Archive for finite mixtures

Jeffreys prior with improper posterior

Posted in Books, Statistics, University life with tags , , , , , , , , , , on May 12, 2014 by xi'an

In a complete coincidence with my visit to Warwick this week, I became aware of the paper “Inference in two-piece location-scale models with Jeffreys priors” recently published in Bayesian Analysis by Francisco Rubio and Mark Steel, both from Warwick. Paper where they exhibit a closed-form Jeffreys prior for the skewed distribution

\dfrac{2\epsilon}{\sigma_1}f(\{x-\mu\}/\sigma_1)\mathbb{I}_{x<\mu}+\dfrac{2(1-\epsilon)}{\sigma_2}f(\{x-\mu\}/\sigma_2) \mathbb{I}_{x>\mu}

where f is a symmetric density, namely

\pi(\mu,\sigma_1,\sigma_2) \propto 1 \big/ \sigma_1\sigma_2\{\sigma_1+\sigma_2\}\,,

where

\epsilon=\sigma_1/\{\sigma_1+\sigma_2\}\,.

only to show  immediately after that this prior does not allow for a proper posterior, no matter what the sample size is. While the above skewed distribution can always be interpreted as a mixture, being a weighted sum of two terms, it is not strictly speaking a mixture, if only because the “component” can be identified from the observation (depending on which side of μ is stands). The likelihood is therefore a product of simple terms rather than a product of a sum of two terms.

As a solution to this conundrum, the authors consider the alternative of the “independent Jeffreys priors”, which are made of a product of conditional Jeffreys priors, i.e., by computing the Jeffreys prior one parameter at a time with all other parameters considered to be fixed. Which differs from the reference prior, of course, but would have been my second choice as well. Despite criticisms expressed by José Bernardo in the discussion of the paper… The difficulty (in my opinion) resides in the choice (and difficulty) of the parameterisation of the model, since those priors are not parameterisation-invariant. (Xinyi Xu makes the important comment that even those priors incorporate strong if hidden information. Which relates to our earlier discussion with Kaniav Kamari on the “dangers” of prior modelling.)

Although the outcome is puzzling, I remain just slightly sceptical of the income, namely Jeffreys prior and the corresponding Fisher information: the fact that the density involves an indicator function and is thus discontinuous in the location μ at the observation x makes the likelihood function not differentiable and hence the derivation of the Fisher information not strictly valid. Since the indicator part cannot be differentiated. Not that I am seeing the Jeffreys prior as the ultimate grail for non-informative priors, far from it, but there is definitely something specific in the discontinuity in the density. (In connection with the later point, Weiss and Suchard deliver a highly critical commentary on the non-need for reference priors and the preference given to a non-parametric Bayes primary analysis. Maybe making the point towards a greater convergence of the two perspectives, objective Bayes and non-parametric Bayes.)

This paper and the ensuing discussion about the properness of the Jeffreys posterior reminded me of our earliest paper on the topic with Jean Diebolt. Where we used improper priors on location and scale parameters but prohibited allocations (in the Gibbs sampler) that would lead to less than two observations per components, thereby ensuring that the (truncated) posterior was well-defined. (This feature also remained in the Series B paper, submitted at the same time, namely mid-1990, but only published in 1994!)  Larry Wasserman proved ten years later that this truncation led to consistent estimators, but I had not thought about it in very long while. I still like this notion of forcing some (enough) datapoints into each component for an allocation (of the latent indicator variables) to be an acceptable Gibbs move. This is obviously not compatible with the iid representation of a mixture model, but it expresses the requirement that components all have a meaning in terms of the data, namely that all components contributed to generating a part of the data. This translates as a form of weak prior information on how much we trust the model and how meaningful each component is (in opposition to adding meaningless extra-components with almost zero weights or almost identical parameters).

As a marginalia, the insistence in Rubio and Steel’s paper that all observations in the sample be different also reminded me of a discussion I wrote for one of the Valencia proceedings (Valencia 6 in 1998) where Mark presented a paper with Carmen Fernández on this issue of handling duplicated observations modelled by absolutely continuous distributions. (I am afraid my discussion is not worth the $250 price tag given by amazon!)

Bayesian Data Analysis [BDA3 - part #2]

Posted in Books, Kids, R, Statistics, University life with tags , , , , , , , , on March 31, 2014 by xi'an

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.

