M-Bias

M-bias is a tricky problem that happens when researchers control for a variable that seems helpful but actually makes their analysis worse. The variable appears in the middle of an M-shaped diagram, with two invisible causes pointing down to it. The safest approach is controlling for variables that are known causes of either the factor you're studying or the result you're measuring, not just anything correlated with both. Notice this: the most dangerous variables in your analysis are the ones that seem most innocent.

M-bias reveals a counterintuitive truth: sometimes controlling for more variables makes your answers less accurate. In a classic M-structure, your exposure and outcome each have their own unmeasured causes, and those two causes both affect some third variable—the collider. This creates an M-shape when drawn as a diagram. The collider sits at the bottom center, receiving arrows from two unmeasured variables that each also point to either exposure or outcome.

The practical implications extend far beyond academia. In policy evaluation, controlling for 'neighborhood conditions' when assessing a housing program might create bias if those conditions are caused by separate pre-treatment factors affecting both program placement and outcomes. The key insight: variables downstream from your causal system are dangerous adjustment targets.

Ding and Miratrix's sensitivity analysis shows that in linear models, M-bias magnitude tends to be relatively small compared to actual confounding bias. This suggests that when uncertain about a variable's role, you're usually safer adjusting for potential confounders than avoiding adjustment entirely. However, the bias can still be substantial in nonlinear settings or when the collider is strongly associated with its causes. The safest strategy follows VanderWeele's 'disjunctive cause criterion': adjust for variables that are known causes of the exposure, outcome, or both. The criterion itself is to adjust for variables known to cause the exposure, the outcome, or both. Handling of known pure instruments is a separate modeling choice, not what makes the criterion M-bias-safe. The R ggdag documentation illustrates similar bias structures, including M-bias patterns.

M-bias represents a specific DAG structure where a pretreatment collider creates bias when conditioned upon, formalized in the causal inference literature by Pearl. The structure consists of two unmeasured variables: one causes the exposure and the collider, and the other causes the outcome and the collider. When this collider is adjusted for, it opens a backdoor path A ← U1 → C ← U2 → Y, where C is the collider and U1, U2 are unmeasured.

• Ding and Miratrix (2015): In linear structural equation models, the magnitude of M-bias tends to be relatively small compared to confounding bias, suggesting that adjusting for a potential M-structure collider is generally less harmful than failing to adjust for a true confounder [1].
• VanderWeele (2019): The 'disjunctive cause criterion'—controlling for variables that cause the exposure, outcome, or both—avoids M-bias while still controlling confounding, as it excludes colliders that are not causes of either exposure or outcome [2].
• Elwert and Winship (2014): Collider bias represents a broader category of selection bias that includes M-bias as a special case; conditioning on colliders (or their descendants) always induces association between their causes [3].

The M-Structure can be distinguished from classical confounding because the collider (M) has no causal effect on either exposure or outcome—it is merely a common effect. This differs from 'butterfly bias' or 'bow-tie bias,' where the variable is both a collider and a genuine confounder.

M-bias assumes you can correctly identify the causal structure, which is rarely possible in practice. Real-world variables often play multiple roles simultaneously—something might be a slight confounder AND a slight collider. The theoretical results about M-bias being 'small' rely on linear models and specific distributional assumptions that may not hold in your data. When effects are nonlinear or interactions exist, collider bias can be substantially larger. Additionally, these results assume perfect measurement; if variables are measured with error (which they almost always are), controlling for even genuine confounders may not fully block backdoor paths, and the distinction between confounders and colliders becomes blurrier. Most critically, you rarely know whether unmeasured variables exist or what the true causal structure is—DAG-based reasoning is only as good as your domain knowledge.