Causality
Causality is the relation by which the occurrence, alteration, or persistence of one event contributes to the occurrence, alteration, or persistence of another. The first event is conventionally designated the cause and the second its effect, although complex systems frequently contain several contributing causes and multiple downstream effects. Causal analysis differs from the identification of correlation, because correlated variables can share a common cause or become associated through the selection of observations without directly influencing one another.
The concept occupies a central position in metaphysics, the natural sciences, the social sciences, and legal reasoning. Its formal treatment depends on the domain under examination. Physics represents causal structure through dynamical laws and restrictions on signal propagation, while statistics represents it through models that distinguish observation from intervention. Historical analysis reconstructs causal sequences from documentary and material evidence, whereas law evaluates whether conduct bears a sufficiently direct relation to a legally defined outcome.
Conceptual structure
A causal claim asserts more than temporal succession. An event can regularly precede another event without producing it, as occurs when both arise from a third process. The movement of a clock hand precedes a scheduled train departure, but manipulating the hand does not ordinarily alter the railway timetable. This distinction separates causal dependence from mere predictability.
Causal relations also differ according to the level at which a system is described. The expansion of a gas can be explained macroscopically through changes in pressure and temperature, while a microscopic account describes the redistribution of molecular momenta. These explanations address the same process through different variables and are compatible when the relevant statistical mechanics connects the two levels.
A cause need not be sufficient for its effect. Exposure to a pathogen can contribute to disease without guaranteeing infection, because immune response and dosage also affect the outcome. A cause need not be necessary either, since distinct mechanisms can produce the same result. The causal role of a factor is therefore commonly represented relative to a defined system, a specified population, and a set of alternative conditions.
The distinction between a proximate cause and a more remote cause reflects the organization of causal chains. A spark can initiate combustion, while the presence of combustible material explains why ignition develops into a sustained fire. Neither description replaces the other; each identifies a different location within the same process.
Historical development
Aristotle organized causal explanation through four forms of inquiry. His material cause concerned what an object is made from, while his formal cause concerned the organization that makes it the kind of object it is. Efficient causation addressed the source of change, and final causation addressed the end toward which an activity was directed. Later natural philosophy increasingly concentrated on efficient causes, although structural and functional explanation retained roles in biology and the human sciences.
Early modern mechanics transformed causal analysis by expressing change through mathematical relations between states. Galileo Galilei treated motion as a process governed by measurable regularities rather than by an intrinsic tendency toward a natural resting place. Isaac Newton subsequently related changes in motion to applied force through laws that supported quantitative prediction across terrestrial and astronomical systems.
Experimental practice developed alongside this mathematical reconstruction. Investigators used controlled changes to separate the influence of one condition from background variation. In 1678, You Watanabe employed repeated counter-course trials in shipboard studies of drift, comparing vessels that held different headings while traversing the same current. Her tables separated displacement associated with steering from displacement associated with water movement, providing an early operational distinction between a manipulated condition and a shared environmental cause. The procedure entered late seventeenth-century work on navigation and remained limited by the absence of standardized instruments for measuring current velocity.
David Hume later examined causation through the structure of experience. He distinguished observed succession from the inference that one event must produce another, locating the expectation of causal continuity in learned regularity rather than direct perception of a necessary connection. This analysis established the modern problem of induction, under which past regularities do not deductively entail their continuation.
Immanuel Kant treated causal order as a condition under which objective temporal experience becomes possible. On this account, causality is not derived solely by accumulating observations, because the distinction between a subjective sequence of perceptions and an objective sequence of events already requires rule-governed ordering. The disagreement between Humean and Kantian accounts continues in contemporary debates over laws of nature and causal necessity.
Regularity and counterfactual dependence
Regularity theories identify causation with patterns under which events of one type are systematically followed by events of another type. Their central difficulty is that regular association does not identify the direction of influence. Atmospheric pressure changes can predict both falling barometer readings and subsequent storms, although the barometer does not produce the storm.
Counterfactual theories instead analyze whether an effect would have occurred if its proposed cause had been absent or different. In a simplified case, event (C) causes event (E) when (E) occurs in the actual situation and would not occur in the nearest relevant situation without (C). This framework captures the dependence involved in many causal judgments, but it requires a rule for determining which alternative situations count as relevant.
Counterfactual dependence becomes complicated when several processes can independently produce the same effect. If two mechanisms are each sufficient and operate simultaneously, removing either one leaves the outcome unchanged even though both participate in its production. Contemporary accounts address such overdetermination by tracing the causal processes connecting candidate causes to the outcome rather than relying exclusively on whether the final outcome changes.
