restat-robustness

Robustness Suite (restat-robustness)

When to trigger

• The main estimate exists but its stability under reasonable alternatives is untested
• A referee could ask "does this survive [specification / sample / inference] choice?"
• Inference rests on conventional SEs without checking clustering / few-cluster / multiple-testing
• The result might be an artifact of how a variable was measured

The REStat robustness bar

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).

The five robustness dimensions

Building the suite

1. Start from the threats, not the menu. List the 4–6 objections this exact design invites; each gets a robustness exhibit. (restat-referee-strategy)
2. One headline, many checks. Keep a single main estimate; show alternatives orbit it in a robustness table or a coefficient-stability plot.
3. Report, don't bury, fragility. If an estimate weakens under a defensible alternative, say so and bound it — referees trust honest authors.
4. Specification curve where appropriate. For results sensitive to many small choices, a specification curve shows the full distribution rather than cherry-picked rows.
5. Inference last and seriously. Cluster at the assignment level; with few clusters use wild-cluster bootstrap; adjust for multiple outcomes (Romano–Wolf / sharpened q-values).

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. REStat is applied econometrics/empirical micro — the home of careful identification; DiD/IV/RDD with weak-IV-robust CIs.

• 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.

Checklist

• Headline estimate stable across the controls/FE/functional-form a skeptic would try
• Sample robustness: leave-one-out / subperiod / trimming shown; no single group drives it
• Measurement robustness: alternative measure(s) and/or error correction reported
• Alternative design-appropriate estimators agree; placebo/falsification tests null
• Inference: clustering justified; few-cluster fix applied; multiple-testing handled
• Fragility, where it exists, is reported and bounded — not hidden
• Robustness exhibits map to the specific threats this design invites

Anti-patterns

• A robustness "kitchen sink" unconnected to the design's actual threats
• Reporting only the specifications that work; omitting the obvious skeptical one
• Conventional SEs with a handful of clusters (mechanical over-rejection)
• Many outcomes, no multiple-testing correction (referees will recompute)
• Ignoring measurement-robustness — a REStat-specific gap referees catch
• A specification curve presented as decoration without reading off what it implies

Worked vignette: the measurement-robustness check a referee demanded (illustrative)

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.

Output format

【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