This is basically six questions stapled together and presented as one.
Start by clarifying the business goal and defining a primary success metric with guardrails, then design the experiment with user-level randomization and justify it. Address network effects by detecting and quantifying spillover, and plan sample size, duration, and stopping rules. Finally, outline instrumentation for ITT and TOT estimates.
Pro tip: At Uber, network effects are critical due to the marketplace; consider using cluster randomization or switchback tests to mitigate spillover, and always pre-register your analysis plan to avoid p-hacking.
Choose a primary success metric (e.g., incremental rides or revenue per user) and guardrail metrics (e.g., cancellation rate, driver utilization, customer satisfaction) to ensure the promotion doesn't harm the ecosystem.
Randomize at the user level to avoid contamination, but justify if cluster randomization (e.g., by city) is needed due to network effects. Define eligibility criteria (e.g., new users, specific geographies) and ensure balanced assignment.
Use techniques like cluster randomization, switchback tests, or spatial separation to detect spillover. Quantify spillover by comparing treatment effects across different exposure levels or using instrumental variables.
Calculate sample size based on minimum detectable effect, power, and significance level. Set duration to capture full behavior cycle, and define stopping rules (e.g., sequential testing, futility) to avoid peeking.
Log user-level exposure, treatment assignment, and outcomes. Use intent-to-treat (ITT) as the primary analysis, and treatment-on-treated (TOT) via instrumental variables or compliance-adjusted analysis to estimate the effect on those who actually used the promotion.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.