Taxation and Migration of Peasants: Evidence from the Tax-for-Fee Reform in Rural China

Abstract This paper studies the effect of rural taxation on Chinese peasants’ mobility towards urban areas. We find that the tax alleviation due to the tax-for-fee reform significantly reduces rural-to-urban migration, with a 10% decline in tax leading to a 10% reduction in migration. We provide compelling evidence that the result is not driven by unintended policy effects of the tax-for-fee reform in fostering rural non-agricultural job opportunities. The effect of taxation is more pronounced and precisely estimated on migration across provinces or at older ages. To the best of our knowledge this is the first study to identify the tax-induced migration among low-skilled and low-income laborers.


Introduction
Tax-induced mobility of human capital has become a central public policy issue, especially since economic integration lowers barriers to migration. Many European countries have introduced preferential tax schemes for highly skilled foreign workers, creating a de facto tax competition for talent. Some subnational governments regularly advertise their favorable tax environments in an explicit attempt to attract investments and skilled laborers, indicating that tax competition occurs not only across but also within national borders. In addition to the mobility of the talented and the rich, tax policies may also influence the migration of low-skilled and low-income laborers. For example, in order to promote outflow of labor in return for foreign exchange remittances, in 1995 the Philippine government exempted Filipinos working overseas from income tax, most of whom are low-skilled laborers working in manufacturing or service industries for low wages.
Although anecdotes suggest that people may vote with their feet for low taxes, there remains a frustrating consensus among scholars that little empirical evidence is available to support this claim (Moretti & Wilson, 2017). Existing literature on the influence of taxation on the labor market focuses primarily on how taxes discourage labor supply, reporting high elasticities among low-income workers and retiring older workers (see OECD, 2011, and the extensive references therein). Regarding mobility effects, although many studies find that corporations and investments move in response to tax policy (Giroud & Rauh, 2019), scarce literature has explored the effects of taxation on individual labor mobility, yielding little conclusive evidence (Feldstein & Wrobel, 1998;Young, Varner, Lurie, & Prisinzano, 2016).
Drawing upon a unique dataset constructed from the Chinese Household Income Project in 2002, and a series of county level statistics, our investigation of the tax-for-fee reform in rural China yields three main empirical findings in regards to migration. First, we uncover compelling evidence of a negative effect of the reform-induced tax cut on migration, with a 10% decline in tax leading to a 10% reduction in migration; Second, we show that as the tax-for-fee reform relieves rural households from the formerly burdensome tax obligations, rural laborers tend to farm more and migrate less, with their local non-farm employment unaffected; Third, we find that migration bearing higher costs, such as those across provinces or at older ages, are more responsive to rural tax reduction.
Our study adds to a longstanding literature that examines migration and its drivers (see surveys by Massey et al., 1993). In particular, our finding of the vital impact of taxation sheds new light on the neoclassical equilibrium explanation of migration, which sees laborers as rational actors whose mobility decisions are made based on cost-and-benefit calculations of migration relative to non-migration, according to the changing 'push and pull' factors (Hare, 1999;Harris & Todaro, 1970). Our study also contributes to a small but emerging literature on the sensitivity of domestic migration to taxation, which mostly concerns the developed countries in Europe and America, and uncovers generally mixed results. Complementary to this existing literature, our research considers the internal mobility of low-skilled and low-income peasants within a developing country such as China.

The tax-for-fee reform
Rural tax burden was one of the deepest social problems facing Chinese society in the early 1990s (Bernstein & L€ u, 2003). Before the tax-for-fee reform peasants were required to pay a wide array of taxes and fees including: (i) agricultural tax, (ii) township-and-village-retained fees (santi wutong) which were collected essentially to finance rural public goods, and (iii) ad hoc fees that local cadres imposed in the name of village governance (Bernstein & L€ u, 2003). While the tax was levied according to the amount of land that a rural household contracted, the fees were levied upon simple headcount. Local cadres were given great autonomy to determine the amount of 'head tax', namely, the fees that each rural resident should pay. This tax scheme was criticized harshly for being both arbitrary and excessive, impairing the economic interests of peasants.
