ab-test-analysis

A/B Test Analysis

When to use

• An experiment has finished and the team needs a ship / no-ship recommendation
• Results look directionally positive but the team is unsure if they're statistically significant
• A test has been running for weeks without a clear winner and someone needs to decide whether to continue
• A new experiment needs sample-size planning before launch
• Results are disputed and need a rigorous, documented analysis

Process

1. Confirm test design — verify the hypothesis, the control and treatment definitions, the randomisation unit (user/session/device), the primary metric, any guardrail metrics, and the target split ratio.
2. Check for sample ratio mismatch (SRM) — run a chi-square test on the actual vs. expected split. If SRM is detected, stop and investigate the randomisation pipeline before interpreting results. Use scripts/ab_test_analyzer.py --check-srm.
3. Calculate per-variant metrics — compute the rate (or mean) and 95% confidence interval for the primary metric in each variant. Document absolute and relative difference.
4. Run the significance test — execute a two-proportion z-test (for rates) or Welch's t-test (for means). Record z-score, p-value, and 95% CI for the effect. Use references/statistical_tests_reference.md if unsure which test applies.
5. Check guardrail metrics — run the same significance test for each guardrail metric. A significant degradation on any guardrail is a blocker regardless of primary metric results.
6. Produce the recommendation — synthesise SRM result, power, significance, and guardrail checks into a clear ship / no-ship / extend decision. Quantify the expected business impact if shipped. Record in assets/ab_test_report_template.md.

Inputs the skill needs

• Test plan or hypothesis document (variant definitions, randomisation unit, primary metric)
• Data with at minimum: user_id, variant assignment, primary metric outcome
• Optional: guardrail metric values per user, daily aggregate data for temporal validity checks
• Target split ratio (e.g., 50/50)
• Minimum detectable effect or business threshold for "worth shipping"

Output

• scripts/ab_test_analyzer.py — runs SRM check, significance test, power analysis, and guardrail checks from a CSV or summary stats input
• references/statistical_tests_reference.md — which test to use and when
• references/ab_test_design_guide.md — SRM causes, power planning, peeking and multiple testing
• assets/ab_test_report_template.md — structured report: design, results, checks, recommendation, expected impact