A Pattern-Mixture Model Framework for Assessing Informative Censoring in Survival Data: A Multiple Imputation Approach
This presentation will feature Dr. Mohamed Elsaid, Assistant Professor in the Division of Biostatistics and Population Health, Department of Biomedical Informatics at Ohio State.
About
The Biostatistics seminar series invites researchers from across the nation to discuss methodological research and its implications for a variety of health issues.
Abstract
Many time-to-event analyses lean on the unverifiable assumption that censoring is non-informative. When loss to follow-up relates to prognosis, effect estimates can drift meaningfully, biasing the results. This seminar presents a pragmatic, multiple imputation framework built on pattern mixture models to stress test departures from the noninformative censoring assumption in both trials and observational studies. We will review censoring mechanisms (CCAR, CAR, CNAR), then show how to treat censoring due to loss to follow-up as missing event times and impute them under clinically interpretable assumptions, delta-adjusted (tipping point) and control-based imputations, followed by Rubin’s rules for inference. An applied case study in a MarketScan cohort of adults with NAFLD comparing bariatric surgery vs. non-surgical care will be discussed. We will close with implementation tips, reporting recommendations, and near-term extensions (subjectspecific deltas; integration with IPTW/propensity methods; plans for software support), aiming to make sensitivity analysis for informative censoring routine rather than rare in practice.
Contact
Andy Ni