Prediction, Causation and Policy: A Review of Machine Learning in Development Economics and Public Policy

Abstract

Background. Machine learning (ML) has moved from the margins of economics to the centre of applied work. It is used to measure poverty from satellites and mobile phones, to target social transfers, to estimate treatment effects with many covariates and to study who benefits most from public programmes. Objective. This review asks four questions: which policy problems are prediction problems; how far new data and ML can substitute for traditional measurement; how ML can be combined with causal inference; and what limits apply when predictions are used to allocate public resources. Methods. We synthesise methodological work in econometrics and empirical applications in development and public economics, giving priority to peer-reviewed studies that report out-of-sample performance or policy-relevant counterfactuals. Figures report published results; diagrams are the authors' design. Findings. ML performs well when the task is prediction and when the target is aggregate, such as village wealth, where models explain most of the variation. It adds less when the target is an individual life outcome. Phone-based targeting improved on feasible alternatives in an emergency programme but fell short of a full social registry. Double/debiased ML and causal forests make high-dimensional causal inference more credible, but do not replace identification. Implications. The value of ML for policy depends on matching the tool to the question, validating predictions against the outcome policy actually cares about, and anticipating strategic responses, distribution shift and distributional consequences.