Direct Likelihood Estimation for the Commonly Used Pattern Mixture Models in Clinical Trials

• Activity Number: 421 - Missing Data Handling and Consideration
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Jitong Lou* and Yongming Qu, PhD
• Companies: Eli Lilly and Company and Lilly
• Keywords: Missing data; Multiple imputation; Return-to-baseline imputation; Jump-to-reference imputation; Retrieved dropout imputation

Abstract

Pattern mixture models (PMMs) have been receiving increasing attention as they are commonly used to assess treatment effects as the primary analysis or sensitivity analyses for clinical trials with non-ignorable missing data. PPMs are commonly implemented using multiple imputation (MI). The variance estimation may be a challenge as Rubin's approach of combining between- and within-imputation variance may not provide the consistent variance estimation and bootstrap methods may be time-consuming. Methods using direct likelihood-based approaches have been proposed in literature and implemented for some PMMs, but the assumptions are restrictive. In this article, we propose an efficient direct likelihood estimation with relaxed assumptions and with a novel approach to adjust for baseline covariates for PMMs including return-to-baseline, retrieved dropout imputations, and jump-to-reference. Results from simulation studies demonstrate the proposed method provides consistent estimators and outperforms MI-based methods in terms of efficiency.