The Nature of Causality

IDENTITY, APPEARANCE, AND INTERVENTION

Aristotle, Rand, Kant, and Pearl on the Nature of Causality

A Scholarly Essay

September 2026

Abstract.  This essay compares four accounts of causality: Aristotle’s fourfold scheme of explanation, Ayn Rand’s thesis that causality is the law of identity applied to action, Immanuel Kant’s treatment of causality as an a priori category of appearances, and Judea Pearl’s structural account of cause and effect. Rand’s formulation is Aristotelian in spirit: entities act as they do because of what they are. Kant, by contrast, severs causal necessity from knowledge of things as they are in themselves and relocates it in the conditions of possible experience. Pearl’s The Book of Why and the technical work that precedes it argue that causality cannot be reduced to statistical association. His ladder of causation, structural causal models, graphical criteria, and do-calculus supply a precise language in which interventions and counterfactuals can be distinguished from mere correlation—and in which the conditions for identifying causal relations from data can be stated and, when they fail, shown to fail.

I.  The Question

Causality is at once a metaphysical claim about how the world hangs together and a methodological problem about how we may come to know that hanging-together. The four thinkers treated here divide along both axes. Aristotle and Ayn Rand treat cause as a feature of entities and their natures. Kant treats the principle of cause and effect as a condition under which objects of experience can be cognized as objectively successive. Judea Pearl treats causality as a relation among mechanisms that can be represented, queried, and—under stated assumptions—identified, but that is not recoverable from probability distributions alone.

Rand’s compact formula is the hinge of the comparison. “The law of causality,” she writes in Galt’s speech, “is the law of identity applied to action. All actions are caused by entities. The nature of an action is caused and determined by the nature of the entities that act; a thing cannot act in contradiction to its nature.”[1] The claim is that what a thing does follows from what it is. That is a recognizably Aristotelian thought. It is not a Kantian one. Kant’s Second Analogy does secure a principle of causality for the objects of possible experience, but it does so without granting cognition of the identities of things in themselves.[2] Pearl’s project is neither metaphysics in Rand’s sense nor transcendental philosophy in Kant’s. It is an attempt to say, with mathematical explicitness, what a causal claim means and when data plus assumptions can support it. The result, set out for a general audience in The Book of Why, is that “correlation is not causation” is not a slogan but a theorem: the associational layer of a probability distribution underdetermines the interventional and counterfactual layers.[3]

The essay proceeds in two movements. The first compares Aristotle, Rand, and Kant, and locates Pearl against that background. The second describes Pearl’s principal innovations and shows how they explain the concept of causality rather than merely offering techniques for estimating effects.

II.  Aristotle: Causes as Answers to Why

Aristotle’s account in Physics II.3 and Metaphysics V.2 does not isolate a single relation of “causing” in the modern, efficient-cause sense. It distinguishes four aitia—four modes of answering “why?”—that together make a thing intelligible.[4] The material cause is that out of which a thing comes to be (the bronze of a statue). The formal cause is the essence or structure that makes the thing the kind of thing it is (the arrangement by which this bronze is a statue of a given subject). The efficient cause is the source of the change (the art of bronze-casting, or the artisan insofar as he possesses that art). The final cause is that for the sake of which the change occurs (the completed statue, or, in nature, the mature form of the organism).

Two features of this scheme matter for later comparison. First, causal knowledge is knowledge of what a thing is. “We think we have knowledge of a thing only when we have grasped its cause.”[5] To know the cause is not merely to record that one event followed another; it is to grasp the principles that make the sequence the sequence of this kind of thing. Second, in natural generation the formal, final, and efficient causes often coincide. A human being generates a human being: the parent is the efficient cause by virtue of already possessing the form that the offspring will realize, and that form is also the end of the process.[6] Identity—what it is to be a human being—is not an extra metaphysical ornament hung on a sequence of events. It is what explains why the sequence has the shape it has.