Bayesian Data Analysis [BDA3]

Posted in Books, Kids, R, Statistics, University life with tags , , , , , , , , on March 28, 2014 by xi'an

Andrew Gelman and his coauthors, John Carlin, Hal Stern, David Dunson, Aki Vehtari, and Don Rubin, have now published the latest edition of their book Bayesian Data Analysis. David and Aki are newcomers to the authors’ list, with an extended section on non-linear and non-parametric models. I have been asked by Sam Behseta to write a review of this new edition for JASA (since Sam is now the JASA book review editor). After wondering about my ability to produce an objective review (on the one hand, this is The Competition  to Bayesian Essentials!, on the other hand Andrew is a good friend spending the year with me in Paris), I decided to jump for it and write a most subjective review, with the help of Clara Grazian who was Andrew’s teaching assistant this year in Paris and maybe some of my Master students who took Andrew’s course. The second edition was reviewed in the September 2004 issue of JASA and we now stand ten years later with an even more impressive textbook. Which truly what Bayesian data analysis should be.

This edition has five parts, Fundamentals of Bayesian Inference, Fundamentals of Bayesian Data Analysis, Advanced Computation, Regression Models, and Non-linear and Non-parametric Models, plus three appendices. For a total of xiv+662 pages. And a weight of 2.9 pounds (1395g on my kitchen scale!) that makes it hard to carry around in the metro…. I took it to Warwick (and then Nottingham and Oxford and back to Paris) instead.

We could avoid the mathematical effort of checking the integrability of the posterior density (…) The result would clearly show the posterior contour drifting off toward infinity.” (p.111)

While I cannot go into a detailed reading of those 662 pages (!), I want to highlight a few gems. (I already wrote a detailed and critical analysis of Chapter 6 on model checking in that post.) The very first chapter provides all the necessary items for understanding Bayesian Data Analysis without getting bogged in propaganda or pseudo-philosophy. Then the other chapters of the first part unroll in a smooth way, cruising on the B highway… With the unique feature of introducing weakly informative priors (Sections 2.9 and 5.7), like the half-Cauchy distribution on scale parameters. It may not be completely clear how weak a weakly informative prior, but this novel notion is worth including in a textbook. Maybe a mild reproach at this stage: Chapter 5 on hierarchical models is too verbose for my taste, as it essentially focus on the hierarchical linear model. Of course, this is an essential chapter as it links exchangeability, the “atom” of Bayesian reasoning used by de Finetti, with hierarchical models. Still. Another comment on that chapter: it broaches on the topic of improper posteriors by suggesting to run a Markov chain that can exhibit improperness by enjoying an improper behaviour. When it happens as in the quote above, fine!, but there is no guarantee this is always the case! For instance, improperness may be due to regions near zero rather than infinity. And a last barb: there is a dense table (Table 5.4, p.124) that seems to run contrariwise to Andrew’s avowed dislike of tables. I could also object at the idea of a “true prior distribution” (p.128), or comment on the trivia that hierarchical chapters seem to attract rats (as I also included a rat example in the hierarchical Bayes chapter of Bayesian Choice and so does the BUGS Book! Hence, a conclusion that Bayesian textbooks are better be avoided by muriphobiacs…)

“Bayes factors do not work well for models that are inherently continuous (…) Because we emphasize continuous families of models rather than discrete choices, Bayes factors are rarely relevant in our approach to Bayesian statistics.” (p.183 & p.193)

Part II is about “the creative choices that are required, first to set up a Bayesian model in a complex problem, then to perform the model checking and confidence building that is typically necessary to make posterior inferences scientifically defensible” (p.139). It is certainly one of the strengths of the book that it allows for a critical look at models and tools that are rarely discussed in more theoretical Bayesian books. As detailed in my  earlier post on Chapter 6, model checking is strongly advocated, via posterior predictive checks and… posterior predictive p-values, which are at best empirical indicators that something could be wrong, definitely not that everything’s allright! Chapter 7 is the model comparison equivalent of Chapter 6, starting with the predictive density (aka the evidence or the marginal likelihood), but completely bypassing the Bayes factor for information criteria like the Watanabe-Akaike or widely available information criterion (WAIC), and advocating cross-validation, which is empirically satisfying but formally hard to integrate within a full Bayesian perspective. Chapter 8 is about data collection, sample surveys, randomization and related topics, another entry that is missing from most Bayesian textbooks, maybe not that surprising given the research topics of some of the authors. And Chapter 9 is the symmetric in that it focus on the post-modelling step of decision making.

