Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› jebo-robustness
jebo-robustness
GitHub针对JEBO论文，按行为威胁（如需求效应、多重比较、机制混淆）组织稳健性检验。覆盖实验、观察及仿真设计，旨在验证结果对规范选择或参数调整的鲁棒性，并区分效应真实性与特定机制，不重做识别或撰写正文。
Trigger Scenarios
实验存在未校正的多重比较或潜在的需求效应
观察性结果对控制变量或估计量敏感
仿真结果可能受参数或种子影响
需要建立从稳健性检查到具体行为威胁的映射
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill jebo-robustness -g -y
SKILL.md
Frontmatter
{
"name": "jebo-robustness",
"description": "Use when a Journal of Economic Behavior & Organization (JEBO) result may be fragile to demand effects, multiple comparisons, specification, or tuning. Organizes robustness by the behavioral threat each check addresses — for experiments, observational designs, and simulations; it does not redesign the identification or write the prose."
}
Robustness Strategy (jebo-robustness)
When to trigger
- An experiment has several treatment arms / outcomes and you have not corrected for multiplicity
- A referee could attribute the effect to experimenter demand, confusion, or order effects
- An observational result moves with controls, sample windows, or estimator choice
- An agent-based result might be an artifact of grid, seed, or tuning choices
- You have a long, unstructured "robustness" appendix and no map from check to threat
Organize robustness by behavioral threat, not by appendix list
At JEBO the right question is never "did we run enough checks?" but "for each way this could not be the behavioral mechanism, did we show it survives?" Build a threat → check map. The threats differ sharply across the four archetypes.
Experiments (lab/online/field)
| Threat | Check JEBO referees expect |
|---|---|
| Experimenter demand | demand-treatment bounds (de Quidt-style), neutral framing, obfuscated objective |
| Multiple comparisons | pre-registered primary outcome; MHT correction (Romano–Wolf, List–Shaikh–Xu, BH) across arms/outcomes |
| Comprehension / confusion | results hold among subjects passing comprehension checks |
| Order / sequence effects | randomize order; show within-order stability |
| Subject pool / platform | replicate across pools (student vs. Prolific vs. field); attention screens on online samples |
| Bots / inattentive online subjects | attention checks, completion-time filters, duplicate-IP screening |
| Selection / attrition | balance among completers; Lee bounds if differential |
Observational behavioral empirics
- Specification curve / multiverse over reasonable controls and windows; show the headline is not a knife-edge.
- Inference robustness: clustering level, wild-cluster bootstrap with few clusters, randomization inference where natural.
- Placebo / falsification: effect absent where the mechanism predicts none; pre-trend tests for DID.
- Sensitivity to unobservables (Oster δ; Rambachan–Roth honest-DID for parallel-trend violations).
Simulation / agent-based
- Parameter sweeps over the behavioral-rule space; report the region where the result holds.
- Seed sensitivity (many runs, report distribution not one path); grid/step-size invariance.
- Sensitivity of emergent regularities to the behavioral rule chosen (e.g., reinforcement vs. EWA learning).
Distinguish "the effect is real" from "the mechanism is the claimed one"
JEBO's distinctive robustness demand is mechanism robustness: even a real, replicable effect can be driven by a different behavioral channel than claimed. Where possible, add a check that separates your mechanism from the leading alternative (a moderation test the rival channel does not predict, a mediation analysis with the caveats stated, or a treatment that shuts the rival channel off).
Execution bridge (StatsPAI / Stata MCP)
Run the battery, don't just enumerate it. Full map:
execution-with-mcp. JEBO spans behavioral/experimental and applied micro; randomization inference for experiments, DiD/IV for observational claims.
- Many outcomes / specifications:
romano_wolf(step-down FWER) orbenjamini_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 exactsuggest_functionfor each. - Exhibits:
etable/did_summary_to_latexfrom the handle — no retyped numbers.
Decisive checks in the body, exhaustive battery in the appendix. JF execution walkthrough.
Checklist
- Every robustness exhibit is labeled with the specific threat it neutralizes
- Experiments: demand effects bounded; primary outcome pre-registered; MHT correction reported
- Comprehension/order/subject-pool/attention threats addressed for the relevant design
- Observational: spec-curve + inference robustness + placebo/falsification + unobservables sensitivity
- Simulation: parameter sweeps + seed/grid sensitivity reported
- At least one check separates the claimed mechanism from the leading alternative
- The headline magnitude is reported with honest uncertainty across specifications
Anti-patterns
- A robustness appendix that lists 20 regressions without saying what threat each rebuts
- Reporting only the cell that survives MHT, omitting the corrected p-values across all arms
- Treating "the effect replicates" as proof the mechanism is the claimed one
- Online experiments with no attention/bot screening
- An agent-based headline shown for a single seed and a single grid
- Hand-picked control sets that quietly maximize the coefficient
Output format
【Archetype】experiment / observational / simulation
【Threat → check map】
- <threat 1> → <check>
- <threat 2> → <check>
【Multiplicity】primary outcome pre-registered? MHT method:
【Mechanism vs. alternative】<test separating claimed channel from rival>
【Headline stability】<range of estimate across specs + uncertainty>
【Next step】jebo-tables-figures
Version History
- 1839142 Current 2026-07-05 13:32
