Advances in Causal Machine Learning for Health Management

Abstract

Recent advances in causal machine learning leverage observational data to estimate heterogeneous treatment effect (HTE), which is crucial for individual-level decision-making in many application areas such as precision medicine. However, estimating HTEs from observational data entails several challenges, given that inferring causal effects goes beyond correlations observed in data. These challenges include violation of important assumptions needed for identification of causal effects (e.g., violation of unconfoundness in the presence of hidden confounders), as well as covariate shifts caused by treatment selection bias. Failure to address these problems can result in unreliable estimates of HTEs. In this dissertation, we develop novel causal machine learning methods for estimating HTEs that address the aforementioned challenges in different settings. In the second chapter, we analyze HTE estimation with missing treatment information, where unique challenges arise in the form of covariate shifts. In the third chapter, we deal with violation of unconfoundness in the longitudinal setting where, instead of true hidden confounders, we observe only noisy proxies thereof. In the fourth chapter, we leverage causal machine learning to solve an important practical problem by informing optimal allocation of development aid towards ending HIV/AIDS epidemic. Finally, we outline general limitations and obstacles that arise when using artificial intelligence algorithms and, hence, also causal machine learning in practice. Our work demonstrates the enormous potential of causal machine learning to improve individual-level decision-making in practice. By developing tailored methods, we address several challenges pertaining to HTE estimation from observational data. As a result, we make significant contributions towards more reliable HTE estimation in different settings. This has important practical implications across many disciplines, where our methods can be used to support managers in making better decisions, thereby leading to more favorable outcomes.