PayPal·Data Scientist·Technical Phone Screen·Senior

Senior

Jun 2026

Summary

PayPal data scientist interview that leaned heavily into causal inference theory. One meaty question that covered a lot of ground, felt more like a stats exam than a typical DS screen.

Questions Asked (1)

Using Uber or Uber Eats data as context: define confounding in causal inference, give a concrete Uber-specific example with a clear treatment, outcome, and confounder, explain the direction of bias, and describe at least two methods to detect or mitigate confounding including what assumptions each requires.

A/B Testing & ExperimentationProduct Analytics & MetricsRoot Cause Analysis

Author's notes

This question has three parts and I kind of fumbled the transition between them.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining confounding in causal inference clearly, then walk through a concrete Uber-specific example with treatment, outcome, and confounder explicitly labeled. Explain the direction of bias and describe at least two methods to detect or mitigate confounding, including the assumptions each requires.

Pro tip: Choose an example where the confounder is not obvious (e.g., driver experience) to show depth, and explicitly state the direction of bias (over- or underestimation) to demonstrate rigor.

1. Define confounding

Explain that a confounder is a variable that affects both the treatment and the outcome, creating a spurious association. Emphasize that it must be associated with treatment and independently affect the outcome.

2. Provide a concrete Uber example

Choose a realistic scenario, e.g., effect of surge pricing (treatment) on rider wait time (outcome), with driver availability (confounder) affecting both. Clearly label each component.

3. Explain the direction of bias

Describe how the confounder biases the estimate. For example, if surge pricing occurs when driver availability is low, and low availability increases wait times, the naive estimate overstates the effect of surge pricing on wait times.

4. Describe detection/mitigation methods

List at least two methods, e.g., randomization (assumption: no unmeasured confounders), stratification (assumption: confounder measured), or instrumental variables (assumption: instrument affects treatment but not outcome except through treatment).

5. State assumptions for each method

For each method, explicitly state the key assumptions required for validity, such as no unmeasured confounding for randomization, or correct model specification for regression adjustment.

Key Points to Mention

Definition of confounding: variable affecting both treatment and outcome
Concrete Uber example with treatment, outcome, and confounder clearly identified
Direction of bias: whether the effect is over- or underestimated
At least two methods to detect or mitigate confounding (e.g., randomization, stratification, regression adjustment, instrumental variables)
Assumptions required for each method (e.g., no unmeasured confounders, correct model specification)
Potential limitations or challenges in real-world Uber data (e.g., unmeasured confounders, dynamic pricing)

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