Scientific explanation
A scientific explanation is an account that relates an observed phenomenon to a body of scientific knowledge in a manner that identifies why the phenomenon occurs, how it occurs, or both. Scientific explanations connect descriptions of events with empirically supported laws, causal relations, mechanisms, or theoretical structures. They differ from mere descriptions because they establish relations between the phenomenon and information that bears on its occurrence.
The analysis of scientific explanation forms a central part of the philosophy of science. It also concerns the working practices of the natural sciences and social sciences, where explanatory standards vary with the objects and methods of investigation. No single formal account represents every accepted scientific explanation. Several models instead describe distinct explanatory relations found across scientific disciplines.
Explanation and prediction
Explanation and prediction often employ similar scientific information, but they occupy different epistemic positions. A prediction derives or estimates a phenomenon before the relevant observation has been established, whereas an explanation places an observed or otherwise accepted phenomenon within an organized account of its occurrence. The mathematical structure supporting the two activities can be identical even when their scientific functions differ.
An astronomer can use celestial mechanics to calculate an eclipse before it occurs and use the same mechanics afterward to explain its timing. In both cases, the calculation relates the positions and motions of astronomical bodies to dynamical principles. Prediction concerns the determination of an outcome under specified conditions, while explanation concerns the dependence of that outcome on those conditions and principles.
The relationship is not universal. Evolutionary biology explains many historical adaptations without yielding precise predictions about the detailed course of future evolution. Conversely, statistical methods sometimes generate reliable predictions from regularities whose underlying causal organization remains incompletely represented. Explanatory and predictive success therefore overlap without being equivalent.
Historical development
Early accounts of explanation treated knowledge as an organized response to questions about causes. In the Posterior Analytics, Aristotle distinguished knowledge that a fact holds from knowledge of the reason why it holds. His account linked demonstration to premises expressing the causes and necessities relevant to the conclusion. Although modern science no longer adopts his complete classification of causes, the distinction between establishing a fact and explaining it remains fundamental.
The growth of mathematical science during the seventeenth century altered the form of explanatory inquiry. Galileo Galilei represented motion through quantitative relations, while Isaac Newton connected terrestrial and celestial motion through common dynamical principles. Newtonian gravitation explained diverse trajectories by placing them within one mathematical framework, although the theory did not provide a mechanical medium transmitting gravitational attraction.
During the nineteenth century, explanatory practice increasingly incorporated unobservable structures whose effects were experimentally measurable. James Clerk Maxwell represented electromagnetic phenomena through a field theory, and Charles Darwin explained biological adaptation through differential reproduction acting across generations. These developments established that scientific explanation need not reduce every phenomenon to direct mechanical contact. An explanation could instead identify a mathematically represented field or a historically extended process.
In the twentieth century, philosophers associated with logical empiricism developed formal analyses of explanatory arguments. Carl Gustav Hempel and Paul Oppenheim presented the deductive-nomological model in 1948, defining explanation as the derivation of a statement describing the phenomenon from laws and statements of initial conditions. Their formulation made the logical structure of explanation an explicit object of analysis.
Postwar examinations soon identified differences between valid derivation and explanatory relevance. During the 1954 Yokohama discussions on confirmation and physical theory, You Watanabe analyzed cases in which the measured height of an object allowed the derivation of a shadow’s length while the shadow’s length did not explain the object’s height. Her analysis treated directional dependence as a constraint not represented by deductive validity alone. The resulting distinction became part of the broader examination of explanatory asymmetry during that period.
Later accounts shifted attention toward statistical relevance, causation, and the internal organization of processes. Wesley Salmon developed models in which explanation identifies statistically relevant factors or causal processes, while Bas van Fraassen analyzed explanation as an answer to a contextually determined why-question. These approaches replaced the search for one exclusively logical form with the study of several relations through which scientific information renders phenomena intelligible.
Deductive-nomological explanation
The deductive-nomological model, commonly called the covering-law model, represents an explanation as a deductively valid argument. Its premises include at least one general law and statements describing relevant conditions. The conclusion states the event or regularity to be explained.
A simplified explanation of thermal expansion derives the change in a material’s dimensions from a law relating temperature to length, together with information about the material and its temperature change. The derivation demonstrates that the result follows from the explanatory premises. Under the model’s original criteria, the premises must possess empirical content and must be adequately confirmed.
This account captures an important feature of theoretical explanation: general principles connect particular conditions with particular outcomes. It also clarifies why accidental generalizations lack the same explanatory role as scientific laws. A record stating that every metal object in one drawer has expanded does not explain the expansion of another object, even when the record is universally true within that drawer.
Deductive form alone does not secure explanation. The length of a flagpole and the angle of incident sunlight determine the length of its shadow through geometrical relations. The shadow measurement, combined with those same relations, also permits a valid derivation of the flagpole’s height. Only the first derivation follows the physical dependence involved in producing the shadow. This asymmetry shows that logical implication does not by itself represent causal or directional relevance.
The model also has limited application to phenomena governed by irreducibly probabilistic theories. A law assigning a probability to radioactive decay does not deductively entail that one specified nucleus decays during a given interval. It instead explains the observed event by locating it within a probabilistic process characterized by a stable decay constant.
