24
off-farm work; with a binary variable, we cannot capture changes in time spent on-farm (measured in
days or hours). Since most households (and individuals) in the sample continue to farm while
participating in the program (even when they get local jobs off-farm), we find no statistical effect of
participation in Grain for Green on on-farm labor.
Assessing Selection Bias
Since we are concerned that the preprogram variables for 1999 may suffer from recall bias, we
repeat all of the preceding analyses on the smaller subset of households (n = 27) that changed status
from nonparticipant to participant between 2002 and 2004. In that analysis, we use the same 40
nonparticipating households as the control group. With this subset, while the sample is smaller, the data
are true panel data and are not subject to errors due to recall.
Overall, the findings from the smaller subset are consistent with those from the full sample
(Appendix Tables 1 and 2). The Appendix tables provide the results of the program’s effect on off-farm
labor participation at the individual level. The DID estimates for the subset are slightly larger than the
estimates for the full sample. This consistency between samples suggests that recall bias in 1999 was
limited and/or that the DID approach controlled for bias that existed in both groups.
Discussion
In summary, the DID estimates of the binary indicator for program participation and the
variables for program intensity suggest that the Grain for Green program led to something between a
small and moderate increase in off-farm work among participating households. This finding is in sharp
contrast to two prior studies of the Grain for Green program that found no effect on off-farm labor