RESEARCH DOSSIER
Biopsychosocial Causation in Health and Disease
Abstract. Biopsychosocial causation concerns the mechanisms through which biological states, psychological processes, social relationships, and institutional conditions jointly influence disease incidence, illness experience, functional limitation, clinical care, and recovery. This monograph reconstructs the intellectual history of multilevel health explanation from Engel's 1977 proposal to contemporary causal inference; defines exposures, mediators, modifiers, confounders, and outcomes; and critically integrates experimental, longitudinal, mechanistic, and intervention evidence. The central scientific question is how effects originating at one level of organization become expressed through processes at other levels while remaining empirically identifiable. Respiratory-virus challenge studies, diabetes prevention trials, chronic-pain interventions, observational social-relationship research, psychosocial stress biology, and social determinants provide complementary test cases. The evidence supports specific, context-dependent causal pathways and exposes important limits of inference concerning mediation, generalization, measurement, feedback, and treatment response. A biologically instantiated mechanism can be socially patterned; a psychologically mediated effect can be tested by intervention; and a clinical outcome can reflect both pathophysiology and circumstances of measurement and care. The resulting synthesis treats biopsychosocial explanation as a program of explicit, falsifiable causal models, comparative designs, and clinically meaningful outcomes rather than an undifferentiated inventory of possible influences. This is a critical integrative research monograph based on identifiable scholarly studies and methodological sources, not a protocol-driven systematic review or meta-analysis.
Keywords: biopsychosocial model; causal inference; health psychology; behavioral medicine; disease etiology; psychoneuroimmunology; stress physiology; social determinants of health; mediation; effect modification; causal diagrams; longitudinal epidemiology; health behavior; chronic pain; clinical outcomes.
The Scientific Problem of Multilevel Health Causation
An explanation of health and disease identifies how a particular outcome comes about under specified conditions. Biological mechanisms such as viral entry, insulin signaling, vascular injury, and immune regulation are central to this undertaking. The same processes unfold in organisms whose exposures, behaviors, care opportunities, and interpretations of symptoms are shaped by relationships and institutions. A scientifically adequate account must determine which of these conditions are causal, which operate through measurable intermediates, and which merely accompany the outcome because of common causes or selection.
Biopsychosocial causation names this explanatory undertaking across biological, psychological, and social domains. Its unit of inquiry is a causal question: for which population, exposure or intervention, comparator, outcome, and time horizon would a change in one condition alter what happens? Distinguishing the question from a broad commitment to comprehensive care is decisive. Broad clinical attention can improve communication, while scientific explanation demands a more restrictive account of the causal pathways under examination. A clinician can appropriately consider housing insecurity during a consultation without thereby establishing the exact contribution of housing to a particular biomarker. Conversely, a molecular mechanism can be demonstrated in an experiment while leaving open why its triggering exposure is distributed unequally across populations.
Engel, G. L (1977) defined the biopsychosocial model as a program for scientific research, medical education, and clinical action. The contemporary question is how to turn that program into models that generate distinguishable predictions. von Känel, R (2026) identifies the persistence of descriptive domain lists as a central obstacle and argues for testable regulatory relations across levels. This monograph uses that challenge to organize its evidence: explanatory adequacy depends on interventions, temporal ordering, mechanism measurements, competing causal structures, and the bounds of each claim.
Conceptual Definitions and Units of Analysis
A biological factor is an observed or hypothesized process involving organisms and their material operations: genetic variation, endocrine signaling, cellular immunity, metabolic regulation, organ function, or physiological injury. A psychological factor is a process characterized at the level of perception, appraisal, learning, motivation, affect, memory, decision, or behavior. A social factor includes patterned relations among people and institutions, including caregiving, labor conditions, socioeconomic position, norms, discrimination, treatment access, and public policy. These are analytic descriptions of phenomena that frequently share physical realizations and common historical causes. They are neither exclusive classifications of individual events nor interchangeable indicators.
An exposure is a condition whose effect is under study. A mediator lies along a proposed causal pathway from exposure to outcome. A confounder causes or predicts both exposure and outcome in a manner that distorts the association of interest. A moderator or effect modifier identifies conditions under which an effect differs. A collider is a variable with two or more causal inputs; restricting an analysis to a collider can create a spurious association between its causes. The meaning of a variable therefore depends on the causal question and time indexing, not on its intrinsic label. For example, depressive symptoms may be an outcome of illness in one study, a determinant of treatment participation in another, and a time-varying confounder affected by prior treatment in a third.
Causation may be investigated at different units: cells, physiological systems, individual trajectories, families, clinics, neighborhoods, and populations. Aggregate patterns cannot automatically establish individual-level mechanisms, and individual-level trials do not automatically demonstrate a population policy effect. Pearl, J (1995) provides a formal language for making some of these assumptions explicit. It does not replace substantive knowledge about which arrows are plausible or whether an intervention is defined well enough to be meaningful.
Disease, Illness, Function, and Care as Distinct Outcomes
Disease may be operationalized as a specified pathological state or clinically classified condition. Illness designates the lived experience of symptoms, disruption, and interpretation. Function refers to capacities and participation in activities; care outcomes include detection, treatment initiation, adherence, continuity, access, and safety. These quantities can change together, but their alignment is neither necessary nor stable. A therapy may reduce symptom distress without eliminating tissue pathology. An environmental change may decrease exposure to a pathogen before any symptomatic illness occurs. Improved case finding may increase diagnosed prevalence even while true risk remains unchanged.
The distinction determines what evidence can establish. If perceived stress predicts self-reported colds, the association could reflect infection susceptibility, symptom perception, memory, reporting practices, or their combination. Virologic challenge studies with documented infection offer a more discriminating test than self-report alone. Similarly, self-reported improvement in chronic pain is a patient-important outcome, but it is not evidence that the anatomical source of pain has been reversed. A biomarker response can be scientifically informative without guaranteeing an improvement in daily function.
For causal analyses, investigators must therefore specify the outcome's measurement procedure and its relation to the phenomenon claimed. The biopsychosocial perspective gains explanatory precision by preserving these distinctions. A multilevel study that measures biologic disease activity, perceived symptoms, functional ability, and participation in care can investigate both their relationships and their separate determinants. When a single composite index is used instead, the assumptions that make the composite meaningful deserve independent examination.
Historical Formation of Biopsychosocial Inquiry
The history of biopsychosocial explanation includes traditions in psychosomatic medicine, social medicine, behavioral science, epidemiology, and systems-oriented physiology. These traditions asked overlapping questions with different evidentiary tools: how experience is embodied; how material deprivation distributes illness; how habits affect exposure; and how interacting regulatory processes generate physiological stability and breakdown. Their historical relevance lies in the causal problems they made visible, rather than in a single uninterrupted succession of universally accepted theories.
The development of microbiology, laboratory physiology, and clinical experimentation produced extraordinary causal specificity. Pathogens, drugs, physiological pathways, and pathological lesions became amenable to controlled comparison. At the same time, the prevalence and outcomes of many conditions were demonstrably patterned by working conditions, housing, nutrition, and care access. These observations required research designs capable of connecting mechanisms at the organism level to causes of exposure at the population level. They also warned against treating social categories as if they were direct physiological agents.
During the second half of the twentieth century, psychosomatic and behavioral medicine increasingly investigated the effects of stress, emotion, learning, and social experience on physiological responses. Subsequent developments in psychoneuroimmunology and epidemiologic causal inference expanded the available questions and tools. The history is therefore best reconstructed through changing methods: descriptive observations, psychometric assessment, challenge experiments, prospective cohorts, randomized interventions, mechanistic assays, and formal causal models. Their strengths differ, and each reveals a different part of the larger explanatory structure.
Engel's Medical Model and Its Scientific Legacy
In Engel, G. L (1977), published in Science in 1977, George L. Engel argued that disease and medicine must be investigated with attention to psychological, behavioral, and social dimensions. His argument was framed around an enlarged scientific and clinical model. It linked questions of disease to the clinical encounter and situated persons within systems of interrelated processes. The paper's historical achievement was to make multilevel explanation an explicit organizing problem for medical science, rather than only a concern of selected psychosomatic specialties.
The model has subsequently served several functions: philosophical orientation, clinical interviewing guide, teaching framework, and research program. Borrell-Carrió, F., Suchman, A. L., & Epstein, R. M (2004) clarified that the clinical method requires attention to subjective experience and to the relationship between patient and professional. They also observed that circular causal stories often need temporally ordered approximations when choosing interventions. This is a useful disciplinary correction: a satisfactory clinical relationship and a valid scientific effect estimate are complementary achievements that require distinct evaluation criteria.
A major contemporary risk is treating the biological, psychological, and social categories as mandatory boxes to complete in every case. Merely finding at least one variable from each column supplies neither mechanism nor a reason to act. von Känel, R (2026) advances an explicitly causal reading of the model, emphasizing computational approaches, network analysis, and measurable interlevel regulation. Such proposals should be judged by whether they discriminate among plausible accounts, generate transferable predictions, and improve decisions under uncertainty.
Systems, Regulation, and Cross-Level Mechanisms
A system consists of components whose interactions affect the evolution of states over time. Physiological systems exhibit feedback, delays, compensation, thresholds, and different scales of response. A psychological process can alter the input to such a system through behavior or stress-related signaling, and a physiological state can alter psychological processes through fatigue, pain, inflammation, or interoceptive cues. Social arrangements can alter both: an employer's shift schedule changes sleep opportunity, which can change activity patterns and physiological regulation; a family caregiver changes medication organization and the practical ability to follow a care plan.
This multilevel description becomes a causal account once the intervening links can be investigated. Suppose unstable employment increases sleep disruption and changes exposure to environmental hazards. If subsequent disease risk differs, the investigator needs evidence for both the total effect of employment conditions and the proposed mediating pathways. A biomarker correlation alone cannot decide whether a pathway was necessary, sufficient, or quantitatively dominant. The biological mechanism may be well established in laboratory work while the population exposure-to-mechanism chain remains uncertain.
System-level thinking also generates questions about intervention location. The same downstream biological variable might be changed by a pharmaceutical intervention, a behavioral intervention, or a policy that removes a recurrent exposure. Their feasibility, distributional consequences, timing, and unintended effects differ. Explanation should identify which change is realistic, how outcomes would be measured, and where evidence exists for the particular population. This is the form in which systems concepts become operational rather than metaphorical.
Counterfactual and Mechanistic Conceptions of Cause
A counterfactual causal effect asks how an outcome would differ under alternative conditions for the same target population. In an adequately randomized trial, treatment assignment can make groups comparable before intervention, enabling estimation of effects on prespecified outcomes. In observational settings, identifying the corresponding effect generally requires substantive assumptions about confounding, selection, exposure definition, and measurement. The point is not that causation equals randomization; the point is that claims of causal influence must be tied to a coherent comparison.
Mechanistic explanation asks how a change propagates. An exposure may alter perceived threat, autonomic activity, endocrine signaling, sleep behavior, immune regulation, and susceptibility. These links can be measured separately, with different designs and temporal resolutions. A complete mechanistic chain requires stronger evidence than proof of a total exposure effect: the mediator must be affected by the exposure; the mediator-outcome relation must be causally interpretable; and alternative paths need to be addressed. VanderWeele, T. J (2016) shows why decomposition rests on specific identification assumptions.
The two conceptions are mutually informative. Demonstrating a causal total effect without fully identifying its mechanism may be sufficient to guide a practical intervention. Establishing biological plausibility without an estimated exposure effect remains valuable for theory but is not a demonstration of population-level causality. A central task of biopsychosocial science is to keep these forms of evidence connected while recording what each actually establishes.
Directed Acyclic Graphs and Explicit Assumptions
A directed acyclic graph represents hypothesized relations between variables in a temporally ordered model. Consider material deprivation, occupational exposure, chronic stress, smoking, inflammation, and disease. The hypothesis that deprivation contributes to disease through several pathways can be drawn with arrows; an additional variable, such as baseline health, may influence both employment and later outcome. The graph helps investigators decide which baseline factors to adjust for, which intermediates to preserve when estimating a total effect, and where selection could bias a sample.
A graph is acyclic because simultaneous feedback is represented through distinct times rather than arrows looping within the same time point. A time-indexed model can include illness at baseline affecting employment at the next observation, which then affects stress and health at a later observation. Flattening these stages into a cross-sectional diagram could obscure reverse causation. Graphs also force decisions about constructs: the label “social adversity” must be operationalized as an observable or well-defined latent quantity before its arrows can be tested.
Pearl, J (1995) showed how graph structure and statistical assumptions can support the identification of causal effects. Yet two different graphs may produce the same observed associations, and unmeasured variables can defeat otherwise elegant identification. Graphical analysis therefore documents assumptions for scrutiny. It does not verify the truth of the graph by its visual appearance, nor does it give causal status to a variable simply because it is named in a biopsychosocial framework.
Time, Reciprocity, and Feedback
The relationships between health, behavior, and environment often change over time. Symptoms can reduce social participation; reduced participation can impair access to support; diminished support can change care engagement; altered care can affect symptoms. Describing this process as reciprocal captures an important feature but leaves causal estimation unresolved. The observational researcher must determine the relevant time scale, measurement intervals, prior states, and interventions that could interrupt a proposed sequence.
