How to Perform Repeated Measures ANOVA in Medical Research

Q: What is the difference between repeated measures ANOVA and a paired t-test?

A paired t-test compares exactly two related measurements taken from the same subjects. Repeated measures ANOVA extends the same within-subject logic to three or more related measurements — three or more time points or conditions on the same subjects — and tests whether there is an overall difference across all of them in a single model, rather than requiring multiple separate pairwise tests that would inflate the false-positive rate.

Q: What is sphericity and why does it matter?

Sphericity is the assumption that the variances of the differences between every possible pair of the repeated measurements are approximately equal — for example, the variance of (Week 4 minus Baseline) should be similar to the variance of (Week 12 minus Week 8). It only becomes relevant with three or more repeated measurements, since with exactly two there is only one pair of differences and nothing to compare. When sphericity is violated, the standard F-test's p-value becomes anti-conservative, meaning it's more likely to produce a falsely significant result than the stated alpha level suggests.

Q: What happens if Mauchly's test is significant?

A significant Mauchly's test (conventionally p less than .05) indicates that the sphericity assumption has been violated, meaning the standard, uncorrected repeated measures ANOVA p-value can no longer be trusted at face value. The standard response is to apply a correction to the degrees of freedom — most commonly the Greenhouse-Geisser correction, or the Huynh-Feldt correction when the estimated sphericity value (epsilon) is closer to 1 — which produces an adjusted, more conservative and valid p-value without requiring a different test altogether.

Q: Can repeated measures ANOVA handle missing data?

The traditional repeated measures ANOVA model handles missing data poorly — it typically requires complete cases, meaning any subject missing even one of the repeated measurements is excluded entirely from the analysis (listwise deletion), which can substantially shrink the effective sample size in longitudinal studies with dropout. A linear mixed-effects model is a modern alternative that can include subjects with partial data, as long as the reason for the missingness doesn't itself depend on the unobserved outcome value, and is increasingly preferred for exactly this reason in longitudinal clinical research.

Q: What is the non-parametric alternative to repeated measures ANOVA?

The Friedman test is the standard non-parametric alternative, used when the outcome is ordinal, or when the assumption of approximately normal residuals is severely violated and cannot be reasonably addressed with a data transformation. Like repeated measures ANOVA, the Friedman test compares three or more related measurements from the same subjects, but it works on ranks rather than raw values and does not carry a sphericity assumption, at some cost in statistical power compared to a correctly specified parametric model.

Q: How many time points do I need for repeated measures ANOVA?

A minimum of three repeated measurements (time points or conditions) on the same subjects is required — with only two, the analysis reduces mathematically to a paired t-test, and sphericity becomes a non-issue since there is only one pair of differences to consider. There is no strict upper limit on the number of time points, though statistical power to detect differences at any specific pair of time points, and the number of pairwise comparisons needed afterward, both increase as more time points are added.

Q: What effect size should I report for repeated measures ANOVA?

Partial eta squared (partial η²) is the most commonly reported effect size for the overall time effect, representing the proportion of variance in the outcome explained by time after accounting for between-subject variability. Using Cohen's widely cited rough benchmarks, partial η² around 0.01 is considered a small effect, around 0.06 a medium effect, and around 0.14 or above a large effect, though these thresholds are conventions rather than fixed clinical cutoffs, and effect sizes in repeated measures designs tend to run larger than in between-subjects designs because each subject serves as their own control.

Q: Do I need post hoc tests after repeated measures ANOVA?

Yes, whenever the overall (omnibus) time effect is statistically significant and there are three or more time points, since the omnibus F-test only tells you that at least one time point differs from at least one other, not specifically which ones. Pairwise comparisons between specific time points, corrected for multiple comparisons using a method such as Bonferroni, are needed to identify exactly where the significant differences occur — for example, whether the change from baseline to week 4 was significant, separately from the change from week 4 to week 12.

Q: What is the difference between repeated measures ANOVA and a mixed-design ANOVA?

