Mixed Models and Longitudinal Analysis

MMRM, linear mixed-effects models, and their application in clinical trials and observational studies

2026-06-07 18:16 PDT

Overview

The mixed-effects model for repeated measures (MMRM) has become the primary analysis tool for longitudinal endpoints in confirmatory clinical trials, and for good reason: it accommodates arbitrary covariance patterns, handles missing data under the missing-at-random assumption without imputation, and is robust to certain forms of model misspecification. Despite wide adoption, important questions remain about MMRM’s behavior at boundary conditions – when the true covariance structure departs from the fitted model, when dropout is informative, or when the underlying trajectory is nonlinear.

This program examines mixed models both as analytic targets and as simulation-generating processes, with application to Alzheimer’s disease trials (ADCS, ADNI), PTSD randomized controlled trials, and post-traumatic headache studies.

Current work

  • MMRM linearity and robustness. Characterizing the consequences of misspecifying the functional form of the longitudinal trajectory in MMRM, including the power and Type I error behavior under nonlinear true effects.

  • MMRM versus survival analysis in Alzheimer’s disease. Comparing the operating characteristics of MMRM on continuous endpoints versus time-to-event analysis when both are available, with application to ADCS trial data.

  • Mixed-model R² and effect-size measures. Developing and evaluating R²-analogues for linear mixed models, for use in effect-size reporting and meta-analysis in clinical research.

  • Prazosin for post-traumatic headache: longitudinal analysis. Primary analysis of the Murray/Yeilim prazosin PTH RCT using MMRM, with marginal contrast estimation and mediation analysis of secondary outcomes.

  • CSP-563 prazosin reanalysis (NEJM format). Reanalysis of the VA Cooperative Studies Program prazosin PTSD trial using contemporary mixed-model methods applied to the raw SAS-format trial data.

  • Empirical covariance matrices from ADCS trials. Systematic extraction and documentation of variance-covariance matrices from completed ADCS trials, providing realistic simulation inputs for the N-of-1 methodology program.

Methods

MMRM (unstructured, compound symmetric, AR(1), Toeplitz covariance structures); linear mixed-effects models via lme4, lmerTest, and nlme; emmeans-based marginal contrasts; zzlongplot for longitudinal visualization; permutation tests via coin as nonparametric sensitivity analyses.

Software

  • zzlongplot – Longitudinal plotting for clinical trials, supporting MMRM-style trajectory displays with confidence ribbons and individual-profile overlays.

Publications

Selected publications are accessible through the full publications list by filtering on the keywords mixed-models or longitudinal. The publication record spans methodological papers on MMRM behavior and applied analyses of Alzheimer’s and PTSD trial data.