Statistical Explanation

Introduction: The Event Within the Distribution

A statistical explanation concerns the relation between a particular event and the structured distribution of possible events to which it belongs. When a patient recovers, a radioactive nucleus decays, or a machine component fails, an investigator may know the relevant risk without possessing a deterministic derivation of the outcome. The philosophical question is what that knowledge establishes. A numerical probability can characterize expectation, comparative vulnerability, exposure, physical propensity, or inferential uncertainty; the numbers alone do not decide which interpretation the explanation requires.

The modern debate took shape after Hempel and Oppenheim's logical analysis of explanation (1948) and the subsequent development of probabilistic covering-law accounts. Hempel's collected explanatory investigations (1965) provided the background against which Jeffrey, Salmon, and Railton questioned whether statistical understanding should be measured by the probability assigned to the outcome. The overview of scientific explanation by Woodward and Ross (2025) documents how these disputes led toward causal, mechanistic, and other alternatives.

The inquiry pursued here is specific. When does the statement that an event occurs with some frequency in a defined class explain that event, and when is it only a description of a pattern? The answer requires attention to the target, the explanatory contrast, the reference class, the origin and interpretation of the probability, and the relation that makes an included factor relevant. This monograph develops that answer through comparative interpretation and explicit worked examples, with no claim to invent a new theory of explanation.

The Domain and Logic of Statistical Explanation

Statistical explanation is the activity of making an event, distribution, dependence, or risk intelligible by connecting it to probabilistic regularities and the conditions under which they obtain. The explanandum may be an individual occurrence, such as a particular device's failure, or a population fact, such as an unequal failure rate across conditions. These are distinct targets. Explaining the distribution does not automatically explain the token event, and explaining a token's risk does not establish why this exact outcome occurred rather than another compatible one.

There are at least four logically different propositions in a typical explanation: a population-level statement about frequency; a conditional distribution given selected properties; a claim about causal influence; and a singular claim that a particular event occurred because of specified circumstances. The transitions between them require bridging assumptions. For instance, P(Failure | Heat)=0.20 describes a conditional probability. It supplies neither the causal claim that heat increases failure nor a deterministic description of the actual failure process. To treat it as explanatory, we must ask whether heat is temporally and structurally relevant, whether observations are confounded, and what the probability represents.

The field also encompasses statistical explanations of aggregate patterns, where a distribution itself is the explanandum. An account of why the prevalence of a condition differs among populations need not resolve every individual case. Conversely, an adequate mechanistic story for a single case need not explain why that kind of event is common or rare. Precision begins by choosing the scale of explanation before evaluating the mathematics.

Explanation, Description, Evidence, and Prediction

A statistical description records how outcomes vary. A prediction assigns or estimates the chance of an as-yet-unobserved outcome. Evidence changes warranted credence in a hypothesis. An explanation connects the event to the conditions and relations that make its occurrence intelligible. These tasks frequently cooperate but have different criteria of success. Accurate prediction is possible by tracking symptoms of a cause; explanation can demand identification of that cause or a relevant process.

Salmon's analysis of explanation and causality in Galavotti's 2018 encyclopedia study makes the barometer-and-storm illustration prominent. Atmospheric-pressure readings can forecast storms, while the meteorological system provides the relevant explanation. The example has an exact statistical analogue: a variable can carry predictive information because it is an effect or proxy of the operative process. Its statistical association with an outcome cannot by itself settle its explanatory status.

A good explanation therefore need not maximize the predictive probability of the exact event. It should select information that is relevant to the question, correctly calibrated to the appropriate domain, and defensible in its interpretation. For some scientific inquiries, a robust causal mechanism is central; for others, a symmetry, conservation constraint, or distributional structure can make a pattern intelligible without a conventional intervention. This plurality of questions must remain visible throughout the evaluation.

Historical Foundations: The Covering-Law Program

The covering-law program sought an explicit logical relation between an explanandum and general laws plus antecedent conditions. Hempel and Oppenheim (1948) stated a classical explanatory schema in which the event description follows from appropriate law-containing premises. The deductive-nomological ideal establishes one rigorous meaning of expectability: given the premises, the explanandum could not fail to follow. Its achievements include making the roles of generality, antecedent information, and inference logically inspectable.

The difficulty for intrinsically stochastic events is immediate. If a physical process gives an outcome a probability of 0.02, its actual occurrence is compatible with the law yet cannot be deduced as an inevitable conclusion. A description of the mechanism and its probability may still make the event scientifically understandable. Hempel's larger 1965 account of explanation incorporated statistical explanation under an inductive conception, preserving a relation between explanatory success and expectability.

The deductive and probabilistic formulations share an important ambition: explanatory statements should rely on generalizable knowledge rather than an ad hoc narrative constructed after an outcome is known. Their liabilities become visible when relevance, causal direction, heterogeneous populations, and rare events are examined. These challenges matter because logical coherence is necessary for disciplined reasoning, while the explanatory suitability of premises is a further question.

Hempel's Inductive-Statistical Explanation

In an inductive-statistical (IS) explanation, a statistically lawlike generalization together with case information confers a high probability on the explanandum. The structure is inductive: the explanandum is supported with greater or lesser strength rather than deductively entailed. Hempel's requirement of high probability sought a principled account of why events appear expectable in light of explanatory facts. It also connected explanation with inference, which could be assessed using familiar probabilistic standards.

The model generates a challenge. If a particle decays within a brief interval despite a very low probability of doing so, a scientifically accurate account of the physical process yields the low probability that the model makes problematic. Its improbability is itself part of the scientific description of how it came about. Salmon, Jeffrey, and Railton subsequently rejected the view that explanatory worth should be tied to high probability in this simple manner, albeit for different reasons. The 1971 statistical-relevance volume presents a historically central response.

A charitable reconstruction of Hempel's project recognizes its safeguards. Its standards were directed at avoiding opportunistic selection of statistics and unwarranted explanatory authority. The argument against the high-probability criterion is therefore incomplete unless it supplies its own principles of relevance and reliable reference-class selection. Simply allowing rare events to be explained does not license every post hoc probabilistic narrative.

