← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Meta DS interview that went deep into statistics pretty fast. The whole thing was built around a single dataset and they kept pulling threads from it, so if you're shaky on distribution intuition or quantile math you'll feel it.

Questions Asked (4)

Q1

Given a right-skewed watch time dataset, describe the shape of the distribution verbally. What does the tail look like and why?

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

I knew it was right-skewed the moment I saw 180 sitting next to a bunch of single-digit values, but I fumbled the verbal description.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly stating that the distribution is right-skewed, meaning it has a long tail extending to the right. Then describe the tail's shape (e.g., heavy, long) and explain why it occurs in watch time data—because a small fraction of users watch extremely long content, while most watch shorter durations. Finally, connect this to implications for analysis, such as using median over mean or log transformations.

Pro tip: Mention that the tail is often heavy (e.g., power-law-like) and that this affects metric selection and experiment analysis—showing you understand both the statistics and the product context.

1. Identify the skew

State that the distribution is right-skewed (positively skewed), meaning the bulk of the data is concentrated on the left with a long tail to the right.

2. Describe the tail

Explain that the tail is long and extends toward higher values, indicating a small number of observations with very large watch times.

3. Explain the cause

Discuss why this happens: user behavior is heterogeneous—most users watch short sessions, but a few binge-watch or leave videos running, creating extreme values.

4. Discuss implications

Mention how this shape affects summary statistics (mean > median) and analysis choices (e.g., use median, log-transform, or robust methods).

5. Connect to product context

Relate to Meta's context: heavy-tailed watch time impacts A/B tests (variance, outliers) and product decisions (e.g., optimizing for engaged users vs. casual viewers).

Key Points to Mention

  • Right-skewed (positive skew) with a long right tail
  • Mean is greater than median due to extreme high values
  • Tail is heavy—some users watch extremely long durations
  • Causes: binge-watching, background play, varied user intent
  • Implications: use median or trimmed mean, consider log transformation
  • Impact on A/B testing: increased variance, need for robust methods

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

Q2

Compute the sample median, mode, and 95th percentile using linear interpolation for empirical quantiles. Walk through each step.

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

This is where I lost some time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by sorting the data and clarifying the definitions of median, mode, and 95th percentile. For median and percentile, use the linear interpolation formula for empirical quantiles, and for mode, identify the most frequent value(s). Walk through each calculation step-by-step with a small example dataset to illustrate the process.

Pro tip: Mention that linear interpolation for quantiles is the default in many tools (e.g., numpy's percentile with method='linear') and that the median is a special case of the 50th percentile. This shows practical awareness and can preempt follow-up questions about implementation.

1. Sort and understand the data

Sort the dataset in ascending order. Clarify the size (n) and whether the data is discrete or continuous, as this affects mode calculation.

2. Compute the median

If n is odd, the median is the middle value. If n is even, it's the average of the two middle values. Alternatively, use the 50th percentile formula with linear interpolation.

3. Compute the mode

Count the frequency of each value. The mode is the value(s) with the highest frequency. If all values are unique, there is no mode.

4. Compute the 95th percentile using linear interpolation

Calculate the rank: r = (p/100)*(n-1) + 1, where p=95. Let k = floor(r) and d = r - k. The percentile is x_k + d*(x_{k+1} - x_k), with 1-based indexing. Handle edge cases when r is an integer.

5. Verify and interpret

Check calculations with a small example. Interpret the results in context: median as central tendency, mode as most common, and 95th percentile as a threshold below which 95% of data falls.

Key Points to Mention

  • Definition of empirical quantiles and the linear interpolation formula: Q(p) = x_k + d*(x_{k+1} - x_k) where r = (p/100)*(n-1)+1, k=floor(r), d=r-k.
  • Median is the 50th percentile; for even n, it's the average of the two middle values, which aligns with linear interpolation.
  • Mode can be unimodal, bimodal, or multimodal; if no value repeats, there is no mode.
  • Edge cases: when r is an integer, d=0, so the percentile is exactly x_k; when p=0 or 100, the percentile is the min or max.
  • Use of 1-based indexing in the formula; clarify indexing to avoid off-by-one errors.
  • Practical implementation: mention that numpy.percentile with method='linear' uses this approach, and pandas quantile defaults to linear interpolation.

