← Upstart Interview Insights

Upstart·Data Scientist·Onsite - Multi Round·Intermediate

Intermediate
May 2026

Summary

Statistics-heavy onsite round at Upstart for a data scientist role. Three questions, all technical, all with some real bite to them. Not the kind of round where you can coast on intuition.

Questions Asked (3)

Q1

A population has one-, two-, and three-child families. Given a sample of 100 children (not families), estimate the proportion of each family type and construct a 95% confidence interval.

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

This one tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that sampling children introduces size-biased sampling: the probability of selecting a child from a family is proportional to family size. Use the observed distribution of family sizes among sampled children to estimate the underlying family-size distribution via a weighted estimator, then construct a confidence interval for the proportion of each family type using the delta method or bootstrap.

Pro tip: Explicitly state the size-biased sampling assumption and show how ignoring it leads to overestimating larger families. Mention that the estimator is essentially a weighted average with weights inversely proportional to family size.

1. Define the problem and assumptions

Clarify that the sample is of children, not families, so each child is sampled with probability proportional to their family size. Assume families are independent and the sample is random.

2. Derive the estimator for family-type proportions

Let p1, p2, p3 be the true proportions of one-, two-, and three-child families. The probability a randomly selected child comes from a family of size k is k*pk / (p1+2p2+3p3). Use the observed counts of children from each family size to solve for p1, p2, p3.

3. Compute the estimates from the sample

Given the sample counts n1, n2, n3 of children from one-, two-, and three-child families, compute the estimated proportions: p1_hat = (n1/100) / (n1/100 + 2*n2/100 + 3*n3/100) * (1/1)? Actually, solve the system: n1/100 = p1 / (p1+2p2+3p3), n2/100 = 2p2 / (p1+2p2+3p3), n3/100 = 3p3 / (p1+2p2+3p3). Then p1_hat = (n1/100) / (n1/100 + n2/100 + n3/100) * (1/1)? Wait, the denominator is the average family size. More precisely, let S = n1/100 + 2*(n2/100) + 3*(n3/100)? No, the observed proportions of children from size k families are qk = nk/100. Then q1 = p1 / (p1+2p2+3p3), q2 = 2p2 / (p1+2p2+3p3), q3 = 3p3 / (p1+2p2+3p3). So p1 = q1 * (p1+2p2+3p3), p2 = q2/2 * (p1+2p2+3p3), p3 = q3/3 * (p1+2p2+3p3). Since p1+p2+p3=1, we have (p1+2p2+3p3) = 1 + p2 + 2p3. But easier: Let D = p1+2p2+3p3. Then p1 = q1 D, p2 = q2 D /2, p3 = q3 D /3. Sum: D(q1 + q2/2 + q3/3) = 1 => D = 1 / (q1 + q2/2 + q3/3). Then p1 = q1 / (q1 + q2/2 + q3/3), p2 = (q2/2) / (q1 + q2/2 + q3/3), p3 = (q3/3) / (q1 + q2/2 + q3/3).

4. Construct 95% confidence intervals

Use the delta method to approximate the variance of the estimated proportions, or use bootstrap resampling of children to obtain empirical confidence intervals. For the delta method, compute the gradient of the estimator with respect to the observed proportions and use the multinomial covariance matrix.

5. Interpret and validate

Check that the estimated proportions sum to 1 and lie in [0,1]. Discuss potential biases if the sampling scheme is not truly random or if family sizes are correlated with other factors.

Key Points to Mention

  • Size-biased sampling: children from larger families are overrepresented.
  • The estimator for family proportions is a weighted average with weights inversely proportional to family size.
  • Use of the delta method or bootstrap for confidence intervals due to the non-linear transformation.
  • Assumption of independent families and random sampling of children.
  • The sum of estimated proportions should equal 1.
  • Potential need to account for finite population correction if sampling without replacement from a small population.

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

Q2

Why does regressing Y on X give a different slope coefficient than regressing X on Y, and how does this relate to causal direction?

