← Apple Interview Insights

Apple·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Apple data scientist interview that went pretty deep into statistical theory. Four questions, all probability and inference, no product sense or SQL anywhere. Felt more like a stats PhD oral exam than a job interview.

Questions Asked (4)

Q1

You have 200 independent minute-level event counts with sample mean 14.5 and sample variance 16.2. Under a Poisson model, derive the MLE for lambda, its asymptotic variance, and build a 95% Wald confidence interval.

Product Analytics & MetricsTechnical Trade-offs
Author's notes

MLE derivation was fine, lambda-hat equals x-bar, straightforward log-likelihood stuff.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by stating the Poisson MLE for λ is the sample mean, then derive its asymptotic variance using the Fisher information. Use the plug-in estimator for variance and construct the Wald interval with the normal approximation, ensuring to mention the delta method if needed for transformations.

Pro tip: Emphasize that the Poisson model assumes mean equals variance; here the sample variance (16.2) is close to the mean (14.5), so the model is reasonable. Also, note that the Wald interval can be improved with transformations like log or square root for better coverage, but stick to the question's request for Wald.

1. State the Poisson MLE

For independent Poisson counts, the MLE for λ is the sample mean, so λ_hat = 14.5.

2. Derive asymptotic variance

The Fisher information for a single observation is 1/λ, so for n=200, the asymptotic variance of λ_hat is λ/n. Plug in λ_hat to get 14.5/200 = 0.0725.

3. Compute standard error

The standard error is the square root of the variance: sqrt(0.0725) ≈ 0.2693.

4. Construct 95% Wald interval

Use the formula λ_hat ± 1.96 * SE. This gives 14.5 ± 1.96*0.2693, resulting in approximately (13.972, 15.028).

5. Interpret and check assumptions

Interpret the interval and note that the Poisson assumption seems plausible since sample variance (16.2) is close to mean (14.5).

Key Points to Mention

  • MLE for Poisson λ is the sample mean.
  • Asymptotic variance of MLE is λ/n, estimated by λ_hat/n.
  • Standard error is sqrt(λ_hat/n).
  • Wald interval uses normal approximation: λ_hat ± z_{α/2} * SE.
  • Check mean-variance relationship: sample variance ≈ sample mean supports Poisson.
  • Mention potential improvements like transformations for better coverage.

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

Q2

Test whether the data follows a Poisson distribution (equidispersion) versus being overdispersed. Specify the test statistic, its reference distribution, and give a decision at alpha=0.05 using the provided sample mean and variance.

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

This is where I actually blanked for a second.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that the test for equidispersion versus overdispersion in a Poisson context is based on the dispersion index (variance-to-mean ratio). Then, compute the test statistic as (n-1)*sample_variance/sample_mean, which follows a chi-square distribution with n-1 degrees of freedom under the null hypothesis of equidispersion. Finally, compare the test statistic to the critical value from the chi-square distribution at alpha=0.05 to make a decision.

Pro tip: When presenting the decision, explicitly state the null and alternative hypotheses: H0: variance = mean (equidispersion) vs. H1: variance > mean (overdispersion). Also, mention that this test assumes the sample size is large enough for the chi-square approximation to hold.

1. State the hypotheses

Define the null hypothesis (equidispersion: variance equals mean) and the alternative hypothesis (overdispersion: variance greater than mean).

2. Compute the test statistic

Calculate the dispersion test statistic as (n-1)*sample_variance/sample_mean, where n is the sample size.

3. Determine the reference distribution

Under the null hypothesis, the test statistic follows a chi-square distribution with n-1 degrees of freedom.

4. Find the critical value and make a decision

At alpha=0.05, find the critical value from the chi-square distribution (or compute the p-value) and compare it to the test statistic. Reject H0 if the test statistic exceeds the critical value (or if p-value < 0.05), indicating overdispersion.

Key Points to Mention

  • The dispersion index (variance-to-mean ratio) is the key quantity; a value greater than 1 suggests overdispersion.
  • The test statistic (n-1)*s^2/mean follows a chi-square distribution with n-1 degrees of freedom under the null.
  • This is a one-sided test because we are specifically testing for overdispersion (variance > mean).
  • The test assumes that the data are a random sample from a Poisson distribution under the null.
  • If overdispersion is detected, consider using a negative binomial or quasi-Poisson model instead of Poisson.
  • For large samples, the chi-square approximation is robust, but for small samples, exact tests may be needed.

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

Q3

For a single Poisson random variable with lambda=120, approximate P(100 <= X <= 140) using the Normal approximation with continuity correction. Write out the exact Normal integral and explain why the correction matters.

