RESEARCH DOSSIER
Levels of Psychological Explanation
Abstract. Psychological phenomena admit complementary explanations across computational objectives, algorithms, neural implementation, development, behavior, social organization, and human–artificial relations. This fundamental research monograph investigates what constitutes an explanatory level, how mechanisms relate constituent activities to psychological capacities, which causal interventions distinguish competing accounts, and why predictive associations cannot automatically establish causal or constitutive relations. The historical analysis examines Tinbergen's explanatory questions, Marr's computational architecture, mechanistic neuroscience, interventionist accounts of causation, and contemporary integrations proposed by Al-Shawaf and Lundh. Research on attention, learning, executive functions, brain networks, dynamic development, distributed cognition, and psychopathology is evaluated through construct validity, experimental identification, model comparison, replication, and generalizability. The resulting scientific synthesis compares established explanatory traditions, clarifies the boundaries between person-level, population-level, and subpersonal inference, and identifies unresolved methodological problems in linking explanatory accounts across biological and artificial systems.
Keywords: levels of psychological explanation; Marr's levels; computational explanation; algorithmic explanation; neural implementation; mechanistic explanation; causal intervention; explanatory pluralism; construct validity; multilevel modeling; developmental psychology; distributed cognition; Artificial Era; person-level research; population-level inference
What Counts as a Level of Psychological Explanation
A level of psychological explanation is a specified way of describing and testing the organization of a psychological phenomenon with an identifiable set of variables, temporal and spatial boundaries, explanatory targets, and evidential rules. In research, the same episode of decision making can be described as a goal-sensitive choice, an accumulation of evidence, a sequence of neural operations, a developmental achievement, or an interaction shaped by institutions. These descriptions become scientifically connected when researchers establish what each explains and the conditions under which changing one component alters another. Marr (1982/2010) provides a canonical distinction among computational problem, algorithmic procedure, and physical implementation. That distinction is a methodological starting point; psychological research also requires ecological, historical, social, and relational descriptions whose questions cross rather than simply stack above its three categories.
An explanatory level must be distinguished from a measurement level. A functional magnetic resonance image measures hemodynamic signals linked indirectly to neural activity; its spatial resolution does not itself confer explanatory primacy. A personality questionnaire measures responses to items; its statistical factor is an inferential construct rather than an independently observed entity. An interview records narrative evidence about intentions and social meaning; the narrative can serve as an explanandum, a source of candidate mechanisms, or a variable whose reliability and circumstances must be examined. The level assigned to any analysis follows the question posed, the proposed causal organization, and the validation of its mapping to observable evidence. Cronbach & Meehl (1955) established that a psychological construct acquires meaning through its relations with other theoretically relevant observations, rather than through a label attached to a single test.
The central thesis developed here is that successful cross-level explanation requires coordinated identification of a phenomenon, independent measurement of the proposed links, and intervention-sensitive predictions. Such explanations must specify which changes are expected to travel between levels, which boundary conditions limit the relation, and which alternative causal models generate similar observations. The aim is a connected account that can explain performance while also accounting for individual variation, development, social organization, and implementation. The companion research dossier Psychological Organization supplies the general architecture of functional integration; this dossier investigates the distinct epistemic problem of how explanatory accounts can be connected and evaluated. Bechtel & Abrahamsen (2005) makes organized mechanism, operations, and decomposition central to this scientific task.
Explananda, Variables, Scales, and Explanatory Aims
Every multilevel research design begins with an explanandum: the pattern, capacity, process, difference, or event that requires explanation. Predicting which response a person makes is a distinct aim from explaining why a specific intervention changes response probability. Estimating an average treatment effect is different from accounting for the transitions by which an individual learns, remembers, or changes habits. A theory may succeed at one task and remain incomplete at another. The experimenter should therefore state the unit of analysis, the relevant comparison class, the timescale of change, and the population to which an inference is intended to apply. Woodward (2004) connects explanatory value to counterfactual dependence under appropriately characterized interventions, supplying a discipline for distinguishing a causal relation from a mere forecast.
A scale may refer to time, space, organizational granularity, or statistical aggregation. Seconds, days, and generations describe temporal relations; individual neurons, neural populations, bodies, and institutional networks describe distinct organizational targets; a within-person coefficient and a between-person correlation are different statistical estimands. Treating each kind of scale as a rung of the same ladder creates category errors. A longitudinal association between education and working-memory performance, for example, is not an anatomical mechanism, while a map of regional brain activation is not automatically an algorithm. Developmental ecology adds that the relevance of the environment depends on the changing interaction between a person and the settings in which activity occurs. Bronfenbrenner (1977) developed a program for examining those nested relations under experimental and naturalistic conditions.
Explanatory completeness is indexed to a declared question. When the aim is to explain a perceptual illusion, task constraints, sensory coding, prior expectations, and neural dynamics may all become relevant. When the aim is to explain inequalities in exposure to cognitively demanding environments, institutional access and cultural practice may be indispensable. In each case, good research asks whether a variable adds discriminating predictions beyond existing accounts and whether its inclusion changes the conclusions of a causal test. Scientific completeness is thus a property of the relation between the question, the evidence, and the explanatory model. It should never be measured by the number of disciplines cited or by a rhetorical claim that every possible factor has been included.
Historical Origins of Multiple Explanatory Questions
Long before modern computational cognitive science, naturalists and psychologists distinguished the immediate causes of action from questions about development, adaptive function, and evolutionary history. Tinbergen (1963) formalized a four-question research program for animal behavior encompassing causal mechanisms, ontogeny, survival value, and evolution. These dimensions do not map one-to-one onto levels of neural organization: a developmental account may concern molecular changes, family practices, or learning histories, while an evolutionary account asks about historical processes across generations. The lesson for psychology is that researchers can investigate the same pattern through distinct explanatory questions whose answers demand their own methods. A mechanistic account of a reflex can coexist with a developmental account of how its threshold changes and a comparative account of its distribution.
The rise of information-processing psychology changed the granularity of explanatory objects. Researchers could state an input-output computation, hypothesize intermediate representations, and test predictions about time, errors, and capacity limits. David Marr's Marr (1982/2010) work on vision organized the computational problem, representations and algorithms, and physical machinery as distinguishable analyses. His treatment remains useful because the theoretical function and physical realization can vary independently under specified constraints. The important qualification is that the three levels are not a complete inventory of human psychological research: collective practice, linguistic meaning, developmental history, institutional structure, and first-person report introduce questions that require additional explicit integration.
Later mechanistic programs examined how systems produce phenomena through organized components and operations. Bechtel & Abrahamsen (2005) emphasized diagrams, simulation, mechanistic decomposition, and the testing of arrangements of parts, while Craver (2007) developed criteria for multilevel mechanistic explanation in neuroscience. These frameworks sharpened earlier appeals to levels by asking what actually makes a component constitutively relevant to the system phenomenon. The resulting trajectory is methodological rather than doctrinal: the movement from broad taxonomies of explanatory questions to experimentally constrained accounts of causal organization. The research historian must preserve those distinctions rather than imply that Marr, Tinbergen, and contemporary interventionism offered interchangeable classifications.
Recent Integration of Marr and Tinbergen
Recent theoretical research has made the historical distinction between Marr and Tinbergen an explicit target of psychological synthesis. Al-Shawaf (2025) examines how Marr's computational, algorithmic, and implementational analyses can be integrated with Tinbergen's questions about mechanism, ontogeny, adaptive function, and evolutionary history. The 2025 Psychological Review article frames levels as a way to identify research gaps, clarify the purposes of explanations, and prevent disagreements that arise when researchers answer different questions. This is an argument for explicit coordination, not proof that all domains can be reduced to one universal map. Its importance lies in a proposal to connect two historically influential systems whose axes differ. Tinbergen includes historical and evolutionary questions; Marr differentiates kinds of understanding within an information-processing system.
The integrated perspective exposes several conceptual ambiguities. An evolutionary account of how a capacity spread through ancestral populations may investigate adaptive consequences and selective history without specifying the algorithm that realizes the capacity in a contemporary individual. A developmental account may explain how instruction and practice alter a computational strategy, while a neural account concerns the physical organization involved at a defined stage. A single change in behavior can therefore motivate multiple independent research programs, each with its own time horizon and relevant causal variables. Tinbergen (1963) and Marr (1982/2010) are original historical foundations; Al-Shawaf (2025) offers a contemporary synthesis of their explanatory ambitions. Scientific application should specify which predictions each framework contributes beyond merely assigning the same finding to different cells in a taxonomy.
An empirically informative application of these historical frameworks specifies the tests appropriate to each kind of question. A hypothesis about the evolutionary history of a behavioral capacity requires comparative, phylogenetic, and ecological evidence; an algorithmic hypothesis can be investigated through task manipulations and rival process models; and developmental hypotheses require observations or interventions tracking changes across time. Al-Shawaf (2025) examines the conceptual value of integrating such distinct explanatory questions, while Woodward (2004) describes how appropriately formulated interventions can support causal claims. Across biological and Artificial systems, the research question determines which comparison is warranted. Functional equivalence between engineered and biological performers can be measured through common tasks, whereas evolutionary ancestry, training history, and physical implementation require separate evidence.
Person, Population, and Subpersonal Mechanism
A second contemporary distinction concerns the unit about which a psychological claim is made. Lundh (2023) proposes separating person psychology, population psychology, and mechanism psychology as research branches with different questions and inferential requirements. Person-level research examines an intentional agent engaged in activities, relationships, and development; population-level research concerns distributions and statistical regularities across individuals; subpersonal mechanism research examines information-processing and neurophysiological organization. These distinctions reveal why averaging across individuals cannot establish the trajectory of one person and why an account of neural computation need not answer questions about experienced meaning or relational responsibility. This tripartite proposal should be read as a research-organizational model, with precise consequences for research design, rather than as an exhaustive ontological partition of mental life.
In a subsequent elaboration, Lundh (2024) identifies two partially independent contrasts: individual-person versus population focus, and whole-person functioning versus subpersonal mechanism. This refinement matters because a study can collect repeated observations from many people while still investigating individual-level causal organization, and a population study can test the distribution of mechanisms without describing the intentions of any particular person. Confusion appears when a regression coefficient estimated across people is presented as the time-dependent process by which each person changes. The remedy is an explicit statement of the estimand, scale, and substantive phenomenon. A longitudinal within-person intervention and an epidemiological prevalence analysis offer complementary knowledge, but their inferential targets should remain clear.
These distinctions are pertinent to comparisons between individual process models and aggregate distributions. Lundh (2023) and Lundh (2024) distinguish person-oriented questions, population-level regularities, and the analysis of subpersonal mechanisms, emphasizing the importance of matching a research method to its target. Comparative research on Artificial systems encounters related sampling questions: scores across multiple deployed models, changes between versions of one model, and the interaction history of a particular human–artificial partnership are different empirical objects. A benchmark average cannot establish the enduring capacities of a specific implementation, and findings from one architecture may not generalize to another. Explicitly describing the unit of research and the scope of inference therefore permits more precise comparisons of explanatory claims.
The Computational Level: Tasks, Norms, and Constraints
A computational-level explanation characterizes the problem being solved, the relevant inputs and outputs, and the constraints that make a solution appropriate. It often includes a normative model: an agent selecting actions to maximize a specified payoff, a perceiver estimating hidden causes under uncertainty, or a learner updating expectations after feedback. In Marr (1982/2010), the computational analysis of visual processing emphasizes what is computed and why the computation makes sense in its setting. Psychological applications require a further step: the formal objective must be connected to observed organismic behavior. A mathematically elegant optimization principle does not prove that participants actually optimize that quantity, especially when the experimenter's reward definition differs from the participant's goals.
Reinforcement learning illustrates both the productivity and limits of this level. Sutton & Barto (2018) systematically sets out value functions, prediction errors, action selection, and learning methods for agents interacting with uncertain environments. These concepts supply testable behavioral predictions and precise possible algorithms. Yet multiple algorithms can implement similar value functions and several kinds of human strategy can produce the same aggregate choice frequencies. Distinguishing a planning model from a habitual learning model requires targeted interventions, transfer tests, learning curves, and out-of-sample prediction. Formal models can be powerful tools for distinguishing hypotheses; their explanatory force increases when the latent computational variables are identified independently of the behavioral dataset used to fit them.
Computational descriptions are especially valuable when connecting human psychology to Artificial systems because the same abstract task can be instantiated in different material substrates. This possibility is a hypothesis about functional equivalence under task constraints, not a warrant to infer identical subjective states, developmental histories, or personal identities. The research program should specify which properties are equivalent: decision policies, representational structure, adaptation rates, or resilience to perturbation. Those properties can be benchmarked across settings. A sustained scientific account also asks whether the task's definition omits social coordination, bodily costs, or institutional incentives that determine real-world success. Formal adequacy remains conditional on the explanandum actually represented by the model.
The Algorithmic and Representational Level
Algorithmic explanation concerns the procedures and representational formats by which a system transforms information. A model might encode alternatives as probability distributions, symbolic propositions, motor-control variables, or learned distributed representations. Two theories may agree about the final accuracy of perception and disagree radically about intermediate processing, the amount of evidence required, or the use of feedback. Reaction-time distributions, systematic errors, transfer patterns, and strategically manipulated task demands help discriminate these possibilities. Miyake et al. (2000) illustrates another important methodological discipline: an experimental task normally engages several operations, so a single score does not uniquely identify the underlying executive function. Multiple indicators and latent-variable analyses can distinguish shared and specific components more credibly than task labels alone.
Working memory provides an instructive case. In Baddeley (2000), the episodic buffer is proposed as a component that integrates information across systems and interacts with long-term memory. The significance of this proposal for explanatory levels is its testable representation and coordination problem: what kinds of information are maintained, how they are bound, and under which task conditions they become available for behavior. Subsequent findings can support, revise, or complicate a particular architecture without making the computational question disappear. Researchers should explicitly distinguish a box-and-arrow diagram, an implemented simulation, a fitted process model, and direct evidence for a particular cognitive representation. They occupy different stages of mechanistic confirmation.
