4.1. Likelihood of Participation in Grain for Green
We employ econometric models to investigate factors affecting participation in Grain for Green. We distinguish between extensive and intensive margins of participation, where the extensive margin refers to whether the household is enrolled in the program, and the intensive margin refers to the area of land enrolled in the program conditional on enrollment. Many factors affect participation, including age, gender, education, household size, land area, and so on. As we focus on whether being an ethnic minority household affected participation, all other characteristics of households and land being equal, we included as explanatory variables ethnicity and all other observed characteristics of households and their land holdings. Village dummies are also included in the regressions in order to compare households in the same village.
For the extensive margin, we employ a probit model to address the problem of a dummy dependent variable, using the full sample of 119 households before the program (in 1997). For the intensive margin, we employ a multiple linear regression (MLR) estimated by ordinary least square (OLS), using the subsample of 49 participants before the program. Regression results are listed in
Table 3.
The first two columns show the results of the extensive margin. Column (1) includes all factors affecting participation except for ethnic characteristics, while Column (2) includes ethnic characteristics. The regression results show that, ceteris paribus, the probability of ethnic minorities participating in Grain for Green is 69% higher than the probability of ethnic Han participating. Compared with Column (1), the coefficients of other variables change, which illustrates the possible estimation bias if the ethnic factor is ignored. For example, when the ethnic factor is omitted, the effect of the subsidy on participation is overestimated (0.0069 and statistically significant in Column (1), compared with −0.0026 and not statistically significant in Column (2)).
Similarly, we estimate the intensive margin in Columns (3) and (4) without and with the variable of ethnic characteristics, respectively. The coefficient of ethnic minority is estimated to be negative, and not statistically significant. The coefficient is not economically significant either. Because compared with the average area of land for the treatment group before the program (2.902 ha), the estimated effect, −0.0288 ha, is small. This indicates that enrolled ethnic minorities are similar to enrolled Han in terms of area of enrolled land. The possible explanation is that the program is quasi-voluntary [
14]. There were program officers who were in charge of selecting participants and enrolled land, and they made decisions based on characteristics of each household’s land holdings. Since minority is not explicitly targeted, it is not surprising to see similar land enrollment between the majority and the minorities with similar characteristics of land holdings.
Based on the estimation results on the extensive and the intensive margins, we expect to find a positive total effect on the area of enrolled land. This is confirmed in Columns (5) and (6), the results from a tobit model using the full sample of 119 households. Tobit model is chosen to take account of the truncation issue that the area of enrolled land for non-participants is zero. The results show that the minorities enrolled 0.2776 ha more cultivated land in the program, compared to the majority with the same characteristics in the same village.
Besides ethnic characteristics, other factors are also found to affect participation likelihood or enrolled land area. They include number of children in the household and subsidy level. The estimation results are as expected: Households with more children have a higher incentive to participate in the Grain for Green program, as it can free labor from plowing and the household can hence put more labor into taking care of children. Households with a higher level of subsidy also have a higher incentive to enroll in the program. These are both household characteristics. In contrast, the characteristics of the household head (such as education, age) do not seem to have much impact on the participation decision, except for the ethnicity.
In sum, the difference in the participation between the minorities and the majority is a total effect of the differences in the observed characteristics between the two groups, as well as the differences in the unobserved characteristics captured in the ethnic minority dummy.
Table 4 displays the characteristics of households, the differences in participation between the two groups, and the contribution of each characteristic to the difference in participation. It shows that about 70.6% of the ethnic minorities and 19.4% of the ethnic majority are enrolled in the program, and that the difference is mainly driven by the difference in ethnicity. Ethnic minority alone contributes 68.68 percentage points to the difference in participation, which means that a minority household is 68.68% more likely to participate in the program, compared to a majority household with all other household and land characteristics the same. Compared to ethnicity, the contribution of other characteristics is much smaller.
4.2. Effect on Off-Farm Labor of the Grain for Green Program
We investigate the effect of
Grain for Green on off-farm labor supply using a difference-in-difference method (DID). A DID estimator is generally obtained as:
where
=
and
is the off-farm labor employment of household
in group
in year
, where
equals 1 for the year after
Grain for Green becomes effective and 0 for the year before, and s equals 1 for households taking part in the program, referred to as the treatment group, and 0 for households not in the program, referred to as the control group. We estimate the treatment effect of the program,
, in the following regression:
where
is the group dummy, which equals 1 if household
is in the treatment group, and 0 if in the control group;
is the year dummy, which equals 1 if the observation is in the year after the program is in effect, and 0 if before; and
is the treatment indicator, which is the interaction term of the group dummy and the year dummy. The coefficient of the treatment indicator is the treatment effect of the program on
.
includes the observable variables which affect the off-farm labor supply of a household, including household characteristics (e.g., household size, education, etc.) and land characteristics (e.g., land area, distance from land to residence/road, etc.).
is the residual term.