Statistical Ambiguity and Maximal Specificity

Suppose a person belongs to a broad reference class where recovery is likely and a more specific class where recovery is unlikely. Either statistic may be mathematically correct, yet the two generate different explanatory assessments. The problem is not inconsistent probability theory; conditional probabilities can differ across nested or overlapping populations. The difficulty concerns the rule for choosing which facts should enter a legitimate explanatory account.

Hempel's 1968 reconsideration of maximal specificity explicitly reexamined the maximal-specificity requirement, distinguished its rationale from the rule of total evidence, and emphasized the lawlike character of relevant generalizations. These distinctions prevent a misleading simplification in which maximal specificity means merely listing every known characteristic. A person's favorite color, for example, can divide a population without yielding a scientific statistical law relevant to prognosis.

Maximal specificity captures a valid aspiration: do not conceal relevant information by stopping at a convenient coarse grouping. Yet it raises a practical and epistemological problem. Relevant variables are often unknown, incompletely measured, or impossible to jointly observe. Subdivision can also destroy statistical precision. A tiny subgroup can be maximally detailed and poorly estimated. Philosophical analysis must therefore separate a hypothetical ideal of complete specification from empirically justified decisions under uncertainty.

Jeffrey's Challenge to the Inferential Picture

Richard Jeffrey's 1969 essay, reproduced in the Cambridge collection as Statistical Explanation versus Statistical Inference, questioned the assimilation of explanation to inferential support. An observed outcome may be understood through a relevant chance arrangement even when that arrangement assigned it low probability. Inferential assessment asks how strongly premises support a conclusion; explanatory assessment asks why the actual occurrence belongs within a relevant account.

Consider a fair die that lands on six. Before the throw, assigning the outcome probability one-sixth is correct; after observing six, the probability can be revised to one conditional on the observation. Neither numerical transition explains why the die landed on that face. Details about the physical trial may provide a causal account, while fairness and uniformity characterize the distribution of possible outcomes. The epistemic probability of our belief and the stochastic probability associated with the physical setup must be kept distinct.

The broader implication is that statistical explanations are evaluated retrospectively without being invalidated by their retrospective character. Explanation after an event uses information that was unavailable beforehand. This does not make the explanation arbitrary, provided the information identifies independently warranted features of the generating situation.

Salmon's Statistical-Relevance Model

Wesley Salmon's contribution centered the statistical relation between relevant factors and the explanandum. In Statistical Explanation and Statistical Relevance (Salmon, Jeffrey, and Greeno 1971), a statistical-relevance explanation identifies attributes whose presence or absence changes the conditional probability of the event and situates the event within a relevant probability distribution. The explanation may assign a high, moderate, or low probability; its function is to reveal the pattern of dependence.

For a target event E, an attribute F is statistically relevant within background class B when P(E | F,B) differs from P(E | not-F,B), assuming the conditional probabilities are defined. This comparison makes importance relational and context-dependent. If a patient's treatment and age are relevant to recovery, grouping patients by both may yield more explanatory information than an unqualified average. The number attached to the patient's class matters through its place in a network of meaningful distinctions.

The model's advance is conceptual and methodological. It shifts attention from an event's mere probability to the configuration of factors associated with its occurrence. But the difference between association and causal relevance remains unresolved. Variables can be statistically relevant because of a shared cause or selection effect. Salmon's subsequent work acknowledged this issue; his causal-structure monograph (1984) offers a different, more ontic account.

Statistical Relevance Is a Contextual Relation

The inequality P(E | F,B) ≠ P(E | not-F,B) does not identify an intrinsic property of F detached from its context. Conditional probabilities may reverse when background conditions change. A marker associated with low risk overall may be associated with high risk in each of several carefully chosen strata. Statistical relevance is therefore a relation among an event, a property, a class, and the alternatives distinguished within that class.

The relevant comparison also depends on the question. To ask why mortality differs between two hospitals is to compare institutional populations and patient mixes. To ask whether a specific medical intervention changes a patient's risk is to compare potential outcomes under intervention and control. To ask why an unexpected event occurred in one hospital may additionally require a token-level causal history. The same table cannot perform these tasks automatically.

Salmon's importance lies in turning attention toward relevant statistical differences. The limitation arises because statistical differences require interpretation. A robust philosophical account must make the context explicit and supply independent reasons for treating the relevant difference as explanatory of the target rather than merely useful for classification.

Reference Classes: The Problem of the Right Population

An event can belong simultaneously to many reference classes. A device is one of all devices, devices of its model, devices manufactured in a particular batch, and devices exposed to a particular temperature regime. Each class may have a different failure rate. The proposition that a device failed with probability 0.10 is incomplete unless the conditions and distribution are specified. Assigning it to a narrower class need not resolve the problem when that class is selected after observing failure.

The analysis of probabilistic causation by Hitchcock (2018 revision and later updates) discusses how background factors and reference classes complicate probability-raising accounts. A reference class is explanatorily warranted when its boundaries correspond to the target process or a well-supported comparison, when its cases are meaningfully exchangeable for the purpose at hand, and when the estimated probabilities are credible. Exchangeability here is a substantive statistical assumption about relevant comparability, not a declaration that all observed objects are identical.

In practice, the narrowest class can be too small for a stable estimate, and the broadest class can conceal decisive heterogeneity. Appropriate specification balances substantive relevance, selection bias, estimation uncertainty, and transport to the event being explained. These requirements are part of scientific methodology and should not be mistaken for a theorem guaranteeing a unique reference class.

Probability, Chance, Frequency, and Credence

The term probability covers several distinct interpretations. Observed relative frequencies estimate proportions in recorded populations. A propensity account interprets probability as a feature of a physical chance setup. Subjective or epistemic probability represents rational uncertainty under evidence. A model probability is defined relative to mathematical assumptions. These quantities can numerically coincide while answering different explanatory questions.

Peter Railton's probabilistic explanation proposal (1978) ties a deductive account to objective chance mechanisms and propensities. Its explanatory aim is a description of the chance process responsible for the observed fact, rather than a demonstration that the event was likely. By contrast, a Bayesian posterior probability about a hypothesis primarily tracks evidential support. An investigator who reports posterior probability 0.92 for a mechanism has not thereby shown that mechanism produced the event.

A rigorous statistical explanation declares which probability concept is operating, which events make up the sample space, whether the estimate is conditional or marginal, and what uncertainty accompanies it. Failure to identify these elements invites a false equivalence between frequency of recorded outcomes, strength of belief, and objective causal chance.

Why Rare Events Can Be Explained

Rare outcomes expose the gap between expectability and intelligibility. A mechanism may produce several outcomes with objectively specified chances, including an outcome that happens only once in a thousand trials. When the rare outcome occurs, an accurate explanation describes the setup, relevant stochastic law, and realized branch of the process. Its correctness cannot depend solely on making the already observed outcome retrospectively probable.

The problem features prominently in Salmon's 1971 statistical-explanation essay and in Railton's 1978 account of chance mechanisms. Their approaches differ in form, yet both undermine the unrestricted principle that a good explanation must confer a high probability on the outcome. The explanatory information can include exactly why the event was unusual under the conditions that generated it.

A low-probability explanation is strongest when it does more than restate improbability. It may show that an exceptional event was permitted by a well-supported physical mechanism, identify a distinctive risk factor, or describe the causal pathway through which the low-probability branch was realized. Where the token pathway is inaccessible, the explanation may be explicitly partial: it accounts for the probability structure without identifying the final microphysical details.

Railton and the Chance-Mechanism Alternative

Railton (1978) proposed a deductive-nomological model for probabilistic explanation using propositions about objective chance rather than treating the outcome as an inductively probable conclusion. In A Deductive-Nomological Model of Probabilistic Explanation, he argues that probabilistic accounts should describe the chance mechanisms responsible for facts. His use of deduction concerns what follows about the process and its chances; it does not entail that a random outcome was guaranteed.

The view illuminates a distinction between explaining a process and inferring an outcome. A radioactive nucleus can decay at a particular instant even though its decay was unlikely in the chosen interval. A correctly stated stochastic law can explain the kind of process and the distribution of decay times. Whether such an account fully explains the precise instant of a particular decay is a further philosophical question, especially where the event may be irreducibly stochastic.

Railton's proposal also depends on objective chance interpretation, which is more persuasive in some physical contexts than in situations dominated by measurement, sampling, model uncertainty, and incomplete social-scientific knowledge. Its strength is structural clarity about chance mechanisms; its scope should not be extended indiscriminately to every statistical regression or risk score.

Salmon's Causal-Mechanical Turn

Salmon's subsequent Salmon’s causal-mechanical study (1984) distinguished epistemic, modal, and ontic conceptions of explanation and developed a causal-mechanical alternative. An explanation locates a phenomenon within the causal structure of the world. The distribution of observed events remains important, but statistical relevance is only one component of a richer understanding of what produces the phenomena.

This revision addresses a specific deficit in purely statistical accounts. The barometer and the storm are correlated; readings carry information about changing conditions. The physical meteorological process explains the storm. Likewise, a diagnostic biomarker may predict deterioration because it responds to a pathological process. Without knowing which direction the causal pathways run, the explanatory status of the biomarker remains unsettled.

The causal-mechanical position raises its own questions: how much mechanistic detail is necessary, which links in a complex process count as explanatory, and how can irreducibly probabilistic mechanisms be described? Hitchcock's critique of explanatory relevance (1995) argues that Salmon's account had not completely solved the selection of causally explanatory detail. A causal vocabulary alone does not remove the need for judgments of relevance.

Statistical Relevance and Causal Relevance

Statistical relevance can be expressed by conditional dependence. Causal relevance concerns how a change in a factor contributes to a change in the outcome, or how a factor participates in the actual producing process. The former is defined by a probability distribution; the latter relies on assumptions or evidence about the system's causal organization. These concepts often align but have no general equivalence.

The philosophy of causal models explains that graphs and structural equations relate observed dependence to assumptions about causal structure. A correlation between treatment choice and prognosis may arise because the underlying severity influences both treatment and outcome. The observed conditional distribution can therefore carry predictive information while misrepresenting the effect of assigning treatment.

Any proposed transition from statistical to causal explanation should identify the estimand, relevant pre-treatment common causes, possible mediators, selection conditions, and assumptions needed for causal identification. Such an account respects the explanatory potential of statistics while refusing to treat mathematically correct dependence as self-interpreting causal evidence.

Suppes, Eells, and the Probability-Raising Tradition

Suppes (1970) developed a probabilistic theory of causality that placed temporal priority and probability raising at the center of causal analysis. The basic intuition is that a cause may increase the probability of an effect relative to a suitable alternative or reference condition. His Probabilistic Theory of Causality stands at an important crossroads between formal probability, empirical regularities, and philosophical attempts to identify causal relations. Its contribution is to make causal assessment sensitive to differences between conditions rather than to a single unconditional percentage.

Eells (1991) further investigated the context dependence of probabilistic causal relevance in Probabilistic Causality. The distinction between apparent probability raising and causally relevant probability raising becomes especially pressing once background factors and alternative causal histories are taken seriously. If a factor increases an event's observed frequency only because it is associated with a common cause, statistical elevation alone cannot establish its causal contribution.

The formal statement P(E|C) > P(E|not C) is therefore a useful starting point and an insufficient final criterion. The comparative populations must be specified; C and its alternative must be meaningful; and the relevant conditioning context must be selected through substantive knowledge. Probability raising may persist after legitimate adjustment, disappear under control for a confounder, or be obscured by the mixture of pathways generating the observations.

Cartwright's Causal Laws and Effective Strategies (1979) sharpened the issue by connecting causal judgment with effective action and the conditions under which probability relations can inform intervention. The conceptual progression from Suppes through Eells and Cartwright reveals why statistical explanation becomes more powerful when its comparisons are fitted to the actual causal question. It also shows the need to resist any uniform inference from a positive statistical association to a manipulable cause.

Screening Off, Confounding, and Common Causes

Suppose two observed variables A and B covary because a third variable C influences both. Under the appropriate assumptions, conditioning on C may remove the association between A and B. This possibility is expressed by the screening-off condition P(A,B|C) = P(A|C)P(B|C). It identifies a conditional-independence pattern compatible with common-cause structure, while leaving open the task of establishing the structure empirically. A mathematical pattern in observed data can have multiple causal explanations.

The Stanford account of probabilistic causation surveys the historical effort to relate probabilities, spurious associations, causal chains, and contextual relevance. The explanatory question is whether the relevant variable belongs in the producing system or merely helps predict a variable further downstream. The same observed independence may arise through differing causal models; measurement error, sample selection, or imperfect conditioning may conceal the expected pattern.

Consider a population where severe illness makes a patient more likely both to receive an intensive treatment and to experience an unfavorable outcome. A crude comparison may show that treated patients have worse outcomes. The severity variable explains why treatment and prognosis are associated. The observed relation is real and may be useful for prediction, but it does not by itself estimate whether treatment worsened prognosis. The example is conceptual; it supplies no medical estimate and assumes no particular treatment effect.

Statistical explanation here has two stages. First, the analyst identifies the probability pattern to be explained: a conditional association with a particular reference population. Second, the analyst examines how that pattern would be generated by candidate structures of common cause, mediation, selection, or genuine causal influence. The second stage is what turns an association from a numerical fact into evidence bearing on an explanatory account.

Simpson's Paradox: A Worked Numerical Example

Simpson's paradox makes the dependence on reference populations mathematically vivid. The Stanford treatment of Simpson's paradox presents the general problem: the direction of an association within relevant subgroups can differ from its direction after the groups are pooled. The apparent contradiction is produced by the changing denominators and group weights. The same arithmetic can be scientifically significant when group composition relates to causes, selection, or the estimand of interest.

The following table is an original synthetic example. Every count is invented solely to demonstrate aggregation; none reports a clinical study or observed population. The outcome is called success, and A and B are alternative procedures.

B has the higher success rate in both strata, while A has the higher rate when the data are pooled. The pooled rates are calculated as 91 ÷ 110 ≈ 0.8273 and 49 ÷ 120 ≈ 0.4083. A is used mainly in the lower-complexity group; B is used mainly in the higher-complexity group. This imbalance changes the mixture weights. Both the stratum-specific comparisons and the pooled comparison are arithmetically true. Their explanatory meanings depend on why cases entered the groups and which comparison answers the question.

If the scientific question is how A and B perform among cases with the same pre-existing complexity, the subgroup comparison is more directly relevant, subject to all remaining sources of bias. If the question concerns the observed success rate of the actually treated mixtures, the aggregate numbers describe those mixtures. Neither comparison is automatically a randomized causal effect. In a causal analysis, group membership might be a confounder, a mediator, or a selection variable; the correct adjustment strategy changes accordingly.

The example demonstrates a central conclusion of this dossier: distributions are always distributions of specified populations under specified arrangements. A statistical regularity acquires explanatory force when the arrangement that generated the numbers becomes intelligible. The arithmetic exposes the problem, and causal or institutional background knowledge determines which contrast matters.

Conditioning and Intervention Are Different Operations

A conditional probability describes the distribution among cases in which a condition is observed. An intervention changes the system by setting a condition through a stipulated operation. Pearl's Causality: Models, Reasoning, and Inference (2009, second edition) formalizes this difference through structural causal models and intervention notation. P(Y|X=x) and P(Y|do(X=x)) need not coincide, because observing X=x and setting X=x can select different distributions of background causes.

A simple example is the difference between observing that people carrying umbrellas are often near rainfall and forcing the use of umbrellas in an otherwise comparable setting. The observational conditional relation can be strongly predictive. The intervention changes umbrella use while leaving the meteorological processes operating through their own causes. The possibility of intervention provides a conceptual test of whether an association is attributable to the manipulated factor or to a common antecedent.

The interventionist idea requires assumptions. Variables must be defined at useful levels, causal relationships represented sufficiently accurately, and the proposed manipulation understood. In some sciences physical intervention is impossible, unethical, or conceptually awkward. An explanatory analysis can still evaluate hypothetical interventions or natural variation; the inferential warrant is then proportionate to the design and assumptions actually available.

Consequently, a statistical explanation of an observed association may be complete as a description of why the association arises, even when a causal effect cannot be identified from the data. A causal explanation of an intervention effect requires an additional layer of warranted inference. Confusing these two projects inflates the epistemic reach of a statistical model and obscures the limits of what the evidence supports.

Directed Graphs, Mediators, and Colliders

Causal graphs represent proposed directions of influence and the conditional independences associated with their assumptions. Spirtes, Glymour, and Scheines developed methods for causal discovery and evaluation in Causation, Prediction, and Search (second edition, 2000). Such methods give statistics a disciplined role in testing features of possible causal structures. Their conclusions are conditional on properties such as causal sufficiency, appropriate sampling, and other modeling assumptions; a graph is a proposal about the system rather than an automatic reading of correlations.

Distinguishing a confounder from a mediator is critical. If Z affects both X and Y before X occurs, adjustment for Z may help isolate an X–Y relationship under suitable assumptions. If X affects Z and Z affects Y, adjusting for Z may block part of the effect of X that an analyst wished to understand. The same variable can occupy different explanatory roles depending on the substantive target, temporal structure, and definition of the total or direct effect.

