📅 Uses 73 Years of Historical Weather to Rethink Rainfall Randomization
Rainfall is widely used as an as-if random shock in political science and economics, but county-level rainfall assignment is highly correlated across space. Because this clustered assignment does not align with clear unit boundaries, conventional variance estimation for rainfall effects on political outcomes is challenging and may understate uncertainty.
🧪 How the Test Was Set Up — Replaying National Rain Patterns
- Treats 73 years of historical weather patterns as potential randomizations for election-day rainfall.
- Replicates the influential U.S. presidential turnout analysis by Gomez, Hansford, and Krause (2007) and computes the estimated average treatment effect (ATE) of rainfall on turnout.
- Compares that ATE to a sampling distribution of estimates generated under the sharp null hypothesis of no effect by drawing alternate national rainfall configurations on actual and would-be election days.
- These alternate randomizations preserve spatial clustering in treatment assignment and avoid simulating weather or imposing arbitrary clustering boundaries.
🔎 Key Findings
- The ATE from the replicated analysis is evaluated against the distribution produced by historical random draws under the sharp null.
- The estimated effect of rainfall on turnout exhibits greater sampling variability than indicated by conventional standard errors.
💡 Why It Matters
- Using historical national weather patterns for randomization inference preserves the empirical clustering of rainfall assignment and removes the need for boundary assumptions when estimating variance.
- This approach suggests that prior studies relying on standard error conventions may understate uncertainty about rainfall’s impact on political behavior, with implications for causal claims that use weather as an exogenous shock.