Aristotle’s efficient cause is therefore not a bare prior event. It is a source of change whose efficacy is specified by the nature of the agent and the patient. Hypothetical necessity binds matter to form: if there is to be a house, there must be bricks and beams arranged thus. Regularity in nature is not a habit of our expectations; it is the expression of forms that operate for the most part toward characteristic ends.[7] Later philosophy would strip this scheme down to efficient causation alone. That contraction is the background against which both Humean skepticism and Rand’s return to identity must be read.[8]

III.  Rand: Causality as Identity in Action

Rand presents her view as a completion of Aristotle’s formula “A is A.” Existence is identity; consciousness is identification. Causality is not a second law added to identity but identity as it governs action:

The law of identity does not permit you to have your cake and eat it, too. The law of causality does not permit you to eat your cake before you have it. — Atlas Shrugged

The same speech continues: all actions are actions of entities; the nature of an action is determined by the nature of the entities that act; a thing cannot act in contradiction to its nature.[9] Leonard Peikoff later classified the principle as a corollary of the axiom of identity rather than as an independent axiom: once it is granted that every existent is something definite, and that action is always action of an existent, it follows that an entity can act only as its identity permits.[10] In any given set of circumstances, the action open to an entity is the action expressive of what it is.

Several consequences follow. First, there are no causeless events and no action by a “zero.” An action not caused by an entity would be an action caused by nothing, which Rand treats as a contradiction, not as a conceivable limiting case.[11] Second, chance, in the sense of an occurrence unconstrained by identity, is excluded from metaphysics. Statistical unpredictability may mark our ignorance or the complexity of initial conditions; it does not mark a hole in being. Third, volition is not an exemption from causality. Choice is “not chance”; it is “a type of causation”—the type proper to a conceptual consciousness that can select among alternatives by a process of identification.[12] The human being does not float free of identity; the human being acts as the kind of entity that must choose in order to live.

Rand’s target is therefore double. Against the mystic, she denies miracles, the unearned, and the reign of the zero. Against the Humean empiricist, she denies that causality is a habit induced by the repeated pairing of impressions. Against the Kantian, as the next section argues, she denies that we are barred from knowing the identities that make causal claims true. The world is not a sequence of appearances awaiting a category to stitch it together. It is a world of entities, and entities act.

IV.  Kant: Causality as a Condition of Appearances

It is important not to collapse Kant into Hume. Hume argued that the idea of necessary connection cannot be copied from any impression of the objects themselves; we find only constant conjunction and a determination of the mind to pass from one object to its usual attendant.[13] Kant accepted the negative claim that necessity is not given in the mere empirical sequence, and then inverted the moral. If objective experience of events is possible at all, the understanding must already supply the rule that determines the order of states as necessary. Causality is a pure category, schematized as necessary succession in time.

The Second Analogy states: all alterations occur in accordance with the law of the connection of cause and effect.[14] Kant’s argument distinguishes the subjective sequence of perceptions from the objective sequence of states of an object. I may look at the parts of a house in any order; I cannot perceive a ship moving downstream as first downstream and then upstream. The irreversibility of the latter sequence is not a fact I first notice and then generalize. It is thinkable as an event only because I place the manifold under a rule according to which the later state must follow the earlier. The principle of causality is thus, for Kant, a principle of possible experience. Without it there are no objects of empirical cognition in temporal relation, only a play of representations.

The price of this rescue is the restriction of the categories to appearances. Space and time are forms of our intuition, not properties of things in themselves. The categories, including cause and effect, have legitimate application only within possible experience. Of things as they are in themselves we can think, but we cannot cognize. Whatever necessity we find in the causal order of nature is necessity for us: necessity as a condition of the unity of experience, not necessity read off the inner constitution of mind-independent entities.

The contrast between Rand and Kant is therefore almost right, and must be stated with care. Kant does not hold the crude thesis that correlation is causation. He holds the opposite of Hume’s regularity theory: causal connection is necessary, universal, and a priori. What he does hold, and what Rand rejects, is that we cannot know the true identities of the things that appear to us, and that the causal order we report is therefore the order of appearances as synthesized by the understanding.[15] If “correlation is causation” is used as a compressed description of that position, it names the consequence, not the official doctrine. The consequence is that the investigator is confined to sequences as they show up in experience, and cannot treat causal necessity as a disclosure of what the entities themselves are. From an Aristotelian or Objectivist standpoint, that is already to have surrendered the thing that makes causality intelligible.

V.  A Brief Comparison

The four positions can be aligned on three questions. What is a cause? What is the relation between cause and the identity of things? How do we know causal relations?

