Expecon Identification

Experimental Control & Identification (expecon-identification)

When to trigger

• A referee questions whether the treatment effect is causal or an artifact of an uncontrolled difference
• Payoffs may not be incentive compatible — subjects could earn more by misreporting, or stakes are hypothetical/trivial
• Any procedure risks tripping the ESA no-deception norm (false feedback, fake co-players, rigged draws, undisclosed payoff manipulation)
• Randomization, session structure, or the order of treatments could confound the comparison

Identification at ExpEcon = control + a clean contrast

In observational economics, identification is an argument about why selection does not bias the estimate. In the lab, you manufacture identification by design: randomization plus tight control means the only systematic difference across conditions is the manipulated dimension. The job here is to verify that claim holds, on four fronts.

1. The two gates (binary; check first)

• No deception (hard gate). Experimental Economics only considers studies that do not deceive participants (检索于 2026-06；以官网为准). This is the single most common cause of an ExpEcon desk reject. Deception includes: false information about other participants or their choices, fabricated feedback, rigged "random" draws, misrepresenting payoffs or the true purpose in a payoff-relevant way, and confederates posing as subjects. Acceptable practices that are not deception: withholding (not misstating) information, abstract/neutral framing, the strategy method, and not naming the hypothesis. If a design needs deception to work, it cannot be saved by disclosure — redesign it.
• Salient real incentives. Choices must be incentivized with real consequences. Verify the payment mechanism is incentive compatible: BDM/random-lottery for valuations and risk, strategy-method payoffs that match the decision being elicited, one-randomly-paid-round to avoid wealth/hedging effects, and truthful-reporting mechanisms where beliefs are elicited (e.g., a proper scoring rule, ideally binarized/BSR to be robust to risk preferences). State the ECU→money conversion and the realized average payment.

2. Randomization & control

• Random assignment to treatment, and document the unit (individual, group, session). Report balance on observables and on a comprehension measure.
• Hold everything else fixed: identical instructions except the manipulated clause, same interface, same matching protocol, same subject pool and recruitment (e.g., ORSEE/hroot), same physical/online conditions.
• Beware session-level confounds: if a treatment was run in different sessions/cohorts than the control, session is confounded with treatment — randomize within session or run treatments interleaved.

3. Comprehension, order, and learning

• Comprehension checks before play; report pass rates and pre-specify how failures are handled (exclude vs. retain). A treatment difference driven by differential comprehension is not the mechanism.
• Order/learning effects: if within-subject, counterbalance order and test for order effects; if between-subject, justify the loss of power against the gain in clean identification.

4. The estimand and inference unit

• State the estimand in one sentence: the average treatment effect of [manipulation] on [primary outcome], for [population]. Distinguish it from any structural parameter.
• The independent unit is usually the session or matching group, not the individual decision (decisions within a group are not independent). Inference must respect this — see expecon-robustness.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Experimental Economics is lab/field experiments; randomization inference, romano_wolf for many treatments/outcomes, and power are decisive — observational tools secondary.

• 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.