← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
Jun 2026Remote

Summary

A deep-dive stats question at Meta for a Data Scientist role. The whole interview was basically one massive multi-part problem about modeling skewed count data, and it went places I did not fully expect.

Questions Asked (5)

Q1

Daily comment counts per post on a large social app are highly skewed with many zeros. Which discrete distribution would you choose among Poisson, Negative Binomial, Poisson-lognormal, and discrete power-law, and why? Justify based on overdispersion, tail behavior, and zero inflation.

Data ModelingTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This is where I spent most of my time and honestly fumbled the ordering of my reasoning.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the characteristics of the data: many zeros, overdispersion, and heavy tail. Then evaluate each distribution against these features, ultimately recommending the Negative Binomial as a strong baseline, but noting that Poisson-lognormal or discrete power-law may be better if the tail is extremely heavy or zero inflation is severe. Justify your choice by discussing how each distribution handles overdispersion, tail behavior, and zero inflation.

Pro tip: Mention that in practice, you would validate the choice with diagnostic plots (e.g., rootogram, tail index estimation) and consider zero-inflated or hurdle models if zeros are excessive. This shows you balance theory with empirical checks.

1. Characterize the data

Identify key features: high frequency of zeros, variance greater than mean (overdispersion), and heavy right tail. These properties violate the assumptions of a simple Poisson model.

2. Evaluate Poisson

Poisson assumes mean equals variance and cannot handle overdispersion or excess zeros. It is inadequate for this data.

3. Evaluate Negative Binomial

Negative Binomial allows variance to exceed the mean via a dispersion parameter, making it suitable for overdispersed counts. It can capture some zero inflation if the dispersion is large, but may still underfit if zeros are excessive.

4. Evaluate Poisson-lognormal

Poisson-lognormal introduces a lognormal random effect, leading to overdispersion and a heavier tail than Negative Binomial. It is more flexible for heavy tails but may be computationally intensive and less interpretable.

5. Evaluate discrete power-law

Discrete power-law (e.g., Zipf) is designed for extremely heavy tails, often seen in social media engagement. However, it may not naturally accommodate zero inflation and can be challenging to fit.

6. Make a recommendation

Recommend Negative Binomial as a practical starting point due to its balance of flexibility and interpretability. If diagnostics show extreme tail or zero inflation, consider Poisson-lognormal or a zero-inflated/hurdle extension. Discrete power-law is an alternative if the tail is the primary concern.

Key Points to Mention

  • Overdispersion: Poisson fails because variance > mean; Negative Binomial and Poisson-lognormal handle it.
  • Tail behavior: Poisson-lognormal and discrete power-law have heavier tails than Negative Binomial; power-law is best for extreme heavy tails.
  • Zero inflation: All standard distributions may struggle with excess zeros; zero-inflated or hurdle models are often needed.
  • Negative Binomial as a baseline: It is widely used, interpretable, and can be extended to zero-inflated versions.
  • Model validation: Use diagnostic plots (e.g., rootogram) and compare AIC/BIC to select the best model.
  • Business context: For social media comment counts, heavy tails and zeros are common; the choice should align with the goal (e.g., prediction vs. inference).

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

Q2

Walk through a step-by-step model selection plan for this count data problem, covering dispersion tests, MLE fitting, likelihood-ratio tests, AIC/BIC comparisons, Vuong's test, power-law tail estimation via KS minimization, and visual diagnostics like QQ-plots.

Data ModelingTechnical Trade-offsA/B Testing & Experimentation
Author's notes

I knew most of the pieces but couldn't sequence them cleanly under pressure.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the goal: to select the best model for count data by balancing statistical rigor with practical interpretability. Walk through a logical sequence: first test for overdispersion to choose between Poisson and negative binomial, then fit models via MLE and compare using likelihood-ratio tests, AIC/BIC, and Vuong's test for non-nested models. Finally, assess tail behavior with power-law estimation and validate with visual diagnostics like QQ-plots.

Pro tip: Emphasize that model selection is iterative and context-dependent—always tie statistical results back to the business problem and interpretability. Mention that at Meta, where data is massive, computational efficiency and scalability of MLE and Vuong's test matter, so consider approximations or subsampling when necessary.

