Characterizing Geographic Variation in GLP-1 Receptor Agonist Prescribing Using Interpretable Machine Learning

Characterizing Geographic Variation in GLP-1 Receptor Agonist Prescribing Using Interpretable Machine Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Short Report Characterizing Geographic Variation in GLP-1 Receptor Agonist Prescribing Using Interpretable Machine Learning Zhouzhou Chu, Sherry Yun Wang, Ang Li, Richard Pitts, Ryan Stofer, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8803134/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Geographic disparities in glucagon-like peptide-1 receptor agonist (GLP-1 RA) prescribing remain poorly characterized, with prior studies yielding inconsistent findings regarding rural-urban patterns. The objective of this article is to characterize geographic variation in GLP-1 RA prescribing across the United States and identify neighborhood-level predictors using interpretable machine learning models. We conducted a retrospective analysis of IQVIA PharMetrics® Plus claims data (2010–2022) among commercially insured adults with type 2 diabetes or obesity. Prescriptions were aggregated to 667 three-digit ZIP code (ZIP-3) areas. We compared ridge-penalized linear regression, generalized additive models (GAMs), random forest, and gradient boosting to identify predictors of log-transformed prescribing rates per 100,000 population using 5-fold cross-validation. Among 44,150 individuals contributing 64,281 prescriptions, prescribing rates varied substantially across ZIP-3 areas (median: 2,487 per 100,000). GAMs achieved strong predictive performance (cross-validated R² = 0.335) while maintaining interpretability. Health uninsured rate emerged as the most influential predictor, showing a sharp inverse association with prescribing that attenuated beyond 8% uninsured. Education attainment and urban core fraction were also strongly predictive, with prescribing rates lower in more urbanized areas. Feature importance rankings showed moderate concordance across methods (Spearman ρ = 0.51–0.57). GLP-1 RA prescribing varies significantly across US regions, with patterns shaped by insurance coverage, educational environment, and urbanicity rather than rural-urban status alone. Interpretable machine learning methods provide actional insights for formulary planning and access equity monitoring. Medical Informatics Figures Figure 1 Figure 2 Introduction The rapid growth of glucagon-like peptide-1 receptor agonist (GLP-1 RA) prescribing has raised urgent questions about geographic variation and its alignment with population health needs 1 , 2 . Although prior studies have explored rural-urban differences in the use and distribution of newer diabetes medications, findings remain inconsistent. An analysis using the Medical Expenditure Panel Survey (MEPS) 3 found no significant difference in the use of newer diabetes medications between metropolitan and non-metropolitan areas; however, individuals residing in non‑metropolitan areas were more likely to incur diabetes‑related medication expenditures and tended to spend more on non‑insulin diabetes medications, although these differences did not reach statistical significance. Moreover, Analysis of Veterans Health Administration electronic health record data 4 showed that rural patients with obstructive sleep apnea (OSA) and excess weight had approximately 10% lower adjusted odds of receiving GLP-1 RAs compared to urban peers. Another study from the Wisconsin Health Information Organization 5 reported that patients with type 2 diabetes living in higher-income, majority-White areas had significantly higher adherence to GLP-1 RA medications, while those in rural, lower-income, and majority-Black areas had lower adherence. A national county-level study of Medicare beneficiaries demonstrated that lower prescribing of GLP‑1RAs was associated with rural location, after adjusting for demographics and clinical risk factor 6 . A national Medicare study showcased substantial geographic variation in initiation of newer glucose‑lowering drugs (including GLP‑1RAs) among US older adults, with rural residents less likely to initiate these therapies compared with urban residents (after adjustment) 7 . Collectively, these findings suggest persistent geographic and sociodemographic disparities in GLP-1 RA access and use. However, the patterns remain unclear and may be shaped by varying data sources, populations, and definitions of rurality. To address this gap, we conducted a retrospective study using the IQVIA PharMetrics® Plus for Academics Health Plan claims database from 2010 to 2022. Our objective was to characterize geographic variation in GLP-1 RA prescription volume across the United States and identify neighborhood-level predictors using interpretable machine learning methods. Methods We utilized de-identified claims from IQVIA PharMetrics® Plus for Academics Health Plan claims database, spanning 2010–2022. The database captured longitudinal medical and pharmacy claims for commercially insured individuals across the United States and has been widely used for pharmacoepidemiologic research. Neighborhood-level sociodemographic characteristics were derived from linked external data sources, including the US Zip Codes Database (Pareto Software™, version 2023) 8 and Rural-Urban Commuting Area (RUCA) codes{Onega, 2020 #328}. Study Population and Outcome Adults with type 2 diabetes or obesity were identified using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), and Tenth Revision (ICD-10) Diagnosis Codes (as shown in Supplementary Table 4) and required to have at least one GLP-1 RA prescription (NDC codes shown in Supplementary Table 3) during the study period. A total of 44,150 eligible individuals contributed 64,281 prescriptions aggregated to 667 ZIP-3 areas. The primary outcome was log-transformed prescribing rate per 100,000 population, which was approximately normally distributed after transformation and used as the continuous outcome for all machine learning models. Initial covariates included 68 ZIP-3–level measures spanning demographics, socioeconomic conditions, insurance coverage, housing characteristics, and transportation (Supplementary Table 1). Urban-rural classification used RUCA codes collapsed into binary urban core versus