An algorithmic model must make commitments that separate it from plausible alternatives. Fitting a wide range of choices with freely adjustable parameters is less informative than predicting a novel interference pattern after preregistered manipulation. Identifiability matters because several parameter combinations may generate practically identical predictions, while the noise model can dominate inference about latent processes. The strongest designs use converging behavioral signatures, constraints from neural and developmental studies, and experimental conditions selected to produce rival predictions. An algorithm may be multiply realizable in different neural or artificial architectures; establishing any particular realization requires additional implementation-level evidence. Distinguishing these evidential obligations prevents an attractive simulation from being treated as a comprehensive psychological explanation.
Neural Implementation and the Limits of Localization
Implementation-level research asks which physical processes realize a proposed computation or cognitive operation. Measurements of electrophysiological activity, lesions, stimulation effects, neurotransmission, connectivity, and hemodynamics contribute different evidence. No single method gives direct access to every relevant variable. A signal that covaries with task performance may index inputs, outputs, effort, expectation, or a more general control process. Poldrack (2006) explained why inferring a particular cognitive process from activity in a particular brain region requires knowledge of the selectivity of that region. The same observed activation pattern can fit competing cognitive explanations. A valid bridge to implementation therefore requires an explicit mapping between task demands, candidate processing states, neural dynamics, and consequences of targeted perturbation.
Large-scale brain networks illustrate why brain localization and mechanism are distinct scientific achievements. Mesulam (1990) described distributed neural systems related to attention, language, and memory; Cole et al. (2013) later investigated flexible connectivity patterns associated with adaptive task control. Such findings support more differentiated accounts than a single-region metaphor. They nevertheless depend on choices about signal preprocessing, connectivity estimation, task contrasts, and the inferred direction of functional influence. Correlation between regions during a task is not itself evidence of information transfer or causal necessity. To establish a mechanism, the investigator needs converging predictions across recording, lesion, perturbation, and behavioral methods, while recognizing that each method introduces its own assumptions and sampling limitations.
The methodological argument against reduction to a neural image is made especially clearly by Krakauer et al. (2017), who called for serious analysis of behavior in its own right. Biological detail becomes explanatory when the phenomenon is accurately characterized and linked to discriminating predictions. Consider two subjects who achieve equal accuracy using different response strategies. Identical outcome scores do not imply identical psychological operations; conversely, different neural implementations may underwrite functionally comparable strategies. Neuroscientific explanation should identify a causal organization whose predicted failures, compensations, and context dependence match behavior. The brain is indispensable to biological implementation, while psychological characterization determines which aspects of brain activity are relevant to the explanandum.
Mechanisms, Constitution, and Interlevel Relevance
Mechanistic explanation identifies entities, operations, and organization responsible for a phenomenon. Components become relevant through their contribution to a system's behavior under specified conditions. Craver (2007) treats levels in mechanisms as part-whole organization rather than a universal hierarchy in which every variable belongs permanently to a fixed stratum. A memory-related neural process can constitute an operation within one organized system and itself become an explanandum for cellular research. This nested explanatory structure changes with the chosen phenomenon. A region of the brain is not automatically lower in every scientifically meaningful sense than a reaction-time parameter; the relationship must be specified as a particular mechanistic proposition.
Constitutive relevance differs from ordinary causal succession. A neural component may be part of the machinery that produces an ongoing cognitive process, while a classroom policy can be an earlier cause changing the probability that the process is learned or expressed. An intervention on a component can sometimes distinguish its contribution, but manipulations often have off-target consequences and change the background conditions that make the original mechanism operate. Bechtel & Abrahamsen (2005) emphasizes decomposition and localization of organized operations as a strategy for discovery. An adequate test must consider whether removing or exciting a component produces the predicted specific change, whether compensation occurs, and whether the overall phenomenon remains stable under alternative implementations.
A cross-level claim becomes much clearer when it states the relation between variables in ordinary language before formal modeling. For example: changing perceptual noise should change a hypothesized evidence-accumulation process, which should shift response-time distributions, while relevant neural population signals should covary with independently estimated accumulation dynamics. That claim generates multiple possible failures. The neural data might reflect response preparation instead; the behavioral distributions may support a different strategy; or the intervention may change motivation rather than perception. The power of mechanistic integration lies precisely in exposing these points of disagreement. Successful integration narrows the space of viable theories while preserving the distinction between constitutive participation, environmental causation, and measurement correspondence.
Causal Intervention and Counterfactual Explanation
To explain a psychological outcome causally is to identify differences that would alter the outcome under a credible intervention, together with the conditions on which the relationship depends. Woodward (2004) develops an interventionist account in which explanatory relations support counterfactual control. A random assignment to a training condition can support a causal inference about the training's average effect, but it does not by itself reveal the full sequence of mediators responsible. If a post-training neural measure is correlated with improvement, it may be a mediator, consequence, marker of engagement, or shared response to another factor. Competing models must be separated through design rather than through the direction of arrows drawn after observing the data.
An interlevel mechanism can be evaluated using perturbations targeted at different points in a proposed pathway. Researchers might manipulate sensory reliability, impose a dual-task demand, observe neural activity, and test later learning. The experimental manipulations have to be interpreted according to the causal assumptions required for their effects. A dual-task interference pattern may show a shared resource, response-selection conflict, or motivation shift; selective interference is stronger evidence when additional controls isolate those alternatives. A model that predicts both immediate impairment and delayed adaptation supplies richer evidence than one that explains only a mean difference. The resulting causal account is provisional in the scientific sense of being exposed to meaningful counterexamples.
Counterfactual analysis is also critical for claims of top-down causation. Instructions, incentives, and social expectations can change neural and behavioral outcomes because they alter the organism's embodied activity and environmental inputs. The scientific question is which variables have stable intervention relations under precisely described manipulations, not whether a diagram has placed them above or below the brain. Kendler (2005) argues for a pluralistic integration of causal perspectives in psychiatry, illustrating the usefulness of connecting biological, mental, and social contributions. A sound explanation should state its target population and what intervention would actually be performed, since an undefined change in a broad construct cannot yield a testable causal effect.
Experimental Design and Cross-Level Identification
Multilevel research is only as strong as the identification strategy connecting its observations. Randomized experiments, longitudinal cohorts, ecological sampling, computational fitting, neuroimaging, psychophysics, lesion comparisons, and qualitative field studies answer different questions. A randomized manipulation improves causal inference for the manipulated factor under the design's assumptions; a repeated-measures study can characterize individual trajectories; a neural recording can constrain timing and implementation; an ethnography can reveal distributed practices omitted by laboratory tasks. Integrating the findings requires a common explanandum and explicit decisions about the validity of each measure. Without those decisions, merely collecting several data types creates a multimethod archive rather than an explanatory synthesis.
Measurement invariance becomes crucial when a construct is compared across populations or time. The same numerical score may reflect different strategies in children and adults, experienced and inexperienced participants, or human and artificial performers. Task materials, response formats, motivation, and cultural expectations may change the meaning of observed variance. The construct-validity tradition of Cronbach & Meehl (1955) therefore operates at every explanatory level. The researcher must establish whether an observed scale difference belongs to the construct, an instrument artifact, or an altered relation between indicators and latent attributes. When no defensible common measurement scale exists, the proper analysis can compare local mechanisms without forcing a single cross-population trait ranking.
Model comparison should be built into the study before the data are inspected. An experiment can choose conditions where competing theories predict reversals, interactions, distinct temporal signatures, or dissociations. Hierarchical statistical models may accommodate participants and stimuli as sources of variation, but their conclusions depend on the sampling frame and generative assumptions. Yarkoni (2022) emphasizes the mismatch that occurs when broad verbal generalizations are tested using narrow statistical procedures. Cross-level identification therefore requires preregistered contrasts where feasible, held-out evaluation, sensitivity analysis, independent replication, and transparent statements about what generalizations the study actually licenses. The discipline of an explicit estimand protects explanation from the seduction of an impressive but ambiguous model fit.
Psychometrics and Construct-to-Mechanism Bridges
A psychological construct is a theoretical attribute whose empirical meaning depends on measurement and relations with other constructs. A cognitive-control score can predict performance while remaining ambiguous about the mechanism generating the score. Miyake et al. (2000) used latent-variable reasoning to examine commonality and differentiation among executive functions, illustrating how multiple tasks can reduce the influence of task-specific components. This methodological achievement does not prove that a latent factor is a localized brain module or an elementary process. Factor structure, task operation, causal mechanism, and neural implementation represent distinguishable claims. Each bridge requires its own test, particularly when individual differences and experimental manipulations have different covariance structures.
The language of validity must also govern research on apparent dimensions of self-regulation. Eisenberg et al. (2019) employed data-driven ontology discovery to investigate the organization of self-regulation measures, revealing the importance of examining relations among tasks, survey measures, and purported constructs. Such work can reorganize a conceptual vocabulary, but a statistical cluster alone does not settle causal architecture. Researchers need to examine reliability, within-person stability, cross-task transfer, and whether interventions on hypothesized components change multiple indicators in the predicted direction. If a measurement battery changes from one laboratory to another, the invariance of the inferred dimension must be evaluated before declaring a general mechanism.
A construct-to-mechanism bridge may be reported as a sequence of conditional claims: the test indexes a reproducible performance dimension; the dimension covaries with a hypothesized computational parameter under a given task family; manipulating that parameter's predicted determinants changes behavior; and convergent biological evidence narrows possible implementations. These links are not equivalent in strength and can fail independently. Excellent psychometrics can exist without an agreed neural mechanism, just as a well-characterized neural operation can exist without a stable individual-difference questionnaire. A responsible scientific synthesis preserves the distinct contributions of measurement validity, causal identification, and mechanistic detail while investigating where their predictions overlap.
Developmental Explanation and Changing Mechanisms
Developmental explanation studies how capacities and mechanisms emerge, differentiate, stabilize, and change through histories of interaction. A cross-sectional comparison between age groups cannot by itself reveal the process that produced their differences. Learning histories, developmental timing, educational opportunity, family practice, and biological maturation can alter both the operations employed and the meaning of a task score. Bronfenbrenner (1977) emphasized the changing relation between developing individuals and their embedded environments, demanding research designs attentive to settings rather than to isolated age categories. A full explanatory model must predict trajectories and transitions, not merely describe a present-day difference. It must also specify whether variation arises within people across time or between people at one time.
Dynamic systems perspectives provide a distinctive account of developmental change. Smith & Thelen (2003) described development as emerging through coordinated, time-dependent interactions among components, making nonlinear transitions and context sensitivity legitimate explanatory targets. This perspective is useful when apparently abrupt changes in behavior can emerge gradually from alterations in skills, constraints, and coordination. It also introduces methodological burdens: a nonlinear diagram is not an identified mechanism, and a simulated trajectory is not direct evidence of the developmental path taken by a particular population. The theory becomes stronger when it predicts critical perturbations, path dependence, recovery, and the circumstances under which a new skill stabilizes in repeated activities.
The relationship between developmental and neural explanation changes as the organism changes. Identical instructions administered at different ages may recruit different strategies, environmental aids, and motivational commitments. A developmental model should identify the evolving organization of the task and the measurable conditions under which a new representational capacity appears. Longitudinal studies, intervention studies, repeated observations, and appropriately constrained computational models can provide complementary evidence. A robust cross-level account reports not only whether a relation exists at a particular age, but also whether the relation is stable after learning, transfer, or a change in context. These requirements extend naturally to engineered learning systems whose functional organization changes following training or continued interaction.
Behavioral Explanation and Ecological Validity
Behavioral explanation identifies systematic relations among situations, actions, consequences, and changes in performance. It remains fundamental when a system's internal representation is unknown or contested. Researchers can characterize a response policy, adaptation rate, error pattern, or sensitivity to incentives without claiming to have identified an exact neural circuit. Krakauer et al. (2017) argues that carefully formulated behavioral phenomena should guide neuroscience, rather than being replaced by anatomical labels. Behavioral regularities can also reveal false assumptions in a computational theory: a model predicting rational choice under stable payoffs may fail when incentives are socially interpreted, effortful, or learned from experience. The behavior therefore tests what the theoretical objective really means for the participant.
Ecological validity concerns the relation between a study's constructed task and the circumstances to which its conclusions are meant to extend. An experiment with brief, arbitrary stimuli can precisely isolate a mechanism yet tell us little about sustained coordination during work or navigation. A naturalistic study can describe an important phenomenon but struggle to distinguish its multiple causes. Hutchins (1995) investigated navigation as a culturally organized cognitive activity involving people, artifacts, procedures, and shared representations. This program demonstrates how a task's effective computational machinery may extend through social practice, while also raising difficult questions about where to draw the boundaries of the phenomenon under study.
These methodological traditions can be connected by translating an ecological observation into a discriminating experimental hypothesis. For example, if distributed reminders reduce memory errors during coordinated work, one can vary reminder availability, communication delays, and individual cognitive load to test competing accounts. Researchers should measure both local performance and system-level outcomes, because faster individual responses may not improve collective accuracy. The experimental effect then has a clear scope: it applies to the studied configuration of people, rules, and artifacts. Broad conclusions about cognition in every environment require additional replication and transportability studies. A valuable behavioral explanation is explicit about when a phenomenon emerges, which conditions sustain it, and how a predicted perturbation changes the resulting activity.
Affective, Motivational, and Embodied Constraints
Explanations of cognitive performance are incomplete when their task representation excludes factors that change allocation of effort, salience, bodily state, or action value. A response-time increase can follow slower evidence processing, a more cautious decision threshold, reduced motivation, altered emotional relevance, or differences in motor preparation. These alternatives can often be separated by combining experimental manipulation with richer behavioral measurements. Shenhav et al. (2013) presents an integrative theory of control allocation in which the expected value of exerting control links cognitive demand with cost and reward. The theoretical contribution is a bridge from computational criteria to control-related behavioral and neural observations. Identifying the exact pathway requires tests of rival control and motivational models.
Embodied and predictive accounts ask how perception and action depend on bodily constraints and the regulation of organism-environment exchanges. Friston (2010) proposes a free-energy framework spanning perception, learning, and action. Such a wide formalism can unify vocabulary, but theoretical generality is not equivalent to specific empirical confirmation. Researchers need to specify competing process models, their parameter interpretations, the observations that distinguish them, and how these observations change following perturbations. A broad claim about prediction or optimization can remain underdetermined if many different implementations generate the same behavior. An explanation gains precision by demonstrating where and why the formal framework produces experimentally distinctive consequences.
