In 1946, Joseph Berkson, a biostatistician at the Mayo Clinic, observed something striking in hospital records. Among inpatients, those with diabetes appeared to have a lower rate of gallbladder disease than those without. The pattern was real in the data. It was statistically significant. Several clinicians had taken note, and the discussion in the wards had begun to drift toward causal explanations. Perhaps something about diabetes—a metabolic effect, a change in bile chemistry, an alteration in tissue response—was protective against gallbladder disease. The clinicians thought the next step was to investigate the mechanism. Berkson stopped the discussion before it left the building. The pattern, he showed, was not evidence of a protective mechanism. It was an artifact of looking at hospital inpatients rather than the population at large. Patients are admitted because they have at least one serious illness. If a patient does not have diabetes, the reason for hospitalization must be something else, and that something else is somewhat more likely to involve a gallbladder issue. If a patient does have diabetes, the diabetes itself can be a sufficient reason for hospitalization. The apparent protective effect existed only inside the hospital, where it was an arithmetic consequence of the selection mechanism.

The striking aspect of Berkson's case is not the bias itself, but who got it wrong. The clinicians were among the best-trained physicians in the world. They were experienced. They had thoughtful colleagues to discuss findings with. They had access to high-quality data. They were not careless. They saw a real correlation, formed a plausible causal hypothesis, and were preparing to investigate the mechanism. Berkson's intervention saved them from a research program that would have produced confident but false conclusions. Yet eighty years later, the bias is still invisible to most experts most of the time. The mechanism that catches careful clinicians at the Mayo Clinic also catches careful product managers, careful policy analysts, careful military strategists. It is not a failure of effort or competence. It is a structural feature of specialized cognition under conditions where the structure of the data does not match the structure of the question. The expert has deep knowledge about their domain but no automatic sensitivity to the selection mechanisms that created the dataset they are analyzing.

Collider blindness is the failure to recognize that a variable in the diagram is a common effect of two or more other variables, and that conditioning on it—by stratifying the data, restricting the sample, or adding it to a regression—induces a spurious association between its causes. In the Berkson case, being hospitalized is the collider. Diabetes causes hospitalization. Gallbladder disease causes hospitalization. These are independent causes in the general population. But when we restrict our sample to people in the hospital, we create a statistical dependence between diabetes and gallbladder disease that did not exist before. The mathematics is unavoidable. The bias does not depend on the clinicians' competence or statistical sophistication. The bias is built into the structure of the data once you condition on the collider. An expert looking at correlation within that restricted dataset will naturally form causal hypotheses about the apparent relationship, completely unaware that the relationship is an artifact of the selection mechanism that created the sample.

The causal structure is simple and clean. Diabetes causes hospitalization. Gallbladder disease causes hospitalization. In the population at large, diabetes and gallbladder disease have no causal connection to each other. However, once we restrict our sample to people in the hospital—people for whom at least one of these conditions is true—the mathematics changes. If a patient does not have diabetes, the reason for hospitalization must be something else, and that something else is somewhat more likely to be a gallbladder problem. If a patient does have diabetes, the diabetes alone can explain their hospitalization, making a gallbladder problem less likely. This is not a claim about medical mechanisms. It is a claim about arithmetic. Given only people with at least one cause present, the causes must be negatively correlated. The expert sees the negative correlation in the hospital data and naturally looks for a causal explanation—a mechanism by which diabetes protects against gallbladder disease. The expert is reasoning correctly within the dataset. The problem is that the dataset contains only hospitalized patients, not the population that the causal question is about.

Collider blindness is invisible because the expert is asking a reasonable question about domain causation, and the data is providing a clear, mathematically valid answer to a different question. The expert's question is about the population: what is the relationship between diabetes and gallbladder disease? The data's answer is about the hospital: among patients already admitted, what is the conditional relationship? The expert has domain knowledge that makes them sensitive to certain causal patterns—metabolic mechanisms, tissue interactions, disease pathways—and insensitive to others. They are not automatically attuned to the selection mechanisms that created their dataset. The structure of how patients enter the hospital is not part of medical domain knowledge in the way that pathophysiology is. So the expert matches the empirical correlation to domain-plausible mechanisms and does not think to question whether the correlation itself is an artifact of selection. The expert is not being careless. They are using domain knowledge exactly as domain knowledge should be used—to make fast, confident inferences from patterns. The problem is that the pattern is real, but it reflects the data structure, not the domain structure.

