Choosing an approach for repeated measures
Choose a repeated-measures approach from the clinical contrast, outcome type, visit pattern and missing-data assumptions; correlated visits are not independent observations.
Fit
When it matters, and when it does not
It matters when
- Each patient is measured at several visits, such as symptom scores at 1, 3 and 6 months.
- You want to compare how the outcome changes over time between groups.
- Some patients miss visits, so the choice of method must fit a stated missing-data assumption.
It matters less when
- Each patient is measured once at baseline and once at follow-up. Regress the follow-up value on group and baseline, which is called ANCOVA. Read: ANCOVA
- The outcome is the time until an event, not a value measured at each visit. Use survival analysis, such as Cox regression. Read: Cox proportional hazards regression
- You have already chosen a model and need its fitting details. Use the mixed-effects or GEE method page. Read: Mixed-effects models
Before you start
What to check in your study
- Which difference you will comparesummaryMeasureThe difference at one set visit, the difference in slopes, or the area under the curve. Fix it before seeing the data.
- The visit schedule, and how far real visits may driftrepeatedMeasuresWhen visits drift, use the actual time since baseline, so each value sits at the right time.
- The type of outcomeoutcomeTypeContinuous, yes or no, or a count. It sets the model and the effect measure.
- Why visits might be missedName the assumption, such as missing at random, and choose the method because of it.
- How visits from one patient relateA random intercept, a random slope, or a correlation pattern across visits. The choice affects the standard errors.
- The analysis allows for visits from one patient being alike.How to check: Give each patient a random effect, use GEE with robust standard errors, or model the correlation across visits.
- Missed visits are missing at random: given the recorded values, missing does not depend on the unseen value.How to check: Mixed models are valid under it. If sicker patients may drop out, add a sensitivity analysis, such as a tipping-point analysis.
- The model's shape over time fits the data.How to check: Plot group means and some patients' lines. If a straight line does not fit, treat visits as categories or add a curve.
- With GEE, missed visits depend only on factors in the model.How to check: Plain GEE needs this. If dropout follows earlier results, use weighted GEE or a mixed model.
Plain language
How it works
First decide what you want to compare: an outcome at a set visit, change over time, or a patient-level summary. Then choose an approach that respects visits within patients. Mixed-effects models can describe patient-specific trajectories. GEE targets an average across patients. A prespecified summary can answer a simpler question, but discards the time pattern.
A continuous baseline and one follow-up may suit ANCOVA. Several visits need a plan for time and within-patient dependence. The mixed-effects and GEE pages teach their methods; the Choosing comparison explains their different targets. Missed visits also matter. Name why values are missing before choosing an analysis, and test departures from that assumption.
Worked example
A worked example
- Setting
- 240 adults with asthma, randomized to a new inhaler regimen or usual care, with symptom scores at baseline and at 1, 3 and 6 months. Higher scores mean worse symptoms.
- Question
- Does the new regimen lower symptom scores more than usual care by 6 months?
- Approach
- A mixed-effects model uses every visit recorded, with a random intercept for each patient. The estimand, fixed in advance, is the difference between groups in change from baseline to 6 months.
- Result
- Mean scores fell from 21 to 12 with the new regimen, and from 20 to 16 with usual care.
- What it means
- Scores fell 9 points with the new regimen and 4 with usual care, a difference in change of 5 points. Patients with missed visits still count, which is valid if visits are missing at random.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Two measurements on the same patient are more alike than two measurements on different patients, and an analysis that ignores that treats each visit as new information it is not, producing standard errors that are too small. The other common error runs in the opposite direction: analyzing only the last visit, or only the change from first to last, discards most of the data collected and usually most of the power paid for.
- Choosing the visit or summary to report after seeing the plot.Instead: Fix the estimand in advance. A trial with no overall effect can show one significant visit by chance.
- Using repeated-measures ANOVA by default.Instead: It drops every patient with a missed visit and assumes sphericity. A mixed model or GEE keeps partial records.
- Placing every visit at its planned time when real visits drifted.Instead: Use the actual time since baseline, or show that the drift is small.
Inside Mallard
How Mallard uses this
When your plan records repeated measures of a continuous outcome, Mallard's method list may include a linear mixed model for repeated measurements.
concept.choosing_repeated_measures, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Repeated Measures Designs and Analysis of Longitudinal Data: If at First You Do Not Succeed-Try, Try Again. (opens in a new tab) · Anesth Analg · doi 10.1213/ANE.0000000000003511
- Statistics notes: Analysing controlled trials with baseline and follow up measurements (opens in a new tab) · BMJ · doi 10.1136/bmj.323.7321.1123