Simulating interventions in graphical chain models for longitudinal data

Abstract

We propose a method to simulate the effects of an intervention to a complex system of interacting variables whose association structure is represented by a graphical chain model. Our approach takes into account both the direct and indirect effects of the intervention on the outcome of interest. We apply our approach to assess the effect of a policy to improve the mental health of teenage mothers using a model fitted to data from the 1970 British Birth Cohort Study. Since some of the conditional probabilities required are not directly available, we estimate them using a Gibbs sampler.