My modeled process has binary outcomes and site and region grouping. I am modeling it with a generalized linear mixed model (using GLIMMIX) and a diagonalized covariance structure. Typically I see covariance component based formulas for ICC calculation in the literature. GLIMMIX gives me the covariance of the site indicator and the region indicator variables but not the error variance I need to fill out the formulas. I have seen a suggestion to use 3.29 (pi**2/3) as the error variance when the dispersion is near 1.0, but am not sure this is right.
Hi David,
I suggest you check out Larry Madden's home page and maybe shoot him an email (you can tell him I suggested it - we're old friends). Larry has spent a lot of time developing techniques for handling binomial data with high intra-class correlations. He has some SAS routines and other software on his web page but he is always interested to hear from people looking at this sort of issue:
http://plantpath.osu.edu/madden
He has a ResearchGate profile too:
https://www.researchgate.net/profile/Laurence_Madden
Yours,
N
Instead of using a 'random intercept' statement, try modeling the R-side residuals with 'random _residual_'. This option can be used if dispersion is not equal to 1. Here's the support documentation: http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_glimmix_sect009.htm
Hope that helps!
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms