Chapter 17: Causation & Mechanisms

David Hume’s analysis of causation in the Treatise of Human Nature (1739) and the Enquiry Concerning Human Understanding (1748) remains the starting point for all philosophical discussion of the topic. Hume asks a deceptively simple question: what do we actually observe when we observe one event “causing” another?

Consider the paradigm case: one billiard ball strikes another, and the second ball moves. What do we observe? We observe the first ball moving, the collision, and then the second ball moving. We observe contiguity (the events are spatiotemporally adjacent), temporal priority (the cause precedes the effect), and constant conjunction (events of the first type are regularly followed by events of the second type). But do we observe a necessary connection between the two events? Hume says no:

“When we look about us towards external objects, and consider the operation of causes, we are never able, in a single instance, to discover any power or necessary connexion; any quality, which binds the effect to the cause, and renders the one an infallible consequence of the other. We only find, that the one does actually, in fact, follow the other.”

The idea of “necessary connection” — the idea that the cause makes the effect happen, that the effect must follow — is not given in experience. It is, Hume argues, a projection of our psychological habit of expecting the second event upon observing the first. After repeated experience of one type of event following another, the mind forms a habit of anticipation, and this habitual transition is mistaken for an objective connection in the world.

Hume’s analysis inspired the regularity theory of causation: to say that A caused B is just to say that events of type A are regularly followed by events of type B, with appropriate contiguity and temporal ordering. There is nothing more to causation than constant conjunction.

John Stuart Mill refined the regularity theory by introducing the idea of a cause as the sufficient condition (or set of conditions) for the effect, within a framework of necessary and sufficient conditions. The cause of an event is the “antecedent, or the concurrence of antecedents, on which it is invariably and unconditionally consequent.”

J.L. Mackie’s influential INUS analysis (1965) provides a more sophisticated regularity account. A cause is an Insufficient but Necessary part of an Unnecessary but Sufficient condition. For example, the short circuit caused the fire: the short circuit was not by itself sufficient (oxygen and flammable materials were also needed), but it was a necessary part of a set of conditions that was jointly sufficient for the fire (even though the fire could also have been caused by other sets of conditions).

Regularity theories face well-known problems. They cannot distinguish genuine causes from mere correlations (the rooster’s crow regularly precedes dawn but does not cause it). They cannot handle singular causation (a unique event can be a cause even without a regularity to appeal to). And they struggle with preemption and overdetermination (cases where multiple causes compete or overlap).

David Lewis’s counterfactual theory of causation, presented in his 1973 paper “Causation,” offers an elegant alternative to the regularity approach. The core idea is:

“C causes E if and only if, had C not occurred, E would not have occurred.”

More precisely, Lewis defines causal dependence: E causally depends on C if and only if, in the closest possible world where C does not occur, E does not occur either. Causation is then the ancestral of causal dependence: C causes E if there is a chain of events C, D₁, D₂, ..., E such that each depends on the preceding one.

Lewis analyzes counterfactuals using his theory of possible worlds. The counterfactual “if C had not occurred, E would not have occurred” is true if, among the possible worlds most similar to the actual world where C does not occur, E does not occur either. This requires a similarity metric on possible worlds — a way of ranking worlds by their closeness to actuality.

The counterfactual theory has notable advantages:

But the theory faces serious difficulties. Preemption cases are notoriously problematic. Suppose Billy and Suzy both throw rocks at a window. Suzy’s rock arrives first and breaks the window. Billy’s rock would have broken it if Suzy’s had missed. Intuitively, Suzy’s throw caused the breakage. But the counterfactual test fails: had Suzy not thrown, the window would still have broken (because of Billy’s throw).

Lewis spent much of his career developing increasingly sophisticated versions of the counterfactual theory to handle such cases, including his influential 2000 paper introducing the notion of “influence” as a more nuanced successor to simple counterfactual dependence. Whether these refinements succeed remains contested.

