## about the strong likelihood principle

Posted in Books, Statistics, University life with tags , , , , , , , on November 13, 2014 by xi'an

Deborah Mayo arXived a Statistical Science paper a few days ago, along with discussions by Jan Bjørnstad, Phil Dawid, Don Fraser, Michael Evans, Jan Hanning, R. Martin and C. Liu. I am very glad that this discussion paper came out and that it came out in Statistical Science, although I am rather surprised to find no discussion by Jim Berger or Robert Wolpert, and even though I still cannot entirely follow the deductive argument in the rejection of Birnbaum’s proof, just as in the earlier version in Error & Inference.  But I somehow do not feel like going again into a new debate about this critique of Birnbaum’s derivation. (Even though statements like the fact that the SLP “would preclude the use of sampling distributions” (p.227) would call for contradiction.)

“It is the imprecision in Birnbaum’s formulation that leads to a faulty impression of exactly what  is proved.” M. Evans

Indeed, at this stage, I fear that [for me] a more relevant issue is whether or not the debate does matter… At a logical cum foundational [and maybe cum historical] level, it makes perfect sense to uncover if and which if any of the myriad of Birnbaum’s likelihood Principles holds. [Although trying to uncover Birnbaum’s motives and positions over time may not be so relevant.] I think the paper and the discussions acknowledge that some version of the weak conditionality Principle does not imply some version of the strong likelihood Principle. With other logical implications remaining true. At a methodological level, I am less much less sure it matters. Each time I taught this notion, I got blank stares and incomprehension from my students, to the point I have now stopped altogether teaching the likelihood Principle in class. And most of my co-authors do not seem to care very much about it. At a purely mathematical level, I wonder if there even is ground for a debate since the notions involved can be defined in various imprecise ways, as pointed out by Michael Evans above and in his discussion. At a statistical level, sufficiency eventually is a strange notion in that it seems to make plenty of sense until one realises there is no interesting sufficiency outside exponential families. Just as there are very few parameter transforms for which unbiased estimators can be found. So I also spend very little time teaching and even less worrying about sufficiency. (As it happens, I taught the notion this morning!) At another and presumably more significant statistical level, what matters is information, e.g., conditioning means adding information (i.e., about which experiment has been used). While complex settings may prohibit the use of the entire information provided by the data, at a formal level there is no argument for not using the entire information, i.e. conditioning upon the entire data. (At a computational level, this is no longer true, witness ABC and similar limited information techniques. By the way, ABC demonstrates if needed why sampling distributions matter so much to Bayesian analysis.)

“Non-subjective Bayesians who (…) have to live with some violations of the likelihood principle (…) since their prior probability distributions are influenced by the sampling distribution.” D. Mayo (p.229)

In the end, the fact that the prior may depend on the form of the sampling distribution and hence does violate the likelihood Principle does not worry me so much. In most models I consider, the parameters are endogenous to those sampling distributions and do not live an ethereal existence independently from the model: they are substantiated and calibrated by the model itself, which makes the discussion about the LP rather vacuous. See, e.g., the coefficients of a linear model. In complex models, or in large datasets, it is even impossible to handle the whole data or the whole model and proxies have to be used instead, making worries about the structure of the (original) likelihood vacuous. I think we have now reached a stage of statistical inference where models are no longer accepted as ideal truth and where approximation is the hard reality, imposed by the massive amounts of data relentlessly calling for immediate processing. Hence, where the self-validation or invalidation of such approximations in terms of predictive performances is the relevant issue. Provided we can at all face the challenge…

## Relevant statistics for Bayesian model choice [hot off the press!]

Posted in Books, Statistics, University life with tags , , , , , , on October 30, 2014 by xi'an

Our paper about evaluating statistics used for ABC model choice has just appeared in Series B! It somewhat paradoxical that it comes out just a few days after we submitted our paper on using random forests for Bayesian model choice, thus bypassing the need for selecting those summary statistics by incorporating all statistics available and letting the trees automatically rank those statistics in term of their discriminating power. Nonetheless, this paper remains an exciting piece of work (!) as it addresses the more general and pressing question of the validity of running a Bayesian analysis with only part of the information contained in the data. Quite usefull in my (biased) opinion when considering the emergence of approximate inference already discussed on this ‘Og…

[As a trivial aside, I had first used fresh from the press(es) as the bracketted comment, before I realised the meaning was not necessarily the same in English and in French.]