non-urban. Underlying disease burden was measured using population-level type 2 diabetes and obesity patient counts. Statistical Analysis We employed a comparative machine learning framework to identify neighborhood-level predictors of GLP-1 RA prescribing variation, comparing four modeling approaches of increasing complexity: (1) ridge-penalized linear regression as a linear baseline; (2) generalized additive models (GAMs) as the primary approach, allowing nonlinear effects; and (3–4) random forest and gradient boosting as tree-based predictive benchmarks. All models used the same log-transformed prescribing rate outcome and were evaluated using 5-fold cross-validated R² (CV R²) to ensure comparable, out-of-sample performance assessment. GAMs extend generalized linear models by replacing linear coefficients with smooth functions: g(E[Y]) = \(\:{\beta\:}_{0}+\sum\:{f}_{j}{(X}_{j})\) , where each \(\:{f}_{j}\left(\right)\) is a smooth function estimated nonparametrically. We fitted GAMs to the log-transformed prescribing rate using cubic B-splines with 5 knots per continuous predictor and ridge regularization (α = 10, selected via 5-fold cross-validation). The 20 predictors included in the GAMs were selected based on domain knowledge and prior importance analyses. UrbanCore was entered as a linear (binary) term. This approach was selected because it: (1) captures nonlinear predictor–outcome relationships while maintaining interpretability through partial effect plots; (2) handles multicollinearity through regularization; and (3) provides effect estimates that can be visualized and communicated to clinical and policy audiences. Feature importance in the GAMs was quantified as the effect range, with the difference between the maximum and minimum predicted values across each smooth function reflecting the total magnitude of each predictor’s contribution on the log-rate scale. Tree-based ensemble methods (benchmarks) like random forest and gradient boosting models were fitted using all 68 available predictors. Random forest used 500 trees with hyperparameters optimized via randomized search (50 iterations). Gradient boosting used 200 trees with learning rate 0.05, maximum depth 5, minimum samples per leaf 10, and subsample fraction 0.8. Feature importance was assessed using permutation importance, which measures the decrease in CV R² when each feature is randomly shuffled, providing a model-agnostic importance metric comparable across methods. To assess the robustness of identified predictors, feature importance rankings were compared across all three methods using Spearman rank correlation coefficients. Partial dependence plots were generated for the top 6 predictors across all methods to compare the direction and functional form of associations (Fig. 3). Nonlinearity was assessed by comparing GAM smooth functions against superimposed linear fits, with deviations between the two indicating meaningful nonlinear effects (Fig. 5). Software Analyses were conducted in Python 3.12 using scikit-learn (v1.4) for ridge regression, random forest, and gradient boosting; scipy (v1.12) for statistical tests; and custom B-spline implementations for GAM smooth terms. Geographic patterns were also explored via an interactive Tableau dashboard at http://bit.ly/471TUa7 . Results We identified 64,281 GLP-1 RA prescriptions dispensed to 44,150 adults with diagnoses of type 2 diabetes or obesity and aggregated prescription counts to 667 ZIP-3-level geographic units for area-level analysis. Distribution of GLP-1 RA Prescribing Rates Across ZIP-3 Areas (Fig. 1 ) indicated that raw GLP-1 RA prescribing rates per 100,000 population were highly right-skewed, with substantial heterogeneity across regions. The median prescribing rate was 2,487 prescriptions per 100,000 population, while a small number of ZIP-3 areas exhibited markedly higher rates, exceeding 50,000 prescriptions per 100,000 population. Visual inspection of prescribing rates (Fig. 1 ) identified clusters of high GLP-1 RA use in central and southern states, including Missouri, West Virginia, Oklahoma, and Louisiana. Predictive performance varied systematically by model complexity. Ridge-penalized linear regression explained a modest proportion of out-of-sample variation (mean 5-fold CV R² = 0.248). Incorporation of nonlinear effects through GAMs substantially improved predictive performance (CV R² = 0.335), representing an absolute gain of 0.087 relative to the linear baseline. Tree-based ensemble models achieved higher predictive accuracy, with random forest yielding a CV R² of 0.383 and gradient boosting achieving the highest performance (CV R² = 0.408). However, given the relatively small incremental gain over GAMs and the reduced interpretability of ensemble methods, GAMs were selected as the primary analytic approach to balance predictive accuracy with interpretability. Feature importance rankings (Fig. 2 ) were broadly consistent across GAM, random forest, and gradient boosting models (Fig. 4). Across all three methods, the health uninsured rate emerged as the most influential predictor of GLP-1 RA prescribing variation, followed by educational attainment (percentage of residents with a bachelor’s degree or higher). Urban core fraction consistently ranked among the top predictors in GAM and gradient boosting models, while median home value and labor force participation also showed strong and recurrent importance. Spearman rank correlations of feature importance rankings demonstrated moderate concordance across methods (ρ = 0.51–0.57, all p < 0.05), supporting the robustness of identified predictors despite methodological differences. Partial dependence plots (Supplementary Fig. 1) revealed substantial agreement in the direction of associations across GAM, random forest, and gradient boosting models for the top six predictors, while highlighting meaningful nonlinearities not captured by linear models. Higher uninsured rates were associated with sharply lower GLP-1 RA prescribing at low to moderate levels, followed by attenuation at higher uninsured proportions. Educational attainment exhibited a nonlinear inverse association, with prescribing rates declining more steeply beyond approximately 20% bachelor’s degree prevalence. Urban core fraction demonstrated a monotonic negative