Machine learning for causal inference: developing modern tools for personalised optimal treatments (360G-Wellcome-218554_A_19_Z)

Medical and social scientists evaluating the effects of medicines or social policies (treatments) are interested in finding optimal treatment regimes, as well as identifying the best time for treatment initiation, changes or discontinuation. To achieve such tailored interventions, we need a step-change in our understanding of treatment-effect heterogeneity, which potentially depends on many factors. Reliable inference is this area is notoriously challenging, because these are poorly-understood relationships involving a large number of confounders, effect modifiers and missing data. This has triggered an interest in machine-learning to help obtain models which adapt to accurately capture the relationships of interest. However, quantifying uncertainty is very difficult since machine learning is primarily focused on prediction. Moreover, machine-learning estimates suffer from regularisation bias. Nevertheless, under certain assumptions, doubly-robust estimators can be combined with machine-learning, to obtain consistent estimates with valid confidence intervals (with the bias vanishing faster than for simple estimators). This fellowship aims to exploit this potential, by extending existing doubly-robust estimators to use machine-learning estimation in obtaining valid inferences, by (1) choosing correctly-specified confounder models, optimised for estimating the true causal effects, (2) appropriately accounting for the observation process, (3) identifying the important effect- modifiers, and (4) finding optimal treatment regimes.