Evaluate impact without randomized experiments

When a randomized experiment is not possible, put forward and defend at least three identification strategies for estimating the causal effect of the promotion, covering assumptions, diagnostic checks, and ways each can fail: (1) difference-in-differences or event-study designs for staggered rollouts, using modern estimators rather than the naive two-way fixed-effects specification, along with parallel-trends checks and placebo tests; (2) synthetic control at the geographic level, with pre-intervention fit measures and sensitivity analyses for the donor pool and regularization choices; (3) instrumental variables that exploit exogenous variation such as eligibility rules or random delivery outages, including relevance and exclusion checks and weak-instrument diagnostics; (4) regression discontinuity when eligibility thresholds are present, for instance tenure of at least 30 days, with sensitivity to bandwidth and functional form and McCrary density tests; (5) propensity-score or doubly robust methods such as IPW/AIPW and causal forests, with overlap diagnostics and covariate balance checks. Explain how you would report ATE/ATT with uncertainty, address interference or network effects, and control for seasonality and concurrent campaigns.

Overview: This question assesses a data scientist’s ability to apply causal inference and identification strategies to estimate promotion effects from observational longitudinal data, with particular attention to staggered rollout, interference or spillovers, seasonality, and quantifying uncertainty.