Econometric analysis

We conducted our analysis of determinants of forest clearance at the household level. Another possibility would have been to analyze the determinants at the plot level. Bekele and Mekonnen (2010) point out that decisions of an action (as to clear land) and its intensity may not be made uniformly for the entire farm (all plots) of a household. Additionally, Saint-Macary et al. (2010) argue that a household-level model is unable to

114

capture the effects of soil characteristics, and other plot-specific variables on adoption (clearance in our case). However, the same authors found a much better prediction power by their household-level analysis than by their plot-level analysis regarding the adoption of soil conservation techniques in northern Vietnam.

Our analysis refer to newly acquired plots. Therefore, we assume that the decisions to clear and where to clear are uniformly made for by the households for each new plot. Furthermore, we are primarily interested in the factors which influence the general decision to clear and therefore refer to a household level analysis. The decision to clear natural forest is measured by a binary variable, which is zero if the household did not clear any natural forest since 1999. It takes on the value one if the household cleared forest since this time. We are interested in how the vector of explanatory variables influences the possibility that the binary dependent variable takes on the value 1. The binary response probit model is estimated by Maximum Likelihood Estimation (MLE) using the computer package Stata 10.

A probit model is defined as

(17) P(y_i 0|x_j)(x_jb),

where P is the probability, yi is the dependent variable (in our case forest clearance between 1999 and 2006), xj is the independent variable (in our case household characteristics), Ф is the standard cumulative normal distribution, and xjb is the probit score. The coefficients of the probit model are difficult to interpret, because an increase in x1 by one unit increases the probit score by b standard deviations. Therefore, we display the marginal effects which are based on the change in probability calculated at the mean.

The marginal effect is given as

(18)     1 x ø(xb)b1, where 1 x   

is the probability of a change in x1. Thus the marginal effect is the probability for an infinitesimal change in each independent continuous variable. For dummy variables the discrete change is reported (STATA 2007).

After modeling the household determinants of deforestation, we employed analysis of how much forest is cleared. The dependent variable “size of plots cleared since 1999” is

115

censored at zero, therefore we used a tobit model. A tobit model is a common model to estimate the relationship for limited dependent variables (Tobin 1958, Godoy, et al. 1998).

One major criticism on the tobit model is that it has the restrictive assumption that all zeros arise from factors as economic and demographic characteristics alone and that the model therefore ignores zero-observations due to respondents non-participation decisions (Wodjao 2007).Wodjao argues that in his study on computer and internet use in the US the observed zeros might have two sources. On the one hand he finds behavioral zeros if people do not own a computer (2007). This case is similar to the case of households that do not clear forest in our study. On the other hand he observes random zeros if people own a computer but do not use it the diary (i.e. survey) day. He tests two kind of models which account for the latter problem. On the one side the Heckman model (Heckman 1979) which addresses the problem of zeros due to non- participation decisions. Its argument is that an estimation on a selected sub-sample (a censored estimation) results in a sample selection bias. The also called heckit model therefore estimates a two step estimation procedure, (1) a full sample probit, followed by (2) a censored estimation. Firstly the probability of a positive outcome is estimated and secondly a conditional equation on the level of participation. In heckit it is possible to use different sets of variables in each step. In the Heckman model it is assumed that zeros mainly arise from respondents self-selection and therefore from his or her deliberate choices (Wodjao 2007). Another possibility to overcome the restrictive nature of tobit is the so called double hurdle model (Cragg 1971). This model assumes two hurdles have to be overcome to observe positive outcomes. It accounts both for the ownership or participation decision and for the random circumstances of its intensity of use. Like in the Heckman model different variables can be used in each of the two equations.

In our case the problem of two separate decisions is less likely, as land is scarce in the region and therefore the decision to clear is closely related to the decision on how much land is cleared. Hence it seemed not necessary to take different sets of explanatory variables . Thus we referred to the traditional tobit approach. Following Cong (2000), the tobit model can be written as

116 yi, if a < yi < b

(19) yi*= a, if yi a

b, if yi  b

where yi is a latent variable, a is the upper limit, b is the lower limit. Instead of observing y, we observe yi*, which is bounded between a and b if yi is outside of those bounds (STATA 2007a). In our case the decision of how much land is cleared is obviously bound to the decision regarding if land is cleared. Not all of the households within the sample cleared forest, thus information on the size of the cleared plots is only available for 41 households (referred to as uncensored observations).

For the tobit model we also displayed marginal effects additional to the coefficients, as the ß coefficients (which express the change in the mean of the latent dependent variable) (Cong 2000), are difficult to interpret.

For tobit models, several marginal effects can be of interest. Therefore, we describe shortly the one we used. In the results section the marginal effects for the expected value of the dependent variable (conditional on being uncensored) are displayed,

(20) E(y*|ay*b) x₁,

where E/x₁ represents changes in the conditional expected value of the dependent variable (Cong 2000), a is the lower limit for left censoring and b is the upper limit for right censoring (STATA 2001).

5.4 Results