← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
Aug 2025Remote

Summary

Meta DS interview that went deep into statistical testing for a user engagement scenario. Four questions, all quantitative, no fluff. The kind of round where you either know your NB distribution parameterization cold or you don't.

Questions Asked (4)

Q1

You're comparing daily session counts between new and existing users. The data is integer-valued, skewed, and has lots of zeros. Which test do you use to compare central tendency across the two cohorts and why: two-sample t-test, Welch's t-test, Mann-Whitney U, or a GLM-based approach? Walk through your assumptions and what diagnostics you'd run.

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

This is the kind of question where you can talk yourself into a corner fast.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that the choice depends on the estimand and data properties: with skewed, zero-inflated integer counts, the t-test's normality assumption is violated, so either Mann-Whitney U (for stochastic dominance) or a GLM (e.g., negative binomial or zero-inflated model) is more appropriate. Recommend a GLM-based approach if you need to model the mean and adjust for covariates, but mention Mann-Whitney as a robust non-parametric alternative for simple comparisons.

Pro tip: At Meta, interviewers value practical impact: emphasize that you'd first check if the difference is practically significant (e.g., effect size) and consider bootstrapping for confidence intervals, as it handles zero-inflation and skew well without strong assumptions.

1. Clarify the estimand and goal

Determine whether you're comparing means, medians, or overall distributions, and whether you need to adjust for covariates. This drives the choice of test.

2. Evaluate assumptions of each test

For t-tests: normality and homoscedasticity (Welch's relaxes equal variance). For Mann-Whitney: independent samples and similar distribution shapes (except shift). For GLM: correct distributional family and link function.

3. Diagnose the data

Run descriptive statistics, histograms, QQ plots, and tests for zero-inflation (e.g., proportion of zeros). Check for overdispersion if considering Poisson GLM.

4. Select and justify the test

If the goal is a simple comparison of central tendency and assumptions fail, use Mann-Whitney U. If modeling the mean with covariates or needing interpretable coefficients, use a GLM (e.g., negative binomial or zero-inflated).

5. Validate and interpret

Run the chosen test, check diagnostics (e.g., residual plots for GLM), and compute effect sizes or confidence intervals (e.g., via bootstrap) to assess practical significance.

Key Points to Mention

  • Skewness and zero-inflation violate normality and homoscedasticity assumptions of t-tests.
  • Mann-Whitney U tests stochastic dominance, not specifically means or medians, and assumes similar distribution shapes under the null.
  • GLM (e.g., negative binomial, zero-inflated Poisson) can model the mean directly and handle overdispersion and excess zeros.
  • Welch's t-test only addresses unequal variances, not non-normality or zero-inflation.
  • Bootstrapping can provide robust confidence intervals for differences in means or medians without strong parametric assumptions.
  • Consider the business context: if the goal is to detect any difference in engagement, a non-parametric or GLM approach may be more sensitive.

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

Q2

Daily sessions per user follows a Negative Binomial distribution with mean 2.40 and variance 6.96 for existing users. Using the parameterization where E[X] = r(1-p)/p and Var[X] = r(1-p)/p squared, solve for r and p, then compute P(X = 0).

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

Blanked for a second on the algebra.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, set up the equations for the mean and variance using the given parameterization, then solve for r and p by dividing the variance by the mean to isolate a function of p. After finding p, substitute back to find r, and finally compute P(X=0) using the negative binomial probability mass function.

Pro tip: When solving, remember that the variance-to-mean ratio (6.96/2.40 = 2.9) is greater than 1, indicating overdispersion, which is typical for count data like daily sessions. Also, double-check your algebra because small errors in p can significantly affect P(X=0).

1. Set up equations

Write the given mean and variance formulas: E[X] = r(1-p)/p = 2.40 and Var[X] = r(1-p)/p^2 = 6.96.

2. Solve for p

Divide the variance equation by the mean equation to eliminate r, yielding (1-p)/p = 6.96/2.40 = 2.9. Solve for p: 1-p = 2.9p => 1 = 3.9p => p = 1/3.9 ≈ 0.2564.

3. Solve for r

Substitute p back into the mean equation: r(1-p)/p = 2.40. With p ≈ 0.2564, 1-p ≈ 0.7436, so r ≈ 2.40 * 0.2564 / 0.7436 ≈ 0.827.

4. Compute P(X=0)

Use the negative binomial PMF: P(X=0) = p^r (since (1-p)^0 = 1). Calculate p^r ≈ (0.2564)^0.827 ≈ exp(0.827 * ln(0.2564)) ≈ exp(0.827 * (-1.361)) ≈ exp(-1.125) ≈ 0.325.

Key Points to Mention

  • Negative binomial distribution is often used for overdispersed count data, where variance exceeds the mean.
  • The parameterization given is common in some statistical software (e.g., R's glm.nb) where r is the dispersion parameter and p is the probability of success.
  • Solving for parameters involves using the method of moments: equating sample mean and variance to theoretical expressions.
  • P(X=0) represents the probability of zero sessions, which is a key metric for user engagement.
  • Always verify calculations by checking that the computed mean and variance match the given values.
  • In practice, you might use software to fit the distribution, but understanding the algebra is crucial for interviews.

