O’Bayes 19/1 [snapshots]

Although the tutorials of O’Bayes 2019 of yesterday were poorly attended, albeit them being great entries into objective Bayesian model choice, recent advances in MCMC methodology, and the multiple layers of BART, for which I have to blame myself for sticking the beginning of O’Bayes too closely to the end of BNP as only the most dedicated could achieve the commuting from Oxford to Coventry to reach Warwick in time, the first day of talks were well attended, despite weekend commitments, conference fatigue, and perfect summer weather! Here are some snapshots from my bench (and apologies for not covering better the more theoretical talks I had trouble to follow, due to an early and intense morning swimming lesson! Like Steve Walker’s utility based derivation of priors that generalise maximum entropy priors. But being entirely independent from the model does not sound to me like such a desirable feature… And Natalia Bochkina’s Bernstein-von Mises theorem for a location scale semi-parametric model, including a clever construct of a mixture of two Dirichlet priors to achieve proper convergence.)

Jim Berger started the day with a talk on imprecise probabilities, involving the society for imprecise probability, which I discovered while reading Keynes’ book, with a neat resolution of the Jeffreys-Lindley paradox, when re-expressing the null as an imprecise null, with the posterior of the null no longer converging to one, with a limit depending on the prior modelling, if involving a prior on the bias as well, with Chris discussing the talk and mentioning a recent work with Edwin Fong on reinterpreting marginal likelihood as exhaustive X validation, summing over all possible subsets of the data [using log marginal predictive].Håvard Rue did a follow-up talk from his Valencià O’Bayes 2015 talk on PC-priors. With a pretty hilarious introduction on his difficulties with constructing priors and counseling students about their Bayesian modelling. With a list of principles and desiderata to define a reference prior. However, I somewhat disagree with his argument that the Kullback-Leibler distance from the simpler (base) model cannot be scaled, as it is essentially a log-likelihood. And it feels like multivariate parameters need some sort of separability to define distance(s) to the base model since the distance somewhat summarises the whole departure from the simpler model. (Håvard also joined my achievement of putting an ostrich in a slide!) In his discussion, Robin Ryder made a very pragmatic recap on the difficulties with constructing priors. And pointing out a natural link with ABC (which brings us back to Don Rubin’s motivation for introducing the algorithm as a formal thought experiment).

Sara Wade gave the final talk on the day about her work on Bayesian cluster analysis. Which discussion in Bayesian Analysis I alas missed. Cluster estimation, as mentioned frequently on this blog, is a rather frustrating challenge despite the simple formulation of the problem. (And I will not mention Larry’s tequila analogy!) The current approach is based on loss functions directly addressing the clustering aspect, integrating out the parameters. Which produces the interesting notion of neighbourhoods of partitions and hence credible balls in the space of partitions. It still remains unclear to me that cluster estimation is at all achievable, since the partition space explodes with the sample size and hence makes the most probable cluster more and more unlikely in that space. Somewhat paradoxically, the paper concludes that estimating the cluster produces a more reliable estimator on the number of clusters than looking at the marginal distribution on this number. In her discussion, Clara Grazian also pointed the ambivalent use of clustering, where the intended meaning somehow diverges from the meaning induced by the mixture model.

the philosophical importance of Stein’s paradox [a reply from the authors]

[In the wake of my comment on this paper written by three philosophers of Science, I received this reply from Olav Vassend.]

Thank you for reading our paper and discussing it on your blog! Our purpose with the paper was to give an introduction to Stein’s phenomenon for a philosophical audience; it was not meant to — and probably will not — offer a new and interesting perspective for a statistician who is already very familiar with Stein’s phenomenon and its extensive literature.

I have a few more specific comments:

1. We don’t rechristen Stein’s phenomenon as “holistic pragmatism.” Rather, holistic pragmatism is the attitude to frequentist estimation that we think is underwritten by Stein’s phenomenon. Since MLE is sometimes admissible and sometimes not, depending on the number of parameters estimated, the researcher has to take into account his or her goals (whether total accuracy or individual-parameter accuracy is more important) when picking an estimator. To a statistician, this might sound obvious, but to philosophers it’s a pretty radical idea.

