Causal machine learning for estimating heterogenous treatment effects using time-to-event data

In recent years, great strides have been made in understanding how to effectively us machine learning to estimate causal estimands. A class of estimators which have gained traction are known as de-biased estimators, which attenuate the regularisation bias induced when using machine learning in the estimation process. In this thesis I explore how these estimators are constructed when looking to estimate treatment effect heterogeneity, and I propose new estimators which robustly address the challenges encountered when coarsening mechanisms act on the outcome. I first discuss a simple but commonly overlooked problem of outcome missingness in the binary/continuous outcome setting, highlighting how the presence of outcome missingness can shift the covariate profile of the observed data away from that of the target population. I then develop de-biased machine learning estimators for the conditional average treatment effect (CATE), which robustly handle missing at random outcomes. I then move onto the problem of estimating heterogenous treatment effects when using time-to-event data, where individuals’ follow up times are subject to various coarsening mechanisms, including right censoring, where an individual leaves a study prior to having an outcome, and left truncation, where an individual only enters the study at a time after they become exposed. In each of these settings, I explore how machine learning can be used to estimate the difference in conditional survival probabilities, highlighting the limitations of the existing estimators, and proposing a new estimation framework, surv-iTMLE. Throughout this work, I also aim to provide guidance on how to use these estimators in practice and conclude the thesis by discussing potential further extensions that may help increase the wider adoption of causal machine learning as whole.