Backdoor criterion: Definition & Examples

The backdoor criterion is a graphical test that can justify identifying a treatment's total causal effect by adjusting for a specified set of control variables. In execution research, it helps distinguish a policy that raises costs from a policy chosen for orders that were already harder to execute. The test concerns causal paths, not how well a variable predicts the outcome.

Two conditions before averaging

Let $T$ be a treatment or policy, $Y$ its outcome, and $Z$ a proposed set of observed adjustment variables, excluding $T$ and $Y$ themselves. In an assumed causal directed acyclic graph, $Z$ satisfies Pearl's backdoor criterion for the effect of $T$ on $Y$ when:

No member of $Z$ is a descendant of $T$ : none can be reached by following directed arrows forward from $T$ .
$Z$ blocks every path from $T$ to $Y$ whose first edge points into $T$ .

A path may be traversed against its arrows. Thus

T\leftarrow Z\to Y

is a backdoor path: it associates policy with outcome through a shared cause, rather than transmitting an effect of policy. Conditioning on

Z

blocks that fork. The causal route

T\to Y

remains available. This is the adjustment test in Pearl, Glymour and Jewell, §3.3.

“Blocks” follows the rules of d-separation. A path is blocked if it contains either a conditioned noncollider, or a collider for which neither the collider nor any descendant is conditioned on. A collider is a node where the two arrows on that path meet head-to-head. Conditioning on it, or on its descendant, removes that particular blockage; another node may still block the path. See VanderWeele and Shpitser's Appendix and the worked collider-bias example.

A variable's role is path-specific. Check the whole proposed set, since adding a control can open a route that was previously closed. Even a variable measured before treatment can be a collider. More controls do not automatically mean less confounding.

A forty-point association, a ten-point effect

Consider a fictional population of parent orders. Let $T=1$ mean a specified aggressive execution policy and $T=0$ a specified patient policy. Let $Y=1$ mean execution cost exceeds a fixed, predeclared threshold. Baseline urgency $Z$ is low (0) or high (1), with half the population in each group.

Assume the entire relevant causal graph is $Z\to T$ , $Z\to Y$ , $T\to Y$ , with mutually independent structural disturbances. There are no other shared causes or interference between orders in this toy model. The probabilities are stipulated, not estimated from trades:

Baseline urgency	Population weight	Probability of $T=1$	Probability of $Y=1$ under $T=0$	Probability of $Y=1$ under $T=1$
$Z=0$	50%	20%	10%	20%
$Z=1$	50%	80%	60%	70%

High-urgency orders are more likely both to receive aggressive execution and to exceed the cost threshold. Bayes' theorem gives 80% high-urgency orders among

T=1

, versus 20% among

T=0

. The observed cost-flag probabilities are therefore

\begin{aligned} \mathbb{P}(Y=1\mid T=1)&=0.2(0.2)+0.8(0.7)=0.60,\\ \mathbb{P}(Y=1\mid T=0)&=0.8(0.1)+0.2(0.6)=0.20. \end{aligned}

Their difference is 40 percentage points. It compares populations with different urgency mixes.

Assumed causal graph: urgency Z causes policy T and cost flag Y, while T causes Y. Conditioning on Z blocks the fork. Observed T=1 has 20% low and 80% high urgency; T=0 has 80% low and 20% high urgency. Both interventions use 50% each. Observed outcome probabilities are 60% and 20%; adjusted probabilities are 45% and 35%.

Original stipulated execution model. Blue-filled Z is the adjustment variable. Standardisation uses one target population mix for both policies, rather than each observed policy group's mix.

Here $Z$ satisfies the criterion. The backdoor adjustment formula uses the population weights $\mathbb{P}(Z=z)$ :

\mathbb{P}(Y=1\mid\operatorname{do}(T=t)) =\sum_z\mathbb{P}(Y=1\mid T=t,Z=z)\mathbb{P}(Z=z).

