A mixed model for repeated measures (MMRM) jointly analyzes repeatedly collected endpoint values from a patient. It estimates group differences at the planned visit times and accounts for the fact that measurements from the same person are more correlated than measurements from different people. Missing values are not imputed with single replacement values but are processed by the model under a MAR assumption.
Longitudinal Data Modeling
The MMRM uses the actually observed measurements after study start and links them to the treatment group, visit, often the baseline value, and, if applicable, other pre-specified covariates. The interaction of treatment and visit allows the estimated group difference to vary at different time points. A covariance structure simultaneously describes the dependence of values within a person.
The choice of this structure is part of the analysis plan specification because it influences the precision of the estimates. Patients with some observed follow-up values contribute to the estimation through these values, even if other visits are missing. Patients without any values after study start, however, cannot contribute a post-randomization course to the model.
MAR Assumption and Data Basis
In this context, MAR means that the probability of a missing value can be explained by already observed information, after accounting for the variables used in the model. The assumption cannot be proven from the observed data. Therefore, the reason for discontinuation, previous endpoint values, and prognostic baseline values also determine whether the model is plausible as a primary analysis.
The model does not replace the task of continuing to collect endpoints after treatment discontinuation. The fewer follow-up values available for a discontinuation group, the more the estimation relies on the modeled relationship between observed courses and missing measurements. Supplementary sensitivity analyses can examine the consequences of a divergent course.
For categorical analyses or non-continuous measurement scales, an MMRM is not automatically the appropriate choice. The endpoint type, the target variable, and the planned contrast definition determine whether a linear model for repeated measures reflects the clinical question. Even with an appropriate scale, the model estimates must align with the visit times and intercurrent events defined in the protocol.
Distinction from Imputation and Complete Case Analysis
An MMRM is not an imputation method: it does not generate an individual replacement value for a missing measurement time point and does not combine completed datasets. It is an analysis model for the observed longitudinal data. Reference-based imputation, on the other hand, can generate missing values under an explicitly chosen post-discontinuation assumption and then use an analysis model.
Similarly, LOCF, where an earlier value is carried forward, and an analysis of only completely observed patients answer different questions. Both rely on rigid assumptions about the course or the selection of analyzed individuals. The existing entries ‘regression’ and ‘multivariate-analysis’ are umbrella terms; MMRM refers to the specific model class defined in many analysis plans.
Diagnostic model comparisons can help to check the chosen covariance structure and the visit effect, but they do not replace expert knowledge about the data collection process. They must not lead to the redefinition of the primary specification after reviewing the treatment effect.
Relevance for clinical trials
For scales related to symptoms, function, or laboratory values, the MMRM determines how data from multiple scheduled visits are incorporated into the primary comparison. Before database lock, visit windows, aberrant measurements, baseline definitions, and assignment to intervention phases must be checked. The results tables should transparently present the estimated difference at each time point, along with the underlying patient numbers.
Full-service CROs like Mediconomics support the formulation of the MMRM specification in the statistical analysis plan, the review of the ADaM data structure, and independent programming reviews. The biostatistical documentation can record covariates, visit factors, covariance structure, and planned sensitivity analyses in a way that ensures the longitudinal comparison remains auditable.
For the primary comparison, the model’s convergence, the handling of individual extreme values, and the implementation of visit-specific contrasts must be comprehensibly documented. A model that does not estimate stably computationally is not a viable basis for results. Quality control therefore examines not only the generated tables but also the consistency of specification, data, and program code.
Frequently Asked Questions (FAQ)
Does an MMRM impute missing visit values?
No. The model uses the available measurements to estimate the parameters. While there are model-based expected values, these are not equivalent to individually imputed values in a completed analysis dataset.
Why is the baseline value often a covariate?
The baseline value can strongly predict the later endpoint and thus improves the precision of the group comparison. Its inclusion must be pre-specified and align with the definition of the analyzed change.
Is the MMRM unbiased for all forms of missing data?
No. Its validity depends on the MAR assumption and correct model specification. If missingness still depends on the unobserved endpoint after accounting for the included information, MNAR-oriented sensitivity analyses are required.
Regulatory References
- ICH E9(R1), Addendum to Estimands and Sensitivity Analyses – separates analysis approach and assumptions regarding missing data.
- EMA/CPMP/EWP/1776/99 Revision 1, Guideline on Missing Data in Confirmatory Clinical Trials – explains MAR and the regulatory assessment of missing data.
- EMA/CHMP/295050/2013, Guideline on Adjustment for Baseline Covariates in Clinical Trials – addresses pre-specified covariate adjustment.
