Choose the model around the comparison you need and the way observations are related. Repeated-measures ANOVA can be appropriate for a simple, complete design with common measurement occasions. A mixed model offers more flexibility when participants miss visits, timing varies, or observations have additional clustering.
Neither method turns a weak design into causal evidence. A mixed model also does not make missing data harmless: its conclusions still depend on the model and the assumptions about why observations are missing.
Use the free analysis-planning worksheet to record the question, observation schedule, clustering, and target comparison before choosing software.
Four scores from one participant are not four independent participants. Measurements from the same person tend to be related. If people are also grouped within schools or clinics, the data may have more than one source of dependence.
Traditional univariate repeated-measures ANOVA models within-person comparisons at common occasions. With more than two repeated levels, its usual uncorrected within-person tests require sphericity: a condition about the variances of pairwise differences. Corrections such as Greenhouse–Geisser adjust the test when that condition is not met. They do not solve incomplete follow-up or additional clustering.
A mixed model specifies fixed effects for the comparisons of interest and a structure for the dependence. Depending on the approach, that can involve participant random effects, a residual correlation structure, or both. A random intercept is one possible choice, not a universal specification.
For two repeated occasions, sphericity is automatic, but other assumptions and missing-data issues remain. Do not choose the method solely because a software menu offers a familiar test.
A mixed model can represent different trajectories, unequal observation counts, and additional clustering. Time can be categorical when each scheduled visit needs a separate comparison, or numeric when a particular trajectory is justified. Treating time as numeric assumes a form of change; it should not be a shortcut imposed on a visibly nonlinear pattern.
Likelihood-based longitudinal analysis can use observed outcomes from people with incomplete visits under suitable assumptions, including an appropriate missing-at-random mechanism conditional on variables in the analysis. Missing predictors or identifiers may still exclude records. Dropout related to unobserved outcomes can require sensitivity analysis beyond the primary model.
This is an invented coefficient example, not fitted research results. Suppose participants are randomized to a workshop or comparison group and assessed at baseline and weeks 4, 8, and 12. The question is the difference between groups at week 12.
Consider a model with categorical visit, comparison group and baseline as the reference levels, participant dependence represented appropriately, and these illustrative fixed-effect coefficients:
The model-implied week-12 mean is 50 + 6 = 56 for the comparison group and 50 + 1 + 6 + 4 = 61 for the workshop group. The week-12 group contrast is therefore 5 points. The interaction coefficient alone is 4 points: it represents the between-group difference in change from baseline in this parameterization.
The baseline difference is 1 point. That is why the final group difference and difference in change are not identical here. A baseline-adjusted or constrained model has a different parameterization. State which comparison your model estimates; do not report whichever coefficient has the most appealing p-value.
Actual data are needed for a standard error and confidence interval. The interval for the five-point contrast uses the uncertainty and covariance of the relevant coefficients. Adding separate confidence-interval endpoints is not the calculation.
Review the time specification, covariance structure, residual patterns, influential observations, missingness, and convergence. Check whether the available participants and clusters can support the proposed complexity. A larger model is not automatically a better one, especially with few clusters or sparse follow-up.
For a binary, count, or ordinal outcome, consider a model appropriate to that outcome rather than applying this continuous-outcome example unchanged.
Report the model specification, analysis population, observation counts by visit, missing-data assumptions, and the estimated comparison with uncertainty. If you have only baseline and one follow-up, the pre/post comparison guide explains when ANCOVA is a useful starting point.
DASS helps researchers choose, check, and interpret longitudinal analyses. Bring the question, data dictionary, observation schedule, and assignment design to an initial consultation.
One email whenever we publish something new. No spam, unsubscribe anytime.
{{ subscribeErrorMsg }}