Abstract Using a panel data of China that covers the time period from 1997 to 2011, this research studies the impacts of four factors on inequality income level, emigration, public spending on education, and public spending on infrastructure. This study contributes to the literature with seven key findings. First, there is no strong evidence that supports the Kuznets Hypothesis. Second, the impact of emigration on inequality in the sending area shows significant heterogeneity across municipalities. Third, the effect of emigration depends on the definition: overall, permanent emigration has reducing effect, whereas temporary emigration and emigration of people who sought employment have increasing effects. Forth, the impact of temporary emigration and emigration of people who sought employment vary with the income level: they tend to decrease inequality when income level is low, and increase inequality when income level is high. Fifth, I find public spending on education increases inequality within regions. Sixth, public spending on infrastructure has reducing effect on inequality. Lastly, by using the lagged term, I find the income level affect inequality only after a period of time.
1. Introduction The evolution of social inequality has been a central topic in economics research. Along with economic development, many industrialized economies are characterized by high and increasing income inequality (Lenski, 1984). This trend is alarming because inequality can inhibit growth, slow poverty reduction, and often undermines the political progress (Birdsall, 2001). This paper focuses on the causal factors underlying inequality movements. Researchers began to study this issue since the twenties century. One major theory that characterized industrial and development is the Kuznets curve (1955). Using data from the U.S., England, and Germany in the nineteenth and twentieth century, Kuznets noted that inequality levels first increased as the economies grew and declined after a peak. The underlying mechanism, as explained in his paper, is the shift of population from traditional (agriculture) to modern industries. The income difference between two industries implies that inequality would increase if more people migrate to the urban areas. After the inequality level reaching the peaking point, other factors, such as demand trickle-down and human capital redistribution, start to kick in and reduce the income gap (Ray, 2010). Although this is a wellreasoned hypothesis, Kuznets addressed in his paper that the theory is perhaps 5 percent empirical information and 95 percent speculation. In this study, I employ a quadratic function to test the existence of a Kuznets curve. The test would confirm the Kuznets Hypothesis if the coefficient of income level squared is shown to be
negative. However, I do not find strong evidence for the Kuznets Curve. Instead, the results suggest that the relationship between inequality and income is either linear or a U shape. Nevertheless, it s rather arbitrary to simply reject the Kuznets Hypothesis, because if the data covers only the negative part of the inverted U curve, the test would show a negative linear relationship. By employing the lagged term, I also find the effect of income is time persisting the income level in last time period has significant reducing impact. In addition to income, a number of factors are also shown to be correlated to inequality as well. First, migration is often an endogenous process along with the movement of inequality. On the one hand, inequality changes when either people from the upper or lower tail of the income distribution emigrate (Stark, Taylor, & Yitzhaki, 1988; Jones, 2013). On the other hand, income inequality is widely believed to be the key motivation of migration (Harris & Todaro, 1970; Lewis, 1954). The two-way relationship between migration and inequality makes the dynamic process especially complicated. Moreover, the effect of emigration on the sending community depends on the composition change in income distribution. Using multilevel mixed effect models, I find emigration on average reduces inequality, whereas it shows significant heterogeneity across municipalities. In addition, by interacting income level and emigration, I find that the effect of emigration depends on the income level of the sending community. Emigration tends to raises inequality when income level is
relatively low. On the contrary, emigration has opposite effect when income level is relatively high. Second, public spending on education also affects inequality, since more education attainment is often associated to more income. Specifically, I investigate whether a decline in the share of illiterate or semi-illiterate people can effectively reduce inequality. Although more education coverage among the uneducated workers should reduce inequality, the rapid development in technology is creating high demand of educated people, which would increase inequality. This study shows that public spending on education can effectively reduce inequality within regions despite technology changes. Third, public spending on infrastructure can also play an important role in reducing inequality. This hypothesis is supported by many cross-region research (Ferranti et al. (2004), Fan and Zhang (2004), and Calderon and Serven (2004)). However, little is known about the outcome of within region inequality, which motivates me to include this factor to fill the gap in literature. This study finds that public spending in infrastructure can significant decrease inequality within communities. Due to the lack of high-quality dataset, many economists studied inequality with cross-sectional / cross-country datasets. Anand (1993), for example, used data of sixty countries in 1960 and confirmed the inverted-u shape. However, cross-sectional studies suffer from a number of drawbacks. First, the results may be biased because
