Robustness Suite (jle-robustness)

When to trigger

• The main estimate of a legal/regulatory effect is in hand and you must show it is not an artifact of one specification
• A referee asks "is this robust to alternative controls / sample / which jurisdictions / how you date the rule?"
• The result depends on a bandwidth, a clustering choice, a treatment date, or a sample of included jurisdictions that could be questioned
• You suspect specification-search concerns and want to pre-empt them

The JLE robustness bar

JLE referees — economists who know the institution — probe whether the estimated effect of the rule is stable, honestly inferred, and not the product of researcher degrees of freedom, with special attention to legal-design choices: how you dated the rule, which jurisdictions you treated as controls, whether enforcement was uniform. Robustness here is not a wall of regressions; it is a targeted set of checks each tied to a specific threat to the legal identification. Map every plausible objection to the one check that answers it, and show the point estimate barely moves.

Robustness craft

1. Lock the primary specification first. Everything else perturbs around it; do not present five co-equal specs and let the reader guess the preferred one.
2. One threat → one check. A robustness table should read "here is the worry, here is the evidence it is not a problem."
3. Stress the legal-design choices specifically. Re-dating the rule, swapping the control jurisdictions, and varying enforcement assumptions are the JLE-characteristic checks a referee who knows the institution will demand.
4. Show stability of the point estimate, not just surviving significance.
5. Match inference to the data structure. With few states/jurisdictions, asymptotic clustered SEs over-reject; report a wild-cluster bootstrap or randomization inference — the single most common JLE robustness failure.
6. Be honest about where it weakens. A check that moves the estimate is information; report it and bound the implication.
7. Use placebos on uncovered legal areas. A distinctive and persuasive JLE robustness move is a placebo on an outcome the rule should not affect (an uncapped tort alongside a capped one, an unregulated adjacent market) — a null there isolates the legal channel far more credibly than another control regression.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JLE is empirical law-and-economics — DiD around legal/regulatory change is central.

• 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

• Primary specification declared before perturbations
• Each robustness check mapped to a specific threat, not added for volume
• Treatment-date sensitivity shown (signing vs. effective vs. enforcement)
• Control-jurisdiction sensitivity shown (drop contaminated, leave-one-out, placebo legal area)
• Coefficient-stability evidence (Oster δ or stepwise) for selection on unobservables
• Inference at the legal-variation level + wild-cluster/randomization where jurisdictions are few
• Design-specific sensitivity (honest-DID / RD bandwidth / weak-IV set)
• Stability of the point estimate shown; any check that moves it reported honestly

Separating identification robustness from policy interpretation

Keep the robustness section about whether the estimate of the legal effect is stable, and do not let it drift into re-arguing the rule's normative merits. A referee wants to know the number survives re-dating, control swaps, and correct inference — not your view on whether the rule is good policy. Park the welfare and policy discussion in its own section (see jle-theory-model) so the robustness evidence reads as clean, mechanical stress-testing of the identified effect.

Anti-patterns

• A 20-column robustness table with no map from check to threat ("kitchen-sink robustness")
• Clustering at the firm or case level when the legal variation is at the state level, then claiming robust precision
• Ignoring few-cluster bias with a dozen states and reporting naive clustered SEs
• Hiding the treatment-date choice or the control-jurisdiction choice that breaks the result
• Reporting only that significance survives while the point estimate wanders
• Treating "added more controls and it survived" as sufficient for selection on unobservables

Worked vignette (illustrative)

A DiD estimate that an entry-licensing law raised consumer prices is 6% (s.e. 2). The robustness suite: (i) re-dating from the statute's signing to its effective date shifts the estimate trivially (6.1%); (ii) dropping the three states with simultaneous occupational-licensing reforms leaves it at 5.7%; (iii) a leave-one-state-out sweep stays within [5.2%, 6.4%]; (iv) Oster δ implies selection on unobservables would need to be 2.1× selection on observables to nullify it; (v) with 11 treated states a wild-cluster bootstrap keeps the 95% interval away from zero, whereas naive clustering over-rejects; (vi) a placebo on an unlicensed adjacent service is null. The point estimate barely moves — the JLE target.

The few-clusters problem is the JLE default, not the exception

Most JLE empirical designs exploit variation across a small number of legal units — 50 states, a dozen circuits, a handful of countries, one agency's enforcement regions. With few clusters, conventional clustered standard errors over-reject, so a result that looks significant may not survive correct inference. Treat this as the baseline expectation, not a corner case:

• Report a wild-cluster bootstrap (Cameron–Gelbach–Miller) or randomization/permutation inference as the primary inference when clusters are few, with naive clustered SEs shown only for comparison.
• For a single treated unit or a few, consider synthetic control with placebo-based inference instead of DiD.
• Do not "solve" few clusters by clustering at a finer level (case, firm) that the legal variation does not justify — that manufactures precision the design cannot support.

Referee pushback mapped to the robustness fix

• "You only have 11 states — your standard errors are too small." → Cluster at the state level and report a wild-cluster bootstrap or randomization-inference p-value.
• "Your result depends on when you say the law took effect." → Show estimates under signing, effective, and enforcement dates with a donut around each.
• "Could the control states' own reforms drive this?" → Drop contaminated controls, run leave-one-out, and add a placebo on an unaffected legal area.
• "One treated state can't give you valid inference." → Switch to synthetic control with placebo (in-space) inference rather than asymptotic clustering.

Output format

【Primary spec】declared? [Y/N] — estimate: ___ (s.e. ___)
【Threat → check map】confounders: ___ | date: ___ | controls: ___ | form: ___ | sample: ___ | inference: ___ | design: ___
【Inference】clustering level: ___; few-cluster method: ___
【Design sensitivity】honest-DID / RD bandwidth / weak-IV set: ___
【Estimate stability】range across checks: [___, ___]; checks that move it: ___
【Next step】jle-tables-figures