Robustness Strategy (jru-robustness)

When to trigger

• The headline risk/ambiguity parameter shifts under a different functional form (CRRA vs. CARA vs. expo-power) and you are unsure which to report
• An experimental result might be an artifact of stakes, order, the random-incentive system, or a particular elicitation device
• A referee will ask whether the finding survives EU vs. non-EU specifications, or pooled vs. heterogeneous-type estimation
• Multiple hypotheses are tested across many lottery menus or treatments and no correction is in place

Organize robustness by threat, not by appendix

A JRU robustness section earns its place when every check is tied to a specific threat to the parameter's interpretation. List the threats first, then the check that answers each.

For VSL / insurance empirics, add: alternative risk measures, sample-selection probes, and sensitivity to the publication-selection / meta-analytic benchmark.

Sequencing

1. Rank threats by how badly each would damage the headline claim if true.
2. Run the check that kills the most dangerous threat first; if the result dies there, stop and rethink before polishing anything.
3. Report robustness as "the sign and rough magnitude of [parameter] is stable across [family]," not "Table A12 shows similar results."
4. Distinguish checks that the design demands (incentive-frame tests for experiments) from generic ones (alternate clustering).
5. Hand off to jru-tables-figures once the parameter is stable across the threats that matter.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JRU spans decision experiments and applied risk; randomization inference for experiments, DiD/IV for observational claims.

• 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

• Every robustness exhibit names the threat it addresses
• The risk parameter is shown stable across at least two functional forms
• A model that separates utility from probability weighting is among the specifications
• Experiments: incentive-frame, stake, and order effects probed; second elicitation device where feasible
• Heterogeneity addressed (mixture / random coefficients) rather than masked by a representative agent
• Multiple-comparison adjustment applied when many menus/treatments are tested
• Inference clusters at the subject (or assignment) level; few-cluster issue handled

Anti-patterns

• A robustness appendix that is a pile of tables with no map from threat to check
• Reporting only the functional form that gives the cleanest number
• Treating an EU-only robustness suite as sufficient when the claim is about non-EU behavior
• Pooling across heterogeneous subjects and presenting the average as if it were a type
• Mining many lottery menus and reporting the significant ones without correction

Distinguish robustness from a specification search

JRU referees draw a sharp line between probing a result and searching for one. Stay on the right side of it:

• Pre-commit the headline specification and present alternatives as deviations from it, not as a menu you chose among.
• Report all the forms you ran, including the ones where the estimate weakened — selective reporting reads as a fishing expedition to a specialist.
• State the decision rule for when the result "survives": e.g., the sign holds and the magnitude stays within a stated band across forms and devices.
• For experiments, distinguish pre-registered confirmatory checks from exploratory ones, and label them as such.

Robustness the experiment specifically demands

Lab and field elicitation papers carry threats that generic econometric robustness misses:

• Comprehension and noise: show the result is not driven by subjects who failed comprehension checks; consider a trembling-hand / Fechner noise term rather than dropping "irrational" subjects.
• Incentive realism: compare real vs. hypothetical, and high vs. low stakes, where the claim depends on it.
• Within-subject consistency: report test-retest or internal consistency for the elicited parameter.

Worked vignette (illustrative)

A paper reports loss aversion λ ≈ 2.1 from a choice-list experiment. The most dangerous threat is that λ is an artifact of the list format (multiple switching, framing). The first check replicates the estimate with a second device (matching probabilities); the second re-estimates under CPT vs. a reference-dependent EU baseline; the third splits by a mixture model to confirm λ is not driven by a confused minority. Only after λ survives all three — with the across-device range reported in full — does the paper present it as the headline in jru-tables-figures.

Output format

【Journal】Journal of Risk and Uncertainty
【Skill】jru-robustness
【Verdict】robust / patch / result fragile
【Top threat】<the check that would most damage the claim>
【Threat→check map】<list>
【Parameter stability】sign+magnitude across <families/devices>
【Heterogeneity】mixture / random coefficients / not addressed
【Source status】verified / 待核实 / not asserted
【Next skill】jru-tables-figures