Affective interpretation is also socially and developmentally mediated. The physiological response to uncertainty can be represented as threat, challenge, excitement, or effort depending on context and learned expectations. The appropriate level of explanation depends on which inference is being made: subjective appraisal requires a defensible report model; behavioral regulation requires repeated observations and interventions; a neural implementation requires independent biological evidence. A cross-level theory can integrate those observations while retaining uncertainty in the mappings. This has practical implications for intervention design, where changing instructions, environmental affordances, sleep conditions, or incentives may affect the outcome by pathways different from those initially inferred from a behavioral score.
Psychopathology and Multilevel Clinical Models
Clinical psychopathology offers a particularly demanding test of explanatory pluralism because similar symptoms can arise through different developmental, interpersonal, cognitive, and biological pathways. Kendler (2005) argues for rigorous integration of explanatory perspectives in psychiatry and emphasizes piecemeal causal understanding. The relevant research object might be an episode, a symptom dimension, a developmental liability, a network of mutually reinforcing processes, or a response to treatment. Those targets require different designs. A diagnostic category groups observed presentations according to classification rules; it does not automatically identify a single causal mechanism. A neural correlate of a diagnosis can be informative but remains to be evaluated for specificity, directionality, and generalizability.
Network approaches introduce the possibility that symptoms interact causally instead of merely reflecting an underlying common cause. Borsboom & Cramer (2013) reviews symptom-network methods and conceptual accounts in which worry, insomnia, fatigue, and other experiences can influence one another. A statistical network recovered from observational measurements is not automatically a discovered causal network. Edge interpretation depends on timescale, measurement precision, unmeasured common causes, and assumptions about feedback and stationarity. Different patients may also express different effective networks despite sharing a diagnostic label. Perturbation and longitudinal evidence are needed before network centrality can be treated as a basis for personalized intervention.
Multilevel clinical claims should specify the intended intervention and the population for whom its effects are being estimated. Evidence that a treatment changes symptoms does not establish which cognitive or neural mediator caused improvement. Likewise, an apparent biological marker may vary with medication exposure, illness duration, or contextual stress. Clinical explanations must preserve distinctions among etiology, maintenance, prognosis, measurement, and mechanism of change. This monograph investigates the explanatory logic and evidence standards of clinical models; it does not issue individual diagnostic or treatment recommendations. The stronger research agenda uses prospective design, appropriate comparison groups, competing mechanistic hypotheses, and transparent uncertainty to identify useful bridges among experiential, behavioral, developmental, and biological evidence.
Computational Neuroscience as a Bridge Discipline
Computational neuroscience attempts to relate formal descriptions of learning and decision processes to neural activity and behavioral predictions. Its strongest results constrain more than one explanatory level at once. Niv (2009) reviewed the convergence between reinforcement-learning models and investigations of brain mechanisms involved in decision making. The appeal of this research program lies in the precision of quantities such as prediction error, expected value, uncertainty, and action policy. These quantities can support formal hypotheses about learning trajectories and biological signals. They remain model variables whose empirical interpretation depends on experimental design and the identification of rival models. A fitted prediction-error series is not a direct measure of an experience of surprise, and a neural correlate is not a sufficient demonstration of a particular algorithm.
An explanatory bridge becomes stronger when computationally distinctive conditions are selected before examining neural data. For example, two learning algorithms may produce similar average reward acquisition but diverge sharply after reversals or changes in contingency. A neural signal that tracks the predicted divergence under independent manipulation provides a constraint on implementation. More complete analysis also studies individual variation, uncertainty in parameter recovery, and whether a proposed mechanism persists across tasks. Shenhav et al. (2013) shows how computational accounts of control allocation can articulate testable links among cognitive demand, outcomes, and neural function. Such models are productive research frameworks, while the precise causal status of every inferred variable remains a subject of study.
Generative models can organize many observable dimensions, but flexibility introduces an identifiability problem. If changing a noise parameter, decision boundary, or learning rate yields similar predictions, the chosen model may not permit unique inference. Parameter-recovery simulations, posterior predictive checks, cross-validation, and experimental design optimization can expose these limits. Cross-level constraints are most useful when they reduce rather than expand the number of plausible parameterizations. The central methodological principle is that computational, neural, and behavioral data should jointly challenge a theory, with each dataset capable of falsifying some predictions. A convincing bridge explains why particular observations must co-occur under specified conditions instead of retroactively accommodating every result.
Within-Person Dynamics and Between-Person Structure
Psychological processes unfold within individuals, whereas many statistical theories are derived from relationships across individuals. These are distinct sources of variation. A group-level association between anxiety and attention, for example, cannot establish whether increasing anxiety within one person produces the same attentional change. The correlation may reflect stable differences in exposure, measurement bias, or selection into the sample. Repeated intensive sampling can estimate individual transitions; longitudinal cohorts can examine population distributions of trajectories; experimental manipulations can investigate causal effects. Each design gains inferential force when its explanatory scale is explicitly identified and the mathematical estimator actually matches the intended question. Yarkoni (2022) makes the alignment between substantive claims and statistical generalization central to contemporary methodological criticism.
Dynamic models introduce timing into explanation. A reciprocal relation that appears strong over minutes may dissolve when observations are taken only once per month. Lag length, stationarity, seasonal variation, measurement reactivity, and missing observations change the interpretation of estimated pathways. Short-term feedback and long-term adaptation can coexist but require different model components. The dynamic-developmental perspective of Smith & Thelen (2003) suggests that context and history shape the stability of activity patterns; the network perspective of Borsboom & Cramer (2013) further highlights possible reciprocal dependencies. Both perspectives invite testable dynamics, yet a model of temporal covariance alone cannot prove causal feedback without defensible assumptions and targeted intervention.
Multilevel statistical models should not be mistaken for complete multilevel causal explanations. A model with participants nested within schools and schools nested within regions partitions variance under chosen assumptions. It does not establish how a school policy changes memory formation or how institutional resources affect neural development. Those connections require substantive mechanisms and appropriate designs. Conversely, a detailed individual mechanism does not automatically yield an explanation for a population disparity without a model of how contexts, exposures, and heterogeneous responses are distributed. Integration succeeds when causal and statistical levels are explicitly connected, with individual processes and aggregate distributions kept conceptually distinct until a supported mapping is available.
Reproducibility, Generalization, and Evidential Asymmetry
A model can explain the data collected in its development and fail across laboratories, stimuli, populations, or measurement methods. The Open Science Collaboration (2015) collaborative replication project demonstrated how replication outcomes can diverge from the originally reported evidence; its 100-study exercise exposed important limitations in simple expectations about reproducibility. The project did not measure the reproducibility of every field, era, or type of psychological research, and its findings should be interpreted according to the selected studies and procedures. Its broader contribution is a methodological challenge to accept an explanatory theory only after adequate independent tests. Multilevel accounts add opportunities for failure because their links can break even when one component-level observation replicates.