the datasets only give a snapshot at a single point of time, which cannot represent the movement of inequality over time. Second, the data collection process and their measurements can vary substantially across countries, which could lead to misleading conclusions (Deininger & Squire, 1998). For example, while some official agencies provide data of the population residing in the country, others may report the population working in the country (Nielsen & Alderson, 1997). Third, cross-sectional studies that attempt to plot a single Kuznets curve implicitly assume homogeneity across countries: that is, all economies follow the same inverted-u curve that turns from increasing to decreasing inequality at the same income level (Saith, 1983). The dataset used in this paper comes from The Chinese Health and Nutrition Survey (CHNS), a high quality survey conducted by the Carolina Population Center (CPC) of the University of North Carolina at Chapel Hill, The Institute of Nutrition and Food Hygiene, and the Chinese Academy of Preventive Medicine. The panel dataset covers nine seven waves (1997, 2000, 2004, 2006, 2009, 2011) and nine representative provinces. Dataset was aggregated to municipality-level for this research. This paper aims to contribute to the current body of literature in a number of ways. First, this study will shed lights on the causal relation between inequality and income, emigration, education distribution, and infrastructure. Though many studies have explored the impact of income, less is known about the other three factors. Second, this research will be one of the few that study inequality within regions
instead of across regions. Although a handful of scholars have focused on inequality across states, provinces or between urban and rural, few studies focus on inequality within individual economies. Third, although China has been one of the most important economy in the world, studies of inequality in China is rare due to lack of high-quality data. The high-quality panel data employed in this study provides a previous opportunity to fill this gap in literature. Moreover, the dataset is not subject to drawbacks of cross-sectional data and measurement inconsistency as I mentioned before. The remainder of the paper is structured as follows: Section 2 discusses previous literature regarding emigration, education, and infrastructure. Section 3 introduces the dataset and measures of variables. Section 4 and 5 demonstrates the empirical models applied and results. Lastly, section 6 gives a conclusion of this study. 2. The Potential Factors That Changes Inequality In this section, I discuss previous literature about the three potential factors: emigration, education distribution, and infrastructure. Emigration Since the 1984 reform in China, an upsurge in the movement of human capital was driven by the rapid growth in manufacturing jobs in the coastal areas. The volume of
migration almost tripled from 12 to 32 million during 1995 to 2000, most of whom were temporary migrants from rural areas (Fan 2008). The growth of literature has increasingly recognized that the movement of inequality within the sending communities is a dynamic process along with emigration. The selectivity nature of migration itself is important in terms of the aggregate impact on inequality in sending communities (Black et al., 2006). If the costs of emigration are sizable and only affordable for the rich, the pioneer migrants are likely to come from the relatively wealthy households, while the poor group of people are trapped in nonproductive activities. This is often the case when the economy is in the early stage of development, when the credit market is usually immature and the relatively poor have little access to migration. In this case, the rich benefit the most as they look for high-return work elsewhere, and the inequality within the sending communities increases (Adams, 1993, 1998; Lipton, 1980; Stark, Taylor, & Yitzhaki, 1988). Conversely, labor mobility may generate some feedback effects on the sending communities. Migration of the rich group of people is likely to induce migration of the relatively poor for two reasons. First, the households relatively positions in the income distribution with respect to their reference group serve as strong motivations of migration (Stark & Bloom, 1985; Stark & Taylor, 1991. As more rich people migrate, the increase in inequality within sending communities often leads to the feeling of deprivation among the non-migrants, which boosts their desire to migrate. Second,
as the economy in the sending community becomes developed and the migration network is established in the destination area, the relatively poor are also be able to migrate and seek for more productive activities, and hence narrows down the income gap. As a result, the initial increase in inequality might be dampened or even be reversed, depending on which group of people benefit more. An important note here is that the impact of emigration depends on the level of development in the sending community. When the average income is low, the rich benefit the most because they can afford the cost, which leads to increase in inequality in the sending communities. When the average income is higher, the poor benefit the most as migration becomes more affordable. Therefore, in the model I employ an interaction term of income level and emigration to investigate this dynamic process. In this study, I hypothesize the sign of the interaction term is negative and the sign of emigration is positive emigration increases inequality when the income level is low, and the effect is the opposite when the income level is high. Public Spending on Education China experienced dramatic expansion in enrollment. The share of total GDP spent on public spending almost doubled over 1970 to 1985. Afterward, public spending on education dropped approximately 30%, which indicates that education increasingly relies on the non-government income sources (Rong, 2011).
