Thomas Lumley • 12/18/2023

Why do the Rao-Scott tests have good size?

This technical article analyzes the performance of Rao-Scott hypothesis tests in regression models for complex survey data. It contrasts them with standard weighted tests, explaining that while both require estimating a variance matrix V, the Rao-Scott formulation is less sensitive to poor estimation of V—a common issue in surveys with few, large clusters. The piece delves into the mathematical rationale, involving eigenvalues and the Satterthwaite approximation, for their superior control of test size (Type I error) in contingency tables and generalized linear models.

0 comments

#Hypothesis Testing #Survey Data #Statistical Inference