← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Meta DS interview focused entirely on a ranking change question for a multi-account product feature. Two-part case: offline analysis first, then full A/B test design. Pretty deep for a single question but they really wanted to see how you think through the whole lifecycle from logs to experiment.

Questions Asked (2)

Q1

A product sorts a user's multiple accounts by most recent visit. There's a proposal to sort by number of notifications instead. Using historical logs and observational data, how would you evaluate whether this change is likely to be beneficial before running an experiment?

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

This part tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the goal: the change should improve user experience and engagement, not just notification handling. Then propose using historical logs to simulate the new sorting and compare key metrics like click-through rates and session time, while controlling for confounders and checking for novelty effects. Finally, discuss limitations and how to validate with a small experiment.

Pro tip: Emphasize that observational data can only show correlation, so you need to carefully define the counterfactual and consider using techniques like propensity score matching or instrumental variables to approximate causality. Also, mention that even if offline metrics look promising, you should run a small online experiment to detect unexpected effects.

1. Clarify the objective and metrics

Define what 'beneficial' means: likely increased user engagement (e.g., more account switches, higher session time) and satisfaction, without harming notification interaction. Identify primary and guardrail metrics.

2. Understand the data and simulate the change

Use historical logs to reconstruct user sessions and simulate the new sorting order. For each session, determine which account would be on top under the new rule and compare to the actual top account.

3. Analyze impact on key metrics

Compare metrics like click-through rate on the top account, time to switch accounts, and overall engagement between the simulated new order and the actual order. Use statistical tests to assess significance.

4. Address confounders and biases

Control for factors like user activity level, number of accounts, and notification frequency. Consider methods like stratification, regression, or matching to isolate the effect of sorting.

5. Assess limitations and recommend next steps

Acknowledge that observational data cannot prove causality. Recommend a small A/B test to validate findings, and discuss potential novelty effects or long-term impacts.

Key Points to Mention

  • Define clear success metrics (e.g., account switch rate, session duration, notification click-through).
  • Simulate the new sorting using historical logs to create a counterfactual.
  • Control for confounders such as user activity, number of accounts, and notification volume.
  • Use statistical methods (e.g., regression, matching) to estimate causal effect.
  • Acknowledge limitations of observational data and the need for an online experiment.
  • Consider novelty effects and long-term user behavior changes.

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

Q2

Design an A/B test to measure the causal impact of switching the account list sort order from most-recent-visit to most-notifications. Walk through unit of randomization, eligibility, metrics, sample size, and any pitfalls you'd watch for.

A/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

Felt more comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the core experiment design components: randomization unit, eligibility, metrics, sample size, and pitfalls. Emphasize causal inference and practical considerations specific to Meta's scale and product context.

Pro tip: Acknowledge that changing sort order can have network effects or cannibalization, and propose guardrail metrics to detect them. Also, mention that you'd run a pre-experiment power analysis and consider sequential testing if peeking.

1. Define the hypothesis and randomization unit

State the hypothesis that sorting by most-notifications increases engagement. Choose the randomization unit (e.g., user-level) and justify why it's appropriate for this change.

2. Determine eligibility and assignment

Specify which users are eligible (e.g., active users with multiple accounts) and how they are assigned to control (most-recent-visit) or treatment (most-notifications).

3. Select metrics and guardrails

Define primary metrics (e.g., notifications clicked, session duration) and guardrail metrics (e.g., user retention, complaints) to monitor unintended consequences.

4. Calculate sample size and duration

Estimate required sample size using power analysis, considering baseline metrics, minimum detectable effect, and traffic. Determine experiment duration to achieve sufficient power.

5. Identify pitfalls and mitigation

Discuss potential pitfalls such as novelty effects, network effects, and metric dilution, and propose ways to mitigate them (e.g., holdout groups, long-term monitoring).

Key Points to Mention

  • Randomization unit: user-level to avoid contamination, but consider account-level if accounts are independent.
  • Eligibility: users with multiple accounts and recent activity to ensure relevance.
  • Primary metric: notification engagement (e.g., click-through rate) and secondary metrics like session time.
  • Guardrail metrics: user retention, unsubscribe rates, and support tickets.
  • Sample size: power analysis with 80% power and 5% significance, accounting for multiple testing.
  • Pitfalls: novelty effect, network effects, and interference between accounts; use A/A tests and holdouts.

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