The primary education coverage is associated with reduction in inequality for two important reasons. First, people with more education attainments are generally more productive (Simpson 1990, Nielson 1994). If more illiterate or semi-illiterate people obtain more education, the income gap between the rich and the poor will be narrowed down, which also means inequality level declines. Second, an increase in the supply of people with advanced skills should increase the competition, which imposes pressure on income of skilled workers (Card & DiNardo, 2002). However, more researchers have recognized the increasing effects of higher education on income inequality: technological change is causing the steady increases in relative demand for more-educated labor, which leads to increasing rate of return of higher education has been known as the Skilled Biased Technical Change hypothesis (Card & DiNardo, 2002). Therefore, even if the portion of illiterate people keeps shrinking, inequality can still increase if the educated workers gain substantially return from technological changes. For the above two reasons, the general effect of education distribution on inequality level is ambiguous. Public Spending on Infrastructure China s infrastructure has experienced dramatic growth since 1980. For instance, the investment in transportation account for over 6% of GDP in 1998 (China Statistical Yearbook 1999); the government announced ten major infrastructure
construction projects to stimulate domestic demand and increase employment in 2000 (China Securities Report, 2000). Economic growth theory suggests that infrastructure construction is associated to economic development (Barro, 1990). Better transportation is able to facilitate the mobility of human capital which stimulate trade and make specialization possible. Access to electricity also facilitates industries and boosts productivities. However, the impact of infrastructure on income distribution is less clear. It is widely believed that physical infrastructure development is expected to be more beneficial to poorer areas than to richer area, but most time this expectation on policymakers is taken for granted. Some studies have confirmed the negative relationship between infrastructure and income inequality. For instance, instance, Ferranti et al. (2004), Fan and Zhang (2004), and Calderon and Serven (2004) find that public investment in roads, dams, and telecommunications has contributed toward the alleviation of inequality and poverty in China and Latin America. However, Khandker and Koolwal (2007) find that access to paved roads has a limited distributional impact in rural Bangladesh. The diversity of the empirical findings underscores further exploration on the role of infrastructure. Moreover, to my best knowledge, existing studies have either focused on the impact on income inequality across countries (Chong & Calderon, 2001) or across regions within countries (Ferranti et al. (2004), Fan and Zhang (2004), and Calderon and Serven (2004)). This paper attempts to contribute
to the current body of literature by estimating whether government spending reduces inequality within a community. 3. Data In this section, I introduce the dataset, the measures for inequality, income level, emigration, public spending on education and on infrastructure. 3.1 The Chinese Health and Nutrition Survey (CHNS) To address my research question, I use the The Chinese Health and Nutrition Survey (CHNS). The project is conducted by the Carolina Population Center (CPC) of the University of North Carolina at Chapel Hill, The Institute of Nutrition and Food Hygiene, and the Chinese Academy of Preventive Medicine. This panel dataset covers nine survey waves (1989, 1993, 1997, 2000, 2004, 2006, 2009, 2011) and nine provinces. The sampled provinces were randomly chosen and vary substantially in geography, economic development, and public resources. The survey is conducted at both the community level and the household level. A community is defined as a rural village or a neighborhood in an urban or suburban area. A multistage random cluster process was used to draw the sample surveyed in each province. There were 190 primary sampling units in 1989 to 1993. A new province and its sampling units were added in 1997, and another two were added in 2011. In each community, between 20 to 35 households were drawn, and the
household survey cover all the members with permanent residence of hukou. Currently, there are around 4,400 households and 19,000 individuals in the survey. The dataset provides valuable information of income level, demographic characteristics, emigration, education level, and infrastructure. This study uses data since 1997, when the data of emigration became available. The data were cleaned and aggregated into municipality and provincial level. 3.2 Measures I would introduce the definitions, the method to construct each variable, the method to choose measures when there are multiple choices, and some summary statistics. Each variable is constructed at municipality level. Income Level A variety of income measurements can be derived from the household and individual survey. In the household survey, there are two definitions of income: total net household income and gross household income. Total net household income is defined as the sum of all sources of income and revenue minus expenditures. The survey summarizes potential sources of income into nine categories: business, farming, fishing, gardening, livestock, non-retirement wages, retirement income, subsidies, and other income. Gross household income is defined as the sum of total income, which equals to total net household income plus expenditures. For both total
net household income and gross household income, I consider for inflation, imputed data and normalization. First, I construct both nominal income and real income inflated or deflated to 2009. Second, I construct measures with and without the inclusion of imputed data. Third, I normalize household income by per capita and by equivalence scales. Note that equivalence scale is a scale factor that takes into account of both the size of household and the age of household members. In particular, the method assigns a value of 1 to the first household member, of 0.5 to each additional adult and of 0.3 to each child (OECD, 2011). Overall, there are 16 income measures derived from the household survey. There are two definitions of income in the individual survey. Similarly, individual net income is defined as the sum of all sources of income and revenue minus expenditures. Another definition is annual individual wage, which includes bonus and other income from jobs. For both definitions, I consider for inflation and imputed data in a similar manner. Six income measures are derived from the individual survey. To select from the 22 income measures, I calculated Gini coefficient using each of them and compare the overall trend. The comparisons show that imputed data and inflation have little effect on the results. Household net income (adjusted for equivalence scales) individual net income, and individual wage shows distinct patters, so I choose them (nominal measures without imputed data) as preferred income measures. Figure 1 shows the general trend of the three income measures:
