aer-robustness

# AER Robustness ## Overview A modern AER referee report contains three predictable demands: 1. **Robustness** — does the result survive specification changes? 2. **Heterogeneity** — where does the effect concentrate, and is that consistent with the proposed channel? 3. **Mechanism** — *why* does X cause Y? This skill anticipates all three so that the referee finds the answer already in the paper. Skipping this step turns a referee report into a 6-month delay. ## When to Use - The main result table exists but the rest of the empirical section is thin - An R&R demands "additional checks" without specifying which - Drafting the appendix before submission - Diagnosing why a prior submission drew a "needs more robustness" rejection ## The Referee-Anticipating Battery Every empirical AER paper should report, at minimum: ### Robustness 1. **Alternative specifications** — drop covariates one at a time; include fixed effects at finer/coarser granularity; weight observations differently 2. **Alternative samples** — drop the largest unit; drop the most influential time period; restrict to balanced panel; restrict to comparable subsets 3. **Alternative outcome definitions** — log vs. level; winsorized at 1% / 5%; alternative deflators 4. **Alternative clustering** — cluster at the next-higher level (e.g., state if main is county); two-way cluster; wild cluster bootstrap if few clusters 5. **Alternative estimators** — if main is OLS, show IV; if main is TWFE, show Callaway-Sant...