Controlling for colliders

Hello dears,

I have a problem of understanding. Why shouldn't I control for colliders in DAG's, both statically and in design?

The whole thing is not clear to me. Can someone explain it in simple terms?

Thanks for your help,



Less is more. Stay pure. Stay poor.
Imagine you have an exposure of interest and outcome, and the exposure does not directly affect the outcome. But there is a third variable that is a common effect of the exposure and outcome (inverted fork). Controlling for it opens a statistical path between the exposure and outcome.

There is something called the Berkson paradox that may help you understand how a spurious association can happen in general. I have a party and invite attractive and funny people. In the Berkson paradox these variables are independent, but if you know one you then likely know the value of the other variable. So if you are good looking you are likely not that funny. This isn't exactly what happens in all collider bias. In it you are usually dealing with selection bias, so allowing certain folks in a study.
Last edited: