What problem does it solve? Regression coefficients are often reported as causal effects without a stated identification strategy, producing confident but confounded results. This Skill forces the identification assumptions to be named and tested before estimation, so silent confounding is caught before it reaches a result. ## Core Features & Use Cases - Design Card gate: Requires a signed-off design document stating the estimand, source of variation, untestable assumption, diagnostics, and primary specification before any model is estimated. - Per-design diagnostics: Covers difference-in-differences, event studies, IV, regression discontinuity, matching, panel fixed effects, synthetic control, and ML-based estimators, with the specific testable implications of each (pre-trends, first-stage F, McCrary density test, covariate balance). - Bad-control and robustness discipline: Flags post-treatment controls, colliders, and selection on the outcome, and gates placebo/sensitivity checks behind a user-approved shortlist of roughly three threat-relevant tests. - Use Case: A researcher asks to "run a DiD" on staggered state policy adoptions. The Skill requires a Design Card, warns that vanilla TWFE is biased under staggered timing, and directs estimation to Callaway–Sant'Anna or Sun–Abraham estimators with pre-trend and anticipation diagnostics. ## Quick Start Ask the assistant to estimate the causal effect of a policy on an outcome using difference-in-differences and have it state the identification assumptions and diagnostics before running the regression.
