After over a quarter of a century of experience working as a statistical consultant in a wide array of settings, it’s safe to say that a large proportion of the statistics presentations cover topics I have been over before. Still, even if it is a technique I’ve used many times, I almost always come across some bit of information I didn’t know.
For example, every graduate student learns that when you have repeated continuous measures, like a test score, you use repeated measures ANOVA. You can do PROC GLM or PROC MIXED.
When you have a categorical dependent variable, you use logistic regression.
Well, what if you have a repeated measure on a categorical variable? I haven’t done that kind of design. If I’m doing an outcome study it’s usually with something like mortality or graduation as the dependent variable and we do a one-time analysis. In the events that subjects are measured repeated times for something like health or education studies, I’d do a survival analysis looking at how long the person lived or stayed in school.
Logistic regression assumes you have independent observations, but that is not always the case. Certainly if you have measured the same person multiple times, the observations are likely to be correlated.
Correlated measures aren’t only on one person over multiple time points. It could be that you are measuring several people in the same family to see if they have a certain illness.
Although this hasn’t come up yet (and it kind of surprises me, now that I think about it), if it DOES, I can use PROC GENMOD to do a logistic regression with correlated data, as I learned in this presentation by Dachao Liu.
I mentioned this to a couple of people who were surprised I did not know this already, but honestly, it had just never come up.
I guess if you had asked me if JMP did non-linear models I would have said, “I assume so” but I wouldn’t know for sure. Now I know for sure. I had kind of forgotten that JMP does some pretty complex models. If you’d asked me if you could save your prediction formula in JMP, I would say, “I assume so,” but again, I wouldn’t have known for sure. As you can see, I don’t use JMP all that much, but it is good to know since I have the occasional client using it.
Even though I won’t be using JMP for this next project, the presentation by McCormack got me to thinking about saving equations and gave me some really good ideas about running the analyses nightly against the larger database to create the equation and then applying it on the fly when users access our application.
So, in short, at SAS Global Forum, I learned some things I didn’t know, was confirmed on some things I was pretty sure I knew and remembered to use some things I already knew.
Was it worth spending four days in Orlando?
Since after thirty years I am still learning, yes. Yes, it was.