Google·Data Scientist·Technical Phone Screen·Intermediate

IntermediatePrefer not to say

Jul 2026

Summary

Google stats/ML interview, just one question but it was a conceptual one that required more nuance than I expected going in.

Questions Asked (1)

Why might a biased estimator sometimes be preferable to an unbiased one?

Technical Trade-offsProduct Analytics & Metrics

Author's notes

I knew the textbook answer about variance and the bias-variance tradeoff, but I fumbled explaining when you'd actually make that call in practice.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining bias and variance, then explain that the goal is often minimizing total error (MSE) rather than just bias. Use concrete examples like ridge regression or James-Stein estimator to illustrate when bias reduces variance and improves predictive performance.

Pro tip: Mention that in practice, especially at Google, we care about business metrics and generalization, so a slightly biased but lower-variance model often leads to better decisions and more stable A/B test results.

1. Define bias and variance

Briefly explain that bias is the difference between expected estimate and true value, while variance is the variability of the estimate. Unbiasedness only addresses bias, not variance.

2. Introduce the bias-variance trade-off

Explain that total error (e.g., MSE) = bias^2 + variance + noise. Sometimes increasing bias can significantly reduce variance, lowering overall error.

3. Provide concrete examples

Give examples like ridge regression (L2 regularization) which introduces bias but reduces variance, or James-Stein estimator which dominates the unbiased sample mean in multiple dimensions.

4. Connect to practical scenarios

Discuss situations where prediction accuracy or decision-making matters more than unbiased parameter estimation, such as high-dimensional data, small sample sizes, or when multicollinearity is present.

5. Summarize trade-offs

Conclude that the choice depends on the goal: if inference and unbiasedness are critical, use unbiased; if prediction and stability are key, a biased estimator may be preferable.

Key Points to Mention

Bias-variance trade-off and mean squared error decomposition
Regularization techniques (ridge, lasso) that introduce bias to reduce variance
James-Stein estimator as a classic example where biased estimator dominates unbiased
High-dimensional settings where unbiased estimators have high variance
Small sample sizes where variance reduction is crucial
Practical implications for A/B testing and business metrics at Google

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