The Effects of the Sloping Land Conversion Program on Poverty Alleviation in the Wuling Mountainous Area of China

Abstract

The Sloping Land Conversion Program (SLCP) is an important part of the Chinese ‘Priority Forestry Programs’ which has special effects on mountainous areas, ecologically and economically. This study was carried out in the Wuling mountainous area in Hunan Province of China. Based on data collected in a household survey of a sample of 375 households, a quantile regression model, Tobit regression model and weighted least square model were used to analyze the influence of participating in the SLCP on farmers’ sources of income (especially the poorest farmers), and the influence of forest type in SLCP land on farmers’ incomes. It was found that: (1) participating in the SLCP reduced farmers’ crop production income, but increased their forest income and subsidies; (2) participating in the SLCP increased farmers’ total income on average by 11.7 % and that of the poorest farmers by 24.7 %; and (3) choosing economic forest in the SLCP land increased farmers’ forest income, although its influence on total income was not statistically significant. These findings indicate that the SLCP has a positive effect on poverty alleviation in the Wuling mountainous area in the short run. However, the poverty alleviation effect is still unclear in the long run because the effect of the SLCP on changing farmers’ income structures is not obvious by observation, the subsidy term will have expired in 5 or 8 years and the forest revenue from the SLCP is low.

Appendices

Appendix 1: The Problem of Heteroscedasticity

In the linear regression model y = Xβ + ɛ, the Gauss–Markov assumption can be stated as E(ɛ|X.) = E(ɛ) = 0 and Cov(ɛ|X.) = E(ɛɛ ^′) = σ ² I _N . If heteroscedasticity exists, the residual ɛ _i values will have differing variances with regard to different values of i, and thus violate the Gauss–Markov assumption. Although heteroscedasticity cannot lead to estimation bias or inconformity of β _j , the OLS estimator is not valid: the variances of the OLS estimator are no longer minimal, and the variable t test is meaningless. When the model confirms the existence of heteroscedasticity, weighted least squares (WLS) can be used to estimate the model (Wooldridge 2012). The new model is as follows:

$$\sum {W_{i} } \varepsilon_{i}^{2} = \sum {W_{i} } [Y_{i} - (\hat{\beta }_{0} + \hat{\beta }_{1} X_{1} + \ldots + \hat{\beta }_{k} X_{k} )]^{2}$$

(3)

where W _i is the weight to reduce the heteroscedasticity. The new model satisfies the assumption of homoscedasticity, and so the estimator is unbiased and efficient.

Appendix 2: Merit of Choosing the Quantile Regression Model

Least squares estimation only estimates conditional means of y given x (Heckman 1979). For all conditional distributions of y given x, variances, skewness and kurtosis are assumed to be identical. However, the conditional distribution of y can be skewed or fat-tailed. Despite such restrictive and naive assumptions, most applied econometric analyses are concerned with conditional means (Fitzenberger et al. 2002). The quantile regression approach overcomes this limitation of the location shift model by allowing for various forms of the loss function that allow for potential scale shifts. In this study, the distribution of dependent variables shows a heavy tail (Fig. 3), hence estimates of the mean value through the use of OLS are misleading. Although this study took the logarithm of total household income, the assumption of normality still could not be satisfied.

Fig. 3

Frequency distribution of rural household annual income

Another advantage of quantile regression evident in is that quantile regression can provide a more nuanced view of the stochastic relationship between variables, and therefore, a more informative empirical analysis (Koenker and Bassett 1978). Different coefficient estimates at different quantiles would be a manifestation that an Ordinary Least Square model is inadequate to explain the underlying relationship between the variables of interest. In reality, policy-makers care more about the state of the poorest households compared to those of median-income households, and so this study adopts the quantile regression model to estimate the coefficients in various percentiles of y. The quantile regression model is:

$$\hat{y}_{Q} = a_{Q} + b_{Q} x .$$

(4)

Unlike in the linear regression model, $\hat{y}_{Q}$ in this model indicates the estimates of the Q-quantile position (0 < Q < 1) of y under the condition of given x. The estimation method is the weighted least absolute estimation (WLA), which minimizes $\sum {\left| {\hat{y}_{iQ} - a_{Q} - b_{Q} x_{i} } \right|h_{iQ} }$, here:

$$\begin{aligned} h_{i} & = 2Q,\quad \text{if} \;y_{iQ} > a_{Q} + b_{Q} x_{i} ; \\ h_{i} & = 2(1 - Q),\quad \text{if} \;y_{iQ} \le a_{Q} + b_{Q} x_{i} . \\ \end{aligned}$$

(5)

The initial value uses the estimates derived through use of the WLA method (Hao and Naiman 2007).

Appendix 3: Merit of Choosing the Tobit Model

The dependent variables in this study—including forest income, crop production income and nonfarm work income—can all take observed values of zero, which is known as censoring data. Consequently, the use of ordinary least squares (OLS) estimation can result in bias errors. A solution to the problem with censoring at zero was first proposed by Tobin (1958) as the censored regression model, which became known as Tobit model. Another way to solve this problem is treating the censored data as a sample selection process, using the selection equation of the two-stage estimation to correct the sample selection bias, and then estimating the regression model (Heckman 1979).

Endogenous selection bias may occur when SLCP participants are not assigned randomly, for instance, when household participation decisions are determined by some unobservable factors. However, in this study, households in assigned areas had the right to choose whether to participate in the program. The right is generally protected by SLCP regulations which have been implemented since 2002 (Feng et al. 2005). Besides, Linders et al. (2006) demonstrated through the Monte Carlo method that the Heckman selection model was more suitable for handling the problem of dependent variable having a large sum of zero observed data, which is not the case in this study. Thus this study treats participants as being assigned randomly, and adopts the Tobit method to estimate the model (D, E, F, J). The Tobit model is expressed as follows:

$$y^{*} = \beta_{0} + x\beta + u,u|x \sim Normal(0,\delta^{2} ),\quad y = \hbox{max} \left( {0, y ^{*}} \right)$$

(6)

The latent variable y* satisfies the classical linear model assumptions; in particular, it has a normal, homoskedastic distribution with a linear conditional mean. When y* ≥ 0, the observed variable y equals y*, and where y* < 0, y = 0. x is matrix of explanatory variables, β are coefficients, and u is the residual error, which follows the standard normal distribution (Wooldridge 2012).