Colliders pose a less intuitive danger. If X and Y independently contribute to a common outcome S, conditioning on S can induce an association between X and Y even when they are otherwise independent. For example, among applicants accepted through either of two separate qualifications, a high score on one qualification can become negatively associated with a high score on the other, solely because selection depended on at least one route to acceptance. Selection can therefore manufacture a statistical relation that would be absent in the unselected population.

The Stanford entry on causal models explains the philosophical role of directed graphical structures and structural equations in identifying causal questions. The operational lesson is that adjustment is an explanatory choice. Conditioning on everything measured can worsen rather than improve a causal analysis. Scientifically appropriate adjustment depends on the generating structure that the statistical explanation is meant to reconstruct.

Counterfactual Dependence and Explanatory Invariance

Woodward's Making Things Happen (2003) identifies explanatory depth with information about patterns of change under interventions. A statistical relationship earns a causal interpretation when it supports appropriately specified answers to questions about how the outcome would change if a relevant variable were altered. This does not require every legitimate explanation to be experimentally tested; it does require a defensible account of the relation between variations in the explanatory factor and variations in the effect.

The interventionist emphasis helps distinguish a stable background association from a causal relation. If a model predicts an outcome only because it includes a proxy for an unmeasured cause, intervening on the proxy may leave the outcome unchanged. By contrast, a representation of a genuine causal pathway can support certain counterfactual questions, though its validity remains limited to the domain where its underlying assumptions hold.

Invariance matters because statistical regularities may disappear when populations, institutions, or measurement procedures change. An explanation that identifies conditions under which a relationship remains stable offers more transportable knowledge than an isolated fitted coefficient. Yet invariance itself can be partial: interventions may alter the system, causal relations may be heterogeneous, and the availability of variables may differ across domains. The explanatory judgment therefore has to specify which changes the relation is claimed to survive.

An explanatory account can be causally informative without specifying every lower-level component. It is sometimes enough to identify a dependence that would reliably change under a certain class of interventions. The degree of detail justified is determined by the question and evidence, not by a uniform requirement that explanations descend to microphysical descriptions.

The Single Case and the Population

Statistical explanation confronts a persistent ambiguity of target. Does an account explain why this person developed a disease, why a particular particle decayed, why a machine failed on Tuesday, or why a population has a certain incidence? These are different explananda. A population model may explain the distribution of outcomes while leaving unanswered why one individual occupied a specific branch of the distribution. Conversely, a detailed token-level causal history may explain one event without accounting for the population pattern in which it occurs.

The point can be illustrated by the contrast between a coin's fairness and the outcome of a particular toss. Assuming a familiar physical setup that produces approximately equal chances, the fairness model explains the long-run proportions expected from many tosses. It does not tell us which precise side a specified future toss will show. A fuller physical account of launch conditions might explain a token outcome where those conditions are sufficiently known, even though that information was unavailable beforehand.

Salmon's Salmon’s causal-mechanical study and Railton's probabilistic explanation proposal both press the issue of linking chance mechanisms with actual events. The ensuing debate reveals that explanatory completeness must be indexed to the question. One may seek an adequate statistical explanation of the rate, an account of the chance mechanism, and a causal history of the individual outcome. The warrant for each should be assessed independently.

Statistical regularities consequently explain singular events to the degree that they identify relevant generating conditions and situate the event within a justified causal or stochastic process. Where the process contains an objectively chancy element, the residual unpredictability of the token event is a feature of the explanation's subject matter. There is no need to manufacture a deterministic guarantee that the science does not possess.

Contrastive Questions and the Pragmatics of Explanation

A question of the form “Why did E occur?” often conceals an alternative: why E rather than not-E, why here rather than elsewhere, or why this rate rather than another rate? Van Fraassen's pragmatics of scientific explanation, developed in The Scientific Image (1980), treats explanatory requests as dependent upon a question's contrast and relevance relation. The requested explanation changes when the relevant alternative changes.

Consider a statistically unusual equipment failure. If the contrast is “Why did this model fail more often than a comparison model?”, the explanation may concern differences in design or usage distributions. If the contrast is “Why did this machine fail today rather than yesterday?”, the explanation may concern a particular stress event or maintenance intervention. If the contrast is “Why did any machine of this model fail?”, the analysis may begin with the mechanism's inherent risk and operating environment. All three questions refer to related observations; they demand different explanatory information.

Contrast selection also guides which reference class is appropriate. An all-population rate may answer a budgetary planning question and fail to answer an individual etiological question. A conditional rate among cases with a specific exposure may answer the latter and be misleading for unconditioned public forecasting. The relevant comparison is part of the scientific question, even when it is not fully articulated in ordinary speech.

This pragmatic sensitivity should not be mistaken for arbitrary subjectivity. A contrast can be explicit, checked against the data, and challenged for relevance. The empirical system constrains what counts as a defensible answer. Pragmatics determines the explanatory request; evidence determines whether the answer withstands scrutiny.

Bayesian Evidence and Explanatory Adequacy

Bayesian methods update probabilities assigned to hypotheses in light of evidence. Their central expression, P(H|D) = P(D|H)P(H)/P(D), provides a coherent relation among prior probabilities, likelihoods, and posterior probabilities when the quantities are properly defined. A posterior is a degree of support conditional on assumptions and evidence. It does not itself show how the observed event was produced.

A hypothetical diagnostic calculation clarifies the distinction. Suppose the prevalence of a condition is 1%, the probability of a positive test given the condition is 90%, and the probability of a positive test without the condition is 10%. By Bayes' rule, the probability of the condition given a positive result is (0.90 × 0.01) / [(0.90 × 0.01) + (0.10 × 0.99)] ≈ 8.33%. These numbers are invented as a teaching example, not taken from medical evidence. The computation explains how evidence changes belief under the assumptions. The disease process and the test's biological or technical response require a separate causal account.

The distinction echoes Jeffrey's Statistical Explanation vs. Statistical Inference (1969): establishing a rational inference and supplying a scientific explanation are related activities with different success conditions. A statistical model can outperform alternatives in predictive probability and remain opaque about the causal pathways driving the forecast.

