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.