Bayesian goodness of fit

 

Persi Diaconis and Guanyang Wang have just arXived an interesting reflection on the notion of Bayesian goodness of fit tests. Which is a notion that has always bothered me, in a rather positive sense (!), as

“I also have to confess at the outset to the zeal of a convert, a born again believer in stochastic methods. Last week, Dave Wright reminded me of the advice I had given a graduate student during my algebraic geometry days in the 70’s :`Good Grief, don’t waste your time studying statistics. It’s all cookbook nonsense.’ I take it back! …” David Mumford

The paper starts with a reference to David Mumford, whose paper with Wu and Zhou on exponential “maximum entropy” synthetic distributions is at the source (?) of this paper, and whose name appears in its very title: “A conversation for David Mumford”…, about his conversion from pure (algebraic) maths to applied maths. The issue of (Bayesian) goodness of fit is addressed, with card shuffling examples, the null hypothesis being that the permutation resulting from the shuffling is uniformly distributed if shuffling takes enough time. Interestingly, while the parameter space is compact as a distribution on a finite set, Lindley’s paradox still occurs, namely that the null (the permutation comes from a Uniform) is always accepted provided there is no repetition under a “flat prior”, which is the Dirichlet D(1,…,1) over all permutations. (In this finite setting an improper prior is definitely improper as it does not get proper after accounting for observations. Although I do not understand why the Jeffreys prior is not the Dirichlet(½,…,½) in this case…) When resorting to the exponential family of distributions entertained by Zhou, Wu and Mumford, including the uniform distribution as one of its members, Diaconis and Wang advocate the use of a conjugate prior (exponential family, right?!) to compute a Bayes factor that simplifies into a ratio of two intractable normalising constants. For which the authors suggest using importance sampling, thermodynamic integration, or the exchange algorithm. Except that they rely on the (dreaded) harmonic mean estimator for computing the Bayes factor in the following illustrative section! Due to the finite nature of the space, I presume this estimator still has a finite variance. (Remark 1 calls for convergence results on exchange algorithms, which can be found I think in the just as recent arXival by Christophe Andrieu and co-authors.) An interesting if rare feature of the example processed in the paper is that the sufficient statistic used for the permutation model can be directly simulated from a Multinomial distribution. This is rare as seen when considering the benchmark of Ising models, for which the summary and sufficient statistic cannot be directly simulated. (If only…!) In fine, while I enjoyed the paper a lot, I remain uncertain as to its bearings, since defining an objective alternative for the goodness-of-fit test becomes quickly challenging outside simple enough models.

Goodness-of-fit statistics for ABC

“Posterior predictive checks are well-suited to Approximate Bayesian Computation”

Louisiane Lemaire and her coauthors from Grenoble have just arXived a new paper on designing a goodness-of-fit statistic from ABC outputs. The statistic is constructed from a comparison between the observed (summary) statistics and replicated summary statistics generated from the posterior predictive distribution. This is a major difference with the standard ABC distance, when the replicated summary statistics are generated from the prior predictive distribution. The core of the paper is about calibrating a posterior predictive p-value derived from this distance, since it is not properly calibrated in the frequentist sense that it is not uniformly distributed “under the null”. A point I discussed in an ‘Og entry about Andrews’ book a few years ago.

The paper opposes the average distance between ABC acceptable summary statistics and the observed realisation to the average distance between ABC posterior predictive simulations of summary statistics and the observed realisation. In the simplest case (e.g., without a post-processing of the summary statistics), the main difference between both average distances is that the summary statistics are used twice in the first version: first to select the acceptable values of the parameters and a second time for the average distance. Which makes it biased downwards. The second version is more computationally demanding, especially when deriving the associated p-value. It however produces a higher power under the alternative. Obviously depending on how the alternative is defined, since goodness-of-fit is only related to the null, i.e., to a specific model.

From a general perspective, I do not completely agree with the conclusions of the paper in that (a) this is a frequentist assessment and partakes in the shortcomings of p-values and (b) the choice of summary statistics has a huge impact on the decision about the fit since hardly varying statistics are more likely to lead to a good fit than appropriately varying ones.

posterior predictive checks for admixture models

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

(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 ẑ, 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…

posterior predictive p-values

Bayesian Data Analysis advocates in Chapter 6 using posterior predictive checks as a way of evaluating the fit of a potential model to the observed data. There is a no-nonsense feeling to it:

“If the model fits, then replicated data generated under the model should look similar to observed data. To put it another way, the observed data should look plausible under the posterior predictive distribution.”

