Dear all,
I wanted to run a model in proc mixed with an arh(1) repeated structure from several years ago, but couldn't. After much thinking on why it was not running, the only difference I could indicate is that the model then was run on sas 9.2 and now on sas 9.3. I checked this on another computer that has 9.2 and the model was able to run. I've seen this issue as well in the topic: Question re: Consistency of GLIMMIX results between 9.3 and 9.4 versions https://communities.sas.com/t5/SAS-Statistical-Procedures/Question-re-Consistency-of-GLIMMIX-results...
in which one version a covariance parameter estimate is very close to 0 compared to the others and in the other version it is 0. I have the exact same thing where in sas 9.2 there is an estimate for a covariance parameter estimate (close to 0; 0.002; no positive definte warnings/notes) and in sas 9.3 it specifies it as 0 and is stopping because of an infinite likelihood. I understand in my data and model why it could have difficulty in estimating it and solved it on 9.3, that's not the issue now. I'm wondering how much is this an issue for such differences between sas versions where one does not indicate a note or warning and on the other it does?
This happens more often than SAS would like to omit (in my opinion). The problem occurs with difficult data sets (difficult for the model that is being used). It is especially an issue with nonlinear variance-covariance structures, such as arh(1). Usually, this means that the model is overparameterized for the data being analyzed. Usually you see the difference when one compares 32 with 64 bit versions of SAS. In principle, one should get the same results with each new version of SAS/STAT if one stayed with the same 32 or 64 bit machine and software (I realize that all newest versions of SAS are now only for 64 bit machines). But there are plenty of examples where with the same 64 bit machine, one gets different results with different releases of SAS/STAT. One should not think of this as one being right and the other being wrong. When this happens, I think all results are suspect -- the model cannot be justified for the data set being analyzed. In principle, the newest version of SAS/STAT should be the best, since the developers are always improving the code. Perhaps the lack of warning in the earlier release was a fault of that release. With that said, I have some data sets where I think an earlier release worked better.
An infiinite likelihood can occur for a number of reasons. A common culprit is when you have duplicate values for a repeated measure, the lowest level in the hierarchy (usually because of coding errors with the data).. This would have been picked up in any release of the software. So, this is probably due to something else.
This happens more often than SAS would like to omit (in my opinion). The problem occurs with difficult data sets (difficult for the model that is being used). It is especially an issue with nonlinear variance-covariance structures, such as arh(1). Usually, this means that the model is overparameterized for the data being analyzed. Usually you see the difference when one compares 32 with 64 bit versions of SAS. In principle, one should get the same results with each new version of SAS/STAT if one stayed with the same 32 or 64 bit machine and software (I realize that all newest versions of SAS are now only for 64 bit machines). But there are plenty of examples where with the same 64 bit machine, one gets different results with different releases of SAS/STAT. One should not think of this as one being right and the other being wrong. When this happens, I think all results are suspect -- the model cannot be justified for the data set being analyzed. In principle, the newest version of SAS/STAT should be the best, since the developers are always improving the code. Perhaps the lack of warning in the earlier release was a fault of that release. With that said, I have some data sets where I think an earlier release worked better.
An infiinite likelihood can occur for a number of reasons. A common culprit is when you have duplicate values for a repeated measure, the lowest level in the hierarchy (usually because of coding errors with the data).. This would have been picked up in any release of the software. So, this is probably due to something else.
@lvm I'll add add that sometime changes occur even when SAS doesn't change any of its code. Major releases (like from 9.2 to 9.3) are when SAS updates compilers and other software development tools. Because optimizing compilers change the instruction set for optimized assembly, and because optimization is inherently a nonlinear dynamical system that exhibits sensitive dependence on initia... small changes in the way that the compilers optimize code can result in different results from MLE methods. As you say, when this happens "all results are suspect."
That said, most differences are probably the result of changes in SAS code, including fixing bugs, changing to better algorithms, and so forth.
Ha ha, Rick. You almost said chaotic behavior there, which I have tried to explain to QA types for 20 years without success. "It's the same data and the same model, so if you two (people/systems/whateever) don't get the same answer, then I have to assume that the results are not valid."
I'll pick up on what Larry said--this happens a lot more when I try to fit an overly complicated model to the data.
Steve Denham
Build your skills. Make connections. Enjoy creative freedom. Maybe change the world. Registration is now open through August 30th. Visit the SAS Hackathon homepage.
Register today!ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.
Find more tutorials on the SAS Users YouTube channel.