Time-varying confounding is a central challenge. A health measure recorded midway through follow-up may influence subsequent treatment and predict outcome, while also being affected by earlier treatment. Adjusting for it using standard regression can remove part of the prior treatment effect or introduce bias. Robins, J. M., Hernán, M. Á., & Brumback, B (2000) developed marginal structural models to address such structures under stated assumptions, using inverse-probability weighting. The method depends on adequate measurement, model specification, positivity, and the availability of relevant confounders.
Realistic measurement intervals matter as much as method choice. Autonomic responses occur quickly, routines shift over days, functional limitations may evolve over months, and structural exposures can accumulate over decades. A single baseline questionnaire cannot reconstruct every transition. Intensive longitudinal assessments can improve temporal resolution, but can also introduce participant burden and nonrandom missing data. Strong designs coordinate the time scales of mechanism, exposure, and outcome before selecting an analytic model.
Stress Appraisal and Neuroendocrine Regulation
Stress research separates potentially demanding events, appraisal of their meaning, affective responses, coping behavior, and physiological regulation. An adverse event does not elicit identical biological responses in all persons; prior learning, expectations, resources, and current bodily state affect response. Psychological appraisal therefore occupies a plausible position in many causal pathways, though causal evidence must show how appraisal changes subsequent biological or behavioral variables. Investigators commonly measure perceived stress, stressful events, or momentary affect; these measures refer to related but distinct constructs.
McEwen, B. S (1998) described the protective and potentially damaging consequences of physiological stress responses. Neuroendocrine and autonomic signaling supports adaptation to challenges, while the frequency, duration, and regulation of responses are relevant to long-term physiological burden. Steptoe, A., & Kivimäki, M (2012) reviewed plausible cardiovascular links through autonomic, neuroendocrine, inflammatory, and behavioral routes. A plausible pathway is not necessarily the dominant mediator of a specific observed association.
The broad synthesis by Cohen, S., Janicki-Deverts, D., & Miller, G. E (2007) links stress research to disease while emphasizing the need to evaluate specific pathways and outcomes. Stress physiology is especially vulnerable to measurement simplification. A single cortisol concentration depends on sampling time, circadian rhythm, recent behavior, medication, health condition, and laboratory procedure. Opposite directions of endocrine change may be found under different forms and durations of exposure. The causal question must therefore specify the stressor, regulatory process, temporal sampling plan, and disease endpoint. Psychological distress, autonomic response, and later disease cannot be treated as equivalent stages merely because all can be categorized as “stress.”
Immune Function, Inflammation, and Disease Susceptibility
Immune processes provide concrete candidates for psychological-to-biological pathways. Neural and endocrine signals influence immune-cell function, and immune signaling can alter behavior and affect. Kiecolt-Glaser, J. K., McGuire, L., Robles, T. F., & Glaser (2002) surveyed links among stress, behavior, immune response, and health in psychosomatic medicine. The broader literature includes experimental immune measures, infection outcomes, wound healing, and intervention studies, each requiring its own assessment of clinical relevance.
The difference between an immunological marker and a clinical endpoint is fundamental. An intervention that changes a cytokine level does not necessarily prevent infection, improve survival, or relieve symptoms. Assays may capture transient responses, and immune activation can have protective or harmful functions depending on context. Investigators should specify whether their outcome is host resistance, inflammatory signaling, immune response to a challenge, or patient-important illness. Broad terms such as “boosting immunity” obscure these distinctions.
A Black, D. S., & Slavich, G. M (2016) examined immune-system outcomes, illustrating the utility and limits of mechanistic intervention synthesis. It does not establish a universal immune benefit or a general reduction in disease risk. Evidence integration needs to consider the heterogeneity of interventions, comparison groups, assays, follow-up periods, and populations. The scientific gain of biopsychosocial reasoning lies in formulating testable links between experience and physiological outcomes while resisting extrapolation beyond the measured endpoint.
Behavior as a Biological and Social Pathway
Behavior is among the most empirically tractable routes linking social conditions and biological health. Smoking, sleep, physical activity, nutrition, occupational exposure, medication taking, and preventive care alter contact with hazards, physiological processes, or treatment. Their organization depends on learning and intention, but also on practical opportunity, economic constraints, built environments, and the availability of assistance. A behavior can consequently be both a psychological action and a mediator of social causation.
Consider a change in work schedules. It may influence sleep opportunity and fatigue, which in turn affect dietary routines, medication timing, injury risk, and metabolic regulation. The causal analysis should distinguish the policy effect from the individual's perceived control and the behavioral pathway. An intervention that instructs people to improve sleep may have limited effect when the decisive exposure is an inflexible shift schedule. By contrast, when a modifiable routine is feasible and supported, behavior-focused intervention can produce meaningful clinical change.
The Diabetes Prevention Program Research Group (2002) supplies unusually strong evidence that a structured lifestyle program can change incidence of a defined disease outcome in adults at elevated risk. It does not identify every behavioral or psychosocial mediator, and the estimated treatment effect belongs to the intervention package as delivered to the study population. Causal explanation can recognize this outcome while treating questions of mechanism, maintenance, access, and transportability as further empirical tasks.
Relationships, Social Integration, and Mortality
Social relationships may influence health through emotional support, behavioral coordination, material assistance, exposure to conflict, shared environments, and access to care. The same social network can contain protective and adverse influences. Measures of marital status, household composition, perceived support, network size, and participation therefore represent different constructs. Their interchangeability cannot be assumed simply because each appears under a social heading.
In a major Holt-Lunstad, J., Smith, T. B., & Layton, J. B (2010) synthesized observational studies of social relationships and mortality. The review reported 148 studies involving 308,849 participants and an overall association corresponding to higher survival among those with stronger social relationships (odds ratio 1.50; 95% confidence interval 1.42–1.59). These estimates establish a robust association across the collected designs, while leaving open the precise causal contribution of networks and the mechanisms connecting relationships to survival. Baseline health can affect social ties as well as mortality, and adjustment may be uneven.
Evidence from socially patterned health cannot be used as a warrant to prescribe one normative model of family or community. Relationships differ in safety, voluntariness, reciprocity, and function. Causal research should distinguish supportive interactions from coercion or social burden and examine who is able to obtain assistance. In clinical care, network assessments can be useful for practical planning; research claims about mortality effects require stronger identification and careful differentiation of relationship characteristics.
Structural Conditions and the Distribution of Exposure
Material and institutional circumstances shape the probability of contact with hazards, the duration of exposure, the capacity to respond, and the timing of treatment. Income and wealth influence housing quality, food security, transportation, and the flexibility available for medical appointments. Employment conditions determine hours, job strain, occupational hazards, and eligibility for leave. Neighborhood features affect air quality, pedestrian safety, access to services, and social opportunities. These are not simply background descriptors. They are components of candidate intervention pathways whose effects can be investigated, although the intervention must be defined and the causal comparison made explicit.
Marmot (2005)71146-6) reviewed evidence of graded socioeconomic inequalities in health and emphasized the organization of opportunity and control. A gradient can be present across multiple levels of advantage rather than confined to a dichotomy between extreme deprivation and affluence. Yet socioeconomic position is often a summary indicator of multiple co-occurring processes. An association between education and disease may involve early-life resources, occupational exposures, knowledge, selection by prior health, and later access to care. Adjustment for one downstream consequence can suppress part of the very effect being investigated.
A strong causal investigation separates structurally manipulable factors from proxies for unmeasured histories. A housing intervention might change indoor hazards; paid sick leave might change infectious contact patterns and care-seeking; reliable transportation might improve treatment attendance. These hypotheses imply concrete outcomes and intermediate measurements. A policy effect cannot be assumed to arise entirely through the psychological experience of disadvantage, because direct exposure and resource pathways may remain. Equally, institutional structures influence health partly through expectations, perceived control, social relationships, and behavioral choices. A multilevel causal model must permit both routes.
The scientific and ethical implications converge on careful attribution. Explaining unequal health outcomes by individual motivation alone risks assigning causal responsibility to people whose options are materially restricted. Explaining all differences by social category without specifying pathways can likewise conceal heterogeneity within groups. The appropriate unit of explanation is a specified causal process in a defined context, with attention to how policy, clinical services, and personal action may interact.
Developmental Timing and Life-Course Accumulation
Health trajectories reflect exposures distributed across development, including prenatal conditions, childhood caregiving, schooling, adolescence, work, family responsibilities, and later-life constraints. The same exposure may have different consequences at different ages because physiology, learned regulation, and available supports change. Cumulative disadvantage can arise through repeated exposure, constrained opportunities for recovery, and the effects of earlier illness on later social positions. Developmental causation therefore requires temporal data rather than treating adult socioeconomic status or adult stress as complete summaries of the past.
Felitti and colleagues' Adverse Childhood Experiences study (1998)00017-8) reported associations between retrospectively assessed childhood adversities and multiple adult health risks and outcomes among members of a health maintenance organization. The study helped establish a program of research into early adversity and adult health. Its retrospective exposure assessment, selection of participants, correlated household adversities, and the many pathways between childhood and adult outcomes constrain causal interpretation. The observed patterns cannot be read as deterministic forecasts for individuals, and aggregate dose–response associations do not identify a single biological injury mechanism.
Life-course hypotheses should distinguish sensitive periods, cumulative exposure, chains of risk, and the possibility of recovery or compensation. A sensitive-period hypothesis predicts that intervention at one developmental stage has a consequence not reproduced by otherwise equivalent intervention later. A cumulative-risk account predicts that burden or duration matters across multiple stages. A chain-of-risk account specifies how one event changes the probability of another. These models imply different designs and often require repeated measurements of exposure and context.
Separating life-course causal explanations from social and biological selection is technically demanding. Baseline health can shape educational attainment, employment, relationships, and subsequent exposure, while earlier exposure can change baseline health. Studies benefit from birth cohorts, sibling or family comparisons where justified, repeated measures, and triangulation across different assumptions. Their inferential strengths should be reported alongside attrition, selective survival, measurement drift, and the limited transportability of historical cohorts to new social conditions.
Symptom Perception, Meaning, and Reporting
Symptoms are experienced through bodily signals, perceptual interpretation, attention, learned expectations, emotional context, and the language available for describing experience. Some symptoms correspond closely to an identifiable physiological event, while others are nonspecific or fluctuate with behavioral and environmental context. Reporting reflects what an individual notices, considers important, remembers, and is willing or able to communicate. The causal science of symptoms therefore requires measurement at both the physiological and experiential levels.
A common inferential error arises when investigators infer changes in a disease mechanism solely from changes in symptom ratings. Pain intensity can improve because of changes in nociceptive processing, appraisal, activity patterns, contextual expectations, or combinations of processes. An improvement is real as a change in experienced suffering and remains clinically important, yet its mechanism requires additional inquiry. Conversely, stable symptom ratings do not prove that disease biology is unchanged when adaptation or response-shift processes occur. Patient-reported and objectively measured outcomes answer different questions.
Research protocols can combine symptom diaries with physiological indicators, clinician assessments, and measures of daily functioning. Sampling intervals should follow the suspected process; short-lived episodes may be missed by monthly questionnaires, while intensive momentary measurement can alter monitoring behavior. Measurement equivalence across languages and cultures also matters, especially when a construct relies on interpretation of evaluative terms. When the same question is interpreted differently across groups, an apparent difference in distress may partly reflect the instrument.
A biopsychosocial model of symptom appraisal should make its alternatives visible. Increased symptom reporting might result from greater physiological disturbance, greater attention to unchanged disturbance, greater readiness to disclose, or a change in what respondents consider abnormal. Those alternatives predict different patterns across physiological measures, behavioral observations, and reporting settings. Distinguishing them allows research to value patient experience without collapsing symptoms into either tissue damage or a purely cognitive explanation.
Clinical Encounters, Expectations, and Contextual Effects
Health care is a social intervention delivered through institutions, language, relationships, and material procedures. The clinician's explanation, the patient's expectations, and the surrounding ritual can influence anxiety, symptom interpretation, adherence, and reported outcomes. The treatment context also affects who agrees to participate, which difficulties are disclosed, and whether care can continue. These pathways belong within scientific evaluation because clinical interventions are rarely experienced as decontextualized substances or procedures.
Kaptchuk, T. J., Friedlander, E., Kelley, J. M., Sanchez, M (2010) randomized patients with irritable bowel syndrome to a short open-label placebo intervention or a no-treatment control, using a defined clinical setting and outcome assessments. The study provides evidence that symptom outcomes may respond to a treatment context even without a deceptive statement that the pills contain active pharmacological ingredients. It does not establish that the same procedure cures irritable bowel syndrome, produces durable structural change, or generalizes to every illness. Small samples, short follow-up, subjective endpoints, and the particular comparison condition matter when interpreting the observed improvement.
Contextual effects can operate alongside active treatments. A patient may experience both the specific pharmacological effect of a medication and an effect associated with explanation, expectation, and clinical support. The task is to estimate their contributions under feasible comparisons. A placebo-controlled drug trial can estimate the incremental effect of a drug under the trial's contextual conditions, while additional designs are needed to isolate components of clinician interaction. It is misleading to infer a patient's symptom origin from the existence of placebo responses, because the pathways by which context changes symptom experience are biological as well as psychological.
Clinical interaction has a second, institutional pathway: clear information and accessible services can influence initiation and persistence in appropriate care. Communication quality should be assessed through comprehension, treatment participation, safety, and patient-important endpoints rather than assumed to produce physiological benefits universally. Ethical evaluation also considers the possibility of induced fear, unrealistic reassurance, and compromised consent. A precise biopsychosocial account joins relational competence to measurable care consequences.
Controlled Viral Challenge as a Causal Case Study
Prospective respiratory-virus challenge research illustrates both the promise and the limits of multilevel causal inference. In Cohen, S., Tyrrell, D. A. J., & Smith, A. P (1991), psychological stress was assessed before exposure; 394 healthy participants received one of five respiratory viruses, and 26 received saline. Participants were quarantined and monitored for laboratory evidence of infection and clinical symptoms. The standardized challenge constrained variation in pathogen exposure and helped distinguish documented infection from the experience and reporting of cold symptoms.
The investigators reported higher rates of infection and clinical colds with higher prespecified stress scores, with a graded pattern after adjustment for a number of measured factors. The result is unusually informative because stress assessment preceded the infection challenge and the outcome included biological confirmation. Nevertheless, stress levels were not experimentally assigned. Unmeasured traits, prior experiences, health status, or exposure histories may still influence both stress and host resistance. Standardizing viral exposure strengthens interpretation without converting every association involving stress into an experimental estimate of the stress effect.
The study is important for another reason: plausible intermediates such as smoking, alcohol consumption, sleep, exercise, diet, and several immune measurements did not statistically account for the reported association in that setting. This is a constraint on candidate explanations, not evidence that behavior and immune regulation are irrelevant in other contexts. Mediator measures may be incomplete, measured on an unsuitable timescale, or unable to capture the relevant pathway. A causal theory becomes stronger when it records null mediation findings and specifies new discriminatory tests.
Subsequent challenge studies examined additional psychosocial characteristics. Cohen, S., Doyle, W. J., Turner, R. B., Alper, C. M., & Skon (2003) evaluated emotional style in relation to the development of colds under viral exposure. The collective program shows how the same experimental infrastructure can examine different predictors while preserving careful distinctions between pathogen exposure, infection, symptoms, and subjective experience. External validity remains bounded by the selected healthy volunteers, challenge organisms, and monitoring conditions. Community infection patterns involve repeated exposure, vaccination, comorbidity, environmental conditions, and access to care, making transportability a separate research question.
Prevention Trials and the Difference Between Outcomes
Randomized prevention trials are powerful because an intervention assignment can establish an identifiable contrast. The Diabetes Prevention Program Research Group (2002) studied adults at high risk of type 2 diabetes and reported that intensive lifestyle intervention reduced diabetes incidence relative to placebo; the trial also studied metformin. The frequently cited 58% reduction for lifestyle intervention refers to relative incidence within the randomized comparison and follow-up, not a guaranteed percentage-point reduction for every participant. The intervention combined behavioral goals with support, opportunities for monitoring, and implementation structures.
The trial supports a causal conclusion about the delivered prevention program in the population recruited, subject to its trial design and follow-up. It does not show that changes in motivation alone caused the outcome, nor does it directly estimate the independent effect of each behavioral technique or the benefit of making the same materials available without supporting services. For biopsychosocial explanation, the trial is both an intervention success and a reason to study which biological, behavioral, and institutional elements made participation possible and durable.
The Look AHEAD Research Group (2013) studied an intensive lifestyle intervention among adults who already had type 2 diabetes, using a primary composite cardiovascular endpoint. The trial did not establish a statistically significant reduction in that primary outcome despite improvements in several other health-related measures. A lifestyle intervention can therefore improve weight, fitness, or other intermediate outcomes without necessarily changing the particular clinical event endpoint that a study was designed to test.
These trials should be read together rather than ranked through a single global label of success. They enrolled different populations, tested different interventions, and used different primary outcomes and follow-up structures. Diabetes incidence among people at risk and cardiovascular events among people with established diabetes are distinct estimands. Integrating them requires specification of disease stage, baseline risk, available treatment, adherence, and clinically relevant endpoints. The comparison makes clear that the word “health” cannot serve as an interchangeable outcome across all research.
Chronic Pain as a Test of Multilevel Explanation
Chronic pain is a demanding case for biopsychosocial explanation because nociceptive processes, nervous-system adaptation, emotional and attentional processes, movement, sleep, social participation, and clinical treatment may all influence the observed outcome. Pain intensity is a patient-reported experience, while disability and participation reflect additional dimensions. Imaging findings, inflammatory biomarkers, and other biological indicators can contribute to diagnosis, but there is no universal one-to-one relation between tissue findings and reported pain across conditions.
The Cherkin, D. C., Sherman, K. J., Balderson, B. H., Cook, A. J (2016) compared mindfulness-based stress reduction, cognitive behavioral therapy, and usual care for adults with chronic low back pain. Outcomes included functional limitation and pain-related measures. The study shows how psychological interventions can be evaluated against clinical and functional endpoints in a condition with biological and behavioral dimensions. It does not imply that every chronic pain syndrome has the same treatment response or that pain is generated solely by belief or emotion.
Mechanisms require additional attention. A structured intervention may influence attentional strategies, activity pacing, avoidance, sleep, self-efficacy, social engagement, and expectations; an observed treatment effect alone cannot determine which mediator changed the outcome. Treatment participation can also be affected by financial costs, transportation, provider availability, and baseline disability. These factors should be measured as possible modifiers of effectiveness and delivery, not relegated to vague context.
The case underscores the scientific and clinical value of multiple endpoints. A reduction in pain intensity, improved walking function, greater participation in work, and reduced use of unhelpful treatment are individually meaningful changes. They may occur on different timelines and need not move together. The research question should specify which outcome is targeted, whether improvement persists, what adverse effects arise, and whether the intervention remains feasible for people with different levels of disability and support.
Observational Cohorts and Selection Mechanisms
Prospective cohorts provide information about exposures that are difficult, impossible, or unethical to randomize. They can follow material conditions, social networks, health behaviors, symptoms, and biomarkers across years. Repeated observation helps establish temporal order and assess change. Yet temporal precedence alone does not remove confounding, selective participation, differential measurement, or reverse influences operating before baseline. A cohort recruited from a clinic may differ systematically from people with similar disease who never enter care.
The problem of confounding is particularly acute for psychosocial predictors. Stressful work can be associated with socioeconomic position, baseline illness, personality, occupational hazards, and access to treatment. Adjusting for measured education and income may reduce one source of bias while leaving others. If investigators adjust for a mediator such as sleep quality when estimating the total effect of employment on disease, the adjusted estimate may no longer answer their original question. If they analyze only people who obtain a diagnosis, conditioning on diagnosis can create selection bias when diagnosis depends on both exposure and underlying illness.
