jbf-identification-strategy

Identification Strategy (jbf-identification-strategy)

When to trigger

• The main result is empirical and could be challenged as endogenous
• The design uses bank panels, regulatory shocks, event studies, IV, RDD, or dynamic panels
• You need to decide which robustness and falsification tests JBF referees will ask for

JBF credibility bar

Finance referees usually ask whether the result is driven by omitted risk, selection, reverse causality, market-wide shocks, or correlated bank/firm traits. The design must make the identifying variation visible.

Design checks by strategy

Bank or firm panels

• State the unit, time period, and identifying variation.
• Use fixed effects that match the claim: bank, firm, borrower, market, time, bank-by-market, industry-by-year, or borrower-by-year as needed.
• Cluster at the treatment or shock level; use two-way clustering when shocks vary by unit and time.

Staggered DID / policy shocks

• Show treatment timing and untreated/control units.
• Avoid relying only on plain TWFE when treatment effects may be heterogeneous.
• Report modern DID/event-study estimates and pre-trend tests.
• Explain why anticipation, selection into treatment, and concurrent shocks do not drive the result.

IV / dynamic panels

• Defend exclusion, not just relevance.
• Report first-stage strength and weak-IV-robust inference where relevant.
• For system GMM, limit instrument proliferation and report serial-correlation and overidentification diagnostics.

Event studies

• Define event, estimation window, event window, benchmark model, and clustering.
• Separate market reaction from real effects.
• Add placebo event dates or unaffected securities when possible.

Regulation-shock design picker

Worked threat audit (illustrative)

Claim: banks crossing a $10 billion supervisory threshold cut small-business lending.

• Naive estimate: −3.1% loan growth for crossers (illustrative). Threat: banks time acquisitions to cross, so crossers differ.
• Fix 1: restrict to organic-growth crossers; suppose the estimate moves to −2.2% — report both.
• Fix 2: donut specification plus a bunching plot of the asset distribution; visible bunching below the cutoff must be discussed, not assumed away.
• Fix 3: placebo thresholds at $8bn and $12bn where no rule changes; effects there should be near zero.

Clustering quick rules for bank data

• Policy varies by state or country → cluster at the policy level; with few clusters, wild-cluster bootstrap.
• Bank-quarter panel with bank-level treatment → cluster by bank; two-way (bank and quarter) when macro shocks load on outcomes.
• Loan-level data with repeat borrowers → cluster by borrower (or two-way with bank), and justify the choice in the table note.

Pushbacks JBF referees raise

• "The regulation applied to everyone — where is the control group?" → switch to an exposure design with pre-period balance evidence.
• "Pre-trends look noisy." → add a joint pre-trend test and a sensitivity bound, not only the plot.
• "Weak banks attract supervision (reverse causality)." → show treatment does not predict pre-period outcomes.
• "Banks adjust before the rule binds (anticipation)." → re-center event time on the announcement date and show both timelines.
• "Other Basel rules phased in simultaneously." → restrict to windows where only one rule changes, or exploit exposure differences across rules.

Minimum referee-proof package

• Main specification table with transparent controls and fixed effects
• Alternative fixed effects, clustering, and sample restrictions
• Placebos / pre-trends / falsification outcomes
• Mechanism or heterogeneity tests tied to the finance theory
• Economic magnitudes, not only statistical significance

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JBF is empirical banking/finance — corporate/bank causal designs around regulation and shocks.

• 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.

Output format

[Claim] causal / predictive / descriptive
[Identifying variation] ...
[Core threat] ...
[Design defense] ...
[Required robustness] ...
[Next step] jbf-data-analysis