who’s afraid of the big B wolf?

Aris Spanos just published a paper entitled “Who should be afraid of the Jeffreys-Lindley paradox?” in the journal Philosophy of Science. This piece is a continuation of the debate about frequentist versus llikelihoodist versus Bayesian (should it be Bayesianist?! or Laplacist?!) testing approaches, exposed in Mayo and Spanos’ Error and Inference, and discussed in several posts of the ‘Og. I started reading the paper in conjunction with a paper I am currently writing for a special volume in  honour of Dennis Lindley, paper that I will discuss later on the ‘Og…

“…the postdata severity evaluation (…) addresses the key problem with Fisherian p-values in the sense that the severity evaluation provides the “magnitude” of the warranted discrepancy from the null by taking into account the generic capacity of the test (that includes n) in question as it relates to the observed data”(p.88)

First, the antagonistic style of the paper is reminding me of Spanos’ previous works in that it relies on repeated value judgements (such as “Bayesian charge”, “blatant misinterpretation”, “Bayesian allegations that have undermined the credibility of frequentist statistics”, “both approaches are far from immune to fallacious interpretations”, “only crude rules of thumbs”, &tc.) and rhetorical sleights of hand. (See, e.g., “In contrast, the severity account ensures learning from data by employing trustworthy evidence (…), the reliability of evidence being calibrated in terms of the relevant error probabilities” [my stress].) Connectedly, Spanos often resorts to an unusual [at least for statisticians] vocabulary that amounts to newspeak. Here are some illustrations: “summoning the generic capacity of the test”, ‘substantively significant”, “custom tailoring the generic capacity of the test”, “the fallacy of acceptance”, “the relevance of the generic capacity of the particular test”, yes the term “generic capacity” is occurring there with a truly high frequency.

estimating a constant (not really)

Larry Wasserman wrote a blog entry on the normalizing constant paradox, where he repeats that he does not understand my earlier point…Let me try to recap here this point and the various comments I made on StackExchange (while keeping in mind all this is for intellectual fun!)

The entry is somehow paradoxical in that Larry acknowledges (in that post) that the analysis in his book, All of Statistics, is wrong. The fact that “g(x)/c is a valid density only for one value of c” (and hence cannot lead to a notion of likelihood on c) is the very reason why I stated that there can be no statistical inference nor prior distribution about c: a sample from f does not bring statistical information about c and there can be no statistical estimate of c based on this sample. (In case you did not notice, I insist upon statistical!)

To me this problem is completely different from a statistical problem, at least in the modern sense: if I need to approximate the constant c—as I do in fact when computing Bayes factors—, I can produce an arbitrarily long sample from a certain importance distribution and derive a converging (and sometimes unbiased) approximation of c. Once again, this is Monte Carlo integration, a numerical technique based on the Law of Large Numbers and the stabilisation of frequencies. (Call it a frequentist method if you wish. I completely agree that MCMC methods are inherently frequentist in that sense, And see no problem with this because they are not statistical methods. Of course, this may be the core of the disagreement with Larry and others, that they call statistics the Law of Large Numbers, and I do not. This lack of separation between both notions also shows up in a recent general public talk on Poincaré’s mistakes by Cédric Villani! All this may just mean I am irremediably Bayesian, seeing anything motivated by frequencies as non-statistical!) But that process does not mean that c can take a range of values that would index a family of densities compatible with a given sample. In this Monte Carlo integration approach, the distribution of the sample is completely under control (modulo the errors induced by pseudo-random generation). This approach is therefore outside the realm of Bayesian analysis “that puts distributions on fixed but unknown constants”, because those unknown constants parameterise the distribution of an observed sample. Ergo, c is not a parameter of the sample and the sample Larry argues about (“we have data sampled from a distribution”) contains no information whatsoever about c that is not already in the function g. (It is not “data” in this respect, but a stochastic sequence that can be used for approximation purposes.) Which gets me back to my first argument, namely that c is known (and at the same time difficult or impossible to compute)!

Let me also answer here the comments on “why is this any different from estimating the speed of light c?” “why can’t you do this with the 100th digit of π?” on the earlier post or on StackExchange. Estimating the speed of light means for me (who repeatedly flunked Physics exams after leaving high school!) that we have a physical experiment that measures the speed of light (as the original one by Rœmer at the Observatoire de Paris I visited earlier last week) and that the statistical analysis infers about c by using those measurements and the impact of the imprecision of the measuring instruments (as we do when analysing astronomical data). If, now, there exists a physical formula of the kind

where φ is a probability density, I can imagine stochastic approximations of c based on this formula, but I do not consider it a statistical problem any longer. The case is thus clearer for the 100th digit of π: it is also a fixed number, that I can approximate by a stochastic experiment but on which I cannot attach a statistical tag. (It is 9, by the way.) Throwing darts at random as I did during my Oz tour is not a statistical procedure, but simple Monte Carlo à la Buffon…

Overall, I still do not see this as a paradox for our field (and certainly not as a critique of Bayesian analysis), because there is no reason a statistical technique should be able to address any and every numerical problem. (Once again, Persi Diaconis would almost certainly differ, as he defended a Bayesian perspective on numerical analysis in the early days of MCMC…) There may be a “Bayesian” solution to this particular problem (and that would nice) and there may be none (and that would be OK too!), but I am not even convinced I would call this solution “Bayesian”! (Again, let us remember this is mostly for intellectual fun!)

structure and uncertainty, Bristol, Sept. 26

Another day full of interesting and challenging—in the sense they generated new questions for me—talks at the SuSTain workshop. After another (dry and fast) run around the Downs; Leo Held started the talks with one of my favourite topics, namely the theory of g-priors in generalized linear models. He did bring a new perspective on the subject, introducing the notion of a testing Bayes factor based on the residual statistic produced by a classical (maximum likelihood) analysis, connected with earlier works of Vale Johnson. While I did not truly get the motivation for switching from the original data to this less informative quantity, I find this perspective opening new questions for dealing with settings where the true data is replaced with one or several classical statistics. With possible strong connections to ABC, of course. Incidentally, Leo managed to produce a napkin with Peter Green’s intro to MCMC dating back from their first meeting in 1994: a feat I certainly could not reproduce (as I also met both Peter and Leo for the first time in 1994, at CIRM)… Then Richard Everit presented his recent JCGS paper on Bayesian inference on latent Markov random fields, centred on the issue that simulating the latent MRF involves an MCMC step that is not exact (as in our earlier ABC paper for Ising models with Aude Grelaud). I already discussed this paper in an earlier blog and the only additional question that comes to my mind is whether or not a comparison with the auxiliary variable approach of Møller et al. (2006) would make sense.

In the intermission, I had a great conversation with Oliver Ratman on his talk of yesterday on the surprising feature that some models produce as “data” some sample from a pseudo-posterior.. Opening once again new vistas! The following talks were more on the mathematical side, with James Cussens focussing on the use of integer programming for Bayesian variable selections, then Éric Moulines presenting a recent work with a PhD student of his on PAC-Bayesian bounds and the superiority of combining experts. Including a CRAN package. Éric concluded his talk with the funny occurence of Peter’s photograph on Éric’s Microsoft Research Profile own page, due to Éric posting our joint photograph at the top of Pic du Midi d’Ossau in 2005… (He concluded with a picture of the mountain that was the exact symmetry of mine yesterday!)

The afternoon was equally superb with Gareth Roberts covering fifteen years of scaling MCMC algorithms, from the mythical 0.234 figure to the optimal temperature decrease in simulated annealing, John Kent playing the outlier with an EM algorithm—however including a formal prior distribution and raising the challenge as to why Bayesians never had to constrain the posterior expectation, which prompted me to infer that (a) the prior distribution should include all constraints and (b) the posterior expectation was not the “right” tool in non-convex parameters spaces—. Natalia Bochkina presented a recent work, joint with Peter Green, on connecting image analysis with Bayesian asymptotics, reminding me of my early attempts at reading Ibragimov and Has’minskii in the 1990′s. Then a second work with Vladimir Spoikoini on Bayesian asymptotics with misspecified models, introducing a new notion of effective dimension. The last talk of the day was by Nils Hjort about his coming book on “Credibility, confidence and likelihood“—not yet advertised by CUP—which sounds like an attempt at resuscitating Fisher by deriving distributions in the parameter space from frequentist confidence intervals. I already discussed this notion in an earlier blog, so I am fairly skeptical about it, but the talk was representative of Nils’ highly entertaining and though-provoking style! Esp. as he sprinkled the talk with examples where MLE (and some default Bayes estimators) did not work. And reanalysed one of Chris Sims‘ example presented during his Nobel Prize talk…