association, with the steepest declines occurring as ZIP-3 areas transitioned from mixed to predominantly urban. Median home value displayed a U-shaped association, with lower prescribing in mid-range home values and higher prescribing at both lower and higher extremes. Labor force participation and Native American population proportion also showed nonlinear patterns, with threshold effects evident across models. Discussion In this national analysis of commercially insured adults, we observed substantial geographic variation in GLP-1 RA prescribing that did not conform to commonly assumed rural-urban gradients. Across ZIP-3 areas, prescribing rates were modestly lower in more urbanized areas, with rural and mixed areas exhibiting higher prescription volume per eligible population. This pattern persisted after accounting for a broad set of neighborhood sociodemographic characteristics and was directionally consistent across modeling approaches. These findings extend prior literature reporting mixed or attenuated rural-urban differences in GLP-1 RA use. While several studies{Spinelli, 2025 #329}{Gavin, 2025 #330} using Medicare or Veterans Health Administration data have reported lower adjusted odds of GLP-1 RA initiation among rural residents, others have found minimal differences or inconsistent patterns across outcomes and populations. Our results suggest that, within commercially insured populations and at finer geographic resolution, prescribing intensity may be shaped by a combination of disease burden, insurance design, and local practice environments rather than rurality alone. The mechanisms underlying higher prescribing rates in rural areas remain uncertain and are likely multifactorial. Rural communities may experience higher prevalence of obesity and type 2 diabetes and may rely more heavily on pharmacologic management in settings with fewer lifestyle or specialty care alternatives. In contrast, urban areas, despite greater healthcare density, may face more restrictive payer utilization management, formulary constraints, or greater prescriber caution related to cost, supply constraints, or off-label use concerns. Our use of effect-range-based feature importance, cross-method robustness assessment, and explicit nonlinearity diagnostics moves beyond black-box prediction to support transparency, auditability, and replication, key governance considerations for machine learning in managed care. For instance, the health uninsured rate shows a sharp inverse relationship with prescribing volume up to approximately 8%, beyond which the association flattens, illustrating that even modest improvements in insurance coverage may yield disproportionate gains in GLP-1 RA access. Similarly, the percentage of residents with a bachelor’s degree or higher is inversely associated with prescribing, with an inflection point around 17–20%, highlighting the potential role of educational environment as a structural determinant of treatment uptake. Interestingly, urban core fraction also exhibits a monotonic negative association, but with a clear nonlinear steepening around the mid-range of urbanicity, indicating that mixed areas may experience unique access dynamics. The U-shaped association for median home value, alongside the nonlinear effects of labor force participation, further underscores the need to consider economic precarity and workforce engagement as multifaceted contributors to therapeutic equity. The Native American population proportion, while exhibiting smaller effect magnitudes, consistently shows lower prescribing across models, raising concerns about entrenched disparities not explained by insurance or socioeconomic status alone. Higher proportions of residents identifying as Black or Hispanic were consistently associated with lower GLP-1 RA prescription volume, independent of urbanicity and socioeconomic indicators. In contrast, area-level poverty was not independently associated with prescribing volume after multivariable adjustment. Together, these findings underscore that sociodemographic gradients in GLP-1 RA use do not map cleanly onto rural-urban classifications and may instead reflect complex interactions among race and ethnicity, insurance coverage, and local prescribing environments. Importantly, these area-level patterns are consistent with prior reports documenting lower uptake or adherence to GLP-1 RAs in predominantly minority communities, even when insurance coverage is present. As GLP-1 RAs expand beyond glycemic control into broader cardiometabolic and preventive indications, understanding the geographic and sociodemographic patterning of their use is increasingly important. Our findings highlight that efforts to promote equitable access should move beyond simple rural-urban dichotomies and instead account for neighborhood-level context, insurance design, and structural factors that shape prescribing behavior across regions. Limitation This study has several important limitations. All analyses were conducted at the area level, and associations should not be interpreted as reflecting individual-level access, prescribing decisions, or treatment receipt. Area-level associations are subject to ecological fallacy, and within-ZIP-3 heterogeneity may be substantial. In addition, the IQVIA PharMetrics® Plus database captures commercially insured populations and does not represent individuals insured through Medicaid alone, the uninsured, or those fully enrolled in Medicare. Accordingly, our findings describe geographic variation in GLP-1 RA prescription volume within the commercially insured population and should not be interpreted as population-wide measures of access or equity. Conclusion These findings highlight substantial geographic heterogeneity in GLP-1 RA prescribing that cannot be fully explained by rural-urban status alone. Instead, prescribing patterns reflect complex, nonlinear relationships with neighborhood-level sociodemographics, insurance coverage, and structural factors such as education, housing, and workforce participation. Importantly, interpretable ML methods like generalized additive models provide a transparent, reproducible framework to uncover these nuanced patterns, supporting more equitable and data-driven decision-making in managed care. As GLP-1 RAs expand in scope from glycemic control