In order to interpret these estimates as unbiased measures of the
Grain for Green program, some assumptions must hold. One key assumption is the conditional unconfoundedness, which requires the similarity in outcome trajectories of the treatment and the control groups absent of any treatment effect. This assumption is to ensure comparability of the treatment and the control groups. If the two groups are not comparable, for example, households who prefer off-farm work are more likely to participate the program and the factors behind the preference are not observable and therefore not controlled in the regression-based DID, the DID estimate in this paper would overestimate the effect of the program on off-farm labor supply. One way to assess the plausibility of the conditional unconfoundedness assumption is to compare the pre-treatment trajectories of the two groups. Given that we have only one period data before the treatment, we are not able to compare the pre-treatment trajectories. Therefore, we rely on the implementation of the program and previous literature to assess the plausibility of the assumption. As we have discussed in the previous section, the program is in practice quasi-voluntary [
14]. The self-selection problem is therefore alleviated. In addition, the program officers who were in charge of selecting participants and enrolled land made decisions based on characteristics of each household’s land holdings, which were all observable. The regression based DID employed in this paper include all these characteristics as control variables, and therefore enhance the plausibility of the assumption of conditional unconfoundedness. Furthermore, Uchida et al., 2009 [
14] studied the effect of the same
Grain for Green program in China and adopted both traditional DID and DID matching methods. They found that the results with and without matching were similar. It implies that the control and the treatment groups in this program are comparable. The other key assumption for the validity of DID method is the assumption of stable unit treatment values (SUTVA), which requires that the off-farm labor supply of one household is independent of the treatment status of other households. This assumption is plausible, because the households are independent in household production and cultivated land management. One household’s decision on off-farm labor supply is not likely to be affected by other households’ decisions of participation.
The regression results are listed in
Table 5, where Columns (1) and (2) show the values with and without village fixed effects, respectively. Since we are especially interested in the heterogeneous effect of the program on Han and ethnic minorities, we add the interaction term of the treatment indicator and the ethnic dummy variable. The coefficient of the interaction term shows the difference in the treatment effect between Han and ethnic minorities. The results show that the effect of the program on Han and ethnic minorities differs (the difference is −0.5958 and −0.9002 without and with village fixed effects, respectively) and that the difference is statistically significant. The effect on Han is positive and statistically significant (0.5642 and 0.7817 without and with village fixed effects, respectively), while the effect of the program on ethnic minorities is negative and not statistically significant (0.5642 − 0.6342 = −0.07,
p-value = 0.7404 and 0.7817 − 0.9299 = −0.1482,
p-value = 0.5870). This indicates that Han have increased off-farm labor supply after participating in the program, while ethnic minorities have not.
In
Table 5, Column (3) through Column (6) present treatment variables providing more detailed information on the treatment. In Columns (3) and (4), the treatment variable is the ratio of enrolled land to total land area owned by the household. In Columns (5) and (6), the treatment variable is the amount of subsidy a household received for participation. The two continuous treatment variables are not likely to bring self-selection problem, because, as discussed above, the program is quasi-voluntary and the selection criteria are observed and controlled in the regressions. In addition, the subsidy per unit of land is set by the government and exogenous to households. The results from the regression with the continuous treatment variables are consistent with those from the regression with the binary treatment variables (whether to participate) in Columns (1) and (2): after participating in the program, Han have increased off-farm labor supply, while ethnic minorities have not.
Besides the
Grain for Green program, there are some other factors affecting the off-farm labor supply. In
Table 5, the estimated coefficients of ethnic minority are negative and statistically significant in the regression without village fixed effects, while they are positive (not statistically significant) with village fixed effects. One possible explanation is that minorities living in isolated villages, not intermingled with Han, tend to provide less off-farm labor on average compared with Han. As shown by our sample, the average off-farm labor per household is 0.62 for the minorities in a village where there is no majority resident, while the figure is 1.04 for the majority, and the difference is statistically significant. Without village dummies, this difference is captured by the coefficient of ethnic minority. Therefore, the estimated coefficient is negative. The reason behind the difference in off-farm labor employment could be that ethnic minorities are relatively isolated in society, and therefore may have less off-farm job opportunities, or they are less willing to participate in the off-farm labor employment [
19]. However, when including village dummies, the comparison is between the households from the same village. Our sample shows that ethnic minorities living intermingled with Han have similar off-farm labor employment with Han households in the same village. Therefore, the coefficient of ethnic minority is expected to be not statistically significant when village dummies are included.
In addition, estimated coefficients of education of the household head are positive and statistically significant in all columns; estimated coefficients of land area are all negative and significant; and coefficients of distance from land to residence/road are also negative and significant in all columns. These results indicate that, ceteris paribus, households with a better educated head provide more off-farm labor on average, as higher education provides more off-farm job opportunities; households with more land and land farther away from residence/road provide less off-farm labor on average, as such households require more on-farm labor.
To distinguish between the treatment effects on different types of off-farm labor, we employ the same regression based DID on low-skilled off-farm labor employment and off-farm labor employment that requires higher education separately. The regression results are presented in
Table 6. It shows that the estimated treatment effect on low-skilled labor is at the same scale with that on total off-farm labor shown in
Table 5, while the treatment effect on the higher-education labor is much smaller and not statistically significant. It implies that labor forces freed from the program are more likely to obtain a low-skilled off-farm job than a job that requires higher education.