Georgi is a doctoral researcher at the University of Luxembourg and Ghent University.
A huge obstacle in evaluating treatment effects is that treated and untreated individuals may be incomparable. To address causality, adjustment for confounders is necessary. But such adjustment is typically based on parametric regression. This is unsatisfactory as model misspecification is likely, and even minor misspecification may induce large bias. There is therefore a rapidly growing interest in the use of machine learning (ML) to assist this task. Naive use of ML is itself problematic as the resulting effect estimator is prone to so-called plug-in bias. This is the result of standard ML tasks not being tuned to the causal problem. This bias invalidates feasible (e.g. bootstrapbased) confidence intervals for the treatment effect. Pioneering work on Targeted/ Debiased Machine Learning (TL/DML) has shown how this can be overcome by combining separate ML predictions of both treatment and outcome, provided both converge sufficiently fast to the truth. TL/DML are major advances that are transforming treatment effect evaluations in practice. However, they are suboptimal by relying on standard ML algorithms that are geared towards prediction, but not treatment effect evaluation. A major potential for improvement remains by cleverly targeting the ML procedures on which DML estimators rely. In his research, optimal DML procedures will therefore be developed.
Illustration of recent simulations