Speaker
Description
We propose a dynamic extension of the Mixture of Latent Trait Analyzers (MLTA) for a bipartite network. Specifically, we move along a Hidden Markov Model framework to account for the dynamic nature of the data and enable a dynamic clustering of sending nodes over time. A multidimensional continuous latent variable (trait) is assumed to account for residual, unobserved, time-constant, latent factors that may affect the way sending nodes relate to the receiving ones. Estimation of model parameters is conducted within a maximum likelihood framework by extending the Baum-Welch algorithm to account for the presence of the continuous latent trait in the model. However, this algorithm requires the solution of multidimensional integrals over the latent trait domain that are not available in closed form. To overcome the issue, a variational approach is employed. This consists of approximating the intractable integrals with lower bounds that are easy to manage and solve. The effectiveness of the proposal is demonstrated via an extensive simulation study, based on a varying number of sending and receiving nodes, as well as a varying number of time occasions. The proposal is also employed for the analysis of data from the Survey of Health, Ageing and Retirement in Europe (SHARE) with a focus on Italian residents. The aim is that of dynamically cluster residents according to their mental health status, while also accounting for possible unobserved latent factors related to their psychological well-being.
Keywords/Topics
Bipartite network
Hidden Markov Model
Dynamic clustering
