← Google Interview Insights

Google·Data Scientist·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Google data scientist interview that leaned heavily on stats fundamentals and experimental design. The whole session felt like a PM-facing consulting exercise where you had to explain gnarly concepts without hiding behind jargon. Tough but fair.

Questions Asked (4)

Q1

Explain Simpson's paradox and walk through a concrete numeric example that a non-technical product manager could actually follow.

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

This is one of those questions where you think you know it until you have to make up numbers on the spot.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start with a plain-language definition of Simpson's paradox, then immediately ground it in a simple, relatable numeric example that a non-technical PM can follow. Use the example to highlight the reversal of conclusions when data is aggregated versus segmented, and connect it to product analytics or A/B testing contexts.

Pro tip: Use a real-world product scenario (e.g., overall conversion rate vs. by-device breakdown) to make the paradox tangible and show you can translate statistical concepts into business impact.

1. Define the paradox simply

Explain that Simpson's paradox occurs when a trend appears in aggregated data but reverses when the data is split into groups. Emphasize that it's a cautionary tale about aggregation.

2. Set up a concrete example

Choose a familiar product metric, like conversion rate, and create two versions (A and B) with two user segments (e.g., mobile and desktop). Ensure the numbers are easy to compute mentally.

3. Show the aggregated result

Calculate the overall conversion rate for each version by combining segments. Highlight that one version appears better overall.

4. Break down by segment

Show that within each segment, the other version actually performs better. Explain that the reversal is due to unequal segment sizes or confounding variables.

5. Connect to product decisions

Discuss implications for A/B testing and metric interpretation: always check for segment-level effects and consider potential confounders before drawing conclusions.

Key Points to Mention

  • Definition of Simpson's paradox: aggregated trend reverses when data is segmented.
  • Concrete numeric example with two groups and two treatments, showing reversal.
  • Explanation of why it happens: confounding variables or unequal group sizes.
  • Relevance to A/B testing: overall metrics can mislead; segment analysis is crucial.
  • Product management takeaway: avoid overgeneralizing from aggregate data; dig into segments.
  • Communication tip: use simple numbers and avoid jargon when explaining to non-technical stakeholders.

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

Q2

You're given a forest with three bird species. Propose and define metrics that quantify how segregated the species are from each other spatially.

Product Analytics & MetricsData Modeling
Author's notes

Did not see this coming.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the data structure (point locations with species labels) and the spatial scale of interest. Then propose a small set of complementary metrics that capture different aspects of segregation: global clustering, pairwise co-occurrence, and local mixing. Finally, discuss how to validate and interpret these metrics in the context of the business or ecological question.

Pro tip: Always tie the metrics back to actionable insights—e.g., if segregation is high, what intervention would that suggest? This shows you think beyond computation to decision-making.

1. Clarify data and assumptions

Confirm that you have point locations for each bird sighting, labeled by species, and discuss any assumptions about sampling effort or spatial boundaries.

2. Define global segregation metrics

Propose metrics like the Index of Dissimilarity or Moran's I to measure overall spatial clustering and segregation across all species.

3. Define pairwise segregation metrics

Introduce pairwise measures such as the Join Count Statistic or the cross-K function to quantify segregation between each pair of species.

4. Define local segregation metrics

Suggest local indicators like Local Moran's I or the Getis-Ord Gi* to identify hotspots where species are more or less segregated at fine scales.

5. Validate and interpret

Discuss how to test significance (e.g., permutation tests) and interpret results in terms of ecological or business implications.

Key Points to Mention

  • Index of Dissimilarity (Duncan index) for global segregation
  • Moran's I and Geary's C for spatial autocorrelation
  • Join count statistics for categorical spatial data
  • Ripley's K function and its cross-species variant
  • Local indicators of spatial association (LISA)
  • Permutation tests for significance testing

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

Q3

A simple linear regression is used to estimate how YouTube ad impressions affect product sales. What statistical problems do you see, and how would you redesign the study?

A/B Testing & ExperimentationTechnical Trade-offsRoot Cause Analysis
Author's notes

This was the meatiest one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by identifying the statistical problems with using simple linear regression for this causal question, such as confounding, endogeneity, and omitted variable bias. Then propose a redesigned study using randomized controlled experiments or quasi-experimental methods to establish causality, and discuss how to measure ad effectiveness properly.

Pro tip: Emphasize that observational data can show correlation but not causation, and that Google's gold standard is randomized experiments. Mention that even with experiments, you need to consider spillover effects and long-term impact.

1. Identify problems with simple linear regression

Discuss issues like confounding variables (e.g., seasonality, promotions), reverse causality, measurement error, and violation of linearity assumptions. Highlight that regression cannot establish causality without exogeneity.

2. Explain why causal inference is needed

Clarify that the goal is to estimate the causal effect of ad impressions on sales, which requires counterfactual reasoning. Simple regression on observational data is biased due to selection effects and omitted variables.

3. Propose a redesigned study using experiments

Suggest a randomized controlled trial (A/B test) where users are randomly assigned to see ads or not. Discuss how to measure incremental sales and control for confounders through randomization.

4. Address practical challenges and alternatives

Mention challenges like ad blocking, spillover, and cost. If randomization is not feasible, propose quasi-experimental methods like difference-in-differences, instrumental variables, or propensity score matching.

5. Define metrics and analysis plan

Specify metrics like incremental sales per impression, ROI, and statistical power. Outline analysis using regression with fixed effects or causal models, and discuss how to handle heterogeneous effects.

Key Points to Mention

  • Confounding variables and omitted variable bias
  • Reverse causality and simultaneity
  • Randomized controlled experiments (A/B testing) as gold standard
  • Quasi-experimental methods (difference-in-differences, instrumental variables)
  • Spillover effects and external validity
  • Incremental sales and ROI measurement

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

Q4

What are the ways to reduce a survey's margin of error? And if you can't change sample size or confidence level, what other levers exist and why might bootstrapping help?

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

Answered the first part fine: reduce population variance, use stratified sampling, tighten the questionnaire.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining margin of error (MOE) and its formula, then discuss standard levers like increasing sample size and lowering confidence level. When those are fixed, pivot to design and estimation levers such as stratified sampling, weighting, and variance reduction techniques, and explain how bootstrapping can provide more accurate uncertainty estimates when parametric assumptions fail.

Pro tip: Emphasize that MOE is a function of both sampling and non-sampling error; in practice, reducing bias through better sampling design often matters more than shrinking variance. Also, mention that bootstrapping is not a magic fix for small samples but can help when the sampling distribution is unknown or skewed.

1. Define MOE and its determinants

State that MOE = z * sqrt(p(1-p)/n) for proportions, and depends on confidence level (z), sample size (n), and population variability (p).

2. List standard levers

Increase sample size, decrease confidence level, or reduce population variability (e.g., through stratified sampling).

3. Explore alternative levers when n and confidence are fixed

Use stratified sampling, post-stratification weighting, ratio estimation, or regression adjustment to reduce variance without changing n.

4. Explain bootstrapping's role

Bootstrapping estimates the sampling distribution empirically, providing more accurate confidence intervals when parametric assumptions are violated or when the statistic is complex.

5. Connect to practical impact

Tie back to product analytics: better MOE leads to more reliable A/B test decisions and metric estimates.

Key Points to Mention

  • Margin of error formula and its components (z, n, p).
  • Standard levers: sample size, confidence level, population variability.
  • Alternative design levers: stratified sampling, weighting, ratio estimation.
  • Variance reduction techniques: regression adjustment, control variates.
  • Bootstrapping: resampling method to estimate sampling distribution and confidence intervals.
  • When bootstrapping helps: non-normal data, complex statistics, small samples (with caution).

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