QC flags are not exclusions¶
A quality-control procedure can detect unusual or potentially problematic observations. That does not make the procedure an automatic deletion rule.
Three different objects¶
- Observed data — what was recorded.
- QC evidence — diagnostics, flags, scores, and warnings derived from those observations.
- Review/exclusion decisions — explicit decisions about whether an observation contributes to a defined analysis.
Collapsing these objects makes it difficult to reconstruct why denominators changed.
What should remain visible¶
A defensible QC workflow records the diagnostic, threshold or rule, affected unit, review status, final decision, rationale, and denominator effect.
Why this matters statistically¶
Exclusion can alter group composition, exposure, event counts, dwell denominators, latency risk sets, and missing-data patterns. A rule applied after inspecting the result can also change the interpretation of uncertainty.
GazeForge rule¶
QC functions should produce evidence for review. They should not silently mutate the source table into a preferred analysis sample.