Skip to content

SDR-0046: Require explicit pairs and shared evidence for multisignal analysis

  • Status: Accepted
  • Date: 2026-07-27

Context

Multi-site and multi-color photometry can produce high correlations for biological and non-biological reasons. Shared movement, task events, bleaching, reference channels, detector paths, spectral bleed-through, and common preprocessing can all create apparent coupling. Separate invalid samples or clocks make the denominator ambiguous. Treating fibers, sessions, windows, or frequency bins as independent animals compounds that ambiguity.

Cross-correlation, coherence, and phase also answer different questions. A peak lag is not automatically a transmission delay; high coherence does not imply causal interaction; and optical contamination cannot be established or excluded from correlation alone.

Decision

Every multisignal analysis begins with an ordered, explicitly identified signal pair and one shared timestamp vector after any separately documented synchronization or resampling.

  • Channel metadata declares site, sensor, role, units, wavelength and optional detector, fiber, and coordinate identity; roles are never inferred.
  • Invalidity in either signal, any declared covariate, or the shared clock splits joint continuity. Cross-channel pairs and spectral windows never cross a gap.
  • Optical/crosstalk diagnostics retain metadata flags, control loading, raw association, and control-residualized association. Status is no_flag or review; absence of a flag is not evidence that crosstalk is absent.
  • Lagged association reports physical lag convention and pair counts. Optional event/behavior residualization fits the declared design separately within each continuity run and retains coefficients, rank, and variance explained.
  • A within-session null may randomly re-pair complete temporal blocks, but pairs are still formed only within blocks and the block duration must exceed twice the lag range.
  • Coherence and phase pool power and complex cross-spectra by complete Welch-window count before deriving coherence. Run-level coherence values are not averaged.
  • State conditioning consumes non-overlapping user-supplied epochs and preserves each epoch as a separate joint-continuity partition.
  • Group inference begins from one declared scalar per session, aggregates sessions within animal-condition cells, and resamples or permutes animals.

Alternatives considered

  • Infer channel meaning from wavelength or column order. Rejected because acquisition conventions do not establish biological role.
  • Correlate independently cleaned signals on compressed timestamps. Rejected because different masks and removed gaps create artificial adjacency.
  • Automatically regress the reference channel. Rejected because correction changes the estimand and the reference may contain biological or wavelength- specific structure.
  • Average coherence across runs. Rejected because runs with one and many Welch windows would receive equal weight and coherence is nonlinear in the spectra.
  • Treat a near-zero peak as crosstalk. Rejected because common biological input and task locking can produce the same pattern.
  • Treat sites or sessions as independent replicates. Rejected because the animal is the experimental unit for ordinary group comparisons.

Consequences

The API requires more metadata and retains more evidence than a two-array correlation helper. Joint missingness can reduce support substantially, and short runs may contribute no coherence windows. Covariate-adjusted association remains conditional on the supplied design. Blocked and animal-level randomization answer different questions and are reported separately.

Coordinates can be stored prospectively, but pairwise results do not yet provide a spatial covariance model for dense arrays. Crosstalk correction, spectral unmixing, and hemodynamic models remain separate unimplemented transformations.

Revisit trigger

Revisit after validation on two independently documented dual-site or dual-color datasets with raw optical metadata and controls, or when a dense-array dataset requires a mouse-aware spatial model rather than pairwise summaries.