← Google Interview Insights

Google·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
Jul 2026Remote

Summary

Stats-heavy technical screen for a DS role at Google. The whole thing was basically applied inference from start to finish, no product sense, no SQL, just probability and estimation. Walked out feeling okay about the first half and less okay about the truncated normal stuff at the end.

Questions Asked (6)

Q1

You collected a sample and want to test whether the population mean differs from 0. What does a p-value of x% mean in this context?

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

Knew this cold but still second-guessed my wording mid-answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the null and alternative hypotheses clearly: H0: μ = 0 vs. Ha: μ ≠ 0. Then explain that a p-value of x% is the probability, assuming H0 is true, of observing a sample mean at least as extreme as the one obtained. Finally, interpret it in context: if x% is less than the significance level (e.g., 5%), you reject H0 and conclude the population mean differs from 0; otherwise, you fail to reject H0.

Pro tip: Emphasize that the p-value is not the probability that H0 is true or that the result is due to chance; it's a measure of evidence against H0 under the assumption that H0 is true. Also, mention that practical significance depends on effect size and context, not just the p-value.

1. State the hypotheses

Clearly define the null hypothesis (H0: μ = 0) and the alternative hypothesis (Ha: μ ≠ 0). This sets the foundation for interpreting the p-value.

2. Define the p-value

Explain that the p-value is the probability of observing a test statistic as extreme as, or more extreme than, the one calculated from the sample, assuming the null hypothesis is true.

3. Interpret the specific p-value

If the p-value is x%, it means that if the true population mean were 0, there is an x% chance of seeing a sample mean at least as far from 0 as the one observed.

4. Compare to significance level

Compare the p-value to a pre-determined significance level (α, often 0.05). If p-value < α, reject H0; otherwise, fail to reject H0.

5. Draw conclusion in context

State the practical conclusion: either there is sufficient evidence to suggest the population mean differs from 0, or there is not enough evidence to conclude that.

Key Points to Mention

  • Null hypothesis (H0: μ = 0) and alternative hypothesis (Ha: μ ≠ 0)
  • Definition of p-value: probability of observing data at least as extreme as the sample, given H0 is true
  • The p-value is not the probability that H0 is true or that the result is due to chance
  • Comparison with significance level (α) to make a decision
  • Effect size and practical significance vs. statistical significance
  • Assumptions of the test (e.g., normality, independence) and potential issues with multiple testing

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

Q2

Given a sample mean of 1 and a standard error of 0.1, construct a 95% confidence interval for the population mean.

A/B Testing & Experimentation
Author's notes

Straightforward.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by stating the formula for a confidence interval: sample mean ± critical value × standard error. For a 95% confidence level, use the z-critical value of 1.96 (assuming a large sample). Then plug in the given values to compute the interval and interpret it in context.

Pro tip: Mention that this assumes the sampling distribution is approximately normal, which typically holds for large samples due to the Central Limit Theorem. Also, briefly note that if the sample size were small, you would use a t-critical value instead.

1. Identify the components

Recognize that the sample mean is 1, the standard error is 0.1, and the confidence level is 95%.

2. Determine the critical value

For a 95% confidence interval, the critical z-value is approximately 1.96 for large samples.

3. Compute the margin of error

Multiply the critical value by the standard error: 1.96 × 0.1 = 0.196.

4. Construct the interval

Add and subtract the margin of error from the sample mean: 1 ± 0.196, yielding (0.804, 1.196).

5. Interpret the interval

Explain that we are 95% confident that the true population mean lies between 0.804 and 1.196.

Key Points to Mention

  • Formula for confidence interval: sample mean ± (critical value × standard error)
  • Use of z-critical value 1.96 for 95% confidence when sample size is large
  • Assumption of normality of the sampling distribution (Central Limit Theorem)
  • Margin of error calculation: 1.96 × 0.1 = 0.196
  • Interpretation: 95% confidence that the population mean is within the interval
  • If sample size is small, use t-distribution instead of z

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

Q3

What sample size would you need to reduce the standard error from 0.1 to 0.01, and what can you do if you can't increase sample size?

