Geographical disparities and spatial non-stationarity in stroke prevalence across China: a Bayesian analysis

Abstract

Background

As stroke remains a major public health challenge in China, numerous studies have characterized the epidemiological features and distribution of stroke prevalence across provinces. However, conventional non-spatial analytical approaches may lack the capability to capture spatial dependency and regional variation in the impact of risk factors. This study aims to estimate province-level stroke prevalence in China and quantify how its association with individual-level risk factors varies across provinces, accounting for spatial dependency.

Methods

In this study, 19,713 adults were included from the fourth China Health and Retirement Longitudinal Study (CHARLS 2018), spanning 28 provinces, autonomous regions, and municipalities. Within each province, prevalence estimates were standardized to the 7th National Census (2020) distribution of age, sex, and residence type. Stroke prevalence and 95% Bayesian credible intervals (BCIs) were estimated by a Bayesian spatially varying coefficient model. Global and local Moran’s I statistics were used to assess spatial autocorrelation and identify clustering patterns. Eight metabolic, lifestyle, and socioeconomic risk factors were considered: lower educational attainment, hypertension, diabetes, heart disease, dyslipidemia, smoking, alcohol consumption, and physical inactivity. The model estimated how the association of each with stroke varied across provinces.

Results

Stroke prevalence at province level in China showed marked geographic disparities, ranging from 1.89% (95% BCI: 0.92%–3.61%) to 8.64% (95% BCI: 6.98%–10.63%). A distinct “North-high, South-low” spatial gradient was observed, with significant positive spatial autocorrelation (I = 0.428, p < 0.001). Local cluster analysis identified high–high clusters in Northeast and North China and low–low clusters in South and East China. Except for low educational attainment, the association between stroke prevalence and smoking, drinking, physical inactivity, hypertension, diabetes, dyslipidemia, and heart disease exhibited significant provincial variation. Hypertension showed the strongest association with stroke, with odds ratios ranging from 2.39 to 3.38 across provinces.

Conclusions

Stroke burden in China is spatially clustered, and the associations between stroke and its risk factors vary markedly across provinces rather than being uniform nationwide. By integrating spatial non-stationarity with census-based demographic standardization, this study provides spatially refined evidence to support region-specific stroke-prevention strategies and optimize the allocation of healthcare resources.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12883-026-05173-0.

Background

Stroke represents a major public health challenge in China, accounting for nearly one-third of total stroke-related mortality (2.19 million) and incident cases (3.94 million) worldwide in 2019 [1, 2]. Over the past decade, the epidemiological profile of stroke in China has undergone a notable demographic transition. While the disease has traditionally been reported among elderly populations, there has been an increase in incidence among younger populations [3, 4], largely driven by earlier onset of metabolic risk factors and rapid lifestyle changes [5, 6]. Apart from this demographic pattern, the incidence of stroke also exhibits regional variability across provinces [7–10]. Significant differences in the accessibility of healthcare facilities, socioeconomic development, and lifestyles can also contribute to the incidence of stroke and the effectiveness of stroke preventive care [11, 12]. In a country with such large geographic and demographic variability, reliance on aggregated national estimates may obscure high-risk populations, which would be detrimental to the equitable distribution of healthcare resources as well as the formulation of targeted public healthcare policies [13, 14]. Therefore, accurately characterizing provincial patterns of stroke prevalence has become increasingly important for identifying high-risk populations and informing region-specific prevention strategies.

Although several national surveys and meta-analyses have provided provincial-level estimates of stroke prevalence [7, 8, 15], these studies have generally focused on describing geographic variation in disease burden while paying limited attention to the spatial processes underlying these patterns. First, conventional non-spatial analytical frameworks treat provinces as independent units and ignore spatial autocorrelation, the tendency for geographically adjacent regions to share similar environmental, climatic, and dietary characteristics [16–19]. As a result, the spatial dependence between provinces is left unmodeled. Second, most published studies assume that the impact of risk factors such as hypertension and smoking are homogeneous across regions. However, the associations between stroke and its risk factors may vary geographically because of differences in environmental, socioeconomic, and healthcare contexts [20]. Third, provincial comparisons often rely on crude prevalence estimates that do not adequately account for differences in demographic composition across regions. As age is a major determinant of stroke, unstandardized estimates may distort cross-province comparisons of disease burden. In addition, conventional approaches often lack sufficient statistical smoothing, making prevalence estimates more vulnerable to sampling variability. Few provincial-level studies of stroke in China have simultaneously addressed spatial dependence, spatially varying risk-factor effects, and census-based standardization.

To address these limitations, this study applied a Bayesian spatially varying coefficient model to the 2018 wave of CHARLS, comprising 19,713 adults across 28 provincial-level regions in China. First, a spatially structured random effect captures spatial dependency in stroke prevalence and reveals clustering patterns. Second, each risk factor is assigned a spatially varying coefficient, allowing its association with stroke to differ across provinces rather than being constrained to a single national value. Third, prevalence estimates were standardized to the age, sex, and residence structure of the 7th National Census (2020) to improve cross-province comparability, and the Bayesian framework borrows strength across neighboring provinces, stabilizing estimates in areas with small provincial samples. Therefore, the aim of this study was: (1) to estimate standardized, province-level stroke prevalence and characterize its spatial clustering; (2) to quantify how the associations between stroke and eight metabolic, lifestyle, and socioeconomic risk factors vary across provinces. By identifying regions with concentrated stroke burden and locally important risk factors, these findings may guide region-specific prevention strategies and efficient allocation of healthcare resources.

Methods

Figure 1 summarizes the analytical workflow of this study, which proceeded in three main stages — data preparation, Bayesian modeling and prevalence estimation, and spatial autocorrelation analysis — followed by a sensitivity analysis.

Fig. 1.

Methodological workflow of the study

Study population

This study is a secondary analysis of de-identified, publicly available data from CHARLS, which was approved by the Ethical Review Committee of Peking University (IRB No. IRB00001052-11015); all participants provided written informed consent at the time of the original survey. The study was conducted in accordance with the Declaration of Helsinki, and no additional ethics approval was required for this secondary analysis.

The data were collected from Wave 4 (2018) of CHARLS [21], a population-based, nationally representative longitudinal household survey. The health information, including socio-demographic characteristics, lifestyle factors, and disease history, was self-reported by participants during face-to-face, computer-assisted personal interviews (CAPI) using standardized questionnaires. Disease variables based on self-reported physician diagnosis rather than clinical or administrative records. Because population density varies substantially across Chinese counties and communities, the survey selected primary sampling units with probability proportional to their population size (PPS). This makes the design approximately self-weighting, giving individuals in both urban and rural areas a similar probability of selection and supporting national representativeness. Because the aim of this study was to characterize geographic heterogeneity in stroke prevalence at the population level rather than within a specific age group, all eligible participants were retained. The 2018 survey covered 28 provinces, municipalities, and autonomous regions and 19,816 individuals in total; after excluding 103 with missing data on one or more covariates, 19,713 complete cases were included in the analysis (Appendix A).

The provinces and municipalities in mainland China were categorized into seven geographical divisions based on the seven administrative geographical divisions: Northeast, North, Central, East, South, Northwest, and Southwest (Appendix B). The STROBE checklist was followed to guide data extraction (Appendix C).

Ethics approval and consent to participate

This study used publicly available data from CHARLS, which received ethical approval from the Ethical Review Committee of Peking University in June 2008 (IRB No. IRB00001052-11015). All methods in this study carried out in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to data collection. The data analyzed in this study were anonymized and aggregated at the areal level, with no identifiable information at the individual or household level. As this research involved a secondary analysis of de-identified data, no additional ethical approval was required.

Definition of stroke

The study outcome was prevalent stroke. In accordance with previous studies [22, 23], stroke was defined as a self-reported history of physician-diagnosed stroke. Specifically, participants were asked the standardized question, "Have you been diagnosed with stroke by a doctor?", and those who responded affirmatively were classified as having prevalent stroke. Although interview surveys rely on self-report, self-reported physician-diagnosed stroke has been shown to have high specificity and reasonable agreement with medical records in large population-based studies [24].

Demographic characteristics and key risk factors

The eight risk factors were selected because they are well-established, guideline-recognized modifiable determinants of stroke rather than variables chosen for convenience. The INTERSTROKE study identified ten modifiable risk factors that together account for approximately 90% of the population-attributable risk of stroke [25], and these are reflected in the AHA/ASA stroke-prevention guidelines [26]. From this evidence base, we included the factors that are reliably measured in CHARLS, spanning three domains: socioeconomic status (lower educational attainment), cardiometabolic conditions (hypertension, diabetes, heart disease, and dyslipidemia), and behavioral factors (smoking, alcohol consumption, and physical inactivity). Other recognized factors, such as diet, adiposity, and psychosocial stress, were not consistently available in the survey and were therefore not included.

Study area and spatial units

The spatial unit of analysis was the province. The 2018 CHARLS sample covered 28 of China's 31 mainland provincial-level administrative divisions (provinces, autonomous regions, and municipalities); Tibet, Hainan, and Ningxia were not sampled. These units are highly heterogeneous in both area and population. Land area ranges from approximately 6,340 km² (Shanghai) to 1,664,900 km² (Xinjiang), a more than 260-fold difference, and the 2020 census population ranges from approximately 5.92 million (Qinghai) to 126 million (Guangdong). Area and population do not track each other: Shanghai and Xinjiang have similar populations (24.84 and 25.85 million, respectively) despite Xinjiang being more than 250 times larger in area. The province was chosen as the spatial scale because it is the level at which the CHARLS sampling frame and census denominators are defined and at which stroke-prevention policy and health-resource allocation are organized in China. Spatial dependence was modeled using a first-order adjacency structure, with two provinces defined as neighbors if they share a common land border. The characteristics of all 28 provincial units (geographic region, land area, population, and sample size) are summarized in Appendix D.

Models and statistical analysis

To estimate provincial-level stroke prevalence while accounting for individual-level risk factors and spatial dependence across provinces, a Bayesian hierarchical logistic regression model was fitted [27], incorporating province-specific random effects [28].

Let denote the stroke status of individual k in province i, where = 1 indicates a history of stroke and = 0 otherwise (i=1,...,28). The outcome was assumed to follow a Bernoulli distribution as , where represents the probability of stroke for individual k in province i.The hierarchical logistic regression model was specified as:

Where α is the overall intercept, is the vector of explanatory spatial covariates for individual k in province i. And β is a vector of fixed regression coefficients that are constant across all provinces. represents the vector of province-specific random slopes, assumed as . The province-specific random slopes allow each risk factor's association with stroke to vary across provinces. A coefficient larger than the national average indicates a stronger local association in that province, and a smaller coefficient a weaker one. We refer to this as spatially varying effects.

Spatially structured heterogeneity was modeled through a province-level random effect , which was assumed to follow a conditional autoregressive (CAR) prior as [29], where is the element of the spatial weights matrix =1 if provinces i and j share a common boundary and =0 otherwise [30], and denotes the number of neighboring provinces of province i. This component captures spatial dependence in stroke prevalence across adjacent provinces. In addition, an unstructured random effect was included to account for spatially uncorrelated heterogeneity. The integration of these structured and unstructured spatial effects constitutes the classic Besag-York-Mollié (BYM) specification [29]. