A Category Is Not a Cause

A label can be accurate and still explain nothing. Calling a storm a category four summarizes measured conditions. Calling a transaction suspicious routes it for review. Calling a cluster of symptoms a syndrome helps people communicate. None of those labels, by itself, identifies the process that produced the event.

Categories are indispensable compression. They let people coordinate, compare cases and decide what to examine next. The error begins when a classification is treated as a mechanism: it happened because it belongs to this category. That sentence often repeats the observation in a more official voice.

Categories compress; causes produce

A category groups cases according to selected features. The features may be visible, measured, reported or inferred. The grouping can be binary, ordered, overlapping or probabilistic. A cause is different: it is part of a claim about how changing one condition would change an outcome, under stated assumptions.

The two can be connected. A category may correlate with a cause, contain a cause in its definition or identify a population in which a mechanism is more common. But the connection must be established. It does not arrive free with the noun.

Consider an error message. “Network failure” may correctly group many incidents. One was caused by a failed cable, another by expired credentials, another by a broken DNS path and another by a remote service outage. The category helps select a troubleshooting branch. It becomes harmful when the team repeatedly “fixes the network” without distinguishing the mechanisms.

Why the confusion is so attractive

Labels are cognitively efficient. Once a case has a name, uncertainty feels smaller. Institutions also need stable terms for records, eligibility, reporting and treatment. A label can therefore acquire authority from the work built around it. The form has a field. The database has a code. The expert recognizes the term. The system can now proceed.

Proceeding is not the same as explaining. A category may be operationally necessary while remaining scientifically provisional. It may be optimized for reliability between raters rather than for a single underlying cause. It may combine several pathways because they require the same immediate response. Or it may split one pathway into different categories because severity or context changes the decision.

The National Institute of Mental Health's Research Domain Criteria program makes this distinction visible in one consequential domain. RDoC is a research framework that studies dimensional constructs across multiple units of analysis rather than assuming that existing disorder categories each correspond to one homogeneous mechanism. That does not invalidate clinical diagnoses. It shows why a diagnostic category and a causal account answer different questions.

Four jobs a label can do

Many disputes dissolve when the job is named. A triage label can be useful because it routes attention, even if it is not a theory of disease. A content-moderation category can define a policy boundary without explaining a speaker's motive. A customer segment can predict purchasing behavior without revealing a stable psychological type.

One category can contain many mechanisms

Cases can look similar at the output while differing upstream. A late project may reflect unrealistic estimation, unclear ownership, a vendor failure, changing requirements or deliberate de-prioritization. The shared category “late” is true. Treating all lateness as a motivation problem is not.

This is causal heterogeneity. It matters because interventions act on mechanisms, not labels. More reminders may help an attention failure and worsen an overloaded queue. More training may help a skill gap and do nothing for broken permissions. A category-level average can even hide opposite effects in important subgroups.

Before acting, list at least three plausible pathways. Then ask which observations would differ among them. The aim is not to invent endless stories. It is to prevent the first available label from becoming the only model.

One mechanism can cross categories

The inverse problem also occurs. One upstream condition can produce several visible outcomes. Poor sleep can affect attention, mood, reaction time and appetite. A failing power supply can cause restarts, data corruption and device dropouts. If every output is handled inside its own category, the shared cause stays invisible.

Cross-category evidence is therefore valuable. Ask whether the same timing, environment, dependency or intervention changes several outcomes together. This does not prove one common cause, but it generates a more testable model than treating each label as an isolated thing.

Prediction is not mechanism

A category can predict well without being causal. An address may predict delivery risk because it is associated with terrain, weather, infrastructure and service patterns. The address itself is not the mechanism. An AI classifier may recognize a disease from a hospital-specific marker rather than from pathology. The prediction can look impressive until the context changes.

This is why the site's right-answer-for-the-wrong-reason problem matters beyond AI. Predictive performance answers whether the rule works under tested conditions. Explanation asks why, where it should transfer and what would happen under intervention. Do not demand a causal explanation when prediction is the honest goal. Do not advertise prediction as explanation when action depends on mechanism.

A six-part category-cause record

1. Definition: write the exact features and thresholds that place a case inside the category.
2. Purpose: state whether the label is meant to describe, predict, route or explain.
3. Variation: list meaningful differences among cases that the label compresses.
4. Pathways: name plausible mechanisms, dependencies and contextual alternatives.
5. Discriminating evidence: identify observations or tests that would favor one pathway over another.
6. Intervention: state what action should change the outcome if the proposed mechanism is real, plus risks and stop conditions.

This record pairs well with the claim-evidence-inference worksheet. The category is usually part of the evidence description. The causal mechanism belongs in the inference until an appropriate design has tested it.

Use verbs that reveal the evidence level

Language can keep the distinction visible. Prefer “is classified as,” “is associated with,” “predicts under these conditions,” “triggers this workflow” or “is consistent with” when that is what the evidence supports. Reserve “causes,” “produces” and “prevents” for a causal claim with an identified comparison and assumptions.

This is not stylistic timidity. It is interface design for reasoning. Precise verbs let a reader see where observation ends and explanation begins.

Categories are still necessary

Rejecting reification does not mean abandoning classification. Without categories, every case becomes incomparable and every decision starts from zero. The useful discipline is lighter: keep the label tied to its purpose, preserve variation inside it and do not let the database field overrule contradictory evidence.

When a category controls access to care, benefits, education, credit, moderation or employment, an appeal path also matters. Borderline cases and unusual mechanisms are exactly where a compressed rule is most likely to fail. Read A Threshold Is a Policy Choice for the separate problem of where a category boundary is placed.

Claims and boundaries

Sourced fact: NIMH describes RDoC as a research framework that studies dimensional constructs and multiple units of analysis rather than relying only on disorder-based categories. Inference: the distinction between operational labels and causal mechanisms generalizes to many technical and institutional systems. Judgment: consequential classifications should state their purpose, uncertainty and review path. Not claimed: categories are unreal, diagnoses are useless, every case has one cause or causal knowledge is always required before action.

Primary sources and further reading

• National Institute of Mental Health: About RDoC, for the distinction between disorder categories and research on dimensional constructs and mechanisms.
• NIMH: RDoC domain and construct definitions, for the framework's evolving constructs and units of analysis.
• Cuthbert and Morris: Three Approaches to Understanding and Classifying Mental Disorder, for classification, dimensionality, thresholds, etiology and comorbidity.
• What Would Change the Outcome?, for a practical anatomy of causal claims.