Targeting impact versus deprivation
||Targeting impact versus deprivation|
||SNIC Medium Compute|
||Johannes Haushofer <firstname.lastname@example.org>|
||2021-05-31 – 2022-06-01|
The objective of this research study is to test whether targeting cash transfers on the basis of deprivation leads to the highest per-dollar impact, as well as to find an optimal policy function based on commonly observed household characteristics. Using generalized random forests, we study the heterogeneity of the treatment effects of a randomized unconditional cash transfer program by the NGO GiveDirectly (GD) in Kenya. This program provided one-time cash transfers of about USD 1,000 to eligible rural households across 653 randomized villages.
Using anonymized survey data from almost 5,000 households, we train random forests to predict 4 pre-specified endline outcomes (consumption, food security index, income and assets) using household observables commonly available to policymakers. For each outcome, households are then classified as most deprived if their predicted endline value is bellow the median. Similarly, using causal forests we predict household-level treatment effects and classify households as most impacted if they are above the median household. Having identified these subgroups and individual predictions, we analyze optimal policies and estimate the parameters of a social welfare function that would rationalize targeting the deprived.
To prevent over-fitting we train 5-fold cross-fitted models and use honest trees. Given the sample variation in the 5-fold cross-fitting process we train several forests for each outcome and report the average statistic of interest. To conduct inference on the averages of the statistics of interest we employ a randomization inference (RI) approach under the null of homogeneous treatment effects. This whole procedure is computationally intensive since it requires training causal forests thousands of times.
We plan to compute the RI simulations using the requested resources.