Against this background, the central government implemented the tax-for-fee reform experimentally in Anhui Province in 2000. The basic rationale of the reform was to streamline the tax structure to reduce the scope of local officials' discretion. In particular, the 'one-issue-one-meeting' (yishi yiyi) system regulated that any public project that required financial contribution from the villagers had to be discussed thoroughly in a village meeting, and only the liability that was approved by the majority was considered legitimate. After expansion to 20 provinces in 2002, 1 the reform was finally nationally completed in 2003. Although the tax-for-fee reform slightly increased the agricultural tax rate, it substantially decreased fees. In fact, the reform was well-acknowledged as a great success in relieving rural households from excessive taxation and improving their economic well-being (Kennedy, 2007;Yep, 2004).

Rural-to-urban migration
Since the mid-1980s rural laborers have increasingly flowed to urban labor markets to pursue non-agricultural employment. As the tax-for-fee reform tremendously reduced rural tax, we speculate whether improvements in rural income opportunities would undermine the relative advantages of urban employment and therefore pose a negative effect on migration. Note that a unique characteristic of Chinese rural-to-urban migration is its temporary nature. On the one hand, as rural cadres may redistribute the right to use collectively owned land among rural households, migrants can be deprived of their use rights due to the long absence from the village (Zheng, Gu, & Zhu, 2020); On the other hand, rural migrants are de facto second-class citizens in urban areas, because the household registration system does not entitle them to the same social benefits as their urban native counterparts (Solinger, 1999).
This temporary nature of Chinese rural-to-urban migration bears two implications on the generalizability of our study. First, the large mobility response we uncover in the Chinese context does not necessarily carry over to other institutional settings, as it could be easier for people to change their decision about temporary migration than about permanent migration. Second, unlike permanent migrants who can flee from high tax rates in their places of origin, Chinese migrant workers still have to pay rural tax, in order to maintain their usage rights to collectively-owned land, as well as other benefits associated with their registration statuses in the countryside (Zheng & Gu, 2021). Therefore, taxation differentials between migration and local employment stay unchanged under the new rural tax scheme. However, the tax cut still deters labor outflows from Chinese villages, mostly because the marginal returns to migration become much less attractive to peasants, relative to the rising rural income levels.

Chinese Household income project
To substantiate our assertion that the formerly burdensome taxes and fees levied upon rural households functioned as an important push factor for migration, we analyze the rural portion of the data from the 2002 Chinese Household Income Project. Jointly conducted by the Rural Survey Group of the National Bureau of Statistics of China (NBSC) and the Institute of Economics at the Chinese Academy of Social Sciences, CHIP is arguably one of the most authoritative data sources on household financial conditions in China. Its survey instruments were carefully tailored to capture detailed information on the labor force activities of all household members, as well as a wide variety of family expenditures including miscellaneous fees that rural households paid to the government. Following a multistage sampling strategy, altogether CHIP 2002 investigated 9,200 rural households scattered throughout 961 villages in 22 provincial level administrative units, offering us a large sample reasonably representative of rural China.

Measures of migration
Our dependent variable is the proportion of laborers who were migrant workers in 2002, for each rural household. To construct this variable, we first define the labor force as people aged from 16 to 60, excluding the retired, disabled and students. We then carefully examine the labor force activities of each rural laborer, determining migrants as individuals whose primary job in 2002 was nonagricultural work outside their native township. In order to ensure that our empirical results are not sensitive to alternative definitions of migration, we further divide migration behavior into three categories based on the administrative borders they cross: (i) cross-township migration of which the destination is beyond the native township but within the native county; (ii) cross-county migration beyond the native county but within the native province; and (iii) cross-province migration outside of the native province. 2 We will conduct seemingly unrelated regressions to distinguish the impacts of financial burdens on these three types of migration, respectively.

