probability d k (a i, a -j ) to denote the likelihood that a low-skilled individual i learn some knowledge from a high-skilled individual that does not work in his industry. a i and a -j are the knowledge levels of the student and the teacher, respectively, which are strictly positive, and a -j > a i. d k is increasing in both a i, and a -j. Number of meetings per period is assumed to be equal and denoted as N(U k ), which is assumed to be increasing in the scale of urbanization economy U k, measured by 38

population per square kilometer (Or diversification of economic activities, if available). In each meeting, for a low-skilled worker, the probability of actual learning from a highskilled individual is hd(a i, a -j ), which is the probability of meeting a high-skilled worker times the probability of learning. By the same argument, the probability of learning (at least once) from a high-skilled person in one period is 1-(1- hd(a i, a -j )) N(U). The benefit of learning and becoming more skilled is denoted as V. Individuals discount the future with discount factor β, and have a survival probability of δ. Students must pay their teachers part of the benefits by accepting lower wages to compensate skilled workers for their input, if they work in the same industry. Thus, the expected present value of the second period benefits for individual i to work in a city and learn can be written as βδ[φv l (1-(1- sc) M(L) ) + V u (1-(1- hd) N(U) )], where V l and V u are benefits from localization and urbanization, respectively. Assuming that the net cost of living (primarily the difference in housing price) in city k is A k, and the cost of commuting to the workplace is G k, increasing on the distance between home and workplace F. Firms in industry e in city k offer wage w ekr in rural area, and wage w eku in urban area. For workers to be indifferent between these two locations, utility levels must be equalized between rural and urban area. Therefore: w u + βδ[φv l (1-(1- sc(a i, a j )) M(L) ) + V u (1-(1- hd(a i, a -j )) N(U) )] -A-G(F) = w r (2.1) B i = w u + βδ[φv l (1-(1- sc(a i, a j )) M(L) ) + V u (1-(1- hd(a i, a -j )) N(U) )] -A-G(F) - w r = 0 (2.2) 39

Differentiating Equation (2.2) with respect to s and h yields, B' s = βδφv l M(L)c(1- sc) M(L)-1 > 0 (2.3) B' h = βδv u N(U)d(1- hd) N(U) -1 > 0 (2.4) Equation (2.3) and (2.4) show that the net benefits of being in the city for the unskilled rise with the fraction of the industry and city that is skilled, this implies that workers can benefit more from both the localization-induced learning process and the urbanizationinduced learning process. Young workers may benefit particularly through these learning processes, since they are more likely to value human capital accumulation than elder workers (Glaeser, 1999). Differentiating Equation (2.2) with respect to a i, a j and a -j also yields positive results, recall that c is increasing in skill levels. This justifies that we should control for individual ability while inspecting the wage growth effect. 2.4 Testing for Sorting Hypothesis Table 2.1 provides some descriptive information of the total sample and two subsamples by location of residence. Urban workers are not observed to be significantly more able than rural workers, given their average AFQT scores. Also, these two groups only differ slightly in their educational attainments, which are both between the 12 th grade and the first year in college. Urban workers are elder and 21% more experienced than rural workers. The number of biological children is taken as a control for both groups, which allows women to substitute between childbearing/childrearing and labor 40

supply. Difference in spouse annual income shows a pattern that is consistent with the assortative mating theory, which claims that female with higher earnings tend to choose her spouse who has a commensurate or higher earning ability. While the presented correlations provide rudimentary evidences on the explanatory power of these independent variables, they are by no means causal. Table 2.1 Selected Descriptive Statistics Total Urban Rural Difference Hourly Wage 13.43 (7.22) 13.63 (8.69) 11.64 (10.44) 1.99 *** AFQT Score 42.42 (28.86) 42.43 (28.86) 42.32 (28.86) 0.11 Age 40.38 (5.56) 40.77 (5.47) 36.65 (4.92) 4.12 *** Nonblack 0.81 (0.39) 0.80 (0.39) 0.85 (0.36) 0.50 *** Education 12.47 (1.96) 12.49 (2.01) 12.28 (1.27) 0.21 *** Work Experience 21.90 (5.61) 22.28 (5.56) 18.38 (4.86) 3.90 *** Marital Status 0.87 (0.53) 0.86 (0.53) 0.89 (0.50) 0.03 ** Bio-Children 1.98 (1.42) 1.99 (1.43) 1.95 (1.38) 0.56 *** Spouse Income 36972 (58762) 39381 (62546) 23202 (24810) 16179 *** Number of Obs 29167 26363 2804 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% We now set up models to test the existence of sorting by ability and the magnitude it contributes to the urban wage premium. We first estimate a model that explicitly control for individual ability using each person s AFQT score. The model takes the form: 41

log(w it ) = X it β + L itδ +θlog(afqt i ) + υ it (2.5) W it is the hourly wage for individual i at time t, X it is the vector of control variables, including human capital, demographic and family background information. L it includes one dummy that equals unity if the individual lives in the central-city metropolitan area, and another dummy that equals unity if the individual lives in the non-central-city metropolitan area. Living in the rural area is set to be the base group. Thus vector δ contains the regression coefficients of interest that capture the urban wage premia. AFQT i stands for individual s time-invariant Armed Forces Qualification Test score, as an explicit control for individual ability. To account for unobserved time-invariant individual characteristics, We also estimate a fixed-effects model: log(w it ) = X it β + L itδ + φ i + u it (2.6) The AFQT score is now one covariate included in X it. φ i is the individual fixedeffects parameter to be estimated. Recall that the two coefficients in δ should decrease nontrivially if sorting by ability matters. Table 2.2 reports the estimation results from five specifications. The basic model specification is gradually augmented with more control variables. Column (1) pertains to a basic specification that includes only indicators of the location of residency, age, race and year dummies. The point estimate for central city is 42

measured at 0.114, suggesting that workers in central cities receive a 11.4% wage premium over workers residing in rural areas. Workers in suburbs receive 18.2% wage premium over their rural counterparts, which is consistent with the fact that the richest households in U.S. live in suburbs rather than in city centers. This naive specification, however, explains only 19% of the wage variation. Column (2) augments the baseline model by incorporating observable measures of educational attainment and individual ability. As predicted by the ability-sorting hypothesis, the urban coefficients fall, but only slightly. The elasticity of wage with respect to education is much larger than the elasticity of wage with respect to personal ability. Column (3) presents the results from the full specification that adds controls for marital status, training history, work experience, family background, preschool education, number of biological children and spouse income. The non-central city coefficient falls by only 0.002, and interestingly, the central city coefficient rises marginally from 0.106 to 0.120. Recall that the identified urban wage premium should fall substantially if measured ability is truly affecting the locational choice of workers. The results clearly show that for female workers, sorting by observed human capital or ability is not driving the bulk of the existing urban wage premia. This finding is consistent with the previous male-sample studies. There seems to be a strong pure marriage wage premium for females, which could be explained by unobserved marriage-related preferences such as positive or negative assortative mating, instead of hypothetical industrial concentration or urbanization. Note that a worker's unobserved skill or ability can change over time through the learning process, thus a 43

wage growth effect can partly be captured by the work experience coefficient, which is 0.032 and significant at 1% level. Table 2.2 OLS and Fixed-Effects Results OLS(1) OLS(2) OLS(3) FE(4) FE(5) MSA-Non-Central City 0.182 *** 0.169 *** 0.167 *** 0.045 *** 0.043 *** (0.017) (0.021) (0.013) (0.021) (0.021) MSA-Central City 0.114 *** 0.106 *** 0.120 *** 0.047 *** 0.048 *** (0.029) (0.021) (0.013) (0.014) (0.016) Age 0.052 *** 0.050 *** 0.050 *** (0.014) (0.014) (0.013) Nonblack 0.112 *** 0.091 *** 0.072 *** (0.014) (0.018) (0.013) AFQT Score 0.003 ** 0.003 ** (0.017) (0.017) Education 0.074 *** 0.042 *** (0.004) (0.004) Work Experience 0.032 *** (0.004) Married 0.152 *** 0.081 *** (0.009) (0.011) Bio-children -0.045 * 0.015 (0.004) (0.006) Log-Spouse Income -0.06-0.03 (0.011) (0.011) Industry No No Yes No Yes Occupation No No Yes No Yes Adjusted R 2 0.19 0.21 0.37 0.14 0.18 Number of Obs 29167 29167 29167 29167 29167 Note: (1) Other control variables in Column (3) and Column (5) include father's education, mother's education, number of siblings siblings' education, occupational training dummy, preschool or head start attendance, and interaction terms between age and education. Year dummies are included in all specifications. (2) Standard errors in parentheses. *** Significant at 1%, ** Significant at 5% * Significant at 10%. 44