The Berkson case illustrates a broader pattern. When experts are asked to specify the structure of cause-and-effect relationships in a domain, three kinds of structural errors cluster together. Collider blindness is one—the failure to recognize a common effect when conditioning on it. Feedback-loop simplification is another—when a causal structure contains cycles or feedback mechanisms, experts often describe it as a linear chain, omitting the return paths that determine whether outcomes are stable, unstable, or oscillating. Domain-matching heuristics is the third—when data shows a correlation, the expert matches it to a familiar causal pattern in the domain and stops investigating, without checking whether the data structure actually supports the causal claim. These are not random mistakes. They appear repeatedly, they catch careful people, and they persist because the expert's attention is on the domain content, not on the data structure. A product manager applying domain knowledge from their field, a policy analyst using causal models from economics, a military strategist using known patterns of conflict will each encounter the same blindnesses in slightly different forms. Each of these errors can produce confident, detailed, plausible conclusions that are systematically wrong.

The critical insight is that these are not random errors. They are structural. They appear in the same form across domains, they catch careful experts, and they persist even when warned because the expert's attention is on domain content, not data structure. The error at Mayo Clinic illustrates this perfectly. A statistician whose entire job is to understand data structure had to explicitly intervene to stop medical experts from proceeding. The medical experts were not less competent. They were operating with different expertise. Moreover, the bias is structural in another sense: it does not disappear when you multiply the number of experts. If every expert in a room shares the same blindness to data structure, consensus becomes evidence of systematic error rather than evidence of truth. You cannot fix this by hiring more domain experts or having them debate. You have to fix it structurally—by making the data structure explicit and testing whether it aligns with the domain structure the expert is assuming. The expert is necessary. Observational data alone cannot identify causal structure. But the expert is also systematically blind to certain kinds of errors, and those errors have to be controlled for architecturally, not by exhorting the expert to be more careful.

Chapter 14 established that the expert is mathematically necessary. Observational data alone cannot identify causal structure. The solution requires knowledge from outside the data, which only the expert can supply. This chapter has shown that the expert also brings systematic blindness—structural errors in how causal relationships are conceptualized when the expert's domain knowledge is applied to data whose structure is not aligned with the domain. The architectural consequence is that you cannot simply extract the expert's causal diagram and use it as a model. You must interrogate the diagram, but not by questioning the expert's competence. The expert has not made a mistake in the domain sense. You interrogate by making the data structure explicit—by asking what selection mechanism created this dataset, what variables determine whether a unit appears in the sample, whether feedback loops have been linearized, whether correlations in the data match the causal claims across different subsets. The interrogation is structural, not adversarial. You are checking whether the expert's model of how causation works in the domain is actually a model of how causation works in this particular data. If the data structure and domain structure are misaligned, the expert's causal claims need to be revised—not because the expert was wrong about the domain, but because the domain model does not apply to this specific data.

The deep claim is that expert causal knowledge is domain knowledge, not data knowledge. An expert knows how causation works in the world—how processes unfold, how mechanisms operate, what variables matter and in what relationships. That knowledge is real and necessary. But it is knowledge about the domain, not knowledge about this dataset. When the dataset has a selection mechanism, a measurement structure, or a temporal alignment that differs from the expert's implicit model of how observations are made, the expert's causal claims will be misaligned with the data. The expert will not realize this misalignment because it is not part of domain knowledge to model how datasets are created. This is not a weakness of expertise. It is a feature of how expertise works. Expertise consists of deep knowledge about one thing—the domain—with corresponding blindness to other things that fall outside the domain. The architectural consequence is that a Living Model cannot simply defer to expert judgment, nor can it ignore expert judgment and use data algorithms alone. It must interrogate expert causal claims against the structure of the data, making both structures explicit and testing whether they are aligned.

The expert brings necessary causal knowledge about how their domain works. But that knowledge is structured around the causal mechanisms and relationships that matter in the domain, not around the mechanisms that created the dataset. Collider blindness, feedback-loop simplification, and domain-matching heuristics are not failures of effort or competence. They are structural features of specialized cognition—inevitable consequences of the way human experts understand their fields. The Mayo Clinic clinicians in 1946 were among the world's best physicians. They saw a real correlation and reasoned from domain knowledge to causal hypothesis. The reasoning was sound. The blindness was invisible because it is always invisible to the person operating within the domain. The architectural solution is not to distrust the expert or to replace the expert with algorithms. The solution is to build systems that make both the domain structure and the data structure explicit, allowing systematic interrogation of whether they are aligned. When they are not aligned, the expert's domain knowledge remains valid, but the causal claims require revision. This is the foundation of Part Three's approach to eliciting, validating, and using expert knowledge in causal modeling.