A Path-Analysis Framework using the All of Us Research Program Dataset
Presenter: Kathleen Morales
Core Focus: Mapping chronic environmental exposures to clinical outcomes through a structural capacity filter — moving beyond outcomes-based resilience toward a mechanistic, physiologically grounded model of adaptive capacity.
The Core Problem: Mechanisms, Not Just Outcomes
Contemporary resilience science has long grappled with a fundamental epistemological asymmetry: the field overwhelmingly documents who survives adversity while largely neglecting the question of at what cost and through what mechanism. This framing gap has concrete consequences for translational research and intervention design.
The Trap: Outcome-Centrism
Current resilience research predominantly defines resilience by observable outcomes — "doing well" under adverse conditions. This conflates the result of adaptation with the process of adaptation, obscuring the biological and structural costs embedded in successful coping. A person may appear functionally resilient while silently accruing physiological debt.
The Flaw: Mindset Reductionism
Dominant public health and psychological frameworks continue to operationalize resilience as a dispositional or cognitive trait — a "mindset" — while systematically neglecting the physiological hardware (autonomic nervous system tone, metabolic reserve) and environmental matrix (neighborhood safety, food access, social density) that either enable or constrain adaptive capacity. Grit narratives, in particular, can obscure structural inequity.
The Goal: Structural, Measurable Adaptive Capacity
This research program proposes a fundamental reorientation: operationalizing Adaptive Capacity as a quantifiable, multi-pillar construct subject to empirical decomposition. Rather than asking "is this person resilient?", the framework asks "what is the current bandwidth of this system's capacity to absorb and recover from allostatic burden?" This reframing enables precision intervention and populational surveillance.
The Path Model: A Logic Flow from Exposure to Clinical Outcome
The proposed structural equation model articulates a four-stage causal chain, linking upstream environmental exposures to downstream clinical failures via two mediating constructs: the Capacity Pillars and Allostatic Load. Each stage is independently operationalizable within the All of Us data environment.
The capacity pillars function as pipe width in this hydraulic metaphor: a larger pipe (higher capacity) can absorb greater stressor throughput before allostatic pressure rises to clinically detectable levels. The model treats allostatic load not as an endpoint but as a mediating mechanism — the thermodynamic exhaust of sustained adaptation — that precedes and predicts discrete clinical diagnoses. This architecture permits decomposition of total effects into direct and mediated pathways, enabling identification of which capacity domain offers the greatest intervention leverage per unit of stressor burden.
Delineating Stress: Separating Force from Firmware
A critical methodological distinction in this framework is the separation of chronic environmental stress (the objective external force acting on the system) from perceived stress (the subjective internal appraisal of that force). Conflating these two constructs — as much prior research has done — introduces construct validity threats and obscures the pathway by which objective burden is transduced into physiological cost.
Chronic Stress: The Objective Force
Operationalized using administrative and geospatial data sources embedded in the All of Us survey architecture:
SDOH Surveys: Structured assessments of housing instability, food insecurity, transportation barriers, and financial strain — validated modules within the PPI instrument battery
Area Deprivation Index (ADI): Census-tract–level composite of income, education, housing quality, and employment; a validated geospatial proxy for structural disadvantage
Family Health History: Polygenic and early environmental loading that establishes baseline allostatic susceptibility independent of current exposures
Perceived Stress: The Subjective Firmware
Operationalized using validated psychometric instruments available within the All of Us PPI battery:
PSS-10 (Perceived Stress Scale): 10-item self-report instrument measuring the degree to which life circumstances are appraised as unpredictable, uncontrollable, and overwhelming; widely normed across demographic groups
GAD-7 / PHQ-9: Generalized Anxiety Disorder and Patient Health Questionnaire instruments capturing the affective and somatic sequelae of chronic stress appraisal — serving as downstream indicators of firmware dysregulation
Separating these constructs allows the model to test whether the appraisal pathway operates as a distinct mediating route or is fully subsumed by objective structural burden — a question with direct implications for intervention targeting.
Pillars I & II: Hardware and Firmware of Adaptive Capacity
The first two capacity pillars address the biological substrate and cognitive-affective architecture through which environmental stressors are processed. These are the most proximal determinants of allostatic cost — the layers of the system that either buffer or amplify incoming burden before it registers as physiological dysregulation.
Pillar I — Physiological Hardware
The physiological pillar operationalizes the biological infrastructure available for mounting and sustaining an adaptive response. Data sources within the All of Us environment include:
EHR Laboratory Values: Metabolic panel components (HbA1c, fasting glucose, lipid fractions), inflammatory markers (CRP, WBC), and renal/hepatic indices — collectively capturing the functional status of major organ systems under allostatic demand
Fitbit-Derived Heart Rate Variability (HRV): A validated passive biomarker of autonomic nervous system balance and cardiovascular reserve. HRV reflects the parasympathetic "brake" capacity that governs stress recovery and serves as a real-time index of allostatic buffer available in the ANS
Resting Heart Rate Trajectories: Longitudinal trends in resting HR from continuous wearable data capture secular shifts in cardiorespiratory fitness and sympathetic tone over time
Pillar II — Psychological Firmware
The psychological pillar captures the cognitive and affective software running on the physiological hardware — the appraisal and regulatory processes that determine how threat signals are interpreted and managed:
Threat Appraisal Scores: Derived from validated PPI survey modules, these scores index the tendency to evaluate ambiguous or challenging stimuli as threatening versus manageable — a core determinant of the magnitude of stress-axis activation per unit of objective stressor
Resilience PPI Modules: The All of Us PPI battery includes validated resilience-adjacent instruments (including items from the Brief Resilience Scale and Connor-Davidson Resilience Scale domains) that capture adaptive self-efficacy, cognitive flexibility, and recovery orientation
Emotional Regulation Proxies: GAD-7 and PHQ-9 scores, re-operationalized not as diagnostic endpoints but as inverse indicators of firmware capacity — higher symptom burden indicating reduced regulatory bandwidth available for allostatic buffering
Pillars III & IV: Agency and Ecological Matrix
The second pair of capacity pillars extends the model beyond the individual organism to capture behavioral agency and the environmental context within which adaptation occurs. These pillars acknowledge that structural capacity is not solely an intrinsic property but is co-determined by the conditions the system inhabits and the daily behavioral patterns that either replenish or deplete its reserves.
Pillar III — Behavioral Agency
Behavioral patterns represent the enacted expression of adaptive capacity — the degree to which an individual can translate intent into health-sustaining behavior. Crucially, behavioral regularity (not just average behavior) is posited as the operative mechanism:
Sleep Regularity Index (SRI): Wearable-derived metric quantifying the day-to-day consistency of sleep/wake timing across 7–14 day windows. SRI demonstrates superior associations with metabolic and cardiometabolic outcomes compared to average sleep duration alone, reflecting the chronobiological cost of circadian disruption
Step Consistency Index: Coefficient of variation in daily step counts derived from Fitbit longitudinal streams — capturing the stability of physical activity behavior over time. High variance (low consistency) may reflect social and occupational instability that constrains behavioral agency
Activity-Rest Coupling: The coherence between activity patterns and circadian light-dark cycles, derivable from Fitbit accelerometry, as an index of behavioral entrainment to environmental rhythms
Pillar IV — Ecological Matrix
The ecological pillar operationalizes the social and physical environment as an active determinant of adaptive capacity — not merely as a background variable but as a structural resource that either expands or contracts the system's buffering range:
Neighborhood Safety: Self-reported safety perceptions combined with ADI-derived neighborhood-level indicators; chronic safety threat activates sustained vigilance circuits that consume allostatic resources independently of acute stressor exposure
Food Security: SDOH-derived food access variables, recognizing that nutritional substrate availability constrains both metabolic hardware performance and behavioral agency (cooking, meal regularity)
Social Support Network Density: PPI-derived indices of social connectedness, perceived support availability, and network diversity — capturing the degree to which relational infrastructure can buffer allostatic demand through co-regulation mechanisms and resource sharing
Measuring the "Tax": The Allostatic Load Index
Allostatic load (AL) occupies the critical mediating position in the path model — it is neither the stressor input nor the clinical failure endpoint, but the cumulative physiological cost of sustained adaptation. In thermodynamic terms, AL represents the "exhaust heat" of the adaptive engine: the irreversible entropic cost of repeatedly mobilizing and failing to fully resolve allostatic responses. Its operationalization must therefore be both biologically grounded and feasible within the constraints of a large observational dataset.
