MMRM or MI under MAR?

Biostats
Author

Imanol Zubizarreta

Published

February 9, 2025

Introduction

Sidiqqu et al demonstrated that under Missing At Random (MAR), MMRM can be already an optimal likelihood-based approach and that Multiple Impation (MI) adds no benefit and can actually be harmful to statistical efficiency.

In this post, I will be summarizing the key points of the paper and providing a brief overview of the comparison between MMRM and MI under MAR.

The paper compares Mixed-Effects Model for Repeated Measures (MMRM) and Multiple Imputation (MI) in handling missing data in clinical trials, specifically for neuropsychiatric drug products. The study evaluates their relative performance in controlling Type I error rates, statistical power, and parameter bias using:

  • Simulated data with missingness under MAR
  • Real-world clinical trial data from 25 New Drug Applications (NDAs) involving 48 Phase III trials.

Before diving in, if you need a quick refresher on MAR vs. MNAR and why it matters in statistical analysis, check out my pre-read here:

Type I Error Control (False Positive Rate)

Both MMRM and MI adequately controlled Type I error at approximately 5% under the null hypothesis. However, MI was more conservative, meaning it tended to have lower false positive rates than MMRM【10】.

Statistical Power (Ability to Detect True Effects)

  • MMRM had higher statistical power than MI when a true treatment effect was present.
  • MI produced larger standard errors, leading to smaller T-values and wider confidence intervals, reducing its ability to detect true treatment effects. MMRM consistently re-estimated the true treatment difference, whereas MI underestimated it.

Real-World Clinical Trials: Comparing NDA Data Analyses

  • 108 treatment comparisons were made across 48 clinical trials.
  • 94.4% of conclusions were the same between MMRM and MI.
  • 5.6% of conclusions differed, but there was no systematic pattern in which method performed better in these cases.

Conclusion & Recommendation

  1. MMRM is the preferred method under MAR because:
  • It better controls Type I error.
  • It maintains higher statistical power than MI.
  • It produces more reliable treatment effect estimates.
  • It does not require choosing the number of imputations (m) like MI, which can introduce additional variability.
  1. MI can be overly conservative in estimating treatment effects, leading to lower statistical power. This is a major disadvantage in detecting true drug benefits.