- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
greetings everyone,
I just have a small question, I am in situation where i have r-family and have n-individuals belonging to these families (r<n), thus am to add a random effect (intercept) to my generalized linear mixed poison model to account for the design. Additionaly, i have don't have repeated measurements, hence i was wondering whether to also include a random slope, does this makes sense? and why?
thanks.
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
proc mixed .........
........
random int/subject=individual(family);
run;
Anyway , @SteveDenham @lvm @StatDave could give you more useful information.
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Some more information would help, especially regarding the repeat measurement issue mentioned by the OP. Are there multiple measurements per subject or not?
SteveDenham
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
For a Poisson model with nested random effects, and repeated measures on the lowest level units, consider this:
proc glimmix data=yourdata method=quad;
class subject family time;
model response=time/dist=poisson s;
random intercept family/subject=subject type=chol;
random time/subject=subject(family) type=vc;
lsmeans time/diff ilink;
store out=glmm_store;
run;
Provided that this converges, you should be able to compare counts at the various time points. If there are other factors in the model, they can be accommodated. This code will return least squares means for each time point, both on the log scale and the observed scale and differences on the log scale and associated ratios on the observed scale. The STORE creates a binary file that, in combination with PROC PLM and the %NLmeans macro, can be used to get the differences on the observed scale and the appropriate standard errors of the difference on the observed scale. Search @StatDave 's posts in this community for links to the use of the %NLmeans macro. And speaking of @StatDave , he may be able to point to the use of PROC GEE if you are interested in marginal results rather than the conditional results from GLIMMIX.
SteveDenham
- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content
Since it is a repeated measure experiment, TIME should be R side random effect ,right?
random time/subject=subject(family) type=vc;
-->
random time/subject=subject(family) type=vc residual ;