Why a Variable Can Block a Path-Specific Effect
A causal effect can travel from treatment to outcome along several directed paths. A path-specific effect asks what would happen if treatment were changed along some of those paths while its influence along other paths were held at a reference level. This distinction matters when a variable lies upstream of more than one route: the value needed to transmit the treatment effect along a selected route may conflict with the value needed to keep an unselected route at its reference level.
The recanting witness criterion names this graphical conflict. It is a diagnostic for whether a path-specific effect can be identified from the observed distribution under a nonparametric causal model represented by a directed acyclic graph (DAG). It does not say that the effect is meaningless or that it cannot be calculated if a sufficiently detailed causal model is known. It says that the observed distribution alone does not generally determine it.
Paths, Edge Assignments, and the Criterion
Let \(A\) be a treatment and \(Y\) an outcome. A directed causal path from \(A\) to \(Y\) is a sequence of arrows leading from \(A\) to \(Y\). Choose a collection of these paths as the selected paths; the other directed paths from \(A\) to \(Y\) are unselected. A path-specific contrast compares a regime that transmits treatment level \(a\) along selected paths and reference level \(a^*\) along unselected paths with a reference regime. For illustration below, \(a=1\) and \(a^*=0\).
In an edge intervention, the treatment assignment is specified along edges rather than by giving a single treatment value to every part of the causal system. As this assignment propagates through the graph, a downstream variable can be required to take its value under \(A=1\) for one outgoing route and its value under \(A=0\) for another. The conflict is about counterfactual values for the same variable, not about two different observed values recorded for one person.
The key question is not simply whether a variable appears on multiple paths. It is whether the path-specific assignment asks that variable to transmit different treatment regimes along different outgoing routes that can affect the outcome. The treatment \(A\) and outcome \(Y\) themselves are not counted as recanting witnesses for this definition.
Worked Examples: Applying the Criterion
Worked Example: A Selected Mediated Path With No Recanting Witness
Consider the DAG \(A \to M \to Y\), together with the direct arrow \(A \to Y\). Select the mediated path \(A \to M \to Y\), leaving \(A \to Y\) unselected. The only intermediate variable on the selected route is \(M\). It transmits treatment influence to \(Y\) through its one relevant outgoing route, \(M \to Y\); there is no second outgoing route from \(M\) that must carry the reference assignment. Thus, this graph has no recanting witness for this selection.
For a numerical illustration, suppose the hypothetical structural relations are \(M(a)=a\) and \(Y(a,m)=2a+3m\), with outcome measured in arbitrary outcome units. The selected-path contrast changes the mediated input while holding the direct treatment input at 0:
Substitute the structural values: \(M(1)=1\) and \(M(0)=0\). Therefore,
The illustrative path-specific contrast is 3 outcome units. This calculation describes the target in the hypothetical model; the absence of a recanting witness is the graphical diagnosis, not by itself a claim that every other identification assumption is satisfied.
Worked Example: An Exposure-Induced Witness on Two Routes
Now consider \(A \to L\), \(L \to M\), \(M \to Y\), and \(L \to Y\). Select \(A \to L \to M \to Y\), and leave \(A \to L \to Y\) unselected. The variable \(L\) is downstream of treatment and has outgoing causal influence along both routes. To transmit the treated regime along the selected route, the assignment requires \(L(1)\) to determine \(M\). To keep the direct route through \(L \to Y\) at its reference regime, it requires \(L(0)\) for that route. Thus \(L\) is a recanting witness.
The conflict can be written explicitly. Suppose, only to illustrate the nested counterfactual target, that \(L(a)=a\), \(M(l)=l\), and \(Y(l,m)=2l+4m\). The selected-route contrast holds the direct input at the reference value and changes the input passed through \(M\):
Here \(L(0)=0\), \(L(1)=1\), \(M(L(1))=M(1)=1\), and \(M(L(0))=M(0)=0\). Substituting gives
The numerical target is 4 outcome units in this hypothetical model. But its first term uses \(L(1)\) to form the mediator and \(L(0)\) for the direct outcome route. The arithmetic shows what the path-specific contrast asks for; it does not remove the identification problem. Without additional restrictions, the observed-data distribution does not generally reveal this cross-route counterfactual combination.
