statistical-qa-of-metricslisted

# Skill: statistical-qa-of-metrics > **Invoked by:** `applied-statistician` (primary) **and** `data-platform/dashboard-builder` (the seam — when a widget shows a comparison/trend that needs a significance/CI annotation). > > **The seam:** `data-platform` owns *"is this number correct?"* — present, in-range, reconciled, fresh (its [`data-quality-tests`](../../../data-platform/skills/data-quality-tests/SKILL.md) skill). **This skill owns *"is this number real?"*** — is the movement signal or sampling noise; does the comparison have enough data; is the trend honest. Non-overlapping by design. > > **When to invoke:** "revenue is up 18% — is that real or noise?"; "this KPI tile shows a WoW change — should we annotate it?"; "the dashboard shows variant B winning — can we trust it?" > > **Output:** a signal-vs-noise verdict + the annotation to put on the widget (CI, significance flag, or "insufficient data") + the caveat. ## Procedure 1. **Classify the widget's claim:** a *point comparison* (this period vs last), a *trend* (slope over time), or a *group comparison* (segment A vs B). 2. **Get the denominators.** A "+18%" on n=40 is very different from n=40,000. No sample size → the only honest annotation is "insufficient data to call." 3. **Attach uncertainty, not just a point estimate:** - **Rate/proportion KPI** → Wilson confidence interval on the rate; flag whether the period-over-period change's CI excludes zero. - **Mean/continuous KPI** → CI on the difference (t-based