← Capital One Interview Insights

Capital One·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Jul 2026

Summary

Capital One data scientist interview that was basically a stats exam. Three heavily computational questions back to back, all requiring you to show your work. Not a vibe check, not a case study, just raw probability and inference.

Questions Asked (3)

Q1

You have an A/B test comparing two variants on conversion rate. Variant A got 410 conversions from 5000 visitors, Variant B got 470 from 5100. Compute a two-sided 95% confidence interval for the difference in proportions using both the normal approximation with unpooled standard error and the Wilson-Newcombe method. Is B statistically different from A under each approach? Also, if you were running 3 variants against A and needed to control family-wise error at 5% via Bonferroni, what confidence level would you use per comparison, and does your conclusion about B change?

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

This was a lot.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by computing the sample proportions and their difference, then calculate the confidence interval using both the unpooled normal approximation and the Wilson-Newcombe method. Compare whether the interval includes zero to determine statistical significance. Finally, apply Bonferroni correction for multiple comparisons and reassess the conclusion.

Pro tip: Always state the assumptions behind each method (e.g., normality, sample size) and note that the Wilson-Newcombe method is more robust for smaller samples or extreme proportions. This shows depth beyond just plugging numbers.

1. Compute sample proportions and difference

Calculate conversion rates for A and B: p_A = 410/5000 = 0.082, p_B = 470/5100 ≈ 0.09216. The observed difference is p_B - p_A ≈ 0.01016.

2. Normal approximation with unpooled SE

Compute standard error as sqrt(p_A(1-p_A)/n_A + p_B(1-p_B)/n_B). The 95% CI is difference ± 1.96*SE. Check if zero is inside the interval.

3. Wilson-Newcombe method

Compute Wilson intervals for each proportion separately, then combine them to get the Newcombe interval for the difference. This method adjusts for skewness and is more accurate for proportions.

4. Bonferroni correction

For 3 comparisons against A, the family-wise error rate is 5%, so per-comparison alpha = 0.05/3 ≈ 0.0167. The corresponding confidence level is 1 - 0.0167 = 0.9833 or 98.33%. Recompute the CI using the adjusted z-value (e.g., z ≈ 2.128 for 98.33% two-sided).

5. Compare conclusions

Determine if the intervals include zero under each method and after Bonferroni adjustment. Discuss whether the conclusion about B changes.

Key Points to Mention

  • Difference in proportions: p_B - p_A ≈ 0.01016 (1.016 percentage points).
  • Unpooled SE: sqrt(0.082*0.918/5000 + 0.09216*0.90784/5100) ≈ 0.00567. 95% CI: (0.01016 ± 1.96*0.00567) = (-0.00095, 0.02127). Includes zero, so not significant at 5%.
  • Wilson-Newcombe method: Compute Wilson intervals for each proportion: for A, (0.0747, 0.0900); for B, (0.0844, 0.1005). Newcombe interval for difference: (0.0844-0.0900, 0.1005-0.0747) = (-0.0056, 0.0258). Includes zero, so not significant.
  • Bonferroni: For 3 comparisons, per-comparison alpha = 0.05/3 ≈ 0.0167, so confidence level = 98.33%. Using z ≈ 2.128, unpooled CI: (0.01016 ± 2.128*0.00567) = (-0.0019, 0.0222). Still includes zero, so conclusion unchanged.
  • Assumptions: Normal approximation assumes large sample and not extreme proportions; Wilson-Newcombe is more robust. Bonferroni is conservative and may increase Type II error.
  • Interpretation: Under all methods, B is not statistically different from A at the 5% significance level, even after Bonferroni correction.

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

Q2

A sample of 40 users has a mean session time of 8.2 minutes and a sample standard deviation of 2.4 minutes. Compute the 95% confidence interval for the population mean using the t-distribution, and state the margin of error using the t value at 39 degrees of freedom.

Product Analytics & Metrics
Author's notes

Straightforward t-interval question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, identify the sample size, mean, standard deviation, and confidence level. Then, use the t-distribution with 39 degrees of freedom to find the critical t-value, compute the standard error, and construct the confidence interval. Finally, state the margin of error clearly.

Pro tip: Always check whether the population standard deviation is known; if not, use the t-distribution. Also, mention that the t-distribution is robust to non-normality for sample sizes around 40, but verify assumptions if the data is heavily skewed.

1. Identify given values

Extract the sample size (n=40), sample mean (x̄=8.2), sample standard deviation (s=2.4), and confidence level (95%).

2. Determine the critical t-value

For a 95% confidence interval with 39 degrees of freedom, find the t-value (approximately 2.023). This can be obtained from a t-table or software.

3. Calculate the standard error

Compute the standard error as s/√n = 2.4/√40 ≈ 0.3795.

4. Compute the margin of error

Multiply the t-value by the standard error: 2.023 * 0.3795 ≈ 0.768 minutes.

5. Construct the confidence interval

The 95% confidence interval is x̄ ± margin of error: 8.2 ± 0.768, resulting in (7.432, 8.968) minutes.

Key Points to Mention

  • Use of t-distribution because population standard deviation is unknown.
  • Degrees of freedom = n - 1 = 39.
  • Critical t-value for 95% confidence and 39 df is approximately 2.023.
  • Standard error calculation: s/√n.
  • Margin of error interpretation: we are 95% confident that the true population mean lies within this interval.
  • Assumption: sample is approximately normally distributed or sample size is large enough (n=40 is often considered sufficient).

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

Q3

What minimum per-group sample size do you need to estimate a proportion with 95% confidence and a margin of error no greater than 0.01, using the conservative p equals 0.5 assumption? Show the formula.

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

The formula is n = z squared times p times (1 minus p) divided by e squared.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by stating the formula for sample size for a proportion with conservative p=0.5, then plug in the values for 95% confidence (z=1.96) and margin of error 0.01. Compute the result and round up to the next integer, explaining that this is the minimum per-group sample size.

Pro tip: Mention that this is a conservative estimate because p=0.5 maximizes variance, and in practice you might use a pilot estimate to reduce sample size. Also, note that for A/B tests you need this sample size per variant, and consider any expected non-response or attrition.

1. State the formula

Write the sample size formula for estimating a proportion: n = (z^2 * p * (1-p)) / E^2, where z is the z-score for the confidence level, p is the estimated proportion, and E is the margin of error.

2. Identify parameters

For 95% confidence, z = 1.96. Use conservative p = 0.5. Margin of error E = 0.01.

3. Plug in and compute

Calculate n = (1.96^2 * 0.5 * 0.5) / 0.01^2 = (3.8416 * 0.25) / 0.0001 = 0.9604 / 0.0001 = 9604.

4. Round up and interpret

Since sample size must be an integer, round up to 9604. This means you need at least 9604 observations per group to achieve the desired precision.

Key Points to Mention

  • Formula for sample size for a proportion: n = (z^2 * p * (1-p)) / E^2
  • Conservative assumption p = 0.5 maximizes variance and gives the largest required sample size
  • z-score for 95% confidence is 1.96
  • Margin of error E = 0.01
  • Calculation yields 9604, which is the minimum per-group sample size
  • In practice, consider design effect, non-response, and multiple testing adjustments

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