Uber·Data Scientist·Technical Phone Screen·Senior

Senior

May 2026

Summary

Uber data science interview with a pretty brutal causal inference question about ride-sharing experimentation. The kind of question that separates people who've read about IV from people who've actually used it under pressure.

Questions Asked (1)

A standard A/B test isn't possible for a new ride-sharing feature because of interference between units. Walk through how you'd use instrumental variables to estimate the causal effect on trip volume, including the full set of IV assumptions, at least two concrete instruments, the first- and second-stage regression equations, and how you'd diagnose weak instruments, test overidentification, handle clustering, and assess exclusion restriction violations under marketplace spillovers. Does an effectively unlimited supply environment make the exclusion restriction stronger or weaker?

A/B Testing & ExperimentationTechnical Trade-offsProduct Analytics & Metrics

Author's notes

This one wrecked me a little.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the interference problem and why IV is appropriate, then systematically cover the IV assumptions, propose two concrete instruments (e.g., driver supply shocks and weather events), write out the first- and second-stage equations, and finally discuss diagnostics (weak instruments, overidentification, clustering, exclusion restriction). Conclude by addressing how unlimited supply affects the exclusion restriction, arguing it weakens it because supply-side shocks can directly affect demand or other outcomes.

Pro tip: Emphasize that in marketplace settings, the exclusion restriction is often violated because instruments like driver supply shocks can influence rider behavior through wait times or prices, so you must carefully argue why your instrument affects trip volume only through the endogenous variable (e.g., driver acceptance rate).

1. Frame the problem and IV assumptions

Explain why A/B testing fails due to interference (spillovers between riders/drivers) and introduce IV as a solution. List the four core IV assumptions: relevance, exclusion, independence, and monotonicity (for LATE).

2. Propose concrete instruments

Suggest at least two instruments: (1) exogenous driver supply shocks (e.g., local events causing driver influx) and (2) weather shocks (e.g., sudden rain) that affect driver availability but not rider demand directly. Justify their relevance and plausibility for exclusion.

3. Specify regression equations

Write the first-stage equation: Endogenous variable (e.g., driver acceptance rate) = α + β*Instrument + γ*Controls + ε. Second-stage: Trip volume = δ + θ*Predicted(Endogenous) + λ*Controls + ν. Clarify the endogenous variable and how instruments shift it.

4. Diagnose and test

Discuss diagnostics: weak instruments (first-stage F-stat > 10, or use Cragg-Donald), overidentification (Sargan/Hansen J test if multiple instruments), clustering standard errors at the market or time level, and assess exclusion restriction violations via placebo tests or sensitivity analysis.

5. Address unlimited supply and exclusion

Argue that effectively unlimited supply weakens the exclusion restriction because supply-side instruments may directly affect demand (e.g., shorter wait times increase rider demand) or other outcomes, violating the exclusion restriction. Explain that with unlimited supply, the instrument might not only shift supply but also influence trip volume through channels other than the endogenous variable.

Key Points to Mention

Interference/spillovers in marketplaces (SUTVA violation) and why IV is needed.
The four IV assumptions: relevance, exclusion, independence, monotonicity (LATE).
Concrete instruments: driver supply shocks (e.g., concerts) and weather shocks (e.g., rain).
First-stage and second-stage regression equations with clear notation.
Diagnostics: weak instruments (F-stat), overidentification (Sargan/Hansen), clustering (by market/time), exclusion restriction tests (placebo, sensitivity).
Effect of unlimited supply on exclusion restriction: weakens it because supply shocks can directly affect demand or outcomes.

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