Population inference and robustness
This category applies across the signal, event, continuous, and multi-signal routes. It defines what counts as an independent unit and whether a conclusion depends on one defensible workflow choice.
Choose the workflow
| Need | Workflow | Status |
|---|---|---|
| Materialize animal estimates and contrast a population | Animal estimates and population contrasts | Supported contract; resampling is experimental |
| Test whether a within-animal condition contrast differs between groups | Group-by-condition animal interactions | Experimental two-group × two-condition contract |
| Apply one population contract to transients, state-band power, or association | Population inference across core workflows | Experimental typed materializers |
| Infer across complete PSD, autocorrelation, lag, or coherence curves | Population inference for frequency and lag curves | Experimental exact-axis vector contract |
| Select the experimental unit, estimand, and resampling design | Experimental design and inference | Supported guidance and APIs |
| Quantify sensitivity to named analytic alternatives | Robustness multiverses | Supported |
| Fit an optional scalar hierarchical sensitivity model | Scalar mixed models | Experimental |
| Inspect validation evidence rather than capability claims | Public-data evidence atlas | Maintained register |
Default hierarchy
Repeated samples and trials belong to sessions; sessions belong to animals. A workflow may model lower levels, but population uncertainty must not pretend those observations are independent animals. Predictive validation likewise holds out the complete group at the level of the intended generalization claim.
The common population contract begins after a domain has constructed auditable animal estimates. It retains source sessions, observation counts, pointwise support, exclusions, and leave-one-animal-out influence. Domain-specific denominators—such as finite recording exposure for transient rates—remain the responsibility of the originating analysis.
The cross-workflow materializers make that boundary executable. They do not make events, spectral windows, valid signal pairs, or sessions exchangeable; those quantities remain inspectable support for an animal-level cell.
For repeated conditions across independent treatment groups, the interaction route first forms one condition difference per complete animal and then compares those animal differences between groups. It does not substitute separate within-group tests for the interaction.
Coverage gaps this category exposes
- functional mixed models for complete time courses;
- small-sample corrections across more hierarchical designs;
- calibrated simultaneous inference after model selection;
- Bayesian hierarchical models with inspectable prior sensitivity; and
- meta-analytic evidence across laboratories and acquisition systems.
Robustness is not a vote across arbitrary analyses. Universes must preserve the estimand and denominator, declare compatibility rules before outcomes, and retain failures alongside successful results.