jru-identification
GitHub针对JRU稿件，分析风险或不确定性参数的识别策略。涵盖实验激励相容性验证、结构模型识别变异及VSL估计假设，旨在应对审稿人对参数映射、混淆变量（如信念与偏好分离）的质疑，确保数据能清晰识别目标原始参数。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill jru-identification -g -y
SKILL.md
Frontmatter
{
"name": "jru-identification",
"description": "Use when identifying a risk or uncertainty parameter is the bottleneck for a Journal of Risk and Uncertainty (JRU) manuscript — incentive-compatible elicitation in an experiment, or structural\/empirical estimation of risk preferences, VSL, or insurance demand. Stress-tests how the data pin the primitive; it does not invent evidence or citations."
}
Identification Strategy (jru-identification)
When to trigger
- An experiment elicits a risk or ambiguity attitude but the mechanism may not be incentive-compatible (truthful revelation in doubt)
- A choice-list / BDM / matching-probability design is used and a referee questions whether it measures the parameter cleanly
- A structural model is estimated on field data and it is unclear what variation identifies the risk parameter (vs. beliefs, vs. constraints)
- A VSL or insurance-demand estimate rests on regressions whose exclusion or selection assumptions are not defended
The JRU identification bar
At JRU "identification" means the mapping from choices to the risk/uncertainty primitive must be explicit and defended — whether that primitive is elicited in the lab or estimated from the field. Because the journal spans theory, experiment, and empirics, identification splits by branch. The unifying demand: the procedure must reveal the intended parameter and not confound it with utility curvature, beliefs, or constraints.
Branch A: Experimental elicitation of risk / ambiguity preferences
- Incentive compatibility. State the mechanism and why it elicits truthfully: Becker–DeGroot–Marschak, multiple price lists / choice lists, the random-incentive (one-task-paid) system. Address the known threats — BDM is only IC under EU; the random-incentive system assumes isolation; multiple-switching in price lists signals confusion.
- Estimand before estimator. Name what the task is meant to recover — a switching point, a certainty equivalent, a matching probability — and the structural parameter it maps to (curvature, w(p), ambiguity index).
- Design that separates u from w. A single risk-attitude number cannot identify utility curvature and probability weighting jointly; use lottery menus designed to break that confound (e.g., varying probabilities at fixed outcomes).
- Stakes, hypothetical vs. real, order, and house-money effects stated and, where they matter, randomized.
Branch B: Structural / empirical estimation (risk preferences, VSL, insurance)
- Name the identifying variation. For VSL hedonic-wage: the wage–fatality-risk tradeoff, conditional on the compensating-differentials assumptions; defend why risk is not proxying for unobserved job disamenities. For insurance demand: the price/loss variation that moves takeup.
- Beliefs vs. preferences. Field choices reflect both; say how the design separates a risk attitude from a subjective belief (e.g., independent belief elicitation, or variation that moves one but not the other).
- Selection and measurement error in risk exposure addressed; report the estimating equation and the inference (clustered appropriately).
- Estimation regularity for structural models: objective (MLE/GMM/MSM), starting values, and recovery of known parameters in simulation.
The confounds JRU referees probe most
Three confounds recur across both branches; name how the design defeats each:
- Utility curvature vs. probability weighting. A single risk-attitude index cannot separate them; only a design that varies probabilities and outcomes independently can.
- Preferences vs. beliefs. Field and even lab choices reflect subjective probabilities; either elicit beliefs separately or use variation that moves price/cost while holding beliefs fixed.
- Preferences vs. constraints. Low takeup or conservative choices may reflect liquidity, not taste; control for or exploit variation in the constraint.
A clean identification section states, for each of these, whether the design breaks the confound or leaves it open — and the honest "leaves it open" entries belong in the limitations, not hidden.
Execution bridge (StatsPAI / Stata MCP)
Estimate and audit the design, don't only describe it. Full map:
execution-with-mcp. JRU spans decision experiments and applied risk; randomization inference for experiments, DiD/IV for observational claims.
detect_design→recommend→ fit withas_handle=true→audit_result.- Observational causal claims: staggered DiD (
callaway_santanna/sun_abraham+bacon_decomposition+honest_did_from_result); IV (effective_f_test+anderson_rubin_ci); RDD (rdrobust+mccrary_test). - Experiments: randomization-based inference +
romano_wolffor many-outcome control. - Sensitivity:
oster_delta/sensemakrfor observational claims.
Report the magnitude in interpretable units; route the full battery to the appendix. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.
Checklist
- Branch chosen; the choices-to-primitive mapping stated in one sentence
- Experimental: mechanism named and its incentive-compatibility (and its assumptions) defended
- Experimental: design separates utility curvature from probability weighting (or risk from ambiguity)
- Empirical: the identifying variation is named; preferences are separated from beliefs and constraints
- VSL: compensating-differentials / exclusion assumptions stated and probed
- Inference appropriate to the design (clustering at randomization or assignment level)
- The estimated parameter is not asked to carry interpretation the identification does not support
Anti-patterns
- Calling a single risk-attitude index "the" risk preference when curvature and weighting are confounded
- Using BDM or a price list and claiming truthful revelation without noting the EU/isolation assumptions it rests on
- A VSL estimate that ignores selection of workers into risky jobs and the publication-selection debate
- Reading a field choice as a pure preference when it also reflects beliefs or liquidity constraints
- "The estimator converged" presented as if it were identification (structural)
Referee pushback mapped to the identification fix
- "Your 'risk preference' is just utility curvature times probability weighting — you can't tell them apart." → Add lottery menus that vary probabilities at fixed outcomes so w(p) is identified separately from u; report both.
- "BDM is not incentive-compatible outside expected utility." → State the IC assumptions; where the paper studies non-EU agents, use a mechanism whose IC does not presume EU, or bound the bias.
- "Low takeup could be misperceived risk, not a preference." → Elicit subjective probabilities independently; use price variation that moves cost holding beliefs fixed.
- "Your VSL is contaminated by selection into risky jobs." → Probe selection (instrument or panel within-worker variation), and benchmark against the meta-analytic VSL distribution rather than a single estimate.
Worked vignette (illustrative)
A field study infers high risk aversion from low flood-insurance takeup. A referee notes this confounds preferences with beliefs (households may think the risk is near zero) and with constraints (premiums vs. liquidity). The JRU fix elicits subjective loss probabilities separately, then uses exogenous premium variation (say a subsidy lottery) to move price holding beliefs fixed — so the demand elasticity identifies a preference, not a misperception. The reported elasticity (illustrative −0.3) now means what the paper claims it means.
Second vignette: separating curvature from weighting (illustrative)
A lab paper reports a single "risk aversion" coefficient from a Holt–Laury price list. A referee points out the coefficient bundles utility curvature with probability weighting, so it cannot speak to whether the behavior is EU or CPT. The JRU revision adds a menu block that holds outcomes fixed while sweeping probabilities; the resulting certainty equivalents trace an inverse-S w(p) that pins weighting independently of u — turning one confounded number into two interpretable primitives.
Stating what is NOT identified
Every honest identification section closes a door it leaves open. Name explicitly what the design cannot recover — a population parameter beyond the experimental sample, a belief you could not elicit, a margin you could not exogenously vary. JRU referees treat a candid "we identify the preference but not the belief" far more kindly than an overclaim that the data cannot support, and it pre-empts the most damaging review verdict: that the headline number means something other than what the paper says.
Output format
【Journal】Journal of Risk and Uncertainty
【Skill】jru-identification
【Verdict】identified / patch design / re-estimate
【Branch】experimental elicitation / structural-empirical
【Choices-to-primitive mapping】one sentence
【Identification evidence】mechanism+IC / identifying variation + belief separation
【What it does NOT identify】<confounds left open>
【Source status】verified / 待核实 / not asserted
【Next skill】jru-robustness
版本历史
- 1839142 当前 2026-07-05 13:56
