Quick Answer

One recent Senior Data Scientist candidate reported an Uber technical phone screen built around an A/B test where driver noncompliance and geographic spillovers complicate causal estimation.
The candidate handled ITT versus LATE and the core 2SLS setup reasonably well, but weakened when pushed on interference, exposure mapping, monotonicity, and spillover sensitivity analysis.
The key lesson is to define the estimand before choosing an estimator: decide whether the question concerns sending the recommendation, changing driver behavior, direct effects, spillover effects, or a combination.
This is one candidate-reported case, not a universal Uber Data Scientist interview format or question set.

Interview Snapshot

Company	Uber
Exact Role Covered	Data Scientist
Level	Senior
Reported Round	Technical Phone Screen
Question Type	Causal inference / experimentation
Case Focus	A/B test with noncompliance and geographic interference
Core Signals	Estimand definition, IV assumptions, SUTVA/interference, sensitivity analysis, marketplace reasoning

The Reported Question

“You ran an A/B test on surge recommendations sent to drivers, but 30% of treated drivers ignored the suggestion and surge in one zone spills over into neighboring zones. How would you estimate the causal effect on completed trips, and what analysis plan would you use to handle both the noncompliance and the interference?”

Candidate Approach: Strong IV Mechanics, Weak Interference Plan

The candidate separated intent-to-treat (ITT) from a local average treatment effect (LATE) and proposed 2SLS with randomized assignment as the instrument. They were comfortable stating relevance, exclusion, and clustered standard errors.

The answer lost structure when interference entered. The candidate recognized that spillovers violate the no-interference component of SUTVA, but admitted, “I had kind of mentally set [the interference piece] aside to deal with later.” They also said the “sensitivity analysis for spillovers I mostly hand-waved.”

Their best self-diagnosis was: “I’d lead with the estimand definition before touching the mechanics.” That would have grounded the rest of the analysis.

Analysis: Define the Decision Before the Estimator

Elena Rodriguez agreed that the estimand should come first. ITT estimates the effect of sending the surge recommendation at scale, including drivers who ignore it. LATE instead targets the effect among drivers whose behavior changes because of assignment, under the IV assumptions.

As Elena put it, “Those answer different questions for a PM deciding whether to invest in making the recommendation more prominent.” The value is connecting the estimand to the product decision, not just naming statistical terms.

For interference, Elena recommended an exposure mapping: define each unit’s exposure using its own assignment plus treatment intensity in nearby zones. This creates a path to separate direct and spillover effects. She also suggested zone-level randomization, while recognizing the power trade-off and that cross-zone spillovers may still require explicit treatment.

On monotonicity, a stronger answer should explain the behavioral assumption. Subgroup compliance diagnostics can be a plausibility check, but monotonicity is generally not directly testable because individual potential compliance behavior is not jointly observed. Once spillovers exist, IV assumptions also need to be stated relative to the exposure structure rather than treated as automatic.

A Stronger Answer Structure

Define the estimands. State whether the business question is the ITT of sending recommendations, a complier effect from changing driver behavior, a direct effect, or a spillover effect.
Report ITT first. Assignment remains randomized even with imperfect uptake.
Address noncompliance with IV. Use assignment as an instrument for actual compliance and discuss relevance, exclusion, monotonicity, and LATE interpretation.
Model interference explicitly. Predefine an exposure mapping, such as own assignment plus treatment intensity in neighboring zones. If design can change, consider cluster or saturation designs.
Stress-test the result. Vary the spillover radius or adjacency rule, compare direct and total effects, use cluster-appropriate uncertainty estimates, and check whether conclusions survive reasonable assumptions.

What This Case May Be Evaluating

This single case may signal senior-level judgment in the following areas:

Signal	What a stronger answer shows
Estimand discipline	Names the causal effect of interest before choosing a method.
IV judgment	Explains noncompliance assumptions and what LATE does and does not identify.
Interference reasoning	Moves beyond binary treatment to direct and spillover exposure.
Sensitivity design	Tests whether conclusions depend on the spillover neighborhood.
Communication	Connects statistical choices to an Uber marketplace decision.

A senior answer should not jump directly to 2SLS. More revealing is whether the candidate knows the estimand, threatened assumptions, and how design choices change interpretation.

FAQ

Does every Uber Data Scientist interview include causal inference?

No. Uber’s public materials make causal inference relevant to Science work, but this is one senior candidate-reported case.

Why report ITT when 30% of treated drivers ignore the recommendation?

ITT answers the operational question: what happens when Uber sends the recommendation under real-world compliance, while preserving the randomized assignment comparison.

Can 2SLS solve both noncompliance and geographic spillovers?

Not by itself. IV methods can address noncompliance under their assumptions; interference requires an explicit exposure model or design strategy.

What should a senior candidate say about SUTVA violations?

Name the violation, define how neighboring treatment changes exposure, identify direct and spillover effects of interest, and propose sensitivity checks around the neighborhood definition.

Uber Senior Data Scientist Interview: Causal Inference, Noncompliance and Spillover Effects