← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Senior

Senior
Apr 2026

Summary

Amazon data scientist interview with a stats-heavy question about p-values in the context of A/B testing. One question, but it had a lot of surface area to cover and I don't think I hit all of it.

Questions Asked (1)

Q1

You're running an A/B test for a new product feature and stakeholders are using p-values to make the ship decision. Walk through at least three real limitations of p-values in this setting and what you'd recommend instead or alongside them.

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

I started with sample size dependence because it felt safest, basically that with enough users you can get a significant p-value on a completely trivial effect.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that p-values are useful for detecting statistical significance but insufficient for product decisions, then systematically explain at least three limitations (e.g., binary thinking, ignoring effect size, multiple testing, peeking, practical significance) and propose complementary methods like confidence intervals, Bayesian methods, and decision-theoretic frameworks. Emphasize that the ship decision should balance statistical evidence with business impact, risk, and cost.

Pro tip: Frame the answer around Amazon's leadership principles, especially 'Customer Obsession' and 'Deliver Results'—show that you prioritize customer impact and long-term value over statistical rituals. Mention that you'd align stakeholders on a pre-registered decision rule that incorporates both statistical and practical significance.

1. Clarify the role of p-values

Briefly state that p-values measure the probability of observing data at least as extreme as the current data, assuming the null hypothesis is true. They do not measure the probability that the null is true, nor the size or importance of an effect.

2. Identify key limitations

Discuss at least three limitations: (1) p-values are binary and don't convey effect size or uncertainty; (2) they are sensitive to sample size and can be gamed via peeking or multiple comparisons; (3) they don't reflect practical or business significance; (4) they assume a single test and no sequential analysis.

3. Recommend alternatives and complements

Propose using confidence intervals to show effect size and uncertainty, Bayesian methods to directly estimate the probability that the new feature is better, and decision-theoretic approaches that weigh costs and benefits. Also suggest sequential testing or alpha-spending to handle peeking.

4. Tie to business context

Explain how to incorporate business metrics (e.g., revenue, customer satisfaction) and practical significance thresholds. Recommend pre-registering the decision rule and involving stakeholders in defining what 'ship' means in terms of expected value.

5. Summarize a holistic decision framework

Conclude that the ship decision should combine statistical evidence (e.g., Bayesian posterior probability, confidence intervals) with business impact, risk assessment, and strategic alignment, rather than relying solely on p-values.

Key Points to Mention

  • p-values do not measure effect size or practical significance
  • p-values are often misinterpreted as the probability that the null hypothesis is true
  • multiple testing and peeking inflate false positive rates
  • confidence intervals provide a range of plausible effect sizes
  • Bayesian methods can directly estimate the probability that the new feature is better
  • decision-theoretic frameworks incorporate costs, benefits, and risk tolerance

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