(miscellaneous bits of information, to be organized later)
- from Doug Bates, explaining why much of the p-value fussing is irrelevant for social scientists with large data sets:
Indeed. I realize that calculation of p-values is immensely important to some people but doing it properly for unbalanced designs is extremely complicated and, in the areas that I am most concerned with (large data sets with complex structure), it's a non-issue. When you have data on a half-million students in 2500 schools whatever calculation you use for the degrees of freedom will produce a number so large that the exact number is irrelevant.
- from Jane Elith:
- mention GAMMs? (More generally, this opens up the whole can of worms of using GEEs to fit GLMMs — how strongly can the analogy with PQL/MQL be drawn?)
last time I wanted to do GLMMs was very frustrated with the lack of info on model selection. Wanted to find something like the lasso, but implemented for GLMMs. Found J. Sunil Rao's website - looks like he's working on it ("fence" procedures). Would it be worth mentioning that area somewhere? - I hope that the lasso / related procedures will become more widely used once the software gets a bit further along. I like the concept.