Experimental Design Details
The resilience and livelihood activities in this study align with similar programs that provide cash payments tied to the condition of undertaking some form of (CALP, 2020). While the nature of the work differs substantially across programs, it typically involves an activity that is designed to generate positive returns for the household, or the community they live in. WFP introduced a livelihood program called Food Assistance for Assets (FFA) in the 1990s to meet the short-term food needs of vulnerable populations (through cash transfers), while promoting long-term resilience through the production of assets (WFP, 2019).
We design an RCT that measures the impact of the WFP livelihood program on welfare dynamics. To this end, villages are randomly allocated to two treatment arms and a control group. In the first treatment arm (UCT), WFP provides unconditional cash transfers to households. In the second treatment arm (CCT), WFP provides their cash-for asset program to selected beneficiaries. We compare CCT to control to isolate the benefits of a livelihood program on welfare dynamics. We compare the UCT to the CCT to understand whether the positive returns from the asset outweigh the costs that come from having to invest additional labor. In doing so we are able to isolate the benefit of assets on their own. This complements existing research on graduation programs, which were designed to understand the effect of assets when combined with other forms of assistance (namely cash and training) (Banerjee et al., 2015).
We hypothesize that the welfare gains associated with these programs could increase if the timing of their own activities were adjusted to accommodate seasonality and shocks. These fluctuations are especially relevant in agricultural economies where households' marginal utility of consumption and opportunity cost of labor are positively correlated. During the pre-harvest season, households have less disposable income and less time to devote to nonfarm activities. In the post-harvest season, households have additional income from selling their crops and fewer demands on their time. It follows that cash transfers should be provided in the pre-harvest season when the marginal utility of consumption is high, and work requirements should be reserved for the post-harvest season when the returns to alternative labor allocations are low. To test this hypothesis, we further subdivide CCT villages into two groups. In the first group, villages receive the Coupled WFP cash for asset program- households are invited to work on the asset while they receive cash payments. In the second group, villages receive a De-coupled WFP cash for asset program, whereby the cash transfers are provided when the marginal utility of consumption is highest (the pre-harvest season), but work requirements are limited to when the marginal cost of labor is low (the post-harvest season). Comparing the de-coupled CCT to the coupled CCT isolates the welfare gains associated with providing cash and labor at times when the MPC is high and the MPL are low, respectively. Importantly, this tests the value of changing the timing of programs to account for seasonal variation in labor calendars and consumption patterns.
While the timing of program implementation is one dimension on which programs account for dynamic adjustments, a second dimension can involve updating targeting and beneficiary selection to reflect changes over time in the welfare rankings of households. To study the implications of these decisions, we further cross-randomize CCT villages into two groups. In both groups, we ask villages to draw up a list of beneficiaries, where the 5 last households will be kept in `reserve'. After the harvest season (when shocks may have shifted new households into poverty), we ask communities to identify 5 additional `postharvest' beneficiaries to include in the program. In non-retargeted villages, the 5 `reserve' households will be enrolled right away in the pre-harvest season. In re-targeted villages, the `post-harvest' beneficiaries will receive the program instead. We apply a difference-in-difference specification to determine the relative treatment effect of the CCT for postharvest beneficiaries relative to the `reserve' households in re-targeted vs. non-retargeted villages. We also we investigate the extent to which communities select beneficiaries based on unobservable characteristics across treatment arms. Finally, we can apply the ML techniques developed by Chernozhukov et al. (2018) to predict who benefits from each program based on observable characteristics, and determine the quality of targeting under each treatment arm, comparing whether the beneficiaries selected under each program match those we predict would derive the highest returns.