Repeated measures ANOVA, as covered in this guide, involves a single group of subjects measured repeatedly, with only a within-subject factor (such as time). A mixed-design ANOVA (also called a split-plot ANOVA) adds a between-subjects factor as well — for example, comparing a treatment group against a control group, with both groups measured repeatedly over the same time points — and additionally tests whether the pattern of change over time differs between the two groups, a time-by-group interaction.

Q: Should I use repeated measures ANOVA or a linear mixed model?

Repeated measures ANOVA remains a widely accepted, well-understood choice for a complete, balanced dataset with no missing time points and a manageable number of repeated measurements. A linear mixed-effects model is generally the more robust and flexible choice when data are missing at some time points, when time is better modeled as a continuous rather than categorical variable, or when the study design is more complex than a single within-subject factor — many methodologists now recommend mixed models as the default for longitudinal clinical data for exactly these reasons.

Longitudinal clinical data — the same patients measured at baseline, week 4, week 8, and beyond — is some of the most common data in medical research, and also some of the most frequently mishandled. When a continuous outcome is measured three or more times on the same subjects, repeated measures ANOVA is the parametric test built specifically for this structure. It comes with one assumption most other tests don't have — sphericity — and this guide covers everything from the underlying logic to a full worked example, a StatClinic walkthrough, and exactly how to interpret and report your results, without assuming you've read anything else first.

Key Takeaways

Repeated measures ANOVA compares three or more measurements taken on the same subjects over time or across conditions — it is the multi-timepoint extension of the paired t-test.
It carries one assumption unique to designs with 3+ repeated measurements: sphericity — roughly equal variances among all the pairwise differences between timepoints.
Mauchly's test checks sphericity formally; when it's violated, the Greenhouse-Geisser correction adjusts the degrees of freedom rather than requiring a different test.
A significant overall (omnibus) time effect only tells you that measurements differ somewhere — pairwise comparisons are needed to identify exactly where.
Missing data is a real practical limitation — standard repeated measures ANOVA needs complete cases; a linear mixed model is often the better choice with dropout.

What Is Repeated Measures ANOVA?

Repeated measures ANOVA is a statistical test that compares the means of a continuous outcome measured three or more times, or under three or more conditions, on the same group of subjects. It tests a single overall question — does this outcome change significantly across the repeated measurements? — while correctly accounting for the fact that measurements from the same person are correlated with each other, unlike independent groups.

Because each subject is measured repeatedly and effectively serves as their own control, repeated measures ANOVA removes between-person variability from the error term used to calculate significance. This typically gives it more statistical power than an equivalent between-subjects design with the same total number of measurements, which is one of the main practical reasons longitudinal designs are so common in clinical research.

In Plain Terms

Instead of asking "do three separate groups of people differ?" (ordinary one-way ANOVA's question), repeated measures ANOVA asks "does the same group of people change significantly across three or more measurements?" — a question that requires modeling the correlation between each person's own repeated values.

When Should You Use Repeated Measures ANOVA?

Repeated measures ANOVA is the right test when your study meets all of the following conditions:

Your outcome is continuous (a lab value, a symptom score, a physiological measurement).
The same subjects are measured at three or more time points, or under three or more related conditions (such as three different drug doses given in sequence to the same patients).
You want to test whether there is an overall change across the repeated measurements, not just a change between two specific points.
The differences between repeated measurements are approximately normally distributed, without severe outliers.

If you only have two measurements per subject, this reduces to a paired t-test. If different subjects are measured in different groups (not the same people repeated), you need one-way ANOVA instead. If you have both a repeated (within-subject) factor and a separate grouping (between-subjects) factor — for example, comparing a treatment and control group, each followed over the same time points — you need a mixed-design ANOVA, which extends the logic covered here.

Repeated Measurements and Within-Subject Designs

A within-subject factor is a variable whose different levels are all experienced by the same subjects — most commonly, time (baseline, week 4, week 8). This is the defining feature of a repeated measures design, and it's different from a between-subjects factor, where different subjects occupy different levels (such as a treatment group versus a control group of different people).