Figure 1: Income Measures Income 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 5000 10000 15000 20000 Year Household Net Income Individual Wage Individual Net Income Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: All three measures are calculated at municipality level. The household net income is adjusted for equivalence scales. Net income is defined as all revenues minus expenditures. Gross income equals to net income plus expenditures. Individual wage includes bonus and other income from jobs. Generally, all three measures show that the income level in China goes up considerably since 1997, and the rate of increase is even higher since 2006. The income level in 2011 is more than four times of that in 1997. The gaps among the three measures increase moderately before 2004, and narrow down afterward. Inequality Level There are a handful of possible measurements for income inequality. Frequently used measures include: Gini index, Theil index, relative mean deviation, standard
deviation of log income, and P90/P10. A more detailed description of these inequality measurements can be found in the Appendix B. I calculate the five inequality measurements first, then estimate the model using each inequality measurement as the dependent variable. For measurements that generate similar results, I only pick one for subsequent estimations. Figure 2 shows the general trend of Gini computed using the two preferred income measures: Figure 2: Gini Indices Computed Using Preferred Income Measures Gini Indices 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011.3.35.4.45.5 Year Gini-Household Net Income GiniIndividual Wage Gini-Individual Net Income Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: All three Gini indices are constructed by averaging Gini at municipality level. The household net income is adjusted for equivalence scales. Net income is defined as all revenues minus expenditures. Gross income equals to net income plus expenditures. Individual wage includes bonus and other income from jobs.
As shown in Figure 2, the Gini indices calculated by the three chosen income measures follow rather different paths, which gives us a second reason to use different income measures. The Gini constructed by individual net income is the highest one. Compared to Gini constructed by individual net income, Gini constructed by household net income shows less inequality, because inequality caused by people who do not work can be average out by people who work in their families. Both Gini indices show considerable increase between 1997-2006, which is followed by a decline afterward. The Gini constructed by individual wage is the lowest and most volatile one. It increases from 0.3 in 1997 to almost 0.45 in 2006. The trends of Gini across provinces and municipalities vary substantially: Figure 3: Gini Indices of 4 Selected Provinces Gini 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011.35.4.45.5.55 Year Guangxi Henan Heilongjiang Jiangsu Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: Household net income is used to calculate Gini indices. Each line represents the Gini index with in the corresponding province.
Figure 4: Gini Indices of Individual Municipalities in Nine Provinces Gini 21 23 32.2.4.6.8 37 41 42.2.4.6.8 43 45 52 1997 2000 2003 2006 2009 2012 1997 2000 2003 2006 2009 2012 1997 2000 2003 2006 2009 2012.2.4.6.8 Graphs by Province ID Year Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: Household net income is used to calculate Gini indices. The province IDs represent: 21=Liaoning, 23=Heilongjiang, 32= Jiangsu, 37=Shandong, 41=Henan, 42=Hubei, 43=Hunan, 45=Guangxi, 52=Guizhou. Each line represents the Gini index of a municipality in the corresponding province. Figure 3 shows the inequality levels of four selected provinces, and Figure 4 shows the Gini of each municipalities within the nine provinces. As we can see, the trend of inequality level in each municipality and each province show considerably different patterns. For this reason, I choose to aggregate data at municipality level to account for heterogeneity. Moreover, I employ the multilevel mixed effect model to measure municipality-specific intercepts and slopes, which will be further explained in the empirical results section.
Emigration There are three available measures for emigration from the CHNS dataset. Because the impact of emigration is caused by the changes in production activities, so I restrict my analysis only to people who are between age 16 to 60. Each measure is aggregated to municipality level. a) Share of Respondents Who Emigrated Permanently This variable is derived from the household survey, in which each household member is asked if he/she lives at home or has moved to another county/city. The family member may move for schooling or for employment. b) Share of Emigrated Labor Force This variable is measured by the share of people who sought employment elsewhere last year, which is estimated by the community head or community accountant in the village. Note that this measurement focuses on temporary emigration, since workers permanent residency permit (hukou) remain at their original province. The share of people who seek employment elsewhere can estimate how many people emigrate to shift to other economic activities. Although the emigration may happen in the same county/city, it still can affect inequality level people shift their occupations or economic activities on inequality.
c) Share of Temporary Emigrants This variable is measured by the share of people in the village who worked out of town for more than a month last year. It is estimated by the community head or community accountant in the village. These estimated values are then aggregated to municipality-level values. Note that this variable only provides a rough estimate of emigration, because it does not distinguish people who emigrate temporarily, permanently, within or outside the municipality. It may also include people who did not change occupation but only changed workplaces temporarily. Figure 5: Emigration Temporary Emigration and Emigration for Employment.02.03.04.05 Permanent Emigration.1.15.2.25.3 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 emigrated for employment emigrated temporary Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: The share of temporary emigration is constructed by the estimated value obtained from the community head or community accountant in the village. These estimated values are then aggregated to municipality-level values. Share of permanent emigration and emigrated labor are calculated by data from the household survey, which asks questions regarding whether the family member has moved elsewhere or sought jobs elsewhere in last year. Again, they are municipalitylevel variables.
