field-experimentlisted

Design, analyze, and write up randomized experiments (field experiments, A/B tests, RCTs), covering stratified and clustered designs, randomization inference, covariate adjustment done right, noncompliance, attrition and gated outcomes, treatment-effect heterogeneity, and interference. Produces advice with citations, R estimation and diagnostics code, and a drafted methods paragraph. TRIGGER on "A/B test", "randomized experiment", "RCT", "field experiment", "holdout", "lift test", "randomization inference", "stratified randomization", "cluster randomized", "geo experiment", "power analysis", "MDE", "encouragement design", "noncompliance", "ITT", "attrition", "Lee bounds", "CUPED", "variance reduction", "uplift", "heterogeneous treatment effects", "causal forest" (experimental heterogeneity; observational causal forests belong to causal-design), "interference", "spillover", "SUTVA". Pre-registration documents belong to the preregister skill; design triage across methods belongs to causal-design.
ericluo04/claude-academic-workflow · ★ 10 · AI & Automation · score 69

Install: claude install-skill ericluo04/claude-academic-workflow

# Field experiments An opinionated experimental workflow grounded in a read canon (references/canon.md, current as of 2026-07-28): the Athey-Imbens handbook chapter as the spine (randomization-based inference first), Freedman's logistic-regression critique and Lin's repair for covariate adjustment, Guo-Basse's generalization to nonlinear outcomes, and Lee's bounds for attrition and gated outcomes. The deliverable is the design or analysis decision with the citation that justifies it, the estimation and diagnostics code in R, and a methods paragraph with the limitation stated in first person at the point of the choice. Refresh path: run the litreview skill on the method since the canon date and fold results into references/canon.md as flagged addenda. ## Design first: decisions that cannot be fixed ex post - Stratify at design time instead of adjusting at analysis time. Ex ante stratification with equal treatment fractions weakly dominates complete randomization in expected squared error, even in small samples and even when the stratifier is useless; ex post regression adjustment can hurt when covariates are unpredictive. Stratify as finely as possible subject to at least two treated and two control units per stratum. - The family's default is fine stratification with at least two treated and two control per stratum, not pairs: within-pair variances are not estimable (Athey-Imbens) and the pair-level variance is conservative for the sample ATE. Pairing remains