(Second part of the review to appear on Monday, leaving readers the weekend to recover!)

Advances in scalable Bayesian computation [day #2]

Posted in Books, Mountains, pictures, R, Statistics, University life with tags , , , , , , , , , , , on March 5, 2014 by xi'an

polyptych painting within the TransCanada Pipeline Pavilion, Banff Centre, Banff, March 21, 2012And here is the second day of our workshop Advances in Scalable Bayesian Computation gone! This time, it sounded like the “main” theme was about brains… In fact, Simon Barthelmé‘s research originated from neurosciences, while Dawn Woodard dissected a brain (via MRI) during her talk! (Note that the BIRS website currently posts Simon’s video as being Dan Simpson’s talk, the late change in schedule being due to Dan most unfortunately losing his passport during a plane transfer and most unfortunately being prevented from attending…) I found Simon’s talk quite inspiring, with this Tibshirani et al.’s trick of using logistic regression to estimate densities as a classification problem central to the method and suggesting a completely different vista for handling normalising constants… Then Raazesh Sainudiin gave a detailed explanation and validation of his approach to density estimation by multidimensional pavings/histograms, with a tree representation allowing for fast merging of different estimators. Raaz had given a preliminary version of the talk at CREST last Fall, which helped with focussing on the statistical aspects of the method. Chris Strickland then exposed an image analysis of flooded Northern Queensland landscapes, using a spatio-temporal model with changepoints and about 18,000 parameters. still managing to get an efficiency of O(np) thanks to two tricks. Then it was time for the group photograph outside in a balmy -18⁰ and an open research time that was quite profitable.

In the afternoon sessions, Paul Fearnhead presented an auxiliary variable approach to particle Gibbs, which again opened new possibilities for handling state-space models, but also reminding me of Xiao-Li Meng’s reparameterisation devices. And making me wonder (out loud) whether or not the SMC algorithm was that essential in a static setting, since the sequence could be explored in any possible order for a fixed time horizon. Then Emily Fox gave a 2-for-1 talk, mostly focussing on the first talk, where she introduced a new technique for approximating the gradient in Hamiltonian (or Hockey!) Monte Carlo, using second order Langevin. She did not have much time for the second talk, which intersected with the one she gave at BNP’ski in Chamonix, but focussed on a notion of sandwiched slice sampling where the target density only needs bounds that can get improved if needed. A cool trick! And the talks ended with Dawn Woodard‘s analysis of time varying 3-D brain images towards lesion detection, through an efficient estimation of a spatial mixture of normals.

Bayesian indirect inference [a response]

Posted in Books, Statistics, Travel, University life with tags , , , , , , on February 18, 2014 by xi'an

Zeeman Building, University of WarwickThis Bayesian indirect inference paper by Chris Drovandi and Tony Pettitt was discussed on the ‘Og two weeks ago and Chris sent me the following comments.

unsurprisingly, the performances of ABC comparing true data of size n with synthetic data of size m>n are not great. However, there exists another way of reducing the variance in the synthetic data, namely by repeating simulations of samples of size n and averaging the indicators for proximity, resulting in a frequency rather than a 0-1 estimator. See e.g. Del Moral et al. (2009). In this sense, increasing the computing power reduces the variability of the ABC approximation. (And I thus fail to see the full relevance of Result 1.)

Taking the average of the indicators from multiple simulations will reduce the variability of the estimated ABC likelihood but because it is only still an unbiased estimate it will not alter the target and will not improve the ABC approximation (Andrieu and Roberts 2009).  It will only have the effect of improving the mixing of MCMC ABC.  Result 1 is used to contrast ABC II and BIL as they behave quite differently as n is increased.

The authors make several assumptions of unicity that I somewhat find unclear. While assuming that the MLE for the auxiliary model is unique could make sense (Assumption 2), I do not understand the corresponding indexing of this estimator (of the auxiliary parameter) on the generating (model) parameter θ. It should only depend on the generated/simulated data x. The notion of a noisy mapping is just confusing to me.

