# ON THE QUESTION OF CAUSATION IN INTELLIGENCE: A CRITICAL EXAMINATION OF METHOD, AUTHORITY, AND THE LIMITS OF DIAGRAMMATIC REASON ## I. THE PRETENCE OF RIGOUR It is a peculiar convention of natural philosophy that the severest show of rigour often masks the deepest abdication of intellectual responsibility. For near a century, empirical science—most notably in the human sciences—adopted a curious discipline: the expulsion of causal language from its vocabulary. Correlation was permitted. Association was honourable. But causation was deemed the province of metaphysics, unfit for the laboratory or the calculation. This was presented as asceticism, as the hard-won wisdom of maturity. We were told that to speak of causes was to indulge in primitive thinking, that rigorous minds must content themselves with the humble description of patterns in data. We now understand this exile for what it was: not rigour, but its theatrical performance. A science that refuses to speak of causes is not more precise—it is merely mute about its deepest commitments. It is rather as if a physician, wishing to appear scientific, were to describe in meticulous detail the statistical association between fever and inflammation, whilst maintaining a studied silence on whether the fever *produces* the inflammation or merely accompanies it. The appearance of objectivity masks a kind of intellectual cowardice. The restoration of causal language to scientific discourse—most notably through the work of Judea Pearl and his formalism of causal diagrams—represents therefore a genuine return to intellectual seriousness. Yet we must be careful here. The restoration of a vocabulary is not the same as the restoration of truth. Pearl has given us tools of considerable power, certainly. The calculus of causal diagrams, the do-operator, the machinery of back-door and front-door adjustment—these are genuine achievements. They permit us to reason about interventions in ways that raw correlation cannot. They are, within their domain, *provably correct*. But here lies the rub, and it is a rub that no amount of mathematical elegance can smooth away: these tools are provably correct *given a diagram you must assume before the data can speak*. The diagram is prior to the evidence. It is, in the most fundamental sense, a *theory* about the world—and theory is not data, no matter how transparently one writes it down. ## II. THE TYRANNY OF THE DIAGRAM We must be explicit about what has occurred. Pearl's formalism has done something remarkable: it has *formalized* the problem of causal inference without solving it. It has shown us that if we can specify the causal structure correctly—if we can draw the right diagram—then we can extract causal conclusions from observational data through mechanical procedures. This is genuinely valuable. But the condition is everything. It is rather as if one had invented a perfect method for calculating the trajectory of a projectile, *provided one knew the initial velocity*. The method is flawless; the problem is that one does not know the initial velocity, and the data will not tell you. The diagram is an *assumption*. Let us be clear about this, for it is the point on which the entire enterprise turns. When a researcher sits down with Pearl's machinery, the first task is not to examine data but to construct a Directed Acyclic Graph (DAG)—a visual representation of which variables cause which others. This diagram embodies causal claims. It says: *X causes Y*. It says: *Z is a confounder*. It says: *this pathway is blocked*. These are not conclusions drawn from data. They are premises. They are what the researcher brings to the analysis. Now, the researcher may be explicit about this. Indeed, modern practice increasingly demands it. One finds in contemporary papers careful exposition of the assumed causal model, sometimes accompanied by sensitivity analyses that ask: what if the diagram were slightly different? This is transparency, certainly. It is a genuine improvement over the older practice of leaving one's causal assumptions implicit, buried in the design of the study or the choice of variables to "control for." But—and this is the crux of the matter—*transparency is not validity*. A researcher can be completely explicit about a model that is completely wrong. One can draw a diagram with perfect clarity and still have misrepresented the causal structure of the world. The diagram is not validated by being stated plainly. Indeed, there is something almost Orwellian in the notion that clarity of expression about an assumption somehow converts that assumption into knowledge. Consider a simple case, one that touches directly on questions of intelligence. Suppose a researcher wishes to understand whether educational intervention *causes* an increase in measured intelligence. The causal diagram might look like this: ``` Intervention → IQ Score ``` But of course, the diagram might be wrong in numerous ways. Perhaps the intervention does not directly affect IQ, but rather affects motivation, which affects IQ. Perhaps there is a confounder—some characteristic of the children selected for the intervention that would have increased their IQ regardless. Perhaps the measured IQ score is not a valid measure of the underlying construct, and what the intervention actually affects is test-taking skill or familiarity with the testing environment. Perhaps there is a feedback loop: higher IQ leads to selection for further intervention. The researcher can be entirely explicit about which of these alternatives they have assumed away. They can write: "We assume no unmeasured confounding. We assume the intervention affects IQ directly, not through intermediate mechanisms. We assume the outcome is a valid measure of intelligence." This is admirable transparency. But it does not make these assumptions true. And the data, no matter how extensive, cannot tell us whether they are true. The data can only tell us what the data shows, under the assumption that the diagram is correct. ## III. THE PROBLEM OF AUTHORITY Here we arrive at the second and perhaps more troubling dimension of the problem. If the diagram is an assumption, then the question becomes: *who decides what the diagram is*? Who has the authority to stipulate the causal structure of the world? In principle, the answer should be: the community of researchers, through argument and evidence and the slow accumulation of understanding. And indeed, in well-developed scientific domains—physics, for instance—there is often broad agreement about the causal mechanisms at work. We do not seriously dispute whether gravity causes objects to fall, or whether the motion of the planets is caused by gravitational attraction rather than the reverse. But in the sciences of intelligence, we are not in such a position. Intelligence itself is a contested concept. There is no consensus on what it is, how it is produced, what its causal structure is. Is intelligence primarily a product of innate capacity, or of education, or of motivation, or of cultural familiarity with the forms of reasoning that tests measure? Does it have a unitary cause (a "general factor") or multiple independent causes? Is it malleable through intervention, or largely fixed? These are not technical questions to be resolved by consulting a diagram. They are substantive questions about human nature and human potential—and they are *political* questions, whether we acknowledge it or not. The danger of Pearl's formalism, wielded in this context, is that it lends an appearance of technical necessity to what are in fact contestable theoretical commitments. A researcher can construct a diagram that embodies a particular theory of intelligence—say, one in which intelligence is largely genetically determined, with only modest room for environmental influence—and then use Pearl's machinery to extract causal conclusions from data that appear to support that theory. The diagram is presented as a neutral representation of the causal structure. But it is nothing of the kind. It is a *reification* of a particular theoretical stance. Worse still, the formalism creates an illusion that disagreement about the diagram is somehow a technical failure, a failure to be sufficiently explicit or rigorous. But often it is not. Often it reflects genuine disagreement about what intelligence is and how it comes to be. To pretend that such disagreements can be resolved by drawing a clearer diagram is to misunderstand the nature of the disagreement. Consider the question: does early childhood education cause increases in intelligence? The answer depends almost entirely on what diagram you assume. If you assume that intelligence is a fixed trait largely determined by genetics, and that early education affects only the expression of that trait through increased motivation or test familiarity, then the answer is: education causes apparent increases in test scores, but not genuine increases in intelligence. If you assume that intelligence is a developable capacity, substantially shaped by early experience, then the answer is: yes, education causes increases in intelligence. These are not differences that can be resolved by being more explicit about one's assumptions. They are differences in fundamental theoretical commitments. And who decides between them? Here we must be honest. In practice, it is decided by whoever has the authority to define the terms, to construct the diagram, to set the research agenda. In the contemporary university, this is often researchers trained in particular traditions, funded by particular institutions, answerable to particular audiences. There is no neutral position from which to draw the diagram. The diagram is always drawn from somewhere, by someone with interests and commitments. ## IV. THE QUESTION OF INTELLIGENCE: A CASE STUDY IN CAUSAL COMPLEXITY Let us apply these observations specifically to the question of intelligence—the very question with which we began. Intelligence, as it is ordinarily understood and measured, is a composite phenomenon. It is the outcome of numerous causal processes: genetic inheritance, certainly; but also prenatal environment, nutrition, early experience, education, motivation, cultural familiarity, test-taking skill, emotional regulation, and much else besides. These causes interact in complex ways. They are not independent. A child with greater genetic potential for learning will, if given adequate opportunity, tend to seek out more stimulating environments, which will further develop their capacities. A child who experiences early success in learning will develop greater motivation and confidence, which will facilitate further learning. The causal structure is not a simple directed acyclic graph. It is a tangle of feedback loops and reciprocal influences. Now, one can certainly construct a DAG that represents some simplified version of this structure. One can say: genetic factors here, early environment there, schooling here, and these interact as follows. But the simplification is severe. And the question of *which* simplifications to make is not a technical question. It is a question about what one takes to be the essential features of the phenomenon. Consider the question: does intelligence *cause* educational attainment, or does educational attainment *cause* intelligence? The conventional answer is: both. Higher intelligence facilitates educational attainment, and education develops intelligence. The causal relationship is reciprocal. But a DAG cannot represent reciprocal causation—it must be acyclic. So one must choose: either break the loop by treating one variable as prior in time, or one must decline to use the DAG formalism altogether. The researcher who chooses the first option—treating intelligence as prior, as a cause of educational outcomes—makes a substantive commitment. They are saying: intelligence is a relatively stable trait that individuals bring to the educational process, and education builds upon that foundation. But this is not a neutral technical choice. It embodies a particular view of intelligence and its role in human development. A researcher who instead treats educational experience as prior—arguing that the stimulation and instruction provided by schooling shapes the development of intelligence—makes a different substantive commitment. Both researchers can be entirely explicit about their diagram. But the diagrams embody different theories. And here is the deeper point: the data will not tell us which diagram is correct. We can observe correlations between intelligence and educational attainment. We can use Pearl's machinery to extract causal conclusions under the assumption that one diagram or the other is true. But the diagram itself is not validated by the data. It is imposed upon the data. ## V. THE DIMENSION OF THE CAUSAL: WHAT IT MEANS TO KNOW A CAUSE We must now attempt to be more philosophical. What does it mean to *know* that something is a cause? What would constitute evidence for a causal claim? There are several possible answers, each with its own implications. **First, the experimental answer**: One knows that X causes Y if one can manipulate X and observe changes in Y, while holding other variables constant. This is the gold standard of causal inference in experimental