Here is [yet!] another Bayesian textbook that appeared recently. I read it in the past few days and, despite my obvious biases and prejudices, I liked it very much! It has a lot in common (at least in spirit) with our Bayesian Core, which may explain why I feel so benevolent towards Bayesian ideas and data analysis. Just like ours, the book by Ron Christensen, Wes Johnson, Adam Branscum, and Timothy Hanson is indeed focused on explaining the Bayesian ideas through (real) examples and it covers a lot of regression models, all the way to non-parametrics. It contains a good proportion of WinBugs and R codes. It intermingles methodology and computational chapters in the first part, before moving to the serious business of analysing more and more complex regression models. Exercises appear throughout the text rather than at the end of the chapters. As the volume of their book is more important (over 500 pages), the authors spend more time on analysing various datasets for each chapter and, more importantly, provide a rather unique entry on prior assessment and construction. Especially in the regression chapters. The author index is rather original in that it links the authors with more than one entry to the topics they are connected with (Ron Christensen winning the game with the highest number of entries). Continue reading
Archive for October, 2011
Due to a new law introduced last May by the French government, it has now become almost impossible for foreign non-EU students who graduate from a French business (e.g., HEC or ESSEC) or engineer (e.g., Polytechnique) school, or from a university, to get a job in France after graduation, even with a firm offer from a company. (This post may sound like a strange complaint since, in some countries, a student visa prohibits its holder to get a permanent job without first exiting the country. But this was not the case in France till last May.) Indeed, those non-EU (post)graduates with a job offer need to apply to local administrations who decide whether or not the job fits a need and whether or not it could not be offered to a French national. (As if those local administrations had the proper expertise.) The procedure takes months, during which the (post)graduates cannot work. Months for no reason other than the administrations being understaffed. And in most cases the answer is no. Meaning these (post)graduates then have to leave the country within a month. And cannot apply to a student visa without first leaving the country…
This sudden change of policy has been heavily discussed in the national and international press (chinese version), on blogs, and by student and professional organisations: I cannot but join the flow of protests against this iniquitous, absurd, and counter-productive action, dictated by electoral motives catering to the rightmost (or just plain xenophobic) part of the electorate. It is counter-productive in that most of those students have been trained in elite public schools, meaning their training has been mostly supported by the State (i.e. the French taxpayer), which would only benefit from the input of highly qualified (post)graduates to the French economy. It is absurd in that those non-EU (post)graduates number in the thousands, hence are unlikely to make a dent in the immigration figures used to frighten the electorate. It is counter-productive because it sends the wrong message to potential students abroad and will thus lower the attractivity of French higher education, an attractivity which is already under pressure from competing countries like Canada and Australia (which just went ahead of France in terms of foreign students). It is absurd since the [former Education and currently Budget] Minister, Valérie Pécresse, has publicly written to the Minister of Interior to ask him to abolish a procedure “going the wrong way”. It is counter-productive because these students graduate from schools (HEC, Polytechnique, Essec, Mines, Ensae, &tc.) where there are more job offers than candidates with the proper training. So the typical xenophobic rethoric of “foreigners stealing jobs from nationals” falls completely off the mark there, even though it was instrumental in passing this law… Now, it is quite probable this law will not survive the elections next May, but le mal sera fait (in terms of attractivity)… Note that postdocs are not impacted by the procedure!
A few weeks ago, I finished the fifth volume of George Martin, A Dance with Dragons, I had bought in Lancaster last summer but could not carry with me to the US (and onto the boat!). It reads wonderfully, just like the previous volumes, and so I wonder why it took the author so long to produce it. (He apologizes about this in the preface to the book. But does not [have to] provide reasons.) Esp. when considering that the story constitutes the “other side” of the previous volume, covering characters and regions that were omitted in the fourth book. Even though the pace is sometimes a wee slow (e.g., the coverage of Tyrion’s travel and mishaps and of his every thought!, or of Daenerys’ procrastination and hesitations), again, it is very pleasant to read. I am actually surprised at how easy it is to launch back into the complex geography and geopolitics of Martin’s universe, given the five year gap with my reading the previous volume. The important and consequential action has to wait a while, but things are moving fast by the end of the book, with surprising and permanent changes of dominance and of rulers. It is a good thing that Martin is eliminating some of his characters as it means he cannot go for ever in writing small prints about them! On another level, it is quite interesting to spot so many readers of the first volume (A Game of Thrones), in the metro and in airports, clearly generated by the TV adaptation on HBO…
I just got the “news” that Dennis Ritchie died, although this happened on October 12… The announcement was surprisingly missing from my information channels and certainly got little media coverage, compared with Steve Jobs‘ demise. (I did miss the obituaries in the New York Times and in the Guardian. The Economist has the most appropriate heading, printf(“goodbye, Dennis”); !!!) Still, Dennis Ritchie contributed to computer science to extents comparable to Steve Jobs’, if on a lesser commercial plane: he is a founding father of both the C language and the Unix operating system. I remember spending many days perusing over his reference book, The C programming language, co-written with Brian Kernighan. (I kept trying programming in C until Olivier Cappé kindly pointed out to me that I was merely translating my Pascal vision into C code, missing most of the appeal of the language!) And, of course, I also remember discovering Unix when arriving at Purdue as a logical and much more modern operating system: just tfour years after programming principal components on punched card and in SAS, this was a real shock! I took a few evening classes at Purdue run by the Computer Department and I still carry around the Purdue University UNIX Pocket Guide. Although I hardly ever use it, it is there on the first shelf on top of my desk… As is The C programming language even though I have not opened it in years!
So we (geeks, computer users, Linuxians, R users, …) owe a lot to Dennis Ritchie and it is quite sad both that he passed away by himself and that his enormous contribution was not better acknowledged. Thus, indeed,
for (i=0; i<ULONG_LONG_MAX; i++) printf("thanks a lot, Dennis")
Jan Hanning kindly sent me this email about several difficulties with Chapters 3, Monte Carlo Integration, and 5, Monte Carlo Optimization, when teaching out of our book Monte Carlo Statistical Methods [my replies in italics between square brackets, apologies for the late reply and posting, as well as for the confusion thus created. Of course, the additional typos will soon be included in the typo lists on my book webpage.]:
- I seem to be unable to reproduce Table 3.3 on page 88 – especially the chi-square column does not look quite right. [No, they definitely are not right: the true χ² quantiles should be 2.70, 3.84, and 6.63, at the levels 0.1, 0.05, and 0.01, respectively. I actually fail to understand how we got this table that wrong...]
- The second question I have is the choice of the U(0,1) in this Example 3.6. It feels to me that a choice of Beta(23.5,18.5) for p1 and Beta(36.5,5.5) for p2 might give a better representation based on the data we have. Any comments? [I am plainly uncertain about this... Yours is the choice based on the posterior Beta coefficient distributions associated with Jeffreys prior, hence making the best use of the data. I wonder whether or not we should remove this example altogether... It is certainly "better" than the uniform. However, in my opinion, there is no proper choice for the distribution of the pi's because we are mixing there a likelihood-ratio solution with a Bayesian perspective on the predictive distribution of the likelihood-ratio. If anything, this exposes the shortcomings of a classical approach, but it is likely to confuse the students! Anyway, this is a very interesting problem.]
- My students discovered that Problem 5.19 has the following typos, copying from their e-mail: “x_x” should be “x_i” [sure!]. There are a few “( )”s missing here and there [yes!]. Most importantly, the likelihood/density seems incorrect. The normalizing constant should be the reciprocal of the one showed in the book [oh dear, indeed, the constant in the exponential density did not get to the denominator...]. As a result, all the formulas would differ except the ones in part (a). [they clearly need to be rewritten, sorry about this mess!]