OLS estimates can be severely biased if a worker's attitude, skills or ability are not adequately measured by the control variables but are correlated with urban residence. Column (4) and (5) presents fixed-effects estimates of the urban wage premium. Analogous to the OLS specifications, column (5) is an augmentation of the basic specification in column (4) by adding a full set of control variables. Accounting for the individual fixed-effects significantly reduces the observed urban wage premia. Noncentral city workers now earn only 4.3% more than rural workers, and central city workers earn 4.8% more compare to their rural counterparts. Time invariant unobserved heterogeneities may include motivation, interests, attitude, people skills, persistence, communication skills and other "soft" skills that are difficult to be objectively measured. 2.5 Testing for Learning Hypothesis We next proceed to test the learning hypothesis which may also help to explain the identified urban wage premia. As noted above, if the agglomeration economy and the wage growth effect exist, workers who come to cities should see their wages gradually increase over time, rather than immediately reach a high wage level. By the same argument, workers who leave the urban area should not immediately receive a wage loss, but see their wages decrease smoothly over time. We adopt the identification strategy proposed by Glaeser and Maré(2001) to study the wage change patterns of rural-urban and urban-rural movers, which are the only cohorts that provide a clear and tractable wage change pattern after a locational change. The mover's model takes the form 45

Δlog(W kt ) = ΔX kt β + ΔU kt δ + υ kt (2.7) This model is created simply by differencing the wage level equation (2.5) across time periods and letting Δ denotes changes between survey years. Vector ΔU now contains three dummy variables designating stayers in urban areas, movers who move into urban areas and movers who move out of urban areas, with the base group consisted of stayers in rural areas. Coefficients in vector δ now capture the return to migration and residency, which exactly provide the needed wage change patterns for wage growth analysis. The dependent variable is the difference in reported log wages across survey periods. In order to track individual's wage growth route, we examine the differences in wages over three time periods after the rural-urban/urban-rural migration occur, which are two-year time span, four-year time span and six-year time span. Table 2.3 presents the results from this mover's model. 21 The estimates of 0.019, 0.021 and 0.027 for the "Enter City" variable show a clear pattern that female workers experience a small but robust post-rural-urban migration wage growth, compare to the workers who stay in rural areas, after controlling for observed and unobserved confounders. However, results given by the estimates for the "Exit City" variable do not reveal a wage reduction pattern that is symmetric to the wage gain pattern, while this symmetry is consistently found in male-sample studies. Female workers receive a wage increase of 0.8% in the first two years after moving out of urban areas compare to rural stayers, and this trivial wage premium converges toward zero in the next four years. The underlying implication is that workers' 21 The results of a pooled sample should be included in future research to account for potential attrition bias. 46

movements are not always random, if they only move out of the cities if they are expecting a wage gain (or at least not a wage loss), this endogenous choice of location might bias the estimates by overstating agglomeration effect generated by city per se. 22 Table 2.3 Results of the Mover's Model Two Years Four Years Six Years Enter City 0.019 ** 0.021 ** 0.027 ** (0.064) (0.033) (0.035) Exit City 0.008 * 0.004 * 0.003 (0.063) (0.023) (0.071) City Stayer 0.021 ** 0.025 ** 0.025 ** (0.001) (0.020) (0.043) Adjusted R 2 0.133 0.131 0.130 Number of Obs 29167 29167 29167 Note: (1) All control variables introduced in the OLS analyses are included. (2) Standard errors in parentheses. *** Significant at 1%, ** Significant at 5% * Significant at 10%. City stayers receive relative stable wage premium over rural stayers over the six-year time span, at the level of 21% to 25%. This pattern shows that in addition to knowledge spillovers, cities can make workers more productive through other channels. In a city where many different industries are presented, a specialized worker should be easier to find a job that fits best with her skills. When a city's employment in a particular industry is large, the existence of a large labor pool makes the workers easier to be replaced. As a result, firms can fire unproductive workers at a low cost of replacement, which 22 Glaeser and Maré(2001) use urbanization of parents' state of birth to instrument urban residency. However this variable fails the weak instrument test. For a structural approach that embeds the choice of residency within a model of career choices over time, see Gould (2007). 47

encourages workers to achieve higher productivity than in an environment in which shirking or "free-riding" on the job is harder to punish with dismissal (Brueckner, 2011). These are examples of the "labor market matching" hypothesis which complements the "learning" hypothesis in explaining the urban wage premium in classic studies. From a psychological perspective, workers in cities are more likely to socialize with other workers from the same industry; they may try to achieve high productivity simply to "look good" in the eyes of a broad social set. 2.6 Gender Difference Analysis In this section we make comparisons between the male-sample findings drawn from two representative studies and the female-sample findings from this study. The first study by Glaeser and Maré (2001) uses four different datasets to test for the sorting and learning hypotheses. Their study provides the most comprehensive evidence on this topic to date. The second study by Yankow (2006) extends Glaeser and Maré(2001)'s study by conducting year-to-year wage growth analysis and between-job wage growth analysis. We also present the male-sample results from the exact same dataset used for female analyses in this study, which is a 9-year panel data with workers surveyed biennially from 1994 to 2010, and compare them with the aforementioned female-sample results. These three male-sample analyses all use the NLSY79 dataset and are very similar in the adopted methodology in terms of identification strategy, model specification and control variables, which make their findings highly comparable. Table 2.4 presents the summarized results from these three studies. 48

For the first two panels in Table 2.4, the first row shows the change of the magnitude of the urban wage premium for central cities (dense metropolitan areas) before and after the basic OLS (explicitly controlling for the AFQT score) or fixed-effects regression, and the second row presents this change for non-central cities (nondense metropolitan areas). For example, in the Glaeser and Maré(2001) study, the identified 24.9% urban wage premium drops slightly to 24.3% for central city workers, after controlling for the AFQT score. In contrast, by running individual fixed-effects estimation, the 24.9% urban wage premium drops significantly to 10.9% for central city workers. The finding that unmeasured ability/skills, rather than the observed ability, are driving the bulk of the urban wage premium for male workers is true for both central city and non-central city workers, and is consistent across all three studies. The rationale behind a similar wage level effect and sorting behavior for male and female workers is that women's labor supply behaviors are becoming more like men's (Kahn and Blau, 2007). With rising labor force participation rates, fewer and fewer women are on the extensive margin. Moreover, increasing divorce rates and increasing career orientation are also making women's labor supply less sensitive to their own wages or their spouses' wages (Goldin, 1990). The third panel of Table 2.4 shows the results from examining male movers' wage change patterns. The first row shows the wage change for rural-urban movers across three time periods, and the second row gives this change for urban-rural movers. A relatively strong wage growth for rural-urban movers is revealed in all three studies, especially in the Yankow (2006) study, which finds the wage premium grow from 3.9% to 8.8% 49