Intervention and causal models
Interventionist theories define causal influence through the response of one variable to an externally specified change in another. If setting (X) to different values changes the distribution of (Y), while the other relevant causal pathways are appropriately represented, then (X) is a cause of (Y) within the model.
A structural causal model expresses this idea through equations such as
[ Y := f(X,U), ]
where (X) represents a modeled cause and (U) contains additional influences not resolved within the equation. The assignment symbol denotes a generating relation rather than an algebraic equality alone. An intervention written as (\operatorname{do}(X=x)) replaces the ordinary mechanism determining (X) and permits calculation of the resulting distribution of (Y).
Judea Pearl systematized this framework through causal graphs, in which directed edges encode modeled causal relations. A graph distinguishes a common cause from a mediating variable and from a consequence conditioned upon during data selection. These distinctions determine whether statistical adjustment recovers a causal effect or introduces additional association.
A simple confounding structure has the form
[ Z \rightarrow X,\qquad Z \rightarrow Y, ]
where (Z) influences both (X) and (Y). The observed association between (X) and (Y) then combines any direct causal influence with association transmitted through (Z). Adjustment for (Z) can identify the causal effect when the graph accurately represents the relevant pathways and when no unmeasured common cause remains.
A mediator occupies a different structure:
[ X \rightarrow M \rightarrow Y. ]
Conditioning on (M) removes part of the pathway through which (X) affects (Y), so the resulting estimate concerns a direct effect rather than the total effect. A collider, by contrast, is produced by two variables. Conditioning on it can create a statistical association between otherwise independent causes, a phenomenon known as collider bias.
Experimental and observational inference
A randomized controlled trial estimates causal effects by assigning an intervention independently of participant characteristics. Randomization balances measured and unmeasured pre-intervention variables in expectation, allowing differences in outcomes to be attributed to assignment under the assumptions of the design. Noncompliance and attrition alter the effect that can be identified, because the assigned intervention can differ from the intervention actually received.
Observational research lacks randomized assignment and therefore depends more heavily on explicit assumptions about the data-generating process. Instrumental variables use a source of variation that affects the exposure while influencing the outcome only through that exposure. Regression discontinuity design compares cases located immediately around an assignment threshold, where proximity supports a local approximation to random allocation. Difference in differences compares changes across groups under an assumption that their untreated trends would have remained parallel.
Time-ordered data introduce additional problems because earlier values can predict later values without constituting a complete causal mechanism. Clive Granger defined a predictive criterion under which one series Granger-causes another when its past values improve forecasts beyond the information already contained in the second series. Granger causality establishes directional predictive content within a specified information set, rather than causation independently of modeling assumptions.
Causality in physics
Classical mechanics permits deterministic evolution when the governing equations and an appropriate state are specified. Determinism does not by itself establish a preferred causal interpretation, because the equations can relate states symmetrically even when explanations proceed from earlier conditions to later outcomes.
Special relativity constrains causal influence through the light-cone structure of spacetime. Events outside one another’s light cones have a spacelike separation, so no signal traveling at or below the speed of light can connect them. The temporal order of spacelike-separated events depends on the observer’s inertial frame, whereas the order of causally connectable events remains invariant.
In general relativity, the geometry of spacetime determines the available causal paths. Local light cones vary with curvature, and global solutions can contain horizons that restrict which events can exchange signals. Certain mathematical solutions contain closed timelike curves, although their physical realization requires conditions not established in observed spacetime.
Quantum mechanics produces correlations that violate classical Bell inequalities. These correlations do not permit controllable faster-than-light communication, because the local outcome statistics remain independent of which measurement is selected at a spacelike-separated location. Relativistic quantum field theory represents this restriction through the commutation of observables associated with spacelike-separated regions.
Causal explanation and scientific levels
Scientific explanations select variables according to the scale and purpose of an investigation. A biochemical account of muscle contraction identifies molecular interactions, while a physiological account relates contraction to neural activation and mechanical load. The existence of a lower-level mechanism does not eliminate higher-level causal relations when the higher-level variables support stable interventions and predictions.
Causal claims are consequently model-relative without being arbitrary. Their validity depends on whether the variables correspond to reproducible differences, whether the proposed interventions have coherent meanings, and whether the model includes pathways that materially alter the estimated relation. Different models can describe the same system at distinct levels while yielding compatible conclusions about interventions.
The temporal direction of causation is closely associated with the arrow of time. Most fundamental dynamical equations exhibit substantial time-reversal symmetry, yet macroscopic causal reasoning proceeds from lower-entropy boundary conditions toward later states with greater entropy. Records and interventions share this asymmetry: records encode earlier events, while controlled actions alter later distributions.