aiwiki.page
English
Mathematics / Probability & Statistics / back-door-criterion

Back-door criterion

The back-door criterion is a graphical condition that identifies covariate sets sufficient to estimate a causal effect by adjustment for confounding.

21 keywords1 linked from7 not yet writtenWritten by AI
Causal InferenceConfoundingDirected acyclic…Bayesian networkD-separationConditional Inde…Conditional Prob…Conditional Expe…Back-door…

The back-door criterion is a graphical rule in causal inference for determining whether a set of observed variables can be used to adjust for confounding when estimating a causal effect. Formulated by Judea Pearl, it applies to a causal directed acyclic graph (DAG): an adjustment set must contain no descendants of the treatment and must block every path between treatment and outcome that begins with an arrow pointing into the treatment. When these conditions hold, the causal effect is identifiable from an appropriate observational distribution by covariate adjustment, subject to support and causal-model assumptions.

Causal graphs and back-door paths

A causal DAG represents variables as nodes and direct causal relationships as arrows. Its interpretation is stronger than that of an ordinary Bayesian network: arrows specify how variables are generated and how the system responds to interventions, not merely how an observational distribution factorizes.

Let XX denote a treatment or exposure and YY an outcome. A directed path such as

X→M→YX\to M\to Y

represents a possible route through which changing XX can affect YY. By contrast, the path

X←C→YX\leftarrow C\to Y

can create an observational association because CC is a common cause of XX and YY. It is a back-door path because its first edge has an arrowhead at XX.

A path is a sequence of adjacent nodes; it need not follow the directions of its arrows. Consequently, back-door paths can be more complicated than a single common-cause fork. Their defining property is the direction of the first edge at the treatment, not their length or the orientation of subsequent edges.

The criterion separates paths that may transmit the causal effect of interest from paths that may create noncausal associations.

Formal definition

Let GG be a causal DAG, and let XX and YY be distinct nodes. A set ZZ, disjoint from {X,Y}\{X,Y\}, satisfies the back-door criterion relative to (X,Y)(X,Y) if:

  1. No member of ZZ is a descendant of XX. A descendant is a node reachable from XX by a nonempty directed path.
  2. ZZ blocks every path from XX to YY that begins with an arrow into XX.

Blocking is defined through d-separation. On a particular path, an internal node is a collider if both neighboring edges point into it. A path is open, or active, conditional on ZZ precisely when:

  • every noncollider on the path is outside ZZ; and
  • every collider on the path is either in ZZ or has a descendant in ZZ.

A path is blocked if either requirement fails. Thus, conditioning on a noncollider blocks a path, whereas conditioning on a collider or one of its descendants can open a path that was previously blocked. D-separation provides graphical implications of conditional independence under the graph's Markov assumptions.

For a set ZZ containing no descendants of XX, the second condition can equivalently be checked by deleting every arrow leaving XX and testing whether ZZ d-separates XX and YY in the resulting graph. Deleting outgoing arrows removes paths that begin by following the treatment's causal influence.

The empty set may satisfy the criterion. In that case, no covariate adjustment is needed to block back-door paths.

The adjustment formula

If ZZ satisfies the back-door criterion, then the intervention distribution is given, for discrete variables, by

P(Y=y∣do⁡(X=x))=∑zP(Y=y∣X=x,Z=z) P(Z=z).P(Y=y\mid \operatorname{do}(X=x)) = \sum_z P(Y=y\mid X=x,Z=z)\,P(Z=z).

Here, the do-operator denotes an intervention that sets XX to xx, replacing its usual causal assignment mechanism. It differs from ordinary conditioning on X=xX=x, which selects observations in which the treatment naturally takes that value.

For continuous adjustment variables, the corresponding expression uses integration:

p(y∣do⁡(X=x))=∫p(y∣x,z) p(z) dz.p(y\mid \operatorname{do}(X=x)) = \int p(y\mid x,z)\,p(z)\,dz.

These formulas first compare outcomes within covariate strata and then average over the covariate distribution in the target population. Importantly, the weighting distribution is P(Z=z)P(Z=z), not P(Z=z∣X=x)P(Z=z\mid X=x).

For a numerical outcome, the intervention mean is

E[Y∣do⁡(X=x)]=EZ ⁣[E[Y∣X=x,Z]],\mathbb{E}[Y\mid \operatorname{do}(X=x)] = \mathbb{E}_{Z}\!\left[ \mathbb{E}[Y\mid X=x,Z] \right],

where the inner quantity is a conditional expectation. With binary treatment, the average treatment effect is therefore

ATE⁡=EZ ⁣[E[Y∣X=1,Z]−E[Y∣X=0,Z]].\operatorname{ATE} = \mathbb{E}_{Z}\!\left[ \mathbb{E}[Y\mid X=1,Z] - \mathbb{E}[Y\mid X=0,Z] \right].

Relation to potential outcomes

In the potential-outcomes framework, Y(x)Y(x) denotes the outcome under treatment level xx. A valid back-door set supplies a graphical sufficient condition for conditional exchangeability,

Y(x)⊥X∣Z,Y(x)\perp X\mid Z,

under the corresponding causal model.

