Common Cause In Causal Inference

Abstract

One of the major impediments to reliably inferring both qualitative and quantitative causal relations from non-experimental data is the possibility that there may be unobserved common causes of observed variables. The possibility of unobserved common causes presents no problems for causal inference when simple randomized experiments are possible, beyond the ordinary statistical problems of inferring a population distribution from a sample. There are many types of statistical/causal models that postulate unobserved common causes. These include principal components models, factor analytic models, item response models, some structural equation models, Rasch models, and finite mixture models. They differ in the families of distributions they represent, and the kinds of constraints they entail on marginal distributions over observed variables—the constraints that help to make possible inferences to the existence and causal roles of unobserved variables. This chapter describes structural equation models and introduces the concept of a manipulation, in order to clearly distinguish several different kinds of causal inference from statistical inference. It discusses problems for reliable inference of qualitative causal relations from observational data and background assumptions, and some approaches that are taken to solving these problems.