during the first year of entering the urban area, and rapidly to 17% after four years of entering the city. Yankow (2006) and this study also find "symmetrical" wage change patterns for urban-rural movers, with their wage premium gradually decrease after leaving cities. 23 These findings further confirm the existence of the agglomeration effect, as predicted by the wage growth hypothesis. Table 2.4 Summary of the Male-Sample Results Glaeser and Maré Yankow This Study Basic OLS 0.249 to 0.243 0.220 to 0.175 0.261 to 0.224 0.153 to 0.141 0.095 to 0.076 0.183 to 0.154 Fixed-Effects 0.249 to 0.109 0.220 to 0.054 0.261 to 0.722 0.153 to 0.07 0.095 to 0.038 0.183 to 0.536 Wage Growth 0.079 to 0.111 to 0.125 0.039 to 0.088 to 0.170 0.044 to 0.102 to 0.115 0.005 to 0.028 to 0.006-0.108 to -0.178 to -0.122 0.003 to -0.014 to -0.022 There are a number of reasons why gender difference exists in terms of their wage change patterns after a change of location. Although women are participating more in the labor market, they still work fewer hours and years in their career lives, with higher job turnovers (Keane, 2011). These career characteristics for female can negatively affect 23 As shown in Table 2.4, Yankow (2006) finds a strong wage decline pattern, with the premium coefficients as low as -17.8%. We find a relatively weak wage decline pattern with the coefficients very close to zero. 50

their learning processes, if there is any, by preventing or interrupting them, thus reducing the potential wage growth in cities. Phimister (2004) suggests that interruptions in women s working history indicates that labor market matching benefits women more than man, which is questionable since we do not observe the counterfactual of how much can men benefit if they share the same job turnover pattern with women. Empirical evidence also suggests that women are less mobile than men since the responsibilities of women as wives and mothers (and the role of men as breadwinners) were thought to influence the decisions of women. These gendered responsibilities are believed to explain why women were less likely than men to participate in migration decisions (Boyd, 2003), and why their urban wage premium are not as evident as their male counterparts. 2.7 Conclusion This study focuses on answering "why do cities pay more" by exploring the sources of the urban wage premium realized by female workers. The results are most likely to be mixed, indicating that both a wage level effect (sorting by unobserved abilities) and a wage growth effect (agglomeration effect) appear to be operating. Controlling for unmeasured fixed personal characteristics eliminates about 70% of the identified urban wage premium, which confirms that cities are attracting workers of higher unmeasured ability or workers with better noncognitive skills that are more advantageously valued on the job market. Wage growth analysis reveals a time-dependent wage change pattern for rural-urban movers, but not for urban-rural movers, which provides weak supportive evidence for the existence of the learning process for female workers. Cities can make 51

workers more productive through other channels such as labor market matching or peer pressure. Wage change pattern of the city stayers can probably be attributed to these hypotheses. Wage level effects are similar for males and females. However, male workers have clearer wage growth/decline patterns after they move into or out of the cities, suggesting that agglomeration effect benefits men more than women. This difference can be explained by some female-specific labor supply behaviors, especially on the intensive margin. Further research should explore what fixed confounders are omitted in the linear regression, how these attributes contribute to the urban wage premium and why they are advantageously rewarded in cities. The female sample can be further divided, for example, by income levels or occupation to study the driving forces of the urban wage premium separately for blue-collar workers and white-collar workers. Finally, new methodologies should be adopted to solve the endogeneity problem that arises when movers nonrandomly choose their residencies and workplaces. 52

Chapter 3 Spatial Disequilibrium or Agglomeration? New Evidence on Intra-Urban Wage Gradient 3.1 Introduction Within many metropolitan areas, wages are observed to be the highest at the central urban area while wages paid at non-central dispersed worksites fall with their distance to the central business district (CBD), a phenomenon known as the (negative) intra-urban wage gradient. 24 The Alonso-Muth-Mills model of urban land use predicts that intraurban wage gradient should exist in monocentric cities with decentralized employment. A dual prediction is that wages paid at noncentral worksites should be positively correlated with worker s home-workplace commuting time, assuming that a worker has shorter commuting time if his/her workplace location is farther from the CBD. The rationale behind is that workers who are employed by the noncentral firms save commuting costs (either monetary cost or time cost, or both) by stopping at a closer job site. Cost-minimizing firms would cut the wage by exactly the amount of commuting cost saved by the workers, making the workers indifferent between noncentral worksites and 24 Metropolitan areas are defined as counties or combinations of counties centering on a substantial urban area. A metro area includes not only the urban area, but also surrounding suburban areas. Hence, the term "intra-urban" wage gradient in this study should be interpreted more broadly as "intra-metro" wage gradient. 53

central ones. With this capitalization of commuting costs into wages, it is possible for workers to achieve spatially uniform utility across work zones. Firms, nevertheless, are paying higher wages to observationally equivalent workers at locations closer to the CBD, which indicate that firms at different sites are bearing different amount of labor costs. If firms are indeed in a spatial equilibrium and are receiving uniform profits everywhere, they must somehow be able to generate more revenue or reduce other costs merely by locating closer to the central urban areas, with positive agglomeration economies in the CBD being one cause. Alternatively, firms' locational choices may not be in a long-run spatial equilibrium and we should see high wage firms moving to places that are easier for workers to access, where they can pay less to compensate for workers' commuting costs. Intra-urban wage gradient would gradually disappear in this case. Intra-urban wage gradient is economically interesting for three reasons. First, the existence of the intra-urban wage gradient in monocentric cities is a confirmation of the predictions made by the standard urban land use model. Second, the identification of the causal relationship between commuting costs and wages can provide important implications for the locational choice and commuting patterns of urban firms and workers. Third, a city with intra-urban wage gradient provides a natural laboratory to test for the existence and magnitude of the agglomeration effect, which is typically manifested by higher firm productivity and higher wage rates at the central city areas. There are three major contributions of this study. First, this paper empirically investigates if the observed intra-urban wage gradient in the U.S. cities is temporary and 54

vanishing over time, or if it is sustained by the economies of agglomeration for the past few decades. To achieve this goal, we compare our findings with the results from some previous influential studies, to incorporate a time dimension. Second, regarding the key explanatory variable, this is the first study to use the actual reported commuting time for individual workers as a measure of their home-workplace commuting costs, which generates the least measurement error among existing literature. 25 Third, using the most updated census data, this paper is the first one to look at the New York MSA, which is undoubtedly an economically important metro area and one that truly does fit the monocentric model (Glaeser and Kahn, 2001). The rest of the paper is structured as follows. Section 3.2 provides a brief review of the existing literature. Section 3.3 introduces a simple conceptual framework of the intraurban wage gradient based on the standard urban land use model. Section 3.4 presents the data, and shows the existence of the intra-urban wage gradient in the three studied metropolitan areas. Section 3.5 details the estimation strategy and examines the causal relationship between commuting costs and wages. Section 3.6 proceeds to test for the economies of agglomeration in urban wage determination. Section 3.7 concludes. 3.2 Previous Empirical Works Mose (1962) proposed the first theoretical framework for the analyses of intra-urban wage differentials. Since then, the lack of data on work location, commuting costs and 25 Madden (1985) used the actual reported commuting distance. However, travel time is considered to be a better measure for commuting cost than distance because of the increasing congestion in metropolitan areas. Studies have shown workers are indeed minimizing their commuting time instead of distance with the presence of congestion (Gordon, et al, 1989). 55

