Data Analysis (jfi-data-analysis)

When to trigger

• Building the empirical analysis on bank/firm/loan data, or its robustness battery
• Building a numerical example or calibration that illustrates a model's mechanism

Empirical track (banking data)

• Sample construction: document the universe (e.g., Call Reports / FR Y-9C banks, DealScan loans, HMDA mortgages), merge keys, and every filter; intermediation samples are sensitive to mergers, charter changes, and reporting breaks.
• Variables: define balance-sheet and credit quantities precisely (levels vs. growth, winsorizing, deflation); state timing relative to the shock to avoid mechanical reverse causality.
• Specifications: high-dimensional fixed effects (reghdfe / fixest); for credit-supply questions, use firm×time effects in matched lender–borrower panels to absorb demand.
• Robustness: alternative samples and windows, placebo periods, leave-one-out by large institutions, alternative clustering, and a balance/parallel-trends check for DID. The expected battery is substantial but there is no fixed robustness-table count; keep the main text compact and push secondary checks to appendices.

Theory track (numerical illustration)

When the paper is a model, "data analysis" is lighter and means reproducible computation:

• A numerical example or calibrated figure showing the mechanism and comparative statics — illustrative, not estimation.
• Keep the code clean and deterministic (fixed seeds/parameters) so a reader can regenerate every figure.

Data sharing (both tracks)

Prepare a Data Statement and link datasets via Editorial Manager; cite data with the [dataset] tag (see jfi-replication-and-data-policy). Under Elsevier Option C, deposit/cite/link research data where possible or explain why sharing is restricted.

Dataset-to-mechanism decision table

Pick data for the intermediation mechanism, not the other way around — JFI referees notice when the dataset cannot carry the claimed channel:

Worked robustness pass: a capital-shock battery (illustrative)

A hypothetical JFI paper estimates that a 1pp rise in required capital cuts loan growth to the same firm by 2.1pp (s.e. 0.6, firm×time FE, clustered by bank). The battery a JFI referee expects, each row tied to a named threat:

• OLS without firm×time FE gives −3.0pp; report both, so the reader sees demand absorption moves the estimate by roughly a third — evidence the design bites, and a sorting fact worth a paragraph.
• Drop the three largest banking groups: −1.9pp — the channel is not one institution.
• Placebo reform date two years earlier: +0.2pp, insignificant — supports timing.
• Extensive margin (relationship termination) rises 1.4pp — the intermediation mechanism shows up beyond intensive-margin amounts.
• Few-cluster check: wild-cluster bootstrap p ≈ 0.03 with 31 banks.
• Multi-bank vs. full sample: re-estimate firm-FE-only specs on both, since the within-firm identifying sample skews toward larger, less bank-dependent borrowers.

Analysis probes specific to this venue

• Referees here routinely ask for the exposure-weighted firm-level aggregation when real outcomes (investment, employment) are claimed — firm×time FE cannot be used there, so pre-shock bank shares must carry the identification.
• Magnitude sanity: convert the loan-level coefficient into aggregate credit terms and benchmark it against the range in the lending-channel literature; an estimate ten times the consensus invites a measurement question before a citation.
• For the theory track, a calibration table listing every parameter, its value, and its source (moment matched, literature, normalization) is the JFI-credible substitute for a robustness battery.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JFI is banking and financial intermediation — typically corporate / bank causal designs built around regulation and shocks.

• Many outcomes / specifications: romano_wolf (step-down FWER, accounts for cross-test correlation) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr — the confounder strength that would overturn the headline.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each — no guessing the battery.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive (now actually-run) battery in the appendix. See the executed chain in the JF execution walkthrough.

Anti-patterns

• Undocumented sample filters that drive the result
• Mixing credit supply and demand without firm×time absorption
• A theory "calibration" presented as if it were estimation
• Non-reproducible figures (random seeds, manual steps)

Output format

【Track】empirical / theory
【Sample / parameters】<universe + filters, or calibration>
【Core spec / example】<FE structure, or the numerical illustration>
【Robustness】<the battery, or seed/determinism notes>
【Next skill】jfi-contribution-framing