← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Senior

Senior
Apr 2026

Summary

Amazon data science interview with a meaty causal inference question about live sports on Prime Video. One question, but it sprawled across matching, IV, and diff-in-diff so it basically felt like three interviews compressed into one.

Questions Asked (1)

Q1

Amazon is considering adding live sports to Prime Video. Using observational data, estimate the causal effect on Prime membership subscriptions and engagement. Walk through your treatment and control definition, primary outcomes, covariate selection, a matching strategy with balance diagnostics, your assumptions and how you'd defend them, an instrumental variable approach using variation in promotion placement, the full 2SLS setup, and finally a diff-in-diff or event study alternative with staggered rollout.

A/B Testing & ExperimentationProduct Analytics & MetricsData Modeling
Author's notes

This was a monster.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the causal question as the effect of adding live sports to Prime Video on Prime membership subscriptions and engagement. Then systematically walk through each requested component: define treatment and control groups, specify outcomes and covariates, outline a matching strategy with balance checks, discuss assumptions and defenses, present an IV approach using promotion placement with full 2SLS setup, and finally describe a diff-in-diff or event study design leveraging staggered rollout. Emphasize the importance of addressing confounding and selection bias throughout.

Pro tip: Acknowledge that observational data may suffer from selection bias and that the IV and diff-in-diff approaches provide complementary evidence; triangulating results across methods strengthens causal claims. Also, mention that you would pre-register the analysis plan to avoid p-hacking and ensure robustness.

1. Define Treatment, Control, and Outcomes

Clearly define the treatment as the introduction of live sports content on Prime Video, and control as regions or time periods without it. Specify primary outcomes: change in Prime membership subscriptions (e.g., new sign-ups, retention) and engagement metrics (e.g., watch time, frequency).

2. Covariate Selection and Matching Strategy

Select covariates that affect both treatment assignment and outcomes, such as regional demographics, prior streaming behavior, and marketing spend. Use propensity score matching or coarsened exact matching to create comparable groups, and assess balance via standardized mean differences and variance ratios.

3. Assumptions and Defenses

State key assumptions: conditional independence (no unmeasured confounding), stable unit treatment value assumption (SUTVA), and positivity. Defend them by arguing that rich covariates capture relevant confounders, and conduct sensitivity analyses (e.g., Rosenbaum bounds) to assess robustness to hidden bias.

4. Instrumental Variable Approach with 2SLS

Use variation in promotion placement (e.g., random featuring of sports content on the Prime Video homepage) as an instrument for treatment. Set up 2SLS: first stage regresses treatment on the instrument and covariates; second stage regresses outcomes on predicted treatment. Discuss exclusion restriction and test for weak instruments (e.g., F-statistic > 10).

5. Diff-in-Diff or Event Study with Staggered Rollout

Leverage staggered rollout across regions or time to implement a difference-in-differences design. Include unit and time fixed effects, and test for parallel pre-trends. Alternatively, conduct an event study to estimate dynamic treatment effects and check for anticipation effects.

Key Points to Mention

  • Propensity score matching and balance diagnostics (standardized mean differences, Love plots)
  • Instrumental variable validity: relevance (first-stage F-stat) and exclusion restriction
  • 2SLS estimation and interpretation of local average treatment effect (LATE)
  • Parallel trends assumption in diff-in-diff and how to test it with pre-period data
  • Staggered adoption and Goodman-Bacon decomposition or Callaway-Sant'Anna estimator to handle heterogeneous treatment effects
  • Sensitivity analysis for unmeasured confounding (e.g., E-value, Rosenbaum bounds)

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