Interdependence-Aware Attribution in Real-World Evidence: A Case Study on Semaglutide Weight Loss Outcomes

Abstract

Real-world evidence (RWE) for semaglutide (Ozempic) shows significant but highly heterogeneous weight loss, typically −4% to −12% at 6–12 months, influenced by persistence, early response, lifestyle, and care intensity. Current RWE analyses describe averages and associations but rarely quantify the attributable share of observed individual outcomes to the drug versus contextual amplifiers.

This article applies a FAIR (Fairness, Attribution, Inference, Reproducibility) interdependence-aware framework to decompose a fixed observed outcome (−12% at 9 months) across patient scenarios calibrated on published RWE. Results reveal drug-attributable shares varying from 55% (−6.6 points) in high-support profiles to 78% (−9.4 points) in low-support ones, highlighting up to 23-point swings due to context.

Persistence alone shifts attribution by ~15 points. This simulation-based approach bridges RWE limitations, aiding regulatory interpretation, labeling, and trial design by providing quantifiable attribution respecting factor interdependencies.

Implications for clinical development emphasize personalized outcome modeling beyond group-level effects.​​

Introduction

Semaglutide, a glucagon-like peptide-1 (GLP-1) receptor agonist approved as Ozempic for type 2 diabetes and as Wegovy for weight management, has demonstrated robust and clinically meaningful weight loss in randomized controlled trials (RCTs), as well as in real-world evidence (RWE) settings. Across observational studies, mean body-weight reductions typically range from approximately −4% to −8% at six to 12 months, with larger losses of −8% to −12% reported among persistent users and early responders.

However, these averages conceal substantial inter-individual heterogeneity, a consistent finding across RWE cohorts and healthcare systems.^1–4,9 Multiple real-world factors have been identified as key drivers of this heterogeneity, including treatment persistence (often operationalized as proportion of days covered [PDC]), early weight response, lifestyle adherence, intensity of clinical follow-up, and titration patterns.

Patients with high persistence and structured care pathways tend to achieve greater weight loss, whereas discontinuation, delayed titration, or limited behavioral support are associated with attenuated outcomes.^1,2,6-8 As a result, observed weight loss in routine practice reflects a composite of pharmacological effects and contextual amplifiers that are largely absent or controlled in RCTs.

Importantly, these contextual amplifiers extend beyond treatment mechanics to include the patient’s broader care environment, social support structures, and behavioral context, which are not captured by weight change alone. Despite the growing regulatory and clinical reliance on RWE, current analytical approaches remain limited in their ability to interpret individual-level outcomes.

Most RWE analyses describe cohort averages, associations, or subgroup differences, but do not address a central interpretive question: how much of a given patient’s observed outcome can reasonably be attributed to the drug itself, as opposed to real-world contextual factors such as adherence, care intensity, or behavioral support? For example, when two patients both achieve a −12% weight loss at nine months, existing methods provide no standardized way to determine whether this outcome primarily reflects pharmacological efficacy, favorable context, or a combination of both.

This attribution gap constrains the utility of RWE for regulators, payers, and clinical trial designers, particularly in areas such as labeling interpretation, external control arms, and enrichment strategies. To address this limitation, we apply an interdependence-aware attribution framework grounded in FAIR principles (Fairness, Attribution, Inference, Reproducibility) to decompose a fixed observed outcome into drug-attributable and context-attributable components.

Rather than estimating a new causal effect, the framework conditions on an observed weight-loss outcome and partitions its explanation across key real-world drivers while explicitly accounting for their interdependencies. Using published semaglutide RWE to calibrate plausible parameter ranges,^1,2,9 we illustrate how identical outcomes (−12% weight loss at 9 months) can yield materially different drug-attributable shares across patient scenarios.

By quantifying these attribution swings, the approach aims to enhance the interpretability of RWE at the individual level and to support more informed use of real-world data in regulatory evaluation and clinical development. Most real-world evidence methods remain population-centric, even when stratified or subgroup analyses are performed.

In contrast, the present framework explicitly shifts the unit of interpretation from the cohort to the individual, conditioning on an observed outcome and asking how that outcome can be plausibly explained across multiple contributing factors. This individual-level perspective represents a methodological step beyond average treatment effects, aligning RWE interpretation with real-world decision-making, which ultimately occurs patient by patient.

This individual-level attribution logic mirrors approaches long used in medico-legal and compensation settings, in which population-level epidemiological evidence must be translated into individual determinations of causation or contribution. In such contexts, including occupational and social insurance medicine, decision-makers routinely start from population-based risk estimates and work toward individualized attribution by accounting for personal exposure histories, contextual modifiers, and interdependencies among factors.