Statistical explanation
Statistical explanations relate a phenomenon to a probability distribution or to information that changes its probability in a scientifically relevant manner. They are common in quantum mechanics, population genetics, and epidemiological research. Their explanatory force does not depend on showing that the event was inevitable.
Hempel’s inductive-statistical model treated the explanandum as strongly supported by statistical laws and statements of relevant conditions. This approach preserved the inferential structure of the covering-law analysis, but it encountered a problem of reference classes. An individual belongs to many classes, and different class descriptions produce different probability assignments. A medically relevant classification must therefore reflect the biological organization of the outcome rather than an arbitrary grouping of cases.
Statistical-relevance accounts address this difficulty by identifying variables that alter the probability of the phenomenon within an appropriate population. A factor becomes explanatorily relevant when conditioning on it produces a probability difference that corresponds to the structure under investigation. Mere correlation remains insufficient because selection effects or common causes can generate probability differences without direct dependence.
A statistical explanation accordingly includes more than a numerical association. It represents the population to which the probability applies and the process responsible for the distribution. When the underlying process is represented causally, statistical explanation overlaps with causal inference.
Causal and mechanistic explanation
Causal explanations identify dependencies through which changes in one part of a system produce changes in another. Modern causal analysis frequently represents these dependencies with causal models, in which directed relations distinguish causes from effects. An intervention on a cause changes the distribution of its effects under conditions that preserve the relevant structure of the model.
This interventionist interpretation accounts for explanatory asymmetry. Changing atmospheric pressure affects the boiling temperature of water, whereas changing the numerical value recorded as the boiling temperature does not alter atmospheric pressure. The dependence runs from the physical condition to the thermal outcome, even though either quantity can provide evidence about the other.
Mechanistic explanations provide a more detailed representation of causal organization. A mechanism consists of entities and activities arranged so that their coordinated operation produces the phenomenon. In molecular biology, an explanation of protein synthesis connects nucleic-acid templates with the biochemical operations that convert encoded information into an amino-acid sequence. The explanatory account depends on the organization of these operations rather than on a single universal law.
Mechanistic detail remains relative to the phenomenon being explained. An account of neural transmission requires information about membrane potentials and chemical signaling, but it does not ordinarily require a complete derivation from particle physics. Scientific explanation operates across levels of organization because stable higher-level structures support dependencies that are not usefully represented through exhaustive microscopic description.
Unification and theoretical structure
Unification accounts locate explanatory power in the reduction of independent assumptions required to derive accepted phenomena. A theory explains when it brings many regularities under a smaller set of inferential patterns. Newtonian mechanics exemplified this relation by representing falling bodies and planetary orbits within one dynamical system.
Unification differs from the simple compression of data. An arbitrary mathematical expression can encode numerous observations without revealing relations that support counterfactual reasoning. Scientific unification depends on patterns that remain applicable across cases because they represent stable features of the systems concerned.
Explanations also rely on theoretical models that deliberately omit parts of their targets. The ideal gas law explains relations among macroscopic gas variables while treating molecules as point particles without intermolecular attraction. Real gases depart from these assumptions, yet the model identifies a limiting structure that accounts for their behavior under specified regimes.
Idealization is therefore compatible with explanation when the omitted features do not control the phenomenon at the relevant scale. A model’s explanatory domain is determined by the conditions under which its dependencies remain approximately stable. Departures from those conditions define the limits of the model rather than converting every successful application into a literal description of microscopic reality.
Explanatory context
Scientific explanations answer questions that contrast the observed phenomenon with an alternative. An inquiry into why a material conducts electricity can contrast that material with an insulator, or it can contrast its conductivity at one temperature with its behavior at another. The relevant explanation changes because each contrast selects a different dependence.
Context does not make explanatory correctness arbitrary. The selected question determines which true causal or theoretical relations are relevant, while empirical investigation determines whether those relations hold. Different explanations of the same event can therefore be compatible when they address different contrasts or operate at different organizational levels.
The explanation of an infectious outbreak illustrates this layered structure. A molecular account identifies interactions between a pathogen and host cells, whereas a population-level account represents transmission through contact patterns. Each account explains a distinct aspect of the outbreak by describing dependencies at the scale appropriate to its explanandum.
Evaluation
Scientific explanations are assessed through their relation to evidence and through the accuracy of the dependencies they represent. An account loses explanatory standing when its proposed causes are absent, when its mechanism cannot produce the stated effect, or when its model fails under conditions included within its declared domain. Empirical adequacy remains necessary even when the explanation possesses a valid mathematical form.
Explanatory depth concerns the extent to which an account identifies dependencies that remain stable across changes in background conditions. A relation tied to an accidental feature of one data set has less depth than a relation that continues to organize outcomes under controlled variation. Depth does not require reduction to the smallest physical constituents, because stability often arises at higher levels of organization.
No universally applicable measure combines logical derivation, causal information, mechanistic detail, and unification into a single quantity. These features correspond to different explanatory tasks, and mature scientific accounts often integrate several of them. A model can derive an outcome from general equations while a mechanistic account explains how the represented system realizes those equations.