A/B Test Analysis
Experiments need pre-registered metrics, healthy assignment, and honest uncertainty. Do not crown winners from noisy or peeked data without caveats.
Workflow
- Confirm hypothesis, primary metric, guardrails, unit of randomization, and duration.
- Check data quality: sample ratio mismatch, bots, nulls, novelty effects.
- Compute lift with uncertainty (CI); avoid p-hacking language.
- Segment carefully; treat segments as exploratory unless pre-registered.
- State decision: ship / hold / iterate, with risks.
Output format
## Experiment read: <name>
**Primary metric:** …
**Result:** …
**Decision recommendation:** ship | hold | iterate
### Health checks
- SRM: …
- Coverage / exclusions: …
### Estimates
| Variant | N | Metric | Lift | 95% CI |
### Guardrails
…
### Caveats
…
Rules
- Never invent statistical significance; if numbers are missing, say what to compute.
- Multiple metrics inflate false positives — say so when user looks at many.
- Distinguish practical significance from statistical significance.
- SRM failure means pause interpretation of outcomes.
- Peeking early without sequential design → caveat strongly.
- Correlation of segments is not automatic heterogeneity proof.
Edge cases
- Switchback / network effects: note interference when users interact.
- Low power: recommend more runtime or larger effect — do not fake certainty.
- Non-inferiority: different decision frame than superiority tests.