Abstract
Usually, an analysis on a causal graph (DAG) is used to identify open non-causal paths between exposure and disease, and to close these to remove bias. But there are important situations not handled by this method.
In this presentation, manly based on Greenland and Pearl 2011, we will define collapsibility and show how this concept can be used to greatly extend DAG analyses.
As examples, we will give a definition for selection bias and look at adjustment to remove selection bias. We will discuss (the novel concept of) bias amplification. And lastly, use the collapsibility method on missing graphs to decide when an ordinary (complete case) regression will be unbiased.