09-03-2014 01:06 PM
I am conducting a matched longitudinal study of hospitalization outcomes (with a binary outcome variable and only time-independent covariates). I was wondering whether the robust standard errors will differ depending on the specified covariance structure. In other words, are Huber-White standard errors (used with PROC GENMOD, logit function) calculated independently of the chosen covariance structure (eg. unstructured, exchangeable etc)? I am unable to verify this for myself given that the model is only converging when an independence correlation structure is specified (given limited data).
Thanks in advance for your support.
09-04-2014 03:29 PM
Because the standard error is a quadratic form involving the variance-covariance matrix, I believe the selection of the covariance structure will have a profound influence on the values obtained. That's the long answer. Short answer--they are not calculated independently.
The fact that the model is only converging under an independent structure, is as you surmise, almost certainly due to limited data.