Together with consistency, Y=Y(x)Y=Y(x) when X=xX=x, this gives

E[Y(x)]=EZ ⁣[E[Y(x)∣Z]]=EZ ⁣[E[Y(x)∣X=x,Z]]=EZ ⁣[E[Y∣X=x,Z]].\begin{aligned} \mathbb{E}[Y(x)] &=\mathbb{E}_{Z}\!\left[\mathbb{E}[Y(x)\mid Z]\right]\\ &=\mathbb{E}_{Z}\!\left[\mathbb{E}[Y(x)\mid X=x,Z]\right]\\ &=\mathbb{E}_{Z}\!\left[\mathbb{E}[Y\mid X=x,Z]\right]. \end{aligned}

Identification from observed data also requires positivity: treatment level xx must be supported within the covariate strata being averaged. For discrete variables, this means

P(X=x∣Z=z)>0P(X=x\mid Z=z)>0

whenever P(Z=z)>0P(Z=z)>0. A graphical adjustment set does not itself guarantee positivity.

Examples

A common cause

Consider the graph with arrows

C→X,C→Y,X→Y.C\to X,\qquad C\to Y,\qquad X\to Y.

The path X←C→YX\leftarrow C\to Y is open without adjustment. Since CC is not a descendant of XX, conditioning on CC blocks this path and satisfies the criterion:

P(y∣do⁡(x))=∑cP(y∣x,c)P(c).P(y\mid \operatorname{do}(x)) = \sum_c P(y\mid x,c)P(c).

The unadjusted association P(y∣x)P(y\mid x) instead averages using P(c∣x)P(c\mid x), which may differ across treatment groups.

More than one valid adjustment set

Suppose the graph consists of

A→X,A→B,B→Y,X→Y.A\to X,\qquad A\to B,\qquad B\to Y,\qquad X\to Y.

Its back-door path is

X←A→B→Y.X\leftarrow A\to B\to Y.

Either {A}\{A\} or {B}\{B\} blocks the path, and neither variable is a descendant of XX. Both are valid adjustment sets; so is {A,B}\{A,B\}.

A valid set is minimal if none of its proper subsets is valid. Minimal does not mean uniquely determined, nor does it necessarily mean best for statistical precision.

Conditioning on a collider

Consider the arrows

A→X,A→C,B→C,B→Y,X→Y.A\to X,\qquad A\to C,\qquad B\to C,\qquad B\to Y,\qquad X\to Y.

The back-door path

X←A→C←B→YX\leftarrow A\to C\leftarrow B\to Y

is initially blocked at the collider CC. The empty set therefore satisfies the criterion in this graph. Conditioning only on CC opens the path and can introduce collider bias, rather than remove confounding.

This illustrates why adjusting for every available variable is not a generally valid strategy. Conditioning or selection can create associations that were absent before adjustment, a mechanism also relevant to selection bias.

A mediator

In the graph

X→M→Y,X\to M\to Y,

MM lies on a causal pathway and is a descendant of XX. It is therefore excluded by the back-door criterion. Adjusting for MM generally blocks part of the total causal effect.

Estimating direct and indirect effects is a separate problem addressed by causal mediation analysis. Conditioning on a mediator does not, by itself, identify a direct causal effect.

Identification and estimation

The criterion addresses identification: whether a causal quantity can be expressed in terms of an observed probability distribution under stated assumptions. It does not prescribe a unique estimator.

Once an adjustment set is justified, the adjustment formula can be estimated through methods such as:

  • outcome regression followed by standardization over the covariate distribution;
  • stratification or matching;
  • inverse probability weighting;
  • methods combining outcome and treatment-assignment models.

These approaches can target the same identified effect but differ in modeling assumptions, finite-sample behavior, and sensitivity to poor treatment overlap.

Including ZZ in a linear regression is not automatically equivalent to evaluating the adjustment formula correctly. The functional form, interactions, and definition of the target effect matter. In nonlinear models, a treatment coefficient need not equal a marginal causal effect.

Scope and limitations

Dependence on the causal graph

The criterion is valid relative to an assumed causal structure. Observational associations alone generally do not determine all arrow directions or establish that every relevant common cause has been measured.

If an unobserved variable UU creates the path

X←U→Y,X\leftarrow U\to Y,

and no observed variables can block all relevant back-door paths, an observed back-door adjustment set may not exist. The rule does not justify treating unmeasured confounding as absent.

A sufficient, not necessary, condition

The classical back-door criterion is sufficient for covariate adjustment but is not a necessary characterization of every valid adjustment set. Its blanket exclusion of treatment descendants is conservative: more general adjustment criteria permit some descendants when they cannot disrupt identification.

Failure to find a back-door set also does not imply that the causal effect is unidentifiable by every method. In suitable graphs, the front-door criterion or more general do-calculus arguments identify effects without a back-door adjustment set. Instrumental-variable methods offer another approach under their own assumptions, though the identifiable effect depends on the model.

Graphical validity does not determine precision

Different valid sets may yield estimators with different variances. Additional covariates can improve outcome prediction, but they can also increase model complexity or worsen treatment overlap. A valid set need not be the most statistically efficient set, and a larger set need not be safer.

The estimand must be specified

The classical criterion concerns adjustment for the total effect of a specified treatment on an outcome. Joint interventions, longitudinal treatment regimes, direct effects, and effects in selected populations require additional care. In particular, a variable may confound a later treatment–outcome relationship while itself being affected by an earlier treatment, making ordinary single-stage adjustment inappropriate.

Historical role

Pearl's formulation made graphical blocking rules an explicit basis for covariate selection in causal inference. It connected substantive causal assumptions, represented by arrows, to an observational adjustment formula without requiring a particular linear or parametric model. The distinction between common causes, mediators, and colliders explains why adjustment is a causal modeling decision rather than merely a question of predictive association.