A0303
Title: Autoregressive dynamic network modelling with serial and cross-sectional dependence
Authors: Jonathan Flossdorf - TU Dortmund University (Germany) [presenting]
Daniel Dzikowski - TU Dortmund University (Germany)
Carsten Jentsch - TU Dortmund University (Germany)
Abstract: A flexible dynamic network modelling approach is proposed based on a class of generalised binary vector autoregressive (gbVAR) models. Originally designed for multivariate binary data with serial and cross-sectional dependence, it is also well suited for modelling dynamic networks characterized by a time series of binary adjacency matrices. This is due to the fact that gbVAR models are parsimoniously parametrized, well interpretable, and also allow for the modelling of negative dependence. As they are autoregressive in nature and satisfy the classical Yule-Walker equations, the recently developed toolbox for penalised estimation of (continuous) vector autoregressive (VAR) models enables parameter estimation of reasonably large dynamic networks observed at moderately many time points under sparsity constraints. In this context, we consider a lasso-penalised estimation procedure that particularly allows the incorporation of additional information in the form of linear restrictions for further dimension reduction. We evaluate the proposed approach and illustrate our theoretical findings with extensive simulations. The applicability is further demonstrated by some real-world examples.
