## inferential models: reasoning with uncertainty [book review]

Posted in Books, Statistics, University life with tags , , , , , , , , , on October 6, 2016 by xi'an

“the field of statistics (…) is still surprisingly underdeveloped (…) the subject lacks a solid theory for reasoning with uncertainty [and] there has been very little progress on the foundations of statistical inference” (p.xvi)

A book that starts with such massive assertions is certainly hoping to attract some degree of attention from the field and likely to induce strong reactions to this dismissal of the not inconsiderable amount of research dedicated so far to statistical inference and in particular to its foundations. Or even attarcting flak for not accounting (in this introduction) for the past work of major statisticians, like Fisher, Kiefer, Lindley, Cox, Berger, Efron, Fraser and many many others…. Judging from the references and the tone of this 254 pages book, it seems like the two authors, Ryan Martin and Chuanhai Liu, truly aim at single-handedly resetting the foundations of statistics to their own tune, which sounds like a new kind of fiducial inference augmented with calibrated belief functions. Be warned that five chapters of this book are built on as many papers written by the authors in the past three years. Which makes me question, if I may, the relevance of publishing a book on a brand-new approach to statistics without further backup from a wider community.

“…it is possible to calibrate our belief probabilities for a common interpretation by intelligent minds.” (p.14)

Chapter 1 contains a description of the new perspective in Section 1.4.2, which I find useful to detail here. When given an observation x from a Normal N(θ,1) model, the authors rewrite X as θ+Z, with Z~N(0,1), as in fiducial inference, and then want to find a “meaningful prediction of Z independently of X”. This seems difficult to accept given that, once X=x is observed, Z=X-θ⁰, θ⁰ being the true value of θ, which belies the independence assumption. The next step is to replace Z~N(0,1) by a random set S(Z) containing Z and to define a belief function bel() on the parameter space Θ by

bel(A|X) = P(X-S(Z)⊆A)

which induces a pseudo-measure on Θ derived from the distribution of an independent Z, since X is already observed. When Z~N(0,1), this distribution does not depend on θ⁰ the true value of θ… The next step is to choose the belief function towards a proper frequentist coverage, in the approximate sense that the probability that bel(A|X) be more than 1-α is less than α when the [arbitrary] parameter θ is not in A. And conversely. This property (satisfied when bel(A|X) is uniform) is called validity or exact inference by the authors: in my opinion, restricted frequentist calibration would certainly sound more adequate.

“When there is no prior information available, [the philosophical justifications for Bayesian analysis] are less than fully convincing.” (p.30)

“Is it logical that an improper “ignorance” prior turns into a proper “non-ignorance” prior when combined with some incomplete information on the whereabouts of θ?” (p.44)

## Approximate Bayesian computation via sufficient dimension reduction

Posted in Statistics, University life with tags , , , , , on August 26, 2016 by xi'an

“One of our contribution comes from the mathematical analysis of the consequence of conditioning the parameters of interest on consistent statistics and intrinsically inconsistent statistics”

Xiaolong Zhong and Malay Ghosh have just arXived an ABC paper focussing on the convergence of the method. And on the use of sufficient dimension reduction techniques for the construction of summary statistics. I had not heard of this approach before so read the paper with interest. I however regret that the paper does not link with the recent consistency results of Liu and Fearnhead and of Daniel Frazier, Gael Martin, Judith Rousseau and myself. When conditioning upon the MLE [or the posterior mean] as the summary statistic, Theorem 1 states that the Bernstein-von Mises theorem holds, missing a limit in the tolerance ε. And apparently missing conditions on the speed of convergence of this tolerance to zero although the conditioning event involves the true value of the parameter. This makes me wonder at the relevance of the result. The part about partial posteriors and the characterisation of limiting posterior distributions stats with the natural remark that the mean of the summary statistic must identify the whole parameter θ to achieve consistency, a point central to our 2014 JRSS B paper. The authors suggest using a support vector machine to derive the summary statistics, an idea already exploited by Heiko Strathmann et al.. There is no consistency result of relevance for ABC in that second and final part, which ends up rather abruptly. Overall, while the paper contributes to the current reflection on the convergence properties of ABC, the lack of scaling of the tolerance ε calls for further investigations.

