|
Abstract
|
The notions of direct and indirect effects are often used informally to describe phenomena where mediating variables play an important role.
Examples can be found in many disciplines: sociology (e.g. indirect effect of gender on employment mediated by education), biometry (e.g.
direct and indirect effects of vaccinations), ecology (direct and indirect effects of tourism on wildlife), public health (direct and indirect effects of cost-sharing in the use of preventive services), also the well known placebo effect can be regarded as an indirect effect of treatment on outcome. However, much of the statistical theory in this context sticks to the framework of linear structural equation models, where direct effects can be equated to path coefficients. Pearl (2001) and Robins (2003) have proposed a non-parametric definition of (in)direct effects which will be the basis for this talk. In order to get a more formal grasp of direct/indirect effects, I will introduce some basic concepts of causal reasoning, including counterfactuals, Pearl's "do"-operator and (causal) graphical models. There are in fact different approaches to defining direct effects depending on how we choose to 'hold constant' the mediating variable (Didelez et al., 2006). These will be discussed and illustrated with examples. Finally the question of when these (in)direct effects can actually be identified from observational data will be addressed.
References:
Didelez, V, Dawid, A.P., Geneletti, S. (2006). Direct and indirect effects of sequential treatments, Proc. of the 22nd Conference on Uncertainty in Artificial Intelligence, pp.138-146.
Pearl, J. (2001). Direct and indirect effects. Proc. of the 17th Conference on Uncertainty in Artificial Intelligence, pp. 411-420.
Robins, J. (2003). Semantics of causal DAG models and the identification of direct and indirect effects. In: Highly Structured Stochastic Systems, eds. Green, P., Hjort, N., and Richardson, S. OUP, pp. 70-81.
|