The Backdoor Criterion and Candidate Adjustment Sets

Why a Covariate Set Needs a Causal Rationale

A multiple regression model can include many measured variables, but adding a variable does not automatically improve a causal comparison. A covariate is useful for adjustment when the causal structure gives a reason to include it. The backdoor criterion provides a graphical rule for checking whether a measured set blocks noncausal paths from an exposure to an outcome.

The rule is applied to a directed acyclic graph (DAG): a diagram whose arrows represent assumed causal directions and which contains no directed cycles. The earlier material on tracing paths through a causal diagram provides the path language used here. This tutorial focuses on identifying candidate adjustment sets, not on choosing a set from data alone.

Definition: Let \(X\) be an exposure, \(Y\) an outcome, and \(Z\) a set of measured covariates. A backdoor path from \(X\) to \(Y\) is any path between them that begins with an arrow pointing into \(X\). The set \(Z\) satisfies the backdoor criterion for \(X\) and \(Y\) if (1) no variable in \(Z\) is a descendant of \(X\), and (2) \(Z\) blocks every backdoor path from \(X\) to \(Y\).

A path is blocked by conditioning on \(Z\) when it contains a noncollider in \(Z\), or when it contains a collider and neither the collider nor any of its descendants is in \(Z\). A collider is a point on a path where the two arrows both point inward. Thus, conditioning can block some paths and open others. Checking the whole path, rather than just asking whether a variable is associated with \(X\) or \(Y\), is essential.

What the Criterion Assumes

A DAG is a set of causal assumptions, not a pattern that can usually be read directly from correlations. To use the backdoor criterion, the arrows and omitted arrows must be a reasonable representation of the relevant causal structure. In particular, if an important common cause of \(X\) and \(Y\) is missing from the diagram, a set that appears to block all drawn backdoor paths may still leave confounding in the real process.

The criterion is a graphical condition for selecting covariates. A causal interpretation of an adjusted comparison also relies on assumptions about the data and measurements. In potential-outcome notation, let \(Y(x)\) be the outcome that would be observed under exposure value \(x\). For the adjustment formula below, the key conditions include consistency (the observed outcome equals \(Y(x)\) for a unit receiving \(x\)), conditional exchangeability (within levels of \(Z\), exposure assignment carries no remaining information about \(Y(x)\)), and positivity (each exposure value being compared is possible at every relevant level of \(Z\)). The DAG supports the exchangeability claim only if its causal structure is adequate and relevant common causes are represented and controlled.

When these assumptions hold and \(Z\) satisfies the backdoor criterion, a population mean outcome under exposure level \(x\) can be obtained by standardizing the conditional outcome means over the covariate distribution:

E[Y(x)] = \sum_z E[Y \mid X=x,Z=z]P(Z=z).

For continuous covariates, the sum is replaced by an integral. The formula averages the outcome comparisons within covariate strata using the same distribution of \(Z\) for each exposure level. A contrast between two exposure levels can then be formed from the resulting means. The formula does not make the assumptions true; it shows what follows if they hold.

A Practical Path-Checking Method

For each candidate set, first mark the arrows touching \(X\), then trace every path from \(X\) to \(Y\) that starts with an arrow into \(X\). For each such path, determine whether conditioning on the proposed covariates blocks it. Finally, check that none of the covariates is a descendant of \(X\). A set that passes these checks is a candidate adjustment set under the stated graph.

State the exposure and outcome.
Keep the causal question fixed while checking candidate sets; changing the target can change which variables are appropriate.

List actual backdoor paths.
Trace paths that begin with an arrow into the exposure. A route that requires an arrow absent from the diagram is not a path in that diagram.

Check blocking path by path.
Apply the collider and noncollider rules to the proposed conditioned variables and their descendants.

Exclude descendants of the exposure.
Even if conditioning on a descendant seems to block a route, it violates the backdoor criterion.