Measures of decrease in financial burden
The explanatory variable of interest is the amount of decrease in tax liability for each rural household between 1998 and 2002, a time span that witnessed the beginning and the extensive expansion of the tax-for-fee reform. We recognize that in order to assess the responsiveness of migration to tax scheme differentials, a more ideal research setting would be a difference-in-differences analysis, which would compare migration behavior before and after the tax-for-fee reform, between households that were affected and unaffected by the new scheme. However, because the migration status of rural laborers before the tax-for-fee reform is not available in CHIP 2002, we resort to focus on the impact of changes in financial burdens on migration, with controlling for the migration rate at the village level in 1998, when the tax-for-fee reform had not yet begun. 3 Details on the construction of this independent variable are presented in the Supplementary materials.

Control variables
The effect of the reduction in financial burdens could be confounded by some overarching socioeconomic characteristics that impact decisions to migrate. Therefore, we control for a wide array of village characteristics including population size, per capita land endowment, total sales revenue of TVEs, and the distance of the village from the nearest transit station. Three dummy variables are incorporated, measuring respectively whether the village collective organized the rural labor force to migrate out for work, whether the proportion of the largest group of villagers sharing the same family name exceeded 50%, and whether there is occurrence of natural calamities (one if yes and zero otherwise). As the crux for our research is the impact of tax-forfee reform on reducing rural financial burdens and decelerating rural-to-urban migration, we also include a dummy variable indicating whether the village had experienced any financial burden alleviation policies other than the tax-for-fee reform (one if yes and zero otherwise).
Household and labor characteristics may also affect migration behavior. We thus include the share of males, the average age, the average schooling years, and the dependency ratio of household laborers in our equations. Moreover, we employ a set of dummy variables to indicate if a given household belonged to the largest surname group in the village, whether any household member was a Communist Party member, or was a rural cadre, respectively. We also control for the amount of land that a rural household contracted and the household size. Details on the descriptive statistics can be found in the Appendix (Supplemental Material).

Baseline estimates
We begin the analysis by estimating the relationship between the decrease of financial burdens from 1998 to 2002 and rural household migration behavior, using the baseline estimating equation as follows: where h indicates rural households, v villages, and p provinces. The dependent variable Migration h, v, p is measured by the proportion of migrant workers among laborers, for each rural household. Decrease h, v, p is a proxy for the amount of taxes and fees that had been relieved in 2002, relative to what was paid in 1998, varying across rural households. Therefore b 1 is the coefficient of our research interest, denoting the impact of the reduction in a rural household's financial obligations, on its strategic decision-making pertaining to laborers' migration.
Migration98 v, p indicates the migration rate of laborers for village v in 1998, operating as a rough baseline given that the household level estimation of migration in 1998 is not available. The vector X v, p denotes a set of village level covariates that are likely to correlate with migration, while the vector X h, v, p is a series of household level covariates, as we discussed in Section 3.4. h p stands for province fixed effects, which are included to capture some province specific Taxation and migration of peasants 311 characteristics that may influence migration, such as economic development and local governance, to take into account the regional disparity in terms of population outflows. e h, v, p is the error term. We use OLS to estimate the equations, although our empirical results hold when tobit models are employed to account for the truncation of our dependent variable between 0 and 100.
We report the baseline estimates of Equation (1) in Table 1. Column (1) reports the estimate of the correlation between financial burden decrease and migration, for specification that only includes province fixed effects. The estimate is negative as we expect, however small in magnitude and statistically insignificant. Columns (2) shows that the estimated coefficients for financial burden decrease stay statistically insignificant, after controlling for baseline migration rate in 1998 at the village level.
The estimated coefficient for the decrease of financial burdens is negative and statistically significant at the 10% level in column (3) and column (4), of which the specification additionally controls for the household and village level covariates. This is consistent with our hypothesis that the proportion of labor that rural households allocate to migration work negatively correlates to the alleviation of tax burden. In terms of magnitude, reducing financial burdens by 100 RMB from 1998 to 2002 decreases the rural household labor migration rate in 2002 by 0.2 percentage points, holding the village level migration rate in 1998, as well as province, village and household characteristics constant. To assess the magnitude of this implied baseline estimate of the effect of financial burden decrease on migration behavior, we note that the sample mean of rural household financial burden is approximately 362 RMB in 1998 and the average rural household migration rate is 16.81% in 2002. Therefore, for a typical rural household that bears the sample mean financial burden in 1998, our baseline estimate implies that reducing its financial burden by 10% from 1998 to 2002, leads to a decrease in the proportion of migrant workers in household labor by approximately 0.07 (362 Â 10%Â0.002) percentage points, which equals approximately 0.42% (0.07/16.81) of the sample mean in 2002.

Instrumental evidence
Some concerns arise when the association between the changes in financial burdens and migration behavior based on the estimates of Equation (1) is interpreted as a causal relationship. Specifically, we are concerned about possible measurement errors pertaining to our proxies for both the tax-and-fee expenditures and migration behavior, which could probably attenuate our estimation toward the null. Moreover, the residual term in Equation (1) might not necessarily be exogenous to the reduction in financial burdens, giving rise to the omitted variables problem. A notable case is the quality of local governance, which is likely to promote the reduction in tax-and-fee burdens on the one hand, and deter rural-to-urban migration through higher quality of public goods provision on the other, in which case our estimation of Equation (1) would be biased upwards. Moreover, the village leaders might treat rural households under their governance discriminatorily, based on their economic status, political capital, and their personal relations with village cadres, among others. If rural households that receive preferential treatments tend to migrate less and tend to bear lower tax initially, therefore enjoying smaller amount of reductions in their financial burdens, then the emerging omitted variables problem could bias our estimation of Equation (1) downwards. Furthermore, if a better local financial situation facilitates public goods provision and dampens migration, it could bias our baseline estimation in directions that are ambiguous ex ante. On one hand, estimates based on Equation (1) may be biased upwards, if affluent local finance improves the alleviation of financial burdens; on the other hand, downward bias emerges if in the villages with large public revenue, the financial burden for rural households was not heavy in the first place in 1998, leaving little space for alleviation. The possibility of the latter arises when the village committee considers radical fundraising from rural households as unnecessary and politically costly. In sum, we should be cautious about the endogenous nature of our key independent variable, the financial burden decrease, as it may result in biased estimates from Equation (1).
Although the reduction in the financial burdens is endogenous, the tax-for-fee reform offers a policy intervention to the economic opportunities of Chinese rural households for the following reasons. First, the tax-for-fee reform is a from-top-to-bottom campaign, for which the provincial level governments basically made all the decisions regarding to the depth and the scope of the reform, leaving the grassroots cadres and rural residents very little space for negotiation. Second, the tax-for-fee reform is a campaign that applies to all. As long as the reform was adopted in a village, a new tax-and-fee scheme would apply to all rural households, regardless of any household or individual characteristics. Third, the expansion of the tax-for-fee reform was largely, if not entirely, unanticipated by the rural households. As described in Section 2, after a whole year of stagnancy due to the State Council's decision to postpone the experiment, the central government unexpectedly extended the tax-for-fee reform to 20 provinces in 2002. All in all, the tax-for-fee reform is plausibly uncorrelated with most of the other factors that may affect the labor allocation of rural households; Also, it is unlikely that rural households would change their migration behaviors in advance with the knowledge of future implementation of the reform. We thus employ the implementation of the tax-for-fee reform between 1998 and 2002 as the instrumental variable for the decrease in financial burdens, to investigate the policy impact of the reform on alleviating the financial burdens levied upon rural households, and then examine its effect on rural-to-urban migration.
A widespread consensus that helps strengthen the validity of this instrumental variables strategy is that, the tax-for-fee reform achieved little except for the relief in peasants' fiscal burdens, neither overturning the county's structural policy bias against the peasantry, nor fundamentally improving rural governance (Yep, 2004). Previous studies even found that the reform eroded village-level public revenue and undermined public services (Kennedy, 2007;Luo, Zhang, Huang, & Rozelle, 2007). In this regard, our instrumental variables strategy probably yields a lower bound of the effects from rural tax cut, if peasants tend to migrate out when public services fall.
