Proc GenMod/HLM questions

[joebacon]

Hi everyone. I am trying to teach myself about Proc Genmod/Proc Glimmix and the intricacies of higher level models. I am using online resources almost primarily and ran into some walls. I was hoping this forum would help me get some more information or know what to look into to help myself.

This is a multilevel analysis to look at whether trial retention (RETENTION) is predicted by intervention condition (CONDITION), attendance at an orientation session (IMPLEMENTATION), controlling for study intervention group – there were 8 of these (STUDYGROUP).

I am working through some code and figured out how to do a two-level model, but I want to add one more layer.

procgenmoddata=AOOF.INTACT_LONGdescending;

 class ID visit(desc) CONDITION (desc) IMPLEMENTATION (desc);

 model RETENTION = visit CONDITION IMPLEMENTATION/dist=binomiallink=logtype3;

 repeatedsubject=id;

run;

I’ll need to add an additional level of STUDYGROUP. I was reading a few different things and one of the ways I could handle this was using GEE framework something like:

proc genmod data=AOOF.INTACT_LONG descending;
  class ID visit(desc) CONDITION (desc) IMPLEMENTATION (desc)STUDYGROUP;
  model RETENTION = visit CONDITION IMPLEMENTATION CONDITION*IMPLEMENTATION /dist=binomial link=log type3;
  repeated subject=id(studygroup);
  lsmeans CONDITION*IMPLEMENTATION /  exp CL ilink;
run;

I can't seem to find a way to get the following which I believe to be the main reporting statistics usually for this type of model (please forgive me if I'm being ignorant or looking over something):

            -adjusted estimates, CIs

            -estimated differences

            -ORs with  CI

How would I achieve these or are there other main reporting statistics that I am overlooking?

One other question I had was how exactly do you represent a nested structure in genmod?  I want to model repeated measures for people who are within study groups for one outcome (retention). For another outcome (attendance) there isn’t a repeated measure but people are still without study group.

[SteveDenham · Accepted solution]

You have an LSMEANS statement, with exp, CL and ilink options.  You should add a diff option on there, and I think you will get what you are looking for.

The other question raises some theoretical issues - do you want marginal responses (averaged over the random effects) or conditional responses (values conditional on the random estimates)?  For the first, look at PROC GEE (or try PROC GENMOD with a REPEATED statement).  For the latter, use PROC GLIMMIX.  There you would have two RANDOM statements - the first would cover any clustering variables while the second would cover the repeated variable.  And then you have to decide whether to make that variable R-side (residual option) or G-side.  

I would really recommend getting a copy of Walt Stroup's book: Generalized Linear Mixed Models: Modern Concepts, Methods and Applications.  It has a lot of really good examples in SAS and covers both applied and theoretical considerations.

SteveDenham

[SteveDenham]

See other posts in this community by @StatDave  regarding macros to get differences and CL's on the original scale.  My mind has been greatly expanded today.

SteveDenham

[joebacon]

Steve,

Thank you for taking the time to suggest some literature and help with my model. I ended up going with Genmod with a repeated statement and the diff option. I have the book ordered and will definitely give it a read- it looks to have good reviews across the board.

Thanks again,

Joe

[StatDave]

GENMOD cannot be used for multilevel models. It does not allow for random effects. See this note regarding nested correlation structures. The SUBJECT= effect is not part of the model specification - it is only used to determine which observations are in a cluster.