I am off to New York City for two days, giving a seminar at Columbia tomorrow and visiting Andrew Gelman there. My talk will be about testing as mixture estimation, with slides similar to the Nice ones below if slightly upgraded and augmented during the flight to JFK. Looking at the past seminar speakers, I noticed we were three speakers from Paris in the last fortnight, with Ismael Castillo and Paul Doukhan (in the Applied Probability seminar) preceding me. Is there a significant bias there?!
Archive for Andrew Gelman
no starch press (!) sent me the pdf version of this incoming book, Statistics done wrong, by Alex Reinhart, towards writing a book review for CHANCE, and I read it over two flights, one from Montpellier to Paris last week, and from Paris to B’ham this morning. The book is due to appear on March 16. It expands on a still existing website developed by Reinhart. (Discussed a year or so away on Andrew’s blog, most in comments, witness Andrew’s comment below.) Reinhart who is, incidentally or not, is a PhD candidate in statistics at Carnegie Mellon University. After apparently a rather consequent undergraduate foray into physics. Quite an unusual level of maturity and perspective for a PhD student..!
“It’s hard for me to evaluate because I am so close to the material. But on first glance it looks pretty reasonable to me.” A. Gelman
Overall, I found myself enjoying reading the book, even though I found the overall picture of the infinitely many mis-uses of statistics rather grim and a recipe for despairing of ever setting things straight..! Somehow, this is an anti-textbook, in that it warns about many ways of applying the right statistical technique in the wrong setting, without ever describing those statistical techniques. Actually without using a single maths equation. Which should be a reason good enough for me to let all hell break loose on that book! But, no, not really, I felt no compunction about agreeing with Reinhart’s warning and if you have reading Andrew’s blog for a while you should feel the same…
“Then again for a symptom like spontaneous human combustion you might get excited about any improvement.” A. Reinhart (p.13)
Maybe the limitation in the exercise is that statistics appears so much fraught with dangers of over-interpretation and false positive and that everyone (except physicists!) is bound to make such invalidated leaps in conclusion, willingly or not, that it sounds like the statistical side of Gödel’s impossibility theorem! Further, the book moves from recommendation at the individual level, i.e., on how one should conduct an experiment and separate data for hypothesis building from data for hypothesis testing, to a universal criticism of the poor standards of scientific publishing and the unavailability of most datasets and codes. Hence calling for universal reproducibility protocols that reminded of the directions explored in this recent book I reviewed on that topic. (The one the rogue bird did not like.) It may be missing on the bright side of things, for instance the wonderful possibility to use statistical models to produce simulated datasets that allow for an evaluation of the performances of a given procedure in the ideal setting. Which would have helped the increasingly depressed reader in finding ways of checking how wrongs things could get..! But also on the dark side, as it does not say much about the fact that a statistical model is most presumably wrong. (Maybe a physicist’s idiosyncrasy!) There is a chapter entitled Model Abuse, but all it does is criticise stepwise regression and somehow botches the description of Simpson’s paradox.
“You can likely get good advice in exchange for some chocolates or a beer or perhaps coauthorship on your next paper.” A. Reinhart (p.127)
The final pages are however quite redeeming in that they acknowledge that scientists from other fields cannot afford a solid enough training in statistics and hence should hire statisticians as consultants for the data collection, analysis and interpretation of their experiments. A most reasonable recommendation!
Marco Banterle, Clara Grazian, Anthony Lee, and myself just arXived our paper “Accelerating Metropolis-Hastings algorithms by delayed acceptance“, which is an major revision and upgrade of our “Delayed acceptance with prefetching” paper of last June. Paper that we submitted at the last minute to NIPS, but which did not get accepted. The difference with this earlier version is the inclusion of convergence results, in particular that, while the original Metropolis-Hastings algorithm dominates the delayed version in Peskun ordering, the later can improve upon the original for an appropriate choice of the early stage acceptance step. We thus included a new section on optimising the design of the delayed step, by picking the optimal scaling à la Roberts, Gelman and Gilks (1997) in the first step and by proposing a ranking of the factors in the Metropolis-Hastings acceptance ratio that speeds up the algorithm. The algorithm thus got adaptive. Compared with the earlier version, we have not pursued the second thread of prefetching as much, simply mentioning that prefetching and delayed acceptance could be merged. We have also included a section on the alternative suggested by Philip Nutzman on the ‘Og of using a growing ratio rather than individual terms, the advantage being the probability of acceptance stabilising when the number of terms grows, with the drawback being that expensive terms are not always computed last. In addition to our logistic and mixture examples, we also study in this version the MALA algorithm, since we can postpone computing the ratio of the proposals till the second step. The gain observed in one experiment is of the order of a ten-fold higher efficiency. By comparison, and in answer to one comment on Andrew’s blog, we did not cover the HMC algorithm, since the preliminary acceptance step would require the construction of a proxy to the acceptance ratio, in order to avoid computing a costly number of derivatives in the discretised Hamiltonian integration.
