A0778
Title: Combining Bayesian latent traits and topic models for identifying risky behavioral profiles in Italian population
Authors: Angela Andreella - Ca Foscari University of Venice (Italy)
Lorenzo Schiavon - Ca Foscari University of Venice (Italy)
Stefano Campostrini - Ca Foscari University of Venice (Italy)
Mattia Stival - Ca Foscari University of Venice (Italy) [presenting]
Abstract: Understanding the relationships between risk factors is vital in health economics and policy making, facilitating targeted campaigns for vulnerable population subgroups. Behavioral and risk factors surveillance systems (BRFSS) provide extensive data, including demographics, socio-economics, and behavioral habits. Among these, employment information is collected as textual data. Despite its relevance, linking employment information with other risk factors is challenging, and few attempts have been observed in the literature. A model is introduced that integrates latent Dirichlet allocation (LDA) for textual analysis with latent trait models (LTM) for categorical data. LDA identifies topics that describe unobservable job macro-groups. These macro-groups are likely characterized by heterogeneous propensities to risky behavior, aspects captured by the LTM. To better describe this variability, covariates are included in the LTM and let it depend on the topics identified by LDA. A modular inferential procedure is proposed that uses the state of-the-art methods from both approaches, combining their results using importance sampling. The inferential procedure leads to a valid posterior inference strategy, fully leveraging the methodological advancements of both models without the need to develop new cumbersome algorithms from scratch.
