Vine copula based structural equation models

Abstract

Gaussian linear structural equation models (SEMs) are often used as a statistical model associated with a directed acyclic graph (DAG) also known as a Bayesian network. However, such a model might not be able to represent the non-Gaussian dependence present in some data sets resulting in nonlinear, non-additive and non Gaussian conditional distributions. Therefore the use of the class of D-vine copula based regression models for the specification of the conditional distribution of a node given its parents is proposed. This class extends the class of standard linear regression models considerably. The approach also allows to create an importance order of the parents of each node and gives the potential to remove edges from the starting DAG not supported by the data. Further uncertainty of conditional estimates can be assessed and fast generative simulation using the D-vine copula based SEM is available. The improvement over a Gaussian linear SEM is shown using random specifications of the D-vine based SEM as well as its ability to correctly remove edges not present in the data generation using simulation. An engineering application showcases the usefulness of the proposals.