← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

Amazon stats interview, probably for a data or analytics role. Two questions, both about t-tests. Short but the follow-up is the one that actually matters.

Questions Asked (2)

Q1

What assumptions does the t-test make?

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

Ran through the usual list: normality, independence, roughly equal variances for the two-sample version.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by listing the core assumptions of the t-test, then explain why each matters and how violations can be detected and addressed. Emphasize practical implications in A/B testing at Amazon, showing you understand both theory and application.

Pro tip: Mention that with large sample sizes (common in Amazon experiments), the t-test is robust to some violations, but always check for practical significance and consider alternatives like bootstrapping or non-parametric tests when assumptions are severely violated.

1. State the core assumptions

List the main assumptions: independence, normality (or large sample), homogeneity of variance, and for paired t-tests, paired differences are normal.

2. Explain each assumption

Briefly describe what each assumption means and why it is important for the validity of the t-test.

3. Discuss violations and detection

Explain how to check each assumption (e.g., plots, Levene's test) and what happens if violated (e.g., inflated Type I error).

4. Provide remedies and alternatives

Suggest solutions: transformations, Welch's t-test, non-parametric tests, or bootstrapping. Mention robustness with large samples.

5. Relate to Amazon context

Connect to A/B testing: large samples often make t-test robust, but always consider practical significance and business impact.

Key Points to Mention

  • Independence of observations
  • Normality of data or sampling distribution (CLT for large samples)
  • Homogeneity of variance (especially for two-sample t-test)
  • Robustness of t-test with large sample sizes
  • Welch's t-test as a solution for unequal variances
  • Non-parametric alternatives (e.g., Mann-Whitney U) when assumptions fail

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

Q2

When would you use a different statistical test instead of a t-test?

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

This is where I stumbled a bit.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that the t-test is a parametric test with specific assumptions (normality, independence, homogeneity of variance, continuous data). Then systematically walk through scenarios where these assumptions are violated or the data type is different, explaining which alternative test to use and why. Finally, tie it back to practical A/B testing contexts at Amazon, emphasizing the importance of choosing the right test to avoid false positives/negatives.

Pro tip: Mention that in large-scale A/B tests, the t-test is often robust due to the Central Limit Theorem, but you must still check for ratio metrics or skewed distributions where alternatives like bootstrapping or Mann-Whitney U are more appropriate. This shows you understand both theory and real-world application.

1. State the assumptions of the t-test

Briefly list the key assumptions: independence, normality (or large sample), homogeneity of variance, and continuous data. This sets the stage for when to use alternatives.

2. Identify violations and data types

Discuss common violations: non-normal data (especially small samples), ordinal data, unequal variances, paired vs. independent samples, and multiple groups. Also consider non-continuous outcomes like counts or proportions.

3. Map violations to alternative tests

For each violation, name the appropriate test: Mann-Whitney U or Wilcoxon for non-normal/ordinal, Welch's t-test for unequal variances, paired t-test or Wilcoxon signed-rank for paired data, ANOVA or Kruskal-Wallis for multiple groups, chi-square for categorical, and bootstrapping for complex metrics.

4. Relate to A/B testing and product analytics

Explain how these alternatives apply in A/B testing: e.g., using Mann-Whitney for skewed revenue data, bootstrapping for ratio metrics like CTR, or chi-square for conversion rates. Highlight the impact on decision-making.

5. Conclude with practical guidance

Summarize that the choice depends on data characteristics and business context, and emphasize the importance of checking assumptions before defaulting to a t-test.

Key Points to Mention

  • Non-normal data or small sample sizes: use non-parametric tests like Mann-Whitney U or Wilcoxon signed-rank.
  • Unequal variances between groups: use Welch's t-test.
  • Paired or dependent samples: use paired t-test or Wilcoxon signed-rank test.
  • More than two groups: use ANOVA or Kruskal-Wallis.
  • Categorical data: use chi-square test or Fisher's exact test.
  • Complex metrics (e.g., ratios, percentiles) or skewed distributions: use bootstrapping or permutation tests.

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