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

Q3

Compare the mean and median for this dataset. Which is a more reliable location estimator for decision-making, and why?

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

Pretty comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the mean and median and explaining how each is calculated. Then compare their values for the given dataset, discussing what the difference (or lack thereof) indicates about the distribution. Finally, argue which estimator is more reliable for decision-making, considering the context of product analytics and A/B testing at Meta.

Pro tip: Always relate the choice of estimator to the business decision at hand—for example, if the metric is revenue per user, the mean is often more relevant despite outliers, but for engagement time, the median may better represent typical user behavior.

1. Define and compute

Clearly define mean and median, and compute both for the dataset (or explain how you would compute them).

2. Compare and interpret

Compare the two values and interpret any difference in terms of skewness, outliers, or distribution shape.

3. Assess reliability

Discuss the robustness of each estimator: mean is sensitive to outliers, median is resistant. Consider the impact on decision-making.

4. Contextualize for decision-making

Relate to the specific business context (e.g., product metrics, A/B testing) to argue which estimator is more reliable for the decision at hand.

5. Conclude with recommendation

Provide a clear recommendation on which estimator to use, and possibly suggest additional statistics (e.g., trimmed mean) if appropriate.

Key Points to Mention

  • Mean is sensitive to outliers and skewness; median is robust.
  • In symmetric distributions, mean and median are similar; in skewed distributions, they diverge.
  • For decision-making, consider the business metric: revenue (mean) vs. typical user experience (median).
  • In A/B testing, the choice of estimator affects power and interpretation; median may be better for skewed metrics like session time.
  • Meta often deals with large-scale data where outliers are common; robust estimators can prevent misleading conclusions.
  • Always visualize the distribution (e.g., histogram) to understand the data before choosing an estimator.

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

Q4

If you winsorize watch time at the 99th percentile before running an A/B test, what happens to variance and statistical power qualitatively?

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

Winsorizing caps the extreme values so variance drops, which tightens the distribution.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, explain that winsorizing at the 99th percentile caps extreme values, which reduces the variance of the metric. Then, discuss how lower variance generally increases statistical power, but also consider the trade-off of potentially biasing the estimate and the impact on the metric's sensitivity to true effects.

Pro tip: Mention that winsorizing can be particularly useful when the metric has a heavy-tailed distribution, but always check if the treatment effect is concentrated in the tail; if so, winsorizing might mask the effect.

1. Define winsorizing

Briefly explain that winsorizing at the 99th percentile replaces values above the 99th percentile with the value at the 99th percentile, reducing the influence of outliers.

2. Effect on variance

State that winsorizing reduces the variance of the metric because extreme values are capped, leading to a more stable distribution.

3. Effect on statistical power

Explain that lower variance typically increases statistical power (i.e., the ability to detect a true effect) for a given sample size, as the standard error decreases.

4. Consider trade-offs

Discuss potential downsides: winsorizing may bias the estimate of the mean and reduce sensitivity to effects that occur in the tail, potentially lowering power for those specific effects.

5. Contextual recommendation

Conclude that winsorizing can be beneficial for heavy-tailed metrics but should be applied carefully, ideally with sensitivity analysis to ensure it doesn't distort the treatment effect.

Key Points to Mention

  • Winsorizing reduces variance by capping extreme values.
  • Lower variance decreases standard error, increasing statistical power.
  • Potential bias in mean estimate due to capping.
  • Reduced sensitivity to treatment effects in the tail.
  • Trade-off between robustness and power for tail effects.
  • Importance of checking metric distribution and effect location.

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