Multivariate subset autoregressive modelling with zero constraints for detecting 'overall causality' (Q1069258)

It is shown that the canonical reduced autoregressive form is a straightforward and effective form of multivariate autoregressive models to test Granger-causality and instantaneous causality among time series variables. Necessary and sufficient conditions for the absence of Granger-causality and instantaneous causality (therefore for ''overall'' causality) are presented for the reduced autoregressive form. The recursive form of the model is also considered and the equivalent reduced form is obtained via appropriate transformations. It is shown that the conditions for the absence of Granger-causality are only sufficient for the bivariate autoregressive model if the recursive form is used. For trivariate and higher order autoregressive systems it is shown that testing for Granger-causality and overall causality from the recursive autoregressive form is complicated whereas the reduced form is a more straightforward and effective means of testing. An algorithm is proposed to select the most appropriate subset of autoregressive models with zero constraints and to detect Granger- causality and/or instantaneous causal relations. To establish whether there is instantaneous causality, it is proposed to select the best subset in the residual regression system by using ordinary least-squares estimation in conjuction with the model selection criteria. The Canadian money, income, and bank rate series are examined using the proposed procedures and results are compared with those found by analyzing only the money and income series. It is shown that the procedures are not computationally demanding to perform the investigations required to establish the pattern of causality.