Jebo Identification

Identification Strategy (jebo-identification)

When to trigger

• A lab/field experiment's treatment, incentives, or comprehension are not pinned down
• An effect could be an experimenter-demand effect rather than the claimed behavioral mechanism
• An observational behavioral claim rests on OLS + controls, or TWFE on staggered timing
• A study uses deception, or runs many treatments, and the inference/ethics implications are unaddressed
• You are unsure the design isolates the behavioral mechanism, not a confound

The JEBO identification bar

JEBO judges identification through a behavioral-mechanism lens: the design must isolate the psychological or institutional channel the paper claims, not merely produce a significant difference. Because JEBO treats experimental design as a first-class identification branch alongside observational causal designs, "identification" means different things by branch — pick the branch and make the channel transparent. Inference must match the design (clustering at the level of randomization or assignment; few-cluster corrections).

Branch A: Lab / online experiment (the JEBO core)

• Incentive compatibility: payoffs must make truthful/effortful behavior the dominant strategy for the elicited object (e.g., BDM, strategy method, incentivized beliefs). State the mechanism and the stakes.
• No-deception norm: the experimental-economics convention is no deception; if you deviate, justify it and expect scrutiny — many referees treat deception as disqualifying for an incentivized study.
• Comprehension & attention checks: report comprehension quizzes, control questions, and how failures were handled (drop / re-instruct), so the effect is not confusion.
• Experimenter demand: rule out demand effects — neutral framing, between-subject where within-subject would cue the hypothesis, obfuscated objectives, or an explicit demand-treatment (e.g., Mummolo–Peterson / de Quidt-style bounds).
• Randomization & balance: show balance on observables; report the randomization procedure and unit.
• Multiple treatments: if several treatment arms, plan the comparisons and correct for multiplicity (see jebo-robustness).
• Pre-registration: pre-register the design and primary analysis (AEA RCT Registry / AsPredicted / OSF) and report deviations. (JEBO does not currently mandate it — 待核实 — but referees increasingly expect it.)

Branch B: Field experiment

• ITT vs. LATE/TOT stated; randomization unit and stratification described; spillovers and SUTVA addressed.
• Attrition examined and bounded (Lee bounds if differential); compliance documented.
• Ethical clearance / consent noted; external-validity scope stated (the field setting's generality).

Branch C: Observational behavioral empirics

• DID / event study with staggered adoption: move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); show clean event-study leads; Goodman-Bacon decomposition.
• IV: strong first stage (effective F); weak-IV-robust sets (Anderson–Rubin) when needed; defend exclusion in institutions/theory + falsification.
• RDD: density/manipulation test (McCrary / Cattaneo–Jansson–Ma); local-linear, data-driven bandwidth, bias-corrected robust CIs.
• The behavioral interpretation must be argued, not assumed — the design identifies an estimate; the mechanism connecting it to a behavioral channel needs its own evidence (see jebo-theory-model).

Branch D: Agent-based / simulation

• Document the data-generating process and behavioral rules; set and report seeds; show the result is not an artifact of grid/tuning choices (see jebo-robustness).

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JEBO spans behavioral/experimental and applied micro; randomization inference for experiments, DiD/IV for observational claims.

• detect_design → recommend → fit with as_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_wolf for many-outcome control.
• Sensitivity: oster_delta / sensemakr for 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 design-to-mechanism mapping stated in one sentence
• Experiment: incentive-compatible elicitation; no deception (or justified); comprehension checks reported
• Experimenter-demand effects ruled out or bounded
• Randomization unit, balance, and (for field) attrition/spillovers handled
• Observational: design-appropriate diagnostics; modern estimator where TWFE/2SLS would bias
• Inference clustered at the randomization/assignment level; few-cluster issue addressed
• The behavioral claim never exceeds what the design isolates

Anti-patterns

• Reporting a treatment difference as a behavioral mechanism without ruling out demand effects
• Using deception in an incentivized study without justification (referees may reject outright)
• Within-subject designs that cue the hypothesis, presented as if between-subject-clean
• TWFE on staggered treatment with no heterogeneity-bias discussion
• Calling a significant coefficient "evidence of [bias]" when a non-behavioral confound survives
• Significance asterisks standing in for clustered SEs or a pre-registered primary outcome

Worked vignette (illustrative)

A lab study claims that public visibility raises cooperation via image concerns. A referee asks whether subjects merely inferred the experimenter wanted more cooperation in the visible arm. The JEBO fix: add a demand-effect treatment (explicitly tell one cell "we expect more cooperation"), show the visibility effect (illustrative: +0.6 contributions, s.e. 0.2) is an order of magnitude larger than the pure demand response, and pre-register cooperation as the primary outcome — turning "could be demand" into a bounded, mechanism-level claim.

Output format

【Branch】lab / field / observational / simulation
【Design-to-mechanism mapping】one sentence
【Behavioral channel isolated】<image concern / loss aversion / learning / norm / ...>
【Identification evidence】[incentives+comprehension+demand-bound / balance+attrition / pre-trends+density+first-stage / DGP+seeds]
【Estimator + inference】estimator; clustering level; weak-IV/honest-DID if any
【What it does NOT identify】[...]
【Next step】jebo-theory-model