CMStatistics 2015: Start Registration
View Submission - CMStatistics
B0550
Title: Estimating discrete latent models for two-way data arrays: A composite likelihood approach Authors:  Prabhani Kuruppumullage Don - Dana-Faber Cancer Institute (United States) [presenting]
Francesco Bartolucci - University of Perugia (Italy)
Francesca Chiaromonte - The Pennsylvania State University (United States)
Bruce Lindsay - The Pennsylvania State University (United States)
Abstract: Composite likelihood is a likelihood modification useful in instances where maximum likelihood estimation (MLE) is computationally infeasible. We present two discrete latent variable models for two-way data arrays, in which MLE is intractable due to the complex structures of the models. The first model aims to sort a two-way array of observations in blocks, each corresponding to a fixed row cluster and a fixed columns cluster (block mixture model). The second model aims to simultaneously produce clusters along one of the data dimensions and contiguous groups along the other (clustering by segmentation). In both cases, we construct composite likelihoods as a computationally tractable alternative to the full likelihood. We also discuss how to evaluate the number of components for each model, and demonstrate the performance of our methods via simulations. Finally, we illustrate the use of our approach through applications to genomic data.