We begin our instrumental variables approach by comparing the household-level calculations of financial burdens in 2002 to those in 1998, in the villages that experienced the tax-for-fee reform relative to the villages with no such policy change. Panel A of Table 2 shows that tax-Taxation and migration of peasants 313 for-fee reform moderately raises rural taxes in exchange for tremendous reduction of fees, overall resulting in substantial alleviation of rural financial burdens. As reported in Column (3) of Panel A, the fees paid by an average household in villages that have experienced tax-for-fee reform are reduced by approximately 194 RMB although the taxes are slightly increased by approximately 31 RMB. Altogether, tax-for-fee reform relieves the financial burden by approximately 163 RMB, which accounts for a reduction in household taxation rate-the share of financial obligations in household yearly income-of approximately 3.22 percentage points. Panel B of Table 2 shows that financial burden also declined from 1998 to 2002 in villages with no tax-for-fee reform, again as a result of the decrease in fees, however to a much lesser extent. This is probably due to the central government's efforts to mitigate local fees through means other than tax-for-fee reform. As shown in Column (3) of Panel B, the financial burden of an average rural household in villages that did not experience the tax-for-fee reform is decreased by approximately 49 RMB, accounting for a reduction in household taxation rate of approximately 1.32 percentage points. Taken together, Panel A and Panel B in Table 2 illustrate that the tax-for-fee reform leads to greater mitigation in financial burden, 4 namely, there is a positive relationship between tax-for-fee reform and financial burden decrease.
Conceptually, our identification strategy compares migration behavior in the villages where tax-for-fee reform had been implemented during 1998 to 2002 to the villages where it had not. Therefore, before turning to the instrumental estimation, we present the estimates of the reduced-form effects of the tax-for-fee reform on rural household migration outcomes. The independent variable here is a dummy, coded as one if the village carried out the tax-for-fee reform between 1998 and 2002, and zero otherwise, based on the information that CHIP 2002 gathered in the investigation of village cadres. The results presented in Panel A of Table 3 reveal that, with or without the control variables, there is a strong negative relationship between  (1) and (2) report means, with standard deviations in parentheses; Column (3) reports changes between 1998 and 2002, with standard errors in parentheses; Total financial burden is the sum of taxes and fees; Household taxation rate is measured by the share of household yearly income paid as taxes and fees. ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1. the tax-for-fee reform and rural households' migration behavior. As shown in Column (4), in the model that includes a whole set of controls, the implementation of the tax-for-fee reform is associated with a reduction of household migration rate of 4.67 percentage points, which then is translated into 28% when evaluated at the mean household labor migration rate in 2002. As our explanatory variable of interest does not vary across individuals but rather vary at the village level, given the potential for within-village correlation of the residuals, we report in square brackets the standard errors adjusted for clustering effects of observations of the same village. Although this method produces larger standard errors, the effect of the tax-for-fee reform on migration decision stays statistically significant at the 5% level across all columns. We then simultaneously estimate a system of two equations using 2SLS, instrumenting for the financial burden decrease with the tax-for-fee reform. The first stage equation is as follows: where Reform v, p is a dummy variable that indicates whether village v had performed the taxfor-fee reform during 1998-2002, and all other variables have the same definitions as in Equation (1), namely our second stage regression. In order to address the concern of whether the exclusion restriction assumption is satisfied, we perform a number of regressions to predict a comprehensive set of village and household characteristics, including the labor migration rate at the village level as well as the household financial obligations in 1998, using the implementation of the tax-for-fee reform, as shown in Supplementary materials. None of these confounding factors seems to correlate with the reform, providing further confidence in the validity of our instrumental variable strategy. The first stage estimates in Panel C of Table 3 provide evidence for a strong positive correlation between the instrument, namely the tax-for-fee reform, and financial burden decrease between 1998 and 2002. The Kleibergen-Paap F-statistic for the excluded instrument ranges from 37 to 41 across the models, showing that it is very unlikely that our estimates are biased by weak instrument problem. In terms of magnitude, the estimated coefficient in Column (4) suggests that the reduction of rural household financial burden is significantly more pronounced in villages that experienced the tax-for-fee reform, by approximately 99 RMB. Evaluated at the sample mean of financial burdens in 1998, the tax-for-fee reform is predicted to alleviate total taxation by approximately 27%. To gauge the plausibility of this effect, it is useful to compare the magnitude to estimates from other studies. According to Yep (2004), only after the first year of tax-for-fee reform, the financial burden for rural households throughout Anhui province was trimmed by 31%. It appears from this comparison that as intended by the central government, tax-for-fee reform does achieve tremendous success in the relief of rural tax-and-fee burdens. For similar reasons as we described when presenting reduced-form estimates, we also report standard errors adjusted for clustering effect in square brackets. This adjustment inflates standard errors, however the positive relationship between the tax-for-fee reform and financial burden decrease stays statistically significant across all regressions.
Panel B of Table 3 presents 2SLS estimates of Equation (1). Unlike the baseline estimates reported in Table 1, the 2SLS estimates remain fairly stable in magnitude and statistically significant as we introduce a large set of controls across Column (1) to (4). According to the estimates using the full set of covariates reported in Column (4), a 100 RMB decrease in financial burden reduces the proportion of migrant workers among rural household laborers by 4.7 percentage points, an effect that is statistically significant at the 5% level. This implied 2SLS estimate of the effect of financial burden decrease suggests that, for a rural household that bears the average financial burden in 1998, a 10% decrease in its financial burden (36.2 RMB) reduces its migration rate by approximately 1.70 percentage points, which is approximately 10.11% of the sample mean rural household migration rate in 2002. 5 Thus the underlying changes in migration behavior associated with rural financial burden alleviation are not only statistically significant, but also substantively meaningful.
Comparing with the baseline estimates reported in Table 2, the 2SLS estimates are more than 20 times larger, suggesting that our baseline estimates are possibly biased downwards by the measurement errors in our calculations of financial burden decrease and household migration. A plausible alternative explanation is that rural households that were treated more favorably by village leaders may be less likely to benefit from the reform and migrate out simultaneously; or that affluent local financial budgets may constrain the space for burden mitigation on one hand, and discourage migration on the other, as we discussed earlier.

Seemingly unrelated regressions
A natural concern over the validity of our estimation strategy is that rural households' migration decisions may be correlated with other options in terms of family labor allocation. As rural migrants suffer from inferior work conditions and unequal pay, not to mention they are denied access to public services in the cities due to the hukou system, Chinese peasants usually consider migration as the second-best choice that they settle for only after they fail to achieve comparable local employment (Song & Knight, 2003;Zhao, 1999). Specifically, rural laborers' willingness to migrate may be offset by the availability and attractiveness of local non-farm economic opportunities. To address this concern, we employ three simultaneous equations to investigate 316 B. Zheng and Y. Gu alternative rural household labor allocation choices, including migration, local non-farm employment, and farming. Given that the disturbance terms are likely to be highly correlated across equations, we perform the seemingly unrelated regressions (SUR) method to estimate the following equations simultaneously: where Y 1 , Y 2 and Y 3 stand for the shares of rural household laborers who were migrant workers, local off-farm workers and farmers in 2002, respectively. Dec is an abbreviation for Decrease h, v, p , denoting the amount of financial obligations that had been reduced during 1998-2002, while C refers to the whole set of covariates that are included in Equation (1). We again use the tax-for-fee reform as an instrument to yield policy-induced variations in the reduction of rural financial burdens between 1998 and 2002. Instrumental estimates of the SUR model based on Equation (3) are reported in Columns (1), (5) and (6) in Table 4 respectively. We omit the first-stage regression results in Table 4, as they are identical to the estimates in Panel C, Table 3. We only report the SUR estimates for the most extensive specification, although our results are robust to the reduction of any controls. The negative and statistically significant coefficient in Column (1) provides confirmation of the braking effect of financial burden alleviation on rural-to-urban migration. However as reported in Column (5), we observe no relationship between financial burden decrease and rural laborers' engagement in local nonfarm employment: the coefficient is negative and statistically insignificant. In contrast, as the tax-for-fee reform mitigates financial burdens, rural laborers flow back to work on their family farm plots, which is manifested in the strong positive correlation between financial burden decrease and farming behavior, as shown in Column (6). 6 The estimates of our SUR model help relieve the concern that the tax-for-fee reform may yield some unintended policy effects that foster rural non-farm job opportunities and therefore decelerate migration, in which case our instrumental estimation strategy may violate the exclusion restriction assumption. They also suggest that our analytical finding is robust to the concern that the slowdown of rural-to-urban migration may correlate with the development of rural non-  (Lai, 2002).

