Visiting scholar in Bocconi, Milan

I am currently visiting Bocconi University for two weeks (following Pierre’s visit to Vancouver). The team here is very active in Bayesian nonparametrics, no surprise with Italy being the cradle of exchangeability with de Finetti’s work. I talk with Sonia Petrone and Pietro Muliere, and can interact with many PhD students in topics related with mine (Maria Anna di Lucca, Steffen Ventz and Sara Wade among others). The whole team plans to attend to this summer BNP conference in Veracruz, Mexico, for which I submitted a poster abstract today on Multidimensional covariate dependent Dirichlet processes. By the way, I am looking for someone for sharing a (cheaper rate) double room in Veracruz…
Bocconi’s building is very sophisticated, though looking like a bunker from outside, the architect had the good idea to separate the rooms with glass walls. Which makes it possible, and quite nice, to scribble equations all around on the wall. By chance I will give a talk for the PhD student seminar next Friday. I shall present our work with Robin and Nicolas on the Estimation of a regression for rank data with an application to the Eurovision Song Contest.

Update on the parallel IMH article

Hey,

Last October, I blogged about an article written by Christian P. Robert, Murray Smith and myself about parallel computation, Independent Metropolis-Hastings and Rao-Blackwellization.

The article advocates the use of parallel computation and the method described, called “block IMH”, can be done fully in parallel, which makes the whole thing pretty much costless compared to standard Independent Metropolis-Hastings, while significantly decreasing the variance of any estimator based on the generated Markov Chain.

Just a quick update: since then, the reviewers from JCGS got back to us with useful comments and we made a second version, available on arXiv as well. It’s hopefully clearer and the graphs are more compact and informative (thank you, my dear dear ggplot2 package).

Also I’ve put the code online on Google Code:

https://code.google.com/p/py-block-imh/

It’s not a very user-friendly package, still it allows to reproduce the graphs shown in the article and would perhaps help people who want to try the method out!

Opinion polls for the presidential elections… with margins of error!

In the last few days, a lot of opinion polls have been released about the next presidential elections in France, to be held in April and May 2012. They feature three oponents, Nicolas Sarkozy, National Front’s Marine Le Pen, and one of the Socialist Party leaders, either Dominique Strauss-Kahn, Martine Aubry, Ségolène Royal or François Hollande. Here are some of the results cited in a recent Le Monde article

Harris Interactive:

- Marine Le Pen 24 %, DSK 23 %, Nicolas Sarkozy 21 %
- Marine Le Pen 24 %, Nicolas Sarkozy 21 %, François Hollande 20 %
- Marine Le Pen 23 %, Nicolas Sarkozy 21 %, Martine Aubry 21 %.

IFOP:

- DSK 29 %, Nicolas Sarkozy 23 %, Marine Le Pen 21 %
- Martine Aubry 24 %, Nicolas Sarkozy 24 %, Marine Le Pen 22 %
- Nicolas Sarkozy 24 %, François Hollande 23 %, Marine Le Pen 22 %
- Nicolas Sarkozy 24 %, Marine Le Pen 22 %, Ségolène Royal 19 %.

A crazy thing is that most attention of the debates is focused on the rank of the candidates, but a look on margins of error shows that the relative positions are not significant in most of the cases.

A standard opinion poll has a sample of size n=1000. Let the true score of a candidate be , the choice of voter i be (=1 if intends to vote for the candidate, 0 otherwise), their mean . Then the 95%-confidence interval is approximately (which means that it contains the computed score with a probability of  (approximately) 95%).

The score p is around 20%, so the margin of error is . Most of the polls cannot really tell who is gonna reach the second round…