← Upstart Interview Insights

Upstart·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
Jun 2026Remote

Summary

Upstart data scientist interview with a heavy quant focus. Two big analytical prompts covering causal inference without an experiment and a three-arm A/B test, plus follow-up threads on guardrails and production forecasting. Dense material for a single session.

Questions Asked (8)

Q1

Describe a product or policy change where you need to measure causal impact but cannot run a randomized experiment. Walk through your treatment definition, population, metric, identification strategy, assumptions, and bias risks.

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

This is the kind of question where you can spiral fast if you don't anchor on the estimand first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Choose a realistic product or policy change where randomization is infeasible, such as a pricing model update or a credit policy adjustment. Structure your answer by clearly defining the treatment, population, and metric, then explain your quasi-experimental identification strategy (e.g., difference-in-differences, propensity score matching, instrumental variables) and explicitly state the assumptions and potential biases. Emphasize how you would test assumptions and mitigate biases to build confidence in causal claims.

Pro tip: Demonstrate awareness that causal inference without randomization requires strong assumptions, and show how you would quantify the sensitivity of your results to violations of those assumptions (e.g., placebo tests, Rosenbaum bounds).

1. Define the treatment, population, and metric

Clearly specify what the treatment is (e.g., a new underwriting model), who is affected (e.g., applicants in certain states), and the primary metric (e.g., loan approval rate or default rate).

2. Choose an identification strategy

Select a quasi-experimental method such as difference-in-differences, regression discontinuity, instrumental variables, or synthetic control, and justify why it fits the context.

3. State key assumptions

Articulate the assumptions required for the chosen method, such as parallel trends, exclusion restriction, or no unmeasured confounding, and explain why they might hold.

4. Identify and address bias risks

Discuss potential biases (e.g., selection bias, omitted variable bias, spillover effects) and how you would test for or mitigate them (e.g., placebo tests, sensitivity analysis, matching).

5. Validate and interpret results

Describe how you would validate the model (e.g., pre-trend tests, robustness checks) and interpret the causal effect in business terms, acknowledging limitations.

Key Points to Mention

  • Difference-in-differences or other quasi-experimental methods
  • Parallel trends assumption and how to test it
  • Propensity score matching or weighting to balance covariates
  • Instrumental variables and the exclusion restriction
  • Sensitivity analysis (e.g., Rosenbaum bounds) for unmeasured confounding
  • Placebo tests or negative controls to check for spurious effects

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

Q2

For your causal design, what placebo test would you run, and what would you do if the parallel trends assumption fails?

A/B Testing & ExperimentationRoot Cause Analysis
Author's notes

Blanked for a second on the placebo framing.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the causal design and the specific placebo test you would run to validate it, such as a pre-treatment placebo or a negative control outcome. Then, explain how you would diagnose and address a parallel trends violation, including robustness checks and alternative methods. Emphasize the importance of transparency and sensitivity analysis in causal inference.

Pro tip: Demonstrate that you understand the trade-offs between different causal methods and that you can communicate uncertainty to stakeholders. Mention that you would pre-register your analysis plan to avoid p-hacking and ensure credibility.

1. Clarify the causal design and assumptions

Briefly restate the design (e.g., difference-in-differences) and the key identifying assumption, such as parallel trends. This shows you understand the foundation of the analysis.

2. Propose a placebo test

Describe a specific placebo test, such as using a pre-treatment period as a placebo or a negative control outcome that should not be affected by the treatment. Explain how it would detect violations.

3. Diagnose parallel trends failure

Outline steps to test for parallel trends, such as visual inspection and formal tests (e.g., event study). If it fails, consider methods like synthetic control, matching, or adding covariates.

4. Address the failure with robustness checks

Discuss sensitivity analyses, such as varying the pre-period or using alternative control groups, to assess how much the results depend on the assumption.

5. Communicate limitations and next steps

Explain how you would communicate the uncertainty to stakeholders and suggest alternative designs or data collection to strengthen causal inference.

Key Points to Mention

  • Placebo tests: pre-treatment placebo, negative control outcomes, and placebo treatments
  • Parallel trends assumption: definition, importance, and how to test it (event study, visual inspection)
  • Alternative methods when parallel trends fails: synthetic control, matching, instrumental variables, regression discontinuity
  • Sensitivity analysis: how robust are results to violations? (e.g., Rambachan & Roth bounds)
  • Communication: transparently reporting limitations and uncertainty to stakeholders
  • Pre-registration and analysis plan to avoid p-hacking and ensure credibility

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

Q3

You have a three-arm experiment with the following data: Variant A has 150 visits and 43 purchases, Variant B has 200 visits and 48 purchases, and Variant C has 100 visits and 15 purchases. Which variant is winning right now? Show your work using confidence intervals or a hypothesis test done by hand.

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

Point estimates are straightforward: A is around 28.7%, B is 24%, C is 15%.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Calculate the conversion rate for each variant and compute confidence intervals (e.g., 95% CI) for each rate using the normal approximation. Compare the intervals to see which variant is statistically distinguishable, and consider both absolute and relative differences to determine the current winner.

Pro tip: Always check the assumptions for the normal approximation (np and n(1-p) > 5) and mention that with small sample sizes, exact methods like Wilson score interval or Fisher's exact test may be more appropriate. Also, emphasize that 'winning right now' is a snapshot and not a final decision without considering practical significance and business impact.

1. Calculate conversion rates

Compute the conversion rate for each variant: A: 43/150 = 28.67%, B: 48/200 = 24.00%, C: 15/100 = 15.00%.

2. Compute standard errors

For each variant, calculate the standard error of the proportion: SE = sqrt(p*(1-p)/n). For A: sqrt(0.2867*0.7133/150) ≈ 0.0369, B: sqrt(0.24*0.76/200) ≈ 0.0302, C: sqrt(0.15*0.85/100) ≈ 0.0357.

3. Construct 95% confidence intervals

Use p ± 1.96*SE to get intervals: A: 28.67% ± 7.23% = (21.44%, 35.90%), B: 24.00% ± 5.92% = (18.08%, 29.92%), C: 15.00% ± 7.00% = (8.00%, 22.00%).

4. Compare intervals and determine winner

Check for overlap. A and B intervals overlap (A lower bound 21.44% < B upper bound 29.92%), so not statistically distinguishable. C's interval is entirely below A's and B's, so C is significantly worse. A has the highest point estimate, so it is currently winning, but not significantly better than B.

5. Discuss statistical significance and next steps

Mention that with these sample sizes, the difference between A and B is not statistically significant. Suggest running a hypothesis test (e.g., two-proportion z-test) for A vs B, and note that more data may be needed to declare a winner.

Key Points to Mention

  • Conversion rate calculation for each variant
  • Standard error and confidence interval formula for proportions
  • Interpretation of overlapping confidence intervals
  • Statistical significance vs practical significance
  • Assumptions of normal approximation and alternatives (e.g., Wilson interval, Fisher's exact test)
  • Sample size and power considerations for conclusive results

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

Q4

Should Variant C be stopped early given its current performance?

A/B Testing & ExperimentationRoadmap Prioritization
Author's notes

Short answer I gave: probably yes, but you need a pre-registered stopping rule to avoid peeking bias.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying what 'current performance' means—whether it's a statistically significant negative effect, a safety issue, or just underperformance relative to expectations. Then evaluate the decision through a framework that balances statistical rigor (e.g., sequential testing, stopping rules) with business impact (e.g., opportunity cost, risk). Conclude with a recommendation that considers both the data and the broader strategic context.

Pro tip: Demonstrate maturity by acknowledging that early stopping decisions are often made under uncertainty and that the cost of continuing a harmful variant may outweigh the cost of a false stop. Mention that you would consult stakeholders and consider pre-registered stopping rules to avoid p-hacking.

1. Clarify the objective and stopping criteria

Ask what the primary metric is, what constitutes 'failure' (e.g., statistically significant negative effect, guardrail metric breach), and whether there are pre-defined stopping rules. This ensures alignment on what 'should be stopped' means.

2. Assess statistical evidence

Check if the observed performance is statistically significant and practically meaningful. Consider sequential testing adjustments, confidence intervals, and whether the result is likely due to chance or a real effect.

3. Evaluate business impact and risk

Quantify the potential harm of continuing (e.g., revenue loss, user churn) versus the cost of stopping early (e.g., missed learning, false negative). Consider opportunity cost and strategic priorities.

4. Consider alternative explanations and data quality

Rule out confounding factors, sample ratio mismatch, or implementation issues. Ensure the data is reliable and the experiment was run correctly before making a decision.

5. Make a recommendation and next steps

Based on the analysis, recommend whether to stop, continue, or modify the variant. Suggest actions like deeper dive, stakeholder discussion, or implementing a holdout to monitor long-term effects.

Key Points to Mention

  • Statistical significance vs. practical significance
  • Pre-registered stopping rules and sequential testing to control error rates
  • Guardrail metrics and safety thresholds
  • Opportunity cost and business impact of continuing
  • Data quality checks (e.g., sample ratio mismatch, novelty effects)
  • Stakeholder alignment and decision-making under uncertainty

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

Q5

How would your recommendation change if you also had data on revenue per visitor, average order value, refund rate, retention, and page latency across the three variants?

A/B Testing & ExperimentationProduct Analytics & MetricsPricing & Monetization
Author's notes

This is where I think I actually did well.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the primary decision metric and how these additional metrics relate to it, then evaluate each variant on the full metric set to identify trade-offs. Finally, recommend a variant based on the net impact on long-term business value, considering both statistical and practical significance.

Pro tip: Always tie metrics back to the company's north star (e.g., revenue or profit) and consider the time horizon—short-term gains might hurt long-term retention. Also, check for metric interactions and segment-level effects before making a final call.

1. Clarify the decision context

Identify the primary goal of the experiment (e.g., maximize revenue, profit, or user experience) and how the new metrics inform that goal.

2. Analyze each metric across variants

Compare the variants on all metrics, noting statistically significant differences and effect sizes. Look for patterns and trade-offs.

3. Assess trade-offs and net impact

Quantify the combined effect on the primary metric (e.g., revenue per visitor) by considering how changes in AOV, refund rate, retention, and latency interact.

4. Check for segment-level and temporal effects

Examine if the impact varies by user segments or over time, and whether short-term gains might harm long-term retention.

5. Formulate recommendation

Recommend the variant that maximizes the primary goal, acknowledging any uncertainties and suggesting further tests if needed.

Key Points to Mention

  • Define a primary decision metric (e.g., revenue per visitor or profit) and use others as guardrails or secondary metrics.
  • Consider the interplay between metrics: e.g., higher AOV might increase refunds, or faster page load might improve retention.
  • Evaluate both statistical significance and practical significance (effect size) for each metric.
  • Look for segment-level heterogeneity to avoid Simpson's paradox.
  • Consider long-term impact: retention and refund rates affect customer lifetime value.
  • Recommend based on net impact, and if trade-offs are unclear, propose a follow-up experiment or deeper analysis.

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

Q6

What would make you choose Variant B over Variant A even though A has a higher conversion rate?

A/B Testing & ExperimentationProduct Strategy
Author's notes

Revenue per visitor or profit margin differences are the obvious answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that conversion rate is a key metric but not the only one; emphasize that the decision should be based on the overall business objective and long-term impact. Discuss how Variant B might be superior on other metrics such as revenue, retention, or risk, and how statistical significance and practical significance both matter.

Pro tip: Show that you think like a business owner, not just a statistician: quantify the trade-offs and consider the cost of a false positive or negative. Mention that sometimes a lower conversion rate can lead to higher lifetime value or lower default rates, which is crucial in lending.

1. Clarify the primary metric and guardrails

Confirm that conversion rate is the primary metric, but identify other key metrics (e.g., revenue, retention, default rate) that serve as guardrails or secondary objectives.

2. Evaluate statistical and practical significance

Check if the difference in conversion rate is statistically significant and if the lift is practically meaningful. Also assess if Variant B shows significant improvements in other metrics.

3. Consider long-term and downstream effects

Think about how each variant impacts long-term customer behavior, such as retention, lifetime value, and risk. A higher conversion rate might bring in lower-quality customers.

4. Align with business strategy and risk tolerance

Determine which variant better supports the company's strategic goals, such as profitability, risk management, or market expansion, and fits within the risk tolerance.

