I’m speaking Monday 14 April at 4:30 on weakly informative prior distributions and models with interactions. I’ll try to make things accessible to a general audience of people who might not know much about statistics in general or Bayesian methods in particular.
Andrew,
It is not clear to many why interactions (non-additivity) are "bad". Could you expand on that? (I'd love to copy your stuff)
Bill,
Where did I say that interactions and nonadditivity are bad?
this is a weird, totally off-topic question.
how does one make a constraint matrix?
you seem to know a bit about stats ;) and i'm desperately searching for something that makes sense to a non-statistical person about these things. i'm trying to use an R function, QPmat, based on solve.QP and keep goofing it up. i was trying to find stuff about estimating theta and found your blog. obviously, i've got issues. :)
well, sorry for wasting your time, but i'm grasping at straws.
john
Andrew,
You're right, you never said "bad". I should have asked "what are the consequences of ignoring them?"
I didn't like the striptease display (incremental revelation of the content on each slide). Your audience may run out of the auditorium screaming (apologies to Tufte).
Why not show everything that's on one slide in one shot. I somehow doubt that the audience at MIT has such limited processing capacity that their visual attention needs to be controlled ;-).
Also, the graphics (presumably done in R) have not had justice done to them ("Regularization in Action!" slide). It looks like you did a screen grab of the graphic? You probably know this, but I would pipe the R output to a file using the pdf("filename.pdf") command and then includegraphics it in the Beamer source.
hth,
Bill,
One consequence of ignoring interactions is that you can miss the most important features of your data (as in our redistricting example).
John,
I have no idea what a constraint matrix is. You could try the R-help list.
Shravan,
(1) Clicking through the slide gives me something to do with my hands when I'm talking.
(2) Actually, I'm worried about the MIT audience. I have no idea how much or how little they know about statistics.
(3) Unfortunately, I was chopping up a graph that I'd already made awhile ago and using my meager image-processing skills.
Andrew,
I don't know about your audience, but I work in product development in the CPG (Consumer Packaged Goods) industry, so decisions can have a real impact (e.g. spend money for product roll-outs). The product developers (e.g. engineers) have a tendency to ignore interactions, and require constant reminders on "why interactions can be bad." Simply showing them that the IA exist is insufficient and leads to a "yeah but I'm special.." response. I repeatedly need to show them how this can hurt them, as in "your product improvement only works for people who already buy you and they won't buy any more." Lists of disasters are helpful here.