to broader cardiometabolic prevention, this analytic approach offers a valuable tool for formulary planning, access equity monitoring, and benefit design targeted to population needs. Declarations Conflicts of Interest None Guarantor and Funding statements Dr. Sherry Yun Wang served as the guarantor and had full access to all study data, taking responsibility for the integrity of the data and the accuracy of the analyses. No external funding was received for this work. Author Contributions Sherry Yun Wang conceived the study concept and drafted the manuscript. Zhouzhou Chu and Ryan Stofer performed the analyses and ran all code. Ang Li provided high-level methodological guidance. Anna Alber provided technical support. Tannaz Moin and Richard Pitts reviewed and critically edited the manuscript. All authors reviewed and approved the final version of the manuscript. Acknowledgement Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the IQVIA Inc. References Spinelli KJ, Oakes AH (2025) Glucagon-Like Peptide 1 Receptor Agonists and the Deepening Health Equity Divide in America. Am J Health Promotion 39(5):832–836 Hackler-Moorman CJ (2025) Rural and Urban Disparities in Access to GLP-1 Agonists for Patients with Obesity Zhu B, Ding D, Luo J, Glied S (2025) Rural-urban disparities in the uptake of new diabetes medications. Diabetes Spectr 38(1):49–57 Leonhard A, Josey K, Plumley R et al (2025) 1310 Receipt of Glucagon-Like Peptide-1 Receptor Agonists Among Rural Patients with Obstructive Sleep Apnea and Excess Weight. Sleep 48(Supplement1):A564–A565 Gavin KL, Amjad R, Makris V, Neuner JM (2025) Differences in GLP-1 RA medication adherence across place-based variables in patients with diabetes living in Wisconsin. J Managed Care Specialty Pharm 31(12):1311–1319 Hanna J, Nargesi AA, Essien UR et al (2022) County-level variation in cardioprotective antihyperglycemic prescribing among Medicare beneficiaries. Am J Prev Cardiol 11:100370 Chen W-H, Li Y, Yang L et al (2024) Geographic variation and racial disparities in adoption of newer glucose-lowering drugs with cardiovascular benefits among US Medicare beneficiaries with type 2 diabetes. PLoS ONE 19(1):e0297208 Software P (2025) US Zip Codes Database. ; https://www.unitedstateszipcodes.org/ Supplementary Fig 1 Partial Dependence Plots of Top Neighborhood-Level Predictors of GLP-1 RA Prescribing Across Three Modeling Approaches Additional Declarations The authors declare no competing interests. Supplementary Files SupplementaryFiguresTables.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8803134","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":586743240,"identity":"3d988bbd-73c0-4504-abcb-2c53e1e464f2","order_by":0,"name":"Zhouzhou Chu","email":"","orcid":"","institution":"School of Pharmacy, Chapman University","correspondingAuthor":false,"prefix":"","firstName":"Zhouzhou","middleName":"","lastName":"Chu","suffix":""},{"id":586749591,"identity":"e03ef6ad-40f9-4de6-afb6-6bca4264472b","order_by":1,"name":"Sherry Yun Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA10lEQVRIiWNgGAWjYBADOQYJODuBOC3GpGtJbCBai8GN3GMSP3fUpvfPbn726EbNHQZ+9hwDAlry0iR7zxzPnXHnmLlxzrFnDJI9bwhpyTG7wdt2LHeDRIKZdA7bYZAIYS03/7YdSzeQSP8mnfPvMIM9MVpu87bVJBhI5JhJ57YBbZEgoEXyzBvz37JtBwxn3MgpN87tO8wjceZZAV4tfMdzjA3fttXJ889I3/Y459thOf725A14tSgcAFOHQQQbiODBqxwE5BvAVB1cyygYBaNgFIwCDAAAOvBMGhoGzgAAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0001-9438-5586","institution":"School of Pharmacy, Chapman University","correspondingAuthor":true,"prefix":"","firstName":"Sherry","middleName":"Yun","lastName":"Wang","suffix":""},{"id":586749592,"identity":"ac79fb87-43b3-43ec-849d-07a0767c7f8e","order_by":2,"name":"Ang Li","email":"","orcid":"","institution":"Department of Electrical and Computer Engineering, University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Ang","middleName":"","lastName":"Li","suffix":""},{"id":586749593,"identity":"19f3403b-3e86-4326-b9a5-46c085712cb6","order_by":3,"name":"Richard Pitts","email":"","orcid":"","institution":"CalOptima Health","correspondingAuthor":false,"prefix":"","firstName":"Richard","middleName":"","lastName":"Pitts","suffix":""},{"id":586749594,"identity":"7355b29c-064d-431e-8d8d-db8c5cf4c0cf","order_by":4,"name":"Ryan Stofer","email":"","orcid":"","institution":"School of Pharmacy, Chapman University","correspondingAuthor":false,"prefix":"","firstName":"Ryan","middleName":"","lastName":"Stofer","suffix":""},{"id":586749595,"identity":"7fea2cff-e465-42d0-bd49-da60d73792f2","order_by":5,"name":"Anna Alber","email":"","orcid":"","institution":"IS\u0026T Research Computing, Chapman University","correspondingAuthor":false,"prefix":"","firstName":"Anna","middleName":"","lastName":"Alber","suffix":""},{"id":586749596,"identity":"3a4b9a72-f923-4d15-8962-f3b5f61f7f09","order_by":6,"name":"Zhen Cong","email":"","orcid":"","institution":"Department of Health Sciences, Chapman University","correspondingAuthor":false,"prefix":"","firstName":"Zhen","middleName":"","lastName":"Cong","suffix":""},{"id":586749597,"identity":"5469b4db-ea6d-4f88-bc81-09ce565ef182","order_by":7,"name":"Tannaz Moin","email":"","orcid":"","institution":"Department of Medicine, University of California Los Angeles, David Geffen School of Medicine, Los Angeles, CA, USA; VA Greater Los Angeles Healthcare System, Los Angeles, CA, USA","correspondingAuthor":false,"prefix":"","firstName":"Tannaz","middleName":"","lastName":"Moin","suffix":""}],"badges":[],"createdAt":"2026-02-06 06:04:50","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8803134/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8803134/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102174708,"identity":"c0141695-33b7-47d5-890a-32c2f9c0da20","added_by":"auto","created_at":"2026-02-09 05:48:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":561190,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNational Patterns in GLP-1 Receptor Agonist Prescribing Among Adults With Type 2 Diabetes or Obesity in the United States, 2010–2022\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8803134/v1/e4108e0c1b9ed1f156f9a137.png"},{"id":102174709,"identity":"b54b5eb6-d001-4730-b530-49e465a3ef73","added_by":"auto","created_at":"2026-02-09 05:48:28","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1028999,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCross method feature importance comparison\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8803134/v1/303dd95841f3b598f629433d.png"},{"id":102297520,"identity":"d96c9b02-ff59-46af-a15e-14a9ea9a3d38","added_by":"auto","created_at":"2026-02-10 10:27:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2056671,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8803134/v1/4473ff78-93d5-4abf-97b6-51437d27fa8d.pdf"},{"id":102174707,"identity":"149b4bef-b995-493d-addc-fe70745092a8","added_by":"auto","created_at":"2026-02-09 05:48:28","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":959932,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFiguresTables.docx","url":"https://assets-eu.researchsquare.com/files/rs-8803134/v1/905ab094ea8ea82dc71366fb.