Patrick Rubin-Delanchy and Daniel Lawson [of Warhammer fame!] recently arXived a paper we had discussed with Patrick when he visited Andrew and I last summer in Paris. The topic is the evaluation of the posterior predictive probability of a larger discrepancy between data and model
which acts like a Bayesian p-value of sorts. I discussed several times the reservations I have about this notion on this blog… Including running one experiment on the uniformity of the ppp while in Duke last year. One item of those reservations being that it evaluates the posterior probability of an event that does not exist a priori. Which is somewhat connected to the issue of using the data “twice”.
“A posterior predictive p-value has a transparent Bayesian interpretation.”
Another item that was suggested [to me] in the current paper is the difficulty in defining the posterior predictive (pp), for instance by including latent variables
which reminds me of the multiple possible avatars of the BIC criterion. The question addressed by Rubin-Delanchy and Lawson is how far from the uniform distribution stands this pp when the model is correct. The main result of their paper is that any sub-uniform distribution can be expressed as a particular posterior predictive. The authors also exhibit the distribution that achieves the bound produced by Xiao-Li Meng, Namely that
where P is the above (top) probability. (Hence it is uniform up to a factor 2!) Obviously, the proximity with the upper bound only occurs in a limited number of cases that do not validate the overall use of the ppp. But this is certainly a nice piece of theoretical work.
After a rather intense period of new simulations and versions, Juong Een (Kate) Lee and I have now resubmitted our paper on (some) importance sampling schemes for evidence approximation in mixture models to Bayesian Analysis. There is no fundamental change in the new version but rather a more detailed description of what those importance schemes mean in practice. The original idea in the paper is to improve upon the Rao-Blackwellisation solution proposed by Berkoff et al. (2002) and later by Marin et al. (2005) to avoid the impact of label switching on Chib’s formula. The Rao-Blackwellisation consists in averaging over all permutations of the labels while the improvement relies on the elimination of useless permutations, namely those that produce a negligible conditional density in Chib’s (candidate’s) formula. While the improvement implies truncated the overall sum and hence induces a potential bias (which was the concern of one referee), the determination of the irrelevant permutations after relabelling next to a single mode does not appear to cause any bias, while reducing the computational overload. Referees also made us aware of many recent proposals that conduct to different evidence approximations, albeit not directly related with our purpose. (One was Rodrigues and Walker, 2014, discussed and commented in a recent post.)
In the latest Sunday Review of the New York Times, the Norwegian novelist Jo Nesbo has a tribune on revenge against misdeeds and law as institutionalized revenge. Somewhat hidden in the current justifications of the legal system(s). (As an aside, he mentions the example of the Icelandic Alþingi where justice was dispensed once a year, resulting in beheadings, stake burnings, and drowning in the pond depicted above…) This came a few days after another tribune on a similar topic by Charles Blow, following the “botched Oklahoma execution of Clayton Lockett”, entitled “Eye-for-eye incivility” (an understatement if any!), and arguing about the economic inefficiency of the death penalty. Besides the basic moral quandaries of taking someone else’s life, perfectly summarised by Franquin in the following dark strip:
This sequence of tribunes links to one of my pet theories, which is that imprisonment is the most inadequate way of addressing crime and law breaking in (modern?) societies. Setting fully aside the moral notions of revenge and punishment, which aim more at the victim or victim’s relatives than at the perpetrator, and of redemption and remorse, which are at best hypothetical and inspired by religious considerations, I do wonder why economists have not tried to come up with more rational and game-theoretic ways of dealing with law-breakers than locking them up all together and expecting them to behave forever after the end of their term. More globally, I find it quite surprising that no one ever seems to question the very notion of sending people to jail. Indeed, it does bring any clear benefit to society as a whole. One of the usual arguments I receive in those occasions is that imprisonment keeps dangerous people away. But that seems a fairly weak notion: (i) most violent offenders are not dangerous in an absolute berserker sense but only because local circumstances made them violent at a given occurrence in space and time, (ii) those offenders are only put away for a while (in most civilised countries), (iii) they are not getting any less dangerous while in prison, and (iv) it does not apply to the vast majority of people jailed. Furthermore, from a pure offer-versus-demand perspective, this may be counterproductive: e.g., putting some thieves away in jail for a while simply gives an opportunity to other thieves to take advantage of the “thieving market”.