A/B Testing & ExperimentationTechnical Trade-offs
Author's notes

The first part is just algebra on se = s divided by root n, so you need 100x the sample size.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining the relationship between standard error and sample size: SE = σ/√n, so to reduce SE by a factor of 10, you need 100 times the sample size. Then discuss practical alternatives when increasing sample size isn't feasible, such as reducing variance, using sequential testing, or leveraging Bayesian methods.

Pro tip: Quantify the trade-offs: increasing sample size by 100x may be infeasible due to cost or time, so emphasize variance reduction techniques (e.g., stratification, CUPED) that can achieve similar precision gains without more data.

1. Derive the sample size multiplier

Use the formula SE = σ/√n to show that reducing SE from 0.1 to 0.01 requires a 100-fold increase in sample size (since SE scales inversely with the square root of n).

2. Discuss feasibility and constraints

Acknowledge that a 100x increase is often impractical due to cost, time, or limited traffic, and that it may not be necessary if the effect size is large enough.

3. Explore variance reduction techniques

Mention methods like stratification, blocking, using covariates (CUPED), or paired designs that reduce σ and thus SE without increasing n.

4. Consider alternative statistical approaches

Suggest sequential testing, Bayesian methods, or using a more sensitive metric to achieve reliable results with smaller samples.

5. Summarize trade-offs and recommendations

Conclude that the best approach depends on the context, and recommend a combination of variance reduction and appropriate statistical methods when sample size is fixed.

Key Points to Mention

  • Standard error formula: SE = σ/√n, so SE reduction by factor k requires n multiplied by k².
  • 100x sample size increase needed to reduce SE from 0.1 to 0.01.
  • Variance reduction techniques: stratification, CUPED, blocking, paired designs.
  • Sequential testing and Bayesian methods allow early stopping or smaller samples.
  • Trade-offs between statistical power, cost, and time.
  • Importance of effect size: smaller SE may not be necessary if effect is large.

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

Q4

Given independent observations from some distribution, propose an estimator for the probability that X exceeds 10, then construct a 95% confidence interval for that probability and interpret the resulting interval.

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

Used the plug-in estimator, just the proportion of observations above 10.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the estimator as the sample proportion of observations exceeding 10, then use the normal approximation to construct a Wald confidence interval. Alternatively, mention exact methods like Clopper-Pearson for small samples, and interpret the interval in terms of the true probability.

Pro tip: Always check the sample size and number of successes to ensure the normal approximation is valid; if not, use exact or Wilson interval. Also, relate the interval to business impact, e.g., 'We are 95% confident that the true probability lies between X and Y, which means...'

1. Define the estimator

Let p be the probability that X > 10. The natural estimator is the sample proportion: p_hat = (1/n) * sum(I(X_i > 10)).

2. Check assumptions

Verify that the sample size is large enough for the normal approximation: n*p_hat >= 10 and n*(1-p_hat) >= 10. If not, consider exact methods.

3. Construct the confidence interval

For large samples, use the Wald interval: p_hat ± z_{α/2} * sqrt(p_hat*(1-p_hat)/n). For small samples, use Wilson or Clopper-Pearson.

4. Interpret the interval

Interpret in context: 'We are 95% confident that the true probability that X exceeds 10 lies between L and U.' Avoid saying 'there is a 95% probability that p is in the interval.'

Key Points to Mention

  • Sample proportion as an unbiased estimator for a probability
  • Normal approximation and its validity conditions
  • Alternatives: Wilson score interval, Clopper-Pearson exact interval
  • Interpretation of confidence interval: frequentist vs Bayesian
  • Impact of sample size on interval width
  • Practical significance vs statistical significance

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

Q5

You have 1,000 binary features and want to estimate an overall conversion rate. How would you design the estimation or sampling strategy?

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

This one felt a bit underspecified and I spent maybe 30 seconds clarifying what 'overall' meant before diving in.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the goal: estimate the overall conversion rate across all features, not per-feature. Then propose a sampling strategy that balances statistical efficiency with practical constraints, such as stratified sampling by feature prevalence or using a model-based approach like logistic regression with regularization to handle high dimensionality.

