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

What do the interval estimates look like for the two methods? Is there perhaps a large interval, such that the two points you have are fair estimates of the midpoint? In other words, if the interval estimates look like (-100, 100), I would be pretty content, but if they were (-0.7, -0.5) frequentist and (-1.1, -1.0) Bayesian, then I might have some suspicions. Have you tried fitting with a non-full rank parameterization (GLM style), and if that is still giving you fits, tried adding a NOINT option to the MODEL statements?   No guarantees on these suggestions, they are just what I would try. I see @jiltao replied so that opens some doors to getting a good solution.   SteveDenham ... View more

If you know the formulas for the log-likelihood, you can, with no small amount of effort, write the supporting code to fit a beta-binomial in NLMIXED or in MCMC, where you make the binomial parameter a function of the beta distribution. Mr. Google might be of some assistance here. I suppose I am saying I know what might work, but I sure don't know how to do it.   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

Yes, it would. It would enable you to account for correlation within ID as well as the correlation/covariance across ID's.   SteveDenham ... View more

Well, there is a "brute force" method. After running PROC SURVEYFREQ with all of the DEPLVL's included, use a WHERE= option on the DATA= part of the SURVEYFREQ statement.   You would then have 5 calls to SURVEYFREQ. The first would use this:   proc surveyfreq data=yourdata; /*remaining code to get what you already have*/ run; proc surveyfreq data=yourdata(where=(deplvl in (0,1)); /*same code as before */ run; proc surveyfreq data=yourdata(where=(deplvl in (0,2)); /*same code as before */ run; proc surveyfreq data=yourdata(where=(deplvl in (0,3)); /*same code as before */ run; proc surveyfreq data=yourdata(where=(deplvl in (2,3)); /*same code as before */ run;  The biggest caveat here is that the weighting will not be the same as for the overall test (I said it was a brute force method, not an elegant one). The second biggest is that you would likely have to run the resulting p values through PROC MULTTEST to get multiplicity adjusted p values.   But there might be another way. Have you considered PROC SURVEYLOGISTIC? You could model all the data, then extract comparisons of interest using a variety of tools (LSMEANS, LSMESTIMATE, SLICE) which enable specific pairwise comparisons and adjust using an ADJUST= option.  For SAS/STAT 15.2 Example 119.2 The Medical Expenditure Panel Survey (MEPS) might be an excellent starting point, given what I can make of your survey design (not really my field).   SteveDenham   ... View more

There are simple ways, and more complete ways. For the simple way try:   proc ttest data=your_data; by question; class appointment_type; var score; run; However, it appears that your data may be clustered by ID, with several questions for each. If there is any chance that the response within an ID is correlated, then perhaps a generalized estimating equation approach would be useful. Here is some example code:   proc gee data=your_data; class appointment_type question ID; model score = appointment_type question; repeated subject=ID/ type=exch covb corrb; run; You may want to include an interaction term, if you want to compare mean scores for each question. That would lead to this code:   proc gee data=your_data; class appointment_type question ID; model score = appointment_type question appointment_type*question; repeated subject=ID/ type=exch covb corrb; lsmeans appointment_type*question; slice appointment_type*question / diff sliceby=question; run; Hope some of this helps.   SteveDenham       ... View more

An alternative to removing the interactions might be to fit ONLY the highest level interaction, and then use specific CONTRAST or ESTIMATE statements to calculate your odds ratios. This will also quickly identify the cells with small sample sizes.   SteveDenham ... View more

What does the "H" designate for the variable ph? Has the plant been harvested? If so, then it should be replaced with a period if it no longer exists, or perhaps with the last observation carried forward (LOCF method). Mixed modeling is robust to missing data when there is adequate replication. You will have to make a scientific decision as to how to proceed, based on what your research question is.   SteveDenham ... View more

This is what I feared when you described the experimental design. I would first try removing the main effects of frt and week from the MODEL statement leaving    model ph =frt*week; This one-way model would test whether any plant at any time point differed from any other plant at any time point. It should get away from the saturated mean/profiled variance.   If that does not work, consider analyzing the data with a generalized estimating equation approach (PROC GEE). Something like this may be possible:   Proc gee Data =first; class frt week; model ph = week; repeated subject=frt/type=ar(1) covb corrb; run; If this runs, we can work on getting marginal means and pairwise comparisons out later.  Additionally, look at Example 49.3 (SAS/STAT 15.2) Weighted GEE for Longitudinal Data That Have Missing Values.   SteveDenham     ... View more

It doesn't mean that - what it does mean is that you have included it in the CLASS statement. If a dependent variable is given in the CLASS statement, GLIMMIX defaults to a multinomial.  So, remove "ph" from the CLASS statement.   SteveDenham ... 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

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   ... View more

Some more information would help, especially regarding the repeat measurement issue mentioned by the OP. Are there multiple measurements per subject or not?   SteveDenham ... View more