worker characteristics has hindered adequate testing of the driving forces of intra-urban wage gradient. Eberts (1981) performed the first empirical investigation on the existence and magnitude of the intra-urban wage gradient, using aggregate wage and distance data from five groups of municipal public employees in the Chicago SMSA. He confirmed the casual relationship between commuting distance and wages for four of the five public labor groups in Chicago. Although Eberts claimed that conclusions from the public sector can be extended to the private sector, Ihlanfeldt (1992) pointed out that the differences in wage determination, unionization pattern, laws on collective bargaining and dispute resolution between the public and private sectors can cause wage gradient to differ substantially. Another limitation of Eberts study is that his data does not allow him to control for any worker productivity or human capital variables, which may induce bias for the estimates. Using the PSID data from 1971 to 1977, Madden (1985) identified a wage increase for movers whose job change resulted in an increase in the distance of commute. However, she was unable to overcome the selection problem that workers change jobs only if they are expecting a higher wage and such a change would make them better-off. Ihlanfeldt (1992) did a thorough investigation on the intra-urban wage gradient using 1980 Census data for the Philadelphia, Detroit, and Boston metropolitan areas. This study divides the sample by race, gender and sector for the first time and confirms the existence of a negative intra-urban wage gradient for white workers. However, it does not 56

find that a positive wage gradient exists for black workers, as theory predicts. 26 A major limitation of the study is the use of airline distance as the measure of home-workplace distance, which creates an unknown amount of measurement error that biases the results. McMillen and Singell (1992) also used 1980 Census data but included more metropolitan areas such as Columbus, Cleveland, Pittsburgh and Indianapolis. They adopted a new approach to derive a worker s optimal work and residence location by predicting the probability that workers live/work in the suburbs using a probit model. Their analyses, however, are restricted to working white males. Timothy and Wheaton (2001) used 1990 Census data for the Minneapolis-St.Paul MSA and the Boston MSA and concluded that the intra-urban wage gradient reflects the capitalization of commuting costs into wages. Moreover, they found that the agglomeration effect does not seem to explain away the negative correlation between wages and commuting costs. Their estimates can still be contaminated by the use of average zonal commuting time to proxy for actual individual commuting time, due to data limitation. Another problem is that while the Minneapolis-St.Paul MSA has its central urban area composed of two relatively independent cities and surrounded by a number of identifiable subcenters, it is more likely to be polycentric rather than monocentric, and the identified relationship between commuting times and wages can be 26 As White (1976, 1978, 1988) has noted, if racial segregation is sufficiently strong, a suburban firm may find it difficult to draw enough laborers form the local labor pool. At the current wage, the firm faces a excess labor demand and has to compensate black workers from central city areas for their cost of outcommuting. 57

spurious due to unobserved factors that lead to wasteful commuting between city centers (Hamilton, 1982). 27 3.3 Conceptual Framework In a standard Alonso-Muth-Mills urban model, a monocentric city with decentralized employment is assumed to be located on a flat featureless plain, with no zoning policies or other land use regulations that might set certain land use rules for certain regions. Individual workers are assumed to be identical with equal productivity, and receive market wages. Workers are assumed to live further from the CBD than their jobs and in the same direction away from the CBD as with firms, i.e. firms and their workers are along the same ray from the CBD. This assumption eliminates the possibilities of reverse commuting and circumferential commuting from the residential location to the workplace location, which are not common commuting modes within a monocentric city (White, 1988). Firms and workers are both assumed to have full mobility. Formally, let y be the prevailing wage offer at the CBD, let x be the distance of a worker s residential location to the CBD, and x * be the distance of a non-cbd firm to the CBD. The worker s net (disposable) income would be y-tx if she chooses to work at the CBD, where t represents the per mile cost of commuting. The noncentral employer could attract a worker from the CBD by providing a wage at least equal to y-tx *. This wage 27 Hamilton (1982) showed that the volume of wasteful commuting is large in the U.S. polycentric cities, while workers choose not to work at the closest urban center. It is probably due to workers' idiosyncratic preferences for particular neighborhoods, or a desire to be close to the spouse's jobsite, or employment uncertainty. None of these factors is recorded by the census data. Hence we avoid studying polycentric cities and focus on monocentric cities with dispersed employment. 58

would give the worker a disposable income of y-tx * -t(x-x * ), which equals to y-tx. Note that the worker saves tx * by working at the noncentral job site instead of working at the CBD, and a cost-minimizing employer would set the wage rate at exactly y-tx *, making the worker indifferent between the two worksites. Two fundamental conditions are required for spatial equilibrium within the urban area (Brueckner, 2011). Consumers/workers must be equally well of at all locations, achieving the same utility regardless of their residential choices and workplace choices. Firms reach spatial equilibrium when each firm is earning the same (normal) profit across different locations. In the case of intra-urban wage gradient, workers receive the same disposable income at different worksites, and meet the equilibrium condition of uniform utility. However, the difference of the wage paid at different worksites implies that spatial equilibrium condition for firms might not be reached since firms locating closer to the CBD are paying higher wages and thus are bearing higher labor costs. As a consequence, we should see firms relocating to areas where commuting is less costly and wages are lower. In this process, one would expect to see the convergence of commuting times and wages, which leads to a vanishing intra-urban wage gradient over time. Alternatively, if firms are indeed in a stable spatial equilibrium, then it must be true that firms can operate more productively and generate more revenue by locating closer to the CBD. The most plausible explanation for this location-productivity relationship is the existence of agglomeration effects, the benefits that come when firms and people locate near one another together in cities and industrial clusters (Glaeser, 2010). 59

3.4 Data Description and Intra-Urban Wage Variation The data used in this study comes from the 5% Public Use Microdata Sample (PUMS) of the 2000 U.S. Census. 28 Three metropolitan areas, the New York MSA, the Boston MSA and the Chicago MSA, were selected for the reason that they have relatively compact central city jurisdictions and a sufficient number of identifiable suburban work zones. 29 One particular advantage of the census data is the identification of residence based on RESPUMA (Residence Public Use Microdata Area) and workplace locations based on PWPUMA (Place of Work Public Use Microdata Area). These areas are Census Bureau-defined areas of contiguous territory containing at least 100,000 residents/workers, and are state-specific. This study uses RESPUMA to control for workers' residence zones, but uses a larger microdata area, the Super-PWPUMA, to capture the intra-urban wage variation. This allows us to avoid identifying and plotting hundreds of PWPUMAs redundantly, while the demonstration of the existence and magnitude of the intra-urban wage gradient is not compromised. Each Super-PWPUMA contains at least 400,000 workers. The sample is restricted to individuals who reported working at least 35 hours per week during the previous year, since full-time workers are less likely to have changed jobs than part-time workers and are expected to receive a consistent hourly wage throughout the previous year. The sample also excludes the workers who reported 28 The American Community Survey (ACS) data is not appropriate for this study since it reports the place of work but does not report the actual travel time to work. Moreover, we use the 2000 Census data to make our results comparable with previous studies using the 1980 and 1990 Census data. 29 Among the three MSAs, New York is the most standard monocentric MSA and Chicago has the most decentralized employment (Glaeser and Kahn, 2001). All these three MSAs have some polycentricity, but they are still monocentric MSAs based on PWPUMAs. 60

working less than 35 weeks during the previous year, for the same reason. The selfemployed, those who worked for non-profit organizations and those who belongs to the army forces are omitted, due to the fact that commuting costs are less likely to be capitalized into wages for these individuals. The sample is further restricted to white and black workers, to focus the analyses on white-black differentials and the advance of black economic status, which are major topics in the literature (Heckman, Lyons and Todd, 2000). The sample keeps only those who are married with spouse present and single individuals. Outliers with reported hourly wage less than 2 dollars and more than 200 dollars are dropped. Table 3.1 reports the intra-urban wage variation for the three MSAs, divided by sector. As aforementioned, public sector workers have their own wage determination processes that do not resemble that of private-sector workers. Each Super-PWPUMA is assigned a serial number, with its corresponding geographical areas listed in Appendix C.1. and mapped in Appendix C.2. Table 3.1 shows that zonal wage differences are much more apparent within the private sector than the differences for the public sector. For the private sector, the wage reaches the highest level at the central urban areas for all the three MSAs. The highest observed central-city wage premia in the Boston MSA and the Chicago MSA are 29.5% and 26.8% respectively, which are much higher than the wage premia of 15% and 18% found one decade ago (Timothy and Wheaton, 2001). The New York MSA shows even larger wage gaps across different work zones, with the maximum wage premium of 71% 61

