Proving Identification With No Recanting Witness

Why the No-Recanting-Witness Condition Matters

A path-specific effect asks what would happen if treatment were changed along some causal paths to the outcome, while other paths carried a reference treatment value. The difficulty is that a variable shared by selected and unselected paths may need to transmit two different treatment-induced values at once. The no-rec anting-witness condition rules out that conflict. Under an acyclic causal graph with observed variables and independent structural errors, it allows the path-specific outcome to be represented by an edge intervention and identified from the observed distribution.

The result has two logically separate parts. First, an edge intervention has a distribution given by an edge g-formula. Second, when there is no recanting witness, this edge-intervention outcome is the desired path-specific outcome. The first part follows from structural factorization; the second is the substantive role of the graph condition.

Setup and Definitions

Definition: Let \(G\) be a finite directed acyclic graph (DAG) with observed variables \(V\), treatment \(A\), and outcome \(Y\). Let \(\pi\) be a set of directed paths from \(A\) to \(Y\), called the target paths. For treatment values \(a\) and \(a^*\), the path-specific assignment sends value \(a\) along edges out of \(A\) that lie on a target path, and value \(a^*\) along the other edges out of \(A\) that lie on directed paths to \(Y\). The resulting outcome is denoted \(Y^\pi(a,a^*)\). Edges out of \(A\) that cannot reach \(Y\) do not affect this outcome.

An edge intervention changes the value supplied along a specified edge, rather than setting a node to different values for different children. For a node \(v\), write \(\mathrm{pa}_v\) for its parents. Under an edge intervention \(e\), let \(x_{\mathrm{pa}_v}^{\,e}\) denote the vector of values supplied to \(v\): for an intervened edge \(A\to v\), use its assigned treatment value; for an edge not intervened on, use the value of that parent in the counterfactual system.

Definition: A node \(W\), other than \(A\) or \(Y\), is a recanting witness for \(\pi\) if it lies on at least one target directed path from \(A\) to \(Y\) and at least one directed path from \(A\) to \(Y\) that is not in \(\pi\). The target path set has no recanting witness if no such node exists. A recanting node would need to carry one treatment version for a target path and another for an unselected path.

The structural assumptions used below are explicit. The variables obey a correctly specified, finite acyclic structural causal model: each variable is a function of its parents and a structural error. The structural errors are mutually independent, and every variable used in the argument is observed. Thus the observational distribution factorizes according to \(G\), and each conditional distribution \(p(x_v\mid x_{\mathrm{pa}_v})\) represents the corresponding structural mechanism. Also assume consistency and positivity: for discrete variables, the required parent-value combinations have positive probability; for continuous variables, assume support conditions and suitable regularity, such as continuity, that identify the required conditional kernels at the intervention values. Variables may be discrete, continuous, or mixed.

Identification Result and Derivation

For each edge out of \(A\) that lies on a directed path to \(Y\), assign \(a\) if the edge lies on a target path and \(a^*\) otherwise. In a no-recanting-witness setting these assignments do not demand incompatible values for a shared edge. The edge intervention fixes the treatment inputs along these edges while leaving all structural mechanisms and other exogenous errors unchanged.

Theorem: Under the stated structural assumptions, if \(\pi\) has no recanting witness, then \(Y^\pi(a,a^*)\) is identified by the edge g-formula. In particular, its distribution at \(y\) is obtained by evaluating each observed conditional mechanism at the parent values supplied by the edge intervention, multiplying the mechanisms, and marginalizing all variables other than \(A\) and \(Y\). For mixed variable types, marginalization means integration over continuous variables and summation over discrete variables.

More explicitly, use \(p\) to denote the appropriate observed conditional density or probability mass function. Let \(\mathcal X_v\) be the state space of \(X_v\), and let \(\nu_v\) be Lebesgue measure for a continuous variable and counting measure for a discrete variable. The formula is

p\bigl(y^\pi(a,a^*)\bigr) = \int_{\prod_{v\in \operatorname{An}(Y)\setminus\{A,Y\}}\mathcal X_v} \prod_{v\in \operatorname{An}(Y)\setminus\{A\}} p\bigl(x_v\mid x_{\mathrm{pa}_v}^{\,e}\bigr) \prod_{v\in \operatorname{An}(Y)\setminus\{A,Y\}} d\nu_v(x_v), \qquad x_Y=y.

