Why can adjusting for more variables create bias?
A collider is a common effect of two variables. Conditioning on it can make its causes associated within the adjusted comparison, opening a noncausal path that was previously closed.
Try it in the experiment
- Start with adjustment for the risk score only.
- Also account for the follow-up score.
- Watch the estimate move even though the data-generating world stays fixed.
In this example, treatment and outcome both cause the later follow-up score. The score does not cause the outcome. Its availability in the data does not make it a valid adjustment variable.