2. “The part connecting Stein with Bayes again starts on the wrong foot, since it is untrue that any shrinkage estimator can be expressed as a Bayes posterior mean. This is not even true for the original James-Stein estimator, i.e., it is not a Bayes estimator and cannot be a Bayes posterior mean.”

That seems to depend on what you mean by a “Bayes estimator.” It is possible to have an empirical Bayes prior (constructed from the sample) whose posterior mean is identical to the original James-Stein estimator. But if you don’t count empirical Bayes priors as Bayesian, then you are right.

3. “And to state that improper priors “integrate to a number larger than 1” and that “it’s not possible to be more than 100% confident in anything”… And to confuse the Likelihood Principle with the prohibition of data dependent priors. And to consider that the MLE and any shrinkage estimator have the same expected utility under a flat prior (since, if they had, there would be no Bayes estimator!).”

I’m not sure I completely understand your criticisms here. First, as for the relation between the LP and data-dependent priors — it does seem to me that the LP precludes the use of data-dependent priors.  If you use data from an experiment to construct your prior, then — contrary to the LP — it will not be true that all the information provided by the experiment regarding which parameter is true is contained in the likelihood function, since some of the information provided by the experiment will also be in your prior.

Second, as to our claim that the ML estimator has the same expected utility (under the flat prior) as a shrinkage prior that it is dominated by—we incorporated this claim into our paper because it was an objection made by a statistician who read and commented on our paper. Are you saying the claim is false? If so, we would certainly like to know so that we can revise the paper to make it more accurate.

4. I was aware of Rubin’s idea that priors and utility functions (supposedly) are non-separable, but I didn’t (and don’t) quite see the relevance of that idea to Stein estimation.

5. “Similarly, very little of substance can be found about empirical Bayes estimation and its philosophical foundations.”

What we say about empirical Bayes priors is that they cannot be interpreted as degrees of belief; they are just tools. It will be surprising to many philosophers that priors are sometimes used in such an instrumentalist fashion in statistics.

6. The reason why we made a comparison between Stein estimation and AIC was two-fold: (a) for sociological reasons, philosophers are much more familiar with model selection than they are with, say, the LASSO or other regularized regression methods. (b) To us, it’s precisely because model selection and estimation are such different enterprises that it’s interesting that they have such a deep connection: despite being very different, AIC and shrinkage both rely on a bias-variance trade-off.

7. “I also object to the envisioned possibility of a shrinkage estimator that would improve every component of the MLE (in a uniform sense) as it contradicts the admissibility of the single component MLE!”

I don’t think our suggestion here contradicts the admissibility of single component MLE. The idea is just that if we have data D and D’ about parameters φ and φ’, then the estimates of both φ and φ’ can sometimes be improved if the estimation problems are lumped together and a shrinkage estimator is used. This doesn’t contradict the admissibility of MLE, because MLE is still admissible on each of the data sets for each of the parameters.

Again, thanks for reading the paper and for the feedback—we really do want to make sure our paper is accurate, so your feedback is much appreciated. Lastly, I apologize for the length of this comment.

Olav Vassend

the philosophical importance of Stein’s paradox

I recently came across this paper written by three philosophers of Science, attempting to set the Stein paradox in a philosophical light. Given my past involvement, I was obviously interested about which new perspective could be proposed, close to sixty years after Stein (1956). Paper that we should actually celebrate next year! However, when reading the document, I did not find a significantly innovative approach to the phenomenon…

The paper does not start in the best possible light since it seems to justify the use of a sample mean through maximum likelihood estimation, which only is the case for a limited number of probability distributions (including the Normal distribution, which may be an implicit assumption). For instance, when the data is Student’s t, the MLE is not the sample mean, no matter how shocking that might sounds! (And while this is a minor issue, results about the Stein effect taking place in non-normal settings appear much earlier than 1998. And earlier than in my dissertation. See, e.g., Berger and Bock (1975). Or in Brandwein and Strawderman (1978).)