The present framework applies the same foundational logic to real-world evidence interpretation, reframing heterogeneous outcomes not as statistical noise but as explainable variation at the individual level.

Methods

The framework decomposes observed weight loss (Δweight = −12% at 9 months) into attributable shares for drug exposure and amplifiers (persistence/PDC, early response proxy, lifestyle proxy, care intensity proxy, titration proxy), accounting for interdependencies via simulation calibrated on RWE literature.

Baseline "standard patient:" PDC=0.85, early response=0.70, lifestyle=0.60, care=0.70, titration=0.60. Scenarios vary one or more factors while conditioning on −12%:

• High/low support: PDC/lifestyle/care/early/titration at 0.80–0.90 vs. 0.30–0.60.
• Persistence high/low: PDC=0.95 vs. 0.55 (others fixed).
• Early responder/late: early=0.85 vs. 0.35 (others fixed).

Median shares (with P10–P90 intervals) are derived from interdependence-aware simulation, yielding % parts and absolute points summing to −12%. No primary data; illustrative yet grounded in RWE magnitudes (e.g., −5% to −7% means, persistence-driven heterogeneity).​​

Methodologically, the framework draws on attribution principles originally developed in settings where individual-level explanation is required, such as medico-legal causation analysis and risk attribution. By combining causal reasoning with game-theoretic Shapley-style allocation, the approach accounts for interdependencies among factors rather than treating them as independent contributors.

This allows a proportionate and internally consistent decomposition of an observed outcome at the individual level.

Results

For the standard profile, drug share is 63% (−7.6 points; P10–P90: −8.2 to −6.8), context 37% (−4.4 points). Decomposition: persistence 15% (−1.82), lifestyle 9% (−1.08), early 6.5% (−0.78), care 3.9% (−0.47), titration 2.1% (−0.25).​

From Individual Attribution to Population Effectiveness: Regulatory and System-Level Simulations

Regulatory approval establishes that a medicine is effective under controlled conditions, but post-approval decision-making increasingly focuses on whether that effectiveness is realized in real-world use. Regulators, payers, and health systems share a common objective: maximizing the real-world benefit of authorized therapies.

However, conventional real-world evidence often conflates pharmacological performance with variability in care delivery, persistence, and patient support, limiting its utility for guiding post-approval action. By enabling attribution at the individual level, interdependence-aware frameworks also make it possible to move beyond individual explanation toward population-level simulation.

Aggregating attributable shares across real-world patient profiles allows estimation of how much of the observed population outcome is currently realized through pharmacological exposure, and how much is mediated by modifiable contextual factors. This perspective shifts the regulatory question from “Does the drug work?” to “Under which conditions does the drug deliver its full potential?”

Simulation Framework

Using the attribution results described above, we simulated population-level outcomes under alternative real-world scenarios. The simulations are not intended to estimate a new causal effect, but to illustrate how changes in persistence and care context—within plausible ranges documented in RWE—would alter the realized effectiveness of semaglutide at the population level.

We considered a stylized population of patients receiving semaglutide, distributed across three persistence profiles reflecting common RWE patterns:

• High persistence/structured support (30%)
• Intermediate persistence/standard care (40%)
• Low persistence/limited support (30%)

Drug-attributable shares for each profile were derived from the individual-level attribution results (approximately 55%, 63%, and 78%, respectively).

Simulation 1: Baseline Real-World Effectiveness

Under current real-world conditions, the aggregated drug-attributable share of observed weight loss across the population was estimated at approximately 65%, with the remaining 35% attributable to contextual amplifiers such as persistence, care intensity, and behavioral support.

This result highlights a critical regulatory insight: a substantial portion of the observed efficacy of semaglutide is already mediated by the surrounding care environment, rather than pharmacology alone. Importantly, this does not diminish the intrinsic value of the drug but clarifies how that value is realized in practice.

Simulation 2: Improving Persistence and Care Pathways

We then simulated a modest, policy-relevant improvement in real-world conditions: a 10-15 percentage point increase in the proportion of patients achieving high persistence and structured follow-up, consistent with achievable gains reported in adherence-improvement programs. Under this scenario, the aggregated drug-attributable share increased to approximately 70%–72%, accompanied by a higher absolute population-level weight loss despite unchanged pharmacological properties.

In other words, the same authorized medicine delivered meaningfully greater real-world benefit solely through changes in how it was supported and implemented.