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

Q3

For new users you have mean 1.85 and variance 4.20, with 5,000 users per cohort. Construct a 95% confidence interval for the difference in mean sessions between cohorts. You can use delta-method or GLM reasoning, just state your approximations.

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

Straightforward once you remember the SE for a difference in means is just the square root of the sum of the two sample variances divided by their respective n's.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Treat the difference in means as approximately normal with variance equal to the sum of the two cohort variances divided by their sample sizes. Compute the standard error, then construct the 95% CI as the difference in means ± 1.96 * SE. State any assumptions, such as independence and large-sample normality.

Pro tip: Mention that with 5,000 users per cohort, the normal approximation is well-justified, and if the metric is a count (sessions), a Poisson or negative binomial GLM might be more appropriate—but the delta method gives a quick, reasonable interval. Also, clarify whether the variance is for individual users or the mean; if it's the variance of individual sessions, the standard error uses that variance divided by n.

1. Identify given parameters

Note the mean (1.85) and variance (4.20) for new users, and the sample size per cohort (5,000). Clarify that these are for one cohort; the other cohort's mean and variance are not provided, so assume they are the same or that the difference is between two independent cohorts with these parameters.

2. Compute standard error of the difference

Under independence, the variance of the difference in means is the sum of the variances of each mean: Var(mean) = variance / n. So SE = sqrt(4.20/5000 + 4.20/5000) = sqrt(2 * 4.20/5000) ≈ sqrt(0.00168) ≈ 0.041.

3. Determine the critical value

For a 95% confidence interval, use the z-score 1.96 (or 2 for approximation) since the sample size is large.

4. Construct the interval

The difference in means is 0 (if comparing identical cohorts) or some value if specified. The 95% CI is difference ± 1.96 * SE. If difference is 0, the interval is approximately (-0.080, 0.080).

5. State assumptions and approximations

Mention that this uses the delta method / normal approximation, assumes independence between cohorts, and that the variance is known or well-estimated. Also note that if the metric is a count, a GLM approach might yield a slightly different interval.

Key Points to Mention

  • Independence between cohorts and large sample size justify normal approximation.
  • Standard error of difference: sqrt(variance1/n1 + variance2/n2).
  • Use z=1.96 for 95% confidence level.
  • If variances are unequal, use Welch's approximation (but here they are equal).
  • Delta method is equivalent to normal approximation for means.
  • Consider Poisson GLM if sessions are counts, but the normal approximation is robust with n=5000.

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

Q4

You run a Welch's t-test and get p = 0.04 with Cohen's d = 0.08. How do you interpret statistical versus practical significance? What do you do about multiple testing if you also segmented by 5 countries? And name one robust effect size metric suited to count data, like a ratio of means, and how you'd get a CI for it.

A/B Testing & ExperimentationProduct Analytics & MetricsRoot Cause Analysis
Author's notes

The p-value vs effect size piece is almost a gimme but d = 0.08 is genuinely tiny so the practical significance answer is pretty clear: statistically detectable, operationally irrelevant unless you're at massive scale.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, explain that statistical significance (p < 0.05) indicates a likely non-zero effect, but practical significance depends on the effect size and business context. With Cohen's d = 0.08, the effect is tiny and likely not practically meaningful. Then, address multiple testing by applying a correction like Bonferroni or Benjamini-Hochberg, and finally, for count data, suggest using the ratio of means (or its log) and obtaining a confidence interval via bootstrap or the delta method.

Pro tip: At Meta, always tie statistical findings to business impact—e.g., even a tiny effect can matter at scale if the metric is revenue, but often it's noise. Also, mention that multiple testing corrections should be pre-registered to avoid p-hacking.

1. Interpret statistical vs practical significance

Explain that p = 0.04 means the result is statistically significant at α = 0.05, but Cohen's d = 0.08 is a very small effect (negligible). Practical significance requires considering the metric's sensitivity and business impact.

2. Assess practical significance in context

Discuss that a small effect might be practically significant if the metric is critical (e.g., revenue per user) and the sample size is large, but often it's not worth acting on. Recommend looking at confidence intervals and raw differences.

3. Address multiple testing

With 5 country segments, you're conducting multiple comparisons, inflating Type I error. Apply a correction like Bonferroni (α/5 = 0.01) or Benjamini-Hochberg to control false discovery rate, and note that p = 0.04 would not survive Bonferroni.

4. Choose robust effect size for count data

For count data (e.g., clicks, purchases), use the ratio of means (or its log) as a robust effect size. It's interpretable as a multiplicative change and less sensitive to outliers than Cohen's d.

5. Obtain confidence interval for ratio of means

Use bootstrap resampling (non-parametric) or the delta method (parametric) to compute a CI for the ratio. Bootstrap is preferred for robustness with count data.

Key Points to Mention

  • Statistical significance does not imply practical significance; effect size and context matter.
  • Cohen's d = 0.08 is considered negligible (small effect sizes: 0.2 small, 0.5 medium, 0.8 large).
  • Multiple testing correction methods: Bonferroni, Benjamini-Hochberg, and the trade-off between Type I and Type II errors.
  • Ratio of means (or log ratio) is a robust effect size for count data, often used in A/B testing at tech companies.
  • Confidence intervals for ratio of means can be obtained via bootstrap or delta method.
  • At scale (e.g., Meta), even tiny effects can be practically significant if the metric is revenue-critical, but must be weighed against implementation costs.

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