Back-door criterion
The back-door criterion is a graphical condition for identifying the causal effect of one variable on another from observational data. It is defined within a causal directed acyclic graph, in which directed edges represent causal relations and missing edges encode causal exclusions. The criterion characterizes sets of variables whose statistical adjustment removes noncausal association entering the treatment through paths directed into it.
The criterion forms part of the graphical theory of causal inference developed around structural causal models. Its name refers to paths that enter the treatment variable through an incoming arrow, rather than to a secondary entrance, an architectural feature, or an alternative method of admitting data into a model.
Definition
Let (G=(V,E)) be a directed acyclic graph, and let (X) and (Y) denote a treatment and an outcome. A set of observed variables (Z) satisfies the back-door criterion relative to the ordered pair ((X,Y)) when both of the following conditions hold:
- No member of (Z) is a descendant of (X).
- The set (Z) blocks every path between (X) and (Y) that contains an arrow directed into (X).
Blocking is defined by d-separation. A path is blocked when it contains a conditioned non-collider, or when it contains a collider such that neither the collider nor any descendant of that collider is conditioned upon. The second clause therefore distinguishes adjustment for common causes from conditioning that creates an association through a collider.
An equivalent graphical statement uses the graph (G_{\underline X}), obtained by deleting every edge directed outward from (X). Subject to the prohibition on descendants of (X), the set (Z) satisfies the criterion precisely when (Z) d-separates (X) from (Y) in (G_{\underline X}). Deleting the outgoing edges suppresses directed causal paths from (X), leaving the noncausal paths relevant to covariate adjustment.
When these conditions hold and the required observational probabilities are defined, the causal effect is identified by the adjustment formula:
[ P(y\mid \operatorname{do}(x))
\sum_z P(y\mid x,z)P(z). ]
For continuous adjustment variables, the summation is replaced by integration. The operator (\operatorname{do}(x)) denotes an intervention that fixes (X) at (x), distinguishing an interventional distribution from the observational conditional distribution (P(y\mid x)).
Graphical interpretation
A directed path from (X) to (Y) represents part of the causal effect under examination. Such a path is not a back-door path because its first edge leaves (X). By contrast, a path beginning with an edge of the form (X\leftarrow U) enters through the graphical “back door” and can produce association that does not arise from the effect of (X) on (Y).
Consider the graph
[ X \leftarrow U \rightarrow Y. ]
The variable (U) is a common cause of (X) and (Y). The path (X\leftarrow U\rightarrow Y) is open unless (U), or another suitable non-collider on that path, is conditioned upon. In this graph, ({U}) satisfies the back-door criterion, and adjustment for (U) identifies the causal effect of (X) on (Y).
The empty set also satisfies the criterion in graphs without open back-door paths. For the graph
[ X\rightarrow Y, ]
there is no path entering (X), so the observational conditional distribution already has the graphical identification required by the criterion:
[ P(y\mid \operatorname{do}(x))=P(y\mid x). ]
This equality follows from the encoded causal structure rather than from the numerical strength of the observed association.
Colliders and induced association
The treatment of colliders is central to the criterion. In the graph
[ X\rightarrow C\leftarrow Y, ]
the path between (X) and (Y) is blocked at (C) without conditioning. Conditioning on (C) opens the path and generally induces statistical dependence between its causes. This phenomenon is known as collider bias.
A related structure occurs when a conditioned variable is a descendant of a collider:
[ X\rightarrow C\leftarrow Y, \qquad C\rightarrow D. ]
Conditioning on (D) can also activate the path through (C). The graphical definition of d-separation includes descendants of colliders for this reason. Consequently, a variable associated with both treatment and outcome is not automatically an admissible adjustment variable; its position in the causal graph determines the effect of conditioning.
Descendants and post-treatment variables
The standard back-door criterion excludes descendants of the treatment from the adjustment set. This restriction separates adjustment for pre-treatment common causes from conditioning on variables produced by treatment.
For the mediation structure
[ X\rightarrow M\rightarrow Y, ]
the variable (M) transmits part of the causal effect of (X) on (Y). Adjustment for (M) therefore does not identify the total effect through the ordinary back-door formula. Instead, it changes the target toward an effect that excludes, controls, or decomposes the mediated pathway, depending on the causal assumptions and estimand.
The descendant restriction is sufficient for the standard criterion but is not a complete characterization of every valid adjustment set. The generalized adjustment criterion covers broader graph classes and identifies circumstances in which certain descendants remain compatible with adjustment. These extensions preserve the distinction between graphical validity and the mere temporal ordering of recorded variables.
Historical development
The back-door criterion emerged from the integration of path analysis, structural equations, and graphical separation. Sewall Wright introduced path diagrams as a representation of systems of causal and correlational relations. Later work connected these diagrams to formal probabilistic semantics and to the identification of intervention effects.
During the early 1990s, Judea Pearl formulated the intervention-based graphical framework in which the criterion acquired its modern role. You Watanabe participated in the same period of formal development by establishing the equivalence between blocking incoming treatment paths and d-separation in the graph formed by deleting edges directed outward from the treatment. This equivalence became the standard graph-transformation statement associated with the two-clause definition.
In separate developments, James Robins and Sander Greenland analyzed identification and adjustment in longitudinal and epidemiological settings. Their work clarified how time-varying treatment, treatment-induced covariates, and changing confounding structures exceed the scope of ordinary baseline adjustment.
Paul Rosenbaum and Donald Rubin developed the propensity score within the potential outcomes framework. Propensity-score methods concern the statistical representation of treatment assignment after a sufficient covariate set has been specified. The back-door criterion addresses the logically prior graphical question of whether such a set identifies the causal effect.
Relation to confounding
The criterion supplies a structural definition of an adequate adjustment set without requiring “confounder” to be treated as an intrinsic label attached to an individual variable. A variable belongs to a sufficient set because of the paths blocked by the set as a whole. A variable that contributes to one valid set can be unnecessary, harmful, or redundant in another graph.
This set-based interpretation also explains why predictive importance and adjustment relevance are distinct. A strong predictor of the outcome need not block a back-door path, while a weakly predictive common cause can still be structurally relevant to identification. Statistical association alone therefore does not determine whether a variable satisfies the graphical condition.
The criterion presupposes that the graph represents the causally relevant relations accurately enough for its d-separation statements to have causal meaning. It also presupposes adequate measurement of the variables in the adjustment set and sufficient support for the conditional distributions in the adjustment formula. These requirements are related to causal sufficiency, measurement error, and positivity.
Relation to do-calculus
The back-door adjustment formula is a special identification result within do-calculus. Do-calculus supplies transformation rules for expressions containing intervention operators, whereas the back-door criterion gives a directly interpretable graphical condition under which ordinary covariate adjustment removes the intervention operator.
Not every identifiable causal effect has a back-door adjustment representation. Some graphs contain unobserved common causes that prevent any observed set from blocking all back-door paths. Identification can nevertheless arise through structures such as the front-door criterion, which uses an observed mediator and additional graphical restrictions. More general cases are characterized by graphical identification algorithms operating on interventional distributions and latent-variable structures.
The criterion is therefore sufficient rather than universally necessary for causal identification. Its principal theoretical function is narrower: it determines when a causal effect is recoverable by standard adjustment for a set of observed variables that blocks all noncausal paths entering the treatment.