Causality and Probability: A View from Bayesian Networks

Abstract

For the past few decades, the statistical analysis of causation, referred to as causal Bayes-net, has gained wide applicability to many scientific areas including economics, biology, psychology and others. Bayes-nets model the causal relationships with a graph-theoretic structure and evaluates the effect of a causal intervention with the aid of the probabilistic pattern called the Markov Condition. This paper firstly reviews this statistical approach to see how it works in tasks such as the prediction of causal consequence or the evaluation of counterfactual propositions. The success of Bayes-nets in these tasks comes not without philosophical implications. Two related interpretations of causality – the probabilistic reductionism and the interventionist theory – are examined in the second half of the paper. Although both of them suffer from the problem of unfaithfulness, i.e. the probabilistic independence not implied by the Markov Condition, it will be argued that resorting to the underlying mechanism, or the ‘invariance’, helps the interventionists to avoid the problem and to justify some assumptions in the use of Bayes-net.