B0287
Title: Leveraging Bayesian ML for causal inference with missing longitudinal data
Authors: Liangyuan Hu - Rutgers University (United States) [presenting]
Abstract: Missing data presents a significant hurdle in the analysis of complex longitudinal datasets, particularly when striving to draw causal inferences regarding longitudinal treatments. Current imputation methods for missing-at-random longitudinal covariates primarily rely on parametric models, which explicitly outline the relationships among the longitudinal response, treatment, and covariates. However, an inaccurate specification of the parametric form can lead to biases due to model misspecification. To address this, flexible semi- and non-parametric Bayesian sequential imputation methods are proposed for these covariates. A novel Bayesian tree mixed-effects model is innovated to nimbly model longitudinal trajectories, followed by an efficient MCMC algorithm that sequentially imputes the missing data using the model developed. The novel methodology for handling longitudinal missing data is then seamlessly integrated with g-computation to examine the causal effect of longitudinal treatment. Extensive simulations are carried out to examine the practical operating characteristics of the proposed methods. Lastly, the methods are applied to an NHLBI study dataset to estimate and validate optimal dynamic rules for initiating antihypertensive treatment.
