rfs-identification

Identification Strategy (rfs-identification)

When to trigger

• The empirical core is OLS + controls with an open endogeneity threat
• A DID uses two-way fixed effects (TWFE) without addressing staggered-adoption bias
• An IV has a weak first stage or a contestable exclusion restriction
• A cross-sectional asset-pricing claim rests on a factor that may be data-mined
• Reviewers will ask "what is your source of variation?" and you lack a crisp answer

The RFS identification bar

RFS applies the same high causal-inference standard as JF and JFE: a claim of causality requires a credible source of exogenous variation, not a richer control set. RFS is more receptive than the others to genuinely new questions and to structural / theoretical identification — but novelty never substitutes for a clean design. Pick the strongest feasible strategy below.

RFS-specific lever — Stage 1 design review. Because RFS pioneered Registered Reports (pre-results review; Karolyi, "Kick-Starting the Review Process," RFS 27(2), 2014), an identification strategy can be refereed before the results exist. If you pursue this route, the design must be airtight on paper: pre-specified sample, treatment definition, estimator, diagnostics, and the exact tables to be produced — because that protocol becomes the binding commitment that earns in-principle acceptance. Even for a standard submission, draft the design as if it had to survive Stage 1 review with no results to fall back on.

Design priority (corporate / household / empirical finance)

1. Natural experiment / policy shock + DID (incl. staggered and continuous treatment)
2. Regression discontinuity (a sharp institutional threshold: index inclusion, rating cutoff, covenant)
3. Instrumental variables (strong first stage + a defensible, finance-grounded exclusion)
4. Event study with clean windows and confound discussion
5. Matching / propensity-score + DID as a supplement, rarely as the sole strategy
6. Structural estimation when the question is about a parameter or counterfactual

Branch A — DID

• Staggered adoption? → diagnose with Goodman-Bacon decomposition; estimate with Callaway–Sant'Anna, Sun–Abraham, or de Chaisemartin–D'Haultfœuille.
• Parallel trends: show an event-study plot with pre-trends, not a single pre-period dummy.
• Placebo: randomize treatment timing / units; report the placebo distribution.
• Continuous/dose treatment: justify the dose measure and its exogeneity.

Branch B — IV

• First-stage F well above conventional weak-IV thresholds; if borderline, report Anderson–Rubin or other weak-IV-robust inference.
• Exclusion restriction defended in three registers: theory, institutional detail, and a placebo/falsification.
• Report the reduced form, not only the second stage.
• Address the instrument's own potential endogeneity explicitly.

Branch C — RDD

• McCrary / density test for manipulation at the cutoff.
• Optimal bandwidth (Calonico–Cattaneo–Titiunik) plus at least three bandwidth-robustness checks.
• Covariate smoothness across the threshold.

Branch D — Asset-pricing identification

• Factor construction: pre-register the sort/breakpoints logic; avoid look-ahead and survivorship bias.
• Standard errors: Fama–MacBeth or panel with errors clustered/adjusted appropriately (e.g., Newey–West, Driscoll–Kraay) — never naive OLS SEs on overlapping returns.
• Multiple testing: when the claim is a new predictor, confront the data-mining critique (see rfs-robustness).
• Out-of-sample and subsample stability for any predictability claim.

Branch E — Structural / theory-driven

• State the identifying assumptions and which moments identify which parameters.
• Provide a counterfactual or decomposition that reduced form cannot deliver.
• Show the model fits untargeted moments.

Execution bridge (StatsPAI / Stata MCP)

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

• 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 source of exogenous variation is stated in one sentence
• Design-appropriate diagnostics run (parallel trends / density / first-stage F / SE choice)
• Placebo or falsification test included
• Standard-error structure matches the data (clustering / overlap / cross-section)
• Endogeneity threats are listed and each is addressed, not waved away
• For asset pricing, multiple-testing and out-of-sample concerns are anticipated
• Design is specified tightly enough to survive a Stage 1 (Registered Report) review with no results

Anti-patterns

• TWFE on staggered treatment with no discussion of heterogeneous-effect bias.
• "We control for many observables, so the effect is causal."
• An IV that is "an exogenous event × a lagged endogenous variable."
• A new return predictor reported without confronting the multiple-testing critique.
• Naive standard errors on overlapping or autocorrelated returns.

Output format

【Strategy】DID / RDD / IV / event study / asset-pricing / structural
【Source of variation】one sentence
【Diagnostics done】[parallel trends, density, first-stage F, SE choice, ...]
【Missing diagnostics】[...]
【SE structure】...
【Next step】rfs-empirical-design