The strongly ignorable treatment assignment assumption (also known as no unmeasured confounding) is an untestable causal assumption which requires a sufficiently large set of covariates being measured to ensure that subjects are exchangeable across the observed exposure given measured covariates in causal inference.
Several Bayesian sensitivity analyses for unmeasured confounding have been developed that use bias parameters to capture the effect of latent confounders on the outcome and exposure in causal modelling. However, there is a lack of considerations to handle time-dependent unmeasured confounders. This project seeks to develop a novel Bayesian sensitivity analysis for unmeasured time-dependent confounding.
The student will conduct literature review on existing parametric and non-parametric Bayesian sensitivity analysis for unmeasured confounding in the point-treatment setting and compare these methods in a series of simulation studies. The student will then, under the guidance of the supervisor, develop a full parametric Bayesian sensitivity analysis for unmeasured time-dependent confounding and evaluate its performance numerically. Through this project, the student will acquire statistical expertise in causal inference and Bayesian modelling. Statistical programming will be carried out using the software R and Stan. Project can be tailored to students’ specific interests.