Technical Trade-offsRoot Cause Analysis
Author's notes

Knew the mechanical answer (OLS minimizes vertical vs horizontal residuals) but the causal direction part is where it gets interesting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining the mathematical asymmetry: OLS minimizes vertical errors, so swapping X and Y changes the objective and yields different slopes. Then connect this to causality: the regression slope reflects the direction of conditioning, and only a correctly specified causal model (e.g., via instrumental variables or experiments) can identify the true causal effect. Emphasize that regression alone cannot determine causal direction; it only quantifies association under a given conditioning set.

Pro tip: Mention that the product of the two slopes equals the squared correlation (r²), which shows they are not reciprocals and highlights the asymmetry. Also note that in a true causal model with no confounding, regressing Y on X gives the causal effect, while regressing X on Y gives the inverse of the causal effect only if the relationship is deterministic and invertible—otherwise it's biased.

1. Define the regression setup

Clarify that OLS minimizes the sum of squared residuals in the direction specified (Y on X vs X on Y). This difference in loss function leads to different slope estimates.

2. Show the mathematical asymmetry

Explain that the slope for Y on X is Cov(X,Y)/Var(X), while for X on Y it is Cov(X,Y)/Var(Y). Since variances differ, the slopes differ, and their product equals r².

3. Connect to causal direction

Discuss that regression captures association, not causation. The slope's interpretation depends on which variable is treated as the outcome and which as the predictor, reflecting an assumed causal direction.

4. Highlight the role of confounding and identification

Explain that without controlling for confounders or using causal inference methods (e.g., instrumental variables, experiments), neither regression reveals the true causal effect. The direction of regression should be guided by causal assumptions.

5. Summarize implications for practice

Conclude that data scientists must be explicit about causal assumptions and choose regression direction based on the causal question, not just statistical fit.

Key Points to Mention

  • OLS minimizes vertical residuals, so swapping variables changes the objective function.
  • Slope(Y~X) = Cov(X,Y)/Var(X) and Slope(X~Y) = Cov(X,Y)/Var(Y); product = r².
  • Regression quantifies association, not causation; causal direction must be assumed or identified externally.
  • Confounding can bias both regressions; causal inference methods are needed for true effects.
  • In a simple linear causal model with no confounding, Y~X estimates the causal effect, while X~Y estimates the inverse effect only under strict conditions.
  • Always align regression direction with the causal question and validate assumptions.

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

Q3

A regression model has no statistically significant coefficients but still shows strong predictive performance. What's going on statistically, and how would you address it?

Technical Trade-offsRoot Cause Analysis
Author's notes

Classic multicollinearity setup.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that this is a common scenario in regression modeling, often due to multicollinearity, high dimensionality, or overfitting. Explain the statistical reasons behind the discrepancy between individual coefficient significance and overall predictive performance, then outline a systematic approach to diagnose and address the issue, focusing on model validation and feature selection.

Pro tip: Emphasize that predictive performance and inferential significance serve different goals; in practice, you might prioritize one over the other depending on the business objective. Mention that at Upstart, where interpretability and regulatory compliance matter, you'd balance both by using techniques like regularization and careful feature engineering.

1. Diagnose the cause

Check for multicollinearity using VIF, examine the number of predictors relative to sample size, and assess potential overfitting by comparing training and validation performance.

2. Evaluate model performance

Use cross-validation to confirm that the strong predictive performance is not due to overfitting and holds on unseen data.

3. Consider alternative metrics

Look beyond p-values: assess overall model fit (e.g., R-squared, RMSE) and predictive metrics (e.g., AUC, lift) to understand the model's utility.

4. Apply remedies

If multicollinearity is present, use regularization (ridge, lasso), feature selection, or dimensionality reduction. If overfitting, simplify the model or gather more data.

5. Communicate trade-offs

Explain to stakeholders that the model may be good for prediction but not for inference, and align the modeling approach with the business goal.

Key Points to Mention

  • Multicollinearity can inflate standard errors, making coefficients insignificant even if the overall model predicts well.
  • High dimensionality (p > n) or many irrelevant features can lead to insignificant coefficients but good predictive performance due to overfitting.
  • Regularization techniques (ridge, lasso, elastic net) can improve prediction and handle multicollinearity.
  • Cross-validation is essential to ensure the model generalizes well.
  • The distinction between explanatory modeling (inference) and predictive modeling: different goals require different evaluation criteria.
  • Domain knowledge and business context should guide whether to prioritize interpretability or predictive accuracy.

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