Using comparative method to explore the effects of poverty reduction policy, a general method on the longitudinal comparison can be used, comparing the difference of poor counties’ economic development level before and after the policy. But it has the influence of the time trend; maybe not have the policy, poverty county’s economic development will also changes over time, the longitudinal comparison is difficult to accurately judge. Of course we can use a horizontal comparison, comparing the difference between poor counties’ and non-poor counties’ economic development level after this policy. But this ignores the individual differences. Horizontal comparison is also difficult to examine the effects of poverty alleviation policy accurately. Hence, this article will use the double difference method (difference in difference), considering the individual difference and time difference, to analyze the effects of policy exactly.

We construct a dummy variable N P i   . If i county is poor county (treatment group), the value is 1, opposite is 0 (control group). Construct a time dummy variable d t again, it is assigned 1 after the implementation of the policy, before is 0. So for poor counties, the effects of policy can be expressed as:

δ = E ( Y i t 1 | N P i = 1 , d t = 1 ) − E ( Y i t 0 | N P i = 1 , d t = 1 ) = E ( Δ Y i 1 | N P i = 1 ) − E ( Δ Y i 0 | N P i = 1 ) . (1)

E is the mathematical expectation in the equation, Y i t 1 is the observations if i county participate policy at t time. Y i t 0 represents not participate policy. Δ Y i 1 shows the difference before and after the policy implementation if i poor county participate policy, Δ Y i 0 shows the difference before and after the policy if i poor county doesn’t participate policy. It is clear that the latter could not be observed, therefore, we can use the non-poor county’s difference before and after the policy to replace it. That is E ( Δ Y i t 0 | N P i = 1 ) = E ( Δ Y i t 0 | N P i = 0 ) , so, Equation (1) is transformed into:

δ = E ( Δ Y i 1 | N P i = 1 ) − E ( Δ Y i 0 | N P i = 0 ) . (2)

Theoretical model should be designed as follow:

Y i t = α + γ N P i + λ d t + δ N P i d t + ε i t . (3)

In the Model (3):

E ( Δ Y i 1 | N P i = 1 ) = ( α + γ + λ + δ ) − ( α + γ ) = λ + δ . (4)

E ( Δ Y i 0 | N P i = 0 ) = ( α + λ ) − α = λ . (5)

E ( Δ Y i 1 | N P i = 1 ) − E ( Δ Y i 0 | N P i = 0 ) = ( λ + δ ) − λ = δ . (6)

As we can see, Equation (6) and (2) is consistent. Accordingly, in this Model (3) based on “difference in difference”, the estimated coefficients of interaction δ measure the effects of poverty reduction policy.

Figure 1 shows the economic growth of two groups over the years. Before the year of 2001, poor counties are lower than the non-poor counties, the difference of economic growth rate between the two groups remained at almost 2.5%. But in the policy year (2001), poor counties faster, gap narrow 1.5%, almost close to the non-poor counties in year 2002, poor counties even exceed them. After 2005 poor counties economic growth is relatively slow, but the gap is diminishing. It is clear that two groups showed obvious difference before and after the implementation of policy.

Two groups remained the same change trend before policy, which is consistent with hypothesis (2). So we can set the model as:

Y i t = α + γ N P i + δ N P i d t + θ X i t + λ t + ε i t . (7)

Y i t is per capita GDP. α indicates Intercept, N P i is poor county dummy, d t is time dummy. X i t indicates the other control variables, including three groups: the first group is the geographical control variable mountain, minority county and the revolutionary areas; the second is administrative region area, rural population rate; the third group is the initial level of per capita GDP. λ t istime fixed effect. ε i t is residual. In order to avoid heteroscedasticity and serial correlation of residual, we will polymerize residual to the county level.

In this paper, data are derived from “China County Statistical Yearbook”, “National City and County Financial Statistics” and government official document (List of 592 poor counties), covers the period from 1997 to 2010. Considering the comparability is very poor between city districts and counties, all the observations are eliminated municipal district. In addition, considering that there are some outliers in the panel data, gross domestic product has been revised. Specifically, use program to identify outliers, and adjust it by the method of interpolation. At the same time, this paper also uses the “shrinking tail method”, only to retain the first percentile to ninety-ninth percentile of the data, to reduce the impact of outliers [16] . According to this, the data of this paper has higher accuracy and representation.

First of all, based on the county-level panel data from 1997 to 2010, we analyze the influence of poverty alleviation policy on economic development in the new period by the Model (7), because of some new poor counties are also benefited in the previous policy, therefore, the impact of “Seven-year priority poverty alleviation program” has been controlled. δ in Model (7) measures the influence of the policy in the new period.

Regression results in Table 1, as we can see, whether or not join control variable, regression results are significantly positive. Specifically, in the first column, other control variables not be added, poverty alleviation policy in the new period significantly improved the poor region’s economic development level, GDP in non-poor counties varied from 84% to 65.7%, higher than poor counties, therefore the gap has narrowed by 18.3%.

Next, in order to explore the time effect of poverty reduction policy, we focus on dynamic change trend of the effect of policy, to examine robustness of the overall effect.

First, we construct 10 dummy variables of year, if the year 200i after policy, value is 1, otherwise 0. Hence, interactive item in Model (7) is decomposed into 10 years after the reform. The model is designed as follow:

Y i t = α + r N P + i δ 1 N P i d t year 2001 + δ 2 N P i d t year 2002 + ⋯ + δ 10 N P i d t year 2010 + θ X i t + λ t + ε i t . (8)

Figure 2 shows the dynamic change trend and significant degree of effect variable coefficients ( δ 1 to δ 10 ) which do not add control variables. As we can see, the estimated coefficients of each year after the implementation of policy are significantly positive, this shows that policy continues to be effective, improve the level of economic development in poor areas every year. On the other hand, we can see that the estimated coefficients have been increasing, which may imply that the resources may be more involved in long-term development projects, rather than short-term item which pursuit an immediate result. This is consistent with regulation issued by National Poverty Alleviation Office.

Next, we test the robustness of the empirical results by adjusting the control group. In the year of 2001, nation determined 592 national poor counties, at the same time also identified 421 provincial poor counties. We take 421 provincial

Note: Robust standard errors in parentheses, *p < 0.10, **p < 0.05, ***p < 0.01.

Note: Robust standard errors in parentheses, *p < 0.10, **p < 0.05, ***p < 0.01.

poor counties as control group. Doing this, on the one hand, we do further observation by reducing the amount of samples. On the other hand, provincial poor counties have not accepted the state support, but accepted the support of the province. Therefore, the two groups are more comparable. In fact, this is also consistent with the assumptions mentioned in the theoretical framework.

First of all, we still control the impact of “Seven-Year Priority Poverty Alleviation Program”, Secondly, taking into account the difference between the two groups in the level of economic development is relatively small, and it is difficult to accurately identify the effect of Policy in the new era. So, we control the initial economic development level in all regressions, in order to control the interference of other factors at large. The results are shown in Table 2, the coefficients of policy effect we have investigated are all significantly positive. It shows that national poor counties grow faster than provincial poor counties under the national poverty alleviation policy. Take the first column as an example, the per capita GDP of the national poor counties increased by 0.410 relative to the provincial poor counties.