DEPLOYEDAI-era interview training
RankBootstrapping
XP 0 / 250
0
0%
Ready
Data Analyst & Analytics Interview · Statistics & Experimentation · Lesson 3 of 3

A/B testing: design & pitfalls

8 min

An A/B test is a randomized controlled experiment: randomly split users into control and treatment, change one thing, and compare a pre-declared metric. Randomization is what lets you claim causation, not just correlation.

Designing it

  1. Pick one primary metric and a hypothesis before you start.
  2. Do a power / sample-size calculation: the minimum detectable effect, alpha (0.05), and power (0.8) determine how many users and how long you must run.
  3. Randomize at the right unit (usually user, not session) and check the split is balanced (an A/A test).
  4. Run for full business cycles (whole weeks) and decide the stop date in advance.

Where it goes wrong

  • Peeking / early stopping: checking daily and stopping at the first 'significant' result inflates false positives dramatically.
  • Multiple comparisons: test 20 metrics at alpha 0.05 and ~1 will look significant by chance — correct for it (Bonferroni / FDR).
  • Novelty & primacy effects: users react to change itself; short tests overstate impact.
  • Sample ratio mismatch (SRM): a 50/50 split that arrives 53/47 signals a broken assignment — invalidate and investigate.
  • Network / spillover effects: treatment leaks to control (marketplaces, social) — use cluster randomization.
Peeking is the #1 killerStopping a test the moment it crosses p < 0.05 inflates the 5% false-positive rate several-fold — easily 20%+ in simulations of daily peeking. Fix the sample size up front, or use a sequential testing method designed for continuous monitoring.
The senior answerWhen asked 'is this result real?', a strong candidate checks: was the sample size pre-computed, was there peeking, is the split balanced (SRM), were multiple metrics corrected, and is the effect big enough to ship — not just statistically significant.

◆ Lock it in

  • Randomization at the user level is what licenses a causal claim.
  • Compute sample size / power up front; run full weekly cycles; fix the stop date.
  • Guard against peeking, multiple comparisons, novelty effects, and SRM.
Feynman drill — say it out loudExplain why 'peeking' at an A/B test daily and stopping when p < 0.05 inflates false positives.
Step 1 rate your confidence · Step 2 pick your answer
Your experiment was meant to be a 50/50 split but assignment logs show 55/45. What should you do first?
Step 1 — how sure are you?