Repeated Measures Model Suggestion

Data with multiple observations per subject over time or across conditions —longitudinal data, crossover trials, pre-post designs —violate the independence assumption of standard ANOVA and t-tests. Repeated-measures ANOVA or linear mixed models are required to correctly handle the within-subject correlation. The standard one-way ANOVA applied to repeated measures is not just suboptimal; it is wrong.

STEP 1 —The Pitfall

Researchers often apply separate t-tests at each timepoint, or collapse repeated measurements into a single summary statistic (mean or AUC), without accounting for the repeated-measures structure. Separate timepoint tests inflate Type I error (multiple comparisons across time). AUC summaries discard the shape of the temporal trajectory. Neither correctly models the within-subject correlation.

The within-subject correlation in biological experiments is typically high (r = 0.6—.9). Ignoring it both inflates SE (losing power) and introduces biased estimates of treatment effects when the effect changes over time.

[Image placeholder: Line plot showing 5 subjects measured at 4 timepoints, two conditions. Individual trajectories drawn per subject. Left: separate t-tests at each timepoint with multiple comparison inflation. Right: linear mixed model with time xtreatment interaction, showing the full trajectory analysis with a single F-test for interaction.]

Separate timepoint tests (left): multiple comparisons inflate FWER. LMM with time xtreatment interaction (right): single model, correct within-subject correlation, single inference on the trajectory.

STEP 2 —Journal Requirement

Repeated-measures ANOVA (when sphericity is met) or linear mixed model (general case) is required for any design with multiple observations per subject. ARRIVE 2.0 requires that within-subject designs be analyzed with appropriate repeated-measures tests. Most methods-oriented journals will request revision if a longitudinal or crossover study is analyzed without accounting for within-subject correlation.

STEP 3 — What Licklider Currently Provides

What the current product does

The analysis_model_guidance_branch can surface a repeated-measures or mixed-model suggestion when the resolved analysis-unit context includes multiple timepoints.
For two timepoints: the branch suggests paired analysis, change scores, or paired-line views (paired_two_timepoint).
For three or more timepoints: the branch suggests repeated-measures model, mixed model, or spaghetti/trajectory views (repeated_measures).
The guidance appears as suggestion text, not as automatic model execution. The user decides whether and how to act on it.

What the product does not do

The product does not automatically classify the design as longitudinal / crossover / pre-post.
The product does not run RM-ANOVA, Mauchly's sphericity test, or Greenhouse-Geisser correction.
The product does not display LMM formulas or recommended model specifications.
The product does not raise a blocking "Design Mismatch" error. The guidance branch suggests safer alternatives but does not prevent the user from proceeding.

Known limitations

This is a guidance branch, not a standalone execution engine. The branch can be deterministic once its inputs (analysis-unit context and timepoint structure) are resolved, but the upstream design signals still depend on data-contract and observation context. It should be described as "can suggest" rather than "automatically selects."

For the detailed quality-checks documentation, see Repeated Measures Model Suggestion in the quality-checks docs.

[Image placeholder — not current product UI. Illustrative concept of a repeated-measures detection dialog.]

Illustrative concept only. The current product surfaces repeated-measures guidance via analysis_model_guidance_branch, not via this specific dialog layout.

STEP 4 — Draft Output (Draft / Needs review)

The following is an example of how a researcher might write a repeated-measures methods section, not output that Licklider generates automatically.

Data were analyzed using a linear mixed model (LMM) to account for
repeated measurements within subjects (lme4 R package; Bates et al., 2015).
The model included time, treatment, and their interaction as fixed effects,
and subject as a random intercept: outcome ~ time * treatment + (1 | subject).
Denominator degrees of freedom were estimated using the Kenward-Roger
approximation. All subjects contributed complete data across all timepoints.