jfe-identification

Identification & Endogeneity (jfe-identification)

When to trigger

• The empirical core is OLS + controls with endogeneity hand-waved away
• A DID uses two-way fixed effects with staggered adoption and you have not addressed the heterogeneous-treatment-effects bias
• Your IV's exclusion restriction or relevance is undefended
• Selection into the sample or treatment is plausible and unaddressed
• A referee could say "your X is endogenous to your Y"

The JFE identification bar

JFE referees expect endogeneity and selection to be treated explicitly, and they expect every plausible alternative explanation to be ruled out — not waved away. Corporate-finance papers are held to a credible-design standard; asset-pricing papers to a disciplined-inference standard (see jfe-empirical-design). This skill covers the corporate-finance causal side; the design/estimator side lives in jfe-empirical-design.

JFE corporate finance descends from Jensen & Meckling (1976), "Theory of the firm: Managerial behavior, agency costs and ownership structure" — the agency-cost foundation and the journal's single most-cited paper. Modern reviewing keeps that demand for an economic mechanism but layers on a hard requirement for credible identification: a correlation between a governance/financing variable and an outcome will not survive review unless the endogeneity is convincingly handled. The best corporate-finance paper each year wins the Jensen Prize.

Design priority (strong -> weaker)

1. Natural experiment / exogenous shock + DID (regulatory change, plausibly random policy, court ruling)
2. Regression discontinuity (a sharp rule with a running variable: index inclusion, covenant threshold, vote share)
3. Instrumental variables (strong first stage + a genuinely defensible exclusion restriction)
4. Matching / entropy balancing + DID (to reduce, not eliminate, selection)
5. Structural estimation (when the question is about deep parameters or counterfactuals)
6. OLS + controls (acceptable only with a frank endogeneity discussion and as a complement, rarely as the headline)

Branches

Branch A — DID / natural experiment

• Is adoption staggered? If so, two-way FE is biased under heterogeneous effects — use a modern estimator (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille, or stacked regression) and report a Goodman-Bacon decomposition.
• Parallel trends: plot the event study; show pre-trends are flat.
• Treatment exogeneity: argue why the shock is unrelated to the outcome's trajectory; address anticipation.
• Placebos: randomize treatment timing/units; falsification on unaffected outcomes.

Branch B — IV

• First-stage strength: report the first-stage F / effective F (Olea–Pflueger); if weak, use weak-IV-robust inference (Anderson–Rubin).
• Exclusion: defend in three registers — theory, institutional detail, and a falsification/placebo.
• Report the reduced form and discuss the LATE/complier interpretation.
• Address whether the instrument itself could be endogenous.

Branch C — RDD

• Manipulation test of the running variable (McCrary / rddensity).
• Optimal bandwidth (Calonico–Cattaneo–Titiunik) plus at least three bandwidth-sensitivity checks.
• Covariate continuity at the threshold; fuzzy-RDD first stage if applicable.

Branch D — selection / sample construction

• State the population and every filter; show how filters could induce selection.
• Heckman / bounds / reweighting where selection is plausible — and say what each assumes.

Branch E — structural

• Make the economic mechanism and identifying assumptions explicit.
• Provide a counterfactual and validate against reduced-form moments where possible.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JFE is finance top-3 (with JF, RFS) — corporate-causal chain for corporate papers, factor-zoo haircut for asset pricing; attribute canon to the correct top-3 journal.

• 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

• The identifying variation is named and its exogeneity argued, not assumed
• Staggered DID uses a heterogeneity-robust estimator + Bacon decomposition
• Parallel-trends / continuity / first-stage-strength evidence is shown
• Placebo and falsification tests are run
• Standard errors are clustered at the level of treatment assignment
• Every alternative explanation a referee would raise has a counter-test
• Anticipation / pre-treatment manipulation is addressed

Anti-patterns

• Two-way FE on staggered adoption with no acknowledgment of the bias literature
• An IV that is "an exogenous event times a lagged endogenous variable" — referees ask why the lag is exogenous
• "We argue the shock is exogenous" with no supporting evidence
• RDD reported at one bandwidth with no sensitivity
• Clustering at the wrong level (e.g., firm when treatment is at the state level)
• Treating endogeneity as a robustness footnote rather than the design's spine

Output format

【Design】natural experiment / RDD / IV / matching+DID / structural / OLS
【Identifying variation】...
【Tests done】[parallel trends, first-stage F, McCrary, placebo, ...]
【Tests missing】[...]
【Cluster level】...
【Alternatives ruled out】[...] | 【Still open】[...]
【Next】jfe-empirical-design