1. Test for Overdispersion and Choose Initial Model Family

Compute the dispersion statistic (e.g., Pearson chi-square/df) to check if variance exceeds mean. If overdispersion is present, consider negative binomial or quasi-Poisson; otherwise, start with Poisson.

2. Fit Candidate Models via MLE and Compare with Likelihood-Ratio Tests

Fit Poisson, negative binomial, and possibly zero-inflated or hurdle models using maximum likelihood estimation. For nested models, use likelihood-ratio tests to assess significance of additional parameters.

3. Compare Non-Nested Models with AIC/BIC and Vuong's Test

Use AIC/BIC for overall fit penalized by complexity. For non-nested models (e.g., negative binomial vs. power-law), apply Vuong's test to determine if one model is significantly closer to the true distribution.

4. Estimate Power-Law Tail via KS Minimization and Validate with Visual Diagnostics

If heavy tails are suspected, fit a power-law distribution to the tail using KS minimization to estimate the scaling parameter. Generate QQ-plots and other diagnostic plots to visually assess goodness-of-fit for all models.

Key Points to Mention

  • Overdispersion tests: Pearson chi-square, deviance, or score test for Poisson vs. negative binomial.
  • MLE fitting: Use packages like statsmodels or R's MASS::glm.nb; discuss convergence and standard errors.
  • Likelihood-ratio test for nested models (e.g., Poisson vs. negative binomial) and its assumptions.
  • AIC/BIC: Balance fit and complexity; BIC penalizes more for large samples.
  • Vuong's test: For non-nested model comparison, especially when comparing count models to power-law.
  • Power-law tail estimation: Use Clauset et al. method with KS minimization; caution about small sample bias.
  • Visual diagnostics: QQ-plots, residual plots, and cumulative distribution function comparisons.

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

Q3

Using your chosen model, how would you estimate P(Y <= 1) and the 95th percentile of comments per post, including bootstrap 95% confidence intervals? How do you handle standard errors and small-sample bias?

Data ModelingProduct Analytics & Metrics
Author's notes

Felt more comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the chosen model (e.g., Poisson, negative binomial, or zero-inflated) and its assumptions. Then, outline a simulation-based approach: fit the model, generate predicted distributions for each post, compute the desired statistics (P(Y<=1) and 95th percentile), and use bootstrap resampling to obtain confidence intervals. Finally, discuss methods to handle standard errors and small-sample bias, such as bias-corrected bootstrap or Bayesian methods.

Pro tip: Emphasize that for skewed count data, the 95th percentile is often more meaningful than the mean, and bootstrap CIs are robust to non-normality. Also, mention that when sample size is small, using a bias-corrected accelerated (BCa) bootstrap or a Bayesian posterior can provide better coverage than naive bootstrap.

1. Clarify the model and assumptions

State the chosen model (e.g., negative binomial) and justify why it fits comment counts (overdispersion, zero-inflation). Mention any covariates if relevant.

2. Estimate the statistics from the fitted model

Use the fitted model to compute P(Y<=1) and the 95th percentile. For P(Y<=1), sum probabilities from the model; for the percentile, invert the CDF or use simulation.

3. Bootstrap for confidence intervals

Resample posts with replacement, refit the model, and recompute the statistics. Repeat many times (e.g., 1000) to get the bootstrap distribution and derive 95% CIs (e.g., percentile method).

4. Address standard errors and small-sample bias

Discuss that bootstrap SEs are empirical standard deviations of the bootstrap replicates. For small samples, use bias-corrected methods (BCa) or Bayesian priors to reduce bias and improve coverage.

5. Validate and interpret

Check model fit (e.g., residual plots, posterior predictive checks) and interpret the results in business context, noting any limitations.

Key Points to Mention

  • Choice of model (e.g., negative binomial vs. Poisson) and handling overdispersion/zero-inflation.
  • Bootstrap procedure: resampling units (posts), refitting model, and computing statistics.
  • Types of bootstrap CIs: percentile, BCa, and when to use each.
  • Standard errors: bootstrap SE vs. model-based SE, and their interpretation.
  • Small-sample bias: bias-corrected bootstrap, Bayesian methods, or penalized likelihood.
  • Computational considerations: number of bootstrap replicates, parallelization, and convergence issues.

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

Q4

How would left-truncation of zero-comment posts, right-censoring from late-arriving comments, and mixing of bot versus human behavior bias your model estimates, and what corrections would you apply?

Data ModelingRoot Cause AnalysisTechnical Trade-offs
Author's notes

The bot/human mixture segment part caught me mid-sentence.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the data-generating process and identifying each bias mechanism (left-truncation, right-censoring, bot/human mixture). Then explain how each distorts model estimates (e.g., selection bias, survival bias, measurement error) and propose concrete corrections such as Heckman selection, inverse probability weighting, survival models, and bot filtering or mixture modeling. Finally, discuss trade-offs and validation.

Pro tip: Quantify the potential bias by comparing model estimates on a small, manually labeled unbiased sample versus the full biased dataset. This demonstrates rigor and helps prioritize which correction matters most.

1. Define the data-generating process and bias mechanisms

Clarify how left-truncation (zero-comment posts excluded), right-censoring (late comments not observed), and bot/human mixing arise in the data collection. Explain the causal structure and which estimand is affected.

2. Analyze the impact on model estimates

For each bias, describe the direction and magnitude of distortion: left-truncation causes selection bias and underestimates low-engagement behavior; right-censoring biases survival or count estimates downward; bot mixing inflates engagement metrics and adds noise.

3. Propose corrections for truncation and censoring

Apply Heckman selection models or inverse probability weighting for left-truncation; use survival analysis (e.g., Cox model) or joint modeling for right-censoring; consider multiple imputation for missing late comments.

4. Address bot versus human behavior

Filter bots using heuristics (e.g., activity patterns, metadata) or model the mixture with a latent class or zero-inflated model. Adjust estimates by weighting or including a bot indicator.

5. Validate and discuss trade-offs

Validate corrections on a small unbiased sample or via simulation. Discuss assumptions (e.g., missing at random), computational cost, and potential overcorrection. Recommend sensitivity analysis.

Key Points to Mention

  • Selection bias from left-truncation and Heckman correction
  • Survival analysis or censored regression for right-censoring
  • Bot detection and mixture modeling (latent class, zero-inflated)
  • Inverse probability weighting (IPW) for non-random missingness
  • Assumptions (MAR, exclusion restriction) and sensitivity analysis
  • Trade-offs between bias correction and variance/complexity

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

Q5

If threaded replies launch next month, how would you predict the qualitative change in your model parameters, and how would you revalidate model fit one week after launch?

Product Analytics & MetricsA/B Testing & ExperimentationAdaptability & Ambiguity
Author's notes

Nice way to end.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product change and its expected impact on user behavior, then reason about how that shifts model parameters qualitatively (e.g., feature importance, interaction effects). Finally, outline a revalidation plan that includes monitoring key metrics, checking calibration, and running a statistical test to detect drift one week post-launch.

Pro tip: Emphasize that you would set up a pre-launch validation framework with guardrail metrics and a holdout group to isolate the launch effect, showing you think proactively about causal inference and not just post-hoc analysis.

1. Clarify the change and expected behavioral impact

Ask clarifying questions about how threaded replies work (e.g., visibility, engagement patterns) and hypothesize how user interactions might change (e.g., more back-and-forth, deeper conversations).

2. Predict qualitative parameter shifts

Reason about which model parameters might change: e.g., features related to reply depth or conversation length could gain importance, while simple like/click features might lose weight; interaction terms may become more relevant.

3. Design pre-launch validation and monitoring

Propose setting up a holdout group, defining guardrail metrics (e.g., engagement, satisfaction), and establishing a baseline for model performance and calibration before launch.

4. Revalidate model fit one week post-launch

Outline steps: compare model predictions vs. actuals on new data, check calibration plots, compute performance metrics (AUC, RMSE), and run a statistical test (e.g., KS test) to detect distribution shift.

5. Decide on retraining or adjustments

Based on validation results, recommend whether to retrain, recalibrate, or add new features; emphasize iterative monitoring and potential A/B tests to measure model impact.

Key Points to Mention

  • Feature importance shifts due to new interaction patterns
  • Potential increase in model complexity or need for new features
  • Importance of a holdout group and guardrail metrics for causal inference
  • Calibration and performance metric monitoring (e.g., AUC, log loss)
  • Statistical tests for drift detection (e.g., KS test, PSI)
  • Iterative retraining and feedback loop with product teams

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