From restat-skills
Builds a robustness suite for REStat manuscripts: tests whether headline estimates survive specification, sample, measurement, identification, and inference alternatives.
How this skill is triggered — by the user, by Claude, or both
Slash command
/restat-skills:restat-robustnessThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
- The main estimate exists but its stability under reasonable alternatives is untested
REStat referees ask whether the headline number is a fact about the world or an artifact of choices. The persuasive paper shows the estimate is stable across the specifications a skeptic would try, and is honest where it is fragile. Because REStat weights measurement, robustness here includes a dimension siblings sometimes skip: robustness to measurement choices (alternative measures, error corrections, construct definitions). Robustness is not a kitchen sink — it is a targeted defense of the specific threats this design invites (route the threat menu via restat-referee-strategy).
| Dimension | What to vary | Pass condition |
|---|---|---|
| Specification | Controls, fixed effects, functional form, sample restrictions | Headline stable in sign and rough magnitude |
| Sample | Subperiods, leave-one-group-out, trimming outliers, alt. universe | No single group/period drives the result |
| Measurement | Alternative measures of outcome/regressor, error corrections, construct defs | Conclusion not an artifact of one measure |
| Identification | Alternative estimators (het-robust DID, alt bandwidth/IV), placebo/falsification | Design-appropriate estimators agree; placebos null |
| Inference | Clustering level, wild-cluster bootstrap (few clusters), randomization inference, multiple-testing correction | SEs valid under the data's dependence; key results survive MHT |
restat-referee-strategy)A health paper estimates the effect of a clinic-opening on infant mortality, using a registry-based mortality rate. The headline is robust to controls, sample, and clustering — but a REStat referee notes the registry under-counts deaths in remote areas, and under-counting is correlated with clinic access (where clinics opened, reporting also improved). This is non-classical measurement error that could create the result. The robustness answer is not another control set: it is an alternative outcome (survey-based mortality from an independent source) plus a bounding exercise under plausible mis-reporting rates. The effect survives the survey measure and the bounds exclude zero — a measurement-robustness defense siblings often skip but REStat expects. This is the dimension that most often separates a REStat accept from a revise.
【Headline estimate】[point + SE], identified by [design]
【Specification】stable across: [controls/FE/form] → [Y/N + range]
【Sample】leave-one-out / subperiod / trimming → [Y/N]
【Measurement】alt measure / error correction → [result]
【Identification】alt estimators agree? placebos null? [Y/N]
【Inference】clustering: [level]; few-cluster: [wild bootstrap?]; MHT: [method]
【Honest fragility】[where it weakens + bound] — or "robust throughout"
【Next step】restat-tables-figures
npx claudepluginhub brycewang-stanford/awesome-journal-skills --plugin restat-skillsBuilds robustness suites for AEJ: Applied manuscripts to show headline estimates survive specification, sample, and inference choices.
Runs a robustness battery for EER-style results: tests specification, sample, measurement, inference, and multiple-hypothesis sensitivity. Use when a referee demands disciplined stress tests beyond the author's preferred specification.
Organizes robustness checks for JEEA manuscripts around threats a general-interest referee would raise, ensuring headline results are stable to specification, sample, inference, and assumption perturbations.