Figure 5 shows that the share of emigrated labor force and temporary emigrants rise up dramatically and almost tripled between 1997 and 2011. The share of permanent emigrants also increases, although in a much smaller scale, due to government restriction and costs. Public Spending On Education This variable is measured by percentage of respondents who obtain less than 6 years of education. It is constructed based on the years of schooling reported by respondents of the household survey. This portion of people are counted as illiterate or semi-illiterate. Figure 6: General Trend of Public Spending On Education.08.1.12.14.16.18 Percentage of illiterate or semi-illiterate 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: People who receive less than 6 years of education are counted as illiterate or semi-illiterate.
As shown in Figure 6, the share of illiterate or semi-illiterate people drops rapidly before 2004. It picks up a little between 2004 to 2006, and then declined steadily to around the lowest point of 0.08. Public Spending on Infrastructure This variable is measured by the transportation score provided by Jones-Smith and Popkin (2010). The score is evaluated based on the most common type of road, distance to bus stop, and distance to train stop. The method to construct the score is introduced in the Appendix A. Figure 7: General Trend of Transportation Score 5.2 5.4 5.6 5.8 6 6.2 Transportation Score 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: the transportation score provided by Jones-Smith and Popkin (2010). The method to constructing the score is shown in Appendix A.
The transportation score shows steady improvement since 1997 with only a small decline between 2004 to 2006. 4. Empirical Models First, I use the baseline model to estimate the presence of the Kuznets Curve. The model contains only income measure, its square term, and other variables that affect overall inequality level as predictors:!"#$ %,' = ) * + ), - %,' + ). -. %,' + /, 0 %,' + /. 1 %,' + 2 %,' (1) where i and t index municipalities and year respectively. Ineq denotes the inequality level, such as Gini coefficient or Theil; Y represents the income level. The Z vector is included to control for demographic structural variables such as the share of female, the share of young people (younger than 16), share of elder people (older than 60), and the share of minority. X represents variables we are interested in: emigration, education distribution, and infrastructure. A detailed description of variable names is included in table 1. If the estimation results yield a quadratic function with a negative β2 and a positive β1, I can conclude the unconditional Kuznets hypothesis is true. In addition, I compare the efficiency among OLS, fixed and random effect models. Note that this model makes a strong assumption that the same model applies to all municipalities. Afterward, I use a multilevel mixed effect regression test to investigate heterogeneity across municipalities. It generates the distribution of coefficient
estimations of individual communities and allows me to compare them with the average estimation. This test is characterized as having both fixed and random effects. The fixed effect is similar to that of a standard regression, as the model removes the time-invariant effect on predictors. The random effect captures the possibilities that the characteristics of individual communities lead to different slope and intercept. The overall error distribution of the model is assumed to be Gaussian. Essentially, the model specifies a municipality-specific random intercept and municipality random slope for the i th observation in the j th municipality:!"#$ %,3,' = ) * + 4,,3 + ), + 4.,3 - %,3,' + (/, + 4 6,3 )0 %,3,' + (/. + 4 8,3 )1 %,' + 2 %,' (2) The model assumes that the covariates 0 %3', 1 %', and the error term 2 %' are independent of 4,,3 and 4.,3. Next, I employ multiple GLS models to explore the lagging effects of income, and test whether the effect of emigration varies with income level. I investigate lagged income because the multilevel mixed effect model shows evidence that the impact of income on inequality is not immediate. First, I compare between models using contemporaneous and lagged income:!"#$ %,' = ) * + ), - %,' + /, 0 %,' + /. 1 %,' + 2 %,' (3)!"#$ %,' = ) * + ), - %,'9, + /, 0 %,' + /. 1 %,' + 2 %,' (4)
Then, I include the lagged dependent variable to equations:!"#$ %,' = ) * + ),!"#$ %,'9, + ). - %,' + /, 0 %,' + /. 1 %,' + 2 %,' (5)!"#$ %,' = ) * + ),!"#$ %,'9, + ). - %,'9, + /, 0 %,' + /. 1 %,' + 2 %,' (6) The results need to be interpreted carefully though, because the strong correlation of lagged and current inequality can suppress the explanatory power of other independent variables when independent variables or disturbance are serially correlated (Achen, 2000). Lastly, I include interaction terms of emigration and (lagged) income level:!"#$ %,' = ) * + ), - %,' + /, 0 %,' + /. ;<=> %,' - %,' + / 6 1 %,' + 2 %,' (7)!"#$ %,' = ) * + ), - %,'9, + /, 0 %,' + /. ;<=> %,' - %,'9, + / 6 1 %,' + 2 %,' (8) As explained in section 2, the reason of including the interaction term is to study whether the effect of emigration varies with the income level. In equation (8), the interaction term implies that people make decision based on past income instead of current income. This assumption is reasonable because contemporaneous is available only at the end of the year. 5. Empirical Results This section presents the main empirical results. I first test the significance of non-linear terms with model 1. Second, I compare the results of OLS, fixed and random effect models. Third, to test for heterogeneity, I allow for multilevel mixedeffect model to compare across municipalities. Forth, I allow for lagged effect of