Observational findings therefore require a structured explanation of why comparison groups should be exchangeable under the intervention contrast of interest. This is partly a design question and partly a theory question. Family comparisons may control some shared factors but can be vulnerable to nonshared confounding and measurement error. Instrumental-variable strategies demand strong assumptions about the instrument's pathways. Mendelian randomization, described by Davey Smith, G., & Hemani, G (2014), offers one family of approaches for specific exposure questions, but genetic proxies rarely encode complex social interventions without ambiguity.
Triangulation is valuable when methods with different sources of bias converge on a narrowly specified claim. A cohort association, a randomized behavioral intervention, and an experimentally measured physiological pathway can provide a stronger integrated case than any one alone. Disagreement is also informative: it may indicate differences in population, exposure definition, follow-up duration, or assumptions. A responsible narrative synthesis retains such disagreements and avoids manufacturing consistency through broad labels.
Natural Experiments and Target-Trial Emulation
Public policy, workplace reforms, changes in treatment access, and environmental interventions can sometimes be investigated through natural experiments. The investigator identifies a change that affects some people or settings and constructs a comparator capable of estimating what would likely have happened without the change. Interrupted time-series, difference-in-differences, regression discontinuity, and related approaches can be informative when their assumptions are plausible. A policy event is not automatically a valid natural experiment simply because it occurred without the researcher's control.
A difference-in-differences analysis, for example, may compare outcomes before and after an exposure change across populations with and without that change. It relies on assumptions about untreated trends, composition, concurrent policies, and measurement continuity. If the intervention changes who seeks care, an observed change in diagnosis counts might reflect detection rather than disease incidence. If implementation differs by institution, the average effect may hide highly variable adherence and service capacity. These are substantive features of the social mechanism.
Hernán, M. A., & Robins, J. M (2016) proposed the explicit emulation of a target randomized trial when using observational data. Their framework requires specification of eligibility, treatment strategies, assignment, follow-up, outcomes, causal contrast, and analytic plan. This discipline is directly applicable to biopsychosocial questions: instead of asking whether “social support causes recovery,” a researcher can ask whether a defined support intervention, initiated at a specified time for eligible patients, changes a prespecified recovery outcome relative to a feasible alternative.
Target-trial emulation does not transform observational datasets into randomized trials. It clarifies which causal question is being attempted and reveals data limitations such as confounders measured only after treatment starts, misaligned time zero, or unavailable intervention histories. Its advantage is analytical transparency. For multilevel health research, the same discipline helps specify interventions on behavior, institutions, and care systems rather than treating domain categories as interventions in themselves.
Mediation, Interaction, and Transportability
Mediation analysis seeks to understand how an exposure affects an outcome through intermediate processes. A multilevel model may hypothesize that housing instability influences sleep, which alters inflammatory regulation, which affects disease outcomes. Estimating the share of an effect carried by this chain is demanding because mediator-outcome confounding, measurement error, exposure-induced confounding, and timing can invalidate simple regression decompositions. The existence of sequential associations does not by itself identify an indirect causal effect.
VanderWeele, T. J (2016) set out distinctions among mediation estimands and the assumptions required for causal interpretations. Researchers can pursue total effects, controlled direct effects, and appropriately defined indirect effects, but these quantities need not answer the same policy question. Intervening on a mediator may be infeasible or may affect outcomes by pathways different from those induced by the exposure. Interventional analogues and sensitivity analyses can make some assumptions more transparent, yet their interpretation remains linked to what the intervention and mediator represent.
Interaction concerns variation in effects across levels of another variable. A support program may have different results in patients with stable housing and patients facing food insecurity. That variation could reflect differences in access, exposure intensity, competing demands, or biological susceptibility. Interaction can be expressed on additive or multiplicative scales, with different implications; a single “significant interaction” label is an incomplete description. Subgroup analyses are particularly susceptible to low power, multiplicity, and selective reporting, so prespecified hypotheses and uncertainty intervals are important.
Transportability asks whether an effect estimated in one trial, clinic, or population would apply to another. Intervention effects depend on baseline risk, treatment fidelity, comorbidity, infrastructure, and social opportunity. A program developed in a highly resourced clinical system may not have the same effect when transportation, staffing, language access, and continuity differ. Transporting causal effects demands knowledge of both how the intervention works and which contextual variables modify its impact. The biopsychosocial framework is most useful when it makes these sources of heterogeneity empirically visible.
Measurement and Construct Validity
Scientific integration begins with valid measurement. “Stress” may denote an event count, an appraisal, a perceived lack of control, a hormone profile, or a composite of related indicators. “Social support” may refer to available helpers, received assistance, perceived responsiveness, or network contact. “Health” may refer to symptoms, morbidity, function, health-related quality of life, or mortality. Calling all of them biopsychosocial variables adds little unless the construct and its operational indicator are defined.
Construct validity concerns whether a measure supports the intended interpretation. Questionnaires can capture beliefs and experience more directly than some physiological assays, but they may also be affected by response styles and recollection. Biomarkers have biological specificity, yet their meaning can depend on timing, assay method, and state variability. Administrative records can show service use or clinical events but often omit unobserved symptoms, informal care, or the reasons an encounter occurred. No instrument grants automatic privileged access to every aspect of a condition.
Multimethod research can compare self-report, clinical assessment, ambulatory monitoring, laboratory assays, records, and direct observation. Agreement is useful when the methods are intended to measure the same construct; disagreement may be meaningful when they capture different dimensions. The analytic plan should explain whether indicators are treated as reflective measurements of one latent construct, formative components of an index, or separate causal variables. Combining them into a composite without a measurement model can hide opposing mechanisms.
Measurement equivalence requires attention across age, language, culture, disability, and clinical stage. A psychological questionnaire validated in a student sample may behave differently in a hospitalized older population. Wearable-device data may systematically exclude people who cannot access or tolerate the device. Health records can miss persons whose care access is limited. Selection and missingness thus affect the substantive causal question as well as statistical precision. Transparent measurement protocols are essential to credible comparisons across levels of analysis.
Digital Phenotyping and Computational Models
Digital devices and administrative systems can measure aspects of activity, sleep, communication, mobility, and treatment engagement at greater temporal density than traditional questionnaires. Computational methods may detect temporal patterns, predict near-term outcomes, and generate hypotheses about causal organization. These capabilities are relevant to biopsychosocial inquiry because many proposed mechanisms involve feedback, time lags, and heterogeneous responses. However, high-frequency data do not automatically resolve causal confounding.
A predictive model learns associations between inputs and outcomes in its training context. It may perform well when the underlying population, measurement practices, and care pathways remain stable. If an institution changes its triage system, a predictive association learned from historical records may no longer hold. Algorithmic features may encode who received care or who possessed a smartphone rather than a biological state. Predictive performance and causal identification therefore require separate evaluations.
von Känel, R (2026) argues that digital phenotyping, computational modeling, and network methods may help make multilevel biopsychosocial relations more testable. This proposal is strongest when paired with designs capable of testing whether predicted links change under interventions. Dynamic models can represent reciprocal pathways, but assumptions about latent states, noise, feedback, and model identifiability must be evaluated against independent data. Greater model complexity can increase apparent fit while making scientific interpretation more fragile.
Privacy and representativeness are central methodological conditions. Continuous monitoring can impose burdens, alter behavior, and expose intimate patterns of movement or social contact. Individuals with restricted connectivity, disability, or unstable housing may be systematically underrepresented. A clinically useful model needs calibration, subgroup performance assessment, external validation, monitoring for drift, and a clearly defined response when an alert is wrong. In this setting, data governance affects the reliability and fairness of the scientific evidence, not only the ethics of implementation.
Critical Scientific Discussion
The biopsychosocial approach offers a compelling inventory of plausible determinants, yet the scientific force of a causal explanation lies in its precision. Any particular claim must identify the exposure or action, the outcome, its timing, and why an observed difference would represent an effect. The framework becomes scientifically productive when a purported interlevel relation can be contradicted by evidence. It becomes harder to evaluate when every observation is interpreted as confirming that biology, psychology, and society are interconnected.
One recurring tension concerns mechanism and intervention. Demonstrating that psychological stress is associated with infection in a controlled exposure setting does not establish that a particular stress-reduction intervention prevents community infection. A robust social gradient in mortality does not identify the effect of one specific income policy. Conversely, a randomized prevention intervention can have a clear total effect without a complete account of the mediator network. Mechanistic research and intervention research should therefore inform each other without treating plausibility as efficacy or efficacy as proof of a unique mechanism.
A second issue concerns levels of explanation. Molecular findings may explain how a disease process operates, but they may say little about why particular populations encounter the initiating exposure. Social variables can predict outcomes across institutions, but they may remain too coarse to specify a physiological mechanism. These are complementary research questions with different units and counterfactuals. Explanatory integration proceeds by connecting them through verified pathways, not by requiring every study to measure every level.
A third challenge involves causal overadjustment and misleading statistical control. A researcher who includes all available biological, psychological, and social covariates in one regression can inadvertently adjust away mediators or condition on colliders. A more elaborate model may look comprehensive while estimating an ill-defined quantity. Explicit diagrams, time ordering, preregistration of causal contrasts, sensitivity analyses, and targeted intervention comparisons are often more valuable than adding further predictors.
The sources examined here also vary greatly in evidentiary strength. Clinical randomized trials can support intervention effects under their assignment and follow-up conditions. Viral-challenge studies constrain exposure and outcome measurement yet leave naturally occurring stress observational. Social-relationship meta-analysis synthesizes observational associations but does not inherit the causal identification of randomized experiments. Mechanistic reviews establish biological plausibility and guide testing but cannot replace patient-important endpoints. These distinctions are substantive results of the synthesis.
Finally, publication bias, selective analysis, incomplete source access, outcome switching, and limited replication can distort any research tradition. Biopsychosocial explanations are especially vulnerable to attractive post hoc narratives because many plausible pathways can be nominated after observing an effect. Prospective specification of competing mechanisms, independent replication, transparent missing-data handling, and negative-control strategies where appropriate are practical safeguards. The scientific purpose is to reduce the number of compatible explanations with each new study.
A Research Agenda for Testable Integration
A cumulative research program can begin with a specific contrast, such as the effect of a defined workplace-scheduling reform on metabolic risk over a stated interval. The investigator then distinguishes the policy exposure from intermediate sleep opportunity, behavioral changes, stress responses, and clinical outcomes. Baseline confounders and time-varying influences should be represented before data collection. This approach can generate a causal model in which direct occupational exposures, behavioral responses, and physiological regulation can be evaluated together without requiring an exhaustive inventory of every social variable.
A second priority is better intervention-component evidence. Multicomponent programs often succeed or fail as packages, making it difficult to know whether their benefits arise through information, practical support, behavioral practice, social coordination, or improved clinical access. Factorial experiments, implementation trials, and appropriately powered mediation studies can discriminate among these pathways. Their interpretation must retain adherence, fidelity, acceptability, and the distribution of opportunity to participate. What matters is which components affect which outcomes for whom and under what conditions.
A third priority is longitudinal mechanism research with validated cross-level timing. Repeated measurement of perceived demands, sleep, physiological markers, and clinically meaningful outcomes can establish temporal patterns that are invisible in cross-sectional work. Such studies should use causal estimands suited to time-varying exposure and confounding. Dynamic models must be compared with simpler alternatives and evaluated in new samples, including settings where interventions perturb at least one key link in the proposed sequence.
A fourth priority is systematic heterogeneity analysis. Studies should report the characteristics of participants excluded from observation or intervention, how outcomes differ across age and disability, and whether effects depend on access to resources. Transportability and equity should be treated as scientific properties of intervention effectiveness, rather than appended as general recommendations. A trial result may be internally valid and still have limited relevance for people living under substantially different material conditions.
Finally, synthesis practices should distinguish evidence tables from causal diagrams, and narrative interpretation from reproducible systematic review. Research teams can register competing models, specify disconfirming outcomes, share analysis code where appropriate, and maintain clear links between primary studies and interpretive claims. A genuinely integrative science accumulates discriminating tests instead of accumulating terminology.
Artificial Era and Human–Artificial Health Systems
The Artificial Era introduces new organizational conditions for psychological and medical care: algorithmic decision support, conversational interfaces, adaptive monitoring, automated triage, and human–artificial collaboration in health information. These systems can influence biopsychosocial pathways by altering how symptoms are interpreted, which options are presented, how attention is allocated, and which professionals or services become accessible. Their causal relevance depends on deployment, reliability, institutional constraints, and the real choices available to patients and clinicians.
One pathway is cognitive and behavioral. A digital health system may provide reminders, explanatory information, or organizational assistance, potentially changing medication routines and treatment participation. Another is relational: repeated interaction may alter perceived support, expectations, or willingness to disclose distress. A third is structural: triage and platform policy may distribute services differently, create new waiting times, or amplify the effects of unequal connectivity. These are research hypotheses and classes of potential effects, not evidence that any particular artificial system has an established therapeutic benefit.
The appropriate evaluation parallels other biopsychosocial interventions. Define the specific artificial-system behavior, comparison condition, relevant human oversight, outcome, follow-up, and population. Examine unintended effects, including inaccurate reassurance, false alarms, overreliance, privacy breaches, and unequal performance. Measure psychological outcomes and clinical outcomes separately, and distinguish improved navigation from improved disease prognosis. Systems that explain their recommendations and permit human correction may be easier to govern, but actual safety and effectiveness still require empirical testing.
