Case studies / Life sciences
Linear Mixed Effects Models for an Intervention Measured Over Six Occasions
Recovering a defensible effect estimate from repeated measures with missing follow-up, where repeated measures ANOVA could not be used.
The challenge
What the organisation came with
A research organisation needed to establish whether an intervention remained effective once participant characteristics were taken into account.
Each participant was measured on six occasions, which creates two problems at once. Repeated measurements on the same person are correlated — someone who scores high at occasion one tends to score high at occasion two — so the observations are not independent, and treating them as if they were understates the standard errors and overstates significance.
The second problem was missing follow-up. Not every participant was measured on every occasion. Attendance drops off, and the resulting dataset is unbalanced. That rules out the method most researchers reach for first.
The approach
How it was analysed, and why that method
Why mixed effects rather than repeated measures ANOVA
Repeated measures ANOVA was rejected outright. It assumes complete, balanced data: every participant measured at every occasion. Faced with missing follow-up it either discards every participant with a single missing measurement, or requires imputation before the analysis can run at all. With 350 participants and six occasions, listwise deletion would have removed a substantial and non-random portion of the sample.
Linear mixed effects models handle unbalanced data directly. Participants contribute whatever measurements they have, and a random effect for participant captures the correlation between that person's repeated measurements — modelling the dependence explicitly rather than pretending it does not exist.
Participants who drop out are rarely a random sample of those who stay. A method that silently excludes them does not just lose power; it changes what the remaining sample represents.
Adjustment and sensitivity
Participant characteristics were included as fixed effects so the intervention estimate could be read as adjusted rather than raw. Sensitivity analyses were then run to establish how far the conclusion depended on the modelling choices — a result that survives a reasonable range of specifications is one you can defend; one that appears under a single specification is not.
The analysis was carried out in SAS and R, with the code annotated so it could be rerun and checked independently.
Delivered
What the client received
The value
What changed as a result
The findings informed the next phase of the project and the decisions that followed it.
The sensitivity work was as important as the headline estimate: it established the range of assumptions under which the conclusion held, which is what a reviewer or funder asks about first.
Written from the assigned statistician's own project notes. The client is not named and no identifying detail, data or figures are published. Scale and timeframe are approximate.
Related services
The services behind this work
Statistical consultancy
Design, analysis and reporting for charities, universities, the NHS and business.
From £950SPSS data analysis
The right tests, annotated output, reusable syntax and APA-ready tables.
From £95Power analysis & sample size
A priori power analysis with the justification an ethics committee will accept.
From £180Send the data. Get a fixed quote.
Attach your dataset or just describe the project. A named statistician replies with a price and a deadline, usually within one working day.