Cluster-robust standard errors — when to use?

hard

Answer

$Observations within clusters (schools, accounts, sessions) are correlated \to default OLS SEs are too small \to false positives.$
$Cluster-robust ('Liang-Zeger', 'sandwich') SEs correct for within-cluster correlation.$
$Rule: cluster at the level of treatment assignment or the highest level of correlation.$
$Needs \geq 30-50 clusters to be reliable — with few clusters, use wild cluster bootstrap.$
$Standard in econometrics / A/B testing on clusters.$

Check yourself — multiple choice

Not needed
Correct within-cluster correlation → wider SEs; cluster at treatment level; needs ≥ 30-50 clusters (else wild cluster bootstrap)
Same as robust SE
Only Bayesian

Cluster SEs: correct within-cluster correlation; need 30+ clusters.

#mixed-effects#regression

215 interview questions in this topic.