Data Science Colloquium Series Event: Professor Susan Athey

ABSTRACT: This talk will review several recent papers which aim to modify popular machine learning methods for problems of causal inference, such as evaluating the impact of a treatment using experimental or observational data. We will focus on estimation of treatment effect heterogeneity (that is, which individuals are predicted, based on their features, to have higher or lower benefits of a treatment) in settings such as A/B testing platforms or medical trials where it is important to provide confidence intervals in addition to estimated effects. We analyze the tradeoffs between evaluating model fit based on observed outcomes, and evaluating model fit based on an estimate of the (unobserved) treatment effect. We build on this work to show how random forests can be modified to provide asymptotically centered estimates of treatment effect heterogeneity in experiments or observational studies.