Investigate cross-country engagement and ads experiments

You are a Data Scientist in an Ads organization. ## Part A — Engagement differs across two countries You observe that **engagement** is meaningfully different in Country A vs Country B. 1) Define at least **two plausible engagement metrics** (e.g., DAU/WAU, sessions/user/day, time spent, D1/D7 retention, ad interactions) and explain tradeoffs. 2) Outline a structured investigation plan to determine **why** engagement differs, covering: - Data integrity/instrumentation and logging parity - Population mix / selection effects (new vs existing users, device mix, traffic sources) - Seasonality/holidays and product-market differences - Statistical issues (Simpson’s paradox, multiple comparisons) 3) Propose a minimal set of analyses (cuts, models, or decompositions) you would run and what “next actions” different outcomes would imply. ## Part B — Find developers interested in advertising (“pay/call to action”) You want to identify which developers (advertisers) are most likely to be interested in adopting ads tools. 1) Define the target outcome (label) and key funnel stages (e.g., visit → create account → create campaign → spend → retained spender). 2) Propose features/signals you’d use and how you’d avoid leakage. 3) Describe an approach to rank developers (rules vs model), and how you would evaluate it online and offline. ## Part C — Evaluate a new ad format A team launches a **new ad format** and asks you to measure whether it is “good”. 1) Propose: - A **primary success metric** (or a small set) and justification - **Diagnostic metrics** to understand mechanism - **Guardrail metrics** (user experience, long-term value, platform health) 2) Describe how you would design randomization and experiment rollout: - Unit of randomization (user, request, session, geo, advertiser) - Interference/spillovers and how you’d mitigate them - Power/MDE considerations and duration 3) If the experiment shows **no effect**, what would you do next? 4) If it shows **positive impact initially** but the effect disappears later, list plausible reasons and how you would test them. Assume you can query logs, run experiments, and partner with engineering/product to change instrumentation if needed.