Abstract:
|
In clinical trials, handling missing data correctly is critical to drawing unbiased inferences on treatment effects. In our research, we focused on longitudinal clinical trials with dropout in which two treatment groups were compared in terms of a continuous endpoint, and evaluated the sensitivity of principled statistical methods assuming missing at random (MAR) for this setting. Specifically direct likelihood methods (e.g. mixed-effects models) and Bayesian imputation methods (e.g. multiple imputation with some imputation models and covariates) were evaluated. Simulation studies based on an actual randomized parallel group clinical trial for patients with major depression were performed under various settings, including missing mechanisms and missing fractions. Sensitivity of analysis results of aforementioned MAR methods to departures from the MAR assumption was compared with that of complete-case analysis, last observation carried forward, and missing not at random methods (e.g. some pattern-mixture models with various restrictions). The details of results will be shown in the presentation. This work was supported by JSPS KAKENHI Grant Number 25240005 and Grant Number 15K08564.
|