A Comparison of Jackknife Estimators of Variance for GEE2

Marginal regression modeling with generalised estimating equations
became very popular in the last decade. While the mean structure is
of primary interest in first-order generalised estimating equations
(GEE1), second-order generalised estimating equations (GEE2) allow
the estimation of both the mean and the association structure. It
has repeatedly been shown that the usual robust variance estimator
for the GEE1 is conservative, especially in small samples. As an
alternative, the jackknife estimator of variance can be used. In
this discussion paper, we extend the different jackknife estimators
of variance to GEE2 models. The variance estimators are compared in
a simulation study. While there is only little difference in the
variance estimates of the mean structure across simulated models,
the results differ substantially with respect to the association
structure. The fully iterated jackknife estimator seems to be the
most appropriate when focusing on the GEE2.