Jeg Identification Strategy

Identification & Argument Strategy (jeg-identification-strategy)

When to trigger

• An empirical growth claim rests on a cross-country regression with endogenous regressors
• A theoretical result depends on an assumption you have not justified or tested for tightness
• You are unsure whether your inferential backbone clears a growth-specialist bar

JEG publishes both theory and empirics, so this skill has two tracks. Pick the one matching your paper; quantitative/calibrated papers use both.

Track E — Empirical identification (causal design for growth)

Growth empirics carry a hard endogeneity problem: most candidate determinants (institutions, human capital, finance, openness) are co-determined with income. The bar:

Cross-country / dynamic-panel growth

• If you run growth-on-determinant regressions, confront reverse causality and omitted deep determinants explicitly. A bare OLS or static panel will not convince.
• Dynamic-panel system GMM (Arellano-Bond / Blundell-Bond) is common, but it is a trap if abused: cap and report the instrument count, report the Hansen-J over-identification test and AR(2) serial-correlation test, and show results are not driven by instrument proliferation.
• Convergence claims: distinguish β- from σ-convergence and address Galton's-fallacy / measurement-error critiques.

Clean causal shock (where one exists)

• Where a credibly exogenous shock to a growth determinant exists, use a sharp design: IV (strong first stage, defended exclusion restriction in theory + institutions + falsification), DID/event study (modern estimators, not naive TWFE on staggered timing; pre-trends), or RDD (density and bandwidth diagnostics).
• Few-country / few-cluster inference: use wild-cluster bootstrap or randomization inference; do not lean on asymptotic t-stats with a handful of clusters.
• State the estimand (ATT / LATE / local effect) and its external validity for the growth question.

Track T — Theoretical argument (assumptions, results, generality)

For a theory paper the "identification" object is the logical structure, not a research design.

• Assumptions: list them explicitly; mark which are substantive (drive the result) vs technical (for tractability). Justify each economically and flag knife-edge conditions.
• Results: state propositions/theorems precisely with their hypotheses; give existence, uniqueness, and stability of the relevant steady state or balanced-growth path, and check transversality.
• Proof exposition: put intuition in the text and full proofs in an appendix; make each step auditable. A growth-theory referee will reproduce the algebra.
• Generality: show how far the result reaches — which assumptions can be relaxed, what breaks if you do, and which comparative statics / testable predictions survive. Generality is the contribution's reach.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JEG (growth) uses cross-country and long-run panels with deep endogeneity; foreground identification and robustness to alternatives.

• 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.

Anti-patterns

• (E) System GMM with hundreds of instruments and no Hansen-J / AR(2) reported.
• (E) Naive TWFE on staggered growth-policy timing; OLS cross-country causal claims with no design.
• (T) A "general" theorem that silently depends on a knife-edge parameter restriction.
• (T) Proofs that assert rather than derive existence/uniqueness/stability.
• Either track claiming more than the argument supports.

Persistence-design defenses (Track E extension)

Historical-persistence and deep-determinants papers face a now-standard referee script at this journal; pre-empt all four lines before submission:

• Spatial autocorrelation: report Conley standard errors at several distance cutoffs alongside clustered SEs, and show the headline estimate survives the widest defensible cutoff.
• Spurious spatial fit: run placebo treatments drawn from spatially correlated noise and report where the true coefficient falls in that placebo distribution.
• Overused instruments: if your instrument (terrain, climate, disease ecology, a historical shock) has already served other outcomes in print, defend exclusion against each published channel it explains — not in the abstract.
• Mechanism opacity: a reduced-form persistence coefficient is a starting fact, not an answer; bring intermediate-period outcomes or a decomposition that traces how the past reaches the present.

A persistence paper that clears only the first two is an economic-history note; clearing all four is what makes it a growth paper.

Output format

【Track】E (empirical) / T (theory) / both
【E: design】GMM-panel / IV / DID / RDD + key diagnostics (Hansen-J, AR(2), first-stage F, pre-trends)
【E: inference】clustering / few-country handling; estimand + external validity
【T: assumptions】substantive vs technical; knife-edge flags
【T: results】existence / uniqueness / stability / transversality checked?
【T: generality】what can be relaxed; surviving predictions
【Next skill】jeg-data-analysis

Source: brycewang-stanford/Awesome-Journal-Skills → Journal-of-Economic-Growth-Skills/skills/jeg-identification-strategy/SKILL.md