causal-designlisted

Triage a causal question to the identification strategy the data can support, then hand off to the owning method skill. Owns the selection-on-observables branch (overlap, doubly robust estimation, double ML, causal forests, policy learning, sensitivity analysis), plain panel fixed effects, and the inference rules shared across designs (clustering, multiplicity, interference routing). TRIGGER on "identification strategy", "which causal method", "research design", "endogeneity", "quasi-experiment", "natural experiment", "selection on observables", "unconfoundedness", "propensity score", "doubly robust", "double machine learning", "causal forest", "policy learning", "sensitivity analysis", "Oster bounds", "overlap", "panel fixed effects", "strict exogeneity", "surrogate index", "mediation", or "how do I estimate the effect of X on Y" with no design chosen yet. Once a design is named, its method skill owns it.
ericluo04/claude-academic-workflow · ★ 13 · AI & Automation · score 70

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

# Causal design triage The router of the family, grounded in a read canon of four sources (references/canon.md, current as of 2026-08-05): Imbens (2024) supplies the assumption axis (what licenses identification), Li, Luo, and Pattabhiramaiah (2024, hereafter AMA) the marketing data-shape axis (how many treated units, how many pre-periods, how rich the covariates), Feder et al. (2022) the text-role axis (which role unstructured data plays in the graph), and Abadie, Athey, Imbens, and Wooldridge (2023) the clustering rules the family shares. The deliverable is a design recommendation carrying four things: the assumption that licenses it, the estimand it actually identifies WITH its subpopulation named, the handoff to the owning skill, and, for the one branch no method skill owns (selection on observables), estimation code and a methods paragraph. Marketing's framing throughout: randomization is the gold standard, and quasi-experimental work substitutes statistical rigor for design rigor (AMA); a design that fails its gate is a verdict, not an obstacle. Refresh path: run litreview on quasi-experimental methods in marketing since the canon date, then propose additions to references/canon.md as flagged addenda. ## The triage: four questions in order 1. Was assignment randomized, or as good as (lottery, randomized rollout)? Yes: field-experiment. Two cautions at this gate: naive sample means from adaptive/bandit experiments are biased (the arm that looked worse early is