Comprehensive implementations of multiple imputation using retrieved dropouts for continuous endpoints

In the metabolic disease area, there has been a long-time debate against using mixed models for repeated measures (MMRM) as the primary analysis of longitudinal continuous endpoints. As missing data arising from missing not at random assumptions are not addressed in MMRM, multiple imputation based on specific assumptions has been brought into play. Among many missing not at random assumptions with varying degrees of conservativeness, multiple imputation based on retrieved dropouts (MIRD) has been accepted by regulatory agencies in several type 2 diabetes and chronic weight management products in recent years, marking the beginning of a new standard for analysis of longitudinal data in this disease area.

On top of the established MIRD approach of which the imputation is based on last on-treatment data of retrieved dropout (RD)s, we propose a new class of MIRD approaches utilizing all available data from RDs. The imputation implementation can be one-step Markov Chain Monte Carlo (MCMC) or two-step (creating monotone missingness, followed by regression approach). ANCOVA can be applied to the complete dataset post imputation and Rubin's rule can be used to combine all estimates into a single estimate. Simulation studies in a wide range of scenarios are conducted to understand the type-I error and power rates of the new class versus the established MIRD approach and other reference statistical methods such as MMRM.

Overall, the new class has very similar performance compared to the established MIRD approach based on last on-treatment data. What's more interesting is the one-step MCMC approach has better controlled type-I error and is more powerful than the established MIRD approach in certain scenarios with the difference gradually diminishing with larger sample size. The data analyses based on two real phase 3 datasets further manifest the power conclusions, with the results based on the new class applied to the larger of the two datasets almost identical to that of on-study MMRM.

We have presented comprehensive implementations of the MIRD approach for continuous endpoints in a longitudinal setting that fully fit within the strategy of treatment policy. The proposed new class based on all observed data of RDs is proved to be as powerful as the established MIRD approach based on last on-treatment visit in most scenarios. The one-step MCMC approach is more powerful than the established MIRD approach in certain scenarios. Since the new class involves less programming derivation of additional flags, it's anticipated to be more easily implemented in clinical trial reporting.