A/B Test on Monthly Churn: Inference, Power, and Testing Choices
You ran a 28-day A/B test to reduce monthly churn among subscribers, randomizing 150,000 users per arm. The observed churn rates were:
-
Control: 5.00%
-
Treatment: 4.85%
A pre-specified two-sided test reported p = 0.10 at α = 0.05.
Tasks:
-
Compute the 95% confidence interval for the absolute churn difference (treatment − control, in percentage points) and interpret it.
-
Compute the post-hoc power to detect a −0.15 pp effect with these sample sizes.
-
Explain why p = 0.10 is not “a 10% chance the null is true,” and state what it does mean.
-
State when a one-sided test would be appropriate if pre-registered, approximate its p-value here, and explain why switching to one-sided post hoc is invalid.
-
Name two design/analysis changes (e.g., covariate adjustment, stratification, CUPED) that could legitimately reduce variance and potentially move p below 0.05, and when they are valid.