How can M-bias make a covariate measured before treatment a harmful control?

## The shape M-bias is named for the picture. Put the treatment `A` on the lower left and the outcome `Y` on the lower right. Above them sit two **unmeasured** variables, `U1` on the left and `U2` on the right. The arrows run: - `U1` to `A` - `U1` to `Z` - `U2` to `Z` - `U2` to `Y` where `Z` is a measured variable sitting between and below the two unmeasured ones. Tracing the arrows gives an M. Crucially, `Z` is a **common effect** of `U1` and `U2` - a collider - and it is measured *before* treatment. ## Why it fools every empirical check `Z` is associated with treatment, because `U1` drives both. `Z` is associated with the outcome, because `U2` drives both. That pattern is the textbook empirical signature of a confounder, and the popular working definition - 'a variable correlated with both treatment and outcome' - flags it for adjustment. Timing does not save you either: the usual protective rule 'only control for things measured before treatment' passes `Z` with no complaint. But nothing is wrong before you touch it. `U1` and `U2` are independent, so treatment and outcome are not linked through this structure at all. The moment you condition on `Z`, the explaining-away effect makes `U1` and `U2` dependent - and since `U1` reaches treatment and `U2` reaches the outcome, that manufactured dependence shows up as bias in the treatment effect. You created the confounding you thought you were removing. ## Why this is a genuine differentiator Two things make M-bias worth knowing beyond trivia value. First, it breaks a rule most practitioners rely on. 'Adjust for everything pre-treatment' is a defensible heuristic precisely because pre-treatment variables cannot be mediators or treatment-affected colliders. M-bias is the counterexample: a pre-treatment variable that is still a collider, of two things you never measured. Second, **no diagnostic on the data can detect it**. The correlation structure of an M-bias `Z` is indistinguishable from that of a real confounder. Balance checks, fit statistics, and coefficient stability all look identical. The only way to tell the two apart is subject-matter reasoning about which variables plausibly cause which - the same knowledge that separates a collider from a confounder anywhere else. ## How much does it matter in practice? This is genuinely contested, and a good answer says so rather than picking a side. The argument for taking it seriously is that it is a real bias with no data-driven defence. The argument for not letting it paralyse you is that the M structure requires a fairly specific configuration - two independent unmeasured causes, each strong on both of its arrows - and simulation work generally finds the induced bias modest compared with the confounding bias you would incur by *omitting* a variable that turns out to be a genuine confounder. Many applied methodologists therefore still adjust for plausible pre-treatment covariates while acknowledging M-bias as a real possibility. The defensible practical position has three parts: - **Stop treating 'pre-treatment' as proof of safety.** It rules out mediators and treatment-affected colliders; it does not rule out collider structure among unmeasured causes. - **Reason about each covariate's role**, using domain knowledge about what causes what, rather than including variables because they correlate with both sides. - **Report robustness across specifications.** If an estimate moves materially when a suspicious covariate enters or leaves, that instability is itself the finding, and a sensitivity analysis over the strength of plausible unmeasured causes is more honest than a single blessed specification. ## In an interview Being able to draw the M, say why `Z` looks like a confounder, and state that conditioning on it opens a link between two unmeasured causes is the whole answer. Adding that the magnitude is debated - and that the real takeaway is the failure of the pre-treatment heuristic rather than a ban on adjustment - is what separates a memorised definition from understanding.