Bridging the rural logistics penalty: quantifying structural costs and designing policy mechanisms for home TCM delivery to older adults with chronic respiratory diseases Authors: Yunanji Zhou, Siyu Xie, Xinyi Qiu, Ting Yuan, Lihua Wang Published: 2026 ABSTRACT BackgroundEquitable access to prescribed therapies remains challenging for older adults with chronic respiratory diseases (CRDs) in rural China. Liquid Traditional Chinese Medicine (TCM) decoctions are widely used in respiratory care, but their large volume, thermal sensitivity, and need for temperature-controlled handling make home delivery particularly difficult in geographically dispersed rural settings. The structural and cost implications of this rural cold-chain delivery burden remain insufficiently quantified.MethodsWe developed a four-stage calibrated simulation framework linking CHARLS 2018 microdata with external facility-density statistics, interpretable machine learning, parameterized micro-costing, and policy scenario modelling. The analytic sample included 11,045 adults aged ≥ 60 years, including 2,351 participants with chronic lung disease or asthma. Because CHARLS does not provide patient-level geocoded coordinates or observed residence-to-facility routes, Stage I generated simulation-derived distance distributions rather than observed travel distances, and no patient-level geocoded empirical validation was possible. Stage II used an IADL-based XGBoost-SHAP model to characterize functional vulnerability more directly relevant to medication access and home-delivery prioritization, with ADL retained as a sensitivity outcome. Stage III estimated cold-chain delivery costs under alternative assumptions for consolidation, route tortuosity, and willingness to pay. Stage IV examined cooperation dynamics across subsidy scenarios.ResultsThe calibrated simulation preserved the facility-density-driven rural-urban contrast by design rather than empirically discovering patient routes. Mean modeled distance was 10.61 km for rural residents versus 2.85 km for urban residents, and 10.8% of rural simulated observations exceeded 15 km. The IADL-based XGBoost model achieved an AUC of 0.668 and an accuracy of 0.631; rural residence ranked eighth and was not the leading predictor in the service-vulnerability model. Rural delivery costs exceeded urban benchmarks under all modeled scenarios, with the rural 95th percentile ranging from 34.28 RMB at η = 10 to 77.63 RMB at η = 3. Higher subsidy, willingness to pay, and consolidation efficiency increased cooperation levels, but no scanned condition produced stable three-party cooperation within the 35–55 RMB subsidy range.ConclusionsThis study provides a transparent calibrated-simulation stress test rather than an empirically validated route-reconstruction or subsidy-threshold study. Under the current assumptions, the findings do not support a single transferable subsidy threshold and instead indicate that local pilot calibration, observed route-cost data, and realistic payoff inputs are required before implementation claims can be made.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12913-026-15006-y. Introduction Driven by rapid population ageing, chronic respiratory diseases (CRDs) have become a major health burden among older adults in China. Systematic reviews have estimated the prevalence of chronic obstructive pulmonary disease (COPD) among Chinese adults aged ≥ 40 years at approximately 9.9% [1]. The nationally representative China Pulmonary Health (CPH) study reported a higher standardized prevalence of 13.6% [2]. Previous studies further suggest that rural populations may have an elevated risk of COPD and chronic respiratory symptoms. This disparity has been associated with structural factors, including high smoking prevalence, solid-fuel exposure, and limited medical resources [3, 4]. Consequently, older adults in rural areas may be especially vulnerable to a cycle of high disease burden and limited access, which may reinforce a health–poverty trap. Chinese herbal decoctions, particularly pre-packaged liquid formulations, remain an important modality in the long-term management of CRDs. These formulations are commonly used as adjunctive treatments for stable COPD and for symptom control during remission [5]. In routine Traditional Chinese Medicine (TCM) practice, they may also be incorporated into seasonal strategies. Because of their large volume, limited preservative use, and thermal sensitivity, these liquid formulations generally require temperature-controlled handling to maintain quality and clinical usability [6]. Temperature excursions during transport may increase the risk of degradation, contamination, or reduced potency [7]. Therefore, delivering these decoctions to underserved rural populations is not merely a logistical issue; it is also relevant to medical safety, service continuity, and health equity. However, the specific challenges of rural pharmaceutical cold chains remain underexamined. Medical logistics studies have predominantly focused on urban or densely populated settings, using approaches such as vehicle routing problem (VRP) models, heuristic algorithms, and time-window optimization to improve efficiency and reduce costs [8, 9]. These models often assume high population density, sufficient order volume, and well-developed road networks. Although suitable for urban hospitals and commercial distribution, these models may have limited applicability in rural contexts. Health services research further suggests that spatial dispersion of demand points and service locations is a central challenge in rural care delivery [10, 11]. This dispersion limits economies of scale and creates structural diseconomies. Such costs cannot be fully compressed through algorithmic optimization alone because they are partly imposed by geospatial constraints. In health services research, explainable artificial intelligence (XAI) and agent-based modeling (ABM) are increasingly used to analyze inequality drivers, evaluate service accessibility, and simulate policy scenarios. For example, XGBoost combined with SHAP values is widely used for interpretable feature-attribution analysis [11, 12], whereas ABM has been applied to community care, vaccination programs, and logistics networks [13] However, few studies have integrated calibrated spatial simulation, XAI, micro-costing, ABM, and evolutionary game theory within a unified framework. Evidence remains limited on the use of such multidisciplinary frameworks to examine last-mile delivery costs and policy mechanisms for temperature-sensitive pharmaceuticals, including TCM decoctions, in rural regions. This study addresses this methodological and policy-relevant gap by developing a transparent calibrated-simulation stress-test framework rather than a direct empirical reconstruction of delivery access. Using the China Health and Retirement Longitudinal Study (CHARLS) 2018 baseline data as the demographic foundation, we estimated geographic accessibility through facility-density calibration rather than direct observation of patient routes. We then used XGBoost-SHAP to characterize service-vulnerability patterns relevant to home delivery, simulated rural micro-distribution costs using 2023–2024 market parameters, and applied evolutionary game theory to examine policy dynamics under alternative subsidy scenarios without asserting a single critical threshold. This study addresses three research questions. First, what is the magnitude of the modeled rural distance burden faced by older adults with CRDs? Second, to what extent could this burden translate into higher delivery costs under the stated assumptions? Third, under what conditions do scenario-based policy simulations suggest that fiscal support may or may not improve stakeholder cooperation? The contribution is methodological: the framework exposes where calibrated spatial assumptions, route tortuosity, consolidation efficiency, WTP, and payoff specification jointly constrain policy claims. Methods To address the limitations of traditional surveys, which often lack logistics parameters, and purely theoretical models, we developed a four-stage calibrated simulation framework that links nationally representative microdata with policy scenario analysis. The framework integrates CHARLS micro-population data, external facility-density statistics, interpretable machine-learning models, and parameterized cost and policy simulations (Fig. 1). Specifically, the study quantifies broad urban–rural structural contrasts in healthcare accessibility and examines the conditional sustainability of policy interventions under stated assumptions. Additional parameter details are provided in the Supplementary Tables. Fig. 1Four-stage calibrated simulation framework. Note: The framework links calibrated spatial simulation, IADL-based service-vulnerability profiling, parameterized micro-costing, and condition-dependent policy dynamics. Stage I estimates model-derived distance distributions rather than observed patient routes Four-stage calibrated simulation framework. Note: The framework links calibrated spatial simulation, IADL-based service-vulnerability profiling, parameterized micro-costing, and condition-dependent policy dynamics. Stage I estimates model-derived distance distributions rather than observed patient routes Study design and data sources The primary micro-level data were obtained from the 2018 wave of CHARLS. CHARLS covers 28 provinces and 150 county-level administrative units and provides a nationally representative sample [14]. We included respondents aged ≥ 60 years with complete records on demographics, health status, and household economic variables, yielding a final analytic sample of 11,045 individuals. Because liquid TCM decoctions used in respiratory management may require stringent cold-chain handling, respondents diagnosed with chronic lung disease or asthma were defined as the target population for home-delivery services (n = 2,351). To improve simulation transparency and address survey-data limitations, we integrated multiple external public sources for parameter calibration. Because the public-use CHARLS Wave 4 documentation provides nationally representative demographic and health-status information but does not provide patient-level geocoded coordinates or observed residence-to-facility travel routes, CHARLS was used as the demographic and health-status foundation rather than as a source of directly observed route data [14]. Medical-facility density inputs were derived from facility statistics reported in the China Health Statistics Yearbook 2019 [15]. Logistics inputs were informed by 2023–2024 public freight-index reports, provider pricing pages, and public fee schedules for freight, packaging, temperature-control consumables, and labor-related proxy components [16–18]. These external sources were used for broad parameter calibration and plausibility anchoring only, not as patient-level geocoded validation or directly observed local route-cost measurements; source classification, proxy status, sensitivity ranges, direct URLs, access dates, and interpretation boundaries are detailed in Supplementary Tables. Calibrated simulation of spatial access burden Because CHARLS does not provide respondent-level geographic coordinates, we used facility-density calibration and Monte Carlo simulation to generate individual-level estimates of distance to medical facilities (di) [19].These distances were model-derived estimates rather than observed patient travel distances [20]. First, we extracted urban and rural primary healthcare facility densities for each province from the China Health Statistics Yearbook. Based on spatial point-process theory, we derived the expected distance to the nearest facility using a theoretical approximation. Rurality was primarily defined by the CHARLS current-residence variable (“rural”), coded as 1 for rural residence and 0 for urban residence in the analytic data, because delivery exposure occurs at the respondent’s place of living. Hukou status (“rural2”) was retained as a sensitivity definition. This operational definition was chosen because the delivery problem is tied to current place of residence rather than household registration alone. However, rural–urban classification is inherently sensitive to boundary definitions, administrative zoning, settlement morphology, and spatial scale; therefore, the current-residence definition should be interpreted as a pragmatic service-location proxy rather than a definitive classification of rurality [21, 22].Because respondent coordinates were unavailable, the implied spatial-evenness assumption should be interpreted as a first-order approximation for generating broad structural distance distributions, not as a fine-grained representation of local geography. Building on these calibration results, we modeled individual distance to medical facilities (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{d}}_{\mathrm{i}}$$\end{document}) as a random variable with distinct distributions for urban and rural settings: Urban settings were modeled using a normal distribution, reflecting denser and less dispersed facility coverage. Rural settings were modeled using a log-normal distribution to accommodate the right-skewed, long-tailed pattern expected under low facility density. In the current workflow, the log-normal form was treated as the primary calibrated distributional assumption. A matched-moment gamma sensitivity analysis produced nearly unchanged summary statistics, supporting the distributional stability of the main results but not constituting external validation. To quantify internal simulation variability, we performed 200 Monte Carlo iterations per individual. In the CRD subsample, the coefficient of variation for the mean modeled distance was 0.031 among rural residents and 0.015 among urban residents. These values indicate internal Monte Carlo stability only and should not be interpreted as external validation. Service-vulnerability attribution analysis To characterize patients who may face compounded vulnerability relevant to home-delivery prioritization, we constructed a risk-stratification model using extreme gradient boosting (XGBoost). The primary outcome was IADL impairment, defined as an IADL score > 0, because instrumental activities of daily living better reflect practical capacity for medication access, shopping, transportation, and household management than basic ADL items alone. This outcome was an observed CHARLS variable independent of the simulated Stage I distance variable. ADL impairment was retained only as a sensitivity outcome. The main predictor set included sociodemographic and health-status covariates, while ADL score, ADL impairment, and simulated distance were excluded from the main IADL predictor set by design. Rural residence was used as the primary geographic attribute, whereas hukou status was retained only for sensitivity analysis. We computed SHAP (SHapley Additive exPlanations) values to quantify the marginal contribution of each feature to model predictions [12]. In the current framework, this analysis was interpreted as a service-vulnerability module rather than as a direct model of real-world access barriers. Models of hospital utilization and doctor visits were retained as supporting sensitivity analyses. Parameterized micro-costing and economies of scale analysis Instead of relying on coarse regional averages, we developed a parameterized micro-costing model. The model integrated simulated individual distances (di) with 2023–2024 market data to estimate delivery costs in urban and rural settings. Cost parameters were calibrated using public industry standards reported in the Supplementary Tables, including cold-chain transportation rates (1.8–2.6 RMB/km), consumables (0.8–2.0 RMB), and rural labor-cost assumptions informed by village-doctor home-visit subsidies (10–15 RMB/order). To account for shared use of logistics resources, the average delivery cost per individual (Ci) was defined as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{C}}_{\mathrm{i}}\mathrm{=}\frac{{\mathrm{C}}_{\mathrm{ve}\mathrm{h}\mathrm{icle}}\cdot{\mathrm{d}}_{\mathrm{i}}\cdot\beta\mathrm{+}{\mathrm{C}}_{\mathrm{cold}}\mathrm{(}\mathrm{t}\mathrm{)}}{\eta}\mathrm{+}{\mathrm{C}}_{\mathrm{labor}}\mathrm{+}{\mathrm{C}}_{\mathrm{packaging}}$$\end{document} Here, Cvehicle denotes the transportation cost per unit distance, adjusted by the road-network tortuosity coefficient β: 1.2 for urban settings and 1.5 for rural settings. β is used as a structural road-to-geodesic scenario adjustment rather than as a locally observed route coefficient. These values were retained as scenario parameters within a sensitivity range and were supported only by route-circuity literature and small-scale public-map plausibility anchors, not by participant-level GPS or route-log data. Ccold(t) represents time-dependent cold-chain costs, including equipment depreciation and energy use; t includes effective travel time plus fixed loading and unloading intervals of approximately 6–10 min. Clabor and Cpackaging fixed per-order costs for labor compensation and packaging, respectively. The model also introduced a consolidation coefficient (η) to parameterize logistics economies of scale. In the current analysis, η was treated as a scenario input rather than as an empirically identified constant. To address uncertainty in parameter assumptions, we simulated optimistic, neutral, and pessimistic cost scenarios and scanned η values of 3, 5, 8, and 10. The 95th percentile of the rural cost distribution under each scenario was used as a scenario-dependent upper-cost summary for subsequent policy simulations, not as a single empirically validated threshold. Policy scenario simulation under calibrated cost assumptions To evaluate the fiscal sustainability of the proposed service models, we constructed a tripartite evolutionary game framework involving the government, public hospitals, and logistics enterprises [23]. Patients were treated as exogenous actors rather than strategic players, because willingness to pay (WTP) was assumed to remain relatively stable in the short term and to lack dynamic strategic adaptation [24]. Financial parameters in the game were linked to the cost-scenario outputs rather than treated as directly observed empirical quantities. Specifically, the subsidy gap was defined as the difference between the simulated upper-tail rural cost summary and patient WTP. Patient WTP was retained as an affordability-based scenario input rather than interpreted as a study-identified preference estimate. Government payoffs were estimated conservatively, with positive utility assigned only when quantifiable public-health gains, such as medical-insurance savings, exceeded subsidy expenditures; this payoff rule should be interpreted as a conservative scenario specification rather than as an empirically validated fiscal-return estimate, as detailed in Supplementary Tables. To capture system dynamics, we used replicator-dynamics equations [25] and introduced a decision-making lag coefficient (α = 0.08). In the current interpretation, α was treated as an illustrative dynamic-tuning parameter rather than as an empirically identified behavioral constant. This framework was used to examine condition-dependent cooperation dynamics under alternative subsidy, WTP, and consolidation settings. In the current analysis, it was not used to assert a single minimum effective subsidy threshold or a robust stable-cooperation tipping point. Ethical statement and reproducibility This study used publicly available, de-identified CHARLS data for secondary analysis. Because the original CHARLS survey was approved by the Institutional Review Board of Peking University (IRB00001052-11015), additional ethical approval was not required for this secondary analysis. To improve computational reproducibility and transparency, a fixed random seed (seed = 42) was used for all stochastic simulation processes. To further protect respondent privacy, we generated a structurally equivalent synthetic dataset for public dissemination. All relevant analytical code and parameter-configuration files will be made available in a public repository upon publication. Results The key findings are presented sequentially according to the four-stage framework: calibrated simulation of spatial access burden, service-vulnerability attribution, parameterized micro-costing, and condition-dependent policy dynamics. Sensitivity analyses showed that the main qualitative patterns remained stable across alternative parameter settings and distributional assumptions. The reported Monte Carlo variability reflects internal sampling stability rather than external validation. Unless otherwise specified, all findings are based on the neutral simulation scenario. Simulation-derived access burden: structural urban-rural contrast Stage I generated simulation-derived travel-distance estimates that preserved the urban-rural contrast implied by the facility-density calibration. (Fig. 2; Table 1). This contrast should therefore be interpreted as a calibrated structural implication rather than as an independently discovered empirical distance gap. These patterns are consistent with broader evidence on regional accessibility gradients across China [10, 26], but they should be interpreted as model-based estimates rather than as directly observed patient routes. Additional uncertainty analyses and distributional sensitivity results are reported in the Supplementary Tables and Supplementary Figures. Fig. 2Simulation-derived distance distribution and rural tail sensitivity. Note: (A) Density distributions of modeled distance to the nearest medical facility for urban and rural CRD participants under the calibrated Stage I framework. The dashed line marks the > 15 km tail. (B) Matched-moment gamma sensitivity check for rural tail statistics, including the 95th percentile, 99th percentile, and share above 15 km. These distances are simulation-derived estimates rather than observed patient travel routes Simulation-derived distance distribution and rural tail sensitivity. Note: (A) Density distributions of modeled distance to the nearest medical facility for urban and rural CRD participants under the calibrated Stage I framework. The dashed line marks the > 15 km tail. (B) Matched-moment gamma sensitivity check for rural tail statistics, including the 95th percentile, 99th percentile, and share above 15 km. These distances are simulation-derived estimates rather than observed patient travel routes Table 1Baseline characteristics of older adults with chronic respiratory diseases by residence groupCharacteristicUrban (n = 808)Rural (n = 1,543)P-value Demographics Age, mean (SD)70.8 (7.7)70.2 (7.0)0.063Male, %54.3%53.7%0.782Living alone, n (%)108 (13.4%)220 (14.3%)0.541 Socioeconomic status Annual income (RMB), median24,0002,910< 0.001 Health status ADL impairment, n (%)272 (33.7%)594 (38.5%)0.021 Spatial accessibility (simulation-derived) Modeled distance to facility (km), mean (SD)2.85 (1.63)10.61 (5.24)< 0.001Modeled distance to facility (km), 95th percentile6.9424.53< 0.001Share with modeled distance > 15 km0.00%10.76%< 0.001Note: Modeled distance measures are simulation-derived under the Stage I framework and should not be interpreted as observed patient routes. The rural > 15 km entry reflects the simulated distribution under the residence-based main calibration rather than an observed patient rate Table 1Baseline characteristics of older adults with chronic respiratory diseases by residence groupCharacteristicUrban (n = 808)Rural (n = 1,543)P-value Demographics Age, mean (SD)70.8 (7.7)70.2 (7.0)0.063Male, %54.3%53.7%0.782Living alone, n (%)108 (13.4%)220 (14.3%)0.541 Socioeconomic status Annual income (RMB), median24,0002,910< 0.001 Health status ADL impairment, n (%)272 (33.7%)594 (38.5%)0.021 Spatial accessibility (simulation-derived) Modeled distance to facility (km), mean (SD)2.85 (1.63)10.61 (5.24)< 0.001Modeled distance to facility (km), 95th percentile6.9424.53< 0.001Share with modeled distance > 15 km0.00%10.76%< 0.001Note: Modeled distance measures are simulation-derived under the Stage I framework and should not be interpreted as observed patient routes. The rural > 15 km entry reflects the simulated distribution under the residence-based main calibration rather than an observed patient rate Baseline characteristics of older adults with chronic respiratory diseases by residence group Note: Modeled distance measures are simulation-derived under the Stage I framework and should not be interpreted as observed patient routes. The rural > 15 km entry reflects the simulated distribution under the residence-based main calibration rather than an observed patient rate The modeled distance distribution among urban patients was tightly concentrated, with a mean of 2.85 km and a 95th percentile of 6.94 km. This compact profile is consistent with denser healthcare-facility networks in urban areas [27]. In contrast, the rural distribution remained positively skewed and heavy-tailed. The mean modeled distance among rural patients was 10.61 km, approximately 3.72 times that among urban patients; 10.8% of rural simulated observations exceeded 15 km. This long-tailed structure is consistent with the expected consequences of low facility density and dispersed settlements [10]. In the distributional tail check, the matched-moment gamma sensitivity produced a rural 95th percentile of 25.75 km versus 24.53 km under the log-normal main simulation and a rural 99th percentile of 31.27 km versus 30.92 km. These results support directional tail stability within the simulation framework, but they do not validate the estimates against observed routes. Stage II: service-vulnerability patterns relevant to home delivery The IADL-based XGBoost model achieved modest discrimination, with an Area Under the Curve (AUC) of 0.668 and an accuracy of 0.631 (Fig. 3). IADL impairment captures practical limitations in instrumental daily activities and is therefore more defensible for service-prioritization questions than basic ADL alone. An ADL impairment model was retained only as a sensitivity analysis and is reported in the Supplementary Tables and Supplementary Figures. Fig. 3Feature importance in the Stage II IADL service-vulnerability model. Note: Features are ranked by mean absolute SHAP value in the IADL-based XGBoost model. Rural residence is highlighted and ranks eighth in the main model. The model is interpreted as a service-prioritization model rather than a causal access-barrier hierarchy Feature importance in the Stage II IADL service-vulnerability model. Note: Features are ranked by mean absolute SHAP value in the IADL-based XGBoost model. Rural residence is highlighted and ranks eighth in the main model. The model is interpreted as a service-prioritization model rather than a causal access-barrier hierarchy In the SHAP feature ranking, chronic disease burden contributed most strongly to model output, followed by age, sex, income, asthma, and education. Rural residence retained a measurable contribution but ranked eighth, below these clinical, functional, demographic, and socioeconomic variables. This hierarchy indicates that delivery-relevant vulnerability was shaped by multiple factors rather than by residence status alone. These findings support a multidimensional service-prioritization interpretation. Rather than relying on rural residence as a standalone marker, home-delivery targeting would need to consider instrumental functional limitation, chronic disease burden, socioeconomic capacity, and residential context together. Stage III cost results: rural cost disadvantage under parameterized assumptions The parameterized micro-costing model showed clear urban–rural delivery-cost disparities under all modeled scenarios (Fig. 4; Table 2). These results should be interpreted as scenario-based estimates rather than as directly measured delivery costs. Expanded parameter-grid and WTP-gap results are reported in the Supplementary Tables and Supplementary Figures. Fig. 4Rural delivery-cost distribution and one-way parameter sensitivity. Note: (A) Distribution of rural per-delivery costs across consolidation-intensity scenarios, shown as horizontal violin and box plots. (B) One-way tornado sensitivity analysis showing changes in mean rural delivery cost under alternative parameter values. Costs are scenario-based estimates and depend especially on consolidation intensity, route tortuosity, transport price, labor, and packaging assumptions Rural delivery-cost distribution and one-way parameter sensitivity. Note: (A) Distribution of rural per-delivery costs across consolidation-intensity scenarios, shown as horizontal violin and box plots. (B) One-way tornado sensitivity analysis showing changes in mean rural delivery cost under alternative parameter values. Costs are scenario-based estimates and depend especially on consolidation intensity, route tortuosity, transport price, labor, and packaging assumptions Table 2Scenario summary of delivery-cost estimates and policy scan resultsUrban scenario (baseline)Neutral base-case (η = 5)Higher-efficiency rural scenario (η = 8)Low-consolidation rural stress case (η = 3) Input parameters Mean modeled distance (km)2.8510.6110.6110.61η25583β1.21.51.51.5 Cost results Mean delivery cost (RMB)19.5131.9825.8742.8395th percentile cost summary (RMB)20.5652.8638.9277.63Cost gap vs. urban mean (RMB)Reference+ 12.47+ 6.37+ 23.33 Policy scan WTP values scanned (RMB)-8 / 12 / 208 / 12 / 208 / 12 / 20Subsidy values scanned (RMB)-35 / 40 / 45 / 47 / 50 / 5535 / 40 / 45 / 47 / 50 / 5535 / 40 / 45 / 47 / 50 / 55Highest mean cooperation within scan-0.333 (WTP = 20, S = 55)0.346 (WTP = 20, S = 55)0.032 (WTP = 20, S = 55)Robust threshold identified within scan?-NoNoNo Policy interpretation Reference benchmarkPersistent gap;no stable equilibriumImproved response;no stable equilibriumSevere burden;weak cooperationNote: Values summarize cost and policy-scan scenarios. Rural η and β are scenario/proxy parameters, not locally observed constants; source details are provided in the Supplementary Tables. No robust subsidy threshold was identified Table 2Scenario summary of delivery-cost estimates and policy scan resultsUrban scenario (baseline)Neutral base-case (η = 5)Higher-efficiency rural scenario (η = 8)Low-consolidation rural stress case (η = 3) Input parameters Mean modeled distance (km)2.8510.6110.6110.61η25583β1.21.51.51.5 Cost results Mean delivery cost (RMB)19.5131.9825.8742.8395th percentile cost summary (RMB)20.5652.8638.9277.63Cost gap vs. urban mean (RMB)Reference+ 12.47+ 6.37+ 23.33 Policy scan WTP values scanned (RMB)-8 / 12 / 208 / 12 / 208 / 12 / 20Subsidy values scanned (RMB)-35 / 40 / 45 / 47 / 50 / 5535 / 40 / 45 / 47 / 50 / 5535 / 40 / 45 / 47 / 50 / 55Highest mean cooperation within scan-0.333 (WTP = 20, S = 55)0.346 (WTP = 20, S = 55)0.032 (WTP = 20, S = 55)Robust threshold identified within scan?-NoNoNo Policy interpretation Reference benchmarkPersistent gap;no stable equilibriumImproved response;no stable equilibriumSevere burden;weak cooperationNote: Values summarize cost and policy-scan scenarios. Rural η and β are scenario/proxy parameters, not locally observed constants; source details are provided in the Supplementary Tables. No robust subsidy threshold was identified Scenario summary of delivery-cost estimates and policy scan results Persistent gap; no stable equilibrium Improved response; no stable equilibrium Severe burden; weak cooperation Note: Values summarize cost and policy-scan scenarios. Rural η and β are scenario/proxy parameters, not locally observed constants; source details are provided in the Supplementary Tables. No robust subsidy threshold was identified In urban settings, costs were narrowly clustered around 19.51 RMB under the neutral scenario. Rural cost distributions remained more right-skewed, with a mean of 31.98 RMB at η = 5 under the neutral scenario. This pattern is consistent with cost escalation under the combined constraints of longer transport distances and lower order-consolidation volumes. To summarize upper-tail cost exposure, we examined the 95th percentile of the cost distribution. Under the neutral scenario, the rural 95th percentile ranged from 34.28 RMB at η = 10 to 77.63 RMB at η = 3, with an estimate of 52.86 RMB at η = 5. This finding indicates