The Effects of Irrigation and Weather on Agriculture in Haiti: A Machine Learning Approach

Low program uptake and the prevalence of highly non-linear relationships between variables can make the evaluation of development interventions challenging. This study highlights the potential for using Machine Learning (ML) as a supplement to the traditional econometric toolkit for causal inference in such complex settings. We illustrate this approach by using data from an RCT in rural Haiti to measure the causal relationship between access to small-scale irrigation, weather and household outcomes. We use ML techniques to generate predictions of productivity, revenue and high value crop production from the baseline and then estimate treatment effects on the residuals. Our Intent to Treat (ITT) program estimates suggest that irrigation increased crop revenue by 24%. We find no statistically significant effects on maize quantity harvested and the production of high value crops, even though the corresponding point estimates go in the right direction.