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eCharacterizing Geographic Variation in GLP-1 Receptor Agonist Prescribing Using Interpretable Machine Learning\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe rapid growth of glucagon-like peptide-1 receptor agonist (GLP-1 RA) prescribing has raised urgent questions about geographic variation and its alignment with population health needs\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Although prior studies have explored rural-urban differences in the use and distribution of newer diabetes medications, findings remain inconsistent. An analysis using the Medical Expenditure Panel Survey (MEPS)\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e found no significant difference in the use of newer diabetes medications between metropolitan and non-metropolitan areas; however, individuals residing in non‑metropolitan areas were more likely to incur diabetes‑related medication expenditures and tended to spend more on non‑insulin diabetes medications, although these differences did not reach statistical significance. Moreover, Analysis of Veterans Health Administration electronic health record data\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e showed that rural patients with obstructive sleep apnea (OSA) and excess weight had approximately 10% lower adjusted odds of receiving GLP-1 RAs compared to urban peers. Another study from the Wisconsin Health Information Organization\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e reported that patients with type 2 diabetes living in higher-income, majority-White areas had significantly higher adherence to GLP-1 RA medications, while those in rural, lower-income, and majority-Black areas had lower adherence. A national county-level study of Medicare beneficiaries demonstrated that lower prescribing of GLP‑1RAs was associated with rural location, after adjusting for demographics and clinical risk factor\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. A national Medicare study showcased substantial geographic variation in initiation of newer glucose‑lowering drugs (including GLP‑1RAs) among US older adults, with rural residents less likely to initiate these therapies compared with urban residents (after adjustment)\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Collectively, these findings suggest persistent geographic and sociodemographic disparities in GLP-1 RA access and use. However, the patterns remain unclear and may be shaped by varying data sources, populations, and definitions of rurality. To address this gap, we conducted a retrospective study using the IQVIA PharMetrics\u0026reg; Plus for Academics Health Plan claims database from 2010 to 2022. Our objective was to characterize geographic variation in GLP-1 RA prescription volume across the United States and identify neighborhood-level predictors using interpretable machine learning methods.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eWe utilized de-identified claims from IQVIA PharMetrics\u0026reg; Plus for Academics Health Plan claims database, spanning 2010\u0026ndash;2022. The database captured longitudinal medical and pharmacy claims for commercially insured individuals across the United States and has been widely used for pharmacoepidemiologic research. Neighborhood-level sociodemographic characteristics were derived from linked external data sources, including the US Zip Codes Database (Pareto Software\u0026trade;, version 2023)\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e and Rural-Urban Commuting Area (RUCA) codes{Onega, 2020 #328}. Study Population and Outcome Adults with type 2 diabetes or obesity were identified using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), and Tenth Revision (ICD-10) Diagnosis Codes (as shown in Supplementary Table\u0026nbsp;4) and required to have at least one GLP-1 RA prescription (NDC codes shown in Supplementary Table\u0026nbsp;3) during the study period. A total of 44,150 eligible individuals contributed 64,281 prescriptions aggregated to 667 ZIP-3 areas. The primary outcome was log-transformed prescribing rate per 100,000 population, which was approximately normally distributed after transformation and used as the continuous outcome for all machine learning models. Initial covariates included 68 ZIP-3\u0026ndash;level measures spanning demographics, socioeconomic conditions, insurance coverage, housing characteristics, and transportation (Supplementary Table\u0026nbsp;1). Urban-rural classification used RUCA codes collapsed into binary urban core versus non-urban. Underlying disease burden was measured using population-level type 2 diabetes and obesity patient counts.