The practical implication is that a within-subject design needs fewer total subjects to achieve the same statistical power as a between-subjects design, since each subject contributes multiple data points and acts as their own baseline comparison. The trade-off is added statistical complexity — the correlation between a subject's own repeated measurements must be explicitly modeled, which is exactly what introduces the sphericity assumption discussed later in this guide.

Repeated Measures ANOVA vs Ordinary ANOVA and the Paired T-Test

Feature	One-Way ANOVA	Paired T-Test	Repeated Measures ANOVA
Groups/timepoints compared	3+ independent groups	2 related measurements	3+ related measurements
Subjects	Different across groups	Same subjects, 2 timepoints	Same subjects, 3+ timepoints
Accounts for within-subject correlation	No	Yes	Yes
Unique assumption	Homogeneity of variance	None beyond normality of differences	Sphericity (3+ levels)
Test statistic	F	t	F

Take-Home Points Repeated measures ANOVA sits directly between the paired t-test and one-way ANOVA conceptually — it uses the paired t-test's within-subject logic, extended to handle three or more repeated measurements the way one-way ANOVA handles three or more groups.

Suitable Data Structure

Repeated measures ANOVA fits a specific data structure: a single group of subjects, each measured on the same continuous outcome at three or more fixed time points or conditions, with no subject appearing in more than one "arm" of the repeated factor. It's the natural test for pre-post-follow-up designs, dose-escalation studies within the same patients, and any longitudinal single-group study tracking change over multiple visits.

It is not the right test when different subjects belong to different groups (use one-way ANOVA), when only two timepoints exist (use a paired t-test), or when the study has both a repeated factor and a separate between-subjects group to compare (use a mixed-design ANOVA).

Assumptions, Especially Sphericity

Continuous outcome

Measured on an interval or ratio scale at every timepoint.

Independence between subjects

One subject's repeated measurements shouldn't influence another subject's.

Approximate normality

The differences between timepoints (or the model's residuals) should be roughly normally distributed.

No severe outliers

Extreme values at any single timepoint can distort that timepoint's mean and the overall result.

Sphericity

The variances of the differences between every pair of timepoints should be roughly equal — unique to designs with 3+ repeated measurements.

Sphericity is the assumption that gives repeated measures ANOVA its reputation for being trickier than other ANOVA variants. With three timepoints (Baseline, Week 4, Week 8), there are three possible pairs of differences: Week4−Baseline, Week8−Baseline, and Week8−Week4. Sphericity requires the variance of each of these three difference scores to be approximately equal across the group. With only two timepoints, there is only one pair of differences, so sphericity is automatically satisfied and never an issue — this is why it only becomes relevant once you have three or more repeated measurements.

Mauchly's Test and the Greenhouse-Geisser Correction

Mauchly's test formally tests the sphericity assumption. A non-significant result (conventionally p ≥ .05) means sphericity can be reasonably assumed and the standard, uncorrected repeated measures ANOVA result can be used directly. A significant Mauchly's test (p < .05) means sphericity is violated, and the standard F-test's p-value becomes anti-conservative — more likely to appear significant than it should.

The standard fix does not require switching to a different test — it corrects the degrees of freedom used to calculate the p-value. The most common correction is the Greenhouse-Geisser correction, which multiplies the original degrees of freedom by an estimated value called epsilon (ε), always between 0 (severe violation) and 1 (perfect sphericity). A smaller epsilon produces a larger downward adjustment to the degrees of freedom, producing a more conservative, appropriately cautious p-value. When epsilon is estimated above 0.75, some statisticians prefer the slightly less conservative Huynh-Feldt correction instead — both are reported automatically by most statistical software alongside the uncorrected result.

Caution Never report the uncorrected p-value once Mauchly's test is significant — doing so overstates how confident you can be in the result, and is one of the most common errors in published repeated measures analyses.

How to Prepare the Dataset

Repeated measures data is usually entered in wide format: one row per subject, with one column for each timepoint.

Patient ID	Baseline SBP	Week 4 SBP	Week 8 SBP	Week 12 SBP
001	158	150	142	134
002	146	140	136	128
003	162	153	145	137
...	...	...	...	...