- I am unsure about the if and only if part of the Theorem 5.15 [namely that the likelihood sequence is stationary if and only if the Q function in the E step has reached a stationary point]. It appears to me that a condition for the “if part” is missing [the "only if" part is a direct consequence of Jensen's inequality]. Indeed Theorem 1 of Dempster et al 1977 has an extra condition [note that the original proof for convergence of EM has a flaw, as discussed here]. Am I missing something obvious? [maybe: it seems to me that, once Q reaches a fixed point, the likelihood L does not change... It is thus tautological, not a proof of convergence! But the theorem says a wee more, so this needs investigating. As Jan remarked, there is no symmetry in the Q function...]
- Should there be a (n-m) in the last term of formula (5.17)? [yes, indeed!, multiply the last term by (n-m)]
- Finally, I am a bit confused about the likelihood in Example 5.22 [which is a capture-recapture model]. Assume that Hij=k [meaning the animal i is in state k at time j]. Do you assume that you observed Xijr [which is the capture indicator for animal i at time j in zone k: it is equal to 1 for at most one k] as a Binomial B(n,pr) even for r≠k? [no, we observe all Xijr's with r≠k equal to zero] The nature of the problem seems to suggest that the answer is no [for other indices, Xijr is always zero, indeed] If that is the case I do not see where the power on top of (1-pk) in the middle of the page 185 comes from [when the capture indices are zero, they do not contribute to the sum, which explains for this condensed formula. Therefore, I do not think there is anything wrong with this over-parameterised representation of the missing variables.]
- In Section 5.3.4, there seems to be a missing minus sign in the approximation formula for the variance [indeed, shame on us for missing the minus in the observed information matrix!]
- I could not find the definition of in Theorem 6.15. Is it all natural numbers or all integers? May be it would help to include it in Appendix B. [Surprising! This is the set of all positive integers, I thought this was a standard math notation...]
- In Definition 6.27, you probably want to say covering of A and not X. [Yes, we were already thinking of the next theorem, most likely!]
- In Proposition 6.33 – all x in A instead of all x in X. [Yes, again! As shown in the proof. Even though it also holds for all x in X]
Thanks a ton to Jan and to his UNC students (and apologies for leading them astray with those typos!!!)
As I was looking at the discussion paper by Yamin Yu and Xiao-Li Meng on improved efficiency for MCMC algorithms, which is available (for free) on-line, I realised the paper on parallel Metropolis-Hastings algorithm we wrote with Pierre Jacob and Murray Smith is now published in Journal of Computational and Graphical Statistics (on-line). This is a special issue for the 20th anniversary of the Journal of Computational and Graphical Statistics and our paper is within the “If Monte Carlo Be a Food of Computing, Simulate on” section! (My friends Olivier Cappé and Radu V. Craiu also have a paper in this issue.) Here is the complete reference:
P. Jacob, C. P. Robert, & M. H. Smith. Using Parallel Computation to Improve Independent Metropolis–Hastings Based Estimation. Journal of Computational and Graphical Statistics. September 1, 2011, 20(3): 616-635. doi:10.1198/jcgs.2011.10167
The [20th Anniversary Featured Discussion] paper by Yamin Yu and Xiao-Li Meng has already been mentioned on Andrew’s blog, it is full of interesting ideas and remarks about improving Gibbs efficiency, in the spirit of the very fine work Jim Hobert and his collaborators have been developing in the past decade, fun titles (“To center or not center – that is not the question”, “coupling is more promising than compromising”, “be all our insomnia remembered”, and “needing inception”, in connection with the talk Xiao-Li gave in Paris two months ago….), and above all the fascinating puzzle of linking statistical concepts and Monte Carlo concepts. How comes sufficiency and ancillarity are to play a role in simulation?! Where is the simulation equivalent of Basu’s theorem? These questions obviously relate to the idea of turning simulation into a measure estimation issue, discussed in a post of mine after the Columbia workshop. This interweaving paper also brings back memories of the fantastic Biometrika 1994 interleaving paper by Liu, Wong, and Kong, with its elegant proof of positive decreasing correlation and of improvement by Rao-Blackwellisation [another statistics theorem!] for data augmentation.