/did-event-study — DiD / event study, Sant'Anna practitioner standard

This is a thin orchestrator over the canonical packages — it never reimplements an estimator. It walks the practitioner workflow from Difference-in-Differences with Multiple Time Periods (Callaway & Sant'Anna 2021), the Doubly Robust DiD estimators (Sant'Anna & Zhao 2020), and the "What's Trending in DiD?" synthesis (Roth, Sant'Anna, Bilinski & Poe 2023), and it follows the replicate-and-verify-against-source discipline.

Actor → Critic. The skill is the Actor: it runs your packages and the diagnostics. It then puts on the Critic hat for Phase 8 — a graded credibility verdict, never a binary "passes." A mismatch with a pre-test is evidence on credibility, not a gate. (This actor/critic + mandatory-diagnostic + graded-credibility shape mirrors .claude/rules/orchestrator-protocol.md and the verification posture of audit-reproducibility.)

Read first: .claude/rules/did-conventions.md — the HARD standards this skill enforces (data coding, DR default, control group, inference, aggregation, verification, and the pitfalls to avoid). Then the canonical resources in §Resources.

The methodological defaults below reflect Pedro Sant'Anna's sign-off: notyettreated control default, HonestDiD led by relative-magnitudes Mbar (also report M), staggered as an option with att_gt the workhorse, TWFE benchmark-only.

When to use

• Staggered or 2×T adoption with panel or repeated cross-sections; a binary absorbing treatment, or a continuous dose.
• Any time someone reaches for a TWFE event study under staggered timing — route here instead.

When NOT to use

• A single 2×2 with one pre / one post and no covariates is a one-liner — still use DRDID::drdid(), but you don't need the full pipeline.
• Reversible / switching treatments (units turning on and off): these packages assume absorbing treatment. Stop and reconsider the design.