Identification of discrete concentration graph models with one hidden binary variable (Q2435228)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Identification of discrete concentration graph models with one hidden binary variable |
scientific article |
Statements
Identification of discrete concentration graph models with one hidden binary variable (English)
0 references
4 February 2014
0 references
This paper studies local identifiability of undirected graph models with one unobserved binary node. Let \(G^K=(K,E)\) be a graph with node set \(K=\{0,1,\dots,n\}\) and edge set \(E\). \(G^S\) is the induced subgraph of \(G\) on a subset \(S\), and \(\bar{G}^S\) is the complementary graph for it. Each node \(v\) is associated a discrete random variable \(A_v\) with finitely many levels. Let \(\beta\) be the vector of parameters of a graph model \(G^K\) over the discrete variables \(A_0,\dots,A_n\) with \(A_0\) latent variable. \((0,u)\) is not an edge for any \(u\in T_1\subseteq K\backslash 0\); \((0,u)\) is an edge for any \(u\in S=K\backslash\{0\cup T_1\}\). It is shown that the rank of the parametrization map is full everywhere in the parameter space if and only if \(\bar{G}^S\) contains at least on \(m\)-clique for \(m\geq3\), and the generalized identifying sequence \(S_S\) exists with all \(S_S\subseteq S\) for each clique \(C_0\subseteq G^S\) containing at least one node.
0 references
conditional independence
0 references
contingency tables
0 references
hidden variables
0 references
identifiability
0 references
finite mixtures
0 references
latent class
0 references
latent structure
0 references
log linear models
0 references
0 references