While the linear regression explanation for the Stein effect is already exposed in Steve Stigler’s Neyman Lecture, I still have difficulties with the argument in that for instance we do not know the value of the parameter, which makes the regression and the inverse regression of parameter means over Gaussian observations mere concepts and nothing practical. (Except for the interesting result that two observations make both regressions coincide.) And it does not seem at all intuitive (to me) that imposing a constraint should improve the efficiency of a maximisation program…

a general framework for updating belief functions [reply from the authors]

Here is the reply by Chris and Steve about my comments from yesterday:

Thanks to Christian for the comments and feedback on our paper “A General Framework for Updating Belief Distributions“. We agree with Christian that starting with a summary statistic, or statistics, is an anchor for inference or learning, providing direction and guidance for models, avoiding the alternative vague notion of attempting to model a complete data set. The latter idea has dominated the Bayesian methodology for decades, but with the advent of large and complex data sets, this is becoming increasingly challenging, if not impossible.

However, in order to do work with statistics of interest, we need to find a framework in which this direct approach can be supported by a learning strategy when the formal use of Bayes theorem is not applicable. We achieve this in the paper for a general class of loss functions, which connect observations with a target of interest. A point raised by Christian is how arbitrary these loss functions are. We do not see this at all; for if a target has been properly identified then the most primitive construct between observations informing about a target and the target would come in the form of a loss function. One should always be able to assess the loss of ascertaining a value of as an action and providing the loss in the presence of observation x. The question to be discussed is whether loss functions are objective, as in the case of the median loss,

or subjective, in the case of the choice between loss functions for estimating a location of a distribution; mean, median or mode? But our work is situated in the former position.

Previous work on loss functions, mostly in the classical literature, has spent a lot of space working out what are optimal loss functions for targets of interest. We are not really dealing with novel targets and so we can draw on the classic literature here. The work can be thought of as the Bayesian version of the M-estimator and associated ideas. In this respect we are dealing with two loss functions for updating belief distributions, one for the data, which we have just discussed, and one for the prior information, which, due to coherence principles, must be the Kullback-Leibler divergence. This raises the thorny issue of how to calibrate the two loss functions. We discuss this in the paper.

To then deal with the statistic problem, mentioned at the start of this discussion, we have found a nice way to proceed by using the loss function . How this loss function, combined with the use of the exponential family, can be used to estimate functionals of the type

is provided in the Walker talk at Bayes 250 in London, titled “The Misspecified Bayesian”, since the “model” is designed to be misspecified, a tool to estimate and learn about I only. The basic idea is to evaluate I by ensuring that we learn about the for which

This is the story of the background, we would now like to pick up in more detail on three important points that you raise in your post:

1. The arbitrariness in selecting the loss function.
2. The relative weighting of loss-to-data vs. loss-to-prior.
3. The selection of the loss in the M-free case.

In the absence of complete knowledge of the data generating mechanism, i.e. outside of M-closed,

1. We believe the statistician should weigh up the relative arbitrariness in selecting a loss function targeting the statistic of interest versus the arbitrariness of selecting a misspecified model, known not to be true, for the complete data generating mechanism. There is a wealth of literature on how to select optimal loss functions that target specific statistics, e.g. Hüber (2009) provides a comprehensive overview of how this should be done. As far as we are aware, we know of no formal procedures (that do not rely on loss functions) to select a false sampling distribution for the whole of x; see Key, Pericchi and Smith (1999).
2. The relative weighting of loss-to-data vs. loss-to-prior. This is an interesting open problem. Our framework shows in the absence of M-closed or the use of self-information loss that the analyst must select this weighting. In our paper we suggest some default procedures. We have nowhere claimed these were “correct”. You raise concerns regards parameterisation and we agree with you that care is needed, but many of these issues equally hold for existing “Objective” or “Default” Bayes procedures, such as unit-information priors.
3. The selection of the loss in M-free. You say “….there is no optimal choice for the substitute to the loss function…”. We disagree. Our approach is to select an established loss function that directly targets the statistic of interest, and elicit prior beliefs directly on the unknown value of this statistic. There is no notion here of a a pseudo-likelihood or where this converges to.

