UitbreidenOS When to activate
- Designing A/B tests with proper sample size and duration
- Analyzing experiment results for statistical significance
- Interpreting metrics lift, confidence intervals, and guardrail metrics
- Setting up experimentation frameworks (Statsig, LaunchDarkly, custom)
When NOT to use
- For qualitative user research
- For feature flagging without measurement
- For simple before/after comparisons
Instructions
- Pre-experiment. Define primary metric, guardrail metrics, minimum detectable effect (MDE), and required sample size.
- Power analysis. Calculate sample size: n = (Z_α/2 + Z_β)² × σ² / δ². Typical: 80% power, 5% significance.
- Run experiment. Randomize at user/session level. Check for sample ratio mismatch (SRM). Monitor guardrails daily.
- Analyze results. Compute p-value, confidence interval, relative lift. Check for novelty/primacy effects.
- Interpret. Statistical significance ≠ practical significance. Consider effect size and business impact.
- Report. Executive summary: metric, lift %, CI, p-value, recommendation (ship/iterate/kill).
Example
Experiment: New checkout flow
Primary metric: Conversion rate
MDE: 2% relative lift
Sample size: 50K users/group (14 days at current traffic)
Results:
Control: 3.21% conversion
Treatment: 3.34% conversion
Lift: +4.1% (95% CI: [+1.2%, +7.0%])
p-value: 0.008 (< 0.05 threshold)
Guardrails: Page load time unchanged, error rate unchanged
Recommendation: Ship. Statistically significant with positive lift.