← Google Interview Insights

Google·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Apr 2026

Summary

Google data scientist interview with a statistics-heavy question about causal inference and regression. The question was framed as a practical scenario but quickly went deeper than I expected.

Questions Asked (1)

Q1

You have data from 1,000 residents of a single city and want to study whether attending college affects income. Is a simple linear regression with a binary college variable appropriate here? Walk through the sampling concerns, whether the model is valid, and what better approaches might look like.

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

I jumped straight into talking about regression assumptions and kind of glossed over the sampling issue, which was probably the more interesting part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the appeal of a simple linear regression but immediately highlight the fundamental issue: college attendance is not randomly assigned, so the binary variable is likely endogenous. Then, systematically discuss the sampling concerns (e.g., single city, selection bias), the validity of the linear model (functional form, omitted variable bias), and finally propose better approaches such as instrumental variables, propensity score matching, or regression discontinuity.

Pro tip: Demonstrate awareness that even with a large sample, causal inference requires addressing confounding; mention that Google often uses quasi-experimental designs like difference-in-differences or instrumental variables when randomization isn't possible.

1. Identify the goal: causal effect vs. prediction

Clarify that the question asks about the effect of attending college on income, which is a causal inference problem, not just prediction. A simple linear regression with a binary variable estimates a difference in means, but only if the variable is exogenous.

2. Discuss sampling concerns

The sample is from a single city, limiting generalizability. More importantly, college attendance is self-selected: individuals who attend college may differ systematically (e.g., motivation, ability, family background) from those who do not, leading to selection bias.

3. Evaluate model validity

A linear regression assumes a linear relationship, no omitted variable bias, and homoscedasticity. With a binary predictor, the model estimates the mean income difference, but if confounding exists, the coefficient is biased. Also, income may not be linear in predictors; consider transformations or non-linear models.

4. Propose better approaches

Suggest methods that address confounding: instrumental variables (e.g., distance to college, policy changes), propensity score matching, regression discontinuity (if a cutoff exists), or panel data with fixed effects. If the goal is prediction, machine learning models could be used, but causal claims require careful design.

5. Conclude with practical recommendations

Summarize that while simple linear regression is a starting point, it is likely insufficient for causal inference. Recommend a quasi-experimental design or at least sensitivity analysis to assess robustness of results.

Key Points to Mention

  • Endogeneity and omitted variable bias (e.g., ability, family income, motivation)
  • Selection bias due to non-random assignment of college attendance
  • Limitations of single-city sample: external validity and potential homogeneity
  • Functional form assumptions: linearity, constant marginal effects, and binary predictor interpretation
  • Alternative causal inference methods: instrumental variables, propensity score matching, regression discontinuity, difference-in-differences
  • Importance of defining the estimand (ATE, ATT) and checking for heterogeneous treatment effects

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