Repeated-measures ANOVA
Repeated-measures analysis of variance (ANOVA) tests whether mean outcomes differ across visits or conditions while keeping measurements from each person together.
Fit
When to use it, and when not to
Use it when
- The same independent people have a continuous outcome measured at a common set of visits or conditions.
- You want an overall within-person mean comparison and have planned how to handle sphericity and missing visits.
Not when
- Visits are incomplete or irregular, or covariate adjustment is needed. Consider a suitable mixed-effects model. Read: Mixed-effects models
- You want population-average effects for repeated binary or count outcomes. Consider generalized estimating equations. Read: Generalized estimating equations
- Each person belongs to only one group and supplies one value. Use one-way ANOVA. Read: One-way ANOVA: testing group means
Before you start
What you need
- Person identifiers, visit labels and continuous valuesrepeatedMeasuresoutcomeTypeKeep each person together and specify the planned occasions.
- The mean question and planned visit contrastsAn overall time test does not identify which visits differ or establish a treatment effect.
- The missing-data assumption and correction planPrespecify sphericity handling and missing-data analysis. Do not quietly discard people with missed visits.
- People are independent; observations within each person may be correlated.How to check: Shared clinics require further clustering beyond the repeated-measures structure.
- The uncorrected test assumes equal variances of every pairwise visit difference, called sphericity.How to check: Assess the covariance structure. A Greenhouse-Geisser correction reduces degrees of freedom when sphericity is doubtful.
- Small-sample F inference assumes a roughly multivariate normal within-person error distribution.How to check: Inspect residuals and within-person contrasts for extreme values. Robustness to skew does not remove sphericity.
- Complete-case inference needs a justified selection assumption, such as missing completely at random.How to check: If missingness depends on observed outcomes, consider a suitable mixed model under missing at random and assess plausible departures.
Plain language
How it works
Separate variation between people from variation within each person. Compare variation between visit means with the remaining person-by-visit error. Their mean-square ratio is F. Greenhouse-Geisser multiplies both reference degrees of freedom by epsilon, a covariance-based correction. It leaves F unchanged.
Blanca and colleagues study sphericity violations and non-normality in separate simulations. Corrected tests limit false positives under the studied covariance patterns. Robustness to non-normality when sphericity holds does not justify ignoring unequal difference variances.
Worked example
A worked example
- Setting
- 6 independent patients have scores at 3 visits, with no missing values.
- Question
- Are the population means equal across visits?
- Approach
- Rows, in visit order: [8,6,5]; [9,7,7]; [6,5,4]; [7,4,3]; [10,8,6]; [8,7,5]. Use the Greenhouse-Geisser correction.
- Result
- Visit means are 8.00, 6.17 and 5.00. F = 42.586; epsilon = 0.945. Corrected degrees of freedom are 1.890 and 9.449; p = 0.000021.
- What it means
- The correction changes degrees of freedom, not F. Means differ across visits; planned contrasts need suitable intervals and multiplicity control. Time alone does not prove treatment benefit.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
The uncorrected repeated-measures F test assumes sphericity: equal variances of all pairwise differences. Complete-case analysis can bias results when missing visits are informative.
- Treating every visit as a different patient.Instead: Keep the person identifier in the repeated-measures model. Independent-group ANOVA uses the wrong error term.
- Using the uncorrected p-value after assuming sphericity without justification.Instead: Report the chosen correction, epsilon and corrected degrees of freedom. Describe the covariance and missing-data assumptions.
- Interpreting a time effect as a difference between treatment groups.Instead: A treatment comparison needs its own prespecified group-by-time contrast and appropriate design.
Inside Mallard
How Mallard uses this
Mallard includes this article in Learn search and the Atlas. It has no repeated-measures ANOVA calculator. Plans do not link directly to this article yet.
analysis.repeated_measures_anova, the id a plan cue records.How to report it
Report visit means, person count, F, corrected degrees of freedom, p-value, epsilon and the correction used. Give prespecified mean contrasts with model-based t intervals and multiplicity control. Describe missingness and covariance assumptions. A reporting guideline is not a design manual: use it to make the study transparent, not as a substitute for careful design. Check the target journal's own requirements as well, which this page cannot know.
Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Repeated measures ANOVA and adjusted F-tests when sphericity is violated: which procedure is best? (opens in a new tab) · Front Psychol · doi 10.3389/fpsyg.2023.1192453
- Non-normal Data in Repeated Measures ANOVA: Impact on Type I Error and Power (opens in a new tab) · Psicothema · doi 10.7334/psicothema2022.292