science. It is the logic behind the randomized controlled trial. And it has genuine force. If one randomly assigns some children to an intensive early education program and others to a control condition, and one observes differences in later intelligence test scores, one has reason to believe that the education caused the difference. Yet even this is not as simple as it appears. The experimental manipulation is always a *specific* intervention at a *specific* time in a *specific* population under *specific* conditions. One can conclude that this intervention, in these circumstances, produced these effects. But one cannot necessarily conclude that the underlying causal mechanism is understood, or that the effect would generalize to other populations, times, or conditions. And one certainly cannot conclude that one understands *all* the ways in which education might affect intelligence, or what the long-term effects might be, or how the effects might interact with other causal forces. **Second, the mechanistic answer**: One knows that X causes Y if one understands the mechanism by which X produces Y. This is perhaps the deepest form of causal knowledge. It is not enough to observe that the intervention works; one must understand *how* it works. In the case of education and intelligence, this might mean understanding the neurological, cognitive, or developmental processes through which educational experience shapes the growth and organization of the brain. It might mean understanding the psychological mechanisms through which motivation and confidence affect learning. It might mean understanding the social processes through which educational credentials shape opportunity and self-conception. This kind of knowledge is difficult to acquire. It requires not just observation but interpretation, not just data but theory. And it is precisely here that the diagram becomes most problematic. A DAG can represent correlations and conditional independencies. But it cannot represent mechanisms. It cannot tell us *how* one variable causes another, only that it does (under the assumption that the diagram is correct). **Third, the counterfactual answer**: One knows that X causes Y if one can say what would have happened had X been different. This is the answer that Pearl's formalism is designed to formalize. The do-operator is meant to capture the counterfactual: do(X=x), meaning "set X to the value x, and see what happens." This is a powerful conceptual tool. It allows one to distinguish between correlation and causation. If changing X changes Y, then X causes Y (in that context, under those conditions). If changing X does not change Y, then X does not cause Y (even if they are correlated). But the counterfactual answer also depends on the diagram. The question "what would have happened had X been different?" is only answerable if one knows the causal structure. If the diagram is wrong, the answer will be wrong. And the diagram is not validated by the data. **Fourth, the pragmatic answer**: One knows that X causes Y if manipulating X is a reliable way to produce changes in Y for practical purposes. This is the answer that guides much of engineering and medicine. One need not understand the underlying mechanism perfectly; one need only know that the intervention works. This is a modest form of causal knowledge, but it is real. If one can reliably increase intelligence through educational intervention, that is valuable, regardless of whether one understands all the mechanisms involved. Now, the question is: which of these forms of causal knowledge do we actually possess regarding intelligence? On the experimental front, we have some solid evidence. Intensive early education programs do appear to increase measured intelligence, at least in the short term. Schooling does appear to increase certain cognitive capacities. But the effects are often modest, and they tend to fade over time. The long-term effects of educational intervention on intelligence are much more uncertain. On the mechanistic front, we have partial understanding at best. We know that education engages the brain, that learning involves the formation of new neural connections, that practice improves performance. But we do not fully understand how these processes relate to the underlying construct of intelligence. We do not know whether education increases the capacity for intelligent reasoning, or merely the knowledge and skills that are tested by intelligence tests. This is not a trivial distinction. On the counterfactual front, we are entirely dependent on the diagram. We can ask: what would intelligence be if education were removed? But the answer depends on what causal structure we assume. If we assume that education directly affects intelligence, the answer is one thing. If we assume that education affects only test performance, the answer is another. The data cannot decide between these diagrams. On the pragmatic front, we know that education matters. It reliably produces changes in measured intelligence, at least in some contexts. But we do not know how durable these changes are, or how they relate to the practical capacities we care about. ## VI. THE HIDDEN POLITICS OF THE DIAGRAM We must now confront the political dimension of this question, for it is inescapable. The question "what is intelligence, and what causes it?" is not merely a scientific question. It is a question with profound implications for how we organize society, how we allocate resources, how we think about human potential and human difference. If intelligence is primarily a product of innate genetic capacity, then educational intervention has limited efficacy, and social inequality in intelligence reflects fundamental differences in human potential. The policy implication is that resources should perhaps be directed toward identifying and nurturing those with the greatest genetic potential, rather than attempting to raise the intelligence of those with less potential. If intelligence is substantially shaped by environment and experience, then educational intervention has greater potential to equalize opportunity and reduce inequality. The policy implication is that resources should be invested in improving the educational environments of all children, particularly those who have been disadvantaged. These are not merely different scientific hypotheses. They are different visions of human nature and human society. And the diagram one draws—the causal structure one assumes—will tend to align with one's prior commitments on these political questions. This is not to say that empirical evidence is irrelevant. It is not. Evidence does constrain what