Why had I not read this already?

Ah, this is the "Join Bayes Nets" paper, reworked...

Gelman vs. Feller.

"I’ll always remember what [Blackwell] told me when I handed him a draft of my thesis. “The best thing about Bayesians is that they’re always right.”"

"Edward Prescott forms a noteworthy exception: under the rubric of "calibration", he has elevated his conviction that his prior guesses are never wrong into a new principle of statistical estimation." -- Cosma's snark about statisticians and economists is the funniest snark around. We could all aspire to have wits that were as subtle and dry as his... (seriously).

"... and many statisticians are uncomfortable with shrinkage." -- (C'mon, Gelman, go for the easy joke!)

One user's notes...

For you to mull over...

"In this paper we derive one- and two-sample multivariate empirical Bayes statistics (the MB-statistics) to rank genes in order of interest from longitudinal replicated developmental microarray time course experiments."

"When deciding between simpler and more complex hypotheses, I generally prefer the more complex hypothesis. When I choose the simpler hypothesis, I view this as a combination of labor-saving device and approximate Bayes, pooling a parameter estimate all the way to zero instead of merely pooling it most of the way. I certainly don't see Bayes factors having any relevance, given the oft-noted problem that Bayes factors can depend decisively on aspects of the prior distribution that have no influence on the posterior distribution under each of the individual models." -- That's right: *all* the tea.

File the comments under, "Department of Missing the Point."

"Motivated by a computer experiment for the design of a rocket booster, this paper explores nonstationary modeling methodologies that couple stationary Gaussian processes with treed partitioning." -- Via a commenter at Andrew Gelman's blog.

"I think part of the problem is that reports of experimental results should not be aimed at presenting conclusions, as may seem most natural from a Bayesian viewpoint, but rather at providing the information with which the readers may draw conclusions. This may be the source of some objections to the prior distribution in Bayesian analysis, which can be seen as corrupting the objective presentation of the experimental results, even though frequentist methods like p-values are not suitable presentations either."

A follow-up to that Gigerenzer and Hoffrage paper!

For an idea about probabilistic models of path connectivity...

Nice title.

Saw Mark Johnson give a talk about "Adaptor Grammars" (man, that 'o' really bothers me) two days ago. It turned out to be ... an extremely boring talk, although the idea itself seems modestly interesting and it included several reasonable animations of hierarchical Chinese Restaurant processes that were modestly illuminating. At any rate, I sat in the back, doodled on my notebook, and started to idly wonder if issues of "frequentist consistency" for this sort of learning process had been examined (or were even worth examining) at all...

"Anyway, I posted the above discussion (basically, all except for the previous two paragraphs, to their blog and got the strangest comments. Not that people were saying anything wrong, just they were coming from a traditional theoretical computer science perspective. For them, Bayesian statistics is all about code lengths; for me it's all about hierarchical models. Which I guess is consistent with my original point. Still, it's frustrating for me (but perhaps frustrating to some of these people from the other side, that statisticians see Bayes as about models rather than philosophy and code lengths). I thought that communicating with econometricians and non-Bayesian statisticians was tough, but this is a whole new level of difficulty!"

A reading list that goes through (it appears) 2007, hitting most of the high points -- broad, but not overly deep.

A basic review of naive Bayes modeling, bag of words representations, and all that.

"Unit-testing for Bayesian models."

"I’m interested in this beyond the application to word prediction because it is relevant to the general normalization problem: If you want to predict the probability of one of a large number of events, often you must compute a predicted score for all the events and then normalize, a computationally inefficient operation. The problem comes up in many places using probabilistic models, but I’ve run into it with high-dimensional regression. There are a couple workarounds for this computational bug: (1) approximate, (2) avoid, (3) [what this paper does] use a self-normalizing structure."

Q-C's publications -- I wish I could find a table of contents for his new book, though.

A model for social data and tagging, yes. But not a *social* model, and certainly not something that models the social aspect of the tagging. It seems like an obvious generalization, no?

"Now, your model might choose not to make recommendations with controversial movies - but this won't help you on Netflix Prize - you're forced to make errors even when you know you're making them. (R)MSE is pre-probabilistic: it gives no advantage to a probabilistic model that's aware of its own uncertainty."

A two-page writeup of his BUGS models.

Assuming that two events are independent when they really aren't *isn't* the same thing as "ignoring Bayes theorem."

Andrew Gelman's new paper on election prediction using state-level polling. Partial pooling from multilevel models.

I read this, and I think: double-counting. (Right?)

I need to move fast(er).

Linked to by Andrew Gelman, I think. "Effectively, BART is a nonparametric Bayesian regression approach which uses dimensionally adaptive random basis elements. Motivated by ensemble methods in general, and boosting algorithms in particular, BART is defined by a statistical model: a prior and a likelihood. This approach enables full posterior inference including point and interval estimates of the unknown regression function as well as the marginal effects of potential predictors."

"...high-throughput devices, such as microarrays, routinely require simultaneous hypothesis tests for thousands of individual cases, not at all what the classical theory had in mind. In these situations empirical Bayes information begins to force itself upon frequentists and Bayesians alike. The two-groups model is a simple Bayesian construction that facilitates empirical Bayes analysis. This article concerns the interplay of Bayesian and frequentist ideas in the two-groups setting, with particular attention focused on Benjamini and Hochberg's False Discovery Rate method."

Gelman's response to his own April Fools joke (in the journal of Bayesian Analysis no less), and its responses. "In a nutshell: Bayesian statistics is about making probability statements, frequentist statistics is about evaluating probability statements."

Video. Via Andrew Gelman's blog. Dynamic programming equations on the third slide -- nice.

"JAGS is Just Another Gibbs Sampler. It is a program for analysis of Bayesian hierarchical models using Markov Chain Monte Carlo (MCMC) simulation not wholly unlike BUGS."

What *would* "topic models for the sake of topic models" look like? "I must publish a paper on topic models, so that no one forgets the name 'Dirichlet'?"

"The key thing in the early chapters is to not obsess on the question of 'where do the priors come from.' They're just models, they come from the same place that likelihoods come from."

The New Scientist is totally worthless. I need to filter my "Overcoming Bias" feed through Pipes, so that I get all the Hanson but none of the Yudkowsky. Seriously!

JMLR 9(Apr):759--813, 2008.

Daphne Koller gets written up in the NYT. Apparently because she just got an award from the ACM?

Gelman is like the anti-Yudkowsky. It's humorous.

How many cultures must we ascribe buffets to? Stop the madness!