prior sensitivity of the marginal likelihood

Fernando Llorente and (Madrilene) coauthors have just arXived a paper on the safe use of prior densities for Bayesian model selection. Rather than blaming the Bayes factor, or excommunicating some improper priors, they consider in this survey solutions to design “objective” priors in model selection. (Writing this post made me realised I had forgotten to arXive a recent piece I wrote on the topic, based on short courses and blog pieces, for an incoming handbook on Bayesian advance(ment)s! Soon to be corrected.)

While intrinsically interested in the topic and hence with the study, I somewhat disagree with the perspective adopted by the authors. They for instance stick to the notion that a flat prior over the parameter space is appropriate as “the maximal expression of a non-informative prior” (despite depending on the parameterisation). Over bounded sets at least, while advocating priors “with great scale parameter” otherwise. They also refer to Jeffreys (1939) priors, by which they mean estimation priors rather than testing priors. As uncovered by Susie Bayarri and Gonzalo Garcia-Donato. Considering asymptotic consistency, they state that “in the asymptotic regime, Bayesian model selection is more sensitive to the sample size D than to the prior specifications”, which I find both imprecise and confusing,  as my feeling is that the prior specification remains overly influential as the sample size increases. (In my view, consistency is a minimalist requirement, rather than “comforting”.) The argument therein that a flat prior is informative for model choice stems from the fact that the marginal likelihood goes to zero as the support of the prior goes to infinity, which may have been an earlier argument of Jeffreys’ (1939), but does not carry much weight as the property is shared by many other priors (as remarked later). Somehow, the penalisation aspect of the marginal is not exploited more deeply in the paper. In the “objective” Bayes section, they adhere to the (convenient but weakly supported) choice of a common prior on the nuisance parameters (shared by different models). Their main argument is to develop (heretic!) “data-based priors”, from Aitkin (1991, not cited) double use of the data (or setting the likelihood to the power two), all the way to the intrinsic and fractional Bayes factors of Tony O’Hagan (1995), Jim Berger and Luis Pericchi (1996), and to the expected posterior priors of Pérez and Berger (2002) on which I worked with Juan Cano and Diego Salmeròn. (While the presentation is made against a flat prior, nothing prevents the use of another reference, improper, prior.) A short section also mentions the X-validation approach(es) of Aki Vehtari and co-authors.

robust inference using posterior bootstrap

The famous 1994 Read Paper by Michael Newton and Adrian Raftery was entitled Approximate Bayesian inference, where the boostrap aspect is in randomly (exponentially) weighting each observation in the iid sample through a power of the corresponding density, a proposal that happened at about the same time as Tony O’Hagan suggested the related fractional Bayes factor. (The paper may also be equally famous for suggesting the harmonic mean estimator of the evidence!, although it only appeared as an appendix to the paper.) What is unclear to me is the nature of the distribution g(θ) associated with the weighted bootstrap sample, conditional on the original sample, since the outcome is the result of a random Exponential sample and of an optimisation step. With no impact of the prior (which could have been used as a penalisation factor), corrected by Michael and Adrian via an importance step involving the estimation of g(·).

At the Algorithm Seminar today in Warwick, Emilie Pompe presented recent research, including some written jointly with Pierre Jacob, [which I have not yet read] that does exactly that inclusion of the log prior as penalisation factor, along with an extra weight different from one, as motivated by the possibility of a misspecification. Including a new approach to cut models. An alternative mentioned during the talk that reminds me of GANs is to generate a pseudo-sample from the prior predictive and add it to the original sample. (Some attendees commented on the dependence of the later version on the chosen parameterisation, which is an issue that had X’ed my mind as well.)

Error and Inference [#3]

(This is the third post on Error and Inference, yet again being a raw and naïve reaction to a linear reading rather than a deeper and more informed criticism.)

“Statistical knowledge is independent of high-level theories.”—A. Spanos, p.242, Error and Inference, 2010

The sixth chapter of Error and Inference is written by Aris Spanos and deals with the issues of testing in econometrics. It provides on the one hand a fairly interesting entry in the history of economics and the resistance to data-backed theories, primarily because the buffers between data and theory are multifold (“huge gap between economic theories and the available observational data“, p.203). On the other hand, what I fail to understand in the chapter is the meaning of theory, as it seems very distinct from what I would call a (statistical) model. The sentence “statistical knowledge, stemming from a statistically adequate model allows data to `have a voice of its own’ (…) separate from the theory in question and its succeeds in securing the frequentist goal of objectivity in theory testing” (p.206) is puzzling in this respect. (Actually, I would have liked to see a clear meaning put to this “voice of its own”, as it otherwise sounds mostly as a catchy sentence…) Similarly, Spanos distinguishes between three types of models: primary/theoretical, experimental/structural: “the structural model contains a theory’s substantive subject matter information in light of the available data” (p.213), data/statistical: “the statistical model is built exclusively using the information contained in the data” (p.213). I have trouble to understand how testing can distinguish between those types of models: as a naïve reader, I would have thought that only the statistical model could be tested by a statistical procedure, even though I would not call the above a proper definition of a statistical model (esp. since Spanos writes a few lines below that the statistical model “would embed (nest) the structural model in its context” (p.213)). The normal example followed on pages 213-217 does not help [me] to put sense to this distinction: it simply illustrates the impact of failing some of the defining assumptions (normality, time homogeneity [in mean and variance], independence). (As an aside, the discussion about the poor estimation of the correlation p.214-215 does not help, because it involves a second variable Y that is not defined for this example.) It would be nice of course if the “noise” in a statistical/econometric model could be studied in complete separation from the structure of this model, however they seem to be irremediably intermingled to prevent this partition of roles. I thus do not see how the “statistically adequate model is independent from the substantive information” (p.217), i.e. by which rigorous process one can isolate the “chance” parts of the data to build and validate a statistical model per se. The simultaneous equation model (SEM, pp.230-231) is more illuminating of the distinction set by Spanos between structural and statistical models/parameters, even though the difference in this case boils down to a question of identifiability.

Error and Inference [#2]

(This is the second post on Error and Inference, again being a raw and naive reaction to a linear reading rather than a deeper and more informed criticism.)

“Allan Franklin once gave a seminar under the title `Ad Hoc is not a four letter word.'”—J. Worrall, p.130, Error and Inference, 2010

The fourth chapter of Error and Inference, written by John Worrall, covers the highly interesting issue of “using the data twice”. The point has been debated several times on Andrew’s blog and this is one of the main criticisms raised against Aitkin’s posterior/integrated likelihood. Worrall’s perspective is both related and unrelated to this purely statistical issue, when he considers that “you can’t use the same fact twice, once in the construction of a theory and then again in its support” (p.129). (He even signed a “UN Charter”, where UN stands for “use novelty”!) After reading both Worrall’s and Mayo’s viewpoints,  the later being that all that matters is severe testing as it encompasses the UN perspective (if I understood correctly), I afraid I am none the wiser, but this led me to reflect on the statistical issue.