Bayesian revision of a prior given prior-data conflict,expert opinion,or a similar insight: a large-deviation approach

Learning from model diagnostics that a prior distribution must be replaced by one that conflicts less with the data raises the question of which prior should instead be used for inference and decision. The same problem arises when a decision maker learns that one or more reliable experts express unexpected beliefs. In both cases, coherence of the solution would be guaranteed by applying Bayes's theorem to a distribution of prior distributions that effectively assigns the initial prior distribution a probability arbitrarily close to 1. The new distribution for inference would then be the distribution of priors conditional on the insight that the prior distribution lies in a closed convex set that does not contain the initial prior. A readily available distribution of priors needed for such conditioning is the law of the empirical distribution of sufficiently large number of independent parameter values drawn from the initial prior. According to the Gibbs conditioning principle from the theory of large deviations, the resulting new prior distribution minimizes the entropy relative to the initial prior. While minimizing relative entropy accommodates the necessity of going beyond the initial prior without departing from it any more than the insight demands, the large-deviation derivation also ensures the advantages of Bayesian coherence. This approach is generalized to uncertain insights by allowing the closed convex set of priors to be random.     

A Graphical Diagnostic for Identifying Influential Model Choices in Bayesian Hierarchical Models

Abstract. Real‐world phenomena are frequently modelled by Bayesian hierarchical models. The building‐blocks in such models are the distribution of each variable conditional on parent and/or neighbour variables in the graph. The specifications of centre and spread of these conditional distributions may be well motivated, whereas the tail specifications are often left to convenience. However, the posterior distribution of a parameter may depend strongly on such arbitrary tail specifications. This is not easily detected in complex models. In this article, we propose a graphical diagnostic, the Local critique plot, which detects such influential statistical modelling choices at the node level. It identifies the properties of the information coming from the parents and neighbours (the local prior) and from the children and co‐parents (the lifted likelihood) that are influential on the posterior distribution, and examines local conflict between these distinct information sources. The Local critique plot can be derived for all parameters in a chain graph model.     

A note on fiducial model averaging as an alternative to checking Bayesian and frequentist models

Just as frequentist hypothesis tests have been developed to check model assumptions, prior predictive p-values and other Bayesian p-values check prior distributions as well as other model assumptions. These model checks not only suffer from the usual threshold dependence of p-values, but also from the suppression of model uncertainty in subsequent inference. One solution is to transform Bayesian and frequentist p-values for model assessment into a fiducial distribution across the models. Averaging the Bayesian or frequentist posterior distributions with respect to the fiducial distribution can reproduce results from Bayesian model averaging or classical fiducial inference.     

Dropping observations without affecting posterior and predictive distributions

Suppose in a distribution problem, the sample information W is split into two pieces W ₁ and W ₂, and the parameters involved are split into two sets, π containing the parameters of interest, and θ containing nuisance parameters. It is shown that, under certain conditions, the posterior distribution of π does not depend on the data W ₂, which can thus be ignored. This also has consequences for the predictive distribution of future (or missing) observations. In fact, under similar conditions, the predictive distributions using W or just W ₁ are identical.     

A GRAPHICAL TECHNIQUE FOR DETECTING INFLUENTIAL CASES IN REGRESSION ANALYSIS

This paper presents a graphical technique for detecting influential cases in regression analysis. The idea is to decompose a diagnostic problem involving higher order dimensional regression problems, into a series of two-dimensional diagnostic sub-problems, such that the diagnoses of influential cases is undertaken by visually inspecting two-dimensional diagnostic plots of these sub-problems. An algorithm for the graphical procedure is proposed to reduce the computational effort. Practical examples are used to illustrate this graphical technique.     

Bayesian forecasting for AR(1) models with normal coefficients

This paper presents a Bayesian solution to the problem of time series forecasting, for the case in which the generating process is an autoregressive of order one, with a normal random coefficient. The proposed procedure is based on the predictive density of the future observation. Conjugate priors are used for some parameters, while improper vague priors are used for others.     