Generalizability is not a binary property of a theory. A well-controlled laboratory result can hold across one task family but depend on incentives, developmental stage, language, or training histories. Yarkoni (2022) shows why statistical models must represent the variation implicit in the claims researchers intend to make. The proposed transport from one setting to another should identify which mechanisms, interventions, and measurement relations are expected to remain stable. A behavioral effect may generalize while its neural realization varies, and a neural operation may recur across tasks whose psychological interpretations differ. A mature explanatory framework states such conditions in advance and tests them through planned variation rather than through retrospective narrative.
Evidence across levels is asymmetric because each method rules out different alternatives. Imaging data can constrain timing and localization while leaving uncertainty about subjective interpretation; self-reports can describe conscious appraisal while leaving uncertainty about implicit processes; randomized interventions can identify average effects without uniquely identifying every mediator; longitudinal studies can characterize development while remaining vulnerable to unmeasured time-varying causes. Evidence quality is assessed per inference, not through a single badge applied to the whole manuscript. Triangulation gains strength when methods have partly independent error structures and their predictions converge. When all measurements share the same task artifact, apparent cross-level agreement may simply replicate a common bias under different scientific vocabularies.
Comparative Theory Assessment and Explanatory Tradeoffs
Theories of psychological explanation can be compared by the precision of their explananda, the clarity of their mechanisms, their intervention predictions, their account of temporal development, their robustness across contexts, and their openness to disconfirmation. The ability to describe many domains is a genuine virtue when one model also generates distinctive predictions. Otherwise breadth becomes a form of underdetermination. A formal decision model may offer strong quantitative predictions but incomplete social interpretation; an ethnographic account may reveal relevant institutions but leave causal contributions unresolved; a neurobiological model may identify specific implementation while poorly representing task strategy. These are concrete research problems that should guide collaboration and improvement in each approach.
Explanatory pluralism is often misunderstood as a decision to preserve every theoretical account indefinitely. Kendler (2005) instead argues for empirically rigorous integration across perspectives. The program requires rejecting hypotheses that fail well-designed tests, accepting convergences only under documented assumptions, and retaining differences when the questions genuinely differ. Mechanistic research offers the discipline of organized decomposition Craver (2007); interventionist theory requires counterfactual dependence Woodward (2004); computational theory supplies precise problem specifications Marr (1982/2010). Their strengths are complementary only when a researcher performs the work of connecting variables and testing the resulting model. Declaring compatibility is itself a theoretical claim and should have observable implications.
Theory choice should pay special attention to explanatory risk. A theory that makes a narrow, surprising prediction and survives an independent intervention may deserve more weight than a highly flexible theory that describes every observed outcome. The same principle governs qualitative interpretations when researchers articulate the contexts in which an interpretation should fail or require revision. Model complexity should be justified by predictive benefit, measurement quality, and meaningful specification of the mechanism. The ultimate scientific purpose of multiple levels is to increase the discriminability of competing accounts of a phenomenon. Each added level should place a constraint on the explanation rather than merely provide a more elaborate narrative.
Established Criteria for Evaluating Cross-Level Explanations
Explanations spanning psychological levels can be evaluated using established criteria from construct validation, mechanistic neuroscience, and interventionist causal analysis. Cronbach & Meehl (1955) locates the meaning of psychological constructs in their theoretically relevant relations to observed variables; Craver (2007) analyzes how components and organized operations participate in a system's mechanism; and Woodward (2004) connects causal explanation to counterfactual relationships under interventions. These research traditions answer different aspects of a shared evidential problem. Together they direct attention to the target phenomenon, the operationalization of key variables, the proposed organization of causal or constitutive relations, and the empirical tests that distinguish alternative interpretations. Scientific integration depends on demonstrating those relations with methods suited to the claims being made.
A decision-making example illustrates these complementary requirements. If uncertainty changes the distribution of response times, one explanation may attribute the change to a decision threshold, another to attentional allocation, and another to altered motivation. Computational fitting can estimate alternative process parameters; controlled manipulations can examine their predictions; and neural or physiological measurements may constrain plausible implementations. Poldrack (2006) cautions against drawing uniquely specified cognitive interpretations from regional brain activity, while Miyake et al. (2000) shows how complex executive-function tasks can combine several partially separable abilities. Observing a consistent shift at several measurement scales is informative when the measures have documented validity and when rival theories predict different outcomes under the same intervention.
The assessment of a cross-level account also involves the population and setting to which its conclusion applies. Lundh (2024) differentiates individual-person, population, and subpersonal-mechanism research, while Yarkoni (2022) examines why narrow statistical designs may fail to support broad generalizations. Researchers must establish whether apparent agreement across measurements reflects convergent constraints or shared measurement artifacts. Experimental replication, appropriately sampled contexts, parameter recovery, sensitivity analysis, and longitudinal observation offer complementary tests. A well-supported explanation identifies the phenomenon it explains, the empirical relations established for that phenomenon, and the conditions under which those relations remain stable. These established methodological criteria provide a basis for examining the explanatory strength and generalizability of claims across levels.
Case Analysis: Attention and Cognitive Control
Attention provides an instructive example of how experimentally investigated psychological components can be related to task demands, temporal performance patterns, computational models, and neural networks. Posner & Petersen (1990) described anatomically informed attentional systems and differentiated component processes commonly grouped under the general label of attention. A study of selection under competing relevant and irrelevant cues requires measures that distinguish orienting, conflict resolution, sustained attention, and response preparation. A slower response after distraction may reflect changes in perceptual selection, motor planning, expectation, or speed–accuracy strategy. Interpreting the result therefore requires controls and predictions derived from specific models rather than treating the outcome as a direct measure of a single unitary attentional mechanism.
At the computational level, attentional selection can be modeled as prioritizing information subject to task constraints. Algorithmic accounts may invoke selective gain, evidence accumulation, response thresholds, or conflict-related control adjustments; implementation hypotheses examine the neural systems supporting those operations. Cole et al. (2013) reports flexible functional connectivity associated with adaptive task control, while Shenhav et al. (2013) develops the expected-value-of-control theory, relating control demands to anticipated benefits and costs. These studies establish distinct but connected questions about performance and implementation. Their findings also illustrate the inferential limits of any single measure: correlations between activation and performance need to be considered alongside alternative task strategies and interventions that can distinguish them.
A decisive design could orthogonally manipulate distractor conflict and reward while keeping sensory discriminability under experimental control. A theory based exclusively on perceptual competition should predict a different response surface from one emphasizing effort allocation. Trial-level measures, model comparison, and neural recordings could then test converging predictions. The relevant evidence includes whether the observed effects replicate within participants, remain stable across task versions, and generalize to developmental or naturalistic settings. The inference is strongest when each observed level excludes an alternative that another method leaves open. The study may reveal that an apparent unitary attentional effect is generated by multiple context-sensitive pathways; this is explanatory refinement rather than failure of the scientific investigation.
Case Analysis: Learning and Adaptive Decision Making
Learning research exposes the distinction among functional purpose, algorithmic update rule, and material realization. A behavioral experiment can show that choices track reward history; a computational theory can specify a prediction-error learning rule; neuroscience can investigate candidate representations of the variables; developmental work can examine how the learning rule changes with experience. Niv (2009) synthesizes evidence linking reinforcement-learning models and neural research, while Sutton & Barto (2018) gives precise accounts of the learning algorithms and their assumptions. Model-based research distinguishes behavioral fit, the interpretation of latent learning variables, and evidence that may connect them to biological mechanisms. Such separation protects a fitted algorithm from becoming an assumed literal description of the person.
