TL22: Practical Issues with MMRM
*Dalong Patrick Huang, FDA/CDER 

Keywords: MMRM model, unstructured covariance, LOCF, MI

Mixed-Effect Model Repeated Measure (MMRM) model has been widely used to analyze incomplete longitudinal clinical trial data. Based on simulation study and 25 new drug application datasets, Siddiqui, Hung, and O’Neil (2009) and Siddiqui (2011) pointed out that MMRM analysis appears to be a better choice in maintaining statistical properties of a test compared to the last observation carried forward (LOCF) and Multiple Imputation (MI) approaches. In addition, the MMRM model with unstructured covariance is robust against misspecifications of covariance structure.

MMRM model with unstructured covariance tends to be the primary analysis of continuous endpoint(s) in pharmaceutical industries and/or requested by regulatory agencies since the publication of the National Academy of Science (NAS) report on missing data. Topics for discussions are: 1) In the literature, some people include time*baseline interaction term in MMRM model; some do not. What is a proper MMRM model with unstructured covariance? Why should time*baseline interaction term be included (or not)? 2) Should time*other pre-randomization covariate interaction term be included in MMRM model? 3) What are the alternative approaches to fit MMRM model when the model with unstructured covariance fail to converge ? 4) What details regarding the alternative approaches should be specified in statistical analysis plan? This roundtable discussion will gather a group of statisticians to share their best practices and view points, and for all to learn from each other’s experiences.