Front-door criterion: Definition & Examples

The front-door criterion is a graphical condition under which an observed mediator identifies a treatment's total causal effect, even when an unobserved common cause confounds treatment and outcome. For a quant studying an execution policy, it offers a possible route from recorded mechanisms to a causal cost comparison. The route works only if the mechanism and its surrounding causal paths satisfy demanding assumptions.

A mechanism must pass three tests

Let $T$ be treatment, $M$ an observed mediator and $Y$ the outcome. In an assumed causal directed acyclic graph, the classical criterion requires:

Complete mediation: $M$ intercepts every directed path from $T$ to $Y$ . There is no direct $T\to Y$ arrow or other causal route bypassing $M$ .
No unblocked treatment–mediator backdoor path: the effect of $T$ on $M$ is unconfounded without adjustment.
Treatment blocks every mediator–outcome backdoor path: conditioning on $T$ removes the noncausal paths from $M$ to $Y$ .

A backdoor path starts with an arrow pointing into the variable whose effect is being studied. Blocking uses the path rules described under the backdoor criterion. These are the conditions of Pearl's original Theorem 2, with

M

allowed to denote a set of mediating variables.

Consider the particular graph

U\to T,\qquad U\to Y,\qquad T\to M\to Y,

where $U$ is unobserved and there are no other arrows or shared disturbances. The path $T\leftarrow U\to Y$ prevents ordinary observed-confounder adjustment. Yet $M$ can satisfy the front-door criterion: the potential backdoor route $T\leftarrow U\to Y\leftarrow M$ is blocked at the collider $Y$ , and $M\leftarrow T\leftarrow U\to Y$ is blocked by conditioning on $T$ .

Calling

M

a mediator does not establish any of these facts. Mediation analysis asks about causal mechanisms; front-door identification adds particular restrictions that permit recovering the total effect despite treatment–outcome confounding.

The formula averages twice

For discrete variables, write $p$ for the observational probability distribution. Under the criterion and the required support conditions,

p(y\mid\operatorname{do}(T=t)) =\sum_m p(m\mid t) \left[\sum_{t'}p(y\mid m,t')p(t')\right].

The

\operatorname{do}

operator means setting treatment, replacing its usual assignment mechanism. The inner index

t'

ranges over the observed treatment values; it is not fixed to the intervention value

t

. This is the front-door adjustment formula, §3.4.

In the displayed graph, the inner sum identifies the outcome distribution under setting

M=m

, by adjusting for

T

. It uses the marginal weights

p(t')

, not

p(t'\mid m)

. The outer sum then weights those corrected outcomes by the mediator distribution induced by setting

T=t

, which equals

p(m\mid t)

here. For continuous variables, use the corresponding density integrals when the conditional quantities exist.

A fictional execution-policy calculation

Suppose

T

selects one of two fully specified execution policies,

M=1

records a high-intensity execution state, and

Y=1

records cost above a fixed threshold. Assume the graph above, including the strong claim that every policy effect on this cost flag operates through the recorded state. The following are stipulated population probabilities, not trade estimates:

p(T=1)=0.5,\qquad p(M=1\mid T=0)=0.2,\qquad p(M=1\mid T=1)=0.8.

Recorded state	$p(Y=1\mid M,T=0)$	$p(Y=1\mid M,T=1)$
$M=0$	0.10	0.30
$M=1$	0.60	0.80

First correct the mediator–outcome comparison using the same 50–50 treatment mix in both rows:

\begin{aligned} q_0&=0.5(0.10)+0.5(0.30)=0.20,\\ q_1&=0.5(0.60)+0.5(0.80)=0.70. \end{aligned}

Next let the intervention determine the mediator weights:

\begin{aligned} p(Y=1\mid\operatorname{do}(T=0))&=0.8q_0+0.2q_1=0.30,\\ p(Y=1\mid\operatorname{do}(T=1))&=0.2q_0+0.8q_1=0.60. \end{aligned}

The average treatment effect is therefore

0.60-0.30=0.30

: policy 1 raises the cost-flag probability by 30 percentage points in this fictional model.

Assumed graph U causes policy T and cost flag Y, with T causing mediator M and M causing Y. Marginal treatment weights of 50% each produce corrected mediator outcome probabilities of 20% and 70%. Policy 0 puts 20% weight on M=1 and gives a 30% outcome probability; policy 1 puts 80% weight on M=1 and gives 60%.

Original stipulated model. The first averaging operation fixes the treatment mix; the second changes the mediator mix with the policy. The graph is an assumption, not a result of these calculations.

The raw policy comparison instead gives

\begin{aligned} p(Y=1\mid T=0)&=0.8(0.10)+0.2(0.60)=0.20,\\ p(Y=1\mid T=1)&=0.2(0.30)+0.8(0.80)=0.70. \end{aligned}

Its 50-point difference combines policy effects with differences in the unobserved common cause. Those raw policy means happen to equal

q_0

and

q_1

numerically; they represent different quantities.

Simply holding

M

fixed fares no better: the table still shows a 20-point treatment association within either mediator state. Under the assumed graph there is no causal effect of

T

with

M

held fixed, because that fixes the only mechanism. The remaining association reflects

U

. Front-door adjustment reconstructs the total intervention distribution through nested standardisation; it does not take the treatment coefficient from a regression that controls for the mediator. This distinction matters when interpreting post-treatment adjustment.

Where an execution application can fail

The arithmetic cannot establish complete mediation. If the policy changes fees directly, or changes timing in a way not captured by the intensity state, a bypass remains. If hidden urgency affects both policy and intensity, condition 2 fails. If unrecorded liquidity affects both intensity and cost after conditioning on policy, condition 3 fails. Measuring a convenient intermediate score cannot remove these threats.

Support also matters: every conditional outcome probability used by the inner average must be observable. A sufficient discrete requirement is $p(T=t',M=m)>0$ for all treatment–mediator pairs needed by the formula. Here the four joint probabilities are 0.40, 0.10, 0.10 and 0.40, so none is empty. Sparse cells would still make estimation unstable.

The intervention must be well defined, and observed outcomes must agree with the corresponding intervention outcomes when the observed policy is received (consistency). The toy model excludes effects between orders; real executions can affect one another. Such spillovers need an explicit causal model rather than automatic reuse of this single-order formula.

For application background, Market Impact distinguishes mechanical book walking, observed price response and causal impact. It helps specify which cost mechanism a proposed mediator actually captures. A public preview is available; the full lesson requires membership.

The classical criterion is sufficient, not an exhaustive test for identification. Its failure means this formula lacks its stated justification; it does not rule out every other design or identification argument.

Sources

Judea Pearl, “Causal diagrams for empirical research”, Biometrika 82(4), 1995, pp. 669–688; author-hosted paper with discussion. §3.2, Figure 3, equation (9) and Theorem 2 give the classical graph, conditions and formula; §2.2 defines intervention semantics.
Judea Pearl, Madelyn Glymour and Nicholas P. Jewell, Causal Inference in Statistics: A Primer, 2016; author-hosted corrected pages, August 2021. §3.4, especially equation (3.15) and Theorem 3.4.1, explains nested adjustment and positive treatment–mediator support. The criterion above follows the original paper's explicit backdoor wording.

A mechanism must pass three tests

The formula averages twice

A fictional execution-policy calculation

Where an execution application can fail

Sources

Related terms

Backdoor criterion

Mediation analysis

Post-treatment bias