PLoS topic page on ABC

A few more comments on the specific entry on ABC written by Mikael Sunnåker et al…. The entry starts with the representation of the posterior probability of an hypothesis, rather than with the posterior density of a model parameter, which seems to lead the novice reader astray. After all, (a) ABC was not introduced for conducting model choice and (b) interchanging hypothesis and model means that the probability of an hypothesis H as used in the entry is actually the evidence in favour of the corresponding model. (There are a few typos and grammar mistakes, but I assume either PLoS or later contributors will correct those.) When the authors state that the “outcome of the ABC rejection algorithm is a set of parameter estimates distributed according to the desired posterior distribution”, I think they are misleading the readers as they forget the “approximative” aspect of this distribution. Further below, I would have used the title “Insufficient summary statistics” rather than “Sufficient summary statistics”, as it spells out more clearly the fundamental issue with the potential difficulty in using ABC. (And I am not sure the subsequent paragraph on “Choice and sufficiency of summary statistics” should bother with the sufficiency aspects… It seems to me much more relevant to assess the impact on predictive performances.)

Although this is most minor, I would not have made mention of the (rather artificial) “table for interpretation of the strength in values of the Bayes factor (…) originally published by Harold Jeffreys^[6] “. I obviously appreciate very much that the authors advertise our warning about the potential lack of validity of an ABC based Bayes factor! I also like the notion of “quality control”, even though it should only appear once. And the pseudo-example is quite fine as an introduction, while it could be supplemented with the outcome resulting from a large n, to be compared with the true posterior distribution. The section “Pitfalls and remedies” is remarkable in that it details the necessary steps for validating a ABC implementation: the only entry I would remove is the one about “Prior distribution and parameter ranges”, in that this is not a problem inherent to ABC… (Granted, the authors present this as a “general risks in statistical inference exacerbated in ABC”, which makes more sense!) It may be that the section on the non-zero tolerance should emphasize more clearly the fact that ε should not be zero. As discussed in the recent Read Paper by Fearnhead and Prangle when envisioning ABC as a non-parametric method of inference.

At last, it is always possible to criticise the coverage of the historical part, since ABC is such a recent field that it is constantly evolving. But the authors correctly point out to (Don) Rubin on the one hand and to Diggle and Graton on the other. Now, I would suggest adding in this section links to the relevant softwares like our own DIY-ABC…

(Those comments have also been posted on the PLoS Computational Biology wiki.)

still confronting intractability in Bristol…

Another definitely interesting and intense day at the Confronting Intractability in Statistical Inference workshop in Bristol: all talks there had a high informational content for me and even those I had heard previously [in no time difference and hence much less chances of my dozing during talks, which, alas!, now gets into an almost certainty for US conferences!) For instance, I am still coming to terms with Gareth’s importance sampling for continuous diffusions. (This was the first time I was hearing Arnaud’s talk on the estimation of the score vector and I definitely to hear it again, given its technicality!) Sumeet Singh gave a talk mixing ABC with maximum likelihood estimation for HMMS, in connection with his earlier paper, and I got more convince  by the idea of using a sequence of balls for keeping pseudo-data close to the true data when I realised it could be implemented sequentially. Nial Friel’s talk on the double intractable likelihoods was covering graphical models and social network models, maybe calling for a comparison with ABC, as done in the recent paper by Richard Everitt. I had too many slides and thus presumably failed to deliver an intelligible message about the selection of ABC summary statistics for testing, even though the population genetics new illustration presumably helped. In connection with our ABC paper, Dennis Prangle and Paul Fernhead presented a poster on using the Bayes factor as a summary statistics in this setup, in the spirit of their Read Paper of last December. And Richard Wilkinson concluded the day with a more philosophical talk on the dual nature of ABC inference, in a quite pleasant perspective (that related to the way ABC was received by econometricians during my talk in Princeton last week). The day ended up quite pleasantly in a south-Indian thali restaurant, a good preparation for Glasgow’s Ashoka tomorrow night!