## Inference for stochastic simulation models by ABC

Posted in Books, Statistics, University life with tags , , , , , on February 13, 2015 by xi'an

Hartig et al. published a while ago (2011) a paper  in Ecology Letters entitled “Statistical inference for stochastic simulation models – theory and application”, which is mostly about ABC. (Florian Hartig pointed out the paper to me in a recent blog comment. about my discussion of the early parts of Guttman and Corander’s paper.) The paper is largely a tutorial and it reminds the reader about related methods like indirect inference and methods of moments. The authors also insist on presenting ABC as a particular case of likelihood approximation, whether non-parametric or parametric. Making connections with pseudo-likelihood and pseudo-marginal approaches. And including a discussion of the possible misfit of the assumed model, handled by an external error model. And also introducing the notion of informal likelihood (which could have been nicely linked with empirical likelihood). A last class of approximations presented therein is called rejection filters and reminds me very much of Ollie Ratman’s papers.

“Our general aim is to find sufficient statistics that are as close to minimal sufficiency as possible.” (p.819)

As in other ABC papers, and as often reported on this blog, I find the stress on sufficiency a wee bit too heavy as those models calling for approximation almost invariably do not allow for any form of useful sufficiency. Hence the mathematical statistics notion of sufficiency is mostly useless in such settings.

“A basic requirement is that the expectation value of the point-wise approximation of p(Sobs|φ) must be unbiased” (p.823)

As stated above the paper is mostly in tutorial mode, for instance explaining what MCMC and SMC methods are. As illustrated by the above figure. There is however a final and interesting discussion section on the impact of estimating the likelihood function at different values of the parameter. However, the authors seem to focus solely on pseudo-marginal results to validate this approximation, hence on unbiasedness, which does not work for most ABC approaches that I know. And for the approximations listed in the survey. Actually, it would be quite beneficial to devise a cheap tool to assess the bias or extra-variation due to the use of approximative techniques like ABC… A sort of 21st Century bootstrap?!

## that the median cannot be a sufficient statistic

Posted in Kids, Statistics, University life with tags , , , , , on November 14, 2014 by xi'an

When reading an entry on The Chemical Statistician that a sample median could often be a choice for a sufficient statistic, it attracted my attention as I had never thought a median could be sufficient. After thinking a wee bit more about it, and even posting a question on cross validated, but getting no immediate answer, I came to the conclusion that medians (and other quantiles) cannot be sufficient statistics for arbitrary (large enough) sample sizes (a condition that excludes the obvious cases of one & two observations where the sample median equals the sample mean).

In the case when the support of the distribution does not depend on the unknown parameter θ, we can invoke the Darmois-Pitman-Koopman theorem, namely that the density of the observations is necessarily of the exponential family form,

$\exp\{ \theta T(x) - \psi(\theta) \}h(x)$

to conclude that, if the natural sufficient statistic

$S=\sum_{i=1}^n T(x_i)$

is minimal sufficient, then the median is a function of S, which is impossible since modifying an extreme in the n>2 observations modifies S but not the median.

In the other case when the support does depend on the unknown parameter θ, we can consider the case when

$f(x|\theta) = h(x) \mathbb{I}_{A_\theta}(x) \tau(\theta)$

where the set indexed by θ is the support of f. In that case, the factorisation theorem implies that

$\prod_{i=1}^n \mathbb{I}_{A_\theta}(x_i)$

is a 0-1 function of the sample median. Adding a further observation y⁰ which does not modify the median then leads to a contradiction since it may be in or outside the support set.