## reliable ABC model choice via random forests

Posted in pictures, R, Statistics, University life with tags , , , , , , , on October 29, 2014 by xi'an

After a somewhat prolonged labour (!), we have at last completed our paper on ABC model choice with random forests and submitted it to PNAS for possible publication. While the paper is entirely methodological, the primary domain of application of ABC model choice methods remains population genetics and the diffusion of this new methodology to the users is thus more likely via a media like PNAS than via a machine learning or statistics journal.

When compared with our recent update of the arXived paper, there is not much different in contents, as it is mostly an issue of fitting the PNAS publication canons. (Which makes the paper less readable in the posted version [in my opinion!] as it needs to fit the main document within the compulsory six pages, relegated part of the experiments and of the explanations to the Supplementary Information section.)

## ABC model choice via random forests [expanded]

Posted in Statistics, University life with tags , , , , , , , , , , , on October 1, 2014 by xi'an

Today, we arXived a second version of our paper on ABC model choice with random forests. Or maybe [A]BC model choice with random forests. Since the random forest is built on a simulation from the prior predictive and no further approximation is used in the process. Except for the computation of the posterior [predictive] error rate. The update wrt the earlier version is that we ran massive simulations throughout the summer, on existing and new datasets. In particular, we have included a Human dataset extracted from the 1000 Genomes Project. Made of 51,250 SNP loci. While this dataset is not used to test new evolution scenarios, we compared six out-of-Africa scenarios, with a possible admixture for Americans of African ancestry. The scenario selected by a random forest procedure posits a single out-of-Africa colonization event with a secondary split into a European and an East Asian population lineages, and a recent genetic admixture between African and European lineages, for Americans of African origin. The procedure reported a high level of confidence since the estimated posterior error rate is equal to zero. The SNP loci were carefully selected using the following criteria: (i) all individuals have a genotype characterized by a quality score (GQ)>10, (ii) polymorphism is present in at least one of the individuals in order to fit the SNP simulation algorithm of Hudson (2002) used in DIYABC V2 (Cornuet et al., 2014), (iii) the minimum distance between two consecutive SNPs is 1 kb in order to minimize linkage disequilibrium between SNP, and (iv) SNP loci showing significant deviation from Hardy-Weinberg equilibrium at a 1% threshold in at least one of the four populations have been removed.

In terms of random forests, we optimised the size of the bootstrap subsamples for all of our datasets. While this optimisation requires extra computing time, it is negligible when compared with the enormous time taken by a logistic regression, which is [yet] the standard ABC model choice approach. Now the data has been gathered, it is only a matter of days before we can send the paper to a journal

## ABC model choice by random forests [guest post]

Posted in pictures, R, Statistics, University life with tags , , , , , , , , , , on August 11, 2014 by xi'an

[Dennis Prangle sent me his comments on our ABC model choice by random forests paper. Here they are! And I appreciate very much contributors commenting on my paper or others, so please feel free to join.]

This paper proposes a new approach to likelihood-free model choice based on random forest classifiers. These are fit to simulated model/data pairs and then run on the observed data to produce a predicted model. A novel “posterior predictive error rate” is proposed to quantify the degree of uncertainty placed on this prediction. Another interesting use of this is to tune the threshold of the standard ABC rejection approach, which is outperformed by random forests.

The paper has lots of thought-provoking new ideas and was an enjoyable read, as well as giving me the encouragement I needed to read another chapter of the indispensable Elements of Statistical Learning However I’m not fully convinced by the approach yet for a few reasons which are below along with other comments.

Alternative schemes

The paper shows that random forests outperform rejection based ABC. I’d like to see a comparison to more efficient ABC model choice algorithms such as that of Toni et al 2009. Also I’d like to see if the output of random forests could be used as summary statistics within ABC rather than as a separate inference method.

Posterior predictive error rate (PPER)

This is proposed to quantify the performance of a classifier given a particular data set. The PPER is the proportion of times the classifier’s most favoured model is incorrect for simulated model/data pairs drawn from an approximation to the posterior predictive. The approximation is produced by a standard ABC analysis.

Misclassification could be due to (a) a poor classifier or (b) uninformative data, so the PPER aggregrates these two sources of uncertainty. I think it is still very desirable to have an estimate of the uncertainty due to (b) only i.e. a posterior weight estimate. However the PPER is useful. Firstly end users may sometimes only care about the aggregated uncertainty. Secondly relative PPER values for a fixed dataset are a useful measure of uncertainty due to (a), for example in tuning the ABC threshold. Finally, one drawback of the PPER is the dependence on an ABC estimate of the posterior: how robust are the results to the details of how this is obtained?

