Scalar mixed-model sensitivity summary v0.1

Enable the optional secondary summary with:

[analysis.inference]
intent = "exploratory"
randomized = false
scalar_mixed_model = true
acknowledged_assumptions = [
  "approximately_gaussian_unit_differences",
  "complete_pairs",
  "estimand_matches_question",
  "independent_aggregation_units",
]

Install published-package support with pip install "fiberphotometry[stats]". The CLI then writes mixed-model.json and a standalone mixed-model.html evidence view, fingerprints both in the manifest, and embeds the JSON in NWB export.

The model uses an event-level fixed contrast, animal random intercept and condition slope, and a nested session random intercept when more than one session occurs within at least one animal. If that session component is not estimable it is omitted with an explicit warning. The optimizer, REML policy, convergence, normal-theory interval, variance estimates, engine version, and input hash remain machine-readable.

This is a sensitivity estimand. It does not replace the primary animal-level analysis or guarantee the same weighting of sessions and trials. Disagreement is reported for scientific inspection.

The frozen v0.2 parity fixture matches direct statsmodels estimates, standard errors, and confidence intervals in unbalanced-event and unbalanced-nested-session scenarios. This validates numerical plumbing, not interval coverage or universal model adequacy. Statsmodels documents the random-effects and variance-component formulation and notes that its standard fixed-effect confidence intervals are normal-theory intervals:

• https://www.statsmodels.org/stable/generated/statsmodels.formula.api.mixedlm.html
• https://www.statsmodels.org/stable/generated/statsmodels.regression.mixed_linear_model.MixedLMResults.html

The method's role is governed by SDR-0010.