StatClinic and most modern statistical software accept this natural wide format directly and reshape it internally as needed. Two practical points matter most before analysis: confirm every subject has a value at every timepoint (standard repeated measures ANOVA needs complete cases), and document how many subjects were excluded due to missing follow-up visits, since this directly affects your sample size and generalizability.

A Realistic Longitudinal Medical Example

A clinic runs a 12-week lifestyle intervention program for patients with stage 1 hypertension, measuring systolic blood pressure (SBP) at baseline, week 4, week 8, and week 12 in the same 24 patients — a single-group, four-timepoint repeated measures design.

Timepoint	Mean SBP (mmHg)	SD
Baseline	152	9.1
Week 4	145	8.7
Week 8	138	8.3
Week 12	130	7.9

What This Pattern Suggests

Mean SBP falls steadily across all four visits, with the largest drop appearing between baseline and week 4. Whether this pattern is statistically reliable, and whether every consecutive pair of visits differs significantly, is exactly what the repeated measures ANOVA and its follow-up pairwise comparisons will confirm.

Step-by-Step Analysis Using StatClinic

Enter your data in wide format

One row per patient, with a separate column for SBP at each of the four visits.

Confirm your study design

Tell StatClinic these are repeated measurements on the same subjects across four timepoints — this routes the analysis correctly.

Let StatClinic run Mauchly's test automatically

StatClinic tests sphericity behind the scenes and flags a violation before you interpret the main result.

Apply the correction if needed

If sphericity is violated, StatClinic automatically applies the Greenhouse-Geisser correction to the degrees of freedom.

Review the omnibus time effect and effect size

You'll get the corrected F statistic, p-value, and partial η² for the overall effect of time.

Get pairwise comparisons and a written interpretation

StatClinic runs corrected pairwise comparisons between timepoints and generates ready-to-use Methods and Results text.

You can also run the calculation directly with the Repeated Measures ANOVA Calculator once your timepoint data are ready.

Interpreting the F Statistic, P Value, Effect Size, Time Effect, and Pairwise Comparisons

For the hypertension example, StatClinic's output might read: Mauchly's test χ² = 9.8, p = .03 (sphericity violated, ε = 0.70), followed by a Greenhouse-Geisser corrected omnibus result of F(2.1, 48.3) = 42.6, p < .001, partial η² = 0.65.

Output	What It Tells You
F statistic (with corrected df)	The ratio of variance explained by time to residual variance, using degrees of freedom adjusted for the sphericity violation
P-value	The probability of seeing this much change across time if there were truly no overall time effect
Partial η² (effect size)	Proportion of variance explained by time; roughly 0.01 small, 0.06 medium, 0.14+ large
Time effect (omnibus result)	Confirms that SBP changed significantly across the four visits overall — not which specific visits differ
Pairwise comparisons	Identify exactly which timepoint pairs differ significantly, with a multiple-comparison correction applied

Bonferroni-corrected pairwise comparisons for this example might show: Baseline vs Week 4, mean difference = 7 mmHg, 95% CI [3, 11], p = .002; Baseline vs Week 12, mean difference = 22 mmHg, 95% CI [17, 27], p < .001; Week 8 vs Week 12, mean difference = 8 mmHg, 95% CI [4, 12], p < .001 — confirming the reduction was significant at essentially every stage, with the cumulative effect by week 12 being by far the largest.

What to Use When Assumptions Are Violated

Violated Assumption	What to Do
Sphericity (Mauchly's test significant)	Apply the Greenhouse-Geisser (or Huynh-Feldt) correction to the degrees of freedom — same test, corrected p-value
Severe non-normality	Use the Friedman test, the non-parametric equivalent for 3+ related measurements
Substantial missing data across timepoints	Use a linear mixed-effects model, which can include subjects with partial data
More than one within-subject factor, or an added between-subjects group	Use a mixed-design (split-plot) ANOVA or a more general mixed model

How to Report the Methods and Results

A complete report states the design, whether sphericity held, the correction applied if needed, and the full result — omnibus test plus pairwise comparisons.