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

I liked this one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, recall that for a Poisson(λ=120), the mean and variance are both 120, so the standard deviation is √120 ≈ 10.954. Apply the continuity correction by converting the discrete bounds to continuous intervals: P(100 ≤ X ≤ 140) ≈ P(99.5 < Y < 140.5) where Y ~ Normal(120, 120). Then standardize the bounds and express the probability as a difference of standard normal CDFs, explaining why the correction improves accuracy.

Pro tip: Mention that without continuity correction, the approximation would use P(100 < Y < 140), which shifts both bounds by 0.5 and can lead to noticeable error, especially near the tails. Also, note that the correction is most impactful when the interval is narrow or when λ is small, but even for λ=120 it improves the approximation by accounting for the discrete nature of Poisson.

1. Identify parameters and conditions

State that X ~ Poisson(λ=120), so mean μ = λ = 120 and variance σ² = λ = 120, hence σ = √120 ≈ 10.954. Confirm that λ is large enough for a good normal approximation.

2. Apply continuity correction

Convert the discrete bounds to continuous: P(100 ≤ X ≤ 140) ≈ P(99.5 < Y < 140.5) where Y ~ N(120, 120). Explain that this adjustment accounts for the fact that a discrete value k corresponds to the interval [k-0.5, k+0.5] on the continuous scale.

3. Standardize the bounds

Compute z-scores: z1 = (99.5 - 120)/√120 ≈ -1.872, z2 = (140.5 - 120)/√120 ≈ 1.872. Write the probability as Φ(1.872) - Φ(-1.872) = 2Φ(1.872) - 1.

4. Express the exact normal integral

Write the probability as ∫_{99.5}^{140.5} (1/√(2π·120)) exp(-(x-120)²/(2·120)) dx. This is the exact normal integral with continuity correction.

5. Explain why continuity correction matters

Discuss that without it, the approximation would use P(100 < Y < 140), which ignores the 0.5 shift and can underestimate or overestimate the probability. The correction improves accuracy by aligning the discrete distribution with the continuous one, especially for small intervals or when λ is not very large.

Key Points to Mention

  • Poisson mean and variance both equal λ = 120, so standard deviation = √120 ≈ 10.954.
  • Continuity correction: adjust discrete bounds by ±0.5 to get 99.5 and 140.5.
  • Standardization: z = (x - μ)/σ, yielding z-scores approximately ±1.872.
  • Exact normal integral: ∫_{99.5}^{140.5} (1/√(2π·120)) exp(-(x-120)²/(2·120)) dx.
  • Why correction matters: it accounts for the discrete nature of Poisson and improves approximation accuracy, especially near the tails or for narrow intervals.
  • The normal approximation is valid here because λ=120 is large (typically λ > 20 is sufficient).

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

Q4

Under what conditions is a Normal approximation to a Poisson valid, when does it break down, and what alternative model would you use when the sample variance is much larger than the sample mean? Name one diagnostic you'd check.

Technical Trade-offsRoot Cause Analysis
Author's notes

Rushed this one because we were running low on time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by stating the conditions for a valid Normal approximation to the Poisson distribution, emphasizing that both the mean and variance must be equal and sufficiently large. Then explain when the approximation breaks down, particularly when the variance exceeds the mean (overdispersion). Finally, introduce the Negative Binomial as an alternative model and mention a diagnostic like the dispersion test or a plot of residuals.

Pro tip: Demonstrate practical experience by noting that in real-world data, overdispersion is common and can lead to underestimated standard errors if ignored. Mention that Apple often deals with large-scale user data where such nuances matter for reliable inference.

1. State conditions for Normal approximation

Explain that the Poisson distribution can be approximated by a Normal distribution when the mean (λ) is large (typically λ > 20) and events are independent. The Normal approximation uses mean = λ and variance = λ.

2. Identify breakdown scenarios

Discuss that the approximation fails when λ is small (e.g., λ < 5) due to skewness and discreteness, and when the variance is much larger than the mean (overdispersion), indicating the Poisson assumption is violated.

3. Propose alternative model

Introduce the Negative Binomial distribution as a flexible alternative that accounts for overdispersion by allowing variance to exceed the mean. Mention that it reduces to Poisson when the dispersion parameter is zero.

4. Name a diagnostic check

Suggest a diagnostic such as the dispersion test (e.g., Cameron & Trivedi test) or examining the ratio of residual deviance to degrees of freedom. A ratio significantly greater than 1 indicates overdispersion.

Key Points to Mention

  • Poisson distribution assumes mean equals variance.
  • Normal approximation requires large λ (e.g., λ > 20) and independence.
  • Overdispersion occurs when variance > mean, violating Poisson assumptions.
  • Negative Binomial is a common alternative for overdispersed count data.
  • Diagnostic: dispersion test or residual deviance/df ratio.
  • Ignoring overdispersion can lead to underestimated standard errors and false positives.

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