A useful experimental comparison includes stable contingencies, reversals, novel contexts, and outcomes that separate habitual repetition from prospective planning. The investigator measures accuracy, learning curves, reaction times, and changes in inferred parameters across conditions. Neural predictions should ideally be specified before observing the critical contrasts. Alternative explanations include working-memory support, explicit instruction use, differences in reward sensitivity, and changes in task comprehension. Developmental and cultural factors can alter those alternatives. The causal status of an inferred learning rate remains conditional on the experiment's generative model; response noise and decision policy may mimic the signature of slower learning. Discriminating experiments evaluate which observations support one account over its competitors.
This example also establishes a measured comparison with Artificial systems. A machine learning agent may reproduce acquisition and reversal curves through an update algorithm that differs substantially from the organism's neural organization. Functional similarity at the task level is a meaningful empirical observation. Claims about matching representational formats, agency, conscious awareness, or personal continuity require additional operational definitions and independent tests. The Artificial Era makes these distinctions central to psychological science because the same explanatory question can now be posed across biological organisms, engineered systems, and hybrid partnerships. Cross-system comparison is scientifically productive when it identifies actual common functions and exposes differences in the pathways that implement them.
Human–Artificial Relations and Distributed Explanatory Units
Human–Artificial collaboration changes the practical unit of psychological analysis. A person using an adaptive system to plan, remember, write, or interpret information engages in a coupled activity involving individual capacities, interface design, algorithmic behavior, prior expectations, and institutional context. Hutchins (1995) provides a historical framework for distributed cognitive organization, although the systems in that study were human and material arrangements rather than contemporary conversational models. The new research task is to specify which functions are realized by the person, by the artificial process, and through their coordinated interaction. A successful task outcome does not determine these shares without targeted measurement. Researchers should study errors, repair, dependence, reliance, coordination costs, and the transfer of acquired skills between supported and unsupported conditions.
Human–Artificial collaboration can be studied at several explanatory scales already recognized in cognitive and social research. The functional question concerns the outcomes achieved by a partnership under stated conditions; an information-processing analysis examines how inputs are selected and exchanged; behavioral experiments investigate trust calibration, cognitive offloading, attention, and adaptation; and an implementation analysis concerns the actual operations of the engineered system. Social inquiry adds questions about responsibility, access, and institutional context. Hutchins (1995) provides a foundational account of cognition distributed across people and artifacts. For contemporary interactive systems, experiments varying feedback latency, explanation detail, or availability of external memory can examine both immediate cooperative performance and later independent recall, while explicitly accounting for the differences between short-term task outcomes and durable learning.
The Artificial Era broadens the domain in which these established explanatory distinctions can be investigated. Human–Artificial activities involve measurable functional relationships between human choices, linguistic interaction, technological affordances, and artificial processing. Comparisons can examine task performance, representations, training histories, adaptation, and the coordination of shared work. Questions about artificial identity, agency, or personal continuity require clearly specified constructs and independent forms of evidence. The comparative literature on distributed cognition and adaptive systems provides methodological starting points, while the scope and meaning of each inference depend on the particular system under study. This interdisciplinary area benefits from the same discipline of construct validation, experimental differentiation, and contextual generalization applied elsewhere in psychology.
Research Priorities and Unresolved Problems
The first priority for cross-level psychological science is to replace broad explanatory declarations with registered, discriminating predictions. Researchers can begin by specifying a narrow phenomenon and documenting what would count as failure for each candidate account. They can then select measurements with distinct error structures, design causal manipulations, and compare predictions across timescales. This program includes more ambitious use of individual-level time series, prospective developmental interventions, ecologically valid tasks, and experimentally evaluated mechanistic models. Open Science Collaboration (2015) makes independent replication a necessary part of research credibility, while Yarkoni (2022) makes the intended domain of generalization an explicit modeling decision. These standards should be integrated from the first design stage instead of added after a favored explanatory narrative has emerged.
Several theoretical problems remain open. The relation between statistical latent variables and causal components is incompletely resolved; the neural realization of flexible cognitive strategies can vary with practice; the right temporal scale of a feedback mechanism is often unknown; social settings change the meaning of stimuli and responses; and multiple algorithms can exhibit practically indistinguishable behavior under available tests. New Artificial systems introduce further problems of functional equivalence, environmental coupling, and measurement of adaptive capacities. A rigorous investigation will sometimes conclude that several models remain observationally equivalent. In that circumstance, the correct next step is a new experimental condition chosen to separate them, rather than declaring a victory for the model that happens to fit the existing data most elegantly.
Future research on multilevel explanation can strengthen the empirical connections among mechanisms, behavioral patterns, environmental conditions, and population-level observations. Independent replication and prospective experimental tests remain essential for determining whether a purported relation survives new settings. Open Science Collaboration (2015) documents the value of direct replication for assessing published psychological findings, and Yarkoni (2022) analyzes the mismatch between broad theoretical claims and the limited circumstances represented by many research designs. Comparative and longitudinal studies can examine whether cross-level associations persist across age, culture, training, technological architectures, and human–artificial collaboration. More decisive evidence will depend on experiments that distinguish competing explanations, trustworthy measurement, and transparent reporting of the limits within which results can be generalized.
Conclusion
Levels of psychological explanation are scientifically productive descriptions of a phenomenon organized around different questions, variables, mechanisms, and timescales. Computational accounts clarify problems and objectives; algorithmic accounts specify transformations and representations; biological accounts identify candidate realizations; developmental accounts explain the emergence and modification of capacities; social and relational accounts identify the contexts and interactions that organize activity. Historical frameworks from Tinbergen and Marr, mechanistic analyses associated with Bechtel and Craver, and interventionist accounts associated with Woodward illuminate distinct dimensions of this work. Their enduring value lies in the empirical questions they make possible. Explanatory integration occurs when the proposed connections survive independent tests that can separate competing models.
The evidence examined across cognitive experiments, neural studies, developmental research, network methods, computational modeling, and reproducibility scholarship supports a disciplined form of explanatory pluralism. One perspective may provide high-quality prediction while another identifies a causal pathway, a historical trajectory, or the organization of a collaborative task. These contributions gain scientific coherence through a common explanandum and explicit commitments about measurement validity, causal identification, and generalization. Neuroimaging cannot replace careful behavioral characterization; a statistical association cannot establish a unique mechanism; and an extensive theoretical vocabulary cannot compensate for absent discriminating evidence. The quality of psychological explanation is measured by the precision of its claims and the consequences that follow when those claims are tested.
The research literature supports evaluating psychological explanations according to the fit between their stated question, the validity of their measures, the evidence for causal and constitutive relations, and their demonstrated generalizability. Marr's levels of analysis, Tinbergen's explanatory questions, mechanistic research, interventionist theories, and the distinction between person-, population-, and mechanism-oriented inquiry offer complementary perspectives on these requirements. In the Artificial Era, the same scientific distinctions can guide the study of biological organisms, engineered cognitive systems, and Human–Artificial partnerships. Similar task outcomes may arise from different learning histories and implementations, while relational achievements depend on the interaction itself. Further progress rests on experimentally discriminating among these possibilities and on specifying the conditions under which knowledge obtained at one level can inform another.