Here the notation on the left refers to the distribution of the counterfactual outcome, not a density evaluated at a random argument. In the integral, each discrete coordinate is summed over its possible values and each continuous coordinate is integrated. The factor for \(Y\) is included in the product and evaluated at \(y\). There is no factor for \(A\), because its outgoing edge inputs have been set directly; all other conditional mechanisms remain in the product.

The derivation has two steps. First, in the observational model, independence of structural errors gives the factorization \(\prod_{v\in V}p(x_v\mid x_{\mathrm{pa}_v})\). Under the edge intervention, the equation for each node is unchanged except that its inputs on intervened edges are replaced by their assigned values. Independence of its structural error from its parents means the conditional mechanism for that node is still the observed kernel, evaluated at the substituted parent vector. Multiplying these kernels gives the joint distribution under the edge intervention. Integrating or summing out all variables except \(Y\) gives the displayed formula.

Second, the graph condition connects that edge intervention to the path-specific counterfactual. Evaluate the structural equations in a topological order, so each node is evaluated only after its parents. When there is no recanting witness, no intermediate node on an \(A\)-to-\(Y\) path belongs to both a target path and an unselected path. Its value therefore does not have to serve as one treatment version for a selected continuation and a different version for an unselected continuation. Recursively substituting the structural equations consequently gives the same outcome as the path-specific assignment. This establishes \(Y^\pi(a,a^*)=Y^e\) for the target outcome, and hence the edge g-formula identifies its distribution and mean.

Worked Applications

Worked Example: A Binary Mediator and a Direct Path

Suppose the graph has \(A\to M\), \(M\to Y\), and \(A\to Y\). The target path is \(A\to M\to Y\); the direct path is unselected. There is no recanting witness: \(M\) lies only on the target path, and the direct edge has no intermediate node. Let \(A\), \(M\), and \(Y\) be binary. Suppose the hypothetical observed conditional probabilities are \(P(M=1\mid A=1)=0.60\), \(P(M=1\mid A=0)=0.20\), \(P(Y=1\mid A=0,M=0)=0.10\), and \(P(Y=1\mid A=0,M=1)=0.50\). Consider \(a=1\) and \(a^*=0\), with independent structural errors and positivity as specified above.

The edge assignment uses \(A=1\) on \(A\to M\) and \(A=0\) on \(A\to Y\). Therefore the identified risk is

P\{Y^\pi(1,0)=1\} = P(M=0\mid A=1)P(Y=1\mid A=0,M=0) +P(M=1\mid A=1)P(Y=1\mid A=0,M=1).

Substituting the values gives \( (1-0.60)(0.10)+(0.60)(0.50)=0.040+0.300=0.340\). For the reference assignment, both edges carry \(A=0\), so the risk is \( (1-0.20)(0.10)+(0.20)(0.50)=0.080+0.100=0.180\). The path-specific risk difference is \(0.340-0.180=0.160\), or 16.0 percentage points. This is a causal interpretation only under the stated graph and structural assumptions; the arithmetic alone does not establish them.

Worked Example: Seeing the Conflict at a Recanting Witness

Consider \(A\to L\), \(L\to Y\), and \(L\to M\to Y\). Let the target path set contain \(A\to L\to Y\), but not \(A\to L\to M\to Y\). The node \(L\) is on both a selected and an unselected path, so it is a recanting witness. To isolate the target path, its value supplied to \(Y\) must be \(L(1)\), while the value supplied to \(M\) must be \(L(0)\).

For a concrete hypothetical structural model, let \(L=A\), \(M=L\), and let \(Y\) be binary with \(P(Y=1\mid L=l,M=m)=0.10+0.60l+0.20m\). These probabilities are valid for binary \(l,m\), with maximum \(0.90\). The target path-specific assignment would require \(L=1\) for the \(L\to Y\) input and \(L=0\) for the \(L\to M\to Y\) input. It would therefore give \(P(Y^\pi(1,0)=1)=0.10+0.60(1)+0.20(0)=0.70\).

