← Upstart Interview Insights

Upstart·Data Scientist·Technical Phone Screen·Intermediate

IntermediatePrefer not to say
Apr 2026

Summary

Upstart data scientist interview had a probability/statistics question that looks deceptively simple but is actually a sampling bias trap. The math itself isn't hard once you see what's going on, but I almost blew it by treating the school survey as a direct household estimate.

Questions Asked (3)

Q1

You survey 100 children at a school and ask how many kids are in their family. 50 say 1, 20 say 2, 30 say 3. Then you knock on a random house in the same town and ask how many children live there. What's the probability it's exactly 1 child?

Product Analytics & MetricsRoot Cause Analysis
Author's notes

My first instinct was 50% because 50 out of 100 kids said they were in a 1-child family.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that the survey data reflects the distribution of children's family sizes (size-biased), while the house question asks for the distribution of family sizes (unbiased). Convert the child-based counts to family-based counts by dividing each count by the number of children in that family size, then normalize to get the probability that a randomly selected house has exactly 1 child.

Pro tip: Explicitly state the assumption that each family has exactly one house and that the town's family size distribution matches the school's. This shows you understand the sampling bias and the need for a clear model.

1. Identify the two different sampling frames

The survey samples children (size-biased), while the house question samples families (unbiased). Clarify that the given counts are from the children's perspective.

2. Convert child counts to family counts

For each family size k, the number of families is the number of children reporting size k divided by k. Compute: 50/1 = 50 families of size 1, 20/2 = 10 families of size 2, 30/3 = 10 families of size 3.

3. Calculate total families and probability

Total families = 50 + 10 + 10 = 70. The probability that a randomly chosen house has exactly 1 child is 50/70 = 5/7 ≈ 0.714.

4. Sanity-check and discuss assumptions

Verify that the probability is higher than the naive 50% from the survey, which makes sense because larger families are overrepresented in the child sample. Mention assumptions like one house per family and no other family sizes.

Key Points to Mention

  • Size bias in the survey: children from larger families are more likely to be sampled.
  • Conversion from child-based to family-based counts by dividing by family size.
  • The naive answer (50%) is incorrect because it ignores the sampling bias.
  • Assumption that each family occupies exactly one house and that the school's family size distribution is representative of the town.
  • The final probability is 5/7 or approximately 71.4%.
  • This is an application of the inspection paradox or length-biased sampling.

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

Q2

Why doesn't the school survey directly tell you the household distribution? What's the statistical issue?

Root Cause AnalysisProduct Analytics & Metrics
Author's notes

This is the part I fumbled first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Explain that the survey's household distribution is not directly observed due to sampling bias and non-response bias, which distort the representativeness of the sample. Discuss how these biases lead to a mismatch between the survey's household distribution and the true population distribution, requiring statistical adjustments like weighting or post-stratification.

Pro tip: Acknowledge that while weighting can correct for known biases, it relies on accurate population benchmarks; if these are unavailable or outdated, the adjusted estimates may still be biased. Always validate with external data sources when possible.

1. Identify the target population

Clarify what the 'household distribution' refers to (e.g., income, size, geography) and the true population of interest. This sets the baseline for comparison.

2. Explain sampling bias

Describe how the survey's sampling method (e.g., random digit dialing, online panels) may over- or under-represent certain household types, leading to a non-representative sample.

3. Discuss non-response bias

Highlight that even with a good sampling frame, differential non-response (e.g., certain households less likely to respond) skews the observed distribution.

4. Describe statistical adjustments

Mention techniques like weighting, post-stratification, or raking that attempt to correct biases by aligning sample distributions with known population margins.

5. Conclude with implications

Summarize that without adjustments, the survey's household distribution is biased and not directly usable; with adjustments, it's an estimate with uncertainty.

Key Points to Mention

  • Sampling bias: non-representative sample due to selection mechanisms
  • Non-response bias: systematic differences between respondents and non-respondents
  • Coverage error: some households have no chance of being selected
  • Weighting and post-stratification to correct biases
  • Importance of external benchmarks for validation
  • Uncertainty in adjusted estimates due to assumptions

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

Q3

How would your answer change if households with zero children existed, or if families larger than 3 were possible but just not observed in the sample?

Root Cause AnalysisAdaptability & Ambiguity
Author's notes

Didn't have a clean answer here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the distinction between structural zeros (impossible values) and sampling zeros (possible but unobserved). Then, discuss how each scenario affects model assumptions, estimation, and inference, and propose adjustments such as zero-inflated models or truncated distributions.

Pro tip: Always connect the statistical adjustment to the business context—e.g., how zero-child households might represent a different customer segment, or how unobserved large families could bias predictions for future growth.

1. Clarify the nature of the zeros

Determine whether zero children are structurally impossible or just unobserved in the sample. This distinction drives whether you need a zero-inflated model or a truncated model.

2. Assess impact on model assumptions

Evaluate how the presence of structural zeros or unobserved larger families violates assumptions of common models (e.g., Poisson, negative binomial) and affects parameter estimates.

3. Choose appropriate modeling strategies

For structural zeros, consider zero-inflated or hurdle models; for unobserved larger families, consider truncated models or Bayesian methods with informative priors.

4. Validate and interpret results

Use cross-validation, posterior predictive checks, or sensitivity analysis to ensure the model generalizes. Interpret coefficients in light of the adjusted data-generating process.

5. Communicate implications

Explain how the adjustments affect business insights, such as targeting zero-child households differently or preparing for potential larger families in future data.

Key Points to Mention

  • Structural zeros vs. sampling zeros
  • Zero-inflated models (e.g., ZIP, ZINB) and hurdle models
  • Truncated distributions and their estimation challenges
  • Impact on model assumptions and parameter interpretation
  • Bayesian approaches with informative priors for unobserved data
  • Business implications of different household compositions

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