When to Use Mixed Models and When RM-ANOVA - An Argument for Statistical Parsimony
Abstract
When should researchers use linear mixed-effects models (LMMs) versus repeated-measures analysis of variance (RM-ANOVA) for analysing repeated-measures data? I first outline the conceptual and statistical differences between the two approaches: RM-ANOVA operates on aggregated difference scores, whereas LMMs model trial-level data using fixed and random effects to account for dependencies among observations. I then describe the key limitations of RM-ANOVA: its inability to handle missing data, partially within-participants factors, continuous within-participants covariates, and complex dependency structures such as crossed random effects designs. I also discuss the conceptual and computational challenges associated with LMMs, such as model specification and convergence issues. Using two simulation studies, I then demonstrate that when both methods are appropriate, RM-ANOVA -- especially with Greenhouse-Geisser correction -- generally provides better control of Type I error rates and statistical power equal to or greater than that of LMMs. Consequently, I recommend using RM-ANOVA in designs where RM-ANOVA is appropriate. LMMs should be reserved for cases that require the additional flexibility, such as complex designs, missing data, or advanced modelling goals.
Related
- Certified Alpha Capacity: Statistical Arbitrage When Learning Takes Time
- When Stress Tests Miss the Risk: Statistical Scenario Analysis for Financial Portfolios
- When is statistical evidence strong enough? Using hypothesis tests to value data collection
- When and Why Do Linear Bias Probes Fail? A Geometric and Statistical Theory of Bias Detectability in Large Language Model Representations
- When Can LLM Digital Twins Reduce Human Measurement? From Behavioral Fidelity to Statistical Substitutability
BibTeX
@misc{singmann2026when,
title = {{When to Use Mixed Models and When RM-ANOVA - An Argument for Statistical Parsimony}},
author = {Henrik Singmann},
howpublished = {PsyArXiv},
year = {2026},
month = oct,
doi = {10.31234/osf.io/e7bjz_v1},
url = {https://osf.io/preprints/psyarxiv/e7bjz_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - When to Use Mixed Models and When RM-ANOVA - An Argument for Statistical Parsimony
AU - Henrik Singmann
PY - 2026
DA - 2026/10/02
PB - PsyArXiv
DO - 10.31234/osf.io/e7bjz_v1
UR - https://osf.io/preprints/psyarxiv/e7bjz_v1/
AB - When should researchers use linear mixed-effects models (LMMs) versus repeated-measures analysis of variance (RM-ANOVA) for analysing repeated-measures data? I first outline the conceptual and statistical differences between the two approaches: RM-ANOVA operates on aggregated difference scores, whereas LMMs model trial-level data using fixed and random effects to account for dependencies among observations. I then describe the key limitations of RM-ANOVA: its inability to handle missing data, partially within-participants factors, continuous within-participants covariates, and complex dependency structures such as crossed random effects designs. I also discuss the conceptual and computational challenges associated with LMMs, such as model specification and convergence issues. Using two simulation studies, I then demonstrate that when both methods are appropriate, RM-ANOVA -- especially with Greenhouse-Geisser correction -- generally provides better control of Type I error rates and statistical power equal to or greater than that of LMMs. Consequently, I recommend using RM-ANOVA in designs where RM-ANOVA is appropriate. LMMs should be reserved for cases that require the additional flexibility, such as complex designs, missing data, or advanced modelling goals.
ER -