that upper-tail cost summaries are highly sensitive to consolidation assumptions and should not be treated as fixed empirical ceilings. Stage III sensitivity and internal heterogeneity Additional Stage III sensitivity analyses showed substantial heterogeneity within rural subsets and clear dependence on key cost parameters (Fig. 4). Additional one-way sensitivity results and scenario summaries are reported in the Supplementary Tables and Supplementary Figures. Illustrative heterogeneity under alternative rural conditions Under more favorable rural conditions, simulated costs remained above the urban benchmark, whereas low-consolidation and long-distance rural settings produced substantially higher mean and upper-tail costs. This stratified cost gradient is consistent with expected variation in route conditions across plain, hilly, and mountainous settings [28, 29]. Scenario analysis: bounds of uncertainty Following multi-scenario approaches commonly used in health-service planning, we evaluated optimistic, neutral, and pessimistic scenarios. The rural mean cost was 17.50 RMB under the optimistic scenario, 31.98 RMB under the neutral scenario with η = 5, and 58.42 RMB under the pessimistic scenario. Across all scenarios, the rural–urban gap persisted. This pattern suggests a persistent rural cost disadvantage under the stated assumptions, although its magnitude remained parameter-dependent. Policy evolution trajectories: condition-dependent cooperation without a robust threshold Given the scenario dependence of both upper-tail costs and willingness-to-pay assumptions, the policy game was re-evaluated across WTP values of 8, 12, and 20 RMB and subsidy values of 35, 40, 45, 47, 50, and 55 RMB. As a weak external anchor, Chinese evidence on home-based services among older adults suggests that per-visit WTP may be low, with values around 8 RMB remaining plausible [30]. The evolutionary-game analysis therefore focused on conditional responsiveness rather than on a single fixed tipping point (Fig. 5). The neutral-scenario payoff matrix is shown in Table 3, and the alpha-sensitivity analysis is shown in Table 4; additional policy-scan details are provided in the Supplementary Tables and Supplementary Figures. Fig. 5Stage IV policy condition map under alternative subsidy, willingness-to-pay, and consolidation scenarios. Note: Tiles show mean final cooperation levels across the scanned subsidy, WTP, and consolidation settings. Higher cooperation levels cluster under higher subsidy, higher WTP, and higher consolidation efficiency, but no scanned condition achieved stable three-party cooperation Stage IV policy condition map under alternative subsidy, willingness-to-pay, and consolidation scenarios. Note: Tiles show mean final cooperation levels across the scanned subsidy, WTP, and consolidation settings. Higher cooperation levels cluster under higher subsidy, higher WTP, and higher consolidation efficiency, but no scanned condition achieved stable three-party cooperation Table 3Payoff matrix used in the neutral Stage IV scenarioStrategy profileGovernment payoffHospital payoffLogistics payoffG: C / H: C / L: C19.864.009.14G: C / H: C / L: D-53.00-4.000.00G: C / H: D / L: C-53.000.009.14G: C / H: D / L: D-53.000.000.00G: D / H: C / L: C0.00-4.00-40.86G: D / H: C / L: D0.00-4.000.00G: D / H: D / L: C0.000.00-40.86G: D / H: D / L: D0.000.000.00Note: G = government, H = hospital, L = logistics enterprise. C indicates cooperation and D indicates defection. Payoffs are scenario-specific values for the neutral η = 5, WTP = 12 RMB, subsidy = 50 RMB setting and should not be interpreted as empirically observed fiscal or institutional payoffs Table 3Payoff matrix used in the neutral Stage IV scenarioStrategy profileGovernment payoffHospital payoffLogistics payoffG: C / H: C / L: C19.864.009.14G: C / H: C / L: D-53.00-4.000.00G: C / H: D / L: C-53.000.009.14G: C / H: D / L: D-53.000.000.00G: D / H: C / L: C0.00-4.00-40.86G: D / H: C / L: D0.00-4.000.00G: D / H: D / L: C0.000.00-40.86G: D / H: D / L: D0.000.000.00Note: G = government, H = hospital, L = logistics enterprise. C indicates cooperation and D indicates defection. Payoffs are scenario-specific values for the neutral η = 5, WTP = 12 RMB, subsidy = 50 RMB setting and should not be interpreted as empirically observed fiscal or institutional payoffs Payoff matrix used in the neutral Stage IV scenario Note: G = government, H = hospital, L = logistics enterprise. C indicates cooperation and D indicates defection. Payoffs are scenario-specific values for the neutral η = 5, WTP = 12 RMB, subsidy = 50 RMB setting and should not be interpreted as empirically observed fiscal or institutional payoffs Table 4α sensitivity under the neutral Stage IV scenarioαGovernmentHospitalLogisticsStable three-party cooperation0.050.00000.05750.3037No0.080.00000.04080.4974No0.120.00000.02550.7476NoNote: Results are shown for η = 5, WTP = 12 RMB, and subsidy = 50 RMB. α changes adjustment speed and final cooperation levels but does not produce stable three-party cooperation under this payoff specification Table 4α sensitivity under the neutral Stage IV scenarioαGovernmentHospitalLogisticsStable three-party cooperation0.050.00000.05750.3037No0.080.00000.04080.4974No0.120.00000.02550.7476NoNote: Results are shown for η = 5, WTP = 12 RMB, and subsidy = 50 RMB. α changes adjustment speed and final cooperation levels but does not produce stable three-party cooperation under this payoff specification α sensitivity under the neutral Stage IV scenario Note: Results are shown for η = 5, WTP = 12 RMB, and subsidy = 50 RMB. α changes adjustment speed and final cooperation levels but does not produce stable three-party cooperation under this payoff specification Null-result policy scan under the current payoff specification Across the evaluated combinations, higher subsidies, higher WTP, and greater consolidation efficiency generally increased the final cooperation share of logistics firms. However, none of the tested parameter combinations produced stable three-party cooperation. Under the current payoff specification, the policy module therefore did not identify a robust subsidy threshold. Conditional responsiveness within the evaluated subsidy range Within the evaluated subsidy range, cooperation outcomes were sensitive to willingness-to-pay assumptions, consolidation efficiency, and the dynamic tuning coefficient α. In the neutral η = 5, WTP = 12, S = 50 setting, α values of 0.05, 0.08, and 0.12 changed