Skip to main content

Preview — Pro guide

You are seeing a portion of this guide. Sign in and upgrade to unlock the full article, quizzes, and interview answers.

Scenarios·Advanced

Scenario Walkthrough: The A/B Test Went Wrong — SRM, Peeking, and Interference

When the experiment is lying but the slide deck is green. Walk the failure modes that dominate production: sample ratio mismatch, peeking and early stopping, novelty and learning effects, network interference, wrong randomization unit, and thin-event metrics read too early. Teaches the remediation playbook: invalidate, debug assignment, or redesign — and how to say that without panicking the room. Grounded in standard OCE practice, chi-square SRM, and the Kohavi, Tang, and Xu body of work on trustworthy experiments.

45 min read 2 sections 1 interview questions
A/B TestingSample Ratio MismatchSRMPeeking ProblemEarly StoppingNovelty EffectInterferenceSUTVACUPEDOCEExperimentationScenario InterviewChi-Square TestStats IQ

The Question Under the Question

A broken test in interviews is usually one of a few classes. (1) Statistics — peeking, optional stopping, many metrics without multiplicity control, or reading a thin revenue line on a short window. (2) Engineering — SRM from bad assignment, eligibility applied after the bucket, redirect drops, or treatment-induced dropout. (3) Product — novelty in week one that inverts later. (4) Interference — shared inventory, ad auctions, or feeds so user-level SUTVA fails.

Weak candidates treat a small p-value as a ship bit. Strong candidates stop when identification breaks, run the SRM check before hero metrics, and separate monitoring from a pre-registered launch readout. They price a re-run or redesign against a bad OKR that can live for quarters, not against looking busy in SQL.

IMPORTANT

Premium content locked

This guide is premium content. Upgrade to Pro to unlock the full guide, quizzes, and interview Q&A.