Wesley Salmon developed a causal/mechanical model that grounds explanation in physical processes rather than logical relations or counterfactuals. Salmon’s key concepts are causal processes and causal interactions.

A causal process is a spatiotemporally continuous process that transmits a “mark” — a local modification that persists without further intervention. A moving ball is a causal process: if you scuff it, the scuff mark persists as the ball continues moving. A shadow, by contrast, is a pseudo-process: if you modify the shape of a shadow at one point, the modification does not persist along the shadow’s trajectory.

A causal interaction occurs when two causal processes intersect and both are modified in ways that persist beyond the point of intersection. When two billiard balls collide, both are modified (their velocities change, they may acquire scuff marks), and these modifications persist after the collision.

On Salmon’s view, to explain a phenomenon is to trace the causal processes and interactions that led to it. This account solves several problems that plague the D-N model:

Phil Dowe later refined Salmon’s account by replacing the mark-transmission criterion with the conserved quantity criterion: a causal process is a worldline of an object that possesses a conserved quantity (energy, momentum, charge), and a causal interaction is an intersection of worldlines involving exchange of a conserved quantity. This avoids certain difficulties with the original mark-transmission criterion.

James Woodward’s Making Things Happen (2003) develops an interventionist or manipulationist account of causation that has become one of the most influential contemporary approaches. The central idea is disarmingly simple:

“X causes Y if and only if there is a possible intervention on X that would change Y (or the probability of Y).”

An intervention on X with respect to Y is an idealized experimental manipulation that changes X in a way that is independent of any other cause of Y. Think of it as an ideal experiment: you intervene to change X, holding everything else fixed, and observe whether Y changes. If it does, X causes Y.

Woodward’s account has several distinctive features:

Critics ask whether the account is truly non-circular, and whether it can handle cases where interventions are physically impossible (can we “intervene” on the mass of a star?). Woodward responds that the interventions need only be logically possible, not physically performable.

Beginning around 2000, a group of philosophers — Peter Machamer, Lindley Darden, and Carl Craver (MDC) — developed the new mechanistic philosophy, which has profoundly influenced the philosophy of biology, neuroscience, and medicine. Their landmark paper “Thinking about Mechanisms” (2000) defined a mechanism as:

“Mechanisms are entities and activities organized such that they are productive of regular changes from start or set-up to finish or termination conditions.”

The key elements are:

On the mechanistic account, to explain a phenomenon is to describe the mechanism responsible for it. Explaining how a cell divides involves describing the entities (chromosomes, spindle fibers, centrioles) and activities (DNA replication, chromosome condensation, spindle attachment, separation) that constitute the mechanism of cell division.

The mechanistic approach resonates powerfully with actual scientific practice, especially in the life sciences. Molecular biologists, neuroscientists, and biomedical researchers routinely describe mechanisms as their primary explanatory activity. The mechanistic framework also illuminates how explanation works at multiple levels: a mechanism at one level (protein folding) is itself explained by mechanisms at a lower level (chemical bonding, hydrophobic interactions), while being part of a higher-level mechanism (gene expression, cellular signaling).

Craver has developed a detailed account of constitutive explanation in neuroscience: explaining a capacity of a system (e.g., spatial memory in the hippocampus) by describing the organized components and activities that constitute the mechanism for that capacity. This involves “looking down” (decomposing the system) and “looking around” (contextualizing the mechanism within the broader system).

Medicine provides a particularly important testing ground for theories of causation. Establishing that a substance, behavior, or pathogen causes a disease is central to medical practice, public health policy, and legal liability. But medical causation involves distinctive philosophical challenges.

The Bradford Hill Criteria

Austin Bradford Hill’s 1965 paper “The Environment and Disease: Association or Causation?” proposed nine criteria for evaluating whether an observed association between a factor and a disease is causal:

Hill himself was careful to note that these are not necessary or sufficient conditions for causation but rather considerations to weigh in making a judgment. The criteria blend statistical evidence with mechanistic considerations — reflecting the fact that medical causal inference typically requires both epidemiological data and biological understanding.

Randomized Controlled Trials

The randomized controlled trial (RCT) is often called the “gold standard” for establishing causal claims in medicine. By randomly assigning subjects to treatment and control groups, the RCT eliminates confounding variables — ensuring that any observed difference in outcomes is attributable to the treatment rather than to pre-existing differences between groups. The philosophical foundation of the RCT is essentially Woodward’s interventionist account: the randomized assignment is an intervention on the treatment variable, and the observed effect establishes a causal relationship. Yet even RCTs face philosophical challenges: the problem of external validity (do results from one population generalize to others?), the problem of mechanisms (an RCT shows that a treatment works but not necessarily how), and ethical constraints on what can be tested experimentally.

Given the diversity of accounts and the difficulty each faces, some philosophers have embraced causal pluralism — the view that there is no single, unified concept of causation but rather a family of related causal concepts. Different causal concepts may be appropriate in different contexts.

Peter Godfrey-Smith has distinguished between “production” causation (where a cause brings about an effect through a physical process) and “difference-making” causation (where a cause makes a difference to whether the effect occurs). These may be genuinely different causal relations, each captured by different philosophical accounts: process theories capture production causation, while counterfactual and interventionist theories capture difference-making causation.

Ned Hall has similarly argued that there are two fundamentally different concepts of causation: dependence (captured by counterfactual theories) and production (captured by process theories). These concepts come apart in cases of preemption, overdetermination, and prevention. Rather than seeking a single account that handles all cases, Hall suggests we recognize that we are dealing with two distinct relations that usually coincide but sometimes diverge.

Whether causal pluralism represents genuine philosophical progress or a counsel of despair remains debated. Monists argue that a unified account of causation should be possible and that the apparent plurality results from our incomplete understanding. Pluralists respond that the diversity of causal concepts reflects a genuine diversity in the world — different types of causal relations exist, and no single analysis can capture them all.

The relationship between causation and explanation varies across the sciences. In physics, where fundamental laws are deterministic (or probabilistic in a well-defined sense), causal explanation may reduce to derivation from laws. In biology, psychology, and the social sciences, causal explanation typically involves citing mechanisms, dispositions, or functional roles rather than strict laws.

Causation in biology raises distinctive issues. Biological causation often involves multiple levels of organization: genes cause traits, but so do environmental factors, developmental processes, and selective pressures. The “nature vs. nurture” debate illustrates the difficulty of assigning causal responsibility when multiple factors interact. The interventionist framework has proved particularly useful in biology, as it provides a clear way to formulate causal questions at different levels: “would intervening on this gene change the phenotype?” is a well-defined causal question that can be tested experimentally.

Causation in the social sciences faces the additional challenge of reflexivity: social actors respond to knowledge of the causal factors that affect them, potentially changing the causal relationships themselves. If people learn that smoking causes cancer, they may quit smoking, changing the statistics on which the causal claim was based. This “looping effect” (Ian Hacking) means that causal relationships in the social sciences may be less stable and harder to identify than causal relationships in the natural sciences.

Many causal relationships are probabilistic rather than deterministic. Smoking causes lung cancer, but not every smoker develops lung cancer. The probability of cancer is higher for smokers than for non-smokers, but the relationship is not deterministic. How should we understand probabilistic causation?

The simplest probabilistic theory says: C causes E if P(E | C) > P(E | ¬C) — if C raises the probability of E. But this faces the problem of Simpson’s paradox: a correlation that holds in the whole population can reverse in every subpopulation. A treatment can appear to raise the probability of recovery in the overall population while lowering it in every subgroup (because of confounding). To handle this, we need to condition on the right set of background variables.

Woodward’s interventionist account handles probabilistic causation naturally: C causes E if there is an intervention on C that changes the probability of E. This avoids Simpson’s paradox because interventions, by definition, break the correlations with confounding variables. It also connects naturally to the logic of randomized controlled trials, which are precisely designed to implement interventions.