the final logistics cooperation share from 0.304 to 0.497 and 0.748, respectively, while government cooperation remained near zero and hospital cooperation remained low (Table 4). Higher WTP and higher η generally produced larger final cooperation shares among logistics enterprises, particularly when the subsidy was set at 55 RMB. However, none of the evaluated parameter combinations generated stable three-party cooperation. These findings indicate that the policy response surface was condition-dependent rather than threshold-like. Therefore, the Stage IV results should be interpreted as identifying parameter regions associated with stronger cooperative tendencies, not as establishing a minimum effective subsidy level. Discussion By integrating calibrated simulation, interpretable modeling, and policy-scenario analysis, this study examines the rural logistics penalty affecting TCM delivery access for older patients with CRDs in rural China. Although this study focuses on liquid TCM decoctions, the identified logistical challenges—high-volume, temperature-sensitive, and high-frequency delivery—are not unique to TCM [31]. Therefore, the proposed collaborative cold-chain framework may also inform rural last-mile delivery of other medical products with strict storage requirements, such as biologics, insulin, and enteral nutrition products. Structural cost disadvantage under geographic constraints This study suggests that rural last-mile delivery involves structural costs that may not be readily eliminated through operational optimization alone. In the current model, these costs arise from sparse population distribution, longer delivery distances, and weaker economies of scale under low order density, as reflected in the Stage III cost results and parameter-source audit. These findings suggest that geographic structure may generate persistent cost disadvantages under the stated assumptions [32]. The simulation-derived estimates imply that the mean modeled distance for rural care was approximately 3.72 times that for urban care, with a substantive rural long tail beyond 15 km. Because this contrast is inherited from the facility-density calibration, it should not be read as an independently observed geographic finding. Stage III results further indicate that even under optimistic assumptions about pooling efficiency, rural delivery costs remained higher than urban benchmarks. These patterns are consistent with broader concerns about spatial disparities in health resources [10, 33], but they should be interpreted as model-based evidence rather than as direct empirical validation of market failure. Comparison with existing literature: from description to quantification The current literature has focused primarily on three domains. First, urban logistics-optimization studies have improved operational efficiency [8, 9], but they generally rely on high order-density assumptions that are poorly suited to rural contexts [34]. Second, spatial accessibility studies have highlighted inequalities in facility access and rural service sparsity [32], yet most remain confined to distance-based metrics rather than translating these patterns into delivery costs. Third, health-equity and universal health coverage studies have emphasized structural barriers in rural service delivery [35–38], but have rarely translated these barriers into explicit subsidy-sensitive delivery scenarios. In contrast, the present framework does not claim to identify a single validated cost threshold for policy action. Its methodological contribution lies in linking calibrated spatial simulation, IADL-based service-vulnerability profiling, parameterized micro-costing, and condition-dependent policy dynamics within a transparent stress-test framework. Under this interpretation, the value of the study lies less in identifying a transferable subsidy value than in showing how facility-density calibration, route tortuosity, consolidation assumptions, payment tolerance, and payoff specification jointly shape the plausibility and uncertainty of rural delivery models. Policy interpretation under condition dependence The Stage IV analysis indicates that the system remained highly condition-dependent. Higher subsidies, higher WTP, and greater consolidation efficiency moved the simulated dynamics toward greater cooperation. However, none of the evaluated combinations produced a stable three-party cooperative equilibrium. The practical implication is not that a single national subsidy value has been identified. Rather, local pilot calibration and stronger payoff evidence would be more defensible than direct scale-up based on a nominal threshold [39, 40]. Illustrative economic scenario analysis This economic section is presented as an illustrative supplementary scenario, not as a formal within-study economic evaluation. Illustrative calculations were based on chained assumptions about dispensing-channel adherence, adherence-related reductions in exacerbations, and hospitalization risk in COPD [41]. These assumptions drew on external evidence rather than on direct identification within the present study [42–44]. Under these illustrative assumptions, the modeled ICER, ROI, and net-benefit estimates should be interpreted only as literature-informed scenario outputs [45]. They should not be interpreted as direct health-economic findings of this study. Accordingly, the economic module is retained only as a secondary scenario layer to illustrate potential policy plausibility under selected assumptions. Strategic alignment with national health initiatives The proposed delivery mechanism may align with several health-system objectives in China [46]. First, under the tiered diagnosis and treatment system, door-to-door delivery may strengthen the chronic disease management capacity of township health centers [47]. Second, in relation to rural revitalization and public service equalization, integrating cold-chain logistics into rural infrastructure could support broader efforts to improve rural public services [48]. Third, under Healthy China 2030, improved primary-level drug accessibility may help reduce unnecessary visits to tertiary hospitals [46]. Fourth, the model highlights potential linkages among medical insurance funds, primary hospitals, and logistics networks that may improve allocative efficiency in healthcare resources [36]. Pilot-calibration priorities after a null threshold result Because the Stage IV analysis did not identify a stable cooperation