The Freakonomics blog has some entries on the topic—somewhat supportive of my notion that most criminals act in an overall rational way for which incentives and decentives could be considered—, but still fails to address the larger picture… I showed this post to Andrew who pointed me (of course!) to his blog, as several entries therein also consider the issue, like this one on the puzzles of criminal justice. Or prison terms for financial fraud? But I would push the argument further and call for an ultimate abolishment of the carceral system, seeking efficient and generalised alternatives to imprisonment. As detailed in this U.N. report I just came across. As I think a time will come when imprisonment will be seen as irrational as witch-burning is considered today.
Here is the second part of my review of Gelman et al.’ Bayesian Data Analysis (third edition):
“When an iterative simulation algorithm is “tuned” (…) the iterations will not in general converge to the target distribution.” (p.297)
Part III covers advanced computation, obviously including MCMC but also model approximations like variational Bayes and expectation propagation (EP), with even a few words on ABC. The novelties in this part are centred at Stan, the language Andrew is developing around Hamiltonian Monte Carlo techniques, a sort of BUGS of the 10’s! (And of course Hamiltonian Monte Carlo techniques themselves. A few (nit)pickings: the book advises important resampling without replacement (p.266) which makes some sense when using a poor importance function but ruins the fundamentals of importance sampling. Plus, no trace of infinite variance importance sampling? of harmonic means and their dangers? In the Metropolis-Hastings algorithm, the proposal is called the jumping rule and denoted by Jt, which, besides giving the impression of a Jacobian, seems to allow for time-varying proposals and hence time-inhomogeneous Markov chains, which convergence properties are much hairier. (The warning comes much later, as exemplified in the above quote.) Moving from “burn-in” to “warm-up” to describe the beginning of an MCMC simulation. Being somewhat 90’s about convergence diagnoses (as shown by the references in Section 11.7), although the book also proposes new diagnoses and relies much more on effective sample sizes. Particle filters are evacuated in hardly half-a-page. Maybe because Stan does not handle particle filters. A lack of intuition about the Hamiltonian Monte Carlo algorithms, as the book plunges immediately into a two-page pseudo-code description. Still using physics vocabulary that put me (and maybe only me) off. Although I appreciated the advice to check analytical gradients against their numerical counterpart.
“In principle there is no limit to the number of levels of variation that can be handled in this way. Bayesian methods provide ready guidance in handling the estimation of the unknown parameters.” (p.381)
I also enjoyed reading the part about modes that stand at the boundary of the parameter space (Section 13.2), even though I do not think modes are great summaries in Bayesian frameworks and while I do not see how picking the prior to avoid modes at the boundary avoids the data impacting the prior, in fine. The variational Bayes section (13.7) is equally enjoyable, with a proper spelled-out illustration, introducing an unusual feature for Bayesian textbooks. (Except that sampling without replacement is back!) Same comments for the Expectation Propagation (EP) section (13.8) that covers brand new notions. (Will they stand the test of time?!)
