Are you referring to the error:   ERROR: QUANEW Optimization cannot be completed.   or variations on this error?  If so, you can invoke the following option:   options errors=0;   However, this will suppress only error messages in the log associated with the data, which is not quite what you want I think.  Otherwise, I don't see an easy way, although a difficult way could be to text process the log file.   SteveDenham ... View more

I am going to assume you mean you would have the following 4 groups (each measured two times)   Legume A monoculture Legume B monoculture Grass and Legume A mixture Grass and Legume B mixture   So that means you would have 8 measurements in total.  There is no power to detect anything except two or more fold differences.  Are there plots within each of these that were measured?  If so, you may have something that will be aided by an inferential statistical analysis.  But if it is a single measurement, you could present the entirety of your information content in a 4x2 table of the observations.   SteveDenham ... View more

Take whatever analysis code you were going to use, and remove any reference to the grass cultivar, as it is no longer a factor in the design.   You have two levels for the main effect, and two time points that should be treated as repeated measures on the legumes.   What I worry about is that it looks like you only have 4 measurements (2 legumes, 2 times), so any analysis will be severely underpowered.   SteveDenham ... View more

In the absence of any statements like LSMEANS or LSMESTIMATE, you could explore the CODE statement.  According to the documentation, it writes DATA step code to a file that can be used to calculate predicted values (probabilities).  If you have that code, then you can plug in the values for which you want predicted probabilities, which could then be used to get ORs.  I suspect it will not easily give you a standard error, so Wald chi-squares might take more coding.  But if you have a predicted probability, you can calculate the variance using well-known formulas, and from there get the denominator as a function of a sum of the variances.   SteveDenham ... View more

I am not sure at all if this is the issue, but the counts are non-monotonic by group 1 to 9, so it could be that for some group or occasion combination, there is something unusual.  In particular, I would examine Group 6, which has a count greater than Groups 3, 4 and 5.  The documentation says "If the pattern of missing data is intermittent (not dropout), the GEE procedure terminates and does not perform an analysis."  I would suspect that there is at least one intermittent missing value for asthma in there.   SteveDenham. ... View more

I fear you may have to hand-edit the gmcovb3 dataset, but first, consider using the canonical logit link rather than the log link. If things still don't work well, then printing out that covb dataset to look for zero or near zero terms may be what you have to do.   Aaaannnddd, have you checked for quasi-separation (on diagnosis, which looks like Parm6) in any of your imputed datasets?   SteveDenham ... View more

I would guess that the reason you don't get anything for the negative binomial is that the negative binomial distribution is not supported in the HISTOGRAM statement.  Looking over the list of supported distributions I see that they are all continuous (no discrete distributions such as binomial, Poisson, geometric, negative binomial are listed).   SteveDenham ... View more

Putting q1 through q32 in the class statement is one way of doing this, but a better way would be to restructure your data so that the variable 'question' took on the values 1, 2, 3, ... , 31, 32.  Currently, I envision your data to look something like this for each line:   idschool idstud itsex tech event q1 q2 q3 ... q32   In this case, idschool has multiple levels as does idstud.  Itsex has 2 levels (presumably), event has 2 levels, tech has 2 levels (for the sake of illustration) and q1 through q32 each have 2 levels, such that for school=6 and student=100 who is male and had events for question 1, 30 and 32, the data would look like:   6 100 M 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1   or something along those lines.  What I would suggest is a "long" version for variables idschool, idstud itsex tech question event:   6 100 M 1 1 1 6 100 M 1 2 0 6 100 M 1 3 0 <keeps repeating to > 6 100 M 1 29 0 6 100 M 1 30 1 6 100 M 1 31 1 6 100 M 1 32 0   With your data structured in this format, your GLIMMIX code would simplify to:   proc glimmix data=final_clean method=laplace; class idschool idstud itsex tech question; model response (Event='1')=itsex|tech|question / Dist=Binary link=logit solution noint ; random intercept / subject=idschool type=vc; random intercept / subject=idstud(idschool) type=vc; lsmeans itsex*tech*question/ilink; run; The lsmeans statement will yield 128 probabilities.  Down near the bottom of each sex by tech section, there will be an estimate for Question 30.   Now if you want to compare these values, see the many posts by @StatDave regarding the use of the  %NLmeans macro to get the differences.  The documentation for the %NLmeans macro is in this note: https://support.sas.com/kb/62/362.html    SteveDenham   ... View more

For the  simple case you present, you can use an LSMEANS statement:   lsmeans itsex*tech*q30/ilink; The ILINK option puts the means back on the original scale. One issue i see is that the model you are using treats all of the q variables as continuous.  I suspect that you have dummy variables in your dataset to accomplish this.  Wouldn't it be easier to have a categorical variable called "question" with 32 levels, where each record has a single entry for question ranging from 1 to 32, rather than 32 variables coded 0/1?   SteveDenham   ... View more