Worked Example: Several Branches Without a Witness
Consider two separate routes: \(A \to L \to M_1 \to Y\) and \(A \to K \to M_2 \to Y\), with no arrows connecting the intermediate variables. Select the route through \(M_1\) and leave the route through \(M_2\) unselected. Although there are multiple causal paths, neither \(L\) nor \(K\) sends influence along both a selected and an unselected route. Neither is a recanting witness.
For a hypothetical numerical illustration, suppose \(L(a)=a\), \(K(a)=a\), \(M_1(l)=l\), \(M_2(k)=k\), and \(Y(m_1,m_2)=5m_1+2m_2\). The selected-route contrast changes \(M_1\) from its value under 0 to its value under 1, while holding \(M_2\) at its value under 0:
The inputs are \(M_1(L(1))=1\), \(M_1(L(0))=0\), and \(M_2(K(0))=0\). Substitution and arithmetic give
The illustrative contrast is 5 outcome units. The graph contains distinct selected and unselected routes, but it does not require \(L\) or \(K\) to take incompatible values across routes. The number of paths alone therefore does not determine whether a recanting witness exists.
Why the Witness Creates an Identification Obstruction
The basic logic is visible by tracing a variable’s value downstream. Along any one route, a structural equation uses the value of a parent variable to determine the next variable. If the selected and unselected routes branch after \(W\), the path-specific assignment may need \(W(1)\) as an input on one branch and \(W(0)\) on the other. The resulting outcome is a nested counterfactual built from two treatment regimes at the same variable.
Observed data show variables under the treatment regime that actually occurred; they do not generally reveal how the same unit’s \(W(1)\) and \(W(0)\) jointly combine to produce an outcome with different branch-specific inputs. Even knowing the separate distributions of \(W(1)\) and \(W(0)\) would not, in general, determine their joint counterfactual relationship. This is why the recanting witness signals a failure of nonparametric identification, not merely a difficult calculation.
The criterion is a graphical test for this particular obstruction. It does not replace the other assumptions required for causal interpretation, such as a suitable causal graph and consistency. Nor does the absence of a witness prove identification without qualification: other problems, including confounding or a lack of support for relevant treatment combinations, can still prevent identification.
A Practical Diagnostic
List the directed paths from treatment \(A\) to outcome \(Y\), and label each as selected or unselected for the target effect.
For each intermediate variable downstream of \(A\), check whether it sends causal influence toward \(Y\) along both kinds of path.
Ask whether the selected route requires the variable’s value under \(a\) while the unselected route requires its value under \(a^*\). If so, the variable is a recanting witness.
A witness means the path-specific effect is not generally identified from the observed distribution alone. It does not mean the target is undefined or that identification is impossible under every added model or assumption.
Common Mistake
A common mistake is to conclude that any mediator affected by treatment is a recanting witness. That reasoning is too broad: in the first example, \(M\) is affected by \(A\), but its influence lies only on the selected route, so it does not have to transmit incompatible values along selected and unselected branches. The correction is to check both parts of the criterion: the variable must be downstream of treatment, and the path-specific assignment must require different treatment-induced values for its influence along the two kinds of route.
A second mistake is to read “recanting” as a claim about a person’s behavior or intentions. It is a technical description of a counterfactual assignment in the graph. The variable is not choosing between values; the target effect requires different potential values of that variable to feed different branches.
Check Your Understanding
For each question, explain your reasoning in terms of the selected paths and the values required along each route.
- In the graph \(A \to M \to Y\) and \(A \to Y\), with only the mediated path selected, why is \(M\) not a recanting witness?
- In the graph \(A \to L\), \(L \to M\), \(M \to Y\), and \(L \to Y\), which variable is the recanting witness when only the route through \(M\) is selected?
- Does the existence of multiple directed paths from \(A\) to \(Y\) automatically imply a recanting witness? Explain.
- What does the recanting witness criterion say about identification from the observed-data distribution, and what does it not say about the meaning of the target effect?
- Suppose a variable is affected by treatment but has outgoing causal influence toward \(Y\) only on selected paths. Is it a recanting witness for that path selection? Why or why not?