Abstract

Longitudinal observational studies are becoming more popular, largely due to the increasing availability of rich administrative databases. In an observational setting, treatment assignment at each clinical visit follows a patient-adaptive decision strategy, where the clinician tailors treatment to the individual patient based on current and past clinical measurements as well as treatment histories. When working with administrative data, high-dimensional covariates and clustered data add to the data complexity for longitudinal causal modelling. In the context of comparative effectiveness studies, Bayesian methods propagate estimation uncertainty, allow direct probability summaries of the treatment effectiveness, and most importantly, afford us the ability to incorporate prior clinical/expert beliefs. Despite their unique estimation features and wide applicability, Bayesian causal inference methods for handling longitudinal data under observational designs have received limited attention in the statistical literature. In this talk, I will present extensions to two Bayesian causal methods to account for time-dependent confounding and time-dependent treatment in longitudinal observational studies that feature latent confounding modelling and Bayesian causal estimation with clustered data and demonstrate their use using clinical data.