The Front-Door Criterion

The usual fix for confounding — a hidden third variable driving both a cause and an outcome — is to measure the confounder and adjust for it. The front-door criterion handles the harder case where the confounder exists but you cannot measure it at all, by routing around it through a mediator: a variable sitting on the causal path between cause and outcome that fully carries the effect you care about.

Imagine trying to measure whether a trading desk's use of a research tool ($X$) improves its P&L ($Y$), but desk skill is an unmeasured confounder pushing both. Suppose the tool works entirely by changing how much a trader checks a signal dashboard ($M$), and nothing else links the tool to P&L. If skill affects P&L directly but does not affect dashboard-checking except through the tool, then you can estimate the tool's effect in two clean steps: estimate how the tool changes dashboard-checking (this comparison isn't confounded by skill, because skill doesn't touch $M$ except via $X$), then estimate how dashboard-checking changes P&L while adjusting for the tool itself (this step blocks the confounded path back through $X$). Combining those two unconfounded pieces recovers the tool's total effect on P&L without ever measuring skill.

The condition doing all the work is that $M$ must fully mediate the effect — no other route from $X$ to $Y$ bypassing $M$ — and $M$ itself must not share an unmeasured confounder with $Y$. Both are strong, unverifiable assumptions about the causal structure, which is why the front-door criterion is mostly a proof that identification is possible in principle, and rarely a tool applied casually to messy trading data without a carefully argued mediator.