I am working with a cross-national data set. It includes observations
for 20 developed countries in 1990. Since these countries are not
randomly sampled, I thought to use the reg option cluster(country).
In fact it specifies that the observations are not independent within
groups
Is this a reasonable solution in this case?
If so, why do I obtain the same result by using robust option?
I wonder if robust standard errors are the statistical remedy both
for no independent observations and heteroskedasticity?
So, what is the relationship between heteroschedasticity and no
indipendent observations?
I ask thesw questions because I didn't find any answers on STATA
manuals.
If you have only those 20 observations, robust and cluster(country)
mean the same thing. The group of observations corresponding to the
country is allowed to have a covariance matrix incorporating intra-
country correlation, and the diagonal of that matrix is country-
specific. But if you have only one obs. per country that matrix is a
scalar with value equal to the "robust" (sandwich/Huber/White)
variance estimator.
Cluster standard errors would make a difference if you have multiple
observations per country (e.g. a panel). But there are other ways of
skinning that cat, e.g. xtgls, xtgee etc.