The 10-Biomarker Composite Index
Following the MacArthur Studies of Successful Aging and subsequent refinements by Seeman, McEwen, and colleagues, AL is computed as a composite of biomarkers spanning four physiological systems. Each biomarker is scored against an at-risk threshold (typically the quartile indicating highest risk), and scores are summed to yield an integer AL index (range 0–10):
Cardiovascular: Systolic and diastolic blood pressure — capturing the sustained hemodynamic cost of sympathoadrenal activation
Metabolic: HbA1c (glycemic dysregulation), BMI (adiposity as metabolic reserve proxy), total cholesterol and HDL ratio (dyslipidemia)
Inflammatory: High-sensitivity CRP — the most sensitive available EHR-extractable marker of low-grade systemic inflammation driven by chronic stress-axis activity
Neuroendocrine/Renal: Additional biomarkers to be specified per EHR data availability, potentially including eGFR, albumin, or waist circumference where available
Feasibility & Data Quality Thresholds
A fundamental challenge in large cohort AL computation is non-random missingness in clinical laboratory data. To preserve index validity while maintaining adequate sample size, the following decision rule is applied:
Minimum Biomarker Threshold: 7 of 10 — Participants must have at least 7 of the 10 designated biomarkers documented in their linked EHR records to be included in AL computation. This threshold follows precedents in the AL literature and balances completeness against exclusion bias
Temporal Window: Biomarker values will be extracted within a defined observation window (e.g., ±18 months from survey completion) to ensure temporal alignment between self-report and laboratory measures
Longitudinal Trajectory: Where multiple time-point EHR observations are available, AL trajectory slope (rate of change per year) will be computed as a dynamic, time-varying outcome — enabling analysis of AL accrual velocity as a function of capacity pillar scores
Sensitivity analyses will evaluate index stability across 7-, 8-, and 9-biomarker subsets to characterize the robustness of AL estimates to missingness patterns.
Cohort Identification: The Triple-Check Protocol
The analytical cohort is drawn from the All of Us Research Program — a longitudinal, precision medicine initiative funded by the NIH with enrollment exceeding n ≈ 400,000 participants across diverse demographic and geographic strata. The program's multimodal data architecture — integrating self-reported surveys, linked EHR records, and passive wearable sensor streams — makes it uniquely suited to the multi-pillar operationalization required by this framework. However, the richness of this resource requires a stringent inclusion protocol to ensure analytical coherence.
1
PPI Survey Completion
Participants must have completed the specific All of Us PPI (Personal and Family Health History, SDOH, and Lifestyle) survey modules required to operationalize Pillars II, III, and IV, as well as the stress exposure constructs. Partial survey completion is evaluated on a module-by-module basis; participants missing critical psychometric instruments (PSS-10, resilience modules) are excluded from pillar-specific analyses but may contribute to subanalyses where their data are complete.
2
EHR Laboratory Data (7/10 Biomarkers)
Participants must have linked EHR records containing a minimum of 7 of the 10 designated allostatic load biomarkers within the defined temporal observation window. EHR linkage quality will be assessed using data completeness metrics available within the All of Us workbench. Demographic characteristics of excluded participants will be tabulated to evaluate selection bias potential and inform missing-data sensitivity analyses.
3
Fitbit Wearable Data Availability
Participants must have a minimum duration of valid Fitbit data (proposed threshold: ≥ 30 days of wear with ≥ 600 minutes of wear per day) to enable computation of HRV indices, Sleep Regularity Index, and step consistency metrics. Fitbit data availability within All of Us is non-random and skewed toward participants with higher socioeconomic resources; inverse probability weighting or propensity score adjustment will be applied to correct for wearable adoption bias in primary analyses.
Expected analytic N: Applying all three inclusion criteria to the ~400,000 participant base is projected to yield an analytic cohort in the range of n = 20,000–60,000 — sufficient for stable path-model estimation with adequate power for subgroup analyses by race/ethnicity, age stratum, and ADI quintile. Exact projected N will be computed via feasibility queries in the All of Us Researcher Workbench prior to protocol finalization.
Statistical Hypotheses
Two primary inferential questions drive the analytical strategy, each corresponding to a distinct modeling approach within the structural equation modeling (SEM) and longitudinal data analysis frameworks available for the All of Us dataset. Both hypotheses treat Adaptive Capacity as a latent construct composed of the four measured pillar composites.
