Glimmix for longitudinal binary outcome

[gaucho]

Previously I have been using GLIMMIX with the LOGIT link to look at the effect of covariates and time on a disease diagnosis (0,1). Subject outcomes were recorded at 4 time points.  The results looked reasonable.  Most recently, I recoded the outcome so that if a subject had the disease at any time point they would also have the disease at later time points.  The output looks terrible- the estimates are outrageously large.  Could this be due to the correlations between timepoints?

-G

[SteveDenham]

YES!  You will probably need to include some kind of error structure in the random statement, and the choices will depend on the spacing of the time points and the amount of data you have to fit the various structures.  For examples, look at Stroup's Generalized Linear Mixed Models.

Steve Denham