Question	Aristotle / Rand	Kant	Pearl
What a cause is	A principle of an entity: form, matter, agent, and end (Aristotle); the identity of an entity as it acts (Rand).	A category of the understanding that determines objective succession among appearances.	A relation among mechanisms, expressed by structural assignments and by the effects of interventions.
Cause and identity	Inseparable. What a thing does follows from what it is. A thing cannot act contrary to its nature.	Identities of things in themselves are not objects of cognition. Causal necessity applies to appearances.	Identity is encoded as functional dependence: each variable “listens” to its parents. No action without a mechanism.
How we know	By grasping natures and forms (Aristotle); by identifying existents and the actions their identities permit (Rand).	The principle is a priori. Particular causal laws are empirical, but they are laws of appearances.	By combining a causal model (assumptions) with data. Association alone never suffices for intervention or counterfactual queries.
Status of statistics	Secondary. Regularity is explained by natures; it does not constitute causality.	Empirical sequences supply the matter; the form of necessity is contributed by the category.	Statistics occupies rung one of the ladder. Causal claims live on rungs two and three.

The alignment is not a forced harmony. Aristotle’s final causes have no counterpart in Pearl’s graphs, and Rand’s axiomatic style is remote from Pearl’s willingness to treat causal structure as an assumption to be drawn, tested, and revised. Kant’s transcendental necessity is not Pearl’s interventional invariance. Even so, a family resemblance is visible. Aristotle, Rand, and Pearl all refuse to identify causality with the observation of sequences. All three treat some specification of how things work—form, nature, mechanism—as prior to the pattern of association. Kant is the outlier, not because he denies causal necessity, but because he relocates that necessity from the identities of things to the architecture of experience.

VI.  Pearl’s Diagnosis: Causality Is Not Statistics

Pearl’s polemic in The Book of Why is directed at a century of statistical culture in which causal vocabulary was treated as scientifically disreputable. Karl Pearson had hoped to replace causation with correlation. Later textbook wisdom reduced the issue to a warning—“correlation is not causation”—without supplying a language in which the difference could be stated. The result, Pearl argues, was that the most important questions in science and policy (What happens if we do this? What would have happened had we done otherwise?) were either smuggled in under associational formulas or dismissed as metaphysical.[16]

The diagnosis is exact. A joint probability distribution P(x, y, z, …) records how variables move together. It does not record what would happen to Y if we set X to a value it did not take, nor what Y would have been in a unit that in fact took another value of X. Those are different queries. They require a different object: a model of the mechanisms that assign values to variables, together with an operation that represents intervening on a mechanism rather than conditioning on an observation. Without that object, the slogan “correlation is not causation” is an unkeepable promise. With it, the slogan becomes a theorem about the underdetermination of higher rungs by lower ones.

Pearl’s contribution is therefore conceptual before it is technical. He insists that causal claims have a semantics distinct from the semantics of conditional probability, and that the right response to Humean and operationalist suspicion is not silence but a calculus.[17]

VII.  The Innovations

1.  The ladder of causation

Pearl organizes causal competence into three rungs, corresponding to seeing, doing, and imagining.[18]

Rung 1: Association. Questions of the form “How does seeing X change my belief about Y?” are answered by conditional probabilities, P(Y | X). Machine-learning systems that detect patterns, estimate regressions, or compute correlations live here. Nothing at this rung distinguishes a common cause from a direct effect, or a chain from a collider.

Rung 2: Intervention. Questions of the form “What if I do X?” are answered by the interventional distribution P(Y | do(X = x)). The do-operator represents an action that sets X to x by a mechanism external to the system, not by passively observing that X happened to be x. Policy, experiment, and most uses of the word “effect” belong here.

Rung 3: Counterfactuals. Questions of the form “What if I had done otherwise?” concern a particular unit or world already observed. They ask for P(Yₓ = y | X = x′, Y = y′): the probability that Y would have been y had X been set to x, given that in fact X was x′ and Y was y′. Attribution, regret, credit, and many questions about harm and responsibility live here. Rung 3 is strictly stronger than rung 2: an answer to every interventional query can be recovered from a sufficiently rich counterfactual model, but not conversely.

The ladder is not a pedagogical ornament. It is a hierarchy of information. Data at rung 1 do not, in general, determine answers at rung 2; data and interventions at rung 2 do not, in general, determine answers at rung 3. Any method that treats “causal inference” as a sophisticated form of curve-fitting has already mistaken the rung.

