Model Unique Recipients with Poisson Distribution and Test Fit
Quick Overview
Evaluates statistical modeling of unique recipient counts per caller using count distributions. Strong answers compare Poisson, binomial, negative binomial, zero-inflated, and segmented models, then test fit with dispersion checks, likelihood metrics, chi-square tests, and simulation-based diagnostics.
Model Unique Recipients with Poisson Distribution and Test Fit
Company: Meta
Role: Data Scientist
Category: Statistics & Math
Difficulty: medium
Interview Round: Technical Screen
##### Scenario
Modelling the distribution of the number of unique recipients each caller contacts to support capacity and growth planning.
##### Question
Which statistical distribution would you use to model the number of recipients per caller and how would you test the goodness-of-fit?
##### Hints
Compare Poisson, binomial, negative-binomial; check over-dispersion; fit via maximum likelihood; validate with chi-square, KS test, or AIC.
Quick Answer: Evaluates statistical modeling of unique recipient counts per caller using count distributions. Strong answers compare Poisson, binomial, negative binomial, zero-inflated, and segmented models, then test fit with dispersion checks, likelihood metrics, chi-square tests, and simulation-based diagnostics.
Model Unique Recipients with Poisson Distribution and Test Fit
Meta
Jul 12, 2025, 6:59 PM
mediumData ScientistTechnical ScreenStatistics & Math
45
0
Model Unique Recipients with a Count Distribution and Test Fit
You need to model the distribution of the number of unique recipients each caller contacts in a fixed time window. The model will support capacity planning, growth forecasts, and product analysis.
Constraints & Assumptions
The outcome is a nonnegative integer count per caller over a fixed window, such as daily or weekly unique recipients.
Counts may have many zeros, heavy users, overdispersion, or segment differences.
The answer should explain how to choose candidate distributions and how to test goodness of fit.
Do not assume a Poisson model is correct without checking its assumptions.
Clarifying Questions to Ask Guidance
What time window is used: day, week, month, or active-caller session?
Are callers with zero recipients included, or only active callers?
Is there a known upper bound on possible recipients per caller?
Will the model be used for average load, tail capacity, forecasting, or user segmentation?
Part 1 - Choose Candidate Distributions
Which statistical distribution would you use to model the number of unique recipients per caller?
What This Part Should Cover Guidance
Poisson as a baseline for unbounded count data with mean approximately equal to variance.
Binomial if there is a known finite set of possible recipients or trials.
Negative binomial for overdispersed counts and heterogeneous caller behavior.
Zero-inflated or hurdle models if there are excess zeros.
Segment-specific models when one global distribution hides different user populations.
Part 2 - Fit the Model
How would you estimate the model parameters?
What This Part Should Cover Guidance
Maximum likelihood estimation for candidate distributions.
For a Poisson model, lambda estimated by the sample mean.
For negative binomial or zero-inflated models, numerical maximum likelihood or standard statistical libraries.
Train/validation splits or time-based holdouts if the model will be used for prediction.
Part 3 - Test Goodness of Fit
How would you test whether the chosen distribution fits the data well enough?
What This Part Should Cover Guidance
Exploratory checks comparing sample mean, variance, zero rate, and tail mass to model expectations.
AIC/BIC or held-out log likelihood to compare candidate models.
Pearson or deviance goodness-of-fit statistics, chi-square tests with binned counts, and simulation-based posterior or predictive checks.
Residual plots and segment-level diagnostics, especially for high-recipient tails important to capacity planning.
What a Strong Answer Covers Guidance
A strong answer treats Poisson as a starting point, checks dispersion and zeros, compares plausible count models, and chooses a model based on fit, interpretability, and the business use case.
Follow-up Questions Guidance
What would you do if the variance is much larger than the mean?
How would you model recipients for new callers with little history?
Which fit metric matters most if the goal is capacity planning for peak load?