Interpretation and Regulatory Implications

From a regulatory perspective, these simulations demonstrate that post-approval efficacy is not a fixed property of the molecule, but an emergent property of the interaction between the drug and the health system. Attribution-based modeling provides regulators with a quantitative framework to:

• Distinguish intrinsic drug performance from system-dependent amplification.
• Assess whether observed underperformance reflects pharmacology or implementation gaps.
• Identify leverage points where non-pharmacological interventions can enhance real-world benefit.

Implications for Sponsors and Developers

While not prescriptive, the same simulations implicitly convey a message to sponsors. If a substantial share of real-world outcomes is mediated by persistence, care pathways, and patient support, then investments in these domains represent a direct extension of efficacy rather than an external add-on.

Attribution modeling thus reframes patient support programs, adherence initiatives, and care integration not as ancillary activities, but as mechanisms that unlock a larger fraction of the drug’s authorized potential. In this sense, interdependence-aware attribution aligns the interests of regulators, health systems, and sponsors: all benefit from understanding not only whether a medicine works, but how to ensure that it works as effectively as possible in real life.

Discussion

At nine months, these findings quantify how substantially real-world context modifies the interpretation of identical weight-loss outcomes with semaglutide. For a fixed −12% weight reduction, the estimated drug-attributable share ranged from 55% to 78% across plausible patient scenarios, with persistence alone shifting attribution by approximately 16 percentage points.

These magnitudes provide quantitative evidence that identical real-world outcomes are not solely determined by pharmacological exposure but are substantially shaped by contextual factors such as treatment persistence, care intensity, and behavioral support, all of which are embedded within the patient’s care environment.⁷

This attribution perspective addresses a key limitation of conventional RWE analyses, which typically conflate pharmacological effects with real-world amplifiers such as adherence, care intensity, and behavioral support. By conditioning on an observed outcome and decomposing its explanation across interdependent drivers, the framework clarifies why similar clinical results may reflect fundamentally different underlying mechanisms.

In this respect, the model operationalizes the widely acknowledged principle that real-world outcomes reflect “treatment plus context” by translating it into quantifiable, patient-level attribution. From a broader health perspective, these findings underscore that weight loss alone is an incomplete descriptor of real-world treatment impact.

Patients achieving similar weight reduction may experience markedly different trajectories in physical functioning, psychological well-being, and social participation, depending on their surrounding environment and support structures. Administrative and social medicine data consistently show that care pathways and social context influence adherence.

They also shape longer-term functional and psychosocial outcomes. By making the contribution of such contextual elements explicit, attribution modeling offers a pathway to interpret pharmacological effectiveness within a more comprehensive conception of health.

A notable implication of the attribution results is the magnitude of contextual contributions. Across plausible patient profiles, factors related to persistence, care intensity, and behavioral context account for a substantial share of observed outcomes—sometimes approaching or exceeding one-third of the total effect.

These findings suggest that social and care-related determinants of health exert a stronger quantitative influence on individual outcomes than is often assumed when interpreting RWE descriptively. From a social and health systems perspective, making contextual contributions explicit reframes them from unmeasured confounders into quantifiable and potentially actionable determinants.

In settings such as social insurance or public health programs, this distinction is critical, as it clarifies which portions of observed outcomes are modifiable through system-level interventions rather than pharmacological innovation alone. From a clinical development standpoint, these findings have several practical implications.

• First, attribution-aware analyses can inform endpoint interpretation by distinguishing drug-driven responses from context-driven amplification, particularly when evaluating high responders in RWE cohorts.
• Second, the framework may support enrichment and stratification strategies by identifying profiles in which observed outcomes are more likely to reflect pharmacological effects.
• Third, such decomposition can enhance the interpretability of external control arms and hybrid RCT–RWE designs, where differences in persistence or care intensity may otherwise bias comparisons.

For regulators and payers, interdependence-aware attribution offers a complementary tool for evaluating RWE-based claims. Rather than relying solely on average effects, this approach supports more nuanced interpretation of individual and subgroup outcomes, helping to distinguish intrinsic drug performance from modifiable contextual factors relevant to labeling, coverage decisions, and real-world implementation.

Beyond individual interpretation, attribution-based decomposition also enables a reverse extrapolation from individual profiles back to population-level insights. By simulating distributions of attributable shares across real-world patient profiles, the framework allows aggregation of individual explanations into population-level projections that remain faithful to heterogeneity.

This creates a bridge between individual outcome interpretation and medico-economic modeling, enabling scenario-based simulations of how changes in persistence, care pathways, or social context could shift population outcomes, resource utilization, and cost-effectiveness.

From Attribution to Action: Regulatory and System-Level Implications Beyond the Molecule

A central implication of interdependence-aware attribution is that real-world effectiveness should not be viewed as a fixed property of the molecule alone, but as the result of an interaction between pharmacological exposure and the health and social system in which the drug is deployed. While regulatory approval establishes efficacy under controlled conditions, regulators increasingly face a second-order question: how to ensure that authorized therapies deliver their maximal potential benefit in routine care.

Attribution-based decomposition provides a quantitative framework to address this question. By aggregating individual-level attributable shares across real-world patient profiles, it becomes possible to estimate how much of the observed population outcome is currently realized through pharmacological exposure, and how much is mediated by contextual factors such as persistence, care pathways, and behavioral support.

Importantly, this approach preserves heterogeneity rather than averaging it away, allowing regulators to distinguish intrinsic drug performance from implementation-dependent amplification. To illustrate this perspective, we conducted population-level simulations grounded in the individual attribution results described above.

Under baseline real-world conditions—reflecting commonly observed distributions of treatment persistence and care intensity—the aggregated drug-attributable share of observed weight loss was approximately two-thirds, with the remaining one-third attributable to contextual amplifiers. This finding suggests that a substantial fraction of real-world effectiveness is already contingent on the surrounding care environment, rather than on pharmacology alone.

We then simulated a modest but plausible improvement in real-world implementation, such as a 10–15 percentage point increase in the proportion of patients achieving high persistence and structured follow-up. Under this scenario, the aggregated drug-attributable contribution increased meaningfully, accompanied by higher absolute population-level weight loss despite unchanged pharmacological properties.

These simulations illustrate that real-world efficacy can be enhanced without altering the molecule itself, solely by modifying system-level determinants that are already known to influence outcomes. From a regulatory standpoint, this reframes post-approval evaluation.

Rather than interpreting deviations between trial efficacy and real-world outcomes as evidence of diminished drug performance, attribution-based modeling allows regulators to identify whether such gaps arise from pharmacology or from modifiable aspects of care delivery. This distinction is particularly relevant for labeling interpretation, post-authorization evidence generation, and the evaluation of external control arms, where differences in persistence or care intensity may otherwise bias conclusions.

The same framework also carries implicit implications for sponsors and developers. When a substantial share of observed outcomes is mediated by persistence, care pathways, and patient support, these elements should be viewed not as ancillary programs but as extensions of therapeutic effectiveness.

Attribution modeling makes visible how investments in adherence support, education, and care integration can unlock a larger fraction of a drug’s authorized potential. In this sense, real-world effectiveness becomes a shared responsibility across regulators, health systems, and industry.

Finally, these findings highlight a broader shift toward a more comprehensive conception of health. By quantifying the contribution of contextual and social determinants alongside pharmacological exposure, attribution-based approaches align outcome interpretation with physical, behavioral, and social dimensions of health.

For regulators and health systems tasked with maximizing population benefit, this perspective provides a pragmatic, quantitative basis for moving beyond “drug versus context” debates toward coordinated strategies that optimize real-world impact.

Implications for Social Insurance Systems and Health Expenditure

Beyond regulatory and clinical development considerations, interdependence-aware attribution has important implications for social insurance and disability systems that manage the downstream consequences of chronic disease. Obesity is a well-established risk factor for a wide range of functional limitations, work disability, and long-term health expenditures.

Social insurance institutions and national payers are therefore less concerned with weight change per se than with how obesity-related trajectories translate into disability risk, benefit claims, and cumulative healthcare costs over time. From this perspective, real-world outcomes associated with weight reduction must be interpreted cautiously.

A given degree of weight loss may correspond to very different functional, occupational, and economic trajectories depending on the surrounding care environment, persistence, and social context. Systems responsible for disability determination and long-term compensation routinely confront this challenge: population-level epidemiological evidence must be translated into individual-level assessments of contribution, risk, and expected future burden.

Attribution-based reasoning, long used in medico-legal and insurance contexts, provides a natural bridge between these domains. Applying interdependence-aware attribution to real-world evidence enables social insurance systems to move beyond binary interpretations of treatment impact.

By decomposing observed outcomes into pharmacological and contextual contributions, institutions can better understand which portions of observed health improvement are likely to translate into sustained functional gains, reduced disability risk, and lower long-term expenditures—and which depend on modifiable system-level factors such as care continuity, adherence support, and workplace or social accommodations.

Importantly, this perspective does not promote any specific therapy. Rather, it reframes post-approval efficacy as a shared system-level outcome. For payers and social insurance institutions, attribution-based simulation offers a way to model how changes in care pathways or support structures—independent of the molecule itself—could alter population-level disability incidence, benefit utilization, and healthcare spending.