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStatistical Analysis\u003c/b\u003e We employed a comparative machine learning framework to identify neighborhood-level predictors of GLP-1 RA prescribing variation, comparing four modeling approaches of increasing complexity: (1) ridge-penalized linear regression as a linear baseline; (2) generalized additive models (GAMs) as the primary approach, allowing nonlinear effects; and (3\u0026ndash;4) random forest and gradient boosting as tree-based predictive benchmarks. All models used the same log-transformed prescribing rate outcome and were evaluated using 5-fold cross-validated R\u0026sup2; (CV R\u0026sup2;) to ensure comparable, out-of-sample performance assessment. GAMs extend generalized linear models by replacing linear coefficients with smooth functions: g(E[Y]) = \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\beta\\:}_{0}+\\sum\\:{f}_{j}{(X}_{j})\\)\u003c/span\u003e\u003c/span\u003e, where each \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{f}_{j}\\left(\\right)\\)\u003c/span\u003e\u003c/span\u003e is a smooth function estimated nonparametrically. We fitted GAMs to the log-transformed prescribing rate using cubic B-splines with 5 knots per continuous predictor and ridge regularization (α\u0026thinsp;=\u0026thinsp;10, selected via 5-fold cross-validation). The 20 predictors included in the GAMs were selected based on domain knowledge and prior importance analyses. UrbanCore was entered as a linear (binary) term. This approach was selected because it: (1) captures nonlinear predictor\u0026ndash;outcome relationships while maintaining interpretability through partial effect plots; (2) handles multicollinearity through regularization; and (3) provides effect estimates that can be visualized and communicated to clinical and policy audiences. Feature importance in the GAMs was quantified as the effect range, with the difference between the maximum and minimum predicted values across each smooth function reflecting the total magnitude of each predictor\u0026rsquo;s contribution on the log-rate scale. Tree-based ensemble methods (benchmarks) like random forest and gradient boosting models were fitted using all 68 available predictors. Random forest used 500 trees with hyperparameters optimized via randomized search (50 iterations). Gradient boosting used 200 trees with learning rate 0.05, maximum depth 5, minimum samples per leaf 10, and subsample fraction 0.8. Feature importance was assessed using permutation importance, which measures the decrease in CV R\u0026sup2; when each feature is randomly shuffled, providing a model-agnostic importance metric comparable across methods. To assess the robustness of identified predictors, feature importance rankings were compared across all three methods using Spearman rank correlation coefficients. Partial dependence plots were generated for the top 6 predictors across all methods to compare the direction and functional form of associations (Fig.\u0026nbsp;3). Nonlinearity was assessed by comparing GAM smooth functions against superimposed linear fits, with deviations between the two indicating meaningful nonlinear effects (Fig.\u0026nbsp;5). Software Analyses were conducted in Python 3.12 using scikit-learn (v1.4) for ridge regression, random forest, and gradient boosting; scipy (v1.12) for statistical tests; and custom B-spline implementations for GAM smooth terms. Geographic patterns were also explored via an interactive Tableau dashboard at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://bit.ly/471TUa7\u003c/span\u003e\u003cspan address=\"http://bit.ly/471TUa7\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe identified 64,281 GLP-1 RA prescriptions dispensed to 44,150 adults with diagnoses of type 2 diabetes or obesity and aggregated prescription counts to 667 ZIP-3-level geographic units for area-level analysis. Distribution of GLP-1 RA Prescribing Rates Across ZIP-3 Areas (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) indicated that raw GLP-1 RA prescribing rates per 100,000 population were highly right-skewed, with substantial heterogeneity across regions. The median prescribing rate was 2,487 prescriptions per 100,000 population, while a small number of ZIP-3 areas exhibited markedly higher rates, exceeding 50,000 prescriptions per 100,000 population. Visual inspection of prescribing rates (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) identified clusters of high GLP-1 RA use in central and southern states, including Missouri, West Virginia, Oklahoma, and Louisiana.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePredictive performance varied systematically by model complexity. Ridge-penalized linear regression explained a modest proportion of out-of-sample variation (mean 5-fold CV R\u0026sup2; = 0.248). Incorporation of nonlinear effects through GAMs substantially improved predictive performance (CV R\u0026sup2; = 0.335), representing an absolute gain of 0.087 relative to the linear baseline. Tree-based ensemble models achieved higher predictive accuracy, with random forest yielding a CV R\u0026sup2; of 0.383 and gradient boosting achieving the highest performance (CV R\u0026sup2; = 0.408). However, given the relatively small incremental gain over GAMs and the reduced interpretability of ensemble methods, GAMs were selected as the primary analytic approach to balance predictive accuracy with interpretability.\u003c/p\u003e \u003cp\u003eFeature importance rankings (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) were broadly consistent across GAM, random forest, and gradient boosting models (Fig.\u0026nbsp;4). Across all three methods, the health uninsured rate emerged as the most influential predictor of GLP-1 RA prescribing variation, followed by educational attainment (percentage of residents with a bachelor\u0026rsquo;s degree or higher). Urban core fraction consistently ranked among the top predictors in GAM and gradient boosting models, while median home value and labor force participation also showed strong and recurrent importance. Spearman rank correlations of feature importance rankings demonstrated moderate concordance across methods (ρ\u0026thinsp;=\u0026thinsp;0.51\u0026ndash;0.57, all p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), supporting the robustness of identified predictors despite methodological differences.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePartial dependence plots (Supplementary Fig.\u0026nbsp;1) revealed substantial agreement in the direction of associations across GAM, random forest, and gradient boosting models for the top six predictors, while highlighting meaningful nonlinearities not captured by linear models. Higher uninsured rates were associated with sharply lower GLP-1 RA prescribing at low to moderate levels, followed by attenuation at higher uninsured proportions. Educational attainment exhibited a nonlinear inverse association, with prescribing rates declining more steeply beyond approximately 20% bachelor\u0026rsquo;s degree prevalence. Urban core fraction demonstrated a monotonic negative association, with the steepest declines occurring as ZIP-3 areas transitioned from mixed to predominantly urban. Median home value displayed a U-shaped association, with lower prescribing in mid-range home values and higher prescribing at both lower and higher extremes. Labor force participation and Native American population proportion also showed nonlinear patterns, with threshold effects evident across models.