Skip to content

SDR-0010: Scalar mixed models are sensitivity summaries

  • Status: Accepted
  • Date: 2026-07-26

Context

An event-level mixed model and an equally weighted animal/session contrast need not estimate the same quantity. Mixed models also require random-effects choices, optimizer convergence, and normal-theory interval assumptions. Treating them as a universally superior replacement would hide rather than resolve those differences.

Decision

The initial scalar mixed model is opt-in and labelled role = "sensitivity_analysis". It fits the declared two-level event contrast using animal random intercepts and condition slopes, with a nested session random intercept when multiple sessions per animal make it estimable.

The artifact reports convergence, optimizer, fixed-effect interval, random-effect variances, warnings, engine version, and input fingerprint. It remains separate from the primary design-aware result.

Consequences

Scientists can inspect agreement without allowing the mixed model to silently change the primary estimand. Nonconvergence remains visible. Current intervals are statsmodels normal-theory fixed-effect intervals, not simultaneous waveform bands or validated small-sample coverage guarantees.

Revisit trigger

Revisit after coverage calibration, alternative degrees-of-freedom methods, or functional mixed-model parity provides evidence for a different default role.