Repeated-measures data can look simple in a wide table: one participant, several visits, and an outcome at each visit. The analysis is more complicated because the timing, selection, correlation, and absence of observations all influence what the model estimates.
The EMA missing-data guideline is a relevant reference for repeated follow-up outcomes. This article proposes a specification and review approach rather than prescribing a single model for every longitudinal study.
1. Define the visit representation
Explain how actual assessment dates map to analysis visits. Specify windows, handling of unscheduled observations, tie-breaking rules, and the baseline definition. These choices should be fixed in controlled materials and checked at participant level.
For a hypothetical weekly symptom measure, an assessment collected late may fall into a different analysis window than its operational visit label suggests. Compare actual timing with the selected analysis visit before treating an unusual trajectory as a clinical change.
2. Make the model specification complete
A mixed model for repeated measures is not one uniquely defined analysis. Fixed effects, baseline adjustment, treatment-by-visit terms, covariance structure, estimation method, degrees-of-freedom method, and contrast definitions can matter. Document the actual specification and software implementation.
A published MMRM simulation study illustrates the importance of evaluating methods under specified conditions. Do not generalize a simulation result beyond its design. Our recommendation is to test the planned implementation on study-relevant scenarios and verify that estimates and uncertainty behave as expected.
3. Review the observation mechanism
Modeling observed repeated outcomes does not make unobserved outcomes harmless. Examine availability by visit and treatment, reasons for missingness, and the clinical events associated with dropout. Identify the assumptions required for the primary estimator.
ICH E9(R1) provides the treatment-effect framework. The handling of treatment discontinuation or rescue therapy should be consistent with the intended question, including which post-event observations are collected and analyzed.
4. Check both model and presentation
Inspect convergence, warnings, estimability, and the consequences of sparse visit patterns. A program that completes without stopping can still produce an unstable or unintended result. Preserve diagnostics in the analysis package.
Compare the model-based display with descriptive availability and trajectory summaries. Explain whether plotted values are observed means or adjusted estimates and what uncertainty interval is shown. Distinguish a descriptive pattern from a formal treatment contrast.
A useful handoff includes the visit-selection specification, participant-level boundary cases, full model definition, diagnostics, and sensitivity plan. Together, these materials make the analysis reviewable and prevent a familiar method name from concealing important implementation choices.
A specification example
A hypothetical visit-based model might include treatment, categorical visit, their interaction, baseline outcome, and a baseline-by-visit interaction, with a prespecified within-participant covariance structure. That sentence is still incomplete without the estimation method, uncertainty calculation, analysis population, selected records, and treatment contrast at the target visit.
Changing visit from categorical to continuous imposes a different mean-trajectory structure. Changing the covariance model changes the assumed within-participant relationships. A fallback after nonconvergence should be predefined and justified rather than selected because it produces a favorable estimate.
Inspect the actual design matrix, target contrast, and covariance specification used by the software. The label “MMRM” on a table cannot demonstrate that the intended model was fitted.
