references/real-vs-noise.md
Is a downstream effect real or noise?
Symptom: a funnel step the feature doesn't touch (e.g. event page → checkout) shows a lift,
often while the step it does touch is flat. The customer suspects a measurement bug.
Why that leg isn't a clean comparison
A rate measured between two mid-funnel steps only counts users who already reached the earlier step, and reaching it happens after randomization and can be nudged by the treatment. So the leg compares two groups shaped by the experiment, not the randomized groups, and a gap can appear there with no real effect on the downstream action. It can even read more significant than the honest metric, because it's a smaller, more-selected denominator.
Trust the randomized endpoint — exposure → final step, counting everyone assigned — over any rate
measured between two mid-funnel steps. PostHog computes significance from the first step to the last
step for exactly this reason (see numbers-vs-sql.md D2 in the diagnosing-experiment-health
library).
Three checks — any one failing points to noise
- Non-user split. Recompute the effect among users who never fired the feature-interaction event. If the between-variant gap is still there, the feature can't be causing it. State it as a between-variant comparison restricted to non-users — not "non-users convert more".
- Dose-response. Rank the variants by actual feature usage. A real effect is strongest where the feature is used most; an effect that's absent in the high-usage arm and present in a barely-used arm is not causal.
- Cohort stability. Split the outcome by week of first exposure. A real effect holds its sign and rough size across cohorts; one that flips ahead/behind or lives in a single cohort is noise.
Recommend
Report the randomized exposure → outcome number (with its win probability) as the verdict, explain the conditioning trap in plain terms, and — if the checks point to noise — advise against shipping on the downstream figure and to keep running to the pre-planned sample.