Identification & Economic Interpretation (ecj-identification)
When to trigger
- The empirical core is OLS + controls with no defended causal claim
- Staggered DID estimated with TWFE without addressing heterogeneity-bias critiques
- IV with a weak first stage or a thin exclusion argument
- Structural estimation where the source of parameter identification is not spelled out
- A clean causal effect exists but its economic interpretation, and its general relevance, are not pinned down
The EJ bar: credible identification AND broad economic meaning
EJ accepts both reduced-form and structural work across all fields, but the bar has two parts that must both clear:
- Credible identification — the estimate isolates the causal/structural object you claim, to a standard a demanding referee accepts.
- Economic meaning of broad interest — the estimate maps onto a parameter or margin that economists outside the subfield care about. A precisely identified but parochial effect is a field-journal paper here, because EJ's defining bar is broad relevance.
Reduced-form work should connect to a model or mechanism (see ecj-theory-model); structural work must make its identification transparent. Because EJ runs a reproducibility check via the EJ Data Editor before final acceptance (DCAS-endorsed; deposit to Zenodo — see ecj-replication-package), every identification claim must come from code that actually executes and reproduces. EJ's exposition premium also applies here: the identifying assumption must be stated in plain words a generalist can evaluate, not hidden in notation.
Design priority (strong → acceptable)
The right design is dictated by the economics, not by fashion. As a rough ordering of what travels well at EJ:
- Quasi-experiment (DID, RDD, event study) mapped to a model prediction — reduced form whose coefficient has a stated, broadly interesting economic interpretation.
- Structural estimation tied to a model — when the question is about a deep parameter, welfare, or counterfactuals; identification of parameters argued explicitly.
- Strong IV with a theory-grounded exclusion restriction — first-stage strength plus an economic story for exogeneity and exclusion.
- RCT / lab evidence interpreted through a mechanism, with external-validity discussion.
- OLS with a serious endogeneity discussion — acceptable in theory-empirics or descriptive-with-model papers, not as the sole causal claim.
Branch paths
Branch A — DID / event study
- Staggered timing? Diagnose negative-weighting with Goodman-Bacon; estimate with a heterogeneity-robust estimator (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille, or Borusyak–Jaravel–Spiess).
- Pre-trends: show the event-study plot; do not lean only on a low-power joint pre-trend test — argue economically why pre-trends are flat.
- Map the coefficient to a model object: what does the ATT mean economically, and for whom does it generalize?
- Placebo: randomize treatment timing/units; report the distribution.
Branch B — IV
- First-stage strength: report effective F (Montiel Olea–Pflueger); if weak, use Anderson–Rubin / weak-IV-robust CIs.
- Exclusion: defend in three registers — theory, institutional detail, and a placebo/over-identification check.
- Report the reduced form, not just 2SLS.
- State the LATE interpretation: whose behavior does the instrument move, and is that the population the economics is about?
Branch C — RDD
- McCrary /
rddensity manipulation test.
- Optimal bandwidth (Calonico–Cattaneo–Titiunik) plus ≥3 bandwidth-robustness checks; bias-corrected CIs.
- Covariate smoothness at the cutoff; placebo cutoffs.
Branch D — Structural estimation
- State the model's microfoundations and the moments/variation that identify each parameter (a "what identifies what" paragraph is expected).
- External validation: do estimated parameters match independent evidence or untargeted moments?
- Provide counterfactuals and welfare, and show sensitivity to key assumptions.
Checklist
Anti-patterns
- TWFE on staggered treatment with no discussion of heterogeneity bias
- A precisely identified effect with no statement of what it means, or of why a generalist should care
- IV exclusion asserted ("we argue the instrument is exogenous") without evidence
- Structural estimates with no "what identifies what" discussion — the model becomes a black box
- Clustering at the wrong level to manufacture significance
- An identification claim resting on numbers the deposited code cannot reproduce
Output format
【Design】structural / DID / IV / RDD / event study / other
【Identifying assumption】one plain sentence
【Economic interpretation of the estimate】... (and why it is of broad interest)
【Diagnostics done】[pre-trends, first-stage F, manipulation, param-ID, ...]
【Diagnostics missing】[...]
【Clustering level】... (justification)
【External-validity / GE caveats】...
【Next】ecj-theory-model (if mechanism not yet formalized) or ecj-robustness