16
and obtain an unbiased program impact of α . However, the participants in the Grain for Green program
were not randomly chosen. In the absence of truly randomized experimental program, the coefficient α
may be contaminated by other unobserved factors that could affect a household’s off-farm labor-supply
decisions. Simple comparisons of preprogram and postprogram outcomes for the participants also may
be biased due to temporal trends in off-farm labor markets and/or by the effects of events other than the
Grain for Green program that occurred between the two periods (and affected each household’s off farm
employment). Systematic differences could arise, for example, because households were selected for the
program based on unmeasured household or village characteristics or because earning levels differed
among different segments of the labor markets in which the participating and nonparticipating
households function. In essence, these are all components of the selection bias that is inherent in data
from nonrandomized programs.
The descriptive statistics underscore the bias that can arise if we estimate the program impact
by a simple regression that uses only data from participating households or only data from the
postprogram period. Although the number of participating households that reported off-farm work
increased between 1999 and 2004, off-farm employment rates for nonparticipating households also
increased. One or more factors, such as deepening of the local off-farm labor markets in regions that
host the Grain for Green program, could contribute to households shifting labor to the off-farm
employment market. Hence, to obtain the least biased estimate of the impact of the Grain for Green
program, we hold constant other observable and unobservable time-variant and time-invariant effects as
much as possible.