← Google Interview Insights

Google·Data Scientist·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

A data science interview at Google that threw a curveball: a quasi-experimental design question built around Google Meet, where you couldn't just run an A/B test. Felt more like a stats seminar than a typical DS screen, which I wasn't fully expecting.

Questions Asked (1)

Q1

A new version of Google Meet was released with the goal of reducing call drops, but no A/B test was run. What analytical approaches would you use to evaluate whether the new version actually worked?

A/B Testing & ExperimentationProduct Analytics & MetricsAdaptability & Ambiguity
Author's notes

This one exposed some gaps for me.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the lack of a controlled experiment and the challenges of causal inference. Then outline a multi-pronged approach: define success metrics, use quasi-experimental methods like difference-in-differences or synthetic control, and validate with qualitative and quantitative evidence. Emphasize the importance of understanding the rollout pattern and potential confounders.

Pro tip: Propose a 'pre-post with comparison group' design using regions or user segments where the new version was not yet rolled out, and discuss how to test for parallel trends. Also, mention the value of instrumenting the product to capture call drop reasons for future analysis.

1. Define success metrics and hypotheses

Clarify what 'reducing call drops' means: define call drop rate, maybe segmented by call type, device, network, etc. State the hypothesis that the new version reduces call drops compared to the old version.

2. Identify comparison groups and time windows

Since no A/B test, find natural comparison groups: users/regions not yet upgraded, or historical data from before the release. Determine pre- and post-release periods, ensuring sufficient data.

3. Apply quasi-experimental methods

Use difference-in-differences, synthetic control, or interrupted time series to estimate the causal effect. Check assumptions like parallel trends and robustness to confounders.

4. Triangulate with additional evidence

Look at qualitative feedback, support tickets, and internal logs to understand why drops occur. Check if other metrics (e.g., call duration, user satisfaction) moved consistently.

5. Communicate findings and limitations

Present the estimated effect with confidence intervals, discuss limitations of observational methods, and recommend next steps (e.g., run a proper A/B test for future changes).

Key Points to Mention

  • Difference-in-differences (DiD) and synthetic control methods for causal inference without randomization
  • Importance of defining clear metrics (e.g., call drop rate per 100 calls) and segmenting by relevant dimensions
  • Checking for parallel trends assumption in DiD and using placebo tests
  • Potential confounders: seasonality, network upgrades, user behavior changes, concurrent product changes
  • Triangulation with qualitative data (user feedback, support tickets) and other quantitative metrics
  • Limitations of observational studies and the value of running a proper A/B test in the future

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