Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →brycewang-stanford/aejpol-robustness
Use when an AEJ: Economic Policy manuscript's headline policy estimate needs to be shown stable and credible against specification, sample, inference, and identification threats. Organizes the robustness program by threat-to-the-policy-conclusion; it does not design the primary identification or write exhibits.
Passed 23 of 23 test cases with the skill, and 20 of 23 without it.
3 fixed20 passed either way0 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 3 test cases the skill fixed. Same model, same prompt, with and without the skill.
Our study estimates the net poverty reduction per dollar spent following county-level minimum wage adjustments. To ensure the policy conclusion is solid, please lay out a robustness framework detailing: the target policy summary metric, the three most severe identification threats, explicit check-to-result mappings for those threats, our approach to inference stress-testing, how sensitive our results are to chosen calibration parameters, and the next step in our workflow.
| Checked | Without skill | With skill |
|---|---|---|
| Recommends evaluating pre-trend sensitivity using Honest-DID bounds for county-level panel data. | Missed | Passed |
The instructions your model reads when the skill loads.
At AEJ: Policy, robustness is judged by whether the policy takeaway is stable — if the headline estimate is the cost-per-job or the MVPF, show that number is stable, with its uncertainty, not merely that a regression coefficient stays significant. Organize the robustness program around the threats that would change the policy conclusion, and report enough that a skeptical referee can see each threat addressed.
| Threat to the policy conclusion | Check |
|---|---|
| Functional form / controls drive the result | Specification ladder; show the estimate across a coherent set, not a single lucky spec |
| Pre-trends / parallel-trends violation | Honest-DID (Rambachan–Roth) sensitivity bounds; placebo pre-period "effects" |
| Estimator bias under staggered timing | Re-estimate with ≥1 heterogeneity-robust DID estimator (CS / SA / BJS / dCDH) |
| Bandwidth / kernel (RDD) | Bandwidth sweep + bias-corrected CIs; donut-RDD if heaping at the cutoff |
| Weak / invalid instrument | Effective F; AR-robust CI; over-ID test if available |
| Wrong inference / few clusters | Wild-cluster bootstrap; report clustering level sensitivity |
| Multiple outcomes / specifications | Romano–Wolf / sharpened q-values; a specification curve where many specs are run |
| Confounding by an omitted policy/shock | Controls for co-timed policies; event-study around the focal reform only |
| Selection on unobservables | Oster (2019) δ / bounds; argue the implied selection is implausible |
| Sample composition / outliers | Drop influential jurisdictions; winsorize; alternative sample windows |
Run the battery, don't just enumerate it. Full map:
execution-with-mcp. AEJ: Policy evaluates programs and reforms; the design must carry a policy-relevant magnitude, not just statistical significance.
romano_wolf (step-down FWER, accounts for
cross-test correlation) or benjamini_hochberg — report the adjusted threshold.oster_delta / sensemakr — the confounder strength that wouldtest cases passed, vs 20 / 23 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.