hu / en

IE Seminar – Kreif Noémi (Centre for Health Economics, University of York)

Weekly seminar series at the Human Research House (1097 Budapest Tóth Kálmán u. 4. ). The lectures start at 2 pm.
Date: 10/14/2019 2 pm
Venue: MTA Human Research House 1097 Budapest Tóth Kálmán u. 4. ground floor 13-14

Kreif Noémi (Centre for Health Economics, University of York)

An application of causal machine learning to explore heterogeneous treatment effects of social health insurance

Researchers evaluating social policies are often interested in identifying individuals who would benefit most from a particular policy. 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. We contrast two such approaches in a study evaluating the effects of enrollment in social health insurance schemes on health care utilisation of Indonesian mothers.  First, we apply a double-machine learning (DML) approach where we estimate both the outcome regression and the propensity score flexibly using an ensemble ML approach. From the individual-level predictions of potential outcomes we calculate individual-level treatment effects and use a Random Forest (RF) procedure to identify the variables that predict these effects. We contrast this exploratory approach to an application of the Causal Forests method (Wager and Athey, 2018 JASA), which has been designed to directly estimate heterogeneous treatment effects, by modifying the standard RF algorithm to maximise the variance of the predicted treatment effects. In both analyses we find that the most important effect modifiers include educational status, age and household wealth. When reporting conditional average treatment effects (CATEs) for subgroups defined by these variables, the methods agree that less well-educated and younger mothers would benefit more from health insurance than well-educated and older ones. The CATEs reported by the Causal Forests have larger confidence intervals than those reported by the DML method, potentially due to the extra sample splitting step employed.