← Data Curiosities

Data Curiosities · 07

M-Bias

"Control for everything you measured before the treatment" sounds safe. Sometimes it's the opposite: adjusting for the wrong pre-treatment variable can create a bias that wasn't there to begin with.

Try it

Does test prep raise your college GPA? Suppose it truly does nothing. Drag from Ignore SAT to Adjust for SAT — the move most analysts would make without thinking.

What's going on

Ignore SAT and there's no link between prep and GPA — which is the truth here. Adjust for SAT (by comparing students with similar scores) and a downward link appears: among students with the same SAT, the ones who did more prep have lower GPAs. It looks like prep hurts. It doesn't.

SAT is a collider. Two hidden things feed into it: family wealth (which also buys test prep) and innate ability (which also lifts GPA). Prep and GPA have no path between them — until you hold SAT fixed. Among students at the same SAT, a high prep count implies they got there partly on wealth, which means less of the ability that drives GPA. Conditioning on the collider opens a back-door path and links two unrelated things. The DAG looks like an "M", which is where the name comes from.

Pooled prep-GPA trend is flat; within each SAT band it slopes down.
The pooled trend is flat (the truth); within each SAT band a spurious downward link appears — manufactured by adjusting.

This is the mirror of confounding. There, adjusting for a common cause removes a fake link; here, adjusting for a common effect creates one. The same keystroke — "control for SAT" — is the cure in one case and the disease in the other.

In the real world

M-bias is the cautionary case against the habit of throwing every available pre-treatment covariate into a regression. Sander Greenland formalised it as "collider-stratification bias" in 2003; later work (Ding & Miratrix) showed that in practice its size is usually modest unless the collider is strongly driven by both hidden causes — but the direction of the lesson stands: which variables you adjust for has to come from a causal model, not from what happens to be in the spreadsheet.

How not to get fooled

"Adjust for everything" is not a strategy. A variable is only safe to control for if you know its place in the causal structure; adjusting for a collider (or anything downstream of the treatment) can do harm. When you can, sketch the diagram first and let it tell you what to condition on.

Sources