State the assumptions.
A valid graphical set supports adjustment only under an adequate causal diagram and the relevant data conditions, including positivity and consistency.

Worked Example: One Common Cause and an Irrelevant Side Branch

Consider a hypothetical diagram with arrows \(S \rightarrow X\), \(S \rightarrow Y\), \(X \rightarrow Y\), and \(S \rightarrow A\). Here \(X\) is a treatment, \(Y\) is a later outcome, \(S\) is a pretreatment health-status measure, and \(A\) is a recorded administrative label. The proposed diagram has no arrow from \(A\) to \(Y\).

The only backdoor path is \(X \leftarrow S \rightarrow Y\). The route \(X \leftarrow S \rightarrow A\) ends at \(A\), not \(Y\), so it is not a path from \(X\) to \(Y\). A route through \(A\) to \(Y\) cannot be listed because the diagram contains no arrow \(A \rightarrow Y\). Conditioning on \(S\) blocks the actual backdoor path, and \(S\) is not a descendant of \(X\). Thus \(\{S\}\) satisfies the criterion. The empty set does not, because it leaves \(X \leftarrow S \rightarrow Y\) open.

For a numerical illustration, suppose the hypothetical data have \(P(S=0)=0.60\) and \(P(S=1)=0.40\). The conditional mean outcomes, in outcome units, are 12 and 18 for \(X=1\) in the two strata, and 10 and 16 for \(X=0\). Standardizing over \(S\) gives

\begin{aligned} \widehat{E[Y(1)]} &= (0.60)(12)+(0.40)(18)=7.2+7.2=14.4,\\ \widehat{E[Y(0)]} &= (0.60)(10)+(0.40)(16)=6.0+6.4=12.4,\\ \widehat{E[Y(1)]}-\widehat{E[Y(0)]} &= 14.4-12.4=2.0. \end{aligned}

If the assumptions for causal interpretation hold, this standardized contrast estimates a mean outcome difference of 2.0 outcome units. Positivity would require observing both exposure levels within each \(S\) stratum; the example assumes that condition. The calculation does not justify adding \(A\): the diagram gives no backdoor path that \(A\) needs to block.

Multiple Backdoor Paths Require a Set That Blocks Them All

When distinct common causes affect both exposure and outcome, adjusting for just one may leave another backdoor path open. A valid candidate set must block every such path, not merely the most obvious one. A set is minimal if no proper subset of it still satisfies the criterion. Minimality can be useful for describing the graph, but it does not by itself establish that a covariate set is available or well measured.

Worked Example: Two Common Causes

Suppose a hypothetical DAG has \(A \rightarrow X\), \(A \rightarrow Y\), \(S \rightarrow X\), \(S \rightarrow Y\), and \(X \rightarrow Y\). Imagine \(A\) represents age and \(S\) represents smoking status. The backdoor paths include \(X \leftarrow A \rightarrow Y\) and \(X \leftarrow S \rightarrow Y\). Conditioning on \(A\) alone leaves the path through \(S\) open; conditioning on \(S\) alone leaves the path through \(A\) open. The set \(\{A,S\}\) blocks both and contains no descendant of \(X\), so it satisfies the criterion in this graph.

For hypothetical data, suppose the four \((A,S)\) strata each have weight \(0.25\). The conditional mean outcomes for \(X=1\), listed in the order \((0,0),(0,1),(1,0),(1,1)\), are 8, 13, 11, and 16. For \(X=0\), they are 6, 10, 8, and 12. Standardizing both exposure groups over the same four-stratum distribution gives

\begin{aligned} \widehat{E[Y(1)]} &=0.25(8+13+11+16)=0.25(48)=12,\\ \widehat{E[Y(0)]} &=0.25(6+10+8+12)=0.25(36)=9,\\ \widehat{E[Y(1)]}-\widehat{E[Y(0)]} &=12-9=3. \end{aligned}

The adjusted contrast is 3 outcome units, provided the causal and data assumptions hold. In particular, positivity requires both \(X=1\) and \(X=0\) to be represented in each of the four strata. The arithmetic illustrates standardization; the graph explains why both \(A\) and \(S\) are needed.