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this national analysis of commercially insured adults, we observed substantial geographic variation in GLP-1 RA prescribing that did not conform to commonly assumed rural-urban gradients. Across ZIP-3 areas, prescribing rates were modestly lower in more urbanized areas, with rural and mixed areas exhibiting higher prescription volume per eligible population. This pattern persisted after accounting for a broad set of neighborhood sociodemographic characteristics and was directionally consistent across modeling approaches. These findings extend prior literature reporting mixed or attenuated rural-urban differences in GLP-1 RA use. While several studies{Spinelli, 2025 #329}{Gavin, 2025 #330} using Medicare or Veterans Health Administration data have reported lower adjusted odds of GLP-1 RA initiation among rural residents, others have found minimal differences or inconsistent patterns across outcomes and populations. Our results suggest that, within commercially insured populations and at finer geographic resolution, prescribing intensity may be shaped by a combination of disease burden, insurance design, and local practice environments rather than rurality alone. The mechanisms underlying higher prescribing rates in rural areas remain uncertain and are likely multifactorial. Rural communities may experience higher prevalence of obesity and type 2 diabetes and may rely more heavily on pharmacologic management in settings with fewer lifestyle or specialty care alternatives. In contrast, urban areas, despite greater healthcare density, may face more restrictive payer utilization management, formulary constraints, or greater prescriber caution related to cost, supply constraints, or off-label use concerns.\u003c/p\u003e \u003cp\u003eOur use of effect-range-based feature importance, cross-method robustness assessment, and explicit nonlinearity diagnostics moves beyond black-box prediction to support transparency, auditability, and replication, key governance considerations for machine learning in managed care. For instance, the health uninsured rate shows a sharp inverse relationship with prescribing volume up to approximately 8%, beyond which the association flattens, illustrating that even modest improvements in insurance coverage may yield disproportionate gains in GLP-1 RA access. Similarly, the percentage of residents with a bachelor\u0026rsquo;s degree or higher is inversely associated with prescribing, with an inflection point around 17\u0026ndash;20%, highlighting the potential role of educational environment as a structural determinant of treatment uptake. Interestingly, urban core fraction also exhibits a monotonic negative association, but with a clear nonlinear steepening around the mid-range of urbanicity, indicating that mixed areas may experience unique access dynamics. The U-shaped association for median home value, alongside the nonlinear effects of labor force participation, further underscores the need to consider economic precarity and workforce engagement as multifaceted contributors to therapeutic equity. The Native American population proportion, while exhibiting smaller effect magnitudes, consistently shows lower prescribing across models, raising concerns about entrenched disparities not explained by insurance or socioeconomic status alone. Higher proportions of residents identifying as Black or Hispanic were consistently associated with lower GLP-1 RA prescription volume, independent of urbanicity and socioeconomic indicators. In contrast, area-level poverty was not independently associated with prescribing volume after multivariable adjustment. Together, these findings underscore that sociodemographic gradients in GLP-1 RA use do not map cleanly onto rural-urban classifications and may instead reflect complex interactions among race and ethnicity, insurance coverage, and local prescribing environments. Importantly, these area-level patterns are consistent with prior reports documenting lower uptake or adherence to GLP-1 RAs in predominantly minority communities, even when insurance coverage is present. As GLP-1 RAs expand beyond glycemic control into broader cardiometabolic and preventive indications, understanding the geographic and sociodemographic patterning of their use is increasingly important. Our findings highlight that efforts to promote equitable access should move beyond simple rural-urban dichotomies and instead account for neighborhood-level context, insurance design, and structural factors that shape prescribing behavior across regions.\u003c/p\u003e"},{"header":"Limitation","content":"\u003cp\u003eThis study has several important limitations. All analyses were conducted at the area level, and associations should not be interpreted as reflecting individual-level access, prescribing decisions, or treatment receipt. Area-level associations are subject to ecological fallacy, and within-ZIP-3 heterogeneity may be substantial. In addition, the IQVIA PharMetrics\u0026reg; Plus database captures commercially insured populations and does not represent individuals insured through Medicaid alone, the uninsured, or those fully enrolled in Medicare. Accordingly, our findings describe geographic variation in GLP-1 RA prescription volume within the commercially insured population and should not be interpreted as population-wide measures of access or equity.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThese findings highlight substantial geographic heterogeneity in GLP-1 RA prescribing that cannot be fully explained by rural-urban status alone. Instead, prescribing patterns reflect complex, nonlinear relationships with neighborhood-level sociodemographics, insurance coverage, and structural factors such as education, housing, and workforce participation. Importantly, interpretable ML methods like generalized additive models provide a transparent, reproducible framework to uncover these nuanced patterns, supporting more equitable and data-driven decision-making in managed care. As GLP-1 RAs expand in scope from glycemic control to broader cardiometabolic prevention, this analytic approach offers a valuable tool for formulary planning, access equity monitoring, and benefit design targeted to population needs.