← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Senior

Senior
Jul 2026

Summary

Amazon DS interview focused on a building-operations case about elevator allocation. The whole thing was one big experiment design question with follow-ups that kept adding wrinkles. Pretty involved for a single question session.

Questions Asked (4)

Q1

A building owner wants to reserve one out of six elevators exclusively for food-delivery couriers to reduce resident wait times. Design an A/B test to evaluate this policy and define how you would measure success.

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

The first thing I fumbled was jumping straight to metric definition without naming who the primary customer even was.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the goal: reducing resident wait times without harming overall building efficiency. Then design a randomized controlled experiment at the building level, with treatment buildings reserving one elevator for couriers and control buildings maintaining current operations. Define success metrics that capture both resident wait times and courier efficiency, and consider potential trade-offs.

Pro tip: In real-world experiments, randomization at the building level can introduce confounding due to building-specific factors; consider using a switchback or matched-pair design to control for building characteristics. Also, ensure that the treatment doesn't inadvertently increase wait times for other residents due to reduced elevator availability.

1. Define Hypothesis and Metrics

Clearly state the null and alternative hypotheses. Identify primary metrics (e.g., average resident wait time, 95th percentile wait time) and secondary metrics (e.g., courier delivery time, elevator utilization, resident satisfaction).

2. Design Experiment

Randomly assign buildings to treatment (reserve one elevator for couriers) or control (no reservation). Ensure comparable building characteristics (size, number of residents, delivery volume) through stratification or matching. Determine sample size and duration to achieve sufficient power.

3. Collect and Validate Data

Instrument elevators to log wait times, usage, and courier arrivals. Validate data quality and check for randomization balance. Monitor for novelty effects and ensure compliance with the treatment.

4. Analyze Results

Compare primary and secondary metrics between treatment and control using appropriate statistical tests (e.g., t-tests, Mann-Whitney U). Conduct subgroup analyses (e.g., peak vs. off-peak) and sensitivity checks.

5. Interpret and Decide

Assess practical significance: does the reduction in wait time justify the cost of reserving an elevator? Consider trade-offs and potential negative impacts. Recommend rollout if benefits outweigh costs and no major harm is detected.

Key Points to Mention

  • Randomization unit: building-level randomization to avoid contamination between residents.
  • Primary metric: resident wait time (average and 95th percentile).
  • Secondary metrics: courier delivery time, elevator utilization, resident satisfaction, and overall building efficiency.
  • Sample size calculation: account for intra-building correlation (e.g., using cluster randomized design).
  • Potential confounders: building size, number of residents, delivery frequency, time of day.
  • Trade-offs: reserving an elevator may reduce wait times for residents but could increase wait times for couriers or reduce overall capacity.

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

Q2

If only 8 towers are available and meal-time peaks dominate the signal, would you still run a between-building test or switch to a switchback design? Walk through the trade-offs.

A/B Testing & ExperimentationTechnical Trade-offs
Author's notes

Eight buildings is not a lot for a cluster-randomized test.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the experimental goal and constraints: with only 8 towers and strong meal-time peaks, a between-building test may suffer from insufficient power and confounding. Evaluate whether a switchback design can isolate the treatment effect by alternating conditions within the same towers over time, but account for carryover and time trends. Recommend a switchback if the goal is to measure short-term causal impact and you can randomize at the time level, otherwise consider a hybrid or blocking approach.

Pro tip: Emphasize that the choice depends on the nature of the treatment: if it's a persistent change (e.g., new UI), switchback may be impractical due to carryover; if it's a transient intervention (e.g., pricing), switchback can be powerful. Also mention that with only 8 towers, you might need to use all towers in both conditions via switchback to achieve adequate power.

1. Clarify the experimental goal and constraints

Identify what you're testing, the expected effect size, and the key constraints: only 8 towers, meal-time peaks, and potential interference. Determine if the treatment is persistent or transient.

2. Assess between-building design feasibility

Evaluate if 8 towers provide enough statistical power to detect the effect. Consider that meal-time peaks may create high variance and that between-building randomization may be confounded by tower-level differences.

3. Evaluate switchback design trade-offs

Consider if switchback can control for tower-level confounders by using each tower as its own control. Weigh the risks of carryover effects, time trends, and the need for washout periods, especially with strong meal-time peaks.

4. Consider alternative or hybrid designs

Explore options like a matched-pair design, blocking by tower characteristics, or a stepped-wedge design. Also consider if you can increase sample size by using more time periods or finer granularity.