Classification

This paper illustrates an important link between ABC and machine learning classification methods: model choice can be viewed as a classification problem. There are some other links: some classifiers make good model choice summary statistics (Prangle et al 2014) or good estimates of ABC-MCMC acceptance ratios for parameter inference problems (Pham et al 2014). So the good performance random forests makes them seem a generally useful tool for ABC (indeed they are used in the Pham et al al paper).

## Bangalore workshop [ಬೆಂಗಳೂರು ಕಾರ್ಯಾಗಾರ]

Posted in pictures, Running, Statistics, Travel, University life, Wines with tags , , , , , on July 30, 2014 by xi'an

First day at the Indo-French Centre for Applied Mathematics and the get-together (or speed-dating!) workshop. The campus of the Indian Institute of Science of Bangalore where we all stay is very pleasant with plenty of greenery in the middle of a very busy city. Plus, being at about 1000m means the temperature remains tolerable for me, to the point of letting me run in the morning.Plus, staying in a guest house in the campus also means genuine and enjoyable south Indian food.

The workshop is a mix of statisticians and of mathematicians of neurosciences, from both India and France, and we are few enough to have a lot of opportunities for discussion and potential joint projects. I gave the first talk this morning (hence a fairly short run!) on ABC model choice with random forests and, given the mixed audience, may have launched too quickly into the technicalities of the forests. Even though I think I kept the statisticians on-board for most of the talk. While the mathematical biology talks mostly went over my head (esp. when I could not resist dozing!), I enjoyed the presentation of Francis Bach of a fast stochastic gradient algorithm, where the stochastic average is only updated one term at a time, for apparently much faster convergence results. This is related with a joint work with Éric Moulines that both Éric and Francis presented in the past month. And makes me wonder at the intuition behind the major speed-up. Shrinkage to the mean maybe?

## posterior predictive checks for admixture models

Posted in pictures, Statistics, Travel, University life with tags , , , , , , , on July 8, 2014 by xi'an

In a posting coincidence, just a few days after we arXived our paper on ABC model choice with random forests, where we use posterior predictive errors for assessing the variability of the random forest procedure, David Mimno, David Blei, and Barbara Engelhardt arXived a paper on posterior predictive checks to quantify lack-of-fit in admixture models of latent population structure, which deals with similar data and models, while also using the posterior predictive as a central tool. (Marginalia: the paper is a wee bit difficult to read [esp. with France-Germany playing in the airport bar!] as the modelling is only clearly described at the very end. I suspect this arXived version was put together out of a submission to a journal like Nature or PNAS, with mentions of a Methods section that does not appear here and of Supplementary Material that turned into subsections of the Discussion section.)

The dataset are genomic datasets made of SNPs (single nucleotide polymorphisms). For instance, the first (HapMap) dataset corresponds to 1,043 individuals and 468,167 SNPs. The model is simpler than Kingman’s coalescent, hence its likelihood does not require ABC steps to run inference. The admixture model in the paper is essentially a mixture model over ancestry indices with individual dependent weights with Bernoulli observations, hence resulting into a completed likelihood of the form

$\prod_{i=1}^n\prod_{\ell=1}^L\prod_j \phi_{\ell,z_{i,\ell,j}}^{x_{i,\ell,j}}(1-\phi_{\ell,z_{i,\ell,j}})^{1-x_{i,\ell,j}}\theta_{i,z_{i,\ell,j}}$

(which looks more formidable than it truly is!). Regular Bayesian inference is thus possible in this setting, implementing e.g. Gibbs sampling. The authors chose instead to rely on EM and thus derived the maximum likelihood estimators of the (many) parameters of the admixture. And of the latent variables z. Their posterior predictive check is based on the simulation of pseudo-observations (as in ABC!) from the above likelihood, with parameters and latent variables replaced with their EM estimates (unlike ABC). There is obviously some computational reason in doing this instead of simulating from the posterior, albeit implicit in the paper. I am however slightly puzzled by the conditioning on the latent variable estimate , as its simulation is straightforward and as a latent variable is more a missing observation than a parameter. Given those 30 to 100 replications of the data, an empirical distribution of a discrepancy function is used to assess whether or not the equivalent discrepancy for the observation is an outlier. If so, the model is not appropriate for the data. (Interestingly, the discrepancy is measured via the Bayes factor of z-scores.)

The connection with our own work is that the construction of discrepancy measures proposed in this paper could be added to our already large collection of summary statistics to check to potential impact in model comparison, i.e. for a significant contribution to the random forest nodes.  Conversely, the most significant summary statistics could then be tested as discrepancy measures. Or, more in tune with our Series B paper on the proper selection of summary variables, the distribution of those discrepancy measures could be compared across potential models. Assuming this does not take too much computing power…