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflicts of Interest\u003c/h2\u003e \u003cp\u003eNone\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eGuarantor and Funding statements\u003c/h2\u003e \u003cp\u003eDr. Sherry Yun Wang served as the guarantor and had full access to all study data, taking responsibility for the integrity of the data and the accuracy of the analyses. No external funding was received for this work.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contributions\u003c/h2\u003e \u003cp\u003eSherry Yun Wang conceived the study concept and drafted the manuscript. Zhouzhou Chu and Ryan Stofer performed the analyses and ran all code. Ang Li provided high-level methodological guidance. Anna Alber provided technical support. Tannaz Moin and Richard Pitts reviewed and critically edited the manuscript. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e \u003cp\u003eAny opinions, findings, conclusions, or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the IQVIA Inc.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSpinelli KJ, Oakes AH (2025) Glucagon-Like Peptide 1 Receptor Agonists and the Deepening Health Equity Divide in America. Am J Health Promotion 39(5):832\u0026ndash;836\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHackler-Moorman CJ (2025) Rural and Urban Disparities in Access to GLP-1 Agonists for Patients with Obesity\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu B, Ding D, Luo J, Glied S (2025) Rural-urban disparities in the uptake of new diabetes medications. Diabetes Spectr 38(1):49\u0026ndash;57\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeonhard A, Josey K, Plumley R et al (2025) 1310 Receipt of Glucagon-Like Peptide-1 Receptor Agonists Among Rural Patients with Obstructive Sleep Apnea and Excess Weight. Sleep 48(Supplement1):A564\u0026ndash;A565\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGavin KL, Amjad R, Makris V, Neuner JM (2025) Differences in GLP-1 RA medication adherence across place-based variables in patients with diabetes living in Wisconsin. J Managed Care Specialty Pharm 31(12):1311\u0026ndash;1319\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHanna J, Nargesi AA, Essien UR et al (2022) County-level variation in cardioprotective antihyperglycemic prescribing among Medicare beneficiaries. Am J Prev Cardiol 11:100370\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen W-H, Li Y, Yang L et al (2024) Geographic variation and racial disparities in adoption of newer glucose-lowering drugs with cardiovascular benefits among US Medicare beneficiaries with type 2 diabetes. PLoS ONE 19(1):e0297208\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoftware P (2025) US Zip Codes Database. ; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.unitedstateszipcodes.org/\u003c/span\u003e\u003cspan address=\"https://www.unitedstateszipcodes.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSupplementary Fig 1 Partial Dependence Plots of Top Neighborhood-Level Predictors of GLP-1 RA Prescribing Across Three Modeling Approaches\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Chapman University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8803134/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8803134/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eGeographic disparities in glucagon-like peptide-1 receptor agonist (GLP-1 RA) prescribing remain poorly characterized, with prior studies yielding inconsistent findings regarding rural-urban patterns. The objective of this article is to characterize geographic variation in GLP-1 RA prescribing across the United States and identify neighborhood-level predictors using interpretable machine learning models. We conducted a retrospective analysis of IQVIA PharMetrics\u0026reg; Plus claims data (2010\u0026ndash;2022) among commercially insured adults with type 2 diabetes or obesity. Prescriptions were aggregated to 667 three-digit ZIP code (ZIP-3) areas. We compared ridge-penalized linear regression, generalized additive models (GAMs), random forest, and gradient boosting to identify predictors of log-transformed prescribing rates per 100,000 population using 5-fold cross-validation. Among 44,150 individuals contributing 64,281 prescriptions, prescribing rates varied substantially across ZIP-3 areas (median: 2,487 per 100,000). GAMs achieved strong predictive performance (cross-validated R\u0026sup2; = 0.335) while maintaining interpretability. Health uninsured rate emerged as the most influential predictor, showing a sharp inverse association with prescribing that attenuated beyond 8% uninsured. Education attainment and urban core fraction were also strongly predictive, with prescribing rates lower in more urbanized areas. Feature importance rankings showed moderate concordance across methods (Spearman ρ\u0026thinsp;=\u0026thinsp;0.51\u0026ndash;0.57). GLP-1 RA prescribing varies significantly across US regions, with patterns shaped by insurance coverage, educational environment, and urbanicity rather than rural-urban status alone. Interpretable machine learning methods provide actional insights for formulary planning and access equity monitoring.\u003c/p\u003e","manuscriptTitle":"Characterizing Geographic Variation in GLP-1 Receptor Agonist Prescribing Using Interpretable Machine Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-09 05:48:23","doi":"10.21203/rs.3.rs-8803134/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c0e52615-8477-44eb-a035-95339b0347b2","owner":[],"postedDate":"February 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62429214,"name":"Medical Informatics"}],"tags":[],"updatedAt":"2026-02-09T05:48:23+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-09 05:48:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8803134","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8803134","identity":"rs-8803134","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}