Pro tip: Emphasize the bias-variance tradeoff and the importance of validating your sampling strategy with a hold-out set or simulation. Mention that in practice, you'd also consider computational cost and interpretability for stakeholders.

1. Clarify the objective and constraints

Confirm whether the goal is a single overall conversion rate or per-feature rates, and understand data size, computational resources, and business context.

2. Choose a sampling or modeling strategy

Decide between simple random sampling, stratified sampling by feature combinations, or a model-based approach (e.g., logistic regression with L1/L2 regularization) to estimate the overall rate efficiently.

3. Address high dimensionality and sparsity

With 1,000 binary features, many combinations may be sparse. Consider dimensionality reduction (e.g., PCA) or feature hashing, and use regularization to prevent overfitting.

4. Validate and estimate uncertainty

Use cross-validation or bootstrapping to assess the stability of your estimate and provide confidence intervals. Compare against a simple baseline (e.g., overall conversion rate without features).

5. Communicate trade-offs and recommendations

Summarize the chosen approach, its assumptions, and limitations. Discuss how the estimate would be used and whether a more granular analysis is needed.

Key Points to Mention

  • Stratified sampling to ensure representation of rare feature combinations
  • Regularized logistic regression (L1/L2) to handle high-dimensional binary features
  • Bias-variance tradeoff and the risk of overfitting with 1,000 features
  • Use of cross-validation or bootstrapping for uncertainty quantification
  • Computational efficiency and scalability considerations
  • Potential need for dimensionality reduction or feature selection

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

Q6

Assume X follows a normal distribution but you only observe values where X exceeds a threshold of 3 (truncated normal). How would you estimate the mean and variance, and how would you construct confidence intervals for them?

A/B Testing & ExperimentationTechnical Trade-offs
Author's notes

This is where I felt the floor drop out a little.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a truncated normal distribution and explain that naive estimates (sample mean and variance) are biased. Then outline a maximum likelihood estimation (MLE) approach using the truncated normal likelihood, and describe how to construct confidence intervals via the asymptotic normality of MLE or bootstrap. Emphasize the importance of accounting for truncation in both estimation and inference.

Pro tip: Mention that the truncation point (3) is known and fixed, so the likelihood is conditional on X > 3; this simplifies the MLE derivation and avoids identifiability issues. Also, note that if the truncation is severe, the information about the mean and variance is limited, leading to wider confidence intervals.

1. Define the truncated normal distribution

Clarify that the observed data come from a normal distribution conditional on X > 3, with density f(x; μ, σ²) / (1 - Φ((3-μ)/σ)) for x > 3. This is the likelihood contribution for each observation.

2. Derive the likelihood and MLE

Write the log-likelihood for the sample: sum log f(x_i; μ, σ²) - n log(1 - Φ((3-μ)/σ)). Maximize numerically (e.g., Newton-Raphson) to obtain MLEs for μ and σ².

3. Construct confidence intervals

Use the observed Fisher information matrix to get standard errors for the MLEs, then form Wald-type confidence intervals. Alternatively, use bootstrap or profile likelihood for more accurate intervals, especially with small samples.

4. Check assumptions and diagnostics

Verify normality of the underlying distribution (e.g., via QQ plot of the truncated data) and assess whether the truncation threshold is correctly specified. Consider sensitivity analysis if the threshold is uncertain.

5. Discuss practical implications

Explain how truncation affects estimation: the sample mean underestimates μ, and the sample variance underestimates σ². Highlight that ignoring truncation leads to biased estimates and invalid inference.

Key Points to Mention

  • Truncated normal distribution and its likelihood function.
  • Maximum likelihood estimation (MLE) for μ and σ² under truncation.
  • Asymptotic normality of MLE and use of Fisher information for confidence intervals.
  • Bootstrap or profile likelihood as alternatives for confidence intervals.
  • Bias of naive estimators (sample mean and variance) when truncation is ignored.
  • Practical challenges: numerical optimization, small sample properties, and sensitivity to threshold specification.

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