Amazon Data Analyst Case Interview — Metrics, Experimentation, and Root-Cause Deep Dive

This 60‑minute Amazon-style case mirrors real interviews and emphasizes customer-obsessed, data-driven decision making. You will frame an ambiguous business problem, define the right input/output metrics (as used in WBRs), propose and assess an experiment, and dive deep into root causes using SQL-level reasoning—all while demonstrating Amazon Leadership Principles (Customer Obsession, Dive Deep, Ownership, Bias for Action, Are Right, A Lot). Format: - 5–10 min: Clarify the problem and assumptions; identify the customer and business impact. - 15–20 min: Metrics design and data approach. Define a North Star metric and key input metrics (e.g., detail page conversion, Buy Box win rate, in-stock rate, delivery promise accuracy, ad-attributed sessions). Outline a minimal schema and how you’d compute metrics (high-level SQL/pseudocode acceptable). - 15–20 min: Experimentation/causality. Propose an A/B test or quasi-experiment, choose the assignment unit, define primary/guardrail metrics, discuss MDE/power at a high level, and call out pitfalls (seasonality like Prime Day/Cyber events, novelty effects, selection bias, interference). - 10–15 min: Root-cause analysis and actions. Generate and prioritize hypotheses (pricing, selection gaps, out-of-stock, search relevance, delivery speed, third‑party vs 1P mix, device/app issues). Show how you’d validate with cuts (traffic source, geo, device, cohort) and propose next steps you’d own. Example prompt (one will be given): "EU Prime trial-to-paid conversion fell 80 bps week-over-week. Identify likely causes, quantify impact, and recommend the fastest high-confidence experiment." or "Detail Page View → Purchase rate for ‘Home’ category dropped post-page redesign; determine if the change should be rolled back." What interviewers evaluate: - Problem framing tied to customer impact and inputs vs outputs - Metric selection rigor and WBR-ready definitions - SQL/analytical depth (joins, funnels, cohorts, attribution, data quality checks) - Experiment design and interpretation with guardrails - Communication under ambiguity, structured readout, and Ownership of next steps Expectations: Use STAR where relevant, push for data you need, state trade-offs, and finish with a concise exec-style recommendation and a phased plan (immediate rollback vs iterate, additional logs/instrumentation, dashboard for WBR).