Thank you again to Christian for his critical observations!

a general framework for updating belief functions

Pier Giovanni Bissiri, Chris Holmes and Stephen Walker have recently arXived the paper related to Sephen’s talk in London for Bayes 250. When I heard the talk (of which some slides are included below), my interest was aroused by the facts that (a) the approach they investigated could start from a statistics, rather than from a full model, with obvious implications for ABC, & (b) the starting point could be the dual to the prior x likelihood pair, namely the loss function. I thus read the paper with this in mind. (And rather quickly, which may mean I skipped important aspects. For instance, I did not get into Section 4 to any depth. Disclaimer: I wasn’t nor is a referee for this paper!)

The core idea is to stick to a Bayesian (hardcore?) line when missing the full model, i.e. the likelihood of the data, but wishing to infer about a well-defined parameter like the median of the observations. This parameter is model-free in that some degree of prior information is available in the form of a prior distribution. (This is thus the dual of frequentist inference: instead of a likelihood w/o a prior, they have a prior w/o a likelihood!) The approach in the paper is to define a “posterior” by using a functional type of loss function that balances fidelity to prior and fidelity to data. The prior part (of the loss) ends up with a Kullback-Leibler loss, while the data part (of the loss) is an expected loss wrt to l(THETASoEUR,x), ending up with the definition of a “posterior” that is

the loss thus playing the role of the log-likelihood.

I like very much the problematic developed in the paper, as I think it is connected with the real world and the complex modelling issues we face nowadays. I also like the insistence on coherence like the updating principle when switching former posterior for new prior (a point sorely missed in this book!) The distinction between M-closed M-open, and M-free scenarios is worth mentioning, if only as an entry to the Bayesian processing of pseudo-likelihood and proxy models. I am however not entirely convinced by the solution presented therein, in that it involves a rather large degree of arbitrariness. In other words, while I agree on using the loss function as a pivot for defining the pseudo-posterior, I am reluctant to put the same faith in the loss as in the log-likelihood (maybe a frequentist atavistic gene somewhere…) In particular, I think some of the choices are either hard or impossible to make and remain unprincipled (despite a call to the LP on page 7).  I also consider the M-open case as remaining unsolved as finding a convergent assessment about the pseudo-true parameter brings little information about the real parameter and the lack of fit of the superimposed model. Given my great expectations, I ended up being disappointed by the M-free case: there is no optimal choice for the substitute to the loss function that sounds very much like a pseudo-likelihood (or log thereof). (I thought the talk was more conclusive about this, I presumably missed a slide there!) Another great expectation was to read about the proper scaling of the loss function (since L and wL are difficult to separate, except for monetary losses). The authors propose a “correct” scaling based on balancing both faithfulness for a single observation, but this is not a completely tight argument (dependence on parametrisation and prior, notion of a single observation, &tc.)

The illustration section contains two examples, one of which is a full-size or at least challenging  genetic data analysis. The loss function is based on a logistic  pseudo-likelihood and it provides results where the Bayes factor is in agreement with a likelihood ratio test using Cox’ proportional hazard model. The issue about keeping the baseline function as unkown reminded me of the Robbins-Wasserman paradox Jamie discussed in Varanasi. The second example offers a nice feature of putting uncertainties onto box-plots, although I cannot trust very much the 95%  of the credibles sets. (And I do not understand why a unique loss would come to be associated with the median parameter, see p.25.)

Watch out: Tomorrow’s post contains a reply from the authors!