Pearl Causality and the Value of Control

We welcome this opportunity to acknowledge the significance of Judea Pearl’s contributions to uncertain reasoning and in particular to his work on causality. In the decision analysis community causality had long been “taboo” even though it provides a natural framework to communicate with decision makers and experts [Shachter and Heckerman 1986]. Ironically, while many of the concepts and methods of causal reasoning are foundational to decision analysis, scholars went to great lengths to avoid causal terminology in their work. Judea Pearl’s work is helping to break this barrier, allowing the exploration of some fundamental principles. We were inspired by his work to understand exactly what assumptions are being made in his causal models, and we would like to think that our subsequent insights have contributed to his and others’ work as well. In this paper, we revisit our previous work on how a decision analytic perspective helps to clarify some of Pearl’s notions, such as those of the do operator and atomic intervention. In addition, we show how influence diagrams [Howard and Matheson 1984] provide a general graphical representation for cause. Decision analysis can be viewed simply as determining what interventions we want to make in the world to improve the prospects for us and those we care about, an inherently causal concept. As we shall discuss, causal models are naturally represented within the framework of decision analysis, although the causal aspects of issues about counterfactuals and causal mechanisms that arise in computing the value of clairvoyance [Howard 1990], were first presented by Heckerman and Shachter [1994, 1995]. We show how this perspective helps clarify decision-analytic measures of sensitivity, such as the value of control and the value of revelation [Matheson 1990; Matheson and Matheson 2005].