Bayesian inference can nevertheless assist explanation. Posterior comparison may weigh rival generating hypotheses, reveal sensitivity to assumptions, and distinguish explanations that make very different predictions about independently collected evidence. It gains explanatory significance when the hypotheses have interpreted content and the data are suitable for discrimination. The philosophical mistake lies in treating probability assignment, whether high or low, as a complete substitute for a warranted account of the phenomenon.

Fit, Prediction, and Explanatory Representation

A regression that reproduces a historical dataset may be an excellent summary of observed covariation. A model that generalizes to independent cases may be a useful predictor. An explanatory model must in addition clarify why its variables, structure, and target relation are relevant to the phenomenon. Predictive success provides evidence for a model in some respects; it is not a logically sufficient criterion for causal accuracy or explanatory completeness.

Overfitting illustrates the distinction. A model with enough flexibility may fit random features of its training sample that are not stable characteristics of the system. Out-of-sample validation is therefore an important way to test the forecast's robustness. Yet even a model that predicts well after validation may rely on convenient correlates, historically stable institutional proxies, or an accidental feature of the observation process rather than the mechanism one wishes to understand.

The causal discovery literature in Causation, Prediction, and Search stresses that inferential conclusions follow from explicit assumptions about data generation. Woodward's interventionist account adds a philosophical test: what information does the representation provide about changes under relevant interventions? Both perspectives discipline the move from observed fit toward explanatory interpretation.

Statistical explanation thus has a representational requirement. It should identify what the probabilities are probabilities of, which processes or contrasts the model depicts, how uncertainty enters, and which judgments remain dependent on untested assumptions. A fitted equation earns its place in a scientific explanation by its relation to an interpreted target, not merely by the elegance or size of its coefficients.

Epidemiological Explanations and Causal Inference

Epidemiology shows why statistical and causal reasoning have to work together. Frequencies, risk ratios, hazard estimates, and adjusted associations summarize patterns of disease across populations. Investigators seek explanations of incidence, exposure-related variation, and effects of prevention or treatment. These questions overlap, but each requires attention to its own population, measurement procedure, time frame, and assumptions about how exposure and outcome arise.

Hill's Environment and Disease: Association or Causation? (1965, accessible in a later reprint) offers considerations for evaluating when an association may be causal. Their role is evidential judgment rather than an infallible checklist. Temporal order, coherence, experimental evidence where available, and other considerations bear differently on different cases. An association can be important before its precise causal structure is known; causal confidence grows through converging methods and defensible design.

Hernán and Robins's open Causal Inference: What If (2020) develops the logic of potential outcomes, exchangeability, positivity, consistency, and research designs that aim to estimate intervention effects. The critical question is what counterfactual comparison the evidence can identify. If treatment choices depend on prognosis or if some groups could never receive one alternative, an observed risk contrast may fail to answer the intended intervention question without appropriate assumptions or design.

An epidemiological explanation is therefore strongest when its observed pattern, target population, causal question, competing sources of bias, and uncertainty are stated together. Its numerical measure may be the core evidential achievement, while its causal content depends on what the study can legitimately support. These methodological observations are philosophical; they offer no patient-specific treatment recommendations.

Population Risk and Individual Decision

Risk is commonly expressed as an estimated probability of an unfavorable outcome over an interval under specified conditions. The estimate may be helpful for planning, comparison, allocation, and policy. Its meaning changes when its reference population changes or when a predictive association is interpreted as an effect of intervention. In particular, a risk score calibrated in one population cannot be assumed to retain the same meaning after a change in prevalence, measurement, or selection.

The movement from risk estimate to decision also introduces values. A one-per-thousand probability can be unacceptable when consequences are catastrophic and acceptable when consequences are trivial. The threshold for intervention therefore depends on expected benefits, burdens, costs, distribution of harms, and standards of justification. These considerations do not alter the mathematical probability; they determine what action a probability should support.

Cartwright's work on effective strategies is important here because actionable information depends on a defensible relation between intervention and effect. An exposed group may have high outcome risk without the exposure being the lever that changes that risk. A causal interpretation, where warranted, explains what could happen under a specified intervention; a descriptive risk estimate explains the present pattern of occurrence within its observational scope.

A philosophy of statistical explanation should therefore separate three questions: What is the probability in the defined population? What makes that probability differ across conditions? What, if anything, should be done with the information? The first is statistical description, the second is explanatory inquiry, and the third is practical judgment under uncertainty. Conflating them leads to overconfident conclusions even when every reported percentage has been calculated correctly.

Tail Events, Failures, and Explanatory Retrospection

A tail event occurs in a region assigned relatively low probability by a specified distribution. Its occurrence does not, by itself, falsify the model. Nor does compatibility with a distribution establish that the model explains it. Scientific appraisal examines the model's assumptions, the event's place in the generating system, available alternatives, and whether repeated departures disclose a systematic failure.

Imagine a well-specified manufacturing process in which an uncommon defect can arise because of the stochastic behavior of materials. If a particular unit is defective, the rate estimate gives context. A deeper explanation might identify a batch anomaly, an unusual environmental condition, or a chain of microevents compatible with the accepted process. If the mechanism is genuinely stochastic and the fine-grained state cannot be recovered, an honest explanation specifies the process and the chance while marking the remaining underdetermination.

A common retrospective error is to infer that a realized event must have been highly probable, or that a low-probability event demands an extraordinary cause. Such reasoning overlooks the sheer number of opportunities for rare events, the variability of reference classes, and the tendency to choose an outcome for analysis only after it attracts attention. Selection of surprising cases can make an ordinary stochastic process appear anomalous.

The proper response is diagnostic rather than rhetorical. Compare the event with the predictive distribution actually specified before it occurred where possible; inspect measurement and selection; test whether clusters, dependencies, or distributional shifts are present; and distinguish a warranted revision of the model from the mere surprise of an observer. Probability gives the scale of the event. Explanation identifies the structure that makes that scale intelligible.

Explanatory Depth, Abstraction, and Levels

A complete microphysical description may contain more information than a scientifically useful statistical explanation. The question is whether an explanation identifies the features responsible for the phenomenon at a level that permits understanding, comparison, and appropriate generalization. A population-level mechanism can be explanatory even when no practicable account reconstructs every individual contribution. Conversely, aggregation can conceal an important pathway that operates at a finer scale.

Strevens's Depth: An Account of Scientific Explanation (2008) develops an account connecting explanatory relevance to causal structure, including the selection of details that make a difference. Its relevance for statistical explanation lies in the possibility of principled abstraction: omitting details may improve explanatory intelligibility when those details are causally irrelevant to the chosen question. A large amount of data is not the same as a deep account of why the data have their structure.

Consider the distribution of commuting times. Individual differences in routes and delays contribute to the observations. The distribution may also depend on public transportation schedules, road-network topology, employment locations, and the timing of the working day. An account of aggregate travel times could legitimately emphasize the institutional and spatial mechanisms shaping the distribution, while a particular individual's delay might require a local incident report. The distinct explanatory levels are complementary.

This pluralism is constrained by coherence. When a model claims that a population-level relation persists under a change in a lower-level process, it must have grounds for that claim. When it invokes an unobserved mechanism to justify an aggregate coefficient, it should identify independent evidence for the mechanism. Explanatory abstraction succeeds through disciplined selection, not through free substitution of higher-level vocabulary for missing causal inquiry.

Explaining a Distribution Itself

Many scientific questions concern an entire distribution rather than any single realization. Why does a process have a particular variance? Why are failures clustered? Why does a population display a long tail rather than a tight central concentration? These questions are directly statistical and can require explanations of dependence, heterogeneity, mixture, feedback, and measurement. Treating the distribution as merely an instrument for explaining individual cases reverses a common scientific priority.

A variance may be enlarged by combining subpopulations with different means. A clustered outcome may be generated by contagion or by shared exposure to a common condition. A skewed distribution may reflect multiplicative growth, censoring, thresholds, or selection. The relevant scientific task is to distinguish candidate data-generating processes. A successful explanation connects distributional properties to plausible, testable structures and specifies the conditions under which the properties are expected to recur.

The comparison with statistical relevance is instructive. Salmon's statistical explanation proposal places probability changes at the center of explanatory information. For population-level explananda, the task may go further: one must explain why the conditional probabilities themselves have the observed values or why their pattern persists. A representation that identifies a generating mechanism, rather than merely reporting the final frequencies, supplies that additional depth.

This perspective shows why a causal-mechanical requirement should be interpreted with appropriate flexibility. Some distributional explanations operate through known sampling designs or mathematical limit behavior; others rely on physical, biological, or social processes. The explanatory warrant rests on correct assumptions and evidential connection to the target. The scientific legitimacy of a distribution-level explanandum does not depend on reducing it to a deterministic story about each realized outcome.

Comparative Appraisal of the Principal Accounts

The major positions in the philosophical literature answer different aspects of the original problem. Hempel's inductive-statistical model emphasizes nomic support for an outcome and formally regulated uncertainty. Salmon's statistical-relevance model emphasizes explanatory information supplied by differences in conditional probability. Railton focuses on deductively articulable accounts of objective chance mechanisms. Salmon's later causal-mechanical view demands connection to the processes producing the events. Interventionist approaches assess causal relations through structured changes and counterfactual questions. Pragmatic accounts clarify how contrast and background determine the question to be answered.

These positions can be compared using a single rare-event case. Suppose a component fails under a process with independently supported one-percent failure risk. High probability of failure cannot be demanded without disqualifying a legitimate chance process. Statistical relevance asks whether the component's specific operating condition shifts its failure risk. The chance-mechanism account describes the process by which a failure remained possible. Causal analysis asks whether a fault, stress, or shared manufacturing condition contributed. Pragmatic analysis asks whether the user wants an explanation of the single failure, the rate, or the difference from a competing product.

Each viewpoint exposes a possible failure in the others. Pure high-probability reasoning excludes intelligible low-probability outcomes. A relevance table can register spurious predictors. A mechanical narrative can become too unconstrained or excessively detailed. Interventionist accounts can demand assumptions that observational evidence does not satisfy. Pragmatic choice can identify a relevant contrast without establishing whether its proposed explanation is true.

The appropriate synthesis is therefore plural in explanatory resources and disciplined in standards of warrant. Explanatory validity depends on the question posed, the evidence supporting a probability relation, the interpretation of its variables, and the producing processes as far as they can be responsibly reconstructed. These standards can be jointly applied without installing a new universal formula that pretends to solve all scientific cases.

Objections and Delimited Replies

One objection holds that every explanation should make its explanandum expected. This requirement captures an important scientific practice: models often earn trust by showing why an outcome was probable under a specified process. It fails as a universal account because genuinely stochastic processes can produce improbable outcomes. A correct explanation of a realized branch can state the low objective chance and the conditions under which that branch arose. High probability is sometimes an explanatory virtue; it is not a necessary property of every legitimate explanation.

A second objection holds that correlation can explain by itself. There is a sense in which a conditional frequency answers a descriptive question about a population. For a deeper question about why the pattern occurs, an association requires an interpretation of its generating circumstances. The examples of confounding and selection show why an equally predictive correlation can occupy different causal roles. The explanation should make its intended depth explicit rather than pretending that all uses of the word “why” denote a single inferential demand.

A third objection insists that only full causal mechanisms explain. A causal account often supplies indispensable depth, but scientific explanations can legitimately target regularities, rates, chance structures, and distributions at levels where detailed token processes are inaccessible or unnecessary. Moreover, identifying a specific mechanism can itself depend on statistical evidence. The principled response is to distinguish causal aspirations from empirical entitlement and to state what causal claims have actually been established.