And it aims at providing an answer to the frustrating (frustrating to me, at least) issue of Bayesian goodness-of-fit tests. There are however issues with the implementation, from deciding on which aspect of the data or of the model is to be examined, to the “use of the data twice” sin. Obviously, this is an exploratory tool with little decisional backup and it should be understood as a qualitative rather than quantitative assessment. As mentioned in my tutorial on Sunday (I wrote this post in Duke during O’Bayes 2013), it reminded me of Ratmann et al.’s ABC_μ in that they both give reference distributions against which to calibrate the observed data. Most likely with a multidimensional representation. And the “use of the data twice” can be argued for or against, once a data-dependent loss function is built.

“One might worry about interpreting the significance levels of multiple tests or of tests chosen by inspection of the data (…) We do not make [a multiple test] adjustment, because we use predictive checks to see how particular aspects of the data would be expected to appear in replications. If we examine several test variables, we would not be surprised for some of them not to be fitted by the model-but if we are planning to apply the model, we might be interested in those aspects of the data that do not appear typical.”

The natural objection that having a multivariate measure of discrepancy runs into multiple testing is answered within the book with the reply that the idea is not to run formal tests. I still wonder how one should behave when faced with a vector of posterior predictive p-values (ppp).

The above picture is based on a normal mean/normal prior experiment I ran where the ratio prior-to-sampling variance increases from 100 to 10⁴. The ppp is based on the Bayes factor against a zero mean as a discrepancy. It thus grows away from zero very quickly and then levels up around 0.5, reaching only values close to 1 for very large values of x (i.e. never in practice). I find the graph interesting because if instead of the Bayes factor I use the marginal (numerator of the Bayes factor) then the picture is the exact opposite. Which, I presume, does not make a difference for Bayesian Data Analysis, since both extremes are considered as equally toxic… Still, still, still, we are is the same quandary as when using any kind of p-value: what is extreme? what is significant? Do we have again to select the dreaded 0.05?! To see how things are going, I then simulated the behaviour of the ppp under the “true” model for the pair (θ,x). And ended up with the histograms below:

which shows that under the true model the ppp does concentrate around .5 (surprisingly the range of ppp’s hardly exceeds .5 and I have no explanation for this). While the corresponding ppp does not necessarily pick any wrong model, discrepancies may be spotted by getting away from 0.5…

“The p-value is to the u-value as the posterior interval is to the confidence interval. Just as posterior intervals are not, in general, classical confidence intervals, Bayesian p-values are not generally u-values.”

Now, Bayesian Data Analysis also has this warning about ppp’s being not uniform under the true model (u-values), which is just as well considering the above example, but I cannot help wondering if the authors had intended a sort of subliminal message that they were not that far from uniform. And this brings back to the forefront the difficult interpretation of the numerical value of a ppp. That is, of its calibration. For evaluation of the fit of a model. Or for decision-making…

Assessing models when all models are false

When I arrived home from Philadelphia, I got the news that John Geweke was giving a seminar at CREST in the early afternoon and thus biked there to attend his talk. The topic was about comparing asset return models, but what interested most was the notion of comparing models without a reference to a true model, a difficulty I have been juggling with for quite a while (at least since the goodness-of-fit paper with Judith Rousseau we never resubmitted!). And for which I still do not find a satisfactory (Bayesian) solution.

Because there is no true model, Durham and Geweke use the daily (possibly Bayesian) predictive

as their basis for model assessment and rely on a log scoring rule

to compare models. (The ‘o’ in the superscript denotes the observed values.) As reported in the paper this is a proper (or honest) scoring rule. If n models are under competition, a weighted (model) predictive average

can be considered and the paper examines the impact of picking the optimal weight vector against the log scoring rule, i.e.

The weight vector at time t-1 is therefore optimising the backward sequence of predictions of the observed values till time t-1. The interesting empirical result from this study is that, even from a Bayesian perspective, the weights never degenerate, unless one of the models is correct (which is rarely the case!). Thus, even after very long series of observations, the weights of different models remain away from zero (while the Bayesian posterior probability of a single model goes to one). Even though I am not yet at the point of adopting this solution (in particular because it seems to be using the data twice, once through the posterior/predictive and once through the score), I find the approach quite intriguing and hope I can study it further. Maybe a comparison with a Bayesian non-parametric evaluation would make sense…