more air for MCMC

Aki Vehtari, Andrew Gelman, Dan Simpson, Bob Carpenter, and Paul-Christian Bürkner have just published a Bayesian Analysis paper about using an improved R factor for MCMC convergence assessment. From the early days of MCMC, convergence assessment has been a recurring (and recurrent!) question in the community. First leading to a flurry of proposals, [which Kerrie, Chantal, and myself reviewwwed in the Valencia 1998 proceedings], and then slowly disintegrating under the onslaughts of reality—i.e. that none could not be 100% foolproof in full generality—…. This included the (possibly now forgotten) single-versus-multiple-chains debate between Charlie Geyer [for single] and Andrew Gelman and Don Rubin [for multiple]. The later introduced an analysis-of-variance R factor, which remains quite popular up to this day, in part for being part of most MCMC software, like BUGS. That this R may fail to identify convergence issues, even in the more recent split version, does not come as a major surprise, since any situation with a long-term influence of the starting distribution may well fail to identify missing (significant) parts of the posterior support. (It is thus somewhat disconcerting to me to see that the main recommendation is to move the bound on R from 1.1 to 1.01, reminding me to some extent of a recent proposal to move the null rejection boundary from 0.05 to 0.005…) Similarly, the ESS may prove a poor signal for convergence or lack thereof, especially because the approximation of the asymptotic variance relies on stationarity assumptions. While multiplying the monitoring tools (as in CODA) helps with identifying convergence issues, looking at a single convergence indicator is somewhat like looking only at a frequentist estimator! (And with greater automation comes greater responsibility—in keeping a critical perspective.)

Looking for a broader perspective, I thus wonder at what we would instead need to assess the lack of convergence of an MCMC chain without much massaging of the said chain. An evaluation of the (Kullback, Wasserstein, or else) distance between the distribution of the chain at iteration n or across iterations, and the true target? A percentage of the mass of the posterior visited so far, which relates to estimating the normalising constant, with a relatively vast array of solutions made available in the recent years? I remain perplexed and frustrated by the fact that, 30 years later, the computed values of the visited likelihoods are not better exploited. Through for instance machine-learning approximations of the target. that could themselves be utilised for approximating the normalising constant and potential divergences from other approximations.

Computational Bayesian Statistics [book review]

This Cambridge University Press book by M. Antónia Amaral Turkman, Carlos Daniel Paulino, and Peter Müller is an enlarged translation of a set of lecture notes in Portuguese. (Warning: I have known Peter Müller from his PhD years in Purdue University and cannot pretend to perfect objectivity. For one thing, Peter once brought me frozen-solid beer: revenge can also be served cold!) Which reminds me of my 1994 French edition of Méthodes de Monte Carlo par chaînes de Markov, considerably upgraded into Monte Carlo Statistical Methods (1998) thanks to the input of George Casella. (Re-warning: As an author of books on the same topic(s), I can even less pretend to objectivity.)

“The “great idea” behind the development of computational Bayesian statistics is the recognition that Bayesian inference can be implemented by way of simulation from the posterior distribution.”

The book is written from a strong, almost militant, subjective Bayesian perspective (as, e.g., when half-Bayesians are mentioned!). Subjective (and militant) as in Dennis Lindley‘s writings, eminently quoted therein. As well as in Tony O’Hagan‘s. Arguing that the sole notion of a Bayesian estimator is the entire posterior distribution. Unless one brings in a loss function. The book also discusses the Bayes factor in a critical manner, which is fine from my perspective.  (Although the ban on improper priors makes its appearance in a very indirect way at the end of the last exercise of the first chapter.)

Somewhat at odds with the subjectivist stance of the previous chapter, the chapter on prior construction only considers non-informative and conjugate priors. Which, while understandable in an introductory book, is a wee bit disappointing. (When mentioning Jeffreys’ prior in multidimensional settings, the authors allude to using univariate Jeffreys’ rules for the marginal prior distributions, which is not a well-defined concept or else Bernardo’s and Berger’s reference priors would not have been considered.) The chapter also mentions the likelihood principle at the end of the last exercise, without a mention of the debate about its derivation by Birnbaum. Or Deborah Mayo’s recent reassessment of the strong likelihood principle. The following chapter is a sequence of illustrations in classical exponential family models, classical in that it is found in many Bayesian textbooks. (Except for the Poison model found in Exercise 3.3!)

Nothing to complain (!) about the introduction of Monte Carlo methods in the next chapter, especially about the notion of inference by Monte Carlo methods. And the illustration by Bayesian design. The chapter also introduces Rao-Blackwellisation [prior to introducing Gibbs sampling!]. And the simplest form of bridge sampling. (Resuscitating the weighted bootstrap of Gelfand and Smith (1990) may not be particularly urgent for an introduction to the topic.) There is furthermore a section on sequential Monte Carlo, including the Kalman filter and particle filters, in the spirit of Pitt and Shephard (1999). This chapter is thus rather ambitious in the amount of material covered with a mere 25 pages. Consensus Monte Carlo is even mentioned in the exercise section.

“This and other aspects that could be criticized should not prevent one from using this [Bayes factor] method in some contexts, with due caution.”

Chapter 5 turns back to inference with model assessment. Using Bayesian p-values for model assessment. (With an harmonic mean spotted in Example 5.1!, with no warning about the risks, except later in 5.3.2.) And model comparison. Presenting the whole collection of xIC information criteria. from AIC to WAIC, including a criticism of DIC. The chapter feels somewhat inconclusive but methinks this is the right feeling on the current state of the methodology for running inference about the model itself.

“Hint: There is a very easy answer.”

Chapter 6 is also a mostly standard introduction to Metropolis-Hastings algorithms and the Gibbs sampler. (The argument given later of a Metropolis-Hastings algorithm with acceptance probability one does not work.) The Gibbs section also mentions demarginalization as a [latent or auxiliary variable] way to simulate from complex distributions [as we do], but without defining the notion. It also references the precursor paper of Tanner and Wong (1987). The chapter further covers slice sampling and Hamiltonian Monte Carlo, the later with sufficient details to lead to reproducible implementations. Followed by another standard section on convergence assessment, returning to the 1990’s feud of single versus multiple chain(s). The exercise section gets much larger than in earlier chapters with several pages dedicated to most problems. Including one on ABC, maybe not very helpful in this context!

“…dimension padding (…) is essentially all that is to be said about the reversible jump. The rest are details.”

The next chapter is (somewhat logically) the follow-up for trans-dimensional problems and marginal likelihood approximations. Including Chib’s (1995) method [with no warning about potential biases], the spike & slab approach of George and McCulloch (1993) that I remember reading in a café at the University of Wyoming!, the somewhat antiquated MC³ of Madigan and York (1995). And then the much more recent array of Bayesian lasso techniques. The trans-dimensional issues are covered by the pseudo-priors of Carlin and Chib (1995) and the reversible jump MCMC approach of Green (1995), the later being much more widely employed in the literature, albeit difficult to tune [and even to comprehensively describe, as shown by the algorithmic representation in the book] and only recommended for a large number of models under comparison. Once again the exercise section is most detailed, with recent entries like the EM-like variable selection algorithm of Ročková and George (2014).

The book also includes a chapter on analytical approximations, which is also the case in ours [with George Casella] despite my reluctance to bring them next to exact (simulation) methods. The central object is the INLA methodology of Rue et al. (2009) [absent from our book for obvious calendar reasons, although Laplace and saddlepoint approximations are found there as well]. With a reasonable amount of details, although stopping short of implementable reproducibility. Variational Bayes also makes an appearance, mostly following the very recent Blei et al. (2017).

The gem and originality of the book are primarily to be found in the final and ninth chapter where four software are described, all with interfaces to R: OpenBUGS, JAGS, BayesX, and Stan, plus R-INLA which is processed in the second half of the chapter (because this is not a simulation method). As in the remainder of the book, the illustrations are related to medical applications. Worth mentioning is the reminder that BUGS came in parallel with Gelfand and Smith (1990) Gibbs sampler rather than as a consequence. Even though the formalisation of the Markov chain Monte Carlo principle by the later helped in boosting the power of this software. (I also appreciated the mention made of Sylvia Richardson’s role in this story.) Since every software is illustrated in depth with relevant code and output, and even with the shortest possible description of its principle and modus vivendi, the chapter is 60 pages long [and missing a comparative conclusion]. Given my total ignorance of the very existence of the BayesX software, I am wondering at the relevance of its inclusion in this description rather than, say, other general R packages developed by authors of books such as Peter Rossi. The chapter also includes a description of CODA, with an R version developed by Martin Plummer [now a Warwick colleague].

In conclusion, this is a high-quality and all-inclusive introduction to Bayesian statistics and its computational aspects. By comparison, I find it much more ambitious and informative than Albert’s. If somehow less pedagogical than the thicker book of Richard McElreath. (The repeated references to Paulino et al.  (2018) in the text do not strike me as particularly useful given that this other book is written in Portuguese. Unless an English translation is in preparation.)

Disclaimer: this book was sent to me by CUP for endorsement and here is what I wrote in reply for a back-cover entry:

An introduction to computational Bayesian statistics cooked to perfection, with the right mix of ingredients, from the spirited defense of the Bayesian approach, to the description of the tools of the Bayesian trade, to a definitely broad and very much up-to-date presentation of Monte Carlo and Laplace approximation methods, to an helpful description of the most common software. And spiced up with critical perspectives on some common practices and an healthy focus on model assessment and model selection. Highly recommended on the menu of Bayesian textbooks!

