Why the puppies?

The posterior distribution is a compromise between the likelihood and the prior: The half-up ears are a compromise between the perky ears and the floppy ears. The marginal likelihood, a.k.a. evidence, is not needed in MCMC methods, so it gets sleepy with nothing much to do.

If the puppies bother you, a solution can be found at this other blog post.

Talk and Workshop at Vassar College, March 23-24

I'll be doing a talk and workshop at Vassar College. Details can be found here.


A list of future and past workshops can be found here. 

(Poster composed entirely
by folks at Vassar.)

Hierarchical Diagram for Pseudoprior Model in Ch. 10

Chapter 10 of the book describes the use of pseudopriors in model comparison to help the model-index parameter be sampled more efficiently (with less clumpy autocorrelation). The method was discussed without a diagram to illustrate where the pseudoprior has its effect in the model structure. This blog entry provides a diagram.

The extended example in Chapter 10 examined two models of the filtration-condensation data. One model used a separate κ_c for each condition. The second model used a single κ₀ parameter for all conditions. The program that implements the model comparison is called FilconModelCompPseudoPriorJags.R (or FilconModelCompPseudoPriorBrugs.R originally). The diagram shown above extends Figures 9.15 and 9.17 of the book, to include the two models' priors on κ. See the top of the diagram, which forks under the modelIdx into gamma distributions for κ_c or κ₀. The main thing that you have to do, that is not shown in the diagram, is imagine how the shape of the gamma distribution changes depending on the state of the model index. This is suggested by the text beside the gamma distributions. When the model index is 1, then the left gamma distribution is actually being used to model the data, and the real prior constants are used for it. On the other hand, when the model index is 1, then the right gamma distribution is not being used to model the data, and the pseudoprior constants are used for it, to keep the κ₀ value in a reasonable range. See the two variations below, that illustrate when the model index is 1 and 2, respectively:

The Book visits Bayes

On his recent visit to Bunhill Fields, Mark Andrews snapped a photo of the book at Bayes' tomb:

Many thanks to Mark!
And see the book at Fisher's remains, too.

Talk and Workshop at the Eastern Psychological Association, March 2 - 3

I'll be doing a talk and workshop at the meeting of the Eastern Psychological Association. Details can be found here.
A list of future and past workshops can be found here.

Complete steps for installing software and programs

November 29, 2014:
For how to get the software, please see the book's web site here (https://sites.google.com/site/doingbayesiandataanalysis/software-installation).

Jocular Disbelief

Posted by Yann LeCun  -  Jan 19, 2012  -  Public Google+

Joke of the day (true story, circa 2004):
Radford Neal (giving a talk): I don't necessarily think that the Bayesian method is the best thing to do in all cases...
Geoff Hinton: Sorry Radford, my prior probability for you saying this is zero, so I couldn't hear what you said.