Analysis of interaction for identifying causal mechanisms

Great attention has been paid, in this book, to questions about the effect of actions (EOA), for example, ‘What is the effect of causal action X?' With the exception of the preceding three chapters, little has been done to address questions about the ‘mechanisms of effects' (MOE), that is, questions concerning the ‘how' of effects, how and why they occur, how can we explain their occurrence and what are the mechanisms through which these effects operate. The following examples should help see the distinction: (EAO) ‘What is the effect of physical inactivity on risk of infarction, in an individual who carries the A variant of gene X?' (MOE 1) ‘Is the effect of the A variant of gene X on infarction entirely mediated by an increase in blood pressure?' (MOE 2) ‘Has frequent physical activity the power of nullifying the effect of the A variant of gene X on infarction?' Questions of type MOE1 are tackled in Chapters 11 and 12 in this volume. Questions of type MOE2 represent the main theme of the present chapter. What they ask is whether, in some individuals or circumstances, two variables of interest interfere with each other's effect on a specific outcome. Such interference we call mechanistic interaction. Looking for variables that interact mechanistically may be useful because it helps us to understand which causal factors (e.g. genes) are part of a common (e.g. physical, biological) mechanism affecting the outcome of interest. In this chapter, we discuss ways of assessing mechanistic interaction on the basis of observational data. Our illustrations are largely epidemiological; the relevance of the ideas is much wider. This chapter owes much to the work of Tyler VanderWeele, who has pioneered the theoretical foundations of mechanistic interaction and the application of the concepts in observational epidemiology.