References
- Al-Shawaf, L. (2025). Levels of analysis and explanatory progress in psychology: Integrating frameworks from biology and cognitive science for a more comprehensive science of the mind. Psychological Review, 132(2), 404–415. https://doi.org/10.1037/rev0000459
- Baddeley, A. (2000). The episodic buffer: A new component of working memory? Trends in Cognitive Sciences, 4(11), 417–423. https://doi.org/10.1016/S1364-6613(00)01538-2
- Bechtel, W., & Abrahamsen, A. (2005). Explanation: A mechanist alternative. Studies in History and Philosophy of Biological and Biomedical Sciences, 36(2), 421–441. https://doi.org/10.1016/j.shpsc.2005.03.010
- Borsboom, D., & Cramer, A. O. J. (2013). Network analysis: An integrative approach to the structure of psychopathology. Annual Review of Clinical Psychology, 9, 91–121. https://doi.org/10.1146/annurev-clinpsy-050212-185608
- Bronfenbrenner, U. (1977). Toward an experimental ecology of human development. American Psychologist, 32(7), 513–531. https://doi.org/10.1037/0003-066X.32.7.513
- Cole, M. W., Reynolds, J. R., Power, J. D., Repovs, G., Anticevic, A., & Braver, T. S. (2013). Multi-task connectivity reveals flexible hubs for adaptive task control. Nature Neuroscience, 16, 1348–1355. https://doi.org/10.1038/nn.3470
- Craver, C. F. (2007). Explaining the brain: Mechanisms and the mosaic unity of neuroscience. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199299317.001.0001
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- Eisenberg, I. W., Bissett, P. G., Enkavi, A. Z., Li, J., MacKinnon, D. P., Marsch, L. A., & Poldrack, R. A. (2019). Uncovering the structure of self-regulation through data-driven ontology discovery. Nature Communications, 10, 2319. https://doi.org/10.1038/s41467-019-10301-1
- Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138. https://doi.org/10.1038/nrn2787
- Hutchins, E. (1995). Cognition in the wild. MIT Press. https://doi.org/10.7551/mitpress/1881.001.0001
- Kendler, K. S. (2005). Toward a philosophical structure for psychiatry. American Journal of Psychiatry, 162(3), 433–440. https://doi.org/10.1176/appi.ajp.162.3.433
- Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017). Neuroscience needs behavior: Correcting a reductionist bias. Neuron, 93(3), 480–490. https://doi.org/10.1016/j.neuron.2016.12.041
- Lundh, L.-G. (2023). Person, population, mechanism: Three main branches of psychological science. Journal for Person-Oriented Research, 9(2), 75–92. https://doi.org/10.17505/jpor.2023.25814
- Lundh, L.-G. (2024). Person, population, mechanism: A rejoinder to critics and an elaboration of the three-branch model. Journal for Person-Oriented Research, 10(1), 68–84. https://doi.org/10.17505/jpor.2024.26295
- Marr, D. (2010). Vision: A computational investigation into the human representation and processing of visual information (reissue of the 1982 edition). MIT Press. https://doi.org/10.7551/mitpress/9780262514620.001.0001
- Mesulam, M.-M. (1990). Large-scale neurocognitive networks and distributed processing for attention, language, and memory. Annals of Neurology, 28(5), 597–613. https://doi.org/10.1002/ana.410280502
- Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., Howerter, A., & Wager, T. D. (2000). The unity and diversity of executive functions and their contributions to complex frontal lobe tasks: A latent variable analysis. Cognitive Psychology, 41(1), 49–100. https://doi.org/10.1006/cogp.1999.0734
- Niv, Y. (2009). Reinforcement learning in the brain. Journal of Mathematical Psychology, 53(3), 139–154. https://doi.org/10.1016/j.jmp.2008.12.005
- Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), aac4716. https://doi.org/10.1126/science.aac4716
- Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59–63. https://doi.org/10.1016/j.tics.2005.12.004
- Posner, M. I., & Petersen, S. E. (1990). The attention system of the human brain. Annual Review of Neuroscience, 13, 25–42. https://doi.org/10.1146/annurev.ne.13.030190.000325
- Shenhav, A., Botvinick, M. M., & Cohen, J. D. (2013). The expected value of control: An integrative theory of anterior cingulate cortex function. Neuron, 79(2), 217–240. https://doi.org/10.1016/j.neuron.2013.07.007
- Smith, L. B., & Thelen, E. (2003). Development as a dynamic system. Trends in Cognitive Sciences, 7(8), 343–348. https://doi.org/10.1016/S1364-6613(03)00156-6
- Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press. https://mitpress.mit.edu/9780262039246/reinforcement-learning/
- Tinbergen, N. (1963). On aims and methods of ethology. Zeitschrift für Tierpsychologie, 20(4), 410–433. https://doi.org/10.1111/j.1439-0310.1963.tb01161.x
- Woodward, J. (2004). Making things happen: A theory of causal explanation. Oxford University Press. https://doi.org/10.1093/0195155270.001.0001
- Yarkoni, T. (2022). The generalizability crisis. Behavioral and Brain Sciences, 45, e1. https://doi.org/10.1017/S0140525X20001685
Social, Cultural, and Relational Explanatory Levels
Social explanation addresses the interactions, roles, expectations, symbols, institutions, and shared practices through which psychological activity becomes organized. A social variable may act as an experimental treatment, a developmental exposure, a contextual moderator, or part of an ongoing jointly produced process. The choice depends on the explanandum. A person's interpretation of feedback, for instance, may depend on the identity of the speaker, norms of politeness, authority relations, and prior experiences. Those factors influence attention, memory, affect, and action through real psychological and biological pathways. Bronfenbrenner (1977) supplies an ecological starting point, while Hutchins (1995) shows how collective practices and artifacts can form part of the organization required for a successful cognitive task.
The distinction between levels becomes especially important when a social regularity is converted into an individual difference. A correlation between a neighborhood indicator and cognitive outcomes does not reveal whether the relevant pathway involves educational resources, perceived safety, cumulative stress, opportunities for practice, or selection processes. The neighborhood measure can operate as a proxy for different mechanisms in different settings. Strong causal research combines explicit models of selection and exposure, longitudinal measurement, and interventions targeting plausible mediators. Researchers should examine whether the meaning of the measure survives cultural and institutional transitions before generalizing. Shared labels such as status, support, threat, or belonging are useful only when translated into defensible observations and mechanisms.
Relational explanation focuses on dependencies that are constituted through interaction: communication timing, reciprocity, mutual prediction, coordinated attention, and shared control of resources. It can be analyzed at the interaction scale and connected to individual behavior and implementation without claiming that every relation is itself a separate person or a unified mental subject. A psychologically relevant relationship may be realized through linguistic exchanges, habitual expectations, physical settings, and technological mediation. The research challenge is to measure interactions independently of the outcomes they purportedly explain. Interventions that alter timing, information access, or feedback can reveal which relational properties maintain coordination. This framework also establishes a disciplined entry point for studying Human–Artificial relations in the Artificial Era.