Coruscation

Generalised Estimating Equations models, proposed by Liang and Zeger
in 1986, are probably the simplest method for analysing data collected
in groups where observations within a group may be correlated but
observations in separate groups are independent. A complete
description of the method is given in their two 1986 papers. The basic
principle of the method is a generalisation of the fact that weighted
least squares analyses give unbiased parameter estimates no matter
what weights are used. Generalised linear models, such as logistic
regression, have similar robustness properties, giving asymptotically
correct parameter estimates even when the data are correlated. This
means that it is possible to estimate regression parameters using any
convenient or plausible assumptions about the true correlation between
observations and get the right answer even when the assumptions are
not correct.

It is only necessary to use a ``model-robust'' or ``agnostic''
estimate of the standard errors. It would be unreasonable to expect
this freedom of choice to be without cost and it turns out that there
is a moderate gain in efficiency resulting from choosing a working
correlation structure close to the true one.

Useful references include the two original papers (Zeger & Liang 1986,
Liang & Zeger 1986) and two recent books: Diggle, Liang & Zeger (1993)
and Fahrmeir & Tutz (1995). As far as I know the most elementary
treatment anywhere in the literature is still Zeger & Liang (1986).

Section 2 gives an overview of the theory and use of Generalised
Estimating Equations. Section 3 describes how to use the Lisp-Stat
code, including diagnostics. Finally there is a brief discussion of
missing data handling and of other software for fitting GEE models.
Appendix A describes some aspects of the implementation, including the
global variables (Table 5) that control many program options.