Alternatives to post‐processing posterior predictive p values

The posterior predictive p value (ppp) was invented as a Bayesian counterpart to classical p values. The methodology can be applied to discrepancy measures involving both data and parameters and can, hence, be targeted to check for various modeling assumptions. The interpretation can, however, be difficult since the distribution of the ppp value under modeling assumptions varies substantially between cases. A calibration procedure has been suggested, treating the ppp value as a test statistic in a prior predictive test. In this paper, we suggest that a prior predictive test may instead be based on the expected posterior discrepancy, which is somewhat simpler, both conceptually and computationally. Since both these methods require the simulation of a large posterior parameter sample for each of an equally large prior predictive data sample, we furthermore suggest to look for ways to match the given discrepancy by a computation‐saving conflict measure. This approach is also based on simulations but only requires sampling from two different distributions representing two contrasting information sources about a model parameter. The conflict measure methodology is also more flexible in that it handles non‐informative priors without difficulty. We compare the different approaches theoretically in some simple models and in a more complex applied example.     

A note on nonexistence of posterior moments

Bayesian analyses often take for granted the assumption that the posterior distribution has at least a first moment. They often include computed or estimated posterior means. In this note, the authors show an example of a Weibull distribution parameter where the theoretical posterior mean fails to exist for commonly used proper semi–conjugate priors. They also show that posterior moments can fail to exist with commonly used noninformative priors including Jeffreys, reference and matching priors, despite the fact that the posteriors are proper. Moreover, within a broad class of priors, the predictive distribution also has no mean. The authors illustrate the problem with a simulated example. Their results demonstrate that the unwitting use of estimated posterior means may yield unjustified conclusions.     

Posterior Predictive p-values in Bayesian Hierarchical Models

Abstract.  The present work focuses on extensions of the posterior predictive p -value (ppp-value) for models with hierarchical structure, designed for testing assumptions made on underlying processes. The ppp-values are popular as tools for model criticism, yet their lack of a common interpretation limit their practical use. We discuss different extensions of ppp-values to hierarchical models, allowing for discrepancy measures that can be used for checking properties of the model at all stages. Through analytical derivations and simulation studies on simple models, we show that similar to the standard ppp-values, these extensions are typically far from uniformly distributed under the model assumptions and can give poor power in a hypothesis testing framework. We propose a calibration of the p -values, making the resulting calibrated p -values uniformly distributed under the model conditions. Illustrations are made through a real example of multinomial regression to age distributions of fish.     

A test for diagnostic utility

A criterion of usefulness for a diagnostic test is suggested. Based on sensitivity and specificity data, Bayesian and likelihood-ratio procedures to examine whether the criterion is satisfied are presented.     

DATA‐DRIVEN SMOOTH TESTS AND A DIAGNOSTIC TOOL FOR LACK‐OF‐FIT FOR CIRCULAR DATA

Two contributions to the statistical analysis of circular data are given. First we construct data‐driven smooth goodness‐of‐fit tests for the circular von Mises assumption. As a second method, we propose a new graphical diagnostic tool for the detection of lack‐of‐fit for circular distributions. We illustrate our methods on two real datasets.     

Multivariate regression analysis of panel data with binary outcomes applied to unemployment data

Summary  In panel studies binary outcome measures together with time stationary and time varying explanatory variables are collected over time on the same individual. Therefore, a regression analysis for this type of data must allow for the correlation among the outcomes of an individual. The multivariate probit model of Ashford and Sowden (1970) was the first regression model for multivariate binary responses. However, a likelihood analysis of the multivariate probit model with general correlation structure for higher dimensions is intractable due to the maximization over high dimensional integrals thus severely restricting ist applicability so far. Czado (1996) developed a Markov Chain Monte Carlo (MCMC) algorithm to overcome this difficulty. In this paper we present an application of this algorithm to unemployment data from the Panel Study of Income Dynamics involving 11 waves of the panel study. In addition we adapt Bayesian model checking techniques based on the posterior predictive distribution (see for example Gelman et al. (1996)) for the multivariate probit model. These help to identify mean and correlation specification which fit the data well. C. Czado was supported by research grant OGP0089858 of the Natural Sciences and Engineering Research Council of Canada.     

A bayesian analysis of the mixed linear model

The mixed model is defined. The exact posterior distribution for the fixed effect vector is obtained. The exact posterior distribution for the error variance is obtained. The exact posterior mean and variance of a Bayesian estimator for the variances of random effects is also derived. All computations are non-iterative and avoid numerical integrations.     

