Abstract

A multi-stage or multiple decision dynamic treatment regime (DTR) is a collection of decision rules, each taking as input patient's observed history up to the decision time point, and returning a treatment alternative as a function of the history. An optimal DTR is a regime that maximizes the expected value of an outcome. While DTRs are typically formulated at predetermined/fixed decision times, longitudinal data to estimate these may come from observational studies with irregular and potentially informative observation/visit times. Two distinct problems present here: optimizing both the decisions and the decision times, or using the irregularly observed data to estimate the DTRs under a fixed time point target regime. We focus on the latter, and formulate a Bayesian approach based on a shared random effect model and g-computation, incorporating a continuous-time model for the visit times. We also investigate an alternative approach where the effect of time-dependent confounders that are not tailoring variables is removed through weighting.