For example, simulations can explore how improving persistence or care integration among high-risk populations might reduce downstream disability claims, even if average weight loss remains unchanged. In this sense, interdependence-aware attribution aligns clinical outcomes with the objectives of social insurance and public health systems: minimizing long-term disability, preserving functional capacity, and containing costs.

By explicitly quantifying the contribution of contextual and social determinants alongside pharmacological exposure, the framework supports a more integrated evaluation of real-world interventions that connects individual outcomes to population-level economic and social impact without conflating treatment efficacy with system performance. A central contribution of the present analysis is to move the role of social and system-level factors from implicit background assumptions to explicit, quantifiable contributors.

Across plausible patient profiles, contextual determinants—such as persistence, care intensity, and structured support—account for a substantial share of observed outcomes, in some cases approaching or exceeding one-third of total weight loss. This magnitude demonstrates that health system organization, follow-up strategies, and patient support mechanisms exert a major influence on real-world effectiveness.

Importantly, these factors largely fall within the remit of health authorities, payers, and care providers rather than the laboratory alone. The attribution results therefore provide a quantitative basis for understanding how post-approval outcomes emerge from the interaction between authorized therapies and the systems that deploy them.

In this sense, real-world effectiveness becomes a shared product of pharmacology and public health implementation, rather than a property of the molecule in isolation. Several limitations warrant consideration.

The analysis is simulation-based and relies on parameter ranges calibrated from published RWE rather than primary patient-level data. While the magnitudes are grounded in empirical literature, the results should be interpreted as illustrative rather than definitive.

Future work should validate and refine the framework using large linked datasets, such as NHANES or MIMIC-IV, and explore probabilistic extensions to more fully propagate uncertainty. In summary, interdependence-aware attribution provides a structured and transparent approach to interpreting heterogeneous RWE outcomes for semaglutide.

By enabling a principled transition from population-level averages to individual-level explanation, interdependence-aware attribution provides a missing methodological bridge between real-world data and real-world decisions.

By quantifying how context alters the apparent contribution of drug exposure at the individual level, this framework enhances the practical utility of RWE for clinical development, regulatory evaluation, and evidence-based decision-making, while opening the door to outcome interpretation aligned with physical, psychological, and social dimensions of health.

By enabling a principled transition from population-level averages to individual-level explanation—and back again to population simulation—interdependence-aware attribution provides a missing methodological bridge between real-world data, medico-legal reasoning, and health system decision-making. In doing so, it aligns real-world evidence interpretation with a more comprehensive conception of health that integrates pharmacological, behavioral, and social dimensions.

The analysis demonstrates that improving real-world outcomes does not depend solely on molecular innovation. Quantitatively, a non-trivial fraction of efficacy is already determined by how health systems organize care, support persistence, and address social and behavioral barriers.

Ignoring these contributions risks misattributing both success and failure to pharmacology alone.

References

1. Internal FAIR calculations, National Insurance Institute, 2026.​
2. Patorno E, Gagne JJ, Pawlowski M, et al. Real-world effectiveness of GLP-1 receptor agonists. Ann Intern Med. 2022;175(6):831-839.​
3. Whitley JA, Stephens M, Lu L, et al. Real-world weight change among adults initiating semaglutide for weight loss. Diabetes Obes Metab. 2023;25(8):2345-2353.​​
4. McGovern AP, Hinton W, Calderara S, et al. Real-world effectiveness of GLP-1 receptor agonists. Diabetes Obes Metab. 2018;20(5):1296-1302.​
5. Overbeek JA, Ament SM, van der Heijden AA, et al. Effectiveness of GLP-1 receptor agonists in routine clinical practice. Diabetes Care. 2021;44(3):567-574.​
6. Kristensen SL, Rørth R, Jhund PS, et al. Cardiovascular outcomes and weight change with GLP-1 RAs in real-world use. Lancet Diabetes Endocrinol. 2019;7(12):947-958.​
7. Carroll R, McDonald TJ, Ryder S, et al. Real-world evidence for GLP-1 receptor agonists: opportunities and limitations. Diabetes Ther. 2023;14(2):289-305.​
8. Saunders K, Li H, Ruan X, et al. Heterogeneity of response to GLP-1 receptor agonists in real-world practice. Curr Diabetes Rep. 2022;22(7):385-394.​
9. European Medicines Agency. Ozempic EPAR summary. London: EMA; 2018.​
10. Wilding JPH, Batterham RL, Calanna S, et al. SCOPE: Semaglutide real-world cohort (12 months). Postgrad Med. 2025;137(4):412-420.​