Primary Submission Category: Applications in Health and Biology
Sparse Group LASSO for Causal Network Discovery in High-Dimensional Multivariate Time Series: An Application to Swine Disease Surveillance
Authors: Zhengyuan Zhu, Alan Moore, Lynna Chu,
Presenting Author: Zhengyuan Zhu*
Learning causal relationships in high-dimensional multivariate time series is essential for understanding complex dynamical systems. In many applications, variables naturally form overlapping groups reflecting shared structure, such as spatial regions or related processes. Standard causal discovery methods often ignore this structure, which can reduce stability and statistical power when the number of possible interactions is large.
We propose a causal discovery framework based on sparse group LASSO that incorporates structured sparsity in a vector autoregressive model with lagged causal effects. The method encourages similar sparsity patterns across related groups while maintaining overall sparsity of causal links. A two-stage procedure uses sparse group LASSO to screen candidate links and ordinary least squares to estimate causal strengths.
We illustrate the method using swine disease surveillance data consisting of weekly case counts for multiple viruses across multiple U.S. states. Some viruses may precede or trigger outbreaks of others, and outbreaks in one state may lead to outbreaks in other states. Identifying such cross-virus and cross-location causal relationships can help improve outbreak monitoring. The learned causal network provides information that can support earlier detection of emerging outbreaks and reduce detection delay.