Common Errors in Statistics (and How to Avoid Them), 4th Edition

VARIANCE ESTIMATION

In addition to model-based variance estimators, fixed-effects models and GEEs also admit modified sandwich variance estimators. SAS calls this the empirical variance estimator. Stata refers to it as the Robust Cluster estimator. Whatever the name, the most desirable property of the variance estimator is that it yields inference for the regression coefficients that is robust to misspecification of the correlation structure.

GEEs require specification of the correlation structure, but the modified sandwich variance estimator (from which confidence intervals and test statistics are constructed) admits inference about the coefficients that is robust to misspecification of that correlation structure. Why then bother with a specification at all? The independence model is an attractive alternative to interpretation of regression coefficients within the more complicated dependence model. Why not then just assume that the observations are independent, but utilize this variance estimator in case the independence assumption is incorrect? This is not a recommended approach because the correct specification yields an estimator that is much more efficient than the estimator for an incorrect specification. This efficiency is an asymptotic property of the estimator dependent on the number of independent panels. Zeger and Liang [1986] demonstrate the advantages of correct specification of the correlation structures for GEEs.

Specification of GEEs should include careful consideration ...