Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Development

did-event-study-diagnostics

Diagnostics for difference-in-differences and event-study designs: treatment-timing inventory, estimator validity under heterogeneous effects, pre-period paths, anticipation, alternative comparison groups, and terminal-period sensitivity. Invoke from `code-and-data` whenever a target result compares units across treatment timing.

isitcredible · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add isitcredible/research_skills --skill did-event-study-diagnostics

About this skill


name: did-event-study-diagnostics description: >- Diagnostics for difference-in-differences and event-study designs: treatment-timing inventory, estimator validity under heterogeneous effects, pre-period paths, anticipation, alternative comparison groups, and terminal-period sensitivity. Invoke from code-and-data whenever a target result compares units across treatment timing.

DiD and event-study diagnostics

Interrogate the estimator, the comparison group, and the timing structure each on its own terms. A two-way fixed-effects regression on staggered adoption can put negative weights on some treatment effects and average them into a number that represents no group's experience — so "parallel trends look fine" plus a TWFE coefficient is not yet a validated design.

Trigger: any target result identified off treatment timing across units. NOT_APPLICABLE when there is no adoption event or timing variation.

Method crib (source-audited): ../../references/methods/did_event_study_guide.md — estimand decomposition, pre-test power, HonestDiD bounds, Conley–Taber routing. ../../references/methods/schmidheiny_siegloch_2023_event_studies.md — endpoint binning is the identifying restriction when there are no never-treated units; an unbinned saturated event study is underidentified and "we bin at ±k" is an assumption to defend (are effects flat beyond k?), not a footnote. Event-study and distributed-lag parameterizations are numerically identical under correct reparametrization. ../../references/methods/callaway_goodmanbacon_santanna_2024_continuous_did.md — continuous/dose treatments: even with two periods a TWFE dose coefficient suffers from "at least one of three problems: negative weights, selection bias, or non-intuitive weighting" (negative weights in its level-effect decompositions), and the dose-response gradient mixes effect with selection into dose — a dose gradient is not causal evidence by itself. ../../references/methods/olden_moen_2022_triple_difference.md (interim) — DDD needs one parallel-trend assumption, not two, but only because equal biases cancel across the third margin; the cancellation is an assumption to defend, not a free robustness argument.

Core checks

  • Adoption and timing inventory. Which units adopt, when, whether adoption is staggered, whether treatment is absorbing or units switch out (reversible treatment invalidates the standard event-study estimators differently — it routes to dCDH's DID_M, which separates joiners from leavers), and whether any units are always- or never-treated. This inventory decides which estimators are valid; do it first. Count always/never-treated from the estimation sample itself (frac_treated = mean(treat) by unit on the exact regression frame) — not from the paper's text or a prior memo. Units that enter the panel already treated are always-treated; that usually means the coding is an external classification, not a discrete intervention — a substantive finding in itself.
  • Treatment intensity and third-margin variants. If the treatment is continuous or multi-valued (dose, exposure, spending level), the binary toolkit does not carry over: even in two periods the TWFE dose coefficient carries at least one of negative weights (level-effect readings), selection bias (cross-dose readings), or non-intuitive weighting, and comparing ATT across doses mixes effect with selection into dose — route to the CGBS estimators or a defensible discretization, and demand the stronger assumption that handles selection into dose. For a staggered dose, the event-study ATT averaged over doses, or within a dose band, needs only ordinary parallel trends and is estimated with the binary staggered estimators (zero-dose units as never-treated, the sample cut to the band; CGBS 2024 AEA P&P, Thm 1); dose-response and causal-response claims still need the stronger assumption. If the design is triple-difference (group × post × eligibility margin), name the single load-bearing parallel-trend assumption and check the bias-cancellation condition — equal bias across the third margin — rather than accepting "three differences so it's robust"; heterogeneous selection across the margin kills the cancellation. And for any event-study figure without a never-treated group, ask what endpoint-binning normalization identifies the path and whether the plotted coefficients include the binned endpoints or suppress them.
  • Estimator validity under heterogeneous effects. With staggered adoption and effects that vary by cohort or over time, re-estimate with an estimator robust to that heterogeneity and compare it with the TWFE number. First name the claimed estimand and its aggregation — overall ATT, event-time path, or cohort×time grid — because Callaway–Sant'Anna, Sun–Abraham, imputation, and stacked estimators embed different control-group defaults and different aggregations; "the robust number" is not one number. State which identifying assumption the chosen estimator adds (no-anticipation, conditional vs unconditional parallel trends) and which control-group convention it defaults to. When the decomposition shows negative-weighted/forbidden comparisons carrying the estimate and the heterogeneity-robust estimates diverge materially from it (the published coefficient outside the robust estimate's confidence interval, or a change of sign; a change of significance with overlapping point estimates is a precision difference, reported as such), flag the published coefficient validity = invalid, naming the convention in its results.csv row's notes (the forbidden comparisons and the weight they carry): it is a non-estimand — an artifact of the estimator, not "a different estimate" — and may be quoted only as such. A gap between the robust estimand and the published number is a finding about the published estimate.
  • **Event-time coefficie
Read the complete source ↗