Bayesian Design and Analysis for Studies with Intercurrent Events and NoncompliancedissertationsSisti, AnthonycreatorGatsonis, ConstantineReaderMcCreedy, EllenReaderMor, VinceReaderGutman, RoeeAdvisorBrown University. Department of Biostatisticssponsor2024XVII, 137 p.born digitalThesis (Ph. D.)--Brown University, 2024thesesRandomized controlled trials (RCTs) are often considered the gold standard for investigating the effects of an intervention. However, selective recruitment criteria often exclude vulnerable populations from RCTs, and tightly controlled designs make it difficult to determine how an intervention would perform in a more practical setting. To address these challenges, researchers typically rely on observational studies or pragmatic randomized controlled trials (PRCTs). Observational studies use data that has already been collected to analyze the effects of an intervention, and PRCTs are designed to assess an intervention’s performance in routine clinical practice. However, because observational studies and PRCTs are not as tightly controlled as traditional RCTs, they may require additional statistical considerations related to intercurrent events and noncompliance. Intercurrent events are events that occur after the study has been initiated that may affect the existence or interpretation of an outcome, and noncompliance occurs when an individual assigned to an intervention does not take it. The objective of this dissertation is to develop Bayesian design and analysis methods that address statistical challenges related to intercurrent events and noncompliance in observational studies and PRCTs. For observational studies, we develop a composite ordinal outcome and Bayesian analysis method to account for death before follow up when analyzing adverse event side effects of medication. For PRCTs, we propose a design and analysis procedure that enables researchers to modify an intervention at an interim stage to improve compliance while preserving statistical power. For pragmatic cluster randomized controlled trials (PCRCTs), we design a latent principal stratification method for analyzing studies with one-sided partial noncompliance at the cluster level and binary noncompliance at the individual level. We implement this method on METRIcAL: a nursing home based, PCRCT evaluating the effects of a personalized music intervention on agitated behaviors and medication use of patients with moderate to severe dementia.causal inferencenoncomplianceExperimental designBayesian StatisticsIntercurrent EventsPragmatic TrialsObservational StudiesEnglishRPB20241015