A0895
Title: Birth-death dynamic graphical models
Authors: Elena Bortolato - Universitat Pompeu Fabra (Spain) [presenting]
David Rossell - Universitat Pompeu Fabra (Spain)
Stephen Hansen - UCL (United Kingdom)
Abstract: In many real-world settings, from economics and biology to the social sciences, datasets often involve units that enter or disappear over time, leading to changing dimensionality and evolving dependence structures. Traditional graphical models struggle in these dynamic, potentially high-dimensional environments, especially when missingness arises not at random but from systematic birth death processes that alter the very composition of the observed system. We propose a Bayesian graphical model framework that explicitly incorporates the appearance and disappearance of units as a core part of the data-generating process, treating these structural changes as primary drivers of evolving dependencies. The model builds on recent advances in Bayesian factor and graphical modeling, employs a low-dimensional decomposition of the time-varying precision matrix, and incorporates latent states to capture distinct regimes, such as periods of stability or sudden structural shocks, through hidden Markov chains. This allows the framework to detect when and how dependencies among units fundamentally change. The result is a flexible tool for analyzing high-dimensional data with systematic structural missingness, providing robust inference on evolving dependence patterns when both the nodes and edges of the underlying graph change over time.
