Jae data analysis

Data Analysis & Inference for JAE (jae-data-analysis)

When to trigger

• The sample is built and it is time to estimate and report
• You are unsure how to specify fixed effects or cluster standard errors
• Reviewers will probe endogeneity, correlated omitted variables, or sample selection
• A reviewer says "the standard errors are understated" or "this is not identified"

Build and document the archival sample first

JAE reviewers expect a transparent sample-construction waterfall: starting population (e.g., Compustat firm-years), each merge (CRSP, I/B/E/S, Execucomp, DealScan, Audit Analytics via WRDS), each exclusion (financials/utilities, missing data, penny stocks), and the final N at every step. Report descriptive statistics and a correlation table. Winsorize continuous variables (commonly at 1%/99%) and say so.

Specify the estimator to match the panel and the design

Data structure / claim	Estimator / specification
Firm panel with unobserved heterogeneity	Firm and year fixed effects (e.g., reghdfe)
Inference with within-firm correlation	Standard errors clustered by firm; often two-way (firm & year)
Regulatory shock / treatment	Difference-in-differences; report pre-trends
Endogenous regressor	2SLS/IV with first-stage diagnostics (F-stat, exclusion)
Self-selection	Heckman (report inverse Mills) or PSM (report balance)
Information event	Short-window CARs; cross-sectional regression of returns
Binary/limited outcome	Logit/probit/Tobit as the outcome dictates

Match the clustering to where correlation lives in the data; a single firm-clustered SE may understate inference when shocks are common across firms in a year — two-way clustering is the JAE norm for many panels.

Execute the identification, not just the regression

• DiD: plot/test parallel pre-trends; report the dynamic (event-time) coefficients, not only the average treatment effect.
• IV: report the first stage, the instrument's strength, and defend the exclusion restriction in words.
• Matching/Heckman: report covariate balance or the selection equation; show results are not an artifact of the procedure.
• Cross-sectional partitions: the theory's mechanism test — show the effect concentrates where the friction (information asymmetry, weak governance, tight covenants) is severe.

Robustness (expected, not optional)

• Alternative proxies for the key construct (e.g., different discretionary-accruals or conservatism measures).
• Alternative specifications (controls in/out, alternative fixed effects, subsamples).
• Placebo/falsification tests and, for DiD, a non-event window.
• Sensitivity to correlated omitted variables (e.g., bounding / coefficient-stability arguments).
• Address economically plausible alternative explanations empirically.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JAE is empirical accounting with an economics lens; treat identification and weak-IV-robust inference as the binding constraints.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley; multilevel data → cluster at the right level.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive battery in the appendix. See the executed chain in the JF execution walkthrough.

Checklist

• Sample waterfall with N at each step; winsorization stated
• Descriptives and correlation table reported
• Fixed effects and clustered (often two-way) SEs match the design
• Identification executed (pre-trends / first stage / balance), not assumed
• Cross-sectional partition supports the economic mechanism
• Robustness: alternative proxies, specifications, placebos, sensitivity
• Economic magnitude (not only significance) reported

Anti-patterns

• Pooled OLS with no fixed effects or clustering on a firm panel.
• One-way clustering when shocks are common across firms within a year.
• Reporting an IV with no first stage or no exclusion-restriction defense.
• DiD with no pre-trend evidence.
• Significance with no economic magnitude ("statistically significant" but trivially small).
• Selective controls that make the result appear.

Output format

【Sample】population → merges → exclusions → final N; winsorized at ...
【Specification】FE (firm/year); SE clustering (firm / two-way)
【Identification executed】pre-trends / first-stage F / balance ...
【Main result】coefficient, t-stat, economic magnitude
【Mechanism (cross-section)】effect concentrated where friction severe
【Robustness】alt proxies / specs / placebo / sensitivity
【Open issues for reviewers】...
【Next step】jae-contribution-framing