- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
I'm fitting a generalized estimating equation-type model by specifying the empirical model option in proc glimmix. I am using this in order to model spatial autocorrelation among residuals. The model response is 0/1 and is modeled with a binomial distribution. I know that proc genmod outputs Pan's QIC (quasilikelihood version of AIC) but proc genmod does not (to my knowledge) include the spatial modeling options of glimmix. Has anyone found a way to extract or calculate QIC based upon glimmix model output? I could potentially "borrow" the fitted correlation structure from glimmix and use it as a user-defined correlation structure in genmod, and then get QIC from genmod. However, my data set is quite large with >100 spatial clusters and so handling this seems it would be quite tricky. The empirical option is being used because without it my machine runs out of memory in its efforts to populate a huge matrix that includes the correlation among all data points, even the very distant. My data is quite "cheap" and I have discarded many points (it is satellite imagery). So, for now I am simply validating with an independent data set for model comparison purposes. I believe this gets at the purpose of AIC, but you know how some reviewers are with their insistence upon including the latest buzz word they've latched ahold of. And for a few, it seems "information theoretic" is it. Any guidance is appreciated.-Seth
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
http://support.sas.com/kb/26/100.html
http://support.sas.com/kb/26/addl/fusion26100_1_qic.sas.txt
The first link provides a description of the macro and demonstrates use of the macro (under the Results tab). The second link provides the macro code.
HTH
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
I have similar problem. I have over 300 clusters, where sizes of clusters varied from 1 (there are many totally independent subjects) to 250 of individuals with total of over 5000 individuals. I have already obtained correlation matrix between indiviuals using R for each of the clusters. I would like to used PROC GENMOD in SAS and use my own correlation structure instead of using say, EXCH for each cluster. There is an option for userdefined matrix in SAS for a single matrix, but how would I specify these several matrices for these clusters? Also, how do I make sure that the SAS picks up the correct correlation between any two individuals from my data in the GENMOD analysis?
Is there any way to specify user defined correlation matrix of all individuals in GENMOD without spefifying clustering variable?
Many thanks in advance.
BNO.