Complete-Association Factorizations of Discrete Copulas: Existence, Tree Characterization, and Log-Linear Representations

For multivariate categorical variables, the discrete-copula and log-linear approaches provide two different representations of dependence: the former uses probability distributions with discrete-uniform one-way margins, and the latter uses interaction parameters computed from full-table or marginal distributions. This article studies the exact division of the dependence represented by a positive discrete copula into factors that are themselves complete associations among disjoint groups of variables treated as composites. Although no nesting, tree, ordering, or binary structure is imposed in defining admissible factors, every rooted tree on the component variables generates a regular complete-association factorization: the map from the positive discrete-copula space to the product of the individual factor ranges is a smooth bijection with a smooth inverse, and the node factors are globally variation independent. Conversely, every regular complete-association factorization necessarily has this rooted-tree structure. The article also establishes several connections with log-linear analysis. Saturated full-table interactions and discrete copulas encode the same margin-free dependence information, while tree-induced marginal log-linear blocks provide complete, smooth, variation-independent coordinates for individual association factors. Such a factorwise representation cannot be obtained by directly partitioning saturated full-table interactions for a nontrivial tree. The tree factorization guarantees that complete-association factors can be manipulated independently, through recursive application of iterative proportional fitting. A regular parameterization permits direct implementation of such manipulations by substituting parameter blocks if and only if each block is a smooth, invertible transformation of the corresponding tree-induced marginal log-linear block. The article also shows that every multifurcating node admits an exact binary refinement without loss of dependence information, and that refinement, local rotation, and collapse provide smooth transformations between the factor values associated with different rooted trees. The article’s results support structured dependence models, sensitivity analysis, simulation, and counterfactual-based decompositions, including Shapley decompositions.

Authors

Publication Details

Journal
SocArXiv (OSF Preprints)
Published
2026-11-12
DOI
https://doi.org/10.31235/osf.io/ef7a5
Primary Topic
Bayesian Modeling and Causal Inference
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Complete-Association Factorizations of Discrete Copulas: Existence, Tree Characterization, and Log-Linear Representations

Pablo A. Mitnik, Junchao Tang
SocArXiv (OSF Preprints)
Bayesian Modeling and Causal Inference
preprint

Complete-Association Factorizations of Discrete Copulas: Existence, Tree Characterization, and Log-Linear Representations

Pablo A. Mitnik, Junchao Tang
preprint en

Abstract

For multivariate categorical variables, the discrete-copula and log-linear approaches provide two different representations of dependence: the former uses probability distributions with discrete-uniform one-way margins, and the latter uses interaction parameters computed from full-table or marginal distributions. This article studies the exact division of the dependence represented by a positive discrete copula into factors that are themselves complete associations among disjoint groups of variables treated as composites. Although no nesting, tree, ordering, or binary structure is imposed in defining admissible factors, every rooted tree on the component variables generates a regular complete-association factorization: the map from the positive discrete-copula space to the product of the individual factor ranges is a smooth bijection with a smooth inverse, and the node factors are globally variation independent. Conversely, every regular complete-association factorization necessarily has this rooted-tree structure. The article also establishes several connections with log-linear analysis. Saturated full-table interactions and discrete copulas encode the same margin-free dependence information, while tree-induced marginal log-linear blocks provide complete, smooth, variation-independent coordinates for individual association factors. Such a factorwise representation cannot be obtained by directly partitioning saturated full-table interactions for a nontrivial tree. The tree factorization guarantees that complete-association factors can be manipulated independently, through recursive application of iterative proportional fitting. A regular parameterization permits direct implementation of such manipulations by substituting parameter blocks if and only if each block is a smooth, invertible transformation of the corresponding tree-induced marginal log-linear block. The article also shows that every multifurcating node admits an exact binary refinement without loss of dependence information, and that refinement, local rotation, and collapse provide smooth transformations between the factor values associated with different rooted trees. The article’s results support structured dependence models, sensitivity analysis, simulation, and counterfactual-based decompositions, including Shapley decompositions.

SocArXiv (OSF Preprints)
Bayesian Modeling and Causal Inference
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon