ecopol-identification

Identification Strategy (ecopol-identification)

When to trigger

• A policy effect rests on OLS + controls, or TWFE on staggered policy rollout
• An IV's exclusion restriction is institutional hand-waving, not an argument
• A structural/quantitative policy model is estimated but it is unclear what in the data pins each parameter
• The counterfactual policy scenario relies on parameters whose policy-invariance is undefended
• You must convince two discussants at once — one who will probe the econometrics, one who needs the design to be legible enough to trust the policy number

The EP identification bar

EP papers are debated by two named discussants and read by policymakers, so identification must be both rigorous and legible. The academic discussant will apply the modern frontier; the policy discussant must be able to follow why the estimate is causal without reading the appendix. The discipline: make the mapping from data/variation to the policy claim explicit in the main text in plain language, and carry the formal defense in a technical appendix (EP's house split — accessible main text, rigorous appendix). Because results feed a policy recommendation, the magnitude — not just the sign or the stars — must be defended; report standard errors and confidence intervals, never lean on significance asterisks for the headline claim.

Branch A: Empirical causal design (most EP papers)

• DID / event study: with staggered policy adoption, move beyond TWFE — Callaway–Sant'Anna, Sun–Abraham, or de Chaisemartin–D'Haultfœuille. Show a clean event-study with pre-trends; a Goodman–Bacon decomposition pre-empts the "negative-weights" discussant.
• IV: strong first stage (report it); with weak instruments use Anderson–Rubin / weak-IV-robust sets. Defend exclusion in three registers — theory, institutions, and a falsification test — because the policy discussant trusts institutional logic.
• RDD: density test (McCrary / Cattaneo–Jansson–Ma), data-driven bandwidth + robustness, covariate smoothness, bias-corrected CIs. State who is at the cutoff and whether they are the policy-relevant population.
• Inference: cluster at the policy-assignment level; address few-cluster problems (wild-cluster bootstrap) — many EP designs have few treated jurisdictions.

Branch B: Structural / quantitative policy model

• Name what identifies each parameter — tie it to a data moment, not "the estimator converged."
• Targeted vs. untargeted moments: report fit to targeted moments and validate against untargeted ones as out-of-sample discipline.
• Counterfactual policy-invariance: the whole point of EP is the counterfactual policy. Argue (Lucas-critique style) that the estimated parameters are invariant to the policy you simulate; if they are not, say so and bound it.
• Numerical credibility: state the objective (MLE/GMM/MSM), multi-start for the global optimum, tolerances; report Monte Carlo recovery of known parameters.

Translate the design for the policy reader

For each design, write the one-sentence plain-language version that goes in the main text: e.g. "Because the reform applied only to firms just above a 50-employee threshold, firms just below serve as a control group — so the difference in their hiring is the reform's effect." The appendix carries the formal estimand and assumptions.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Economic Policy is policy-facing applied economics; foreground a credible design and a policy-relevant magnitude.

• 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

• One plain-language sentence in the main text: what variation identifies the policy effect
• Modern estimator where TWFE would bias (staggered rollout) + clean event-study leads
• IV: first stage reported; exclusion defended in theory + institutions + falsification
• RDD: density + bandwidth robustness + bias-corrected CIs; cutoff population is policy-relevant
• Structural: each parameter tied to a moment; counterfactual policy-invariance argued
• Inference clustered at assignment level; few-cluster correction if needed
• Headline magnitude reported with SE/CI; the policy claim never exceeds what identification supports

Anti-patterns

• TWFE on staggered policy timing with no heterogeneity-bias discussion (the academic discussant pounces)
• An exclusion restriction stated only as "plausibly exogenous" with no institutional or falsification backing
• Running a counterfactual policy scenario on parameters whose invariance is never defended
• Hiding the entire identification argument in the appendix so the policy discussant cannot follow the causal story
• Headlining a policy recommendation off a starred coefficient without reporting the magnitude and its CI

Output format

【Journal】Economic Policy (EP)
【Skill】ecopol-identification
【Branch】empirical causal / structural-quantitative
【Plain-language identification】one sentence for the policy reader
【Frontier diagnostics】[event-study + Bacon / first-stage + AR / density + bandwidth / moments + invariance]
【Inference】SE/CI + clustering level (no asterisk-driven headline)
【What it does NOT identify】[...]
【Next skill】ecopol-theory-model