[R] Difference between GEE and Robust Cluster Standard Errors

[Thomas Lumley]

On Wed, 18 Feb 2009, jjh21 wrote:

>
> Hello,
>
> I know that two possible approaches to dealing with clustered data would be
> GEE or a robust cluster covariance matrix from a standard regression. What
> are the differences between these two methods, or are they doing the same
> thing? Thanks.

There are two components to 'GEE'. The first is the robust cluster (or 'sandwich') covariance, the second is the ability to choose a weight matrix to get higher efficiency ('working correlation').

Using the 'independence working correlation' asks for the same weighting as in ordinary regression, so the estimates are the same as in standard regression, and then the standard errors are the same as the 'robust cluster' ones (up to factors of n/(n-1) and similar implementation details). The standard errors are also the same as the Horvitz-Thompson estimator gives for cluster sampling from an infinite population, and they are also the same as an approximation to the cluster jackknife standard errors where a cluster is downweighted slightly rather than removed.

       -thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle