Causal relationships and replicability (Q1825575)

This paper deals with the formalization of causality on the base of statistical and econometric methodology. Because of its predictivistic nature, the subjectivistic Bayesian approach points out to be ideally suited for formalizing causal relationships. In section 2, a conditional density in the Bayesian formula will be conceived as `causal relationship', the conditioning variables as `causes'. In section 3, the weak exogeneity of these variables, in the sense of \textit{R. F. Engle}, \textit{D. F. Hendry} and \textit{J.-F. Richard} [Econometrica 51, 277-304 (1983; Zbl 0528.62093)] is discussed. The attractiveness from a Bayesian perspective is that predictive densities can be calculated separately. The process of replication with respect to causality is the subject of section 4. Replication describes the process by which the predictive performance of a law is evaluated in a series of data confrontations. \textit{H. Feigl}'s definition of causality [Notes on causality. (1953)] is discussed in a statistical framework. In section 5, the Bayesian approach of section 2 is modified to handle multiple data. Special specifications of the a priori distribution comprises the essential ingredients of a model occurrence framework as described by the author [see J. Econ. Perspectives 2, 121-170 (1988)]. The steps in sections 2 to 5 lead up to a formal definition of a `causal relationship', presented in section 6. This definition characterizes a causal relationship for use in a specific data environment and identifies one with a Bayesian predictive density. The conditions to hold in the definition all impose testable restrictions on the predictive density and the likelihood. The approach builds on conditions already discussed in econometric literature, and is based on weak exogeneity, predictability, lawfullness and replication. Thus, access to a concrete operational definition of causality is available.