Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› jae-data-analysis
jae-data-analysis
GitHub用于JAE实证分析执行与报告，涵盖样本构建、固定效应设定、聚类标准误选择、识别策略实施及稳健性检验。适用于已建好样本需估算报告、不确定模型设定或应对审稿人关于内生性与标准误质疑的场景。
Trigger Scenarios
样本已构建，需进行估计和结果报告
不确定如何指定固定效应或聚类标准误
审稿人质疑内生性、遗漏变量或样本选择问题
审稿人指出标准误被低估或识别不足
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill jae-data-analysis -g -y
SKILL.md
Frontmatter
{
"name": "jae-data-analysis",
"description": "Use when running and reporting the empirical analysis for a Journal of Accounting and Economics (JAE) manuscript — building the archival sample, choosing fixed effects and clustered standard errors, executing the identification design, and demonstrating robustness for large-sample capital-markets\/contracting\/disclosure data. Executes and reports the analysis; it does not design the study (jae-methods) or frame the contribution (jae-contribution-framing)."
}
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) orbenjamini_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 exactsuggest_functionfor each. - Exhibits:
etable/did_summary_to_latexfrom 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
Version History
- 1839142 Current 2026-07-05 13:25
