learn more about statistics

I have been stalling a lot lately in my writing on this subject—-it seems that I am getting distracted by the minor points, and not writing enough on the major ones. For the next few days I want to focus on writing the more substantial stuff, even if the drafts suck. This requires letting go of the grand organizing principles and focusing on writing smaller parts even if I end up doing it differently at the end.

Old programmer’s Maxim: “Plan to throw one away; you will, anyhow.” —Fred Brooks, The Mythical Man-Month.

Well I agreed to teach a graduate introduction to statistics this coming Fall, so that should count for something here. I am also teaching an upper division class in experimental design/analysis focused on ANOVAs and some introductory non-parametric designs for experimental analysis.

So progress. The class notes I am writing up are massive, too. Might be a book when I am done. This goal may be marked done this Fall.

There are a bunch of minimizers in R. Gotta try them all to finish this one problem I have been working on. Then I will have finished my first serious real-world maximum-likelihood problem. I will also be in a position to finalize some more of this software package I am developing for release to the community.

Alas, I will not be able to attend the very cool class this upcoming quarter as I am just too damn busy to sit through it. Hopefully I will move somewhere else where stats classes are available to audit. I want to take a class on dependencies in data someday.

This most recent class was good, in that I plan on studying more on the advanced theory of the linear model and its extensions.

Well the class I sat in on this quarter has been pretty good. Not sure how much I will remember in 6 months, but it did give me some good solid foundations for further studies of Generalized Linear Models. I think I need to read a good book on mathematical statistics sometime. And re-read my big fat book on linear models.

Neural networks seem to be involved.

Huge list of ones and zeros describing an individual unit. About 100 individual units being observed/measured, perhaps 100,000 ones and zeros per unit. Need a soft modeling approach to use these 100,000 ones and zeros and 100 individual units to predict a set of measured (i.e. real) numbers (also from the 100 individuals).

Now that is a problem.

Well, I guess I now understand some of the basics of PLS (partial least-squares regression) but I am not sure if there is a point in learning all the details. I had a theory about how it related to other methods, but now it seems that the work is done or wrong. :-(

Need to look for something else or see something I am not seeing yet.

I suspect that some relationship exists among all three of ICA, PCA, and PLS. I think it allows me to use one to place bounds on the others. If the result is unknown, I want to mathematically establish it. If it is known, I want to publish it within the context of the field in which I work.

If that sounds like gibberish, then just replace it with “I have found a statistical project to work on.” :-)

I read an article on maximum likelihood versus least-squares today, now on to some papers about AIC/BIC and model selection!