Online Program

Return to main conference page
Wednesday, September 25
Wed, Sep 25, 9:45 AM - 10:30 AM
Marriott Foyer
Poster Session

Practice Issues of MMRM Analysis for Longitudinal Data in Parallel or Crossover Studies (300875)

*Zheng (Jason) Yuan, Vertex Pharmaceuticals 

Keywords: MMRM Analysis

When analyzing longitudinal data in clinical trials, it is common to implement Mixed-effect Model Repeat Measurement (MMRM) to estimate least-square (LS) means and standard errors. However, a variety of issues often occur in MMRM analysis in parallel or crossover clinical trials. In this session, we will discuss the following issues and provide insights and recommendations in clinical trial setting: (1) The bias issue of traditional “LSMEANS” approach when adding categorical covariates into MMRM model in randomized clinical trials. Such bias due to categorical covariates might also lead to larger standard errors (reduced power) and substantial deviation from descriptive raw means (causing difficult interpretation). We evaluated these issues by simulations and proposed an improved flexible approach using “ESTIMATE statement” in SAS. In addition, the issue and impact of fitting non-stratified continuous covariates in MMRM model will be discussed if time permits. (2) The impact of an enormous outlier at one specific visit on the MMRM estimates of LS means and standard errors. Motivated from a real clinical trial, we noticed that one single enormous outlier could substantially impact LS means and standard errors across all visits. We evaluated this issue by simulations to assess the impact under different scenarios including sample sizes, covariance structures, amount of missing data. (3) The issue of whether to use study or period baseline to calculate the changes in crossover studies. We noticed that it is often challenging to interpret the results when the changes from study baseline were quite different from the changes from period baseline in crossover studies. We explored and evaluated this issue under different assumption scenarios and compared a variety of analysis models. We will discuss the findings and make recommendation on the use of baseline information and which analysis model is more appropriate in crossover studies.