cross-sectional-diagnostics
Diagnostics for cross-sectional observational designs: functional form, common support, influence and leave-region-out, dependence structure including geography, negative controls, and omitted-variable-bias robustness values benchmarked against named covariates. Invoke from `code-and-data` for cross-sectional claims resting on a control set.
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 cross-sectional-diagnostics技能说明
name: cross-sectional-diagnostics
description: >-
Diagnostics for cross-sectional observational designs: functional form,
common support, influence and leave-region-out, dependence structure
including geography, negative controls, and omitted-variable-bias
robustness values benchmarked against named covariates. Invoke from
code-and-data for cross-sectional claims resting on a control set.
Cross-sectional diagnostics
Measure how much the claim depends on choices the reader cannot see — which controls, which functional form, which observations, which error structure — and how much unobserved confounding it could absorb. A cross-sectional regression with controls is the weakest design in the battery, so audit it on the evidence's terms, not the paper's: the control set is itself a claim to be tested, not a given.
Trigger: a cross-sectional observational claim among the target results.
NOT_APPLICABLE for results whose identification another module covers;
OLS columns the claim leans on stay in scope.
Core checks
- Functional form and common support. Re-estimate under defensible
alternative forms; check that treated and untreated observations overlap
in covariate space, and report where the estimate comes from
extrapolation. When support fails, route the answer: re-estimate on the
supported region (trim or re-weight) and report whether the estimate
survives on common support; an estimate carried by extrapolation is
labelled descriptive, not verdict-bearing. Where a claim rests on an
imposed shape (linearity,
monotonicity), test it rather than eyeball a scatter:
binstest(binscatter) formally rejects or clears the parametric form, andnprobust/lpdensitygive a nonparametric conditional-mean or density alternative with valid inference — see the battery's software examples. - Influence and leave-region/group-out. Cross-country and cross-region
regressions are routinely carried by a continent or a handful of
observations; report the full leave-out range. Leverage is a property of
the functional form, not of the observation: reproduce the exact
published column first, then run DFBETA/influence on that spec — the
influential tail can sit at opposite ends under levels versus logs, and a
"leverage" unit whose deletion strengthens the authors is a fragility
story that self-destructs. Check the sign of
b_full − b_dropbefore pitching the finding. - Dependence structure, including geography. Neighbors resemble
neighbors; unit-level robust errors on spatially smooth data fake
precision. The gate has two tiers: for every spatially exposed result
the SE-side record is Conley and Müller–Watson SCPC — a non-smooth
regressor does not exempt it, because spatially correlated errors fake
precision either way. When the regressor is smooth or interpolated over
space, add the coefficient-side and noise rows: a spatial-lag/SLX refit,
the Kelly matched-noise placebo, and Conley–Kelly–Kozbur
pre-whitening — the "full spatial record" is all five rows. Conley only repairs the standard errors; SLX/SAR test
whether the point estimate is spillover-inflated (in gap audits it has
halved the own-unit coefficient while inference held). An SE-only pass
that reports "spatial dependence doesn't overturn it" can hide a
magnitude twice the honest value. In Stata,
scpcaborts withe(V_modelbased) missingwhenever the preceding regression omittedrobustorcluster()(e(V_modelbased)is only stored under a sandwich VCE): never loge(V_modelbased) missingas "package failed / unavailable" — refit the levels regression withrobust/cluster()and runscpc, and alongside it (or if levels SCPC fails on high-dimensional controls) run contrast-space SCPC (C-SCPC, the $h \to \infty$ benchmark) by applyingscpcto the LBM-GLS transformed model fromspurtransform(or in R viaspuR::spur/scpcR::scpc). The audited space–time protocol behind the spatial record is../../references/methods/spurious_space_time_unified_guide.md. - Collapse test on the point estimate. A cross-country or cross-region
regressor is often one coarse value spread across many finer rows (a
regional figure assigned to each country, a donor-substituted or
interpolated value). Clustering the standard error does not touch the
coefficient; collapse to the genuine independent units — average, or
drop-to-source for donor/imputed values — and report whether the estimate
moves. See
../../references/design-matrix.md, "Collapse-test record". - Transparent covariate sensitivity. Report how the estimate moves
across the defensible control sets — not the paper's preferred ladder
only, and with shrinkage distinguished from standard-error inflation when
a control "kills" the result (see
controls-samples-diagnosticswhen horse races or mediators are in play). - Negative controls where substantively justified: an outcome or exposure the mechanism cannot touch, run in the same design.
Conditional checks
- Omitted-variable-bias robustness value — when the causal claim rests
on the control set. Compute the robustness value and the treatment's
partial R², and benchmark against named observed covariates: the multiple
of a named covariate's strength an unobserved confounder would need to
overturn the claim (
sensemakrimplements this; the OVB-sensitivity record in the battery's../../references/design-matrix.mdlists the fields). A robustness value near what a single named covariate already explains is a different finding from one no plausible confounder could reach — report which. The robustness value is NOT a verdict; it quantifies what the claim can absorb, not whether confounding exists.
Report back
Lead with what each check found and what it costs the claim, not a row for every check. Still account for coverage: say what you ran, what was `NO
阅读完整来源 ↗