跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

panel-diagnostics

Diagnostics for panel designs without a discrete treatment event: unit and time dependence, serial-correlation-robust inference, fixed-effect structure sensitivity, balance and attrition, and group-level leave-outs. Invoke from `code-and-data` for panel variation that is not an adoption-event design.

isitcredible · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add isitcredible/research_skills --skill panel-diagnostics

技能说明


name: panel-diagnostics description: >- Diagnostics for panel designs without a discrete treatment event: unit and time dependence, serial-correlation-robust inference, fixed-effect structure sensitivity, balance and attrition, and group-level leave-outs. Invoke from code-and-data for panel variation that is not an adoption-event design.

Panel diagnostics

Establish what variation actually identifies the coefficient once the fixed effects have taken their share, whether the error structure supports the reported precision, and which units supply the identification. A panel regression borrows its credibility from the fixed effects, and the fixed effects decide what variation is left — the estimate is only as good as the answer to "what varies, for whom, after all this is absorbed?"

Trigger: panel variation among the target results with no discrete treatment event — otherwise use did-event-study-diagnostics.

Method crib (source-audited): ../../references/methods/panel_methods_guide.md — multiway/nested clustering geometry, D-K, Pesaran CD→CCE/interactive-FE routing, Mundlak margin contrast, Nickell bias and AB diagnostics, G-H attenuation, Lee bounds. ../../references/methods/abadie_athey_imbens_wooldridge_2023_clustering.md — the whether-to-cluster frame this guide's how presumes: clustering is a design problem (two-stage sampling or cluster-level assignment), not a residual-correlation test. When the panel holds nearly all clusters (all states, all years), conventional clustered SEs are inflated by the missing finite-population correction only if effects are heterogeneous across clusters and assignment varies within them; with cluster-level assignment, cluster at the assignment level.

Core checks

  • Unit and time dependence — located, then remedied. Identify how errors correlate within units over time and across units within periods; cluster at the level the dependence lives, and report the cluster count. Then remedy by location: within-unit-over-time → unit clustering plus a persistence measurement (below); across-units-within-period → a Pesaran CD diagnostic first, then two-way clustering (Cameron–Gelbach–Miller) or Driscoll–Kraay (needs a long T, ≥ ~20–30) for common shocks uncorrelated with the regressor. If exposure to a common factor plausibly correlates with the regressor, the dependence biases the coefficient, not just the SE — run a CCE (Pesaran) or interactive-FE (Bai; needs both N and T large) specification and report the movement as a point-estimate finding. A single vce(cluster unit) answers only the first location.
  • Collapse test on the point estimate. Clustering repairs the standard error, not the coefficient. Where a regressor is constant within a coarser unit than its rows (an aggregate variable on micro units, coarse-resolution coding), distinguish the two cases before calling anything wrong: genuine rows with a coarse regressor — the estimand-preserving collapse (original weights/controls retained) typically reproduces the coefficient exactly, and equal-weight re-estimation is a named sensitivity with a different estimand, not a correction (but a large gap undercuts a common-effect claim, and with high residual intraclass correlation the equal-weight fit is the more precise one); versus donor-substituted/imputed copies, where drop-to-source IS the correction. See ../../references/design-matrix.md, "Collapse-test record".
  • Serial-correlation-robust inference. Measure the persistence first — residual and regressor AR within units — then report inference under at least one method whose validity does not ride on low persistence (block/ time-clustered bootstrap, or the time-aggregated / long-differences specification) beside the clustered baseline. If the corrected interval loses significance, route through the MDE machinery rather than reporting "no effect"; if no validated correction exists at this persistence, report "uninformative at this T/ρ," not a discounted pass.
  • Dynamics (lagged dependent variable). A within estimator with a lagged outcome carries Nickell bias — O(1/T), downward on the persistence coefficient, whatever N is. At short T, report it beside a difference- or system-GMM estimate whose output carries m1/m2, Sargan/Hansen and the (collapsed) instrument count; flag system GMM when the persistence coefficient is near 1.
  • Continuous treatment with no stayers (temperature, rainfall, prices: every unit's treatment changes each period). There is no untreated comparison, so TWFE imposes a linear, homogeneous effect; identification rests on quasi-stayers, units whose treatment barely moves (de Chaisemartin–D'Haultfœuille–Vazquez-Bare 2024, AEA P&P). Their nonparametric estimator converges at n^(2/5), not √n, and placebo pre-tests are weak; their parametric two-step estimator is √n. Report it beside the TWFE coefficient (in Deschênes–Greenstone's 1997–2002 county data, −0.024, s.e. 0.007, becomes −0.018, s.e. 0.011).
  • Fixed-effect structure sensitivity — adjudicate the margin. Re- estimate across defensible FE structures (unit, time, unit and time, region-by-period where the claim allows it) and report the coefficient path across them. Before reading the path, state which variation margin the claim requires — within-unit change, between-unit levels, or within-region-period — with the reason; estimates on other margins are rebuttals of the paper's framing, not verdicts on the claim. Run the Mundlak / correlated-random-effects between-vs-within contrast: disagreement between the margins is itself a finding (they answer different questions), so name which one the paper's prose invokes. And check the attenuation direction: where the regressor is smooth, interpolated, or constructed, report its within-vs-total variance share and the between estimate beside the within — if the within-variat
阅读完整来源 ↗