But a single edge intervention on \(A\to L\) supplies only one value to \(L\). If that value is 1, then \(L=1\), \(M=1\), and the outcome risk is \(0.10+0.60(1)+0.20(1)=0.90\). If it is 0, then \(L=0\), \(M=0\), and the risk is \(0.10\). Neither edge intervention produces the required split value \(0.70\). This calculation does not prove a general nonidentification result; it shows why the no-recanting-witness condition is needed to equate a path-specific assignment with a single coherent edge intervention.

Worked Example: A Continuous Baseline Variable

Let \(X\) be uniformly distributed on \([0,1]\), with edges \(X\to A\), \(X\to M\), \(X\to Y\), \(A\to M\), \(A\to Y\), and \(M\to Y\). Select \(A\to M\to Y\) and leave \(A\to Y\) at the reference value. There is no recanting witness. Suppose the hypothetical conditional mechanisms are \(P(M=1\mid A=1,X=x)=0.20+0.40x\), \(P(M=1\mid A=0,X=x)=0.10+0.20x\), \(P(Y=1\mid A=0,M=0,X=x)=0.05+0.10x\), and \(P(Y=1\mid A=0,M=1,X=x)=0.10+0.20x\). All these probabilities lie between 0 and 1 for \(0\leq x\leq1\); assume the other structural conditions and positivity hold.

For \(a=1,a^*=0\), write \(q_1(x)=0.20+0.40x\). Conditional on \(X=x\), the edge g-formula gives the risk \((1-q_1(x))(0.05+0.10x)+q_1(x)(0.10+0.20x)\). Since the difference between the two outcome risks is \(0.05+0.10x\), this simplifies to \(0.05+0.10x+(0.20+0.40x)(0.05+0.10x) =0.06+0.14x+0.04x^2\). Because \(X\) has density 1 on \([0,1]\), integrate rather than sum:

P\{Y^\pi(1,0)=1\} =\int_0^1(0.06+0.14x+0.04x^2)\,dx =0.06+0.07+\frac{0.04}{3} =0.1433.

For the reference assignment, \(q_0(x)=0.10+0.20x\). The corresponding conditional risk is \(0.05+0.10x+(0.10+0.20x)(0.05+0.10x)=0.055+0.12x+0.02x^2\). Thus its marginal risk is \(\int_0^1(0.055+0.12x+0.02x^2)\,dx =0.055+0.060+0.006667=0.1217\). Using the unrounded marginal risks, the path-specific risk difference is approximately \(0.0217\), or about 2.17 percentage points. The continuous baseline is averaged over its distribution, while the binary mediator is summed over its two values.

What the Result Does—and Does Not—Establish

The no-recanting-witness result is a graph-based identification argument. It does not say that the target effect is large, that the causal graph is correct, or that the assumptions can be verified from the observed data alone. The independent-error condition rules out unmeasured common causes in the structural model; positivity is needed so the conditional mechanisms required by the edge assignment can be learned from observed parent-value combinations. Violations of these conditions undermine the identification claim even if the graph has no recanting witness.

A useful proof strategy is to keep the two steps distinct: factorize the edge-intervention distribution using the structural mechanisms, then use the path structure to show that the edge intervention represents the target path-specific outcome. The formula handles mixed state spaces through integration and summation over the appropriate variables; it does not assume every variable is discrete.

Common mistake: “The edge g-formula is available, so any path-specific effect is identified.” The formula identifies the distribution under a well-defined edge intervention, but a recanting witness can prevent that intervention from representing the desired path-specific counterfactual.
Correction: Check that no intermediate node lies on both selected and unselected paths before making that identification step.
Common mistake: Writing a sum over all marginalized variables without checking their types. Use sums for discrete variables and integrals for continuous variables, as in the mixed-variable example.

Check Your Understanding

Use the graph condition and the factorization argument to answer the following.

In a graph with paths \(A\to M\to Y\) and \(A\to Y\), which edge carries \(a\) when only the mediated path is selected?
Why does a node on both a target path and an unselected path create a problem for a path-specific assignment?
What assumptions allow the conditional structural mechanisms in the edge g-formula to be represented by observed conditional distributions?
How should the marginalization in the edge g-formula be written when one marginalized variable is continuous and another is discrete?
Does the existence of an edge g-formula alone establish that a target path-specific effect is identified? Explain the role of the no-recanting-witness condition.