The operator $\operatorname{do}(T=t)$ means setting the policy, replacing its usual assignment mechanism. It differs from observing which policy was chosen. Consequently,

\begin{aligned} \mathbb{P}(Y=1\mid\operatorname{do}(T=1))&=0.5(0.2)+0.5(0.7)=0.45,\\ \mathbb{P}(Y=1\mid\operatorname{do}(T=0))&=0.5(0.1)+0.5(0.6)=0.35. \end{aligned}

The average treatment effect on this binary outcome is 10 percentage points,

0.45-0.35=0.10

. That matches the difference within either urgency group. The graph justifies the causal interpretation; the arithmetic alone does not.

What the graph cannot supply

The criterion is a sufficient condition for adjustment, not a necessary condition for every valid adjustment or every form of causal identification. Failing it does not by itself prove that the causal effect is unknowable. VanderWeele and Shpitser, §2 make this distinction explicitly.

For the nonparametric population standardisation above, several substantive requirements remain:

A credible causal graph. Omitting an unmeasured common cause can make the apparent blocking set fail in the real system. A regression fit does not certify arrow directions or absent arrows.
Positivity. Every policy being compared must have positive probability within each adjustment stratum that receives target-population weight. Both 20% and 80% propensities satisfy this in the toy model. Positive probabilities still do not guarantee enough observations for a precise estimate.
Consistency and a defined intervention. An order receiving policy $t$ must have the outcome corresponding to that same intervention. “Aggressive” must specify the policy being compared; grouping materially different algorithms under one label can obscure the target effect. Interference from other orders also needs an explicit treatment in a real study.

These support and intervention-definition issues are distinct from drawing a valid blocking set; Greenland and Cole explain why positivity is method-specific and why treatment definitions and measurement deserve direct scrutiny.

Finally, controlling for a mediator on

T\to M\to Y

removes part of the total-effect mechanism and violates the criterion's descendant restriction. Predictive usefulness cannot override that causal role. For an execution application, Market Impact distinguishes observed response from causal impact and examines urgency-driven comparisons. Its public preview is available; the full lesson requires membership.

Practice: a safe extra control?

Suppose the only arrows are $A\to T$ , $A\to C$ , $B\to C$ , $B\to Y$ , $C\to R$ , and $T\to Y$ . All variables are observed; $C$ and $R$ are not descendants of $T$ . Does adding either $C$ alone or $R$ alone improve on the empty adjustment set for the total effect of $T$ on $Y$ ?

Answer. Neither is a valid backdoor set here. The only backdoor path is

T\leftarrow A\to C\leftarrow B\to Y

. With no adjustment,

C

blocks it as an unconditioned collider, so the empty set satisfies the criterion. Conditioning on

C

, or its descendant

R

, opens the path. The set

\{A,C\}

would block it again at the noncollider

A

. Thus “never condition on a collider” is also too crude: validity depends on the entire set and graph.

Sources

Judea Pearl, Madelyn Glymour and Nicholas P. Jewell, Causal Inference in Statistics: A Primer, 2016, author-hosted corrected pages (August 2021), §§2.4 and 3.3: path blocking, backdoor criterion and adjustment formula.
Tyler J. VanderWeele and Ilya Shpitser, “A New Criterion for Confounder Selection”, Biometrics 67(4), 2011, pp. 1406–1413, §2, Example 1 and Appendix: sufficient versus necessary conditions, pretreatment collider controls and causal DAG semantics.
Sander Greenland and Stephen R. Cole, “Negatives about Positivity and Consistency as Conditions for Causal Inference”, American Journal of Epidemiology 195(9), 2026, pp. 2416–2422, introduction and sections on sample size and potential-outcome models: support, treatment definitions and the limits of assumption labels.

Two conditions before averaging

A forty-point association, a ten-point effect

What the graph cannot supply

Practice: a safe extra control?

Sources

Related terms

Bayes' theorem

Collider bias