Causalité persistante entre séries non-stationnaire: application à l'étude comparée des politiques monétaires des pays du G5

This paper investigates the notion of persistent causality in multivariate first difference stationary autoregressive systems. The causality concept is based on impulse responses as suggested by Sims [1980]. Non-persistent causality is characterized as non-correlation between the canonical innovation of the causal variable and the trend part of the caused variable, as defined from a standard Beveridge-Nelson decomposition. Non-persistent causality simply corresponds to the nullity of a well identified long run multiplier. We show that persistent causality properties may lead to a causal interpretation of a cointegration relation, because non-causality excludes any prediction improvement at an infinite horizon. Testing non-persistent causality requires an auxiliary parametrization of the dynamics, which is provided by a suitable cointegrated VAR model, according to a two-steps procedure. This methodology is applied to analyse the efficiency of the monetary policies in the G5 countries.