Exchangeability in causal effects

Sander Greenland and James M Robins. Identifiability, Exchangeability, and Epidemiological Confounding. IJE 1986.

This is a heavily cited paper regarding the exchangeability condition in identifying causal effects. Sarvet et al. (2014) cite this when defining partial exchangeability. I came here because I am unclear about some consistencies in the Sarvet paper and hope that this paper will clear it up.

Greenland and Robins state that there is historically little consistency in the definition of a confounder. They provide two broad categories:

1. Comparability-based: inherent differences in risk between exposed and unexposed, absent the exposure;
2. Collapsibility-based: differences between stratified/conditional measures of association and the corresponding “collapsed” measure.

The paper goes on to provide a theory of epidemiological confounding based on “a model for individual effects” using a simple deterministic model of effects.