Here is an interesting StackOverflow question, about a model for data distributed as the sum of two uniforms with unknown support. I was surprised how hard it was for me.
I think the future of probabilistic programming should be to make a model for this easy to code.
I’ve had no teaching responsibilities over the last quarter, and I must miss it. I’ve found myself responding to PyMC questions on StackOverflow more than ever before. It is an interesting window into what is hard in Bayesian computation. Checking (and achieving) MCMC convergence is one thing that is hard. Here are some questions and answers that include it:
A question and answer on CrossValidated, which make me reflect on the danger of knowing enough statistics to be dangerous.
Last week, we read Chien et al, Efﬁcient mapping and geographic disparities in breast cancer mortality at the county-level by race and age in the U.S. I’ve been very interested in these sort of “small-areas” spatial statistical methods recently, so it was good to see what is out there as the state of the art. I think I’ve got something to contribute along these lines some day soon.
Maps in small multiples look just lovely, too:
It has been a while since I visited my pymc-examples repository, but I got a request there a few weeks ago about the feasibility of upgrading the Seeds Example of a random effects logistic regression model for PyMC3. It turns out that this was not very time consuming, which must mean I’m starting to understand the changes between PyMC2 and PyMC3.
See them side-by-side here (PyMC2) and here (PyMC3).
Thomas Richardson gave a recent seminar at IHME about how the potential outcomes crowd can make sense of graphical models and vice versa. It also has a CSSS working paper to complement it, a trend in our recent seminars: http://www.csss.washington.edu/Papers/wp128.pdf
This talk had some great graphics, as I would hope for in a talk on graphical modeling:
A recent IHME seminar by Mark Wheldon described a Bayesian approach to estimating past populations and vital rates by age. I like this stuff. The talk is online, and there is a CSSS working paper on it, too: http://www.csss.washington.edu/Papers/wp117.pdf
Last week we read Repeat Bone Mineral Density Screening and Prediction of Hip and Major Osteoporotic Fracture by Berry et al. It argues that repeat screening does not improve predictions. But I think the world needs a better way to measure the quality of predictions like these. Area-under-the-curve doesn’t cut it when you are predicting the unpredictable.
For our final journal club paper of last semester, we read Meat, fish, and esophageal cancer risk: a systematic review and dose-response meta-analysis. I am a vegetarian.
That brings me up to date for fall quarter, and the winter quarter is just finishing its first week. Good!
I just went through the classic paper on WinBUGS, which might or might not be called probabilistic programming. It is listed on the probabilistic programming resource page, and it is certainly interesting. The WinBUGS “hello, world” is a linear regression model: