Second, they estimate the village-level regressions with the two randomly selected towns. The standard errors are adjusted for possible correlation within the GP using the Moulton correction \cite{r1986}.