“Geometrically, if β-space is thought of as a room, the model implied by classical model selection claims that the true β has certain prior probabilities of being in the room, on the floor, on the walls, in the edge of the room, or in a corner.” (p.368)
Part IV is a series of five chapters about regression(s). This is somewhat of a classic, nonetheless Chapter 14 surprised me with an elaborate election example that dabbles in advanced topics like causality and counterfactuals. I did not spot any reference to the g-prior or to its intuitive justifications and the chapter mentions the lasso as a regularisation technique, but without any proper definition of this “popular non-Bayesian form of regularisation” (p.368). In French: with not a single equation! Additional novelty may lie in the numerical prior information about the correlations. What is rather crucially (cruelly?) missing though is a clearer processing of variable selection in regression models. I know Andrew opposes any notion of a coefficient being exactly equal to zero, as ridiculed through the above quote, but the book does not reject model selection, so why not in this context?! Chapter 15 on hierarchical extensions stresses the link with exchangeability, once again. With another neat election example justifying the progressive complexification of the model and the cranks and toggles of model building. (I am not certain the reparameterisation advice on p.394 is easily ingested by a newcomer.) The chapters on robustness (Chap. 17) and missing data (Chap. 18) sound slightly less convincing to me, esp. the one about robustness as I never got how to make robustness agree with my Bayesian perspective. The book states “we do not have to abandon Bayesian principles to handle outliers” (p.436), but I would object that the Bayesian paradigm compels us to define an alternative model for those outliers and the way they are produced. One can always resort to a drudging exploration of which subsample of the dataset is at odds with the model but this may be unrealistic for large datasets and further tells us nothing about how to handle those datapoints. The missing data chapter is certainly relevant to such a comprehensive textbook and I liked the survey illustration where the missing data was in fact made of missing questions. However, I felt the multiple imputation part was not well-presented, fearing readers would not understand how to handle it…
“You can use MCMC, normal approximation, variational Bayes, expectation propagation, Stan, or any other method. But your fit must be Bayesian.” (p.517)
Part V concentrates the most advanced material, with Chapter 19 being mostly an illustration of a few complex models, slightly superfluous in my opinion, Chapter 20 a very short introduction to functional bases, including a basis selection section (20.2) that implements the “zero coefficient” variable selection principle refuted in the regression chapter(s), and does not go beyond splines (what about wavelets?), Chapter 21 a (quick) coverage of Gaussian processes with the motivating birth-date example (and two mixture datasets I used eons ago…), Chapter 22 a more (too much?) detailed study of finite mixture models, with no coverage of reversible-jump MCMC, and Chapter 23 an entry on Bayesian non-parametrics through Dirichlet processes.
“In practice, for well separated components, it is common to remain stuck in one labelling across all the samples that are collected. One could argue that the Gibbs sampler has failed in such a case.” (p.535)
To get back to mixtures, I liked the quote about the label switching issue above, as I was “one” who argued that the Gibbs sampler fails to converge! The corresponding section seems to favour providing a density estimate for mixture models, rather than component-wise evaluations, but it nonetheless mentions the relabelling by permutation approach (if missing our 2000 JASA paper). The section about inferring on the unknown number of components suggests conducting a regular Gibbs sampler on a model with an upper bound on the number of components and then checking for empty components, an idea I (briefly) considered in the mid-1990’s before the occurrence of RJMCMC. Of course, the prior on the components matters and the book suggests using a Dirichlet with fixed sum like 1 on the coefficients for all numbers of components.
“14. Objectivity and subjectivity: discuss the statement `People tend to believe results that support their preconceptions and disbelieve results that surprise them. Bayesian methods tend to encourage this undisciplined mode of thinking.’¨ (p.100)
Obviously, this being a third edition begets the question, what’s up, doc?!, i.e., what’s new [when compared with the second edition]? Quite a lot, even though I am not enough of a Gelmanian exegist to produce a comparision table. Well, for a starter, David Dunson and Aki Vethtari joined the authorship, mostly contributing to the advanced section on non-parametrics, Gaussian processes, EP algorithms. Then the Hamiltonian Monte Carlo methodology and Stan of course, which is now central to Andrew’s interests. The book does include a short Appendix on running computations in R and in Stan. Further novelties were mentioned above, like the vision of weakly informative priors taking over noninformative priors but I think this edition of Bayesian Data Analysis puts more stress on clever and critical model construction and on the fact that it can be done in a Bayesian manner. Hence the insistence on predictive and cross-validation tools. The book may be deemed somewhat short on exercices, providing between 3 and 20 mostly well-developed problems per chapter, often associated with datasets, rather than the less exciting counter-example above. Even though Andrew disagrees and his students at ENSAE this year certainly did not complain, I personally feel a total of 220 exercices is not enough for instructors and self-study readers. (At least, this reduces the number of email requests for solutions! Esp. when 50 of those are solved on the book website.) But this aspect is a minor quip: overall this is truly the reference book for a graduate course on Bayesian statistics and not only Bayesian data analysis.