jeea-robustness

Robustness (jeea-robustness)

When to trigger

• The main estimate moves when a referee imagines a different specification
• Inference rests on assumptions (homoskedasticity, few clusters) you have not stressed
• A theory result might hinge on one functional form or parameter range
• You have a pile of "robustness tables" but no organizing logic

The JEEA robustness bar

A general-interest referee does not want a wall of extra columns — they want to know the headline result survives the threats that matter. Organize robustness by threat to the claim, not by "more specifications." Each robustness exhibit should answer a specific objection a JEEA referee would raise, and the headline magnitude should be stable in sign and roughly in size across the credible perturbations. Report everything as standard errors / confidence sets, never significance asterisks, and ensure every robustness check is reproducible for the JEEA Data Editor check.

Robustness organized by threat

| Threat to the claim | Robustness response |

|---------------------|---------------------|

| Omitted variables / selection | Oster (2019) δ and bounded bias; alternative controls; coefficient-stability discipline |

| Estimator bias (staggered DID) | heterogeneity-robust estimators (CS / SA / BJS / dCDH) as the main result, not an aside |

| Weak / invalid instrument | first-stage strength; Anderson–Rubin / weak-IV-robust sets; over-ID where available |

| Wrong inference | clustering at the right level; wild-cluster bootstrap with few clusters; randomization inference |

| Sample / outliers | winsorize / trim; alternative samples and time windows; influential-observation checks |

| Functional form | alternative specifications; semiparametric or flexible controls |

| Multiple hypotheses | Romano–Wolf / FDR adjustment when many outcomes/subgroups |

| Theory: assumption fragility | relax the driving assumption; show the result degrades smoothly, not knife-edge |

Principles

• One threat, one exhibit, one sentence. Each robustness check maps to a named objection and a one-line conclusion.
• Main result = the credible estimator. Do not bury the heterogeneity-robust DID in an appendix while TWFE leads.
• Stability, not significance-chasing. The point is the magnitude holds, not that it stays starred (and there are no stars at JEEA).
• Pre-empt, don't pad. Cut robustness checks no referee would ask for; add the ones jeea-referee-strategy flags.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map:

[execution-with-mcp](../../../shared-resources/empirical-methods/execution-with-mcp.md). JEEA is a general-interest European economics flagship; credible identification across applied fields.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists missing checks + the exact

suggest_function for each.

• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix.

[JF execution walkthrough](../../../Journal-of-Finance-Skills/resources/worked-examples/02-execution-walkthrough.md).

Checklist

• [ ] Robustness organized by threat, each tied to a specific referee objection
• [ ] Heterogeneity-robust estimator is the main result where staggered timing applies
• [ ] Inference stressed: clustering level justified; wild-cluster bootstrap if few clusters
• [ ] Selection/OVB addressed (Oster δ or design-based defense)
• [ ] Multiple-testing handled when many outcomes/subgroups
• [ ] Theory: driving assumption relaxed; result shown not knife-edge
• [ ] Every check reproducible; no significance asterisks anywhere

Worked vignette (illustrative)

A draft has fifteen robustness columns and a referee still calls the result fragile. The problem is that none of the fifteen addresses the actual threat — selection into treatment. The JEEA-grade fix cuts the padding and adds the three checks the threat demands: an Oster δ showing the bias needed to overturn the result exceeds 1, a placebo on a group the mechanism predicts should show no effect, and wild-cluster inference because there are only twelve treated units. One paragraph, three exhibits, threat neutralized — far more persuasive than fifteen columns of "add another control."

Main-text vs. online-appendix split

JEEA papers carry an online appendix, so use it to keep the main text focused: the headline robustness (the credible estimator, the key placebo, the inference fix) belongs in the main text; the exhaustive grid (every control set, every sample window) belongs in the appendix, referenced in one sentence. Do not bury the result-deciding check in the appendix, and do not pad the main text with checks no referee asked for.

Anti-patterns

• A dozen robustness columns with no statement of which threat each addresses
• TWFE as the headline with the modern estimator hidden in the online appendix
• "Robust to controls" while the real threat (selection, weak IV) is untouched
• Reporting that results "remain significant" instead of that the magnitude is stable
• Robustness checks that quietly change the sample or estimand without flagging it
• The result-deciding robustness check buried in the online appendix

Output format

【Threats addressed】[OVB / estimator / weak-IV / inference / sample / form / MHT / assumption]
【Main result estimator】[the credible one, not TWFE-only]
【Stability verdict】[sign + magnitude hold? Y/N, with the range]
【Inference】[clustering level; wild-cluster/RI if needed]
【Residual fragility】[what still moves the result]
【Next step】jeea-tables-figures