← LinkedIn Interview Insights

LinkedIn·Data Scientist·Technical Phone Screen·Senior

Senior
Jul 2026

Summary

LinkedIn data scientist interview with a meaty A/B testing case built around Simpson's paradox. One question, but it had a lot of layers and the follow-ups kept coming.

Questions Asked (1)

Q1

A marketing team ran an A/B test on two email versions across two cities and two weeks. The pooled data shows version A winning, but version B wins in every city-level and week-level breakdown. How do you figure out which email is actually better? Walk through your choice of primary and guardrail metrics, how Simpson's paradox shows up here, what confounders you'd investigate, how you'd reanalyze the data, whether confidence intervals are still valid, and how you'd redesign the experiment if the original allocation was imbalanced.

A/B Testing & ExperimentationProduct Analytics & MetricsRoot Cause Analysis
Author's notes

This one sprawled in a way I wasn't ready for.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, recognize this as Simpson's paradox and explain that the pooled result is misleading due to confounding between version and city/week. Then, propose a stratified analysis or a model that adjusts for city and week, and discuss how to check for and address allocation imbalance, including whether confidence intervals from the original analysis are trustworthy.

Pro tip: Emphasize that Simpson's paradox often arises from unequal sample sizes or allocation ratios across strata; always check the distribution of users across versions within each city and week before trusting pooled results.

1. Define metrics

Choose a primary metric (e.g., click-through rate or conversion rate) that directly measures email effectiveness, and guardrail metrics (e.g., unsubscribe rate, spam complaints) to ensure no harm.

2. Diagnose Simpson's paradox

Show that the pooled result reverses when stratifying by city and week, indicating confounding due to unequal allocation or different baseline rates across strata.

3. Investigate confounders

Check if version assignment is correlated with city or week (e.g., version A sent more to a city with higher baseline response), and examine other potential confounders like time of day or user demographics.

4. Reanalyze with stratification or modeling

Use stratified analysis (e.g., Mantel-Haenszel) or regression with fixed effects for city and week to estimate the true effect, and compute adjusted confidence intervals.

5. Redesign if imbalanced

If allocation was imbalanced, propose randomization within each city-week block, ensure equal sample sizes, and pre-register the analysis plan to avoid post-hoc bias.

Key Points to Mention

  • Simpson's paradox: aggregated trend reverses when data is partitioned.
  • Primary metric: click-through or conversion rate; guardrail: unsubscribe or complaint rate.
  • Confounders: city, week, and their interaction with version assignment.
  • Stratified analysis or regression with fixed effects to adjust for confounders.
  • Validity of confidence intervals: pooled CIs are biased; use adjusted CIs from stratified or model-based analysis.
  • Redesign: block randomization by city and week, equal allocation, and pre-registration.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.