02-19-2015 01:58 PM
Here is a possible source for an answer: Choosing the Correct Statistical Test in SAS</title><link href="/stat/ats_style.css" type="text/css"...
02-19-2015 09:02 PM
I'll try and repost it here. Remove the two embedded spaces from the following and copy and paste the url to your browser: www.ats.ucla.edu /stat/ sas/whatstat/
02-20-2015 06:50 AM
Arthur--Thanks, I always forget about canonical correlation. That's a good cross-software comparison, focused on OLS. Tom
02-19-2015 04:28 PM
Different possibilities. What is your goal? Simultaneous modeling of all the (possibly correlated) response variables or just having a convenient way of getting separate analyses for several response variables? Check out PROC PLS (partial least squares).
02-19-2015 07:46 PM
Simultaneous modeling...as in MANOVA but extended to approaches such as hierarchical mixture models (e.g., Proc Mixed with multiple DVs which this procedure won't allow), maximum likelihood estimation for mixtures of DV forms, and so on. Proc PLS is kind of the idea but I prefer to obtain estimates that aren't collapsed across a PLS-derived component.
02-19-2015 11:06 PM
Lots of ways of doing this with proc mixed and glimmix for hierarchical models. This old article is technically about repeated measures in GLM and MIXED, but is ultimately a comparison of MANOVA in the two procedures.
Here is a more recent blog on the topic:
The syntax is different, and the data are stacked, for analysis in MIXED and GLIMMIX compared with GLM, but you can handle so many more situations with the mixed-model procedures. You are not even restricted to the same distribution. One of the nice examples in the User's Guide for GLIMMIX deals with a multivariate analysis of two random variables, one binary (Bernoulli) and the other count (Poisson).
02-20-2015 07:01 AM
LVM-Interesting. The different data structures -- e.g., stacking up the "Y" variable and adding an "Age" predictor in Proc Mixed vs using "Y1-Y4" with a REPEATED statement in Proc GLM isn't quite what I have in mind. Proc Mixed (or Glimmix) still only permit a single "Y" or dependent variable. Tom
02-20-2015 12:52 PM
No, you are incorrect. MIXED and GLIMMIX can absolutely handle multiple dependent variables. This is exactly what is demonstrated in the links I sent.
02-20-2015 01:03 PM
Light bulb goes on. It all comes down to identifying a proper SUBJECT=. Then with an unstructured or factor analytic covariance matrix you can literally have your cake (multivariate) and eat it too (hierarchical), and all you need is a proper machine capable of fitting the model.
03-26-2015 10:20 AM
In thinking more about your SUBJECT= suggestion, it's not clear to me how the SUBJECT= option differs from the GROUP= option in the RANDOM statement. In other words, if one is modeling multiple dependent variables, would one approach be to treat the data stacks for the different DVs using the GROUP= option and the actual unit of analysis with the SUBJECT= option?
Some clarification of this would be helpful!
03-30-2015 09:45 AM
Larry has done a lot more in this area than I have, but I would try to avoid the use of the GROUP= option in a multivariate analysis because of the potential blow-up of the number of parameters to be estimated. The Joint Modeling of Binary and Count Data example in the PROC GLIMMIX documentation is what I have been working off of, where 'dist' or 'parm' or something else is a fixed effect, modeled as a repeated measurement on an individual subject. I am having a hard time figuring out how to interpret a GROUP='dist' under this approach. Now if I understand Walt Stroup's book correctly, and model these as Zg inside the linear predictor, then maybe I can wrap my head around a GROUP= approach. But I'm old, and that hurts my brain too much.
If someone has a worked example I could look at, that would help.
03-30-2015 10:57 AM
As stated by Steve, group= would not be appropriate for your problem. With GROUP, you get separate parameter estimates (or random effect prediction) for each group. The groups are treated independently. Your multivariate responses are correlated. So, you should use subject=, the syntax for correlated data. For more advanced modeling, you can combine group and subject, but I won't get into that here.