Clinical trials and causation: Bayesian perspectives.

In addition to the safety, it is essential to establish the causal efficacy of extant and new treatments, and well-designed clinical trials are thought by most to be the 'gold standard' to accomplish this. Contrary to most statisticians' and regulators' views, however, I will argue that the concept of causation involved in clinical trials is not all that clear. I discuss the manipulability approach to causation, interpreted counterfactually, which seems to fit causation as it is found in such sciences as physiology, but it has unclear relations to a concept of causation proposed by a number of epidemiologists. I characterize 'epidemiological causation' as probabilistic and formulated at a population level, and dependent on certain general criteria for causation as well as study-design considerations. I then attempt to clarify the connections between these concepts of causation and Cartwright's views on complexity and causality, a 'Bayesian' framework proposed by Rubin and further elaborated by Holland, and Glymour and his colleagues' recent directed graphical causal modelling approach.