Abstract
Selection bias is a common problem when we estimate the effect of an exposure on an outcome. We seldom get random samples of our target populations, some subjects refuse participation resulting in selection bias in our estimates. We show that there are two types of selection bias, one based on interaction and one based on conditioning on a collider. These two have different structures and quite different properties. We show causal graphs (DAGs) for the two structures and contrast the properties. We then focus on the second type, selection bias caused by conditioning on a collider, and show how different patterns of bias arise from different selection forces.
No mathematics were abused to reach these results, only DAGs and scatterplots were (ab)-used.