The dependence on θ is a little confusing I agree (especially in the context of ABC II methods).  It starts to become more clear in the context of BIL.  As n goes to infinity, the effect of the simulated data is removed and then we obtain the function φ(θ) (so we need to remember which θ simulated the data), which is referred to as the mapping or binding function in the II literature.  If we somehow knew the binding function, BIL would proceed straightforwardly.  But of course we don’t in practice, so we try to estimate it via simulated data (which, for computational reasons, needs to be a finite sample) from the true model based on theta.  Thus we obtain a noisy estimate of the mapping.  One way forward might be to fit some (non-parametric?) regression model to smooth out the noise and try to recover the true mapping (without ever taking n to infinity) and run a second BIL with this estimated mapping.  I plan to investigate this in future work.

The assumption that the auxiliary score function at the auxiliary MLE for the observed data and for a simulated dataset (Assumption 3) is unique proceeds from the same spirit. I however fail to see why it matters so much. If the auxiliary MLE is the result of a numerical optimisation algorithm, the numerical algorithm may return local modes. This only adds to the approximative effect of the ABC-I schemes.

The optimiser failing to find the MLE (local mode) is certainly an issue shared by all BII methods, apart from ABC IS (which only requires 1 optimisation, so more effort to find the MLE can be applied here).  Assuming the optimiser can obtain the MLE, I think the uniqueness assumptions makes sense.  It basically says that, for a particular simulated dataset we would like a unique value for the ABC discrepancy function.

Given that the paper does not produce convergence results for those schemes, unless the auxiliary model contains the genuine model, such theoretical assumptions do not feel that necessary.

Actually, the ABC II methods will never converge to the true posterior (in general) due to lack of sufficiency.  This is even the case if the true model is a special case of the auxiliary model! (in which case BIL can converge to the true posterior)

The paper uses normal mixtures as an auxiliary model: the multimodality of this model should not be such an hindrance (and reordering is transparent, i.e. does not “reduce the flexibility of the auxiliary model”, and does not “increase the difficulty of implementation”, as stated p.16).

Thanks for your comment.  I need to think about this more as I am not an expert on mixture modelling.  The standard EM algorithm in Matlab does not apply any ordering to the parameters of the components and uses a random start.  Thus it can return any of the multiple MLEs on offer, so the ABC IP will not work here.  So from my point of view, any alternative will increase the difficulty of implementation as it means I cannot use the standard software.  Especially considering I can apply any other BII method without worrying about the non-unique MLE.

The paper concludes from a numerical study to the superiority of the Bayesian indirect inference of Gallant and McCulloch (2009). Which simply replaces the true likelihood with the maximal auxiliary model likelihood estimated from a simulated dataset. (This is somehow similar to our use of the empirical likelihood in the PNAS paper.) It is however moderated by the cautionary provision that “the auxiliary model [should] describe the data well”. As for empirical likelihood, I would suggest resorting to this Bayesian indirect inference as a benchmark, providing a quick if possibly dirty reference against which to test more elaborate ABC schemes. Or other approximations, like empirical likelihood or Wood’s synthetic likelihood.

Unfortunately the methods are not quick (apart from ABC IS when the scores are analytic), but good approximations can be obtained.  The majority of Bayesian methods that deal with intractable likelihoods do not target the true posterior (there are a couple of exceptions in special cases) and thus also suffer from some dirtiness, and BII does not escape from that.  But, if a reasonable auxiliary model can be found, then I would suggest that (at least one of the) BII methods will be competitive.

On reflection for BIL it is not necessary for the auxiliary model to fit the data, since the generative model being proposed may be mis-specified and also not fit the data well.  BIL needs an auxiliary model that mimics well the likelihood of the generative model for values of theta in non-negligible posterior regions.  For ABC II, we are simply looking for a good summarisation of the data.  Therefore it would seem useful if the auxiliary model did fit the data well.  Note this process is independent of the generative model being proposed.  Therefore the auxiliary model would be the same regardless of the chosen generative model.  Very different considerations indeed.

Inspired by a discussion with Anthony Lee, it appears that the (Bayesian version) of synthetic likelihood you mentioned is actually also a BIL method but where the auxiliary model is applied to the summary statistic likelihood rather than the full data likelihood.  The synthetic likelihood is nice from a numerical/computational point of view as the MLE of the auxiliary model is analytic.

Follow

Get every new post delivered to your Inbox.

Join 598 other followers