Within the philosophical horizon of Artificial Era · From Homo to Artificial, relations between human and artificial systems constitute an expanding field of inquiry. The scientific question is how particular configurations of human, artificial, and institutional action modify observable health processes. This question can be studied through randomized service comparisons, implementation research, longitudinal use patterns, calibrated prediction, and qualitative analysis of real care encounters. The historical significance of artificial participation in health systems does not replace evidence of clinical benefit.
Conclusion
Biopsychosocial causation in health and disease is the study of pathways through which biological processes, psychological activity, social relations, and institutional conditions influence specified health outcomes over time. Its scientific content emerges when domain labels are translated into variables, interventions, mechanisms, comparators, and testable predictions. Biological, psychological, and social explanations often concern connected portions of the same causal process, but the strength of each connection must be established through its own evidence.
The studies considered here reveal a structured pattern. Controlled viral-challenge research provides evidence of psychosocial predictors under standardized pathogen exposure while retaining observational uncertainty about stress. Randomized prevention and behavioral treatment trials establish the effects of specified programs on their chosen endpoints, without identifying every mediator or guaranteeing transfer across diseases. Longitudinal social and developmental studies reveal patterned risks, but face confounding and selection. Mechanistic work identifies plausible signaling pathways and measurement targets, while formal causal methods clarify the assumptions required to interpret associations.
The resulting scientific synthesis is an explanatory discipline organized around precise causal questions. It treats symptom experience, disease pathology, function, and health-care participation as distinguishable outcomes; assigns explicit roles to time, context, and measurement; and compares rival mechanisms through designs capable of separating them. Future progress depends on longitudinal and interventional evidence, reliable cross-level measurement, transparent uncertainty, and careful transport of results to diverse populations and increasingly artificial health environments.
References
- Black, D. S., & Slavich, G. M. (2016). Mindfulness meditation and the immune system: A systematic review of randomized controlled trials. Annals of the New York Academy of Sciences, 1373(1), 13–24. https://doi.org/10.1111/nyas.12998
- Borrell-Carrió, F., Suchman, A. L., & Epstein, R. M. (2004). The biopsychosocial model 25 years later: Principles, practice, and scientific inquiry. Annals of Family Medicine, 2(6), 576–582. https://doi.org/10.1370/afm.245
- Cherkin, D. C., Sherman, K. J., Balderson, B. H., Cook, A. J., Anderson, M. L., Hawkes, R. J., Hansen, K. E., & Turner, J. A. (2016). Effect of mindfulness-based stress reduction vs cognitive behavioral therapy or usual care on back pain and functional limitations in adults with chronic low back pain: A randomized clinical trial. JAMA, 315(12), 1240–1249. https://doi.org/10.1001/jama.2016.2323
- Cohen, S., Doyle, W. J., Turner, R. B., Alper, C. M., & Skoner, D. P. (2003). Emotional style and susceptibility to the common cold. Psychosomatic Medicine, 65(4), 652–657. https://doi.org/10.1097/01.psy.0000077508.57784.da
- Cohen, S., Janicki-Deverts, D., & Miller, G. E. (2007). Psychological stress and disease. JAMA, 298(14), 1685–1687. https://doi.org/10.1001/jama.298.14.1685
- Cohen, S., Tyrrell, D. A. J., & Smith, A. P. (1991). Psychological stress and susceptibility to the common cold. New England Journal of Medicine, 325(9), 606–612. https://doi.org/10.1056/NEJM199108293250903
- Davey Smith, G., & Hemani, G. (2014). Mendelian randomization: Genetic anchors for causal inference in epidemiological studies. Human Molecular Genetics, 23(R1), R89–R98. https://doi.org/10.1093/hmg/ddu328
- Diabetes Prevention Program Research Group. (2002). Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. New England Journal of Medicine, 346(6), 393–403. https://doi.org/10.1056/NEJMoa012512
- Engel, G. L. (1977). The need for a new medical model: A challenge for biomedicine. Science, 196(4286), 129–136. https://doi.org/10.1126/science.847460
- Felitti, V. J., Anda, R. F., Nordenberg, D., Williamson, D. F., Spitz, A. M., Edwards, V., Koss, M. P., & Marks, J. S. (1998). Relationship of childhood abuse and household dysfunction to many of the leading causes of death in adults: The Adverse Childhood Experiences (ACE) Study. American Journal of Preventive Medicine, 14(4), 245–258. https://doi.org/10.1016/S0749-3797(98)00017-8
- Hernán, M. A., & Robins, J. M. (2016). Using big data to emulate a target trial when a randomized trial is not available. American Journal of Epidemiology, 183(8), 758–764. https://doi.org/10.1093/aje/kwv254
- Holt-Lunstad, J., Smith, T. B., & Layton, J. B. (2010). Social relationships and mortality risk: A meta-analytic review. PLOS Medicine, 7(7), e1000316. https://doi.org/10.1371/journal.pmed.1000316
- Kaptchuk, T. J., Friedlander, E., Kelley, J. M., Sanchez, M. N., Kokkotou, E., Singer, J. P., Kowalczykowski, M., Miller, F. G., Kirsch, I., & Lembo, A. J. (2010). Placebos without deception: A randomized controlled trial in irritable bowel syndrome. PLOS ONE, 5(12), e15591. https://doi.org/10.1371/journal.pone.0015591
- Kiecolt-Glaser, J. K., McGuire, L., Robles, T. F., & Glaser, R. (2002). Psychoneuroimmunology and psychosomatic medicine: Back to the future. Psychosomatic Medicine, 64(1), 15–28. https://doi.org/10.1097/00006842-200201000-00004
- Look AHEAD Research Group. (2013). Cardiovascular effects of intensive lifestyle intervention in type 2 diabetes. New England Journal of Medicine, 369(2), 145–154. https://doi.org/10.1056/NEJMoa1212914
- Marmot, M. (2005). Social determinants of health inequalities. The Lancet, 365(9464), 1099–1104. https://doi.org/10.1016/S0140-6736(05)71146-6
- McEwen, B. S. (1998). Protective and damaging effects of stress mediators. New England Journal of Medicine, 338(3), 171–179. https://doi.org/10.1056/NEJM199801153380307
- Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688. https://doi.org/10.1093/biomet/82.4.669
- Robins, J. M., Hernán, M. Á., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. https://doi.org/10.1097/00001648-200009000-00011
- Steptoe, A., & Kivimäki, M. (2012). Stress and cardiovascular disease. Nature Reviews Cardiology, 9(6), 360–370. https://doi.org/10.1038/nrcardio.2012.45
- VanderWeele, T. J. (2016). Mediation analysis: A practitioner's guide. Annual Review of Public Health, 37, 17–32. https://doi.org/10.1146/annurev-publhealth-032315-021402
- von Känel, R. (2026). Applying a causal biopsychosocial model to guide medicine and psychiatry fifty years after Engel. Communications Medicine, 6, Article 322. https://doi.org/10.1038/s43856-026-01700-9