5. Make a data-driven recommendation

Weigh the trade-offs and recommend the variant that maximizes overall business value, even if it means sacrificing some conversion rate.

Key Points to Mention

  • Statistical significance vs. practical significance
  • Secondary metrics and guardrail metrics (e.g., revenue, retention, default rate)
  • Long-term impact and customer lifetime value
  • Business objectives and strategic alignment
  • Risk management and regulatory considerations in lending
  • Cost of false positives/negatives and opportunity cost

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

Q7

If you launch the winning variant, how would you forecast its actual conversion rate in production going forward?

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

Probably the most interesting part of the whole interview.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that the A/B test provides an unbiased estimate of the treatment effect, but production performance may differ due to novelty effects, seasonality, and population shifts. Then outline a forecasting method that adjusts the observed lift for these factors, quantifies uncertainty, and validates with holdout or post-launch data.

Pro tip: Emphasize that you would not simply extrapolate the test conversion rate; instead, you would model the treatment effect and apply it to a baseline forecast, while monitoring for drift and being ready to update the forecast as new data arrives.

1. Estimate the treatment effect from the experiment

Use the A/B test data to compute the lift in conversion rate (absolute and relative) with confidence intervals, ensuring the experiment was valid (e.g., no SRM, sufficient power).

2. Adjust for factors that differ in production

Consider novelty effects, seasonality, changing user mix, and long-term trends; adjust the lift if necessary using holdout data or historical patterns.

3. Forecast baseline conversion rate

Project the baseline (control) conversion rate for the future period using time-series models, seasonality, and business drivers.

4. Combine to forecast production conversion rate

Apply the adjusted lift to the baseline forecast to get the expected production conversion rate, and quantify uncertainty (e.g., via simulation or Bayesian methods).

5. Monitor and update

After launch, track actual conversion rate against forecast, detect deviations, and refine the model as new data accumulates.

Key Points to Mention

  • Confidence intervals and statistical significance of the lift
  • Novelty and primacy effects
  • Seasonality and time trends
  • Population shift (e.g., new user segments)
  • Bayesian or simulation-based uncertainty quantification
  • Holdout validation and post-launch monitoring

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

Q8

How would you adjust for multiple comparisons across the three variants?

A/B Testing & Experimentation
Author's notes

Said Bonferroni is the conservative floor and Benjamini-Hochberg is more practical for this kind of setup.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the multiple comparisons problem and its impact on Type I error. Then, discuss specific correction methods like Bonferroni or Holm-Bonferroni, and consider alternatives like controlling the False Discovery Rate (FDR) with Benjamini-Hochberg. Finally, emphasize the importance of pre-registering the analysis plan and considering the trade-offs between Type I and Type II errors in a business context.

Pro tip: Don't just default to Bonferroni; consider the cost of false negatives and whether FDR control might be more appropriate for your business goals. Also, mention that if the variants are compared to a control, using Dunnett's test can be more powerful than pairwise corrections.

1. Acknowledge the multiple comparisons problem

Explain that with three variants, there are multiple pairwise comparisons (e.g., each variant vs. control), inflating the family-wise error rate (FWER).

2. Choose an appropriate correction method

Discuss options like Bonferroni, Holm-Bonferroni, or Dunnett's test for FWER control, or Benjamini-Hochberg for FDR control. Justify your choice based on the experiment's goals and error trade-offs.

3. Consider the trade-offs

Highlight that corrections reduce power (increase Type II error). In a business context, weigh the cost of false positives vs. false negatives and choose a method that aligns with business objectives.

4. Pre-register and document

Emphasize the importance of pre-registering the analysis plan, including the correction method, to avoid p-hacking and ensure validity.

5. Communicate results clearly

Explain how you would present the results, including adjusted p-values or confidence intervals, and interpret them in the context of the business decision.

Key Points to Mention

  • Family-wise error rate (FWER) and false discovery rate (FDR)
  • Bonferroni correction and its conservativeness
  • Holm-Bonferroni step-down procedure
  • Dunnett's test for comparing multiple treatments to a control
  • Benjamini-Hochberg procedure for FDR control
  • Trade-off between Type I and Type II errors and business impact
  • Pre-registration of analysis plan to avoid p-hacking

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