found in Manhattan. This can be easily explained by the concentration of the most highpaid jobs in that area, including management, legal, financial, and technical occupations. Table 3.1 Intra-Urban Wage Variation Boston New York Chicago Super-PWPUMA Private Public Private Public Private Public 1 24.01 21.54 26.54 26.18 19.28 19.97 2 20.08 20.48 27.65 27.66 19.41 18.79 3 23.37 23.28 19.88 23.34 23.35 21.89 4 23.78 22.21 33.98 24.66 23.72 20.45 5 26.01 22.96 21.63 23.68 21.41 21.48 6 23.52 21.45 20.56 23.12 24.45 22.12 7 20.52 21.02 21.51 23.95 8 20.90 20.78 Obs 35719 6757 55916 15566 63789 10620 Note: (1) Values in bold are wage rates of the central urban areas for the three MSAs. (2) Wage differences are significantly different from zero at 5% level. Within the public sector, the highest wage is observed at the central urban area for the Chicago MSA, but not for the other two MSAs. The sample for the Chicago MSA is then divided to account for disparities in the wage determination for different categories of public workers. Federal government salaries are set nationally, with some inter-city adjustment to account for cost-of-living differences. Thus, wages for these workers are not expected to change systematically within metro areas. State government salaries are unclear, they can be different between cities, but can also vary within metro areas. Local government employees have a wage determination mechanism that is closer to private 62

workers. Since local government agencies/departments hire most of the workers from the local labor pool, they must act more like private firms to compensate their workers for longer commuting. Table 3.2 reports the wage variation for these three types of public sector workers in the Chicago MSA. Table 3.2 Intra-Urban Wage Variation for Public Sector Workers in the Chicago MSA Super-PWPUMA Federal Gov State Gov Local Gov 1 26.51 18.09 18.97 2 18.28 19.45 18.56 3 23.30 20.79 21.69 4 19.39 20.80 20.77 5 20.46 21.17 21.83 6 23.02 21.40 22.06 Obs 2038 2011 6535 Note: (1) Values in bold are wage rates of the central urban areas for the Chicago MSA. (2) Wage differences are significantly different from zero at 10% level. It is illustrated that in the Chicago MSA, both state government employees and local government employees, who account for 85% of the public employees, receive the highest hourly wage at the central urban area. Similar pattern, however, is not observed for the other two MSAs. For the Boston MSA, the highest public-sector wages are paid at the central urban area only for state government employees. For the New York MSA, none of these three types of workers has their wages peak at the Manhattan borough. 63

Table 3.3 Intra-Urban Commuting Time Variation (in minutes) Boston New York Chicago Super-PWPUMA Private Public Private Public Private Public 1 31.54 26.64 41.31 29.61 27.21 22.38 2 24.75 19.64 42.21 28.79 27.65 22.15 3 33.10 25.23 38.77 41.29 32.23 27.40 4 33.86 27.52 55.95 55.28 31.26 22.55 5 41.48 39.34 31.68 40.31 32.71 27.67 6 33.22 27.47 31.86 42.93 41.10 38.83 7 30.14 24.66 33.71 31.58 8 25.36 22.68 Obs 35719 6757 55916 15566 63789 10620 Note: (1) Values in bold are commuting times of the central urban areas for the three MSAs. (2) Commuting time differences are significantly different from zero at 10% level. 50 Chicago MSA 50 Boston MSA 40 40 30 30 20 20 10 10 0 0 5 10 15 20 25 30 0 0 5 10 15 20 25 30 60 50 40 30 20 10 0 New York MSA 0 5 10 15 20 25 30 35 40 Figure 3.1 Intra-Urban Wage Gradient in the Boston, New York and Chicago MSA for Private Workers (home-workplace commuting time on vertical axis and wage on horizontal axis) 64

As shown in Table 3.3 and Figure 3.1 that plots the positive correlation between commuting time and hourly wage. For the three MSAs and for both the private sector and the public sector, average commuting times of central-area workers are nearly double those of workers in the edge zones. This confirms the existence of the intra-urban wage gradient within the private sector in the studied MSAs. The findings suggest that the intra-urban wage gradient for private sector workers in major cities is by no means weakening or vanishing. The positive correlation between commuting costs and wages persists since Ebert's first empirical findings and as predicted by the standard urban land use model. Since we do not find that the variance in commuting costs and wages across work zones are converging over time, it is safe to claim that the spatial disequilibrium of employment is not sustaining the intra-urban wage gradient. We next investigate the driving forces of the existing intra-urban wage gradient to better understand what makes firms to choose different locales within cities. 3.5 Estimation Strategies and Results We follow the convention of using a log-linear wage equation to estimate the causal relationship between commuting time and wage. The model takes the form: log(w i ) = βc i + X i δ + υ i (3.1) W i is the hourly wage calculated by dividing total wage and salary income by total hours worked in 1999. C i denotes the reported one-way commuting time, in minutes, that it usually took the respondent to get from home to work. X i is the vector of a set of 65

standard control variables including age, age-squared, education, education-squared, race, gender, marital, disability and veteran status, occupation and industry, and workers' residential places based on RESPUMA. The interaction between age and education and the interaction between marital status and gender are also included. Table 3.4 presents the results for private and public sector workers. The coefficients represent the semi-elasticity of the hourly wage with respect to one additional minute of commuting time. For the private sector workers in the Boston MSA, one more minute of commuting time is associated with a 0.05% increase in hourly wage. Thus, an average worker who randomly switches to a new job with the one-way travel time thirty minutes longer would be expected to receive a 3% increase in her hourly wage. This result is close to what Timothy and Wheaton (2001) reported using the 1990 census sample, which is 3.9%. Furthermore, we should note that the wage variation should compensate workers for the entire commuting cost, which includes not only the time cost, but also the direct money cost from gas, maintenance and depreciation (Brueckner, 2011). Small and Verhoef (2007) concluded that the direct money expenses are at least the magnitude of the time costs. Thus, private sector workers in the Boston MSA are expected to have at least a 6% increase in her hourly wage for working at a new site that takes 30 more minutes to reach. This wage premium is 10.8% for the New York MSA and 8.4% for the Chicago MSA. 30 30 This estimation for money cost applies only for those who use private vehicles to get to work. For those who use public transit, ride bicycles or walk only, little money cost is associated with longer travel distance or time. 66

Table 3.4 Travel Time Coefficients for Public and Private Sector Workers Boston New York Chicago Private Public Private Public Private Public 0.0005 *** 0.0003 0.0009 *** 0.0002 0.0007 *** 0.0006 *** (0.0001) (0.0002) (0.0001) (0.0003) (0.0001) (0.0001) Adjusted R 2 0.38 0.28 0.37 0.27 0.40 0.34 Obs 35710 6757 55916 15566 63789 10620 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% For public sector workers, a statistically significant result is found only for the Chicago MSA. Overall, the causal relationship between commuting time and wage rates is much weaker for public sector workers, a finding that is compatible with the fact that the intra-urban wage gradient is less obvious for this cohort. Racial inequality is often measured by observed wage differences in the civilian labor force (Western and Pettit, 2005).We now divide the private-sector worker sample by race to account for potential white-black differentials. Table 3.5 shows a clear result that only the wages of white workers are statistically responding to longer commuting times. No statistically significant result is found for black workers, which indicates that the wages of black workers, whose residences are usually concentrate in central urban areas, are not associated with their commuting times. One reason may be that new suburban jobs are difficult for black workers to obtain since they may have more difficulties in affording vehicles and bearing monetary commuting costs. Ihlanfeldt (1992) found exactly the same result and argued that the black workers cannot find job within the central city at a 67

