Statistics (optional)¶

The genotype comparison of the paper is not part of the automated run. It is two existing stages that can be switched on:

in stride tmaze: steps: {mixed_models: true, classifier: true} plus stats_config (results under stats/). The metadata needs a genotype column;
with the stage-by-stage runner: stages 7-9 and 10 of run_all.py (Stage-by-stage batch runs).

Both read the batch / statistics config (Configuration reference).

Trial phase tagging¶

Each stride is matched to its trial by the video name, using the decision table. A stride whose frame_start lies between gate_frame and entry_frame (inclusive) is a task stride. Every other stride of a matched video is a return stride. A return stride faster than twice the mouse's median task-stride speed is a poked return. Strides of videos without a decision row are unmatched; without a decision table every stride is untagged. The mixed models use task strides by default (--trial-phase task).

Mixed models¶

Two linear mixed models per metric and batch (statsmodels mixedlm, REML), with a random intercept per mouse. The metric and the covariates are z-scored.

model	fixed effects
M1	genotype + body length [+ sex]
M2	genotype + stride speed + body length [+ sex]

Genotype is compared with the reference genotype (statistics.reference_genotype, WT). The labels in statistics.genotype_order are used; a batch may contain any subset that includes the reference, and other labels are left out.
Sex is added when the batch has has_sex: true.
The metrics are the metrics: list of the config (14 columns; step_freq_hz is derived as 1000 / duration_ms). In M2 a metric that is itself a covariate is skipped.
A fit is not produced when fewer than 100 strides have all values present, or when a standard error of the fit is above 10. lmm_fit_log.csv lists every fit, kept or not, and why.
lmm_results_all.csv holds the coefficients (in the metric's units and standardized), standard errors and p-values, plus <genotype>_fdr: Benjamini-Hochberg over all genotype p-values of one batch and model. The heatmaps show -log10(p) next to the standardized coefficients and mark FDR below 0.05.

Classifier¶

A WT-versus-Tau logistic regression on per-mouse features. It needs mice labelled WT and Tau.

Features. Gait (6): Euclidean stride length, step frequency, snout sway, temporal symmetry, stride duration and step width, from the filtered strides with a speed of 15 to 25 cm/s, median per mouse. T-maze (8): accuracy (the share of correct trials, a mean), junction exploration, number of probes, probe duration, probe bias, probe-choice correlation, and the log10 choice and stem latencies (per-mouse medians).
Ranking. Features are ranked by p-value: the values listed under classification.gait_lmm_pvalues / tmaze_lmm_pvalues in the config where given, otherwise a WT-versus-Tau t-test on the per-mouse values. The ranking uses all mice: it is done outside the cross-validation folds.
Model. Standardized inputs, logistic regression with balanced class weights, 5-fold stratified cross-validation (seed 42), evaluated for the top K features, K = 1 to N. The K with the highest mean AUC is reported as optimal.
Outputs. regression_feature_ranking.csv, regression_auc_vs_k.csv with its figure, and ROC curves for the T-maze, gait and combined feature sets.

The shipped config.example.yaml has empty p-value tables, so with it every feature is ranked by the t-test.

Never pool raw centimetres across cameras

Rigs differ by a few percent in scale and by more in tracking noise. Compare genotypes within a rig, or z-score within cohort before pooling. Camera differences never explain biology-sized effects, but they do add offsets.