Probability and causality are two indispensable tools for addressing situations of social risk. Causal relations are the foundation for building risk assessment models and identifying risk prevention, mitigation and compensation measures. Probability enables us to quantify risk assessments and to calibrate intervention measures. It therefore seems not only natural, but also necessary to make the role of causality and probability explicit in the definition of decision problems in situations of social risk. Such is the aim of this thesis.
By reviewing the terminology of risk and the logic of public interventions in various fields of social risk, we gain a better understanding of the notion and of the issues that one faces when trying to model it. We further elaborate our analysis in the case of nuclear safety, examining in detail how methods and policies have been developed in this field and how they have evolved through time. This leads to a number of observations concerning risk and safety assessments.
Generalising the concept of intervention in a Bayesian network allows us to develop a variety of causal Bayesian networks adapted to our needs. In this framework, we propose a definition of risk which seems to be relevant for a broad range of issues. We then offer simple applications of our model to specific aspects of the Fukushima accident and other nuclear safety problems. In addition to specific lessons, the analysis leads to the conclusion that a systematic approach for identifying uncertainties is needed in this area.
When applied to decision theory, our tool evolves into a dynamic decision model in which acts cause consequences and are causally interconnected. The model provides a causal interpretation of Savage’s conceptual framework, solves some of its paradoxes and clarifies certain aspects. It leads us to considering uncertainty with regard to a problem’s causal structure as the source of ambiguity in decision-making, an interpretation which corresponds to a common understanding of the precautionary principle.