A0590
Title: Novel Bayesian methods for sequential decision-making with incomplete information
Authors: Arman Oganisian - Brown University (United States) [presenting]
Abstract: Chronic diseases are managed over time with a sequence of treatment decisions. Time-varying covariates are often used to tailor the decisions, but are usually monitored sporadically, leading to non-monotone missingness. In these settings, the interest is in estimating causal effects of different tailoring rules on a survival time outcome. It is formalized as a dynamic treatment regime (DTR) problem with a joint monitoring-treatment decision. Under specific identification assumptions, causal effects of such joint DTRs can be estimated, but require stratifying models across monitoring patterns. With many sparsely populated patterns, estimates can be unstable. Thus, Bayesian models are developed, equipped with a class of autoregressive priors that 1) smooth the models across time within a pattern and 2) smooth the models across monitoring patterns. The regularized models are embedded in a Bayesian g-computation procedure that draws from the posterior distribution of the causal effect of various DTRs. The model is applied to analyze survival among patients with pediatric acute myeloid leukemia (AML) enrolled in the AAML1031 trial. These patients move through a sequence of treatment courses in which a decision is made to withhold scheduled anthracycline chemotherapy (ACT). Since ACT is cardiotoxic, ejection fraction (EF) is sometimes - but not always - monitored to inform the withholding decision.