Discrepancy measures for item fit analysis in item response theory

Item response theory (IRT) models are commonly used in educational and psychological testing to assess the (latent) ability of examinees and the effectiveness of the test items in measuring this underlying trait. The focus of this paper is on the assessment of item fit for unidimensional IRT models for dichotomous items using a Bayesian method. This paper will illustrate and compare the effectiveness of several discrepancy measures, used within the posterior predictive model check procedure, in detecting misfitted items. The effectiveness of the different discrepancy measures are illustrated in a simulation study using artificially altered simulated data. Using the best discrepancy measure among those studied, this method was applied to real data coming from a mathematics placement exam.     

Posterior Analysis for Normalized Random Measures with Independent Increments

Abstract.  One of the main research areas in Bayesian Nonparametrics is the proposal and study of priors which generalize the Dirichlet process. In this paper, we provide a comprehensive Bayesian non-parametric analysis of random probabilities which are obtained by normalizing random measures with independent increments (NRMI). Special cases of these priors have already shown to be useful for statistical applications such as mixture models and species sampling problems. However, in order to fully exploit these priors, the derivation of the posterior distribution of NRMIs is crucial: here we achieve this goal and, indeed, provide explicit and tractable expressions suitable for practical implementation. The posterior distribution of an NRMI turns out to be a mixture with respect to the distribution of a specific latent variable. The analysis is completed by the derivation of the corresponding predictive distributions and by a thorough investigation of the marginal structure. These results allow to derive a generalized Blackwell–MacQueen sampling scheme, which is then adapted to cover also mixture models driven by general NRMIs.     

Posterior Predictive Comparisons for the Two-sample Problem

The two-sample problem of inferring whether two random samples have equal underlying distributions is formulated within the Bayesian framework as a comparison of two posterior predictive inferences rather than as a problem of model selection. The suggested approach is argued to be particularly advantageous in problems where the objective is to evaluate evidence in support of equality, along with being robust to the priors used and being capable of handling improper priors. Our approach is contrasted with the Bayes factor in a normal setting and finally, an additional example is considered where the observed samples are realizations of Markov chains.     

A Bayesian cluster analysis of election results

A Bayesian cluster analysis for the results of an election based on multinomial mixture models is proposed. The number of clusters is chosen based on the careful comparison of the results with predictive simulations from the models, and by checking whether models capture most of the spatial dependence in the results. By implementing the analysis on five recent elections in Barcelona, the reader is walked through the choice of the best statistics and graphical displays to help chose a model and present the results. Even though the models do not use any information about the location of the areas in which the results are broken into, in the example they uncover a four-cluster structure with a strong spatial dependence, that is very stable over time and relates to the demographic composition.     

Bayesian prediction of unobserved values for Type-II censored data

In this paper, we consider posterior predictive distributions of Type-II censored data for an inverse Weibull distribution. These functions are given by using conditional density functions and conditional survival functions. Although the conditional survival functions were expressed by integral forms in previous studies, we derive the conditional survival functions in closed forms and thereby reduce the computation cost. In addition, we calculate the predictive confidence intervals of unobserved values and coverage probabilities of unobserved values by using the posterior predictive survival functions.     

ESTIMATION, PREDICTION AND INFERENCE FOR THE LASSO RANDOM EFFECTS MODEL

The least absolute shrinkage and selection operator (LASSO) can be formulated as a random effects model with an associated variance parameter that can be estimated with other components of variance. In this paper, estimation of the variance parameters is performed by means of an approximation to the marginal likelihood of the observed outcomes. The approximation is based on an alternative but equivalent formulation of the LASSO random effects model. Predictions can be made using point summaries of the predictive distribution of the random effects given the data with the parameters set to their estimated values. The standard LASSO method uses the mode of this distribution as the predictor. It is not the only choice, and a number of other possibilities are defined and empirically assessed in this article. The predictive mode is competitive with the predictive mean (best predictor), but no single predictor performs best across in all situations. Inference for the LASSO random effects is performed using predictive probability statements, which are more appropriate under the random effects formulation than tests of hypothesis.