Tracing Causal Paths to Set the Mechanistic Boundary

Why Trace Causal Paths?

A biological system can contain far more processes than a particular QSP decision requires. The model boundary is not a catalogue of everything known about the disease or drug. It is a proposed scope for representing the causal processes needed to connect an intervention to the outcome that informs a decision. Causal path tracing makes that scope explicit: start at the intervention, follow plausible causal links, and ask which processes can transmit its effect to the decision-relevant outcome.

The outcome and decision context must be stated before drawing the boundary. For example, a pathway irrelevant to an efficacy endpoint may matter to a safety endpoint, and a mechanism important for one patient population may be inactive in another. The earlier material on decision questions and model output boundaries establishes why that context matters. Here, the focus is the causal diagram used to trace routes within it.

A causal diagram uses nodes for variables or processes and arrows for hypothesized causal influence. A direct path is a single arrow, such as \(D \rightarrow Y\). A mediated path includes one or more intervening nodes, such as \(D \rightarrow M \rightarrow Y\). Several paths may run in parallel. A node affected by the intervention but with no directed route onward to the specified outcome is a branch that does not transmit an effect to that outcome in the diagram.

The arrows are causal claims, not merely observed associations. A diagram organizes a proposed explanation; it does not establish that the links are true. Its use assumes that the intervention is sufficiently well defined, the relevant context and outcome are specified, the temporal ordering of causes is coherent, and the diagram includes the plausible causal routes needed for the question. If important causes or routes are omitted, a boundary drawn from the diagram may be inadequate.

From a Diagram to a Boundary

Begin at the intervention and trace forward along every plausible directed route that reaches the decision-relevant outcome. The mechanisms on those routes are candidates for representation. Then consider whether other variables are needed to specify how those routes operate—for example, a context variable that changes a mediator’s level or a cause of both a mediator and the outcome. Such a variable is not necessarily a drug mechanism, but omitting it may prevent the model from representing the intervention contrast appropriately.

Do not equate “not on a directed path from the intervention” with “always irrelevant.” A variable may be a common cause of a pathway component and the outcome, may define the population, or may be relevant to another decision outcome. Conversely, a biologically interesting process is not automatically required merely because it changes after treatment. The question is whether it can transmit an effect to the stated outcome under the causal structure and context being modeled.

Why Path Coefficients Matter

A simple linear structural model illustrates how paths carry an intervention effect. Suppose \(M\) is a mediator and \(Y\) is the outcome, with equations \(M=aD+U\) and \(Y=bM+cD+V\), where \(U\) and \(V\) are fixed for the intervention contrast. Substituting the first equation into the second gives \(Y=abD+bU+cD+V\). Therefore, changing \(D\) by one unit changes the predicted \(Y\) by \(ab+c\): \(ab\) is the contribution along \(D \rightarrow M \rightarrow Y\), and \(c\) is the contribution along the direct arrow \(D \rightarrow Y\). This calculation relies on the stated linear equations and fixed context; it is not a general formula for nonlinear or confounded systems.

More generally, in a linear acyclic structural model, a directed path contributes the product of its link coefficients to the total intervention effect, and the total effect is the sum of contributions across directed paths. This follows by repeatedly substituting each node’s equation into the outcome equation: each substitution generates terms for the routes through that node, and terms reaching the outcome correspond to directed paths. The result is a way to check a specified model’s implications, not evidence that the diagram or coefficient values are correct.

Parallel Routes and Excluded Branches

Path tracing is especially useful when a drug may affect more than one process. If two plausible routes reach the outcome, excluding either route can change the predicted intervention effect. A branch that changes under treatment but does not lead to the selected outcome may be outside the boundary for that question. The conclusion is specific: if the branch becomes relevant to another endpoint, the boundary must be reconsidered.

Diagnostics, Alternatives, and Limitations

A path-based boundary is a hypothesis to scrutinize, not a guarantee of adequacy. Warning signs include a claimed intervention effect with no route to the modeled outcome, a route that depends on an unsupported causal link, or a process that changes after treatment but has been excluded without considering whether it affects the chosen endpoint. Also check the direction and timing of each proposed link. An observed association between two nodes does not establish which causes which, and a biomarker can track an outcome without mediating the intervention’s effect.

The boundary should also be checked against the model’s intended use. A single-outcome boundary may be narrower than one needed for a joint efficacy-and-safety decision. A population-specific route may fail in another population. If feedback is important, a simple acyclic diagram at one time point may not represent it; time-indexed states or another explicit representation may be needed. These are reasons to revisit the proposed structure, not reasons to include all biology by default.

A broad mechanistic network can represent many candidate routes, but it increases the number of assumptions and quantities that must be supported. A reduced mechanistic model focuses on routes needed for the decision, while retaining causal interpretation for those routes. An empirical exposure–response fit may summarize a relationship in the data without representing the intervening biological pathways; it can be useful for interpolation within its supported setting, but may not answer a counterfactual that depends on changing a mechanism. These approaches answer different questions, and no diagram alone determines which is adequate.

Path tracing has limits. The diagram may omit a route, misstate a causal direction, or oversimplify interactions and context dependence. A process can influence a path without appearing as a mediator on the path itself, and nonlinear effects can make a single set of path contributions inadequate. Thus, use the diagram to make scope and assumptions inspectable, then compare its implied predictions with biological knowledge and available evidence. A defensible boundary is explicit about what is represented, what is excluded, and for which outcome and context that choice is being made.