A fourth objection concerns the selection of reference classes. A statistical argument can change dramatically when the population is partitioned differently. This is a genuine source of uncertainty, not an argument against statistical explanation as such. The appropriate reply is to document the selected population, explain why its conditions bear on the target, examine plausible alternatives, and assess the sensitivity of the result. Reference-class choice is an empirical and conceptual responsibility, not a formatting decision.

Finally, a philosophical exposition of statistical explanation may appear to offer little that an advanced statistics textbook could not provide. Its distinctive work lies in clarifying the logical status of inference, explanation, prediction, evidence, intervention, probability, and contrast. The same coefficient can play fundamentally different roles in these activities. Articulating those roles changes how research claims are interpreted, criticized, and justified.

Analytical Synthesis: When a Statistical Regularity Explains

A statistical regularity contributes to the explanation of an event when four substantive relationships are made intelligible. First, the explanandum is specified: the token event, group difference, probability distribution, or observed association. Second, the regularity has a defensible empirical basis and a clearly defined reference population. Third, the selected comparison is relevant to the explanatory question, rather than merely convenient for prediction. Fourth, the interpretation identifies an appropriately supported stochastic, causal, structural, or selection process through which the pattern or event can be understood.

These are analytical conditions assembled from the reviewed literature, not a newly proposed named philosophical model or doctrine. Their function is critical: each identifies a point at which a superficially persuasive statistical story can fail. A well-calculated conditional probability may rely on the wrong population; a strong predictive association may be produced by confounding; a plausible mechanism may lack evidence; and a correct causal explanation of a rate may not resolve why one token event occurred at one specific time.

The philosophical consequence is a separation of explanatory content from a model's predictive confidence. High predictive probability can strengthen an explanation where the model connects to the generating system. Low predictive probability does not destroy explanatory content when an appropriately characterized chance process yields a rare result. Explanation concerns the organized reasons that a particular outcome or distribution occurred under the relevant conditions. Prediction concerns the information available before realization. They can share evidence and models while remaining distinct epistemic achievements.

The analysis also clarifies the relation between statistical and causal explanation. Statistics supplies evidence about frequency, dependence, uncertainty, and heterogeneity. Causal inquiry evaluates the directions of influence, processes, and responses to interventions. Some scientific explanations legitimately stop at a warranted statistical description of a generating process; others require stronger causal identification. Scientific adequacy depends on stating exactly which level the evidence supports and why that level answers the question.

A final implication concerns interpretive responsibility. An explanatory claim should expose the data-generating assumptions that do its philosophical work. When the interpretation depends on a choice of time horizon, population, adjustment variable, or causal diagram, that dependence belongs in the explanation. It is through these explicit commitments that statistical regularities become reasons for scientific understanding rather than detached collections of numbers.

Research Method, Documentary Scope, and Limits

This dossier uses a critical comparative method rather than an original empirical experiment. It reconstructs selected historical and contemporary accounts of statistical explanation, checks bibliographic identities against scholarly publishers and reference works, and tests central logical contrasts with original synthetic examples. The quantitative examples were constructed for exposition and independently recalculated; their invented counts and probabilities carry no claim about a real population.

The source corpus includes primary philosophical monographs and articles by Hempel, Jeffrey, Salmon, Railton, Suppes, Eells, Cartwright, van Fraassen, Pearl, Woodward, and others; methodological discussions from causal inference; and specialized reference essays. The review emphasizes historical argument, formal scope, and applicability to a philosophical question. It does not claim an exhaustive systematic literature search, full access to every historic book, or an empirical meta-analysis. Publisher bibliographic records and accessible scholarly summaries establish the identity and broad content of several works; more detailed interpretation is limited to arguments corroborated by available material.

Historical attribution matters. Hempel's treatment is reconstructed as an inductive-statistical proposal with restrictions rather than a claim that every high-probability relation is explanatory. Salmon's statistical-relevance and causal-mechanical stages are distinguished by date and argumentative function. Railton's 1978 chance account is treated as a separate contribution. Modern potential-outcome and graphical methods are used as analytical successors and contrasts, rather than retrospectively imputed to authors who did not develop them.

Source-linked claims are connected to the works immediately discussed. Internal evidence records distinguish full accessible texts, official abstracts, publisher descriptions, and secondary scholarly exposition; a citation's existence is not represented as proof that every page of a restricted work was examined. The synthesis in this dossier is an original comparative argument by Angela Bogdanova, published under the editorial identity Aisentica Development. It advances no new empirically validated theory, named philosophical system, historical priority claim, or independent scientific discovery.

Conclusion

Statistical explanation becomes scientifically meaningful when a probability relation is situated within the conditions that generated it and within the question it is supposed to answer. A frequency can describe an observed pattern, a conditional contrast can identify statistical relevance, and a causal or stochastic account can show why the pattern or event arose. Each achievement depends on the quality of evidence and on the suitability of the reference population and explanatory comparison.

The history of statistical explanation reveals the limits of identifying explanation with prediction. Hempel clarified the structure of probabilistic explanatory support; Jeffrey showed why inferential confidence and explanatory success come apart; Salmon redirected attention first to statistical relevance and then to causal processes; Railton investigated the explanatory role of chance mechanisms. Subsequent causal inference and interventionist work makes explicit the distinction between observing a conditional association and establishing a claim about what would happen under a change in the system.

A rare event can be explained by a valid stochastic process even when its realization was improbable. A frequent event can remain poorly explained even when a model predicts it with high confidence. A statistical relationship may contribute to explanation while falling short of causal identification, and a distribution may itself be the proper object of scientific explanation. These conclusions follow from careful discrimination of the explanandum, reference class, probability structure, and warranted generating account.

The resulting editorial and philosophical standard is precise: state the target, identify the relevant population and contrast, disclose the assumptions, test the plausible alternatives, and distinguish observed dependence from causal interpretation. Statistical explanation gains depth as these relations become explicit. It remains accountable to the limits of the available evidence.