2.  Structural causal models

The semantics of the ladder is given by structural causal models (SCMs).[19] An SCM specifies, for each endogenous variable Vᵢ, an assignment

Vᵢ := fᵢ(PAᵢ, Uᵢ)

where PAᵢ are the parents of Vᵢ—the variables that enter its mechanism—and Uᵢ is an exogenous disturbance standing for omitted factors. The collection of assignments is the model of how the world works. It is not a regression. A regression coefficient describes a fact about a joint distribution; a structural assignment describes a mechanism that would continue to operate if other mechanisms were altered.

Two features of this representation do philosophical work. First, causality is local and modular. Each assignment is a little machine. To intervene on X is to replace the assignment for X with the constant X := x, leaving the other assignments intact. That replacement is “graph surgery”: the arrows into X are deleted because X no longer listens to its former causes. Second, counterfactuals receive a three-step evaluation. Abduction updates the exogenous terms U so that they are consistent with what was observed. Action replaces the relevant assignment. Prediction reads off the new value of Y. The procedure is not a metaphor. It is the definition of the counterfactual in the model.

3.  Causal diagrams and d-separation

Directed acyclic graphs translate an SCM into a picture. Nodes are variables; arrows are the “listens-to” relation. The graph is not a decoration of the statistics. It is a compact statement of which mechanisms exist and which independence relations those mechanisms imply.

The graphical vocabulary distinguishes three elementary configurations. In a chain, X → M → Y, association flows along the path unless one conditions on the mediator M. In a fork, X ← C → Y, C is a common cause; conditioning on C blocks the spurious association. In a collider, X → K ← Y, association does not flow through K unless one conditions on K or a descendant of K, in which case a spurious association is opened. These rules, generalized as d-separation, tell the investigator when a path transmits association and when it does not. They also explain a family of paradoxes—Simpson’s paradox, Berkson’s paradox, collider bias in selection—that are unintelligible if one treats every correlation as a candidate cause.

4.  Identification: back-door, front-door, and do-calculus

Once a query is stated as P(Y | do(X)), the question is whether that quantity is a function of the observational distribution P under the assumptions encoded in the graph. Pearl’s criteria answer that question without requiring a fully specified parametric model.

The back-door criterion. A set Z of covariates is admissible for the effect of X on Y if no member of Z is a descendant of X and Z blocks every path from X to Y that begins with an arrow into X.[20] Intuitively, Z intercepts every non-causal route by which X and Y could be associated. When the criterion holds, the interventional distribution is given by the adjustment formula

P(Y = y | do(X = x))  =  Σ_z  P(Y = y | X = x, Z = z) P(Z = z).

Conditioning here is not a general hygienic practice of “controlling for everything.” Controlling for a mediator blocks a causal path. Controlling for a collider opens a spurious one. The graph tells the investigator which variables must be adjusted and which must not.

The front-door criterion. When an unobserved confounder U opens a back-door path from X to Y, adjustment on observed covariates may be impossible. If, however, there is a mediator Z that intercepts all directed paths from X to Y, is itself unconfounded with X, and is such that X blocks all back-door paths from Z to Y, the effect is still identified:[21]

P(y | do(x))  =  Σ_z P(z | x)  Σ_{x′} P(y | x′, z) P(x′).

The formula is surprising on associational grounds. It uses an observed chain X → Z → Y to recover an effect even though X and Y share a hidden common cause. Pearl’s smoking–tar–cancer illustration in The Book of Why is designed to make the surprise vivid: one need not measure the unobserved confounder if the mechanism can be decomposed into identifiable pieces.

Do-calculus. Back-door and front-door are special cases of a more general symbolic engine. The do-calculus consists of three rules that license the exchange of interventions and observations, or the deletion of an intervention, when corresponding separation conditions hold in a mutilated graph.[22] The calculus is complete for nonparametric identification: if a rung-2 query cannot be reduced by these rules to an expression in ordinary conditional probabilities, then no nonparametric method can identify it from the observational distribution given only the assumptions in the graph. Failure is informative. It tells the investigator that more assumptions, more measurements, or an experiment are required.

5.  Simpson’s paradox and the primacy of the story

Simpson’s paradox—the reversal of an association upon partitioning on a third variable—was long treated as a curiosity of contingency tables. Pearl’s resolution is that the paradox is not in the numbers. The same table authorizes opposite actions depending on whether the third variable is a confounder or a mediator.[23] If the third variable is a common cause of treatment and outcome, one must stratify. If it lies on the causal path, stratification disables the mechanism by which the treatment works. Data do not choose between these stories. A causal diagram does. Lindley and Novick had already suspected that the correct decision “depends on the causal story, not on the data.” Pearl supplies the language in which that suspicion becomes a method.

6.  Mediation, explanation, and the mini-Turing test

To explain an effect is often to decompose it. Direct and indirect effects, natural and controlled, are defined in the SCM by nested interventions: one holds a mediator at the value it would have taken under one treatment while setting the treatment itself to another.[24] The resulting formulas make precise a distinction that ordinary regression cannot: the distinction between a variable that transmits a cause and a variable that merely accompanies it. Explanation, in this sense, is not a narrative overlay on an estimate. It is a further causal query, sitting on rungs 2 and 3.

Pearl’s informal test of whether a system understands causality is whether it can take a story, encode it as a diagram, and then answer interventional and counterfactual questions about that story—the “mini-Turing test.”[25] A system that predicts Y from X with arbitrary accuracy may still fail the test, because prediction is rung 1. Understanding, on this criterion, is the capacity to represent mechanisms and to compute the consequences of changing them.

VIII.  How the Innovations Explain the Concept of Causality

It remains to say how this apparatus is an account of causality, not only a toolkit.

First, Pearl replaces a regularity theory with a mechanism theory. A cause is not an event constantly conjoined with another. A cause is a variable that appears on the right-hand side of a structural assignment for another variable: a factor to which the effect “listens.” That is closer to Aristotle’s efficient cause, specified by the nature of the agent, than it is to Hume’s constant conjunction. It is also closer to Rand’s “the nature of an action is determined by the nature of the entities that act.” The structural function fᵢ is a mathematical stand-in for that nature, restricted to the coordinates the modeler has chosen to name.

Second, the do-operator reconstructs the difference between seeing and doing that regularity theories flatten. Conditioning on X = x selects a subpopulation in which X took the value x, carrying with it whatever background factors produced that value. Intervening on X = x forces the value and thereby cuts the influence of those background factors. The difference is the difference between a world observed and a world altered. Causal claims are claims about altered worlds. That is why they cannot be read off correlations.

Third, identification criteria explain why some causal questions are answerable from observation and others are not. The explanation is not psychological (“we feel a necessity”) and not merely experimental (“only randomized trials speak”). It is structural. Given a graph, some queries reduce to associational quantities and some do not. Completeness results make the boundary sharp. Causality, on this view, is not an occult extra property that floats free of our models. It is a feature of a system of mechanisms, and the epistemic problem is to say which features of that system are recoverable from which data.

Fourth, counterfactuals cease to be a vague gesture at possible worlds and become a derived operation on an SCM. “What would have happened” is answered by holding fixed the realized exogenous terms of a unit and changing a mechanism. That operation respects Rand’s insistence that action is always action of an entity: the unit, with its particular U, remains the same unit. It also respects Aristotle’s interest in the particular as well as the kind. Population effects live on rung 2; the fate of this patient, this policy, this historical episode lives on rung 3.

Fifth, the ladder explains the persistent confusion of correlation with causation as a confusion of rungs. The confusion is not a failure of statistical hygiene. It is a failure to notice that P(Y | X) and P(Y | do(X)) are different mathematical objects, and that no amount of conditioning will turn the one into the other unless the graph licenses the conversion. Kant’s official doctrine, as noted, is not this confusion. But a research culture that treats causal language as an anthropomorphic gloss on sequences of appearances will, in practice, behave as if the first rung were the only rung. Pearl’s calculus is a discipline against that collapse.

IX.  Convergences and Distances

Pearl does not restore Aristotelian teleology. Nothing in an SCM requires that mechanisms aim at an end, and the formalism is deliberately silent on whether the functions fᵢ express essences or merely stable regularities of assignment. Pearl also does not restore Rand’s axiomatic epistemology. Causal structure enters as an assumption—elicited from background knowledge, drawn as a graph, tested against the independencies the graph implies, and revised when those independencies fail. The method is fallibilist about particular causal claims even as it is uncompromising about the distinction between those claims and statistical associations.

Those distances matter. A Randian may object that Pearl still treats “identity” as a modeling choice rather than as an axiom disclosed in perception. A Kantian may object that Pearl’s mechanisms are themselves objects of possible experience, organized by the same categories whose reach Kant limited. An Aristotelian may object that efficient-cause graphs omit form and end, and therefore omit the greater part of explanation. Each objection names a real limit. None of them cancels the result that is internal to Pearl’s project: once one asks what would happen if a mechanism were altered, one has left statistics, and one needs a language that can say so.

Nonetheless, there is a deep convergence. Aristotle and Rand hold that causality is unintelligible apart from what things are. Pearl holds that causality is unintelligible apart from how variables are assigned their values—that is, apart from the mechanisms that constitute, within a model, what those variables are doing. Kant holds that what things are in themselves is not an object of theoretical cognition, and that the necessity we call causal is a necessity of appearances. Pearl’s methods do not refute transcendental idealism. They do show that a science which wishes to speak of doing and of what would have been cannot treat the causal order as a mere report of observed sequences. In that precise sense, they side with identity-in-action against sequence-as-cause.

X.  Conclusion

Aristotle taught that to know a thing is to know why it is as it is, and that “why” has more than one register. Rand compressed the efficient-cause register into a single sentence: causality is identity applied to action. Kant, answering Hume, saved necessary connection by making it a condition of experience and, in the same stroke, denied that we know the identities of things as they are in themselves. Pearl, answering a statistical culture that had forgotten how to write the word “cause,” supplied a ladder, a class of models, a graphical language, and a calculus.

The comparison does not require that these four projects be the same project. It requires only that they be mutually illuminating. They are. If entities act according to their identities, then a science of action needs a way to represent identities as mechanisms and to distinguish altering a mechanism from watching it. If, on the other hand, one has only sequences, one has correlation, and correlation is not causation—not as a proverb, but as a fact about the hierarchy of queries a causal model can support. Pearl’s innovations are the most exact modern statement of that fact, and the most usable set of methods for living with it.

Works Cited

Aristotle. Metaphysics. Translated by W. D. Ross. In The Complete Works of Aristotle, edited by Jonathan Barnes. Princeton: Princeton University Press, 1984.

Aristotle. Physics. Translated by R. P. Hardie and R. K. Gaye. In The Complete Works of Aristotle, edited by Jonathan Barnes. Princeton: Princeton University Press, 1984.

Aristotle. Posterior Analytics. Translated by Jonathan Barnes. Oxford: Clarendon Press, 1993.

Falcon, Andrea. “Aristotle on Causality.” Stanford Encyclopedia of Philosophy. Substantive revision 7 March 2023.

Hume, David. An Enquiry Concerning Human Understanding. Edited by Tom L. Beauchamp. Oxford: Oxford University Press, 2000.

Kant, Immanuel. Critique of Pure Reason. Translated by Paul Guyer and Allen W. Wood. Cambridge: Cambridge University Press, 1998.

Pearl, Judea. “Causal Diagrams for Empirical Research.” Biometrika 82, no. 4 (1995): 669–710.

Pearl, Judea. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge: Cambridge University Press, 2009.

Pearl, Judea. “The Causal Mediation Formula—A Guide to the Assessment of Pathways and Mechanisms.” Prevention Science 13 (2012): 426–436.

Pearl, Judea. “Understanding Simpson’s Paradox.” The American Statistician 68, no. 1 (2014): 8–13.

Pearl, Judea. “The Seven Tools of Causal Inference, with Reflections on Machine Learning.” Communications of the ACM 62, no. 3 (2019): 54–60.

Pearl, Judea, and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. New York: Basic Books, 2018.

Peikoff, Leonard. Objectivism: The Philosophy of Ayn Rand. New York: Dutton, 1991.

Rand, Ayn. Atlas Shrugged. New York: Random House, 1957.

Rand, Ayn. For the New Intellectual. New York: Random House, 1961.

Rand, Ayn. Philosophy: Who Needs It. New York: Bobbs-Merrill, 1982.

Shpitser, Ilya, and Judea Pearl. “Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models.” Proceedings of the 21st AAAI Conference on Artificial Intelligence (2006).

[1]Ayn Rand, Atlas Shrugged (New York: Random House, 1957); reprinted in For the New Intellectual (New York: Random House, 1961), 151.

[2]Immanuel Kant, Critique of Pure Reason, trans. Paul Guyer and Allen W. Wood (Cambridge: Cambridge University Press, 1998), A189–211/B232–256 (Second Analogy of Experience).

[3]Judea Pearl and Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (New York: Basic Books, 2018).

[4]Aristotle, Physics II.3, 194b16–195a3; Metaphysics V.2. Translations follow W. D. Ross unless otherwise noted.

