Thank you for your feedback.   You can see below my results and attached the data. For my test, there is no a lot of observations.   Results with logistic classic:     Results with BGLIMM:     ... View more

I see the solution vector, but the sliced least squares mean differences are what would address your research question.   SteveDenham ... View more

Well, I wouldn't call it refereed so much as crowd sourced, but https://en.wikipedia.org/wiki/Cram%C3%A9r%27s_V  provides a lot of insight to those of us who aren't familiar with it. The formulas there use k to index the number of columns and r to index the number of rows, so we are not limited to square matrices. Calculating using the summation formulas given doesn't seem hampered by cells with zero counts, so that isn't an issue. The biggest issue with the use of Cramer's V is that it is severely biased toward 1, and unbalanced data increases this.   Is that of any help to you?   SteveDenham ... View more

That makes a lot of sense, I will do so. Thank you so much StatDave!     ... View more

I've corrected that one in SAS. It is a letter from a treatment name with a space in between. I would like to have a pairwise comparison of treatments ( fertilizers) with letters. The hypothesis is that the fertilizers will increase the plant height.    I tried adding;    Proc gee Data =Waterford; class frt week; model ph=frt; repeated subject=frt/type=ar(1) covb corrb; lsmeans frt/ pdiff=all adjust=tukey LINES; run;   Error: Only CLASS variables allowed in this effect. Is it because we added cov, which is already LS means of covariances?      Thanks Steve!        ... View more

@Gimlet wrote: I am interested in producing main effect F Tests within LSMESTIMATE using the joint option. Could you please point us to some example code? @Gimlet please start a new thread and explain your question and explain the data, in detail, from the beginning. People will be happy to help you in a new thread which contains a clear explanation of your problem/question. Do not discuss this further in this thread. ... View more

Your particular GLIMMIX code assumes separate variance components for each level of outcome, estimated from the various hospitals. This is the default for your RANDOM statement. For a multinomial, I think that is the only sensible structure, as imposing a correlation between levels for a generalized logit strikes me as a dangerous step toward convergence, Hessian and G matrix issues unless you have "close to homogeneous" responses across all of the subjects.   I'll call in @StatDave at this point. I think the assumption in PROC GEE that concerns you is no different than the default being fit by your code in GLIMMIX - that the variance components are interchangeable. If that is the case, i think GEE has a better chance of producing usable results.   SteveDenham ... View more

I am new to this contrast statement. Can you help me explain why the number should be  -5 -3 -1 1 3 5 not -4 -2 0 2 4?  ... View more

There seems to be a solution now. You can check this paper, which is an introduction to the %SURVEYGENMOD macro, a macro that is capable of building zero-inflated Poisson and zero-inflated negative binomial models. However, I am not sure if random effects can be included in the models. I have not read this paper carefully. ... View more

Thanks for your detailed explanations Stave. I am actually modeling raw BP values over time (BP1 (baseline) and BP2(first follow-up). And Var1 is typo. Also, there are few time-variant covariates, for example physical activity; PA_base and PA_follow, and I have created a single variable PA_change with PROC TRANSPOSE. Thanks in advance for your furtther explanations! ... View more

@SteveDenham 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 ; ... View more

From experience, I think you may not have enough data to be able to distinguish and correctly estimate your RANDOM effects in these models, which would lead to the G matrix issues. Covariance parameter = 0 means you will get the non-positive definite message in the output. So, except in certain regulatory situations, you should simplify the RANDOM statement.    SteveDenham ... View more

If you are locked into model selection techniques like stepwise, and you realize that it really isn't suited to mixed models, consider treating the difference between the treated and control subjects for each pair as your response variable. Without knowing your data structure, consider this:   proc hpgenselect data=tab; class paired var1 var2 var3 var4 ; model wbc_diff= var1 var2 var3 var4/ dist = negbin link=log ; selection method=stepwise details=all; run; The key here is in calculating the response variable wbc_diff. There are several ways, but all of them depend on the current structure of the dataset "tab". You might be happier using method=lasso/choose=aicc, as it provides at lest some protection against type I error inflation due to multiple testing.   SteveDenham         ... View more

Proc FMM is not as flexible as some of the other procedures with respect to the CLASS statement and explicitly defining the reference level.  Instead, you can adjust the order of CLASS variable levels with the ORDER= option in the PROC FMM statement.   ... View more

I recommend getting a copy of SAS for Mixed Models. Either the 2nd edition or the newer Introduction and Basic Applications edition. Both have chapters on power and sample size determination (the newer version has two chapters on this topic). In any case, you feed in the variance components and an exemplary data set that builds in the differences you want to be able to detect, and run PROC MIXED holding random effects fixed with a PARMS statement. You then post-process the results to get an approximate noncentrality parameter. With that in hand, you can calculate the power. There are several worked examples.   SteveDenham   ... View more