Heterogeneous effects
Motivated by the concern that our measure of the migration rate at the household level could be unstable due to small denominators, which are the numbers of household laborers, we change the unit of analysis to an individual, or more specifically a rural laborer, to assess whether the financial burden decrease in a rural household would influence an individual laborer's migration behavior. We estimate the following equation using binomial probit model: where Migration i, h, v, p is an indicator variable that equals one if rural laborer i was a migrant worker, and zero otherwise. The vector X i, h, v, p denotes a set of labor characteristics that may correlate with individual migration behavior, including gender, age, years of schooling, and Communist Party membership. All other variables are defined as before. We again exploit the tax-for-fee reform to instrument for the financial burden decrease, with the second-stage results shown in Column (1), Table C1 in Supplementary materials. The negative coefficient of financial burden decrease is statistically significant at the 1% level, even after the correction of standard errors for the within household clustering effect of observations, suggesting that our empirical finding is not sensitive to the household level measure of the migration rate.
Recall that we find rural households transfer their laborers from migrant work back to agriculture after the tax-for-fee reform, according to our instrumental SUR estimates. This is plausibly because rural households reevaluate whether the marginal returns to migration are still worth the cost, when the reform alleviates their financial burdens and improves post-tax income. While the existing literature suggests that the tax evasion effect is most pronounced when migration cost is low (Kleven et al., 2013), we conjecture that in the case of rural China, the adverse effect of relief in financial burden should be more noticeable on the migration behavior of rural laborers with lower marginal returns in the urban labor market. To the extent that the data allow, we test this conjecture by examining the heterogeneous effects of financial burden decrease on migration by rural laborers' age, given that the migration costs rise as age increases (Bodvarsson, Hou, & Shen, 2014;Schwartz, 1976). While allowing for this heterogeneity, Equation (4) becomes: where all other variables take the same meanings as in Equation (4), except that Age i, h, v, p refers to the age of rural laborer i. We then instrument for Decrease h, v, p and Decrease h, v, p Â Age i, h, v, p with the tax-for-fee reform, and the interaction term between the reform and individuals' age, respectively, to establish the causal inferences of b 4 and b 5 : The results are plotted in Figure 1, in which the horizontal axis indicates the age of rural laborers, and the vertical axis measures the change in migration behavior corresponding to financial burden decrease, conditional on province fixed effects and a series of village, household, and individual characteristics (namely, b 4 þ b 5 Â Age i, h, v, p Þ: Our estimation of Equation (4) suggests that the overall treatment effect of financial burden decrease under the homogeneous effect assumption is -0.002, statistically significant at the 1% level (Table C1, Supplementary materials). However, the downward trend of the solid line in Figure 1, together with the dotted lines indicating the 95% confidence intervals of which the standard errors are clustered at the household level, shows that the assumption of homogeneous effect does not hold true: the inhibiting effect of financial burden alleviation on migration is strongly contingent on the age of rural laborers. For rural laborers age 35 or below, the effect of financial burden decrease stays statistically insignificant, although it starts to be negative when age reaches 24 and above. In contrast, the impact of reducing financial obligations is exclusively pronounced among rural laborers over the age of 35, and one can see that the older the rural laborer is, the larger the negative effect becomes. Overall, the findings in Figure 1 confirm that the effect of the rather modest financial burden relief is driven largely by rural laborers with lower marginal returns to migration, in this case older laborers, distinguishing our research context from tax-fleeing migration among high-skilled and high-income laborers that was widely discussed in the existing literature.
In light of this finding, we speculate that the effect of financial burden relief may vary on different types of migration with unequal economic and psychological costs. To explore these potential heterogeneous effects, we further disaggregate rural-to-urban migration based on the administrative borders that laborers cross into three ranks, including migration to nonnative townships (that are within native counties), migration to nonnative counties (that are within native provinces), and migration to nonnative provinces. We then extend the SUR model in Equation (3) to distinguish the impacts of financial burden decrease on three categories of migration, as well as local non-farm employment and farming. The estimates reported in Columns (2)-(4) of Table 4 show that, the decelerating effect of financial burden alleviation is only significant on migration across provincial boundaries, however the impact is muted for migration within native provinces. Assuming the cost of migration increases as it spans higher level administrative boundaries (Carrington, Detragiache, & Vishwanath, 1996;Lucas, 2001), this analysis confirms that the negative effect of financial burden mitigation during the tax-forfee reform is isolated to migration behaviors that are economically or psychologically more costly, namely, those that bear lower marginal returns. Taken together, our findings of the heterogeneous effects across age groups and migration types provide evidence for our previous conjecture, that tax cut deters migration mainly because relative to the rising rural income levels, the marginal returns to migration become less attractive to peasants. This is especially true for migration behaviors bearing higher costs, such as those across provinces or at older ages.

Concluding remarks
In this paper, we analyze the effects of taxation on the rural-to-urban migration of Chinese peasants. We put particular emphasis on the tax-for-fee reform, which posed a policy intervention on rural tax burden, to examine its impact on rural households' decisions to migrate. Our investigation has yielded three main results. First, the rural tax cut, as a result of the tax-for-fee reform, significantly reduces rural-to-urban migration. In terms of magnitude, a 10% alleviation in rural tax can lead to a 10% decline in migration. Second, it is unlikely that the reduction in migration is driven by unintended policy effects of the tax-for-fee reform in fostering rural nonagricultural job opportunities. As the government mitigates rural tax, peasants tend to farm more and migrate less, leaving their local non-farm employment unaffected. Third, migration across higher level administrative boundaries or at older ages is more responsive to taxation.
To the best of our knowledge, our study is the first to show that tax-induced migration does exist among low-skilled and low-income laborers in developing economies. The results of this study suggest that future research should carefully examine the impact of tax policies not only on labor supply, but also on labor mobility. Joining a small and recent literature, our study shows that the migration behaviors of laborers at both ends of the income curve are very sensitive to the changes in tax liability. Estimating the elasticities of migration to taxation is therefore very important for policy debates on optimal tax formulas, and their impacts on the redistributive power of governments, as well as the healthy development of labor markets. Notes 1. They are Anhui, Hebei, Inner Mongolia, Heilongjiang, Jilin, Jiangxi, Shandong, Henan, Hubei, Hunan, Chongqing, Sichuan, Guizhou, Shaanxi, Gansu, Qinghai, Ningxia, Jiangsu, Zhejiang, and Shanghai. Provincial governments were given the autonomy to determine whether to conduct the reform in the entire province or just a part of the jurisdiction. 2. Our results stay robust if we treat rural laborers that work across townships but live out of their households for less than 180 days in 2002 as local non-farm workers. Our definition of migration does not involve the duration of stay at the destination, since the crux is that the formerly predatory taxation in rural China pushed peasants out to search for better income opportunities. 3. 1998 is chosen as the benchmark mostly because of data availability. In addition to rural households, the CHIP 2002 survey instrument was also administered to the leaders of all 961 villages to gather information on the implementation of the tax-for-fee reform, as well as some basic village characteristics in 1998, including the labor migration rate at the village level. 4. The difference is 114 RMB (163-49), and is statistically significant at the 1% level. 5. To suggestively put these numbers into perspective, the elasticities to the net-of-tax rate of the number of foreign soccer players in Kleven et al. (2013), and of the number of foreign superstar inventors in Akcigit et al. (2016) are both around one. 6. Baseline SUR model without an instrumental variable produces qualitatively identical results. Moreover, the correlation of residuals is -0.41 between migration and local non-farming, and -0.37 between migration and farming, suggesting that whatever unobserved factors that lead to migration tend to negatively influence local employment, especially off-farm occupation. This verifies our assertion that rural households make laborallocation decisions based on comparing marginal gains and losses of migration and its local alternatives. 320 B. Zheng and Y. Gu