18
nonparticipants follow parallel trends over time. In other words, it assumes that the coefficients
associated with Yt' (the preprogram outcome) and the covariates in t' (the preprogram period) equal
one. This assumption may be implausible if unobservable preprogram characteristics are thought to be
associated with the dynamics of the outcome variable and the characteristics are different for
participating and nonparticipating groups. We also report DID estimates of the impact of Grain for
Green on off-farm employment (and other outcome variables) for models that include the preprogram
outcome ( Yt' ) and other preprogram control variables (such as household size and total land holdings)
that can increase the probability that the parallel trend assumptions hold.
Employing DID allows us to control for a number of variables in the reduced form of the
model.12 First, DID differences out all the time-invariant variables. We assume that the total land and
time endowment (A, L ), the variable and fixed transaction costs for participation in off-farm-labor and
credit markets (τvo,τo,τB ) and the household characteristics that determine consumption and
production (zc, zf ,zo) are time-invariant.13 Next, DID zeroes out any time-variant variables for which
the two groups change in parallel (i.e., variables that have common trends). We assume that changes in
input and output prices ( p,w, ω) are common to all households, so these effects are captured.
After controlling for time-invariant factors and for time-variant factors that have common
trends, we are left with time-varying observable and unobservable factors for the two groups that affect
changes in off-farm labor participation and that systematically change along nonparallel trends. Among
them, we are able to control for the program compensation rate (δ) , which varies over time. It equals
zero for all households in 1999 and has a positive value only for participating households in 2004. There