SDR-0029: Treat grouped event-kernel intervals as conditional sensitivity
- Status: Accepted
- Date: 2026-07-27
- Decision owners: project maintainers
- Related evidence: public event-kernel reproduction
Context
Ordinary sample-level regression standard errors are inappropriate for densely sampled photometry because observations are temporally dependent and animals are the independent population units. The experimental model also selects a ridge penalty through grouped cross-validation. Uncertainty computed after that selection may omit selection variability, while pointwise intervals across many lags do not control simultaneous waveform error.
The first public diagnostic run found substantial held-out residual autocorrelation and selected the largest declared penalty. Delete-one-animal refits produced broad, structured pointwise intervals, some of which excluded zero.
Decision
Report delete-one-group jackknife intervals as sensitivity summaries for the pooled fixed-penalty estimator. Store the full coefficient, bias-corrected estimate, standard error, interval, omitted identities, confidence level, and conditioning penalty. Label the bands pointwise and non-simultaneous.
Do not convert zero-excluding lags into significance claims. Keep held-out residual diagnostics adjacent to kernel bands. Restart autocorrelation and Durbin–Watson calculations at each session boundary and use them descriptively, without p-values or automatic pass/fail thresholds.
Consequences
Scientists can see animal influence and temporal misspecification without false sample-level precision. The result remains useful for hypothesis generation and workflow comparison, but it does not yet provide calibrated population waveform inference. API schema version 2 makes the added uncertainty and diagnostic ledger explicit.
Alternatives considered
- Naive regression standard errors: rejected because they treat autocorrelated samples as independent and ignore the animal boundary.
- Trial bootstrap: rejected because trials are not independent population units.
- Present jackknife intervals as conventional confidence intervals: rejected because penalty selection and simultaneous-lag multiplicity remain unresolved.
- Suppress all uncertainty until a hierarchical model exists: rejected because transparent group-influence sensitivity is already materially useful when correctly labelled.
Revisit trigger
Revisit after repeated-sampling coverage calibration across realistic animal counts and heterogeneity, selective or nested model-selection treatment, and a validated simultaneous-band procedure.
Evidence added later
The frozen event-kernel interval calibration v0.1 found acceptable marginal pointwise coverage across all six scenarios, but the candidate simultaneous band missed its normalized-progress gate. SDR-0040 retains the candidate as a distinct opt-in without changing the pointwise default.