independent and dependent variables. Finally, I employ fixed effect model that allows for auto-correlated disturbance with AR(1) process and heteroskedastic error. 5.1 Testing for the Inverted U Shape As explained in section 5, I employ equation 1 to test if the coefficient of nonlinear term is significant. Table 3 presents the OLS regression results. In table 3, each column represents the regression result of equation (1) using one combination of income measure and emigration measure. Compared by columns, the results are consistent across three emigration measures, but inconsistent across income measures. In models that employ household net income and individual wage, the coefficients of log of income squared indicate that the relationship between income and inequality is linear, and hence the Kuznets Hypothesis is rejected. In the model where individual net income is employed, the relationship is a U shape instead of an inverted U shape, which also rejects the Kuznets Hypothesis. However, it would be arbitrary to simply reject the Kuznets curve for several reasons. First, we should note that the dataset covers a relatively short period of time, and it is possible that China has not reached the peaking point till 2011 or has already passed the peaking point before 1997. In both cases, the data would only capture the rising or declining part of the Kuznets curve, and the results would be linear. Second, the U shape could be related to another theory derived from the Kuznets Hypothesis. In Rati Ram s study of inequality in U.S. from 1947 to 1978 (1991), he also found the
relationship to be U shape instead of inverted U shape. His finding has led to the S- curve hypothesis, which implies that the inequality in an economy will increase again after passing the lowest point, which means the inverted U curve will repeat itself. Although we are not clear whether either of the arguments is true, we should not reject the Kuznets Hypothesis based on the results. 5.2 OLS, Fixed Effect and Random Effect Models Since income squared is insignificant, I no longer preserve it in the model. Table 4 shows the comparison between OLS, fixed and random effect models using three emigration measures. The income measure used is the household net income In all 3 cases, random effect model is the most efficient one, which implies that the variation across municipalities is likely to be random and uncorrelated with explanatory variables. In addition, the fact that OLS is the least efficient implies that we can not assume homogeneity among municipalities, and hence we should not apply the same model to all observations. Nevertheless, the results of OLS, fixed and random effect models are consistent using any emigration measure. The effect of emigration on inequality in sending areas is unclear, depending on which measure we use. First, the more people emigrate for employment elsewhere, the lower the inequality level. The result suggests that the relatively poor households are the ones who benefits the most. Considering the relationship between income and the incentive to emigrate, households at the top of a community s income distribution
may have little incentive to move elsewhere, since they are likely to be content with current working opportunities and social status. Therefore, emigrants are mainly drawn from middle or lower class. Inequality could be narrowed down when a considerable number of the poor households emigrate and find employment with higher income. The remittance they send back to their families would reduce inequality in the sending communities. The same argument can be applied to permanent emigration. Nevertheless, permanent is more costly compared to temporary emigration for employment. Therefore, people in the lower class are less likely to emigrate, and the reducing effect of permanent emigration on inequality is likely to be minimal. The regression results show that the impact of permanent emigration on inequality is insignificant, which suggests that the relatively poor and rich benefit from emigration equally. The result confirms to the assumption, since fewer people in the lower class benefit from permanent emigration compared to emigration for employment. Temporary emigration, however, seems to have increasing effect on inequality in sending areas. The relationship between the incentive to emigrate temporarily and income level is ambiguous, because the barrier of temporary emigration is minimal. Moreover, because temporary emigration includes all people who live elsewhere for more than one months, we can not infer much about the purpose of emigration and the impact on economic activities. Overall, the result suggests that people at the middle or the top of the income distribution
benefit the most, and the remittance they send back may exceed the remittance of the relatively poor emigrants. The results show that public spending on education can narrow down income gap. As discussed in Section 2, the more education obtained by people in the bottom of income distribution, the more productive they are, and the lower the income inequality. Public spending on infrastructure, however, seem to have little effect on reducing inequality, which that implies people benefit from it approximately equally. 5.3 Testing for Heterogeneity - Multilevel Mixed-Effect Model Based on Equation (2), I test for heterogeneity across municipalities. As explained before, the multi-level mixed effect model allows for slopes and intercept to vary across municipalities. Figure 8-12 presents the density functions of the coefficients of income, emigration, education distribution, and infrastructure across municipalities. The red line represents the mean value of estimation.
