Structural Assumption

A structural assumption is the presupposition that the causal architecture of a system — which variables are connected to which, through what mechanisms, under what boundary conditions — remains stable across contexts of application. It is the unstated scaffolding that makes causal claims transportable. Without it, a mechanism discovered in one population, one laboratory, one historical moment, cannot be assumed to operate elsewhere. The structural assumption is what permits causal inference to escape the particular and aspire to the general.

The concept sits at the intersection of philosophy of science, causal reasoning, and systems theory. It is rarely examined directly because it operates beneath the surface of empirical work: scientists test whether a causal relationship holds, but they rarely test whether the system in which that relationship is embedded has the same structure in their new context. The structural assumption is the epistemic debt that causal science accumulates and almost never repays.

The Problem of Transportability

In medicine, a drug that lowers blood pressure in a clinical trial population is assumed to lower blood pressure in the general population. But this assumption depends on the structural stability of the cardiovascular system across populations. If the trial population lacks a comorbidity that interacts with the drug's mechanism — kidney disease, say, that alters pharmacokinetics — the structural assumption fails, and the causal claim transports poorly. The failure is not of the mechanism itself but of the system's architecture in which the mechanism operates.

This is why external validity is not merely a statistical problem of sampling. It is a structural problem of system identity. The question is not 'Did we sample the right population?' but 'Does the target population have the same causal topology as the source population?' The latter question cannot be answered by statistics. It requires a model of the system's structure, and it is precisely this model that the structural assumption smuggles in without examination.

Structural Assumptions in Causal Models

In the do-calculus and structural causal models of Judea Pearl, the structural assumption is encoded in the directed acyclic graph (DAG) that specifies which variables have direct causal effects on which others. The DAG is not inferred from data; it is assumed. The data can estimate the strength of effects once the graph is given, but the graph itself comes from background knowledge — from theory, from prior studies, from the modeler's judgment about which variables matter.

This means that every causal estimate produced by a structural causal model is conditional on the correctness of the DAG. If the DAG omits a confounder, or misplaces a mediator, or assumes independence where dependence exists, the estimates are biased in ways that the data cannot detect. The structural assumption is the blind spot of causal inference: it is always present, always consequential, and almost always untested.

The systems-theoretic perspective clarifies why this is so. A complex system cannot be fully represented by any finite DAG. The DAG is a truncation — a decision to treat some interactions as negligible, some variables as exogenous, some feedback loops as absent. These decisions are not arbitrary, but they are not determined by the system either. They are determined by the modeler's purposes, the available data, and the theoretical framework within which the modeler works. The structural assumption is thus perspectival in exactly the sense that second-order cybernetics identifies: the observer's framing of the system is part of the system being modeled.

Structural Assumptions and Modularity

The concept of modularity in biology is a structural assumption in disguise. When an evolutionary biologist assumes that a gene's effect on phenotype is modular — that it influences one trait without disrupting others — they are assuming a particular causal architecture: weak coupling between the gene's pathway and other developmental pathways. This assumption is what makes evolvability possible. If the assumption fails — if the gene is pleiotropic, affecting many traits through dense connectivity — then the causal model that treated it as modular fails with it.

The same structure appears in engineering. A control engineer who designs a PID controller for a heating system assumes that the temperature dynamics are approximately linear, that the sensor and actuator are independent, and that the system has no hidden feedback loops from the environment. These are structural assumptions. They are what make the controller design tractable. They are also what make it fail when the system is installed in a building with thermal inertia, solar gain, and occupancy patterns that the original model did not include.

The Cost of the Assumption

The structural assumption is not an error to be eliminated. It is a necessity. No model can include everything. Every causal claim requires some truncation of the world's complexity. The question is not whether to make structural assumptions but whether to acknowledge them — and whether to design methods for detecting when they fail.

This is where contemporary causal inference is weakest. The field has developed extraordinarily sophisticated methods for estimating effects given a structure, but relatively crude methods for testing whether the structure itself is correct. The Rubin causal model addresses this through sensitivity analysis: how much would the estimate change if an unmeasured confounder of a certain strength existed? But sensitivity analysis does not test the structural assumption; it merely bounds its consequences.

A more radical approach would treat structural assumptions as hypotheses to be tested directly — not by estimating effects within a structure, but by comparing the predictive performance of different structures across contexts. This is the approach of ensemble modeling in machine learning and of model selection in statistics, but it has not been systematically applied to causal inference. The reason, perhaps, is that testing structural assumptions requires data from multiple contexts — multiple populations, multiple times, multiple experimental conditions — and such data are expensive and rare.

The synthesizer's claim is that this cost must be paid. Causal science that does not test its structural assumptions is not science but engineering: it works until it doesn't, and the failure modes are discovered only in production — in the clinic, in the economy, in the policy domain — where the costs are highest and the feedback slowest.

Every causal claim carries a structural assumption like a hidden loan. The interest accrues silently until the context changes, and then the debt comes due all at once. The wise modeler keeps track of the debt. The foolish one forgets it exists.