In the first model, nothing seems to be happening--I wonder about the response variable and quasi-complete separation.  I still think you should run with the ITPRINT option, to look at what the likelihood is doing at each iteration. In the second model, the singularity problem rears its head.  Most likely, you have highly collinear variables in the model that convey almost exactly the same information.  What do crosstabs show?  I realize there are a lot of variables to check, but so long as the collinearity problem exists, the singularity problem is likely to arise. Finally, addressing the question Reeza posed regarding number of observations, I think you are probably adequate in the second model, but near the edge for the first: 629 obs distributed over 47 indicator variables should be adequate, but only if the obs are evenly distributed.  If there are many combinations that are extremely sparse, convergence will be a problem. To get to a workable situation, see if some of the variables can be eliminated.  Those that contain essentially identical information, based on crosstabs, can be combined. Good luck. Steve Denham ... View more

I would add /MAXITER=100 ITPRINT; to your model statement.  Examine the iteration history, and then you could address changing the convergence criteria depending on the behavior of the log likelihood.  However, just increasing the maximum number of iterations may solve your problems. It will not if the response surface is relatively flat and not well-characterized, i.e., you have 39 independent variables--the possibility of collinearity and lack of sufficient sample size to efficiently estimate all of these parameters can certainly lead to convergence problems. Good luck. Steve Denham ... View more

Same "answer" I gave in the SAS Procedures community, but I copied it here as well: You can probably get there with some programming steps to model the covariance structure, but I don't think it will be easy.  Using the mvn distribution in the RANDOM statement looks to me like the equivalent of an unstructured matrix in PROC MIXED, but programming to set parameters in the covariance matrix sigma to fit the specified error structure would be needed, and I think will take a lot of trial and error.  Google doesn't seem to have much in the way of help.  Perhaps there is an R package (probably nmle) that has the coding that could be adapted. Good luck on this one.  If you find a good answer, I really, really, really want to know, as I have over thirty studies where this could be applied. Steve Denham ... View more

You can probably get there with some programming steps to model the covariance structure, but I don't think it will be easy.  Using the mvn distribution in the RANDOM statement looks to me like the equivalent of an unstructured matrix in PROC MIXED, but programming to set parameters in the covariance matrix sigma to fit the specified error structure would be needed, and I think will take a lot of trial and error.  Google doesn't seem to have much in the way of help.  Perhaps there is an R package (probably nmle) that has the coding that could be adapted. Good luck on this one.  If you find a good answer, I really, really, really want to know, as I have over thirty studies where this could be applied. Steve Denham ... View more

If you are familiar with surveyreg, I think it is the right choice.  You have cluster variables, strata variables, a weight variable, and what looks like an interesting model.  The data were collected in a survey, and the inference space is the universe (population) where the survey was applied--not the typical mixed model inference space of all possible populations.  Consequently, the tests and confidence bounds produced by surveyreg are more likely to be correct for that particular population. Nickel says it runs without a memory problem, as well. Steve Denham ... View more

I probably have the cluster and the strata variables confused, so bear with me on this.  You have 377 buildingirn that are nested within ~100 strata09_n variables.  I assume that there are multiple records for each of the individual buildingirn's (probably individual student responses). If that is incorrect, then ignore everything from here on, and we'll try again after I get it straight. Given that my assumption is correct, then the best choice would look like: Proc glimmix data= a; *final model; class ctytype untx ow2  strata09_n buildingirn; model perfindex = lowincome racialmin untx ow2 ctytype enrollment3 / solution; weight finalweight; random intercept strata09_n / subject=buildingirn solution; lsmeans ow2/ilink cl; lsmeans untx/ilink cl; where sealprog = 0; *where sealprog = 1; run; I can't guarantee that this will get around the out of memory problem, though.  So you may have to consolidate some strata (which I am going to guess are school districts, and it is vital to get some info on the variability added there).  That may be difficult given your objectives.  You may need to look into PROC HPMIXED, which is designed to handle data with many random effects. But now I am going to ask some questions:  First, why PROC GLIMMIX rather than PROC SURVEYREG?  If you truly conducted a survey, the estimates from GLIMMIX may be biased, even with the weight statement.  Second, I see "ilink" options in the lsmeans statements.  Generally, I don't use this unless I have specified a distribution in the model that is other than normal or lognormal (which have identity links so that the ilink estimate is identical to the estimate).  Is the response variable what you would consider normal, or would another distribution be preferred? Third, I see "final model" in the comments.  I hope that this is the result of looking at the interactions between the continuous variables and the class variables to find a parsimonious analysis of covariance model.  Good luck with this. Steve Denham ... View more

I'll follow on to PGStats question.  I think for any record you have either (noReadm=0, yesReadm=1) or (noReadm=1, yesReadm=0).  If that is the case, what other characteristic would drive this?  Without another characteristic, the probability of noReadm=0 is exactly that of yesReadm=1, and is exactly equal to 1 minus the probability of noReadm=1 (and 1 minus the probability of yesReadm=0).  A second characteristic would lead to a 2 x 2 or a 2 x k table. If you want to test whether the proportion of noReadm is equal to a fixed value, you can do that with: proc freq data=yourdata; tables noReadm/chisq testp=0.5; run; This would test if the proportion of noReadm in your sample is equal to 0.5, using a chi-squared test. Steve Denham ... View more

Well you can still use the parameter estimates, just remember that a shift of 1 due to an X variable, holding all other variables constant, leads to a shift of exp(beta) in the y variable--if you use a log link as is standard for a gamma distribution.  I'll have to think about the dist=lognormal, as the link is the identity there.  And by think, I mean go plug in some data and look at the results. Hope this helps. Steve Denham ... View more

Not sure, but I am willing to bet that the Estimated G matrix is not positive definite message is due to collinearity for some subjects, so that a separate quadratic estimate does not "come out".  Are these repeated measures?  Is there a possibility of using R side measures?  Probably not, in the context that you would like some overall distribution of quadratic slopes.  Other possibilities: standardize agedx, so that some of the collinearity between the linear and quadratic terms is lessened. Good luck. Steve Denham ... View more

Since we are way outside of my field (biostatistics), there may be some terminology problems.  What do you mean when you say "fix the relativities of independent variables"?  If I knew or saw an example, then I might be able to come up with an analogous situation in my area. Steve Denham ... View more

Do not exponentiate any single parameter estimate.  The lognormal distribution refers to the distribution of the dependent variable, and so you would exponentiate the linear combination you get after plugging in appropriate values in the X matrix.  The only time I could understand exponentiating the intercept is if all independent variables are zero.  Given that, try centering any covariates and see what happens to the intercept.  It would be good to think of the intercept then as estimating the median.  See some recent posts in this community on GLIMMIX for more info. Steve Denham ... View more

To answer that is as much art as science, I am afraid.  You need to look at the processes that generate the independent variables--what sort of distribution results, are there collinearities or redundancies, are the variables straightforward so that a parametric distribution is appropriate or would a semiparametric (B spline) be better? A first step might be to look at the residuals after standardizing the independent variables--and I do mean LOOK.  A graphic presentation will tell you far more than anything else.  Plot, plot, and re-plot the data, so that you get a feel for what can be done. I am afraid I haven't given much of an answer, but then I don't really have a good idea of what your variables might be. Steve Denham ... View more