Noemi Kreif (University of York)

Researchers evaluating the effects of health policies are often interested in identifying individuals who would benefit most from a particular programme. Such analyses could provide evidence on whether a programme worked for the intended recipients, and help design the eligibility criteria of future programmes. Traditional approaches such as subgroup analyses are constrained by only considering a few, pre-specified effect modifiers, and can also be prone to cherry-picking by ad-hoc selection of subgroups. Recently proposed causal inference approaches that incorporate machine learning (ML) have the potential to help explore treatment-effect heterogeneity in a flexible yet principled way. In this talk I illustrate such an approach, Causal Forests (Athey et al. 2019), in an evaluation of the effect of public health insurance on health care utilisation of Indonesian women. I highlight the opportunities presented by the approach to identify subgroups where the impacts of having health insurance differ, and to estimate so-called conditional average treatment effects at the level of the individual. I also discuss the challenges of using this approach alongside non-randomised study designs, typical when evaluating large scale health policies.