Hypothesis 1 — Mediation: Capacity as "Pipe Width"
Formal Statement: The relationship between chronic environmental stressor burden (ADI + SDOH composite) and allostatic load (AL index) is significantly mediated by Adaptive Capacity (four-pillar composite), such that the indirect pathway (Burden → Capacity → AL) accounts for a statistically and practically significant proportion of the total effect.
Analytical Approach: Structural equation modeling with latent variable specification for the Adaptive Capacity construct. Bootstrapped confidence intervals (n = 5,000 resamples) will be used for indirect effect estimation following Baron-Kenny decomposition and Hayes PROCESS-equivalent procedures. Covariates include age, sex, race/ethnicity, and enrollment site.
Directional Prediction: Higher Capacity (wider pipe) → weaker Burden–AL association, with capacity operating as a significant partial or full mediator. Differential mediation by pillar is of secondary interest — which pillar accounts for the greatest proportion of the indirect effect?
Hypothesis 2 — Longitudinal: Capacity Predicting AL Accrual Velocity
Formal Statement: Among participants with ≥ 2 EHR observation time points, baseline Adaptive Capacity score significantly predicts the rate of allostatic load accrual (AL trajectory slope) over the follow-up period, controlling for baseline AL and stressor burden. Low capacity at baseline predicts accelerated AL accumulation ("pipe leaking") and earlier clinical diagnosis of T2D, hypertension, or CVD.
Analytical Approach: Latent growth curve modeling (LGC) or mixed-effects regression for AL trajectory estimation. Time-to-first-diagnosis will be modeled using Cox proportional hazards regression with capacity score as a time-varying covariate. The hypothesis specifically tests whether capacity operates as an effect modifier of the Burden → AL relationship over time — i.e., whether the Burden × Capacity interaction term on AL slope is significant and negative.
Clinical Significance Threshold: A meaningful deceleration in AL accrual — operationally defined as ≥ 0.5 SD reduction in annual AL slope per SD increase in capacity — will be considered clinically relevant, guided by existing AL trajectory literature in comparable cohorts.
Conclusion & Next Steps: From Mindset to Structural Engineering
This research program represents a deliberate epistemological shift in how resilience is conceptualized, measured, and ultimately translated into public health intervention. By grounding adaptive capacity in measurable physiological, psychological, behavioral, and ecological constructs — and by treating allostatic load as the quantifiable cost of adaptation rather than a residual confound — the framework creates a platform for precision-grade resilience science.
Impact: Resilience Literacy as a Testable, Data-Driven Construct
The primary scientific contribution is the replacement of resilience as narrative with resilience as measurable system capacity. By operationalizing each pillar within the All of Us data environment, the framework generates empirically falsifiable hypotheses, enabling cumulative, replicable science. Population-level capacity surveillance becomes feasible — identifying communities, demographic strata, and life-course windows where capacity deficits create concentrated allostatic vulnerability. This transforms resilience from an individual-level virtue into a public health surveillance target amenable to structural intervention.
Intervention: Structural Engineering Over Vague Mindfulness
The framework's most direct translational implication is the redirection of intervention resources from cognitive reframing programs (whose population-level efficacy evidence remains modest and whose benefit is likely mediated by the very structural capacity factors modeled here) toward upstream structural engineering: neighborhood-level food and safety environment improvements that expand ecological pillar capacity; chronobiologically informed sleep regularity interventions targeting SRI; autonomic conditioning protocols leveraging HRV as a biofeedback target; and SDOH-responsive clinical workflows that integrate ADI data into risk stratification. Each pillar becomes a distinct, actionable intervention domain.
Immediate Next Steps
Priority actions for moving from conceptual framework to active protocol include: (1) Feasibility query in the All of Us Researcher Workbench to determine analytic N under the Triple-Check inclusion criteria; (2) Variable mapping — aligning each pillar construct to specific All of Us survey fields, EHR data elements, and Fitbit metrics with documented availability; (3) IRB and data access protocol finalization under the All of Us governance framework; (4) Measurement model specification — confirmatory factor analysis of the four-pillar latent capacity construct prior to full SEM estimation; (5) Pre-registration of hypotheses, analysis plan, and primary outcomes on OSF or ClinicalTrials.gov to ensure transparency and replicability.
Closing Frame: The question is not whether people are resilient. The question is: what is the structural bandwidth of their adaptive system, and what are we, as a society, doing to either expand or constrict it? This framework offers the measurement tools to begin answering that question with the rigor it demands.