Modern datasets are often in the form of matrices or arrays, potentially having correlations along each set of data indices. For example, data involving repeated measurements of several variables over time may exhibit temporal correlation as well as correlation among the variables. A possible model for matrix-valued data is the class of matrix normal distributions, which is parametrized by two covariance matrices, one for each index set of the data. In this article we discuss an extension of the matrix normal model to accommodate multidimensional data arrays, or tensors. We show how a particular array-matrix product can be used to generate the class of array normal distributions having separable covariance structure. We derive some properties of these covariance structures and the corresponding array normal distributions, and show how the array-matrix product can be used to define a semi-conjugate prior distribution and calculate the corresponding posterior distribution. We illustrate the methodology in an analysis of multivariate longitudinal network data which take the form of a four-way array.

In computational biology, gene expression datasets are characterized by very few individual samples compared to a large number of measurements per sample. Thus, it is appealing to merge these datasets in order to increase the number of observations and diversify the data, allowing a more reliable selection of genes relevant to the biological problem. Besides, the increased size of a merged dataset facilitates its re-splitting into training and validation sets. This necessitates the introduction of the dataset as a random effect. In this context, extending a work of Lee et al. (2003), a method is proposed to select relevant variables among tens of thousands in a probit mixed regression model, considered as part of a larger hierarchical Bayesian model. Latent variables are used to identify subsets of selected variables and the grouping (or blocking) technique of Liu (1994) is combined with a Metropolis-within-Gibbs algorithm (Robert and Casella 2004). The method is applied to a merged dataset made of three individual gene expression datasets, in which tens of thousands of measurements are available for each of several hundred human breast cancer samples. Even for this large dataset comprised of around 20000 predictors, the method is shown to be efficient and feasible. As an illustration, it is used to select the most important genes that characterize the estrogen receptor status of patients with breast cancer.

In this article we examine two relatively new MCMC methods which allow for Bayesian inference in diffusion models. First, the Monte Carlo within Metropolis (MCWM) algorithm (O'neil, et al. 2000) uses an importance sampling approximation for the likelihood and yields a Markov chain. Our simulation study shows that there exists a limiting stationary distribution that can be made arbitrarily ``close'' to the posterior distribution (MCWM is not a standard Metropolis-Hastings algorithm, however). The second method, described in Beaumont (2003) and generalized in Andrieu and Roberts (2009), introduces auxiliary variables and utilizes a standard Metropolis-Hastings algorithm on the enlarged space; this method preserves the original posterior distribution. When applied to diffusion models, this pseudo-marginal (PM) approach can be viewed as a generalization of the popular data augmentation schemes that sample jointly from the missing paths and the parameters of the diffusion volatility. The efficacy of the PM approach is demonstrated in a simulation study of the Cox-Ingersoll-Ross (CIR) and Heston models, and is applied to two well known datasets. Comparisons are made with the MCWM algorithm and the Golightly and Wilkinson (2006) approach.

Generalized linear mixed models (GLMMs) enjoy increasing popularity because of their ability to model correlated observations. Integrated nested Laplace approximations (INLAs) provide a fast implementation of the Bayesian approach to GLMMs. However, sensitivity to prior assumptions on the random effects precision parameters is a potential problem. To quantify the sensitivity to prior assumptions, we develop a general sensitivity measure based on the Hellinger distance to assess sensitivity of the posterior distributions with respect to changes in the prior distributions for the precision parameters. In addition, for model selection we suggest several cross-validatory techniques for Bayesian GLMMs with a dichotomous outcome. Although the proposed methodology holds in greater generality, we make use of the developed methods in the particular context of the well-known salamander mating data. We arrive at various new findings with respect to the best fitting model and the sensitivity of the estimates of the model components.

A new tree-based graphical model --- the dynamic staged tree --- is proposed for modelling discrete-valued discrete-time multivariate processes which are hypothesised to exhibit symmetries in how some intermediate situations might unfold. We define and implement a one-step-ahead prediction algorithm with the model using multi-process modelling and the power steady model that is robust to short-term variations in the data yet sensitive to underlying system changes. We demonstrate that the whole analysis can be performed in a conjugate way so that the potentially vast model space can be traversed quickly and then results communicated transparently. We also demonstrate how to analyse a general set of causal hypotheses on this model class. Our techniques are illustrated using a simple educational example.

This paper describes an approach for Bayesian modeling in spherical data sets. Our method is based upon a recent construction called the needlet, which is a particular form of spherical wavelet with many favorable statistical and computational properties. We perform shrinkage and selection of needlet coefficients, focusing on two main alternatives: empirical-Bayes thresholding, and Bayesian local shrinkage rules. We study the performance of the proposed methodology both on simulated data and on two real data sets: one involving the cosmic microwave background radiation, and one involving the reconstruction of a global news intensity surface inferred from published Reuters articles in August, 1996. The fully Bayesian approach based on robust, sparse shrinkage priors seems to outperform other alternatives.

We propose a hierarchical Bayesian nonparametric mixture model for clustering when some of the covariates are assumed to be of varying relevance to the clustering problem. This can be thought of as an issue in variable selection for unsupervised learning. We demonstrate that by defining a hierarchical population based nonparametric prior on the cluster locations scaled by the inverse covariance matrices of the likelihood we arrive at a `sparsity prior' representation which admits a conditionally conjugate prior. This allows us to perform full Gibbs sampling to obtain posterior distributions over parameters of interest including an explicit measure of each covariate's relevance and a distribution over the number of potential clusters present in the data. This also allows for individual cluster specific variable selection. We demonstrate improved inference on a number of canonical problems.

This note provides two corrections to the pseudo-code of the algorithm for the Bayesian estimation of the multinomial logit model using auxiliary variables as developed by Homes and Held (2006). After incorporating the two corrections, the algorithm works correctly for the multinomial as well as the binary logit model.