Watch this Ask the Expert session to learn the BGLIMM procedure, allowing you to transition your models to the Bayesian realm.
Watch the Webinar
You will:
-
Discover the difference between using PROC MCMC and PROC BGLIMM for Bayesian analysis.
-
Explore the diagnostic output provided for Bayesian analysis.
-
Determine how a Bayesian approach to analysis can be helpful.
The questions from the Q&A segment held at the end of the webinar are listed below and the slides from the webinar are attached.
Q&A
Now that you mentioned it, could we ask the developers to add the REPEATED statement to the GLIMMIX procedure?
We could, but I don't think that will happen...
Is there no multinomial and ordered multinomial distribution in BGLIMM?
As it is right now, it is not present. They are always adding more options to Bayesian elements as time passes. They started with the core type of distributions. It could likely show up later. Most people deal with binomial. You can make suggestions to the development group.
Are the distributions restricted to the exponential family?
Yes.
For those cases where we need to use PROC MCMC, can BGLIMM be used to export the design matrix? This is because PROC MCMC does not have a CLASS statement.
I don't believe that BGLIMM has an OUTDESIGN option. GLIMMIX does, so use that.
Except for the fixed effects, the estimates are close. What about the VC Site?
It could be a sample size issue. Look out for asymptotic issues. My Bayesian solution might be better.
Does it make sense to have negative age range in the HPD interval? Help me understand that.
Remember that this parameter is a regression term. The inclusion of negative numbers in the interval means that there is a chance that the relationship between this predictor and the response is negative. If age goes up, I expect the response to go down.
Does BGLIMM have a Firth option for logistic regression?
Not right now. What the best thing I would suggest is to go to MCMC and write that out in terms of coding. Again, that is another thing that I do have on the list over to the developers. It's going to be based on their timeline
What release of SAS/STAT will PROC BGLIMM be available in?
It's been available since 15.1.
For effects importance, is it possible to do Bayes factors with BGLIMM?
A former colleague of mine postulated that GL Models are not reliable because of the data manipulation. It essentially invalidates the procedure. Thoughts on that?
I haven't really gotten too much into that data manipulation viewpoint. The one thing that I love about the Bayesian twist is you can do a lot of sensitivity analysis. And sensitivity analysis is what really makes classical statisticians feel a little bit more comfortable with what's going on inside of the Bayesian realm. If you could send me any type of references or any type of papers, that would be great and then I could probably try to chime in on that. I'm not aware of that argument or that conversation.
How to know when to do Bayesian Analysis? What are the conditions in data that lead to Bayesian Analysis?
There are really no conditions in the data other than the possibility of the asymptotics breaking down. But again, the way that you know that you're going to want to possibly move into the Bayesian realm is the type of questions that you're answering. Do you want to ask things that are more probabilistic? Do you want to be able to speak about a credible interval rather than a confidence interval? What if you have a subject area expert that's been working on this for 30 years? They're telling me something, I want to include it. Classical statisticians process with data only. Times where I would start a move to Bayesian would be to include this additional information with the data.
Can I use PROC BGLIMM in SAS 9.4?
Absolutely. If you have 15.1 or more recent.
For your Jupyter notebook, are you using the same official SAS kernel available on GitHub? Or is there an updated version?
I'm using Viya for Learners on mine for that. In terms of the SAS kernel, I do know that there has been some SAS kernel usage of being able to utilize outside and connecting from Viya to Viya a different way, not in a Viya for Learners territory. That is something that we could probably ask to a few people, or I would suggest go to the communities' pages in the SAT in the SAS communities area. You might see people asking questions there that might have already been answered with that.
https://blogs.sas.com/content/sgf/2022/04/18/saspy-and-the-sas-kernel-for-jupyter-notebook/
Can you share the code?
I have a PDF of this entire presentation and the Jupyter notebook. Those will be provided to everybody. You can copy and paste any of the code that I used, and I even included the data for you. I know that you really love to be able to recreate exactly what I was doing here. So, I wanted to make sure that you had that.
Will you teach a class to BGLIMM?
Our current Bayesian class does not have BGLIMM in it. It was written before BGLIMM came out; I am in the works of updating that right now. My goal from the end of this year to the beginning of next year is I am going to try to reorganize our introductory Bayesian class to have a chapter dedicated to BGLIMM. The developers want me to do this, I've wanted to do this, and I'm finally going to try to structure that out and work through that type of an idea. So, keep an ear out and watch for our Bayesian course that's currently available. When you see the description change, that is when that will be ready to go.
Recommended Resources
Please see additional resources in the attached slide deck.
Want more tips? Be sure to subscribe to the Ask the Expert board to receive follow up Q&A, slides and recordings from other SAS Ask the Expert webinars.
