We introduce a new nonlinear model for classification, in which we model the joint distribution of response variable, y, and covariates, x, non-parametrically using Dirichlet process mixtures. We keep the relationship between y and x linear within each component of the mixture. The overall relationship becomes nonlinear if the mixture contains more than one component, with different regression coefficients. We use simulated data to compare the performance of this new approach to alternative methods such as multinomial logit (MNL) models, decision trees, and support vector machines. We also evaluate our approach on two classification problems: identifying the folding class of protein sequences and detecting Parkinson’s disease. Our model can sometimes improve predictive accuracy. Moreover, by grouping observations into sub-populations (i.e., mixture components), our model can sometimes provide insight into hidden structure in the data.

In analyzing longitudinal or clustered data with a mixed effects model (Laird and Ware, 1982), one may be concerned about violations of normality. Such violations can potentially impact subset selection for the fixed and random effects components of the model, inferences on the heterogeneity structure, and the accuracy of predictions. This article focuses on Bayesian methods for subset selection in nonparametric random effects models in which one is uncertain about the predictors to be included and the distribution of their random effects. We characterize the unknown distribution of the individual-specific regression coefficients using a weighted sum of Dirichlet process (DP)-distributed latent variables. By using carefully-chosen mixture priors for coefficients in the base distributions of the component DPs, we allow fixed and random effects to be effectively dropped out of the model. A stochastic search Gibbs sampler is developed for posterior computation, and the methods are illustrated using simulated data and real data from a multi-laboratory bioassay study.

The naive Bayes classifier (NB) has exhibited its “mysterious ” but outstanding classification ability in practice, in spite of its often unrealistic conditional inde-pendence assumption. This simple assumption implies the adoption of a diagonal structure for the underlying class-specific precision matrices. However, the NB leaves covariates interrelationships unrevealed. In this dissertation, we will ex-tend the NB from the perspectives of covariance modeling and classification. Due to the positive definiteness constraint and the rapidly-growing number of parameters with dimensions, covariance estimation in a multivariate normal population has been a classic but challenging statistical problem. Sparse shrinkage covariance/precision matrix estimation has been obeyed as an important principle in covariance/precision matrix modeling. However, many existing models can only shrink the covariance/precision matrix toward a predefined diagonal structure. We model a precision matrix via its Cholesky decomposition in terms of compositional regression coefficient matrix and error precisions. Our approach aims at estimating

Abstract We propose a double-robust procedure for modeling the correlation matrix of a longitudinal dataset. It is based on an alternative Cholesky decomposition of the form Σ = DLL ⊤ D where D is a diagonal matrix proportional to the square roots of the diagonal entries of Σ and L is a unit lower-triangular matrix determining solely the correlation matrix. The first robustness is with respect to model misspecification for the innovation variances in D, and the second is robustness to outliers in the data. The latter is handled using heavy-tailed multivariate t-distributions with unknown degrees of freedom. We develop a Fisher scoring algorithm for computing the maximum likelihood estimator of the parameters when the nonredundant and unconstrained entries of (L, D) are modeled parsimoniously using covariates. We compare our results with those based on the modified Cholesky decomposition of the form LD 2 L ⊤ using simulations and a real dataset.