Markov equivalence of marginalized local independence graphs (Q2176633)

The authors intend to give a graphical framework for handling the presence of unobserved processes and to extend the work on graphical representations of local independence models. It is stipulated that the proposed graphical framework generalizes the one developed in a series of articles on Graphical models published in the years 2000--2008 by \textit{V. Didelez} [Graphical models for event history analysis based on local independence. Berlin: Logos-Verlag; Dortmund: Univ. Dortmund, Fachbereich Statistik (2000; Zbl 0970.62084)]. One starts by introducing abstract independence models and elements of the Graph theory. Notions as Local independence, Local independence graph, Marginalization, separability, directed mixed graphs (DMG), \(\mu\)-separation, directed graphs (DG), directed acyclic graph (DAG), Latent projection are defined and properties are shown. An algorithm to marginalize a given local independence graph is presented. In the section devoted to Markov equivalence of DMGs, as main result, it is proved that each Markov equivalence class of DMGs has a greatest element. To illustrate and understand the importance of local independence for applications an example from the theory of Gateway drugs is discussed. Additional results and proofs are to be found in a supplementary material, under: \url{doi:10.1214/19-AOS1821SUPP}.