competitive wage, so suburban employers do not have to pay them a compensating differential to cover their commuting costs. The finding suggests the existence of spatial mismatch, a phenomenon that employment opportunities for low-income people are located far away from the areas where they live. The suburbanization of low-skilled jobs, combined with the continued suburban housing market exclusion of low-income workers, has apparently made central-locating black workers more disadvantaged. Table 3.5 Travel Time Coefficients for Private Sector Workers by Race Boston New York Chicago Black White Black White Black White 0.0003 0.0005 *** -0.0001 0.0009 *** 0.0001 0.0008 *** (0.0004) (0.0001) (0.0001) (0.0001) (0.0002) (0.0001) Adjusted R 2 0.29 0.38 0.28 0.34 0.33 0.39 Obs 1631 34079 10615 45301 7222 56567 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% The subsample of private-sector white workers is further divided by gender, to examine the potential gender differences in responding to the change of commuting cost. Table 3.6 reports that in the Boston MSA and the Chicago MSA, women s wages show a much stronger response to commuting times because women value their time lost to commuting more than men do, perhaps due to a greater share in household responsibility. 68

Table 3.6 Travel Time Coefficients for Private Sector White Workers by Gender Boston New York Chicago Male Female Male Female Male Female 0.0002 ** 0.0009 *** 0.0008 *** 0.0006 *** 0.0004 *** 0.0012 *** (0.0001) (0.00003) (0.0001) (0.0001) (0.0001) (0.0001) Adjusted R 2 0.36 0.36 0.36 0.33 0.37 0.34 Obs 21778 12301 29474 15827 36973 19594 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% Today, U.S. women still have a higher wage elasticity of labor supply than men (McClelland and Mok, 2012), and should expect higher wages to compensate for the increase in their commuting costs, compared to their male counterparts. Interestingly, this result is not found in the New York MSA, which suggests labor supply behaviors for females are closer to male workers in this area. 3.6 Testing for Agglomeration Effects After accounting for the classic determinants of urban wages, there still remains a sizable portion of unexplained intra-urban wage variation, which might be explained by the equilibrium distribution of agglomeration advantages. In this case, firms and workers are more productive in dense urban areas than elsewhere. This agglomeration effect has long been documented and quantified by studying spatial patterns in wages and land rents. If firms and workers are mobile and wages and land rents differ across space, higher 69

wages and land rents in denser urban environments must reflect some productive advantage (Puga, 2009). Agglomeration economies are commonly decomposed into urbanization economies (Jacobs externality) and localization economies (MAR externality) (Rosenthal, 2004). We first test for the existence of urbanization economies by examining the link between total employment and wages, and also by inspecting the relationship between employment concentration and wages. To measure employment concentration, an import ratio, which is defined by the ratio of workers employed in a PWPUMA to workers residing there (see Appendix C.4 for the import ratio of each SPWPUMA), is constructed for each work zone. Zonal employment and import ratio are then included in the wage equation to examine its impact on wages. Table 3.7 reports the results. Table 3.7 Total Employment and Import Ratio (IR) Coefficients for Private-Sector White Workers Boston New York Chicago Log( Emp) Log(IR) Log( Emp) Log(IR) Log( Emp) Log(IR) 0.151 *** 0.114 *** 0.194 *** 0.235 *** 0.135 *** 0.167 *** (0.02) (0.01) (0.01) (0.01) (0.01) (0.02) Adjusted R 2 0.36 0.36 0.37 0.36 0.38 0.37 Obs 34079 34079 45301 45301 56567 56567 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% 70

Unlike the results reported by Timothy and Wheaton (2001), strong and significant urbanization effect is found for all three MSAs. Within the New York metropolitan area, the elasticity of wages respect to total employment size (and thus urban density) is 0.194, and the elasticity of wages respect to employment concentration is 0.235. The effect of urbanization is relatively smaller for the other two MSAs. Testing for the localization effect requires a variable on the spatial concentration of the same industries. We follow Henderson (1995) to construct a specialization index for each worker as a direct measure of the zonal industrial concentration. The specialization index of a certain PWPUMA is defined as the percentage of PWPUMA employment in the worker s specific industry relative to the percentage of MSA employment in that same industry. For example, suppose the construction share of total employment in the New York MSA is 10%, a PWPUMA with 20% of its workforce in construction would have a specialization index of 2. Table 3.8 reports the results when the log of specialization indices are included in the wage equations. Table 3.8 Specialization Index Coefficients for Private Sector White Workers Boston New York Chicago 0.104 *** 0.162 *** 0.131 *** (0.04) (0.03) (0.03) Adjusted R 2 0.36 0.36 0.35 Obs 34079 45301 56567 Note: Standard errors in parentheses. *** Significant at 1%, ** Significant at 5%, * Significant at 10% 71

It is shown that the localization effect measured by industrial specialization also explains part of the intra-urban wage variations. While the results are statistically significant for all the three MSAs, New York has the strongest localization effect. A doubling of the industrial concentration in the New York MSA is associated with a 16.2% wage increase, lower than the nationwide average number provided by Wheaton and Lewis (2002) using the 1990 census data, which is 27.8%. A key concern when interpreting the above estimates is that the innate ability of workers may vary across work zones. If more able workers sort into more central areas, the intra-urban wage gradient may simply reflect the differences in abilities instead of any locational advantages. Glaeser and Maré(2001) included the AFQT score in the wage regression and argued that the observed ability does not seem to explain away the urban wage premium. However, subsequent fixed-effects estimation eliminates a sizable portion of the urban wage premium, which suggests that workers might be sorting based on unobserved abilities such as attitude, motivation and people skills. While it is quite difficult to fully overcome this ability-sorting problem, it is reasonable to deduce that these ability-related unobserved factors may also explain part of the intra-urban wage variation. 31 31 Ideal instrument variables for urban residency should be those that predict urban status and are orthogonal to unobserved ability (Glaeser and Maré, 2001).Currently we know of no such variables. For a structural approach by estimating a dynamic programming model, see Gould (2007). 72

3.7 Conclusion Intra-urban wage gradient is one of the central predictions of the standard urban land use model. Within a monocentric city with dispersed employment, longer commuting journeys are either compensated by lower housing prices or by higher wages. While the urban rent gradient has long been a familiar analytical tool in urban economics, much less attention has been given to the intra-urban wage variation and its driving forces. This study shows that for the three large metropolitan areas in the U.S., significant intra-urban wage gradient exists for private sector workers, and this wage gradient has been persisting for decades since it was first empirically studied in the 1980s. The consistent existence of the intra-urban wage gradient rejects the hypothesis that urban firms are not in a spatial equilibrium of uniform profit and would eventually move to lower wage areas. The causal relationship between commuting costs and wages is further identified for private sector white workers, which confirms the capitalization of commuting costs into wages predicted by the Alonso-Muth-Mills model. Among white workers, women s wages are more responsive to commuting costs, which is most likely to be the result of the higher labor supply elasticity with respect to hourly wage. However, as documented by Heim (2007), women s labor supply elasticities are shrinking substantially since 1978, both on the extensive margin and on the intensive margin. It is a reasonable expectation that women s labor supply behaviors will become more and more similar to men s, just like what we find in the New York MSA. For the black workers, it seems that many are trapped in the central urban areas and the suburbanization of new 73

jobs is not making them better-off by attracting them to new worksites or compensating them with higher wages. One of the plausible sustaining powers of the intra-urban wage gradient is the agglomeration effect, which can be further divided into the localization effect and the urbanization effect. By constructing measures for employment concentration and industrial specialization, this study finds that both the localization effect and the urbanization effect are operating in the determination of intra-urban wage differences. An interesting observation is that from 1990 to 2000, urbanization effect seems to have become stronger, while the localization effect seems to have turned weaker. Many aspects of intra-urban wage gradient have yet to be investigated. Due to data limitation, we cannot control for workers innate abilities, which may also help to explain the studied wage gradient, as they explain the urban wage premium. Also, as aforementioned, relatively high-skilled white workers opt to live in the suburban area, certain firms must be strongly affected by agglomeration to stay at central areas, because they could have relocated to be near white residents and save on paying commuting premium. Sample should be split by industries in future studies to investigate how does this white flight affect the locational choices of firms in different industries. It is also the task of future research to examine the wage gradients in other types of cities to ascertain more general results of the intra-urban spatial equilibrium. For instance, it might be less costly for black workers to gain access to new jobs in medium cities like Columbus or Cleveland. 74