[5]Aristotle, Posterior Analytics I.2, 71b9–12; Physics II.3, 194b17–20.

[6]Aristotle, Physics II.2, 194b13; II.7, 198a24–27. On the coincidence of form, end, and mover in natural generation, see also Parts of Animals I.1.

[7]Andrea Falcon, “Aristotle on Causality,” Stanford Encyclopedia of Philosophy, subst. rev. 7 March 2023.

[8]On the historical contraction of Aristotle’s four aitia to efficient causation alone, see Eric Schliesser’s discussion of the early-modern polemic and Hume’s role in What to Do When You Encounter Funky Causes in the Historical Wild (2024).

[9]Rand, Atlas Shrugged, Galt’s speech; For the New Intellectual, 151. See also Ayn Rand Lexicon, s.v. “Causality.”

[10]Leonard Peikoff, Objectivism: The Philosophy of Ayn Rand (New York: Dutton, 1991), chap. 1. Peikoff treats causality as a corollary of the axiom of identity rather than as an independent axiom.

[11]Rand, Atlas Shrugged; For the New Intellectual, 150–151. Cf. “Existence is Identity, Consciousness is Identification.”

[12]Ayn Rand, “The Metaphysical Versus the Man-Made,” in Philosophy: Who Needs It (New York: Bobbs-Merrill, 1982). On volition as a type of causation, see the Lexicon entry “Causality,” drawing on Peikoff’s 1976 lecture series.

[13]David Hume, An Enquiry Concerning Human Understanding, ed. Tom L. Beauchamp (Oxford: Oxford University Press, 2000), secs. 4–7. Hume’s two definitions of cause combine constant conjunction with a counterfactual clause (“if the first object had not been, the second never had existed”).

[14]Kant, Critique of Pure Reason, A189/B232: “All alterations occur in accordance with the law of the connection of cause and effect.” The proof turns on the difference between a merely subjective sequence of perceptions and an objective sequence of states of an object.

[15]This is the Objectivist reconstruction, not Kant’s self-description. Kant insisted that the category of causality involves necessary connection, not mere empirical regularity; that is precisely how he took himself to answer Hume. The reconstruction is that, once things in themselves are placed beyond cognition, the necessity we ascribe cannot be read off the identities of the entities themselves.

[16]Pearl and Mackenzie, The Book of Why, chap. 1 (“The Ladder of Causation”) and Introduction (“Mind over Data”). See also Judea Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. (Cambridge: Cambridge University Press, 2009).

[17]Judea Pearl, “The Seven Tools of Causal Inference, with Reflections on Machine Learning,” Communications of the ACM 62, no. 3 (2019): 54–60.

[18]Pearl, Causality, chap. 1; Pearl and Mackenzie, The Book of Why, 27–32. Formally the three rungs query P(y|x), P(y|do(x)), and P(yx | x′, y′).

[19]A structural causal model is an ordered triple ⟨U, V, F⟩, where U is a set of exogenous variables, V a set of endogenous variables, and F a set of functions, one for each endogenous variable, determining its value from its parents and the relevant exogenous terms. See Pearl, Causality, 27–32; Pearl and Mackenzie, The Book of Why, chap. 1.

[20]Pearl, “Causal Diagrams,” 679–682 (Definition 3 and Theorem 1). The back-door criterion is graphically equivalent to Rosenbaum and Rubin’s ignorability condition, but it makes the required covariates inspectable on a diagram.

[21]Pearl, “Causal Diagrams,” 682–684 (Theorem 2). The classic illustration is a hypothesized path Smoking → Tar → Cancer, with an unobserved common cause of Smoking and Cancer.

[22]Pearl, “Causal Diagrams,” 695–698. Completeness results appear in Ilya Shpitser and Judea Pearl, “Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models,” Proceedings of the 21st AAAI Conference (2006).

[23]Pearl and Mackenzie, The Book of Why, chap. 6; Judea Pearl, “Understanding Simpson’s Paradox,” The American Statistician 68, no. 1 (2014): 8–13.

[24]Pearl, Causality, chap. 7; Judea Pearl, “The Causal Mediation Formula—A Guide to the Assessment of Pathways and Mechanisms,” Prevention Science 13 (2012): 426–436.

[25]Pearl and Mackenzie, The Book of Why, 12–18, on the “mini-Turing test”: the capacity to encode a causal story as a diagram and then answer interventional and counterfactual queries.