Sample Methods Text

"Systolic blood pressure was measured at baseline and at weeks 4, 8, and 12 in the same 24 patients. Mauchly's test indicated a violation of sphericity (χ² = 9.8, p = .03); degrees of freedom were therefore corrected using the Greenhouse-Geisser estimate (ε = 0.70)."

Sample Results Text

"There was a statistically significant effect of time on systolic blood pressure, F(2.1, 48.3) = 42.6, p < .001, partial η² = 0.65. Bonferroni-corrected pairwise comparisons showed a significant reduction at every subsequent visit compared with baseline (all p < .01), with the largest reduction observed between baseline and week 12 (mean difference = 22 mmHg, 95% CI [17, 27])."

Common Mistakes

Mistake 1: Running Multiple Paired T-Tests Instead of One RM-ANOVA

Comparing baseline-vs-week4, week4-vs-week8, and baseline-vs-week8 as three separate paired t-tests, inflating the overall false-positive rate.

✓ Fix: Run a single repeated measures ANOVA, followed by corrected pairwise comparisons if the omnibus result is significant.

Mistake 2: Ignoring Sphericity Entirely

Always reporting the uncorrected p-value without ever checking Mauchly's test, risking a falsely significant result.

✓ Fix: Check Mauchly's test every time you have three or more repeated measurements, and apply the Greenhouse-Geisser correction when it's significant.

Mistake 3: Silently Dropping Subjects With Missing Visits

Excluding subjects with any missing timepoint without reporting how many were dropped or why, obscuring potential attrition bias.

✓ Fix: Report the number of subjects excluded for missing data, and consider a linear mixed model if dropout is substantial.

Mistake 4: Stopping at a Significant Omnibus F

Reporting "there was a significant effect of time" without ever testing which specific timepoints actually differ from each other.

✓ Fix: Follow a significant omnibus result with corrected pairwise comparisons between timepoints.

Mistake 5: Confusing a Within-Subject Factor With a Between-Subjects Group

Treating a design that actually compares two different groups over time as a simple repeated measures ANOVA, missing the need for a mixed-design model.

✓ Fix: If a separate group variable exists alongside the repeated factor, use a mixed-design ANOVA instead.

Mistake 6: Not Checking Outliers at Individual Timepoints

Failing to inspect each timepoint separately for extreme values that could be distorting that timepoint's mean and the overall time effect.

✓ Fix: Screen each timepoint for outliers before running the omnibus test, not just the dataset as a whole.

❌ Incorrect

"Blood pressure decreased significantly over time (p < .001)" — reported without checking sphericity or specifying which visits actually differed.

✅ Correct

"There was a significant effect of time, F(2.1, 48.3) = 42.6, p < .001 (Greenhouse-Geisser corrected), with significant reductions at every visit compared with baseline (all p < .01)."

Frequently Asked Questions

What is the difference between repeated measures ANOVA and a paired t-test? +

A paired t-test compares exactly two related measurements from the same subjects. Repeated measures ANOVA extends this to three or more related measurements, testing for an overall difference in a single model rather than requiring multiple separate pairwise tests.

What is sphericity and why does it matter? +

Sphericity is the assumption that the variances of the differences between every pair of repeated measurements are roughly equal. It only applies with three or more timepoints. When violated, the standard F-test's p-value becomes too likely to appear significant, requiring a correction.

What happens if Mauchly's test is significant? +

A significant Mauchly's test (p < .05) means sphericity is violated. The standard response is to apply the Greenhouse-Geisser correction (or Huynh-Feldt when epsilon is closer to 1) to the degrees of freedom, producing a valid, more conservative p-value without switching tests.

Can repeated measures ANOVA handle missing data? +

Poorly — it typically requires complete cases, excluding any subject missing even one timepoint. A linear mixed-effects model can include subjects with partial data and is often preferred in longitudinal studies with dropout.