And this review is likely to appear in CHANCE, in my book reviews column.

Graphical comparison of MCMC performance [arXiv:1011.445]

A new posting on arXiv by Madeleine Thompson on a graphical tool for assessing performance. She has developed a software called SamplerCompare, implemented in R and C. The graphical evaluation plots “log density evaluations per iteration times autocorrelation time against a tuning parameter in a grid of plots where rows represent distributions and columns represent methods”. The autocorrelation time is evaluated in the same way as coda, which is the central package used in the convergence assessment chapter of Introducing Monte Carlo Methods with R because of its array of partial (if imperfect) indicators. Note that there is an approximation factor in the evaluation of the autocorrelation time because the MCMC output is represented as an AR(p) series, with a possible divergence artifact in the corresponding confidence interval if the AR(p) process is found to be non-stationary. When the simulation method (corresponding to columns in the above graphs) allows for an optimal value of its (cyber-)parameters, the performances exhibit a clear parabolic pattern (right graph), but this is not always the case (left graph). Graphical tools are always to be preferred to tables (a point Andrew would not rebuke!), However I do not see the point in simultaneously graphing the performances of different MCMC algorithms for different targets. This “wasted” dimension could instead be used for increasing to at least three the number of cyber-parameters evaluated by the method.

Bayes vs. SAS

Glancing perchance at the back of my Amstat News, I was intrigued by the SAS advertisement

Bayesian Methods

• Specify Bayesian analysis for ANOVA, logistic regression, Poisson regression, accelerated failure time models and Cox regression through the GENMOD, LIFEREG and PHREG procedures.

• Analyze a wider variety of models with the MCMC procedure, a general purpose Bayesian analysis procedure.

and so decided to take a look at those items on the SAS website. (Some entries date back to 2006 so I am not claiming novelty in this post, just my reading through the manual!)

Even though I have not looked at a SAS program since the time in 1984 I was learning principal component and discriminant analysis by programming SAS procedures on punched cards, it seems the MCMC part is rather manageable (if you can manage SAS at all!), looking very much like a second BUGS to my bystander eyes, even to the point of including ARS algorithms! The models are defined in a BUGS manner, with priors on the side (and this includes improper priors, despite a confusing first example that mixes very large variances with vague priors for the linear model!). The basic scheme is a random walk proposal with adaptive scale or covariance matrix. (The adaptivity on the covariance matrix is slightly confusing in that the way it is described it does not seem to implement the requirements of Roberts and Rosenthal for sure convergence.) Gibbs sampling is not directly covered, although some examples are in essence using Gibbs samplers. Convergence is assessed via ca. 1995 methods à la Cowles and Carlin, including the rather unreliable Raftery and Lewis indicator, but so does Introducing Monte Carlo Methods with R, which takes advantage of the R coda package. I have not tested (!) any of the features in the MCMC procedure but judging from a quick skim through the 283 page manual everything looks reasonable enough. I wonder if anyone has ever tested a SAS program against its BUGS counterpart for efficiency comparison.

The Bayesian aspects are rather traditional as well, except for the testing issue. Indeed, from what I have read, SAS does not engage into testing and remains within estimation bounds, offering only HPD regions for variable selection without producing a genuine Bayesian model choice tool. I understand the issues with handling improper priors versus computing Bayes factors, as well as some delicate computational requirements, but this is a truly important chunk missing from the package. (Of course, the package contains a DIC (Deviance information criterion) capability, which may be seen as a substitute, but I have reservations about the relevance of DIC outside generalised linear models. Same difficulty with the posterior predictive.) As usual with SAS, the documentation is huge (I still remember the shelves after shelves of documentation volumes in my 1984 card-punching room!) and full of options and examples. Nothing to complain about. Except maybe the list of disadvantages in using Bayesian analysis:

• It does not tell you how to select a prior. There is no correct way to choose a prior. Bayesian inferences require skills to translate prior beliefs into a mathematically formulated prior. If you do not proceed with caution, you can generate misleading results.
• It can produce posterior distributions that are heavily influenced by the priors. From a practical point of view, it might sometimes be difficult to convince subject matter experts who do not agree with the validity of the chosen prior.
• It often comes with a high computational cost, especially in models with a large number of parameters.

which does not say much… Since the MCMC procedure allows for any degree of hierarchical modelling, it is always possible to check the impact of a given prior by letting its parameters go random. I found that most practitioners are happy with the formalisation of their prior beliefs into mathematical densities, rather than adamant about a specific prior. As for computation, this is not a major issue.