More Than One Set May Satisfy the Criterion

A candidate set need not be the only valid set. Consider a diagram with arrows \(C \rightarrow X\), \(C \rightarrow Y\), \(C \rightarrow D\), \(D \rightarrow Y\), and \(X \rightarrow Y\). The backdoor paths are \(X \leftarrow C \rightarrow Y\) and \(X \leftarrow C \rightarrow D \rightarrow Y\). Conditioning on \(C\) blocks both, so \(\{C\}\) is a minimal set. Since \(D\) is not a descendant of \(X\), \(\{C,D\}\) also satisfies the criterion. But \(\{D\}\) alone does not block \(X \leftarrow C \rightarrow Y\).

Worked Example: Comparing a Minimal Set With a Larger Set

Use the diagram just described, and suppose the hypothetical covariate distribution is \(P(C=0)=0.50\) and \(P(C=1)=0.50\). Conditional mean outcomes for \(X=1\), averaged over \(D\) within each \(C\) stratum, are 9 for \(C=0\) and 15 for \(C=1\). The corresponding means for \(X=0\) are 7 and 11. Using the valid set \(\{C\}\), the standardized means are

\begin{aligned} \widehat{E[Y(1)]}&=0.50(9)+0.50(15)=4.5+7.5=12,\\ \widehat{E[Y(0)]}&=0.50(7)+0.50(11)=3.5+5.5=9,\\ \widehat{E[Y(1)]}-\widehat{E[Y(0)]}&=12-9=3. \end{aligned}

This example shows how a minimal set can block every backdoor path even when one path passes through another measured variable. The conditional means are averaged over \(D\) within each \(C\) stratum, so the calculation uses the \(C\)-adjusted comparison. The example does not establish that \(\{C,D\}\) is preferable: the criterion establishes that both sets meet the graphical rule, not that adding every permissible variable is beneficial in practice.

Common Mistakes and Why the Criterion Matters

Inventing a path. A sequence of variable names is not a path unless every required edge is in the DAG. For example, a route through \(A\) to \(Y\) cannot be claimed without an edge that connects \(A\) onward toward \(Y\).
Checking only one backdoor path. A set must block all paths that begin with an arrow into \(X\). Leaving even one open means the set fails the criterion.
Adjusting for a descendant of \(X\). The criterion explicitly rules out descendants, even if including one appears to close a particular route.
Treating a valid set as proof of causation. A graph and candidate set do not verify that the graph is correct, that all relevant common causes were measured, or that positivity holds.
Choosing variables by association alone. A strong association with the exposure or outcome does not, by itself, determine whether a variable belongs in an adjustment set. The relevant question is how it lies on the causal paths in the assumed diagram.

Key takeaway: A candidate adjustment set must contain no descendants of the exposure and must block every backdoor path to the outcome. The conclusion is only as credible as the causal diagram and the additional assumptions required for adjustment.

Check Your Understanding

For each question, identify the relevant paths or criterion rather than relying only on associations.

A DAG contains \(C \rightarrow X\), \(C \rightarrow Y\), and \(X \rightarrow Y\). Does the empty set satisfy the backdoor criterion? Does \(\{C\}\)? Explain.
A proposed diagram has \(X \leftarrow S \rightarrow A\), but no arrows from \(A\) toward \(Y\). Can \(X \leftarrow S \rightarrow A \rightarrow Y\) be listed as a path in that diagram? Why or why not?
In a DAG with \(A \rightarrow X,Y\) and \(S \rightarrow X,Y\), why does adjusting for \(A\) alone fail to block every backdoor path?
What two graphical conditions define a set that satisfies the backdoor criterion?
Name two data or causal assumptions, beyond the graphical criterion itself, that are needed for an adjusted comparison to support a causal interpretation.