Bibliography [1] Alonso, W., 1964. "Location and Land Use." Cambridge: Harvard University Press. [2] Angrist, J.D., and J.S. Pischke., 2009. "Mostly Harmless Econometrics." Princeton: Princeton University Press. [3] Adams, R., 1989. Worker Remittances and Inequality in Rural Egypt. Economic Development and Cultural Change 38(1): 45-71. [4] Adams, R., 1994. Non-farm Income and Inequality in Rural Pakistan: A Decomposition Analysis. Journal of Development Studies 31(1): 110-133. [5] Barham, B., and S. Boucher.,1998. Migration, Remittances, and Inequality: Estimating the Effects of Migration on Income Distribution. Journal of Development Economics 55(2): 307-331. [6] Beaudry, C., and A. Schiffauerova., 2009. Who s Right, Marshall or Jacobs? The Localization versus Urbanization Debate. Research Policy 38:318-337. [7] Brueckner, J.K., 2011. "Lectures on Urban Economics." Cambridge: The MIT Press. [8] Brueckner, J.K., and S.S.Rosenthal., 2009. "Gentrification and Neighborhood Housing Cycles: Will America's Future Downtowns Be Rich?" Review of Economics and Statistics 91: 725-743. [9] Brueckner, J.K., and A.G.Largey., 2008. "Social Interaction and Urban Sprawl." Journal of Urban Economics 65:18-34. [10] Brauw, A., and J. Giles., 2005. Migrant Opportunity and the Educational Attainment of Youth in Rural China. World Bank Policy Research Working Paper, No.WPS4526. [11] Bound. J., D. Jaeger and R. Baker., 1995. "Problems with Instrumental Variables Estimation When the Correlation Between the Instruments and the Endogenous Explanatory Variable is Weak." Journal of the American Statistical Association 90(430): 443-450. 75

[12] Card, D, and J. DiNarto., 2002. Skill-Biased Technological Change and Rising Wage Inequality: Some Problems and Puzzles Journal of Labor Economics 20(4): 733-83. [13] Chinhui, J., and K, Murphy., 1997. Wage Inequality and Family Labor Supply. Journal of Labor Economics, 15(1):72-97. [14] Ciccone, A and R. Hall.,1996. Productivity and the Density of Economic Activity. American Economic Review 86(1): 54-70 [15] Combes, P., G.Duranton,. L.Gobillon., and S. Roux. 2012. Sorting and Local Wage and Skill Distributions in France. Regional Science and Urban Economics, 42:913-30. [16] Combes, P., G.Duranton,.and L.Gobillon. 2008. Spatial Wage Disparities: Sorting Matters! Journal of Urban Economics, 63(2):723-42. [17] Du,Y., P.Albert, and S.Wang. 2005. Migration and Rural Poverty in China. Journal of Comparative Economics 33(4): 688-709. [18] Davin, D., 1999. Internal Migration in Contemporary China. ST. Martin s Press, New York. [19] Eberts, R.W., 1981. "An Empirical Investigation of Urban Wage Gradients." Journal of Urban Economics 10: 50-60. [20] Edward, T., 1999. The New Economics of Labor Migration and the Role of Remittances. International Migration 37(1): 63-88. [21] Glaeser, E., 1999. Learning in Cities. Journal of Urban Economics 46: 254-77. [22] Glaeser, E.,and D. Mare., 2001. Cities and Skills. Journal of Labor Economics 19(2): 316-342. [23] Glaeser, E., M.E.Kahn., and J. Rappaport., 2008. "Why Do the Poor Live in Cities?" Journal of Urban Economics 63: 1-24. [24] Glaeser, E.,and M. Resseger., 2010. The Complementarity between Cities and Skills. Journal of Regional Science 50(1): 221-44. [25] Glaeser, E., and M. Kahn., 2001. Decentralized Employment and the Transformation of American Cities. NBER Working Paper No.8117. 76

[26] Glaeser, E. 2007., The Economics Approach to Cities. NBER Working Paper No.13696. [27] Gordon, P., A. Kumar and H. Richardson., 1989. "The Influence of Metropolitan Spatial Structure on Commuting Time." Journal of Urban Economics 26: 138-151. [28] Gould, E., 2007. "Cities, Workers, and Wages: A Structural Analysis of the Urban Wage Premium." Review of Economic Studies 74(2):477 506. [29] Goldin, C., 1990. "Understanding the Gender Gap." New York: Oxford University Press. [30] Hamilton, W., and A. Roell. 1982. Wasteful Commuting. Journal of Political Economy 90(5): 1035-1053. [31] Heckman, J., 1979. Sample Selection Bias as a Specification Error. Econometrica 47(1): 153-161. [32] Heckman, J., T. Lyons and P. Todd. 2000. Understanding Black-White Wage Differentials, 1960-1990. American Economic Review 90(2): 344-349. [33] Henderson, V., 2003. Marshall s Economies. Journal of Urban Economics 19(1): 47-70. [34] Henderson, V., A. Kuncoro and M. Turner. 1995. Industrial Development in Cities. Journal of Political Economy 105(51): 1067-1090. [35] Heim, B., 2007. The Incredible Shrinking Elasticities: Married Female Labor Supply, 1978-2002. Journal of Human Resource 42(4): 881-918. [36] Kahn, L., and F. Blau, 2007. Changes in the Labor Supply Behavior of Married Women:1980-2000. Journal of Labor Economics 25(3): 393-436. [37] Inlanfeldt, R., 1992. "Intraurban Wage Gradients: Evidence by Race, Gender, Occupation Class, and Sector." Journal of Urban Economics 32: 70-91. [38] Kimmel, J., and T. Kniesner. 1998. New Evidence on Labor Supply: Employment versus Hours Elasticities by Sex and Marital Status. Journal of Monetary Economics 42: 289-301. [39] Lorene,Y., 1976. Internal Migration and Economic Development in Brazil. Quarterly Journal of Economics 76 (2): 218-245. 77

[40] Lincoln, A., 2008. Gender, Productivity and the Marital Wage Premium Journal of Marriage and Family 70: 806-14. [41] Li, Q., 2011. Urban Migrant Workers and Social Stratification in China. Beijing: Social Sciences Academic Press. [42] Matano, A., and P.Naticchioni 2012. What Drives the Urban Wage Premium? Evidence along the Wage Distribution. Working Paper, Research Institute of Applied Economics, Barcelona. [43] Mose, L.N., 1962. "Towards a Theory of Intra-Urban Wage Differentials and Their Influence on Travel Patterns."Paper and Proceedings of the Regional Science Association 9: 53-63. [44] McMillen, D.P., and L.D. Singell., 1992. "Work Location, Residence Location, and the Urban Wage Gradient." Journal of Urban Economics 32: 195-213. [45] McClleland, R., and S. Mok. 2012., A Review of Recent Research on Labor Supply Elasticities. Congressional Budget Office Working Paper No.2012-12. [46] Madden, J.F. 1985., "Urban Wage Gradients: Empirical Evidence." Journal of Urban Economics 18: 291-301. [47] Mulalic, I., J. Van Ommeren, and N. Pilegaard. 2013. "Wages and Commuting: Quasi-natural Experiments' Evidence from Firms that Relocate." The Economic Journal DOI: 10.1111/ecoj.12074. [48] Puga, D., 2009. The Magnitude and Causes of Agglomeration Economies. Paper Written for the Journal of Regional Science s 50 th Anniversary Symposium. [49] Phimister, E 2004. Urban Effects on Participation and Wages: Are there Gender Differences? Discussion Paper 2004-04, Centre for European Labour Market Research, Aberdeen. [50] Rauch, E. 1993. Productivity Gains from Geographic Concentration of Human Capital: Evidence from the Cities. Journal of Urban Economics 34(3): 380-400. [51] Roback. J. 1982. Wage, Rents and the Quality of Life. Journal of Political Economy 90(6): 1257-78. [52] Rosenzweig. M., 2005. " Consequences of Migration for Developing Countries." Paper of the United Nations Expert Group Meeting on International Migration and Development. UN/POP/MIG/2005/08 78

[53] Rosenthal, S., 2004. "Evidence on the Nature and Sources of Agglomeration Economies." Handbook of Regional and Urban Economics 4:2119-2171 [54] Stark,O., T.Edward., and Y.Shlomo., 1986. Remittances and Inequality. The Economic Journal 96(383): 722-740. [55] Stark,O., T.Edward., and Y.Shlomo., 1988. Migration Remittances and Inequality: A Sensitive Analysis Using the Extended Gini Index. Journal of Development Economics 28(3): 309-322. [56] Stark, O., C. Helmenstein and A. Prskawetz., 1997. A Brain Gain with a Brain Drain." Economics Letters 55: 227-34. [57] Small, K.A., and E. Verhoef., 2007. "The Economics of Urban Transportation." New York: Routledge. [58] Snow, N.B. and NBER., 2014. "Urban Transportation Expansion, Employment Decentralization, and the Spatial Scope of Agglomeration Economies." Working Paper. [59] Tumbe, C., 2011. Remittances in India: Facts and Issues. Indian Institute of Management Bangalore Working Paper, No. 331. [60] Wheaton,W., and M. Lewis., 2002. Urban Wages and Labor Market Agglomeration. Journal of Urban Economics 51: 542-62. [61] Wheeler, H., 2001. "Searching, Sorting and Urban Agglomeration." Journal of Labor Economics 19(4): 879-99. [62] Wheaton, W., 2001. "Intra-Urban Wage Variation, Employment Location, and Commuting Times." Journal of Urban Economics 50: 338-366. [63] White, J., 1976. "Firm Suburbanization and Urban Subcenters." Journal of Urban Economics 3: 232-243. [64] White, J., 1978. "Job Suburbanization, Zoning and the Welfare of Urban Minority Groups." Journal of Urban Economics 5: 219-240. [65] White, J., 1988. "Location Choice and Commuting Behavior in Cities with Decentralized Employment." Journal of Urban Economics 24: 129-152. [66] Wooldridge, F., 2010. "Econometric Analysis of Cross Section and Panel Data." Cambridge: The MIT Press. 79

[67] Yankow, J., 2006. Why do Cities Pay more? An Empirical Examination of some Competing Theories of the Urban Wage Premium. Journal of Urban Economics 60: 139-161. [68] Zhao, Y., 1999. Labor Migration and Earning Differences: The Case of Rural China. Working Paper, University of Chicago. [69] Zhu, N., and Xubai. L., 2009. The Impact of Migration on Rural Poverty and Inequality: A Case Study in China. Agriculture Economics 41:191-204. 80

Appendix A.1: Poverty Map of China This poverty map is extracted from the World Bank GeoiQ System. It is shown that Gansu province and Inner Mongolia province located at the northwest part of China have annual per capita income below national average, especially Gansu Province. Figure A.1: China Annual Per Capita Income by Province 81

Appendix A.2: Test of Overidentifying Restrictions Analogous to the test of the validity of overidentifying instruments in an overidentified model using the Hansen-Sargan test, we should perform an overidentifying restrictions (OIR) test in the Heckit model to account for potential biases that" can do more harm than good" ( Bound, Jaeger and Baker, 1995). The test is conducted by explicitly including the endogenous decision variable R i into the structural income model and perform a Hansen-Sargan test, with the two exclusion restrictions (distance to the nearest bus station and distance to the nearest town center) treated as two instruments for R i. The model takes the form: log(y i ) = ρd i +γr i + βx i + δd i G i + e i (A.1) Y i now stands for the total household income for all the 2200 surveyed households. In STATA we use the postestimation "estat overid" command following the "ivregress gmm" command to produce the Hansen-J statistic. The test statistic is χ²(1) distributed because the number of overidentifying restrictions is 1. The yielded Hansen-J statistic is 1.604 and its p-value is 0.342. Thus we do not reject the null hypothesis and conclude that the overidentifying restriction is valid. 82

Appendix B: Descriptive Statistics for Low-Weeks-Worked-Per-Year Workers Table B.1 shows that nonworkers and the under-5-weeks-per-year workers are similar to the non-worker sample in terms of age, education, work experience, AFQT score, as well as number of biological children and spouse annual income. Table B.1 Descriptive Statistics for Nonworkers and Low-Weeks-Worked-Per-Year Workers U-5-Urban U-5- Rural Non-Urban Non-Rural Age 36.4 32.3 30.3 29.5 (5.1) (6.1) (5.4) (5.7) Education 8.4 7.5 7.7 8.2 (3.2) (3.2) (4.7) (3.4) Work Experience 8.9 10.8 6.5 6.4 (5.4) (4.4) (4.1) (4.1) AFQT Score 41.6 39.1 39.2 40.9 (26.7) (26.5) (30.1) (30.0) Number of Bio-Children 2.1 2.0 2.2 2.1 (1.3) (1.9) (1.4) (1.4) Spouse Income 23320 19340 19646 17880 (49266) (44038) (37821) (49010) Number of Obs 3710 498 4406 632 Note: Standard errors in parentheses. 83

Appendix C.1: Super-PWPUMA Definitions Boston MSA 1 020 Middlesex Co. (part)-malborough-framingham-norfolk Co.(part) 2 040 Esssex Co. (part)-danvers-salem-lynn-peabody 3 050 Middlesex Co. (part)-arlington-watertown-woburn 4 080 Suffolk Co. (part)-medford-cambridge-revere 5 090 Boston 6 100 Norfolk Co. (part)-newton-brookline-quincy 7 110 Norfolk Co. (part)-braintree-plymouth Co. (part)- Brockton 8 120 Plymouth Co. (part)- Barnstable Co. (part)-falmouth-yarmouth New York MSA 1 082 Dutchess Co. (part) - Westchester County (part) -Putnam Co. 2 085 Yonkers -Yonkers - White Plains -New Rochelle 3 090 Bronx Co. (part) - Bronx borough (part) 4 100 New York Co. (part) - Manhattan borough (part) 5 110 Queens Co. (part) - Queens borough (part) 6 120 Kings Co. (part) - Kings County (part) 7 130 Richmond Co. (part) - Staten Island borough (part) 84

Chicago MSA 1 090 Mchenry Co. (part)-elgin-aurora-dundee 2 100 Grundy Co.- Will Co. (part)-joliet-du Page-Frankfort 3 200 DuPage Co. (part)-naperville-milton-lisle 4 300 Lake Co. (part)-waukegan-avon-libertyville-vernon 5 400 Cook Co. (part)-schaumburg-elk Grove-Maine-Thornton 6 500 Chicago 85

Appendix C.2: Maps of the Super-PWPUMAs for the Studied MSAs Figure C.1: Super-PWPUMAs of the New York MSA 86

Figure C.2: Super-PWPUMAs of the Chicago MSA Figure C.3: Super-PWPUMAs of the Boston MSA 87