One of my favorite papers in recent years included this diagram. It shows the impact of controlling for three different types of variables: confounders, colliders, and mediators. With confounders, control is good. With the others, you ruin your result by controlling.

If you have variables with measurement error, you can run into another problematic variable: the proxy. Proxy variables can make all of these distortions much worse and much more difficult to deal with.

The paper makes this simple observation: statistical control requires causal justification. That's actually the title. They gave several DAG-based examples. Consider this one: is edutainment a confounder or a mediator? Should you control for it, or would that bias your estimate?

This is always something you have to consider, but it is, frankly, exceptional to think causally about statistical control. Many papers do things like controlling for irrelevant downstream variables (proxies), or they unintentionally control for mediators. That's the norm!

My article goes over a lot more issues with the use and misuse of controls. For example, it notes that it can take a lot of data and effort to get propensity-scoring to match experiments:

It notes that a given finding may not mean what it says it does:

And it shows that even the gold standard of causal inference—the RCT—needs to be done right, or you'll end up in a situation where your effect estimates are inflated and your conclusions are wrong.