Logistic Regression: A Self-learning Text, Third Edition (Statistics in the Health Sciences)

Correlation structure:

Framework for estimating:

 Correlations

 Regression coefficients

 Standard errors

Primary interest?

Regression
coefficients

Yes

Correlations Usually not

Infant Care Study example:

AR1 autoregressive correlation

structure specified

Other structures possible

To run a GEE analysis, the user specifies a

correlation structure. The correlation structure

provides a frame-work for the estimation of the

correlation parameters, as well as estimation

of the regression coefficients (b 0 ,b 1 ,b 2 ,...,bp)

and their standard errors.

It is the regression parameters (e.g., the

coefficients for DIARRHEA, BIRTWGT, and

GENDER) and not the correlation parameters

that typically are the parameters of primary

interest.

Software packages that accommodate GEE

analyses generally offer several choices of

correlation structures that the user can easily

implement. For the GEE analysis in this exam-

ple, an AR1 autoregressive correlation struc-

ture was specified. Further details on the AR1

autoregressive and other correlation structures

are presented later in the chapter.

In the next section (Sect. III), we present

the general form of the data for a correlated

analysis.

498 14. Logistic Regression for Correlated Data: GEE