Income Level Figure 8: Density Function of Coefficients Estimation Current Income Current Income Density 0 5 10 15 20 -.05 0.05.1 kernel = epanechnikov, bandwidth = 0.0067 The above Figure shows that the impact of current income on inequality shows significant heterogeneity across municipalities, and the average estimate is close to zero. One possible explanation is that income only comes to affect inequality after a period of time. Therefore, I conducted another mixed effect model using the lagged income level:!"#$ %,3,' = ) * + 4,,3 + ), + 4.,3 @. - %,3,'9, + (/, + 4 6,3 )0 %,3,'9, + (/. + 4 8,3 )1 %,' + 2 %,' I apply the same mixed effect model to test the dispersion of coefficient of lagged income. The result is shown in figure 13.
Figure 9: Density Function of Coefficients Estimation Lagged Income Lagged Income Density 0 5 10 15 -.1 -.05 0 kernel = epanechnikov, bandwidth = 0.0065 Compared to the distribution of current income, the coefficients of lagged income of most municipalities now fall in the negative area, as well as the mean estimation. The figure implies that income level has reducing effect on inequality, but the impact occurs only after a period of time. This point will be further investigated in subsequent models. Moreover, there is significant heterogeneity across municipalities, which suggest that unobserved characteristics of municipalities affect the impact of income. Emigration Figure 8-11 shows the density function of the three emigration measures: temporary emigration, permanent emigration, and emigration of labor force.
Figure 10: Density Function of Coefficients Estimation Temporary Emigration Temporary Emigration Density 0 2 4 6 8 0.1.2.3 kernel = epanechnikov, bandwidth = 0.0138 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: The share of temporary emigrants is constructed by the estimated value obtained from the community head or community accountant in the village. The variable is aggregated to municipality level. Only emigrants who are aged between 16-60 are included. Figure 11: Density Function of Coefficients Estimation Permanent Emigration Permanent Emigration Density 0.5 1-1 -.5 0.5 1 kernel = epanechnikov, bandwidth = 0.0994 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: Share of permanent emigration is calculated by the data from the household survey, which asks questions regarding whether the family member has moved elsewhere. The variable is aggregated to municipality level. Only emigrants who are aged between 16-60 are included.
Figure 12: Density Function of Coefficients Estimation Emigration of Labor Force Emigration of Labor Force Density 0 2 4 6 8 -.1 0.1.2.3 kernel = epanechnikov, bandwidth = 0.0166 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011). Notes: The share of emigrated labor force is constructed by the data from household survey, where the respondents are asked whether they sought employment last year. The variable is aggregated to municipality level. Only emigrants who are aged between 16-60 are included. Of the three measures, the distribution of temporary emigration and emigration of labor force are similar. Both of them distribute in the positive area, which suggests that these two types of emigration have increasing effect on inequality in almost all municipalities, although the effect is larger for some than the others. In addition, the ranges of their coefficients are quite similar too, whereas the increasing effect of temporary emigration is slightly higher on average. As explained in Section 2, technological changes can cause steadily high demand for educated workers and increase the income gaps. Even though the relatively low costs of temporary emigration favor the poor group, the graph shows that the effect of technological
changes is dominant, and hence the overall effect of temporary emigration and inequality is positive. The density function of permanent emigration, however, shows that the impact of emigration is ambiguous. In about half of the municipalities, it increases inequality, and it decreases inequality in the other half. As discussed before, usually the uppermiddle class are the first group to emigrate. As this group of people move up to the top of income distribution, inequality would increase. As the migration network becomes more established, people in the lower class are also able to migrate, and therefore inequality would decrease. Therefore, because some economies may have more developed networks than the others, the impact of permanent emigration can be different. Figure 11: Density Function of Coefficients Estimation Public Spending on Education Share of Illiterate and Semi-illiterate Respondents Density 0 1 2 3 -.6 -.4 -.2 0.2 kernel = epanechnikov, bandwidth = 0.0319 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011).
Notes: Public spending on education is measured by the share of survey respondents who are aged between 16-60 and have not received 6-year compulsory education. The variable is aggregated to municipality level. Only emigrants who are aged between 16-60 are included. Figure 11 shows that public spending on education has increasing effect on inequality in almost all areas. In the most extreme case, 1% increase of illiterate or semi-illiterate people would lead to 0.06 reduction in inequality. This result is contrary to our common sense, since we usually expect inequality to decrease when more people obtain 6-year of compulsory education. As discussed in Section 2, the positive relationship between inequality and public spending on education can be caused by technological changes, which raises the income of educated workers. Only in a few areas, more illiterate or semi-illiterate people is correlated to more inequality. Figure 12: Density Function of Coefficients Estimation Public Spending on Infrastructure Public Spending on Infrastructure Density 0 50 100 150 200 250 -.008 -.006 -.004 -.002 0 kernel = epanechnikov, bandwidth = 0.0004 Data source: CHNS data (1997, 2000, 2004, 2006, 2009, and 2011).