Incidentally, if an aside, when looking for examples, I played with the distribution

$\dfrac{1}{2}\mathfrak{U}(0,\theta)+\dfrac{1}{2}\mathfrak{U}(\theta,1)$

which has θ as its theoretical median if not mean. In this example, not only the sample median is not sufficient (the only sufficient statistic is the order statistic and rightly so since the support is fixed and the distributions not in an exponential family), but the MLE is also different from the sample median. Here is an example with n=30 observations, the sienna bar being the sample median:

## improved approximate-Bayesian model-choice method for estimating shared evolutionary history [reply from the author]

Posted in Books, Statistics, University life with tags , , , , , , , , , , , , on June 3, 2014 by xi'an

[Here is a very kind and detailed reply from Jamie Oakes to the comments I made on his ABC paper a few days ago:]

First of all, many thanks for your thorough review of my pre-print! It is very helpful and much appreciated. I just wanted to comment on a few things you address in your post.

I am a little confused about how my replacement of continuous uniform probability distributions with gamma distributions for priors on several parameters introduces a potentially crippling number of hyperparameters. Both uniform and gamma distributions have two parameters. So, the new model only has one additional hyperparameter compared to the original msBayes model: the concentration parameter on the Dirichlet process prior on divergence models. Also, the new model offers a uniform prior over divergence models (though I don’t recommend it).

Your comment about there being no new ABC technique is 100% correct. The model is new, the ABC numerical machinery is not. Also, your intuition is correct, I do not use the divergence times to calculate summary statistics. I mention the divergence times in the description of the ABC algorithm with the hope of making it clear that the times are scaled (see Equation (12)) prior to the simulation of the data (from which the summary statistics are calculated). This scaling is simply to go from units proportional to time, to units that are proportional to the expected number of mutations. Clearly, my attempt at clarity only created unnecessary opacity. I’ll have to make some edits.

Regarding the reshuffling of the summary statistics calculated from different alignments of sequences, the statistics are not exchangeable. So, reshuffling them in a manner that is not conistent across all simulations and the observed data is not mathematically valid. Also, if elements are exchangeable, their order will not affect the likelihood (or the posterior, barring sampling error). Thus, if our goal is to approximate the likelihood, I would hope the reshuffling would also have little affect on the approximate posterior (otherwise my approximation is not so good?).

You are correct that my use of “bias” was not well defined in reference to the identity line of my plots of the estimated vs true probability of the one-divergence model. I think we can agree that, ideally (all assumptions are met), the estimated posterior probability of a model should estimate the probability that the model is correct. For large numbers of simulation
replicates, the proportion of the replicates for which the one-divergence model is true will approximate the probability that the one-divergence model is correct. Thus, if the method has the desirable (albeit “frequentist”) behavior such that the estimated posterior probability of the one-divergence model is an unbiased estimate of the probability that the one-divergence model is correct, the points should fall near the identity line. For example, let us say the method estimates a posterior probability of 0.90 for the one-divergence model for 1000 simulated datasets. If the method is accurately estimating the probability that the one-divergence model is the correct model, then the one-divergence model should be the true model for approximately 900 of the 1000 datasets. Any trend away from the identity line indicates the method is biased in the (frequentist) sense that it is not correctly estimating the probability that the one-divergence model is the correct model. I agree this measure of “bias” is frequentist in nature. However, it seems like a worthwhile goal for Bayesian model-choice methods to have good frequentist properties. If a method strongly deviates from the identity line, it is much more difficult to interpret the posterior probabilites that it estimates. Going back to my example of the posterior probability of 0.90 for 1000 replicates, I would be alarmed if the model was true in only 100 of the replicates.

My apologies if my citation of your PNAS paper seemed misleading. The citation was intended to be limited to the context of ABC methods that use summary statistics that are insufficient across the models under comparison (like msBayes and the method I present in the paper). I will definitely expand on this sentence to make this clearer in revisions. Thanks!

