Intervention and causality in a dynamic Bayesian network
Catriona M. Queen, Casper Johannes Albers · 2008
The use of intervention for time series modelling is a well established technique for on-line forecasting and decision-making in the context of Bayesian dynamic linear models. Intervention has also been recently used in (non-dynamic) Bayesian networks to investigate causal relationships between variables, and in dynamic Bayesian networks to investigate lagged causal relationships between time series. The Multiregression Dynamic Model (MDM) is a Bayesian dynamic model and an example of a dynamic Bayesian network. The focus of this paper is the use of intervention in the MDM. It will be demonstrated that not only is intervention in the MDM a powerful tool for forecasting, but intervention can also aid in the identification of contemporaneous causal relationships between time series, thus going beyond the identification of lagged causal relationships previously addressed in dynamic Bayesian networks.