Empirical & Structural Design (rfs-empirical-design)

When to trigger

• The identification strategy is chosen but sample, variables, and estimator are unsettled
• You must decide between panel FE, Fama–MacBeth, GMM, or a structural estimator
• Portfolio sorts / factor construction choices feel arbitrary
• Measurement of the key variable is contestable (proxy validity)
• A referee will ask "why this sample / this window / this proxy?"

Design decisions that make or break an RFS empirical paper

RFS publishes design-defining empirical templates referees will hold you to — e.g., the q-factor construction in Hou, Xue, and Zhang (2015) "Digesting Anomalies" (RFS 28(3)) and the variance-risk-premium measure in Bollerslev, Tauchen, and Zhou (2009) (RFS 22(11)). Two RFS-specific pressures sharpen every choice below: (1) the public code-release condition means every filter and construction step must be reproducible by a stranger, not just described; (2) the Registered Reports option means a design can be locked at Stage 1, so pre-specify wherever you can.

1. Sample construction

• State the universe, the time span, and every filter, with the resulting N at each step (a sample-attrition table).
• Justify the start/end dates by data availability or regime, not convenience.
• Handle survivorship, look-ahead, and backfill bias explicitly (CRSP/Compustat merge timing, delisting returns, point-in-time fundamentals).
• Winsorize vs. trim: state the rule (e.g., 1%/99%) and apply it consistently.

2. Variable measurement

• For each key variable, give: definition, data source, construction formula, and unit.
• Defend proxy validity — a proxy needs a first-principles or validation argument, not just precedent.
• Avoid mechanical correlation between LHS and RHS (e.g., overlapping accounting items).

3. Estimator choice

4. Fixed effects and controls

• Saturate fixed effects to absorb the right confounders (firm, industry×year, etc.) — but show the result is not mechanical to the FE choice.
• Distinguish controls that are "bad controls" (post-treatment / outcomes) from legitimate covariates.

5. Standard errors

• Cluster at the level of treatment assignment or the unit of correlation.
• For asset pricing, match SE to the return structure (Newey–West for autocorrelation, Driscoll–Kraay for cross-sectional + serial dependence).

Execution bridge (StatsPAI / Stata MCP)

Run the asset-pricing battery, don't just specify 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.

• Factor regressions / time-series alphas: feols with the right SEs (Newey–West / clustered) — read the alpha and t off the return.
• Factor-zoo haircut: after disclosing how many signals were screened, apply romano_wolf / benjamini_hochberg and report the alpha that survives.
• Fama–MacBeth + Shanken EIV are Stata-canonical — run via mcp__stata-mcp__stata_do with the vendored resources/code/ (asreg / xtfmb).
• Exhibits: etable; hand formatting to the tables/figures skill.

Report the economic magnitude (bps/month alpha, Sharpe gain); full factor grid → appendix. JF execution walkthrough.

Checklist

• Sample-attrition table present; every filter justified
• Survivorship / look-ahead / backfill addressed (point-in-time data)
• Each key variable has definition + source + formula; proxy validity argued
• Estimator matches the question; benchmark estimator also reported
• FE structure justified; no bad controls
• SE clustering / adjustment matches the data-generating structure
• Every filter/construction step is reproducible from the to-be-released code (RFS condition)
• Design choices are pre-committed where possible (Stage-1-ready), not chosen ex post

Anti-patterns

• An unexplained sample period or unexplained filters that conveniently strengthen results.
• A proxy defended only by "following prior literature" when its validity is in doubt.
• TWFE reported as if it were a modern staggered-DID estimator.
• Controlling for post-treatment outcomes ("bad controls").
• Standard errors that ignore overlapping returns or treatment-level clustering.

Output format

【Sample】universe / span / filters / final N
【Key measures】variable → source → formula → validity note
【Estimator】... (+ benchmark)
【FE & controls】...
【SE structure】...
【Next step】rfs-robustness