Lastly, my concluding remarks in the paper about full-likelihood methods in this domain are not as lofty as you might think. The likelihood function of the msBayes model is tractable, and, in fact, has already been derived and implemented via reversible-jump MCMC (albeit, not readily available yet). Also, there are plenty of examples of rich, Kingman-coalescent models implemented in full-likelihood Bayesian frameworks. Too many to list, but a lot of them are implemented in the BEAST software package. One noteworthy example is the work of Bryant et al. (2012, Molecular Biology and Evolution, 29(8), 1917–32) that analytically integrates over all gene trees for biallelic markers under the coalescent.

## ABC with indirect summary statistics

Posted in Statistics, University life with tags , , , , , , , on February 3, 2014 by xi'an

After reading Drovandi’s and Pettitt’s Bayesian Indirect Inference, I checked (in the plane to Birmingham) the earlier Gleim’s and Pigorsch’s Approximate Bayesian Computation with indirect summary statistics. The setting is indeed quite similar to the above, with a description of three ways of connecting indirect inference with ABC, albeit with a different range of illustrations. This preprint states most clearly its assumption that the generating model is a particular case of the auxiliary model, which sounds anticlimactic since the auxiliary model is precisely used because the original one is mostly out of reach! This certainly was the original motivation for using indirect inference.

The part of the paper that I find the most intriguing is the argument that the indirect approach leads to sufficient summary statistics, in the sense that they “are sufficient for the parameters of the auxiliary model and (…) sufficiency carries over to the model of interest” (p.31). Looking at the details in the Appendix, I found that the argument is lacking, because the likelihood as a functional is shown to be a (sufficient) statistic, which seems both a tautology and irrelevant because this is different from the likelihood considered at the (auxiliary) MLE, which is the summary statistic used in fine.

“…we expand the square root of an innovation density h in a Hermite expansion and truncate the in finite polynomial at some integer K which, together with other tuning parameters of the SNP density, has to be determined through a model selection criterion (such as BIC). Now we take the leading term of the Hermite expansion to follow a Gaussian GARCH model.”

As in Drovandi and Pettitt, the performances of the ABC-I schemes are tested on a toy example, which is a very basic exponential iid sample with a conjugate prior. With a gamma model as auxiliary. The authors use a standard ABC based on the first two moments as their benchmark, however they do not calibrate those moments in the distance and end up with poor performances of ABC (in a setting where there is a sufficient statistic!). The best choice in this experiment appears as the solution based on the score, but the variances of the distances are not included in the comparison tables. The second implementation considered in the paper is a rather daunting continuous-time non-Gaussian Ornstein-Uhlenbeck stochastic volatility model à la Barndorf -Nielsen and Shephard (2001). The construction of the semi-nonparametric (why not semi-parametric?) auxiliary model is quite involved as well, as illustrated by the quote above. The approach provides an answer, with posterior ABC-IS distributions on all parameters of the original model, which kindles the question of the validation of this answer in terms of the original posterior. Handling simultaneously several approximation processes would help in this regard.

## the likelihood principle (sequel)

Posted in Statistics with tags , , , , , on November 30, 2012 by xi'an

As mentioned in my review of Paradoxes in Scientific Inference I was a bit confused by this presentation of the likelihood principle and this led me to ponder for a week or so whether or not there was an issue with Birnbaum’s proof (or, much more likely, with my vision of it!). After reading again Birnbaum’s proof, while sitting down in a quiet room at ICERM for a little while, I do not see any reason to doubt it. (Keep reading at your own risk!)

My confusion was caused by mixing sufficiency in the sense of Birnbaum’s mixed experiment with sufficiency in the sense of our ABC model choice PNAS paper, namely that sufficient statistics are not always sufficient to select the right model. The sufficient statistics in the proof reduces the (2,x2) observation from Model 2 to (1,x1) from Model 1 when there is an observation x1 that produces a likelihood proportional to the likelihood for x2 and the statistic is indeed sufficient: the distribution of (2,x2) given (1,x1) does not depend on the parameter θ. Of course, the statistic is not sufficient (most of the time) for deciding between Model 1 and Model 2, but this model choice issue is foreign to Birnbaum’s construction.