5. Make a recommendation with caveats

Based on the trade-offs, recommend a design (e.g., switchback if treatment is transient and carryover is manageable) and outline how you would mitigate risks (e.g., randomization at the meal-time level, washout periods).

Key Points to Mention

  • Statistical power and sample size limitations with only 8 towers
  • Confounding and tower-level heterogeneity in between-building tests
  • Carryover effects and time trends in switchback designs
  • Meal-time peaks as a source of variance and potential interaction with treatment
  • Randomization unit and analysis approach (e.g., cluster-robust standard errors)
  • Practical considerations: implementation complexity, cost, and duration

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

Q3

Compliance is only around 60% because many couriers ignore the reserved elevator. How does this affect your estimate, and how would you distinguish between intent-to-treat and treatment-on-the-treated effects?

A/B Testing & ExperimentationAdaptability & Ambiguity
Author's notes

Classic non-adherence problem.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that 60% compliance introduces non-compliance, so the naive estimate is diluted. Explain that ITT estimates the effect of assignment (intent) and TOT estimates the effect of actual treatment, using instrumental variables or CACE analysis to recover the latter. Emphasize that ITT is policy-relevant while TOT is mechanism-relevant, and both should be reported.

Pro tip: Frame the discussion around the business decision: ITT answers 'what happens if we roll out the elevator policy?' while TOT answers 'what is the true effect for those who comply?' This shows you understand both statistical nuance and practical impact.

1. Define the estimands

Clarify that ITT measures the effect of being assigned to the treatment group, regardless of actual behavior, while TOT measures the effect of actually receiving the treatment.

2. Assess impact of non-compliance

Explain that with 60% compliance, the ITT estimate is attenuated by roughly the compliance rate (under monotonicity and no defiers), so the true TOT effect is larger.

3. Choose estimation method

Describe using instrumental variables (e.g., two-stage least squares) or a CACE (Complier Average Causal Effect) framework to estimate TOT, with random assignment as the instrument.

4. Interpret and communicate

Report both ITT and TOT: ITT for policy rollout decisions, TOT for understanding the effect on compliers, and note that TOT may not generalize to non-compliers.

5. Check assumptions

Mention key assumptions: random assignment, exclusion restriction, monotonicity, and no defiers; discuss sensitivity if these are violated.

Key Points to Mention

  • Non-compliance dilutes the ITT estimate; TOT is the effect for those who comply.
  • Instrumental variables or CACE analysis can recover TOT from experimental data.
  • ITT is policy-relevant for rollout decisions; TOT is mechanism-relevant.
  • Assumptions: random assignment, exclusion restriction, monotonicity, no defiers.
  • TOT may not generalize to non-compliers, so external validity is limited.
  • Report both estimates with confidence intervals and discuss practical implications.

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

Q4

Suppose resident wait time improves but courier delivery times get significantly worse, leading to complaints on the delivery app that damage the building's reputation. How do you frame that trade-off and make a recommendation?

Product Analytics & MetricsStakeholder ManagementTechnical Trade-offs
Author's notes

I didn't have a clean answer here and kind of rambled.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge the trade-off as a multi-objective optimization problem, quantify the impact of both metrics on overall customer experience and business goals, and recommend a balanced solution based on data-driven analysis. Frame the recommendation in terms of net impact on key stakeholders and long-term reputation.

Pro tip: Propose a phased approach: test changes on a small scale, measure both metrics, and iterate. This shows you value experimentation and risk mitigation, which is highly regarded at Amazon.

1. Clarify Objectives and Metrics

Define the primary business goal (e.g., customer satisfaction, reputation) and identify how resident wait time and courier delivery time contribute to it. Establish clear metrics and targets for each.

2. Quantify the Trade-off

Use data to measure the magnitude of improvement in resident wait time and degradation in courier delivery time. Assess the impact on overall customer satisfaction and building reputation through correlation analysis or A/B testing.

3. Evaluate Stakeholder Impact

Consider the perspectives of residents, couriers, and building management. Determine which stakeholder group is more critical to the building's reputation and long-term success.

4. Explore Alternatives and Optimize

Investigate solutions that can mitigate the negative impact on courier delivery times without sacrificing resident wait time improvements, such as process changes or technology interventions.

5. Make a Data-Driven Recommendation

Recommend a course of action based on net impact analysis, possibly suggesting a pilot to test the trade-off in a controlled manner before full rollout.

Key Points to Mention

  • Multi-objective optimization and trade-off analysis
  • Quantifying impact using metrics and KPIs
  • Stakeholder analysis and prioritization
  • A/B testing or piloting to validate assumptions
  • Long-term vs short-term consequences
  • Data-driven decision making and communication of trade-offs

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