Mixture-Based Latent Trait Modeling for Biclustering Overdispersed Counts

We propose a mixture of latent trait models for biclustering units and variables in the presence of multivariate, overdispersed count data. Units are grouped into homogeneous clusters called components through a finite mixture model. Simultaneously, within each component, variables are grouped into segments using a flexible specification of the linear predictor. Covariates are incorporated at the latent level of the model to account for their effect on component formation, while residual dependence among variables is captured by a multidimensional latent trait. A simulation study based on a varying number of units and variables is conducted to assess both clustering performance and the ability to correctly estimate model parameters.