In addition to @Ksharp 's observation about model complexity, I would suggest that you look into using the NLOPTIONS statement in your GLIMMIX code. If the iteration history is relatively well behaved (i.e., it is close to monotonic throughout), simply increasing the number of iterations from the default of 20 using the MAXITER= option may result in convergence. I almost always set this at MAXITER=1000, especially when fitting binary data.   SteveDenham ... View more

I realize this is late, but this is one situation where some of the options in GLIMMIX come in handy (see the GLIMMIX documentation). You can test for homogeneity of variance using the COVTEST statement in conjunction with using the Group= option in the RANDOM statement. My advice would be to fit the heterogeneous model no matter the result of the test, AS LONG AS you have sufficient data to fit the  number of parameters needed for the covariance structures under consideration and the model converges without messages or errors.   SteveDenham ... View more

Would there ever be a reason to not include DIFF=method>? Absolutely yes. All of the methods are designed to control type I error (false positives). However, suppose that it is far more important to control type ii error (false negative rate) in whatever field you are working. Then the unadjusted values give a greater ability to declare a difference significant. As an example, suppose you are working on a safety study for a new drug. In that study, it would be much more important to detect a true difference to protect you from declaring the drug "safe" when in fact it causes some change that may indicate a potential safety concern.   SteveDenham ... View more

On your ESTIMATE statements, I assume you wish to find marginal values over Type, Treatment and the Type by Treatment interaction, as there are no terms included in the statement for those factors.   Have you considered using LSMESTIMATE statements, with the AT= option for various interesting values for the covariate?   SteveDenham ... View more

R-side random effect have the similar/same structure with G-side at some time. You could take R-side random 's error term as: error=Zu here Zu has the similar/same structure with G-side random Zu. E.X. the following two model(G-side and R-side) are the same model. G-side: proc imixed; class subject; model y=time; random int/subject=subject; run; R-side: proc mixed; class subject; model y=time; repeated time/type=cs subject=subject; run; Except RANDOM constrain the correlation to be positive,whereas REPEATED leaves the correlation unconstrained. ... View more

Oh ok, but at least it helped me a lot with proc glm and when having a binary outcome in proc logistic. Thanks a lot Ksharp! ... View more

Please be aware that the assumption of normality of the residuals is also required for linear regression of complex survey data. See page 2 of Linear regression diagnostics for survey data for more details. So in theory, a nonparametric test is required. ... View more

I have no idea on the solution to this specific question, but would like to recommend the only monograph I have found on building overdispersion models with SAS: Amazon.com: Overdispersion Models in SAS: 9781607648819: Morel PhD, Jorge G., Neerchal PhD, Nagaraj: Books. I have read part of it and am not sure whether it contains the answer to your specific question. But I am sure it is a good monograph for those who wish to have a systematic understanding on this field. ... View more

We have had luck using the standard error of the lsmean reported in GLIMMIX as the value to plug in for sigma^2 to give a standard error on the arithmetic scale. Logically this seems a bit off, but the values for the standard error (square root of the variance) are reasonable.   Wikipedia presents the following formulas.  in which mu is the expected value (mean) on the log scale and sigma^2 is the variance on the log scale.   I know I must be missing something algebraic as to why the two sources use different representations. I definitely think @Rick_SAS 's statement about using the estimate of the scale parameter from the SOLUTION is helpful.   SteveDenham ... View more

Hello Steve, thank you very much for getting back to me about this. I appreciate it. I ran several models, and compared results, and I believe that I have a clearer picture now.   ... View more

You could try COVB option: proc glimmix data=sashelp.heart ; class sex; model status=weight height sex/dist=binary corrb covb; run; ... View more

Thank yo so much Steve for helping me with this and for the valuable information.  Ok, I used the log statement before the "cards" statement so it worked. If I use the log-transformed Firmness in genmod with a normal dist.  I get AIC: 2.94 and BIC: 17.24. With a normal dist. and no transformation I get 738 and 752.  With a gamma dist. I get 898 and 913.  So I guess I will choose the log-transformed variable with the normal dist. in genmod.    To the second point on making predictions: What do I have to include in my model in order to fit a first degree linear model? What would be the code for that?   Thank you very much ... View more

If race3 was considered as a fixed effect, I would expect the values to sum to zero. In this case, I think it will depend on the distribution of responses to each race, which will differ by hospital (as I see that not all races are observed at every hospital). So perhaps the variances and covariances for each hospital are not homogeneous, and could be fit with a GROUP= option. You would likely need two runs (one homogeneous, ignoring race3 and one heterogeneous, accommodating race3 differences). This approach has the advantage of being applicable to the fixed levels of race3, rather than assuming that the levels were selected randomly from a population. I would try fitting the following:   /* Homogeneous variance-covariance across race3 */ proc glimmix data=dataset method=laplace; class dshospid race3; model died (event='1') = var1 var2 var3 . . . var10 / dist=binary link=logit solution; random intercept / subject=dshospid type=un s; run; /* Heterogeneous variance-covariance across race3 */ proc glimmix data=dataset method=laplace; class dshospid race3; model died (event='1') = var1 var2 var3 . . . var10 / dist=binary link=logit solution; random intercept / subject=dshospid type=un s group=race3; covtest homogeneity; run; SteveDenham   ... View more

Thanks Steve and no problem. I think I'm mostly just trying to wrap my head around the fundamentals here (I've exceeded my own knowledge with this question to be honest). I'm incorporating the weight, strata, and cluster variables in my proc surveyfreq's because the dataset instructions tell me to, but I'm not 100% sure what those variables are actually doing. I understand weighting, clustering, and stratification as concepts, but haven't really been able to wrap my head around what they do/how they function practically speaking (i.e., in SAS). Maybe I don't need to, but I do want to make sure I'm not overlooking something important.   Your earlier point about subsampling throwing off the weighting was a great one and was the perfect example of something I'm worried about overlooking (that hadn't even occurred to me). At first, it made me pause and think I shouldn't be trying to force pairwise comparison - that I should just let it go and go the regression route - but then I realized I'm basically subsampling already by opting to run all of my analyses on the 18-65 year old segment of the survey respondents. Meaning that if I'm tracking your earlier point, the weighting is probably already thrown off to some degree (assuming the survey's weighting strategy was to correlate the whole sample with known population data).   I suppose now all I'm really trying to do is understand how big of a deal that is/would be. If you have any thoughts on this front, I'd appreciate them. And if not, no worries - I appreciate all the help and clarity so far. It's been a real boon. ... View more