Does Poverty Relief Breed Corruption? An Evaluation of China’s Poverty Alleviation Program

Abstract

Does poverty relief breed corruption? This question has been examined in several electoral polities, but not in other political settings. This study engages in this academic discussion by proposing a unique corruption mechanism in the case of China. The dual causes of National Poor Counties’ unique exit mechanism and long distance from economic hubs create a unique local poli-economic environment, in which local bureaucrats share the incentive to stay in the program, which in turn leads to an intra-elite alliance and more corruption. I test the empirical implication of these mechanisms by first building up a new county-level corruption dataset for 881 low-income counties,and then leveraging propensity score weighting to evaluate the impact of the third wave (2001–2010) of China’s poverty alleviation program on county level corruption. The results of this test show some evidence that counties participated in the third wave of the program are associated with approximately 4.4–19.7 percent increase in corruption incidence and 10.5–23.5 percent increase in numbers of convicted persons compared to similar non-participant counties. These findings highlight the importance of proper institutional designs in poverty alleviation regimes to lower the corruption incentive of local bureaucrats.

Appendices

 Appendix 1 Regression results using number of convicted cadres at village level, section leadership level, and division director level separately as the outcome variable

Table 7

Table 7 OLS and propensity score weighted regression estimates of NPC’s effect on county-level corruption with other outcome variables

Fuzzy RDD Estimation

At the very start of this research process, we thought RDD is a suitable identification strategy to estimate the causal effect of NPC designation on corruption, because according to the national policy, counties could enter China’s poverty alleviation programs if they were below certain poverty lines. In other words, the assignment rule for NPCs creates a discrete relationship between a county’s pre-program economic performance (the running variable) and its probability of being allowed to participate in the program (the treatment status). Under this condition, RDD is quite applicable to avoid selection bias of in the assignment of NPC status and estimate its effect.⁵⁰ To make things more complicated, due to various technique and political reasons,⁵¹ in the real NPCs designation process, the selection criteria are not perfectly followed. Because no deterministic relationship exists between the treatment status and the running variable, we have to use fuzzy RDD to estimate the complier average causal effect of the program (2001-2010) on the outcome variable, i.e., county-level corruption in (2014-2016), and conceptualize the RD approach in an instrumental variable (IV) framework.⁵²

The “631” index of the third wave comprises of four economic indicators in 2001: the poverty incidence in the county (weighted at 60%), the rural net income per capita (weighted at 30% with different threshold for different counties), and GDP per capita and local government revenue per capita (weighted at 10%). Because the data on county-level poverty incidence in 2000 has not been made public and there is no threshold attached to this variable, this paper only used data on the other three indicators with clear thresholds in 2000 to construct the running variable. Exploiting these three indicators’ respective weights, we computed the running variable—poverty index to approach the original “631” index for each county:

$$poverty index=6\text{Rural net income p.c.}+GDP p.c+"Government revenue p.c."$$

Moreover, to select eligible poor counties into the third wave, the basic poverty line was set at 1300 Yuan for rural net income per capita, 2700 Yuan for GDP per capita, and 120 Yuan for government revenue per capita for ordinary counties, while the standard rose to 1500 Yuan for rural net income per capita for RMF counties. Based on these thresholds, we computed two cutoff points respectively for ordinary counties and RMF counties.

$$The cutoff point for ordinary counties=6 1300+2700+120=10620$$

$$The cutoff point for RMF counties=6 1500+2700+120=11820$$

The data of the three poverty line indicators and RMF status were collected and kindly shared by Cong Qin and Terence Tai Leung Chong from the following sources: the rural net income data are from China National Knowledge Infrastructure (CNKI) database supported by China Statistics Press; GDP and local government revenue per capita are from University of Michigan’s China Data Online database; lists of old revolutionary areas, minority autonomous areas and frontier counties were obtained from An Outline of Chinese Rural Economic Statistics by County 1988 and the CNKI database.⁵³

We used fuzzy RDD to estimate the complier average causal effect of the program (2001-2010) on the county level corruption. Moreover, because the cutoff point for ordinary counties and RMF counties are different, we estimated the effects in two subsamples. But before estimation, we needed to validate the fuzzy RDD design. Figure 5 and 6 show that, in both of the two subsamples, counties below the respective cutoffs are more likely to participate in the NPC program, while counties above the respective cutoffs are more likely to be excluded from the NPC program. But, the selection criteria are not perfected followed, so the fuzzy RDD is a legitimate method in such conditions.

Then we moved on to assess if there is systematic manipulation on the running variable. A manipulation test in this context amounts to testing whether there is a disproportional number of counties are situated above the cutoff point relative to those present below the cutoff. Figure 7 and 8 show that there is some manipulation in RMF counties, but not in ordinary counties. We used a formal RD Manipulation Test (Table 7) to quantify the extent of manipulation and found that: there is no statistical evidence (p = 0.9411) that there is systemic manipulation in ordinary counties; but there is weak statistical evidence (p = 0.095) that there is systemic manipulation in RMF counties.⁵⁴

Fig. 5

Treatment distribution in ordinary low-income counties.

Fig. 6

Treatment distribution in RMF low-income counties.

Fig. 7

Histogram of poverty index for ordinary low-income counties.

Fig. 8

Histogram of poverty index for RMF low-income counties.

Table 8 RD Manipulation Test using local polynomial density estimation

In RDD, visualization of the estimate is its defining feature. But to conserve space, we only present two RD plots (Fig. 9 and 10) . Both of the plots put the number of filed corruption cases on the y-axis and the poverty index on the x-axis, and they show: in ordinary counties, counties just below the cutoff have a higher number of corruption cases than counties just above it; but in RMF counties, no such difference can be discerned. As our next step, we employed nonparametric methods of fuzzy RDD to estimate the causal effect of the NPC program for compliers.⁵⁵

The results are presented in the Table 8. For the purpose of robustness, we showed the estimates using all the 5 corruption indicators respectively for both ordinary and RMF counties. The estimation of bandwidths and effective numbers of observations was conducted using a local-linear estimator with a local-quadratic bias-correction estimate, with a triangular kernel. For ordinary counties, the RD point estimators with each of the five corruption indicators used as outcome variables were always positive but not statistically significant. The same is true for most of the models for the RMF counties. Only the last two models for RMF counties show negative and insignificant effects. These results suggest that, around the cutoff in our artificial RDD, compliers in treatment group (NPC program participants) might increase corruption compared to compliers in control group (non-participants), but this effect is still questionable and unstable in a statistical sense.

We think these insignificant results arise from the fact that the composite “631” index makes it difficult to evaluate it with RDD, because the index didn’t set a clear cutoff point for the poverty incidence variable which weighs at 60 percent in the index. So, while we tried to approach this index with the other three constituent variables, we are still one step away from the real program assignment process. As a result, the fuzzy RDD only utilized a few complier observations and excluded those non-complier observations based on our arbitrarily constructed running variable and cutoff points. In other words, the fuzzy RDD dropped a few useful observations. Thus, we turned to the propensity score weighting method as the main identification strategy to tackle this problem (Table 9).

Fig. 9

RD plot for ordinary low-income counties

Fig. 10

RD plot for RMF low-income counties

Table 9 Fuzzy RD estimates of NPC’s effect on county-level corruption in two types of counties