SDR-0056: Form interactions from animal-condition contrasts

• Status: Accepted
• Date: 2026-07-28

Context

A common photometry experiment compares treatment and control animals across two repeated conditions—for example pre/post, cue A/cue B, or rewarded/unrewarded. Testing all events in a group-by-condition model without preserving the animal boundary can make trial yield determine apparent population precision.

The scientific estimand is also easy to obscure. A treatment-control difference in one condition is not the same question as whether the within-animal condition difference changes between treatment groups.

Decision

Represent the interaction as an explicit difference in differences:

1. Materialize one estimate for every animal-condition cell.
2. Within each complete animal, subtract the declared condition denominator from the condition numerator.
3. Compare those within-animal differences between two disjoint animal groups.

The outer comparison reuses the independent-group population contract. It therefore resamples animals separately within group, reports a Welch standard error, retains pointwise group support, produces pointwise and simultaneous bands, and records leave-one-animal-out influence.

Group membership is a separate typed assignment and must be constant within animal. The software never infers groups from filenames, balance, or condition labels. Animals missing either condition remain in the cell ledger and exclusion list but do not enter the interaction estimate.

Consequences

Every included animal has equal population weight regardless of its event count. The interaction answers whether a within-animal condition contrast differs between groups. It does not establish a causal treatment effect unless assignment and the rest of the experimental design justify that interpretation.

The initial contract supports two groups and two repeated conditions. It does not fit arbitrary factorial models, partial pooling, continuous moderators, or generalized outcomes.

Alternatives considered

• Compare groups separately at each condition. Rejected because two separate tests do not test the interaction.
• Pool events in a group-by-condition regression. Rejected as the default because lower-level observations would control precision unless the complete hierarchical model were declared and validated.
• Infer group membership from labels. Rejected because assignment semantics are scientific metadata, not a property of the outcome table.

Revisit trigger

Add richer factorial or hierarchical models only with explicit estimands, exchangeability rules, missing-cell policies, and simulation coverage across the designs they claim to support.