threshold, this study should not be read as recommending immediate implementation phases. Instead, the results identify pilot-calibration priorities. First, representative counties would need to measure local WTP, actual logistics costs, route conditions, and service demand before any subsidy value could be defended [40]. Second, collaborative care models, including coordination with village doctors or township health centers, should be treated as hypotheses for improving consolidation rather than as proven mechanisms for increasing η [49]. Third, regional hub consolidation should be evaluated as a possible operational design for increasing order pooling and reducing unit delivery costs, rather than as a validated scale-up pathway [50]. Fourth, digital governance tools, including delivery tracking, temperature monitoring, and subsidy verification, should be assessed as audit and traceability mechanisms rather than assumed prerequisites for immediate deployment [51]. These priorities define the evidence needed before implementation rather than a roadmap for immediate deployment. Institutional implications: advancing Universal Health Coverage (UHC) This study underscores that equity does not imply uniform costs; rather, it requires ensuring that vulnerable populations can access services of comparable quality within affordable limits [46]. Without compensation for structural geographic costs, rural patients with chronic diseases may face practical exclusion from parts of the care pathway [32]. Within this framework, the policy implication is not that a single 50 RMB subsidy has been identified. Instead, local pilot calibration and explicit recognition of structural geographic costs may provide a more defensible basis for government intervention within the UHC agenda [40, 52]. Limitations Data temporality and spatial validation: The demographic baseline relied on 2018 data, whereas the cost parameters were anchored to 2023–2024 market standards. In addition, because CHARLS does not provide respondent-level coordinates, Stage I could not be externally validated against actual patient-level travel routes. The simulated distances should therefore be interpreted as calibrated structural approximations rather than as ground-truth accessibility measurements.Parameter-identification boundaries: Key parameters such as η, β, WTP, and α were not identified as empirical constants in this study. They were treated as structural adjustments, scenario inputs, or dynamic-tuning parameters. Downstream results should therefore be interpreted conditionally on these assumptions.Scenario dependence of policy and economic modules: Willingness to pay was modeled as an affordability-based scenario input, with sensitivity analyses conducted at 8, 12, and 20 RMB [30, 53].Under the current payoff specification, Stage IV did not identify a robust cooperation threshold. The economic analysis remained a literature-informed illustrative scenario rather than a formal economic evaluation [45]. Data temporality and spatial validation: The demographic baseline relied on 2018 data, whereas the cost parameters were anchored to 2023–2024 market standards. In addition, because CHARLS does not provide respondent-level coordinates, Stage I could not be externally validated against actual patient-level travel routes. The simulated distances should therefore be interpreted as calibrated structural approximations rather than as ground-truth accessibility measurements. Data temporality and spatial validation: The demographic baseline relied on 2018 data, whereas the cost parameters were anchored to 2023–2024 market standards. In addition, because CHARLS does not provide respondent-level coordinates, Stage I could not be externally validated against actual patient-level travel routes. The simulated distances should therefore be interpreted as calibrated structural approximations rather than as ground-truth accessibility measurements. Parameter-identification boundaries: Key parameters such as η, β, WTP, and α were not identified as empirical constants in this study. They were treated as structural adjustments, scenario inputs, or dynamic-tuning parameters. Downstream results should therefore be interpreted conditionally on these assumptions. Parameter-identification boundaries: Key parameters such as η, β, WTP, and α were not identified as empirical constants in this study. They were treated as structural adjustments, scenario inputs, or dynamic-tuning parameters. Downstream results should therefore be interpreted conditionally on these assumptions. Scenario dependence of policy and economic modules: Willingness to pay was modeled as an affordability-based scenario input, with sensitivity analyses conducted at 8, 12, and 20 RMB [30, 53].Under the current payoff specification, Stage IV did not identify a robust cooperation threshold. The economic analysis remained a literature-informed illustrative scenario rather than a formal economic evaluation [45]. Scenario dependence of policy and economic modules: Willingness to pay was modeled as an affordability-based scenario input, with sensitivity analyses conducted at 8, 12, and 20 RMB [30, 53].Under the current payoff specification, Stage IV did not identify a robust cooperation threshold. The economic analysis remained a literature-informed illustrative scenario rather than a formal economic evaluation [45]. Conclusion By constructing a multi-source calibrated simulation framework, this study characterized the rural logistics penalty affecting older patients with CRDs who access TCM delivery services. The model suggested a persistent rural cost disadvantage relative to urban settings, driven partly by dispersed residence patterns and weaker economies of scale under the stated assumptions. Under the current assumptions, rural cold-chain delivery services appeared economically fragile because higher distance burdens and limited consolidation constrained cost recovery. The Stage IV analysis did not show that a single subsidy of 50 RMB/order could reliably establish a stable cooperative mechanism. Bridging this logistics gap should therefore be viewed less as a fixed-threshold subsidy problem than as a combined problem of infrastructure development, demand-side support, and local pilot calibration. Future policy formulation may need to integrate targeted subsidies, route consolidation, primary-care coordination, and auditable governance before large-scale implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2