When Should You Adjust Standard Errors for Clustering?

Abstract Clustered standard errors, with clusters defined by factors such as geography, are widespread in empirical research economics and many other disciplines. Formally, clustered errors adjust for the correlations induced sampling outcome variable from a data-generating process unobserved cluster-level components. However, econometric framework clustering leaves important questions unanswered: (i) Why do we some ways but not others, example, state gender, observational studies completely randomized experiments? (ii) Is variance estimator valid if observe large fraction of population? (iii) In what settings does choice whether how to cluster make difference? We address these using novel inference on average treatment effects. addition common component, new incorporates design component that accounts variability assignment mechanism. show that, when number sample is nonnegligible population, conventional can be severely inflated, propose estimators correct this bias.