Notes: The transportation score is constructed from the data obtained from the official record provided by the community head or community accountant. The score is constructed by Jones-Smith and Popkin (2010). More details are presented in Appendix A As expected, in almost all municipalities, better transportation condition is associated with less inequality. The strong heterogeneity across municipalities is possibly the reason why the significance is weakened in previous tests. Nevertheless, the density function shows that public spending has reducing effect on inequality. 5.4 GLS models with lagged income, lagged dependent variable, and interaction of (lagged) income and emigration First, I employ equation (3) and (4) to compare between effect of contemporaneous and lagged income: The results are presented in table 5. Consistent to previous multilevel mixed effect model, only lagged income shows statistical significance, which suggests that income affects inequality only after a period of time. One unit of increase in income in last period would lead to 0.045 to 0.05 decrease in Gini index. Current income, however, has no significant impact on inequality. Both emigration of workers and temporary emigration has increasing effect on inequality, despite which income measure used. 0.01 increase in the share of emigration of labor force would lead to 0.075% increase in inequality, while 0.01 increase in the share of temporary emigration would lead to 0.095% increase in inequality. Similar to the conclusion we get in Section 5.3, the positive relationship
between these two types of temporary emigration and inequality suggests that technological change plays an important role. Permanent emigration has reducing effect when the past income is used. One percent increase in permanent emigrants is associated with 0.17% decrease in Gini. The result suggests that permanent emigrants is mainly drown from the top of the income distribution. As more of them move to elsewhere, the overall inequality in the sending community would decrease. The share of illiterate or semi-illiterate people has no significant effect in inequality. From the two factors that affect the impact of education, the result suggests that while the poor group are able to earn higher income with more education, meanwhile technological changes benefit the rich group and balance the overall impact of education. Public spending in transportation has significant reducing effect. However, it is difficult to interpret the numerical result since the score is constructed based on the access and quality of roads. Then I include lagged dependent variable as shown in equations (5) and (6). The results are presented in table 6. Lagged Gini is positively correlated to current Gini, since they are serially correlated. Consistent to last test, contemporaneous income shows no statistical significance, and lagged income shows significant reducing effect. The effects of emigration are the same when using current income, however they become insignificant when using lagged income. As discussed
before, adding lagged independent variable can suppress the explanatory power of other independent variables, even if the variable is actually substantive. Other results regarding education and infrastructure stay the same. Lastly, I include interaction terms of emigration and (lagged) income level based on Equation (7) and (8). The results are shown in Table 7. Again, only lagged income has significant reducing effect on inequality. In both model (7) and (8), either permanent emigration or its interaction term is statistically significant. The result of temporary emigration and emigration of labor force are similar. Both of them shows negative sign for emigration term, and positive sign for the interaction term, despite which income measure is used. This result indicates that emigration initially has reducing effect on inequality when income level is low. However, the impact is reversed when the economy is more developed. The result suggests a dynamic process of emigration and income. When income level is low, emigration can reduce inequality because the relatively poor group benefit the most. Particularly, because the costs of temporary emigration or temporarily moving out for employment is relatively low, compared to permanent emigration. Therefore, even people in the lower tail of income distribution can afford to emigrate and look for higher income jobs, and the income gap would decrease at the beginning of emigration. As the economy is more developed, however, the beneficiary becomes the rich group. This is when technological changes start to play a role in the model. We can assume that the rich group are only willing to work
elsewhere when there are better job opportunities. If the demands for educated or skilled workers create high-income jobs elsewhere, the rich would be motivated to emigrate temporarily, and hence inequality would increase. While it is difficult to estimate technological changes, it is often positively associate with income level (Krumar, 2002). Therefore, it is a potential factor that determine the effect of emigration. 6. Conclusion This paper aims to study the impact of income, emigration, public spending on education, and public spending on infrastructure. Specifically, I study the existence of the Kuznets Curve, the heterogeneity of the impact of each interested variables across municipalities, the effect of lagged income, and whether the effect of emigration varies across income level. There are several key findings. First, the relationship between income and inequality is shown to be linear or U shape, which fails to support the Kuznets hypothesis. Second, the effect of emigration on inequality depends on the definition of emigration. Overall, I find temporary emigration and emigration of labor force to have increasing effect on inequality, and permanent emigration to have the opposite effect.