← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Meta data scientist interview with a stats-heavy question about comparing two LLM outputs via hypothesis testing. One question, pretty deep on the math side, not a lot of fluff.

Questions Asked (1)

Q1

You ran an A/B test comparing two language models: Model A had a 70% success rate (700 out of 1000) and Model B had 80% (800 out of 1000). State your null and alternative hypotheses, compute a two-proportion z-statistic and p-value to determine if Model B is significantly better at α = 0.05, and provide the 95% confidence interval for the lift.

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

This one took me a second to organize.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly stating the null and alternative hypotheses, then compute the pooled proportion and standard error to calculate the z-statistic and p-value. Finally, compute the 95% confidence interval for the difference in proportions (lift) and interpret the results in the context of the business decision.

Pro tip: Always check the assumptions for the two-proportion z-test (independent samples, random assignment, and sufficient sample size) and mention them briefly to show rigor. Also, interpret the confidence interval in terms of practical significance, not just statistical significance.

1. State Hypotheses

Define null hypothesis H0: pA = pB (or pB - pA = 0) and alternative hypothesis H1: pB > pA (one-sided) or pB ≠ pA (two-sided). Since the question asks if B is significantly better, use a one-sided test.

2. Compute Test Statistic

Calculate the pooled proportion p̂ = (700+800)/(1000+1000) = 0.75. Then compute the standard error SE = sqrt(p̂(1-p̂)(1/nA + 1/nB)) = sqrt(0.75*0.25*(0.001+0.001)) = sqrt(0.000375) ≈ 0.01936. The z-statistic is (0.80 - 0.70)/SE = 0.10/0.01936 ≈ 5.16.

3. Find P-value and Make Decision

For a one-sided test, p-value = P(Z > 5.16) ≈ 0.00000012 (essentially 0). Since p-value < 0.05, reject H0. Conclude Model B is significantly better.

4. Compute Confidence Interval for Lift

Calculate the 95% confidence interval for the difference in proportions (pB - pA). Use unpooled SE for CI: SE_diff = sqrt(pA(1-pA)/nA + pB(1-pB)/nB) = sqrt(0.7*0.3/1000 + 0.8*0.2/1000) = sqrt(0.00021 + 0.00016) = sqrt(0.00037) ≈ 0.01924. The 95% CI is (0.10 ± 1.96*0.01924) = (0.10 ± 0.0377) = (0.0623, 0.1377). So, we are 95% confident the lift is between 6.23% and 13.77%.

5. Interpret Results

Explain that the result is statistically significant and the confidence interval suggests a positive lift. Discuss practical significance: a 10% absolute lift (from 70% to 80%) is substantial and likely meaningful for the business.

Key Points to Mention

  • Null and alternative hypotheses (one-sided vs two-sided)
  • Pooled proportion for z-test and unpooled for confidence interval
  • Assumptions: independence, random assignment, large sample size
  • Z-statistic calculation and p-value interpretation
  • Confidence interval interpretation and practical significance
  • Potential pitfalls: multiple testing, novelty effects, and segment analysis

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