What is the non-parametric alternative to repeated measures ANOVA? +

The Friedman test, used for ordinal outcomes or severe non-normality that can't be addressed with transformation. It compares three or more related measurements using ranks and carries no sphericity assumption, at some cost in power.

How many time points do I need for repeated measures ANOVA? +

A minimum of three. With only two, the analysis reduces to a paired t-test and sphericity is a non-issue. There's no strict upper limit, though power and the number of pairwise comparisons both increase with more timepoints.

What effect size should I report for repeated measures ANOVA? +

Partial eta squared (partial η²) is standard for the overall time effect. Using Cohen's adapted benchmarks, around 0.01 is small, 0.06 medium, and 0.14 or above large — though repeated measures designs often show larger effect sizes than between-subjects designs.

Do I need post hoc tests after repeated measures ANOVA? +

Yes, whenever the omnibus time effect is significant with three or more timepoints. Corrected pairwise comparisons (commonly Bonferroni) identify exactly which specific timepoints differ, since the omnibus test alone doesn't specify that.

What is the difference between repeated measures ANOVA and a mixed-design ANOVA? +

Repeated measures ANOVA involves one group of subjects measured repeatedly, with only a within-subject factor. A mixed-design (split-plot) ANOVA adds a between-subjects factor, such as comparing treatment and control groups both followed over the same time points.

Should I use repeated measures ANOVA or a linear mixed model? +

Repeated measures ANOVA works well for complete, balanced data with no missing timepoints. A linear mixed model is generally more robust when data are missing, when time is better treated as continuous, or when the design is more complex — many methodologists now recommend mixed models as the default for longitudinal data.

Once your repeated measures ANOVA is run, these guides take you through the surrounding decisions:

Ready to Run Your Repeated Measures ANOVA?

Let StatClinic's AI Statistical Assistant confirm the right test, check sphericity automatically, and write your results sentence. Free, no registration required.

Try StatClinic Free →

Patient ID	Baseline SBP	Week 4 SBP	Week 8 SBP	Week 12 SBP
001	158	150	142	134
002	146	140	136	128
003	162	153	145	137
...	...	...	...	...

Patient ID	Baseline SBP	Week 4 SBP	Week 8 SBP	Week 12 SBP
001	158	150	142	134
002	146	140	136	128
003	162	153	145	137
...	...	...	...	...

How to Perform Repeated Measures ANOVA in Medical Research

What Is Repeated Measures ANOVA?

When Should You Use Repeated Measures ANOVA?

Repeated Measurements and Within-Subject Designs

Repeated Measures ANOVA vs Ordinary ANOVA and the Paired T-Test

Suitable Data Structure

Assumptions, Especially Sphericity

Continuous outcome

Independence between subjects

Approximate normality

No severe outliers

Sphericity

Mauchly's Test and the Greenhouse-Geisser Correction

How to Prepare the Dataset

A Realistic Longitudinal Medical Example

Step-by-Step Analysis Using StatClinic

Enter your data in wide format

Confirm your study design

Let StatClinic run Mauchly's test automatically

Apply the correction if needed

Review the omnibus time effect and effect size

Get pairwise comparisons and a written interpretation

Interpreting the F Statistic, P Value, Effect Size, Time Effect, and Pairwise Comparisons

What to Use When Assumptions Are Violated

How to Report the Methods and Results

Common Mistakes

Mistake 1: Running Multiple Paired T-Tests Instead of One RM-ANOVA

Mistake 2: Ignoring Sphericity Entirely

Mistake 3: Silently Dropping Subjects With Missing Visits

Mistake 4: Stopping at a Significant Omnibus F

Mistake 5: Confusing a Within-Subject Factor With a Between-Subjects Group

Mistake 6: Not Checking Outliers at Individual Timepoints

Frequently Asked Questions

Related Articles

Ready to Run Your Repeated Measures ANOVA?

Patient ID	Baseline SBP	Week 4 SBP	Week 8 SBP	Week 12 SBP
001	158	150	142	134
002	146	140	136	128
003	162	153	145	137
...	...	...	...	...