A0504
Title: Modeling residual heterogeneity within the mixture of experts framework
Authors: Francesca Martella - La Sapienza University of Rome (Italy) [presenting]
Dalila Failli - University degli Studi di Perugia (Italy)
Maria Francesca Marino - University of Florence (Italy)
Abstract: The mixture of latent trait analyzers (MLTA) is a model-based clustering framework tailored for multivariate categorical data, combining features from latent class and latent trait models. It allows for both the clustering of units and the modeling of residual within-cluster variability through continuous latent variables. A novel extension of the MLTA model that incorporates the effects of covariates (concomitant variables) is proposed in a flexible and comprehensive way. Specifically, covariates are allowed to influence cluster formation, the distribution of the response variables, both, or neither, mirroring the structure of standard mixtures of expert models. This generalization enhances the interpretability of clustering results and captures more complex data structures, especially in contexts where observed characteristics are expected to affect both group membership and response behavior. The model is estimated via an EM-type algorithm using variational approximations. A motivating example from the biomedical domain illustrates how the proposed model can simultaneously uncover latent patient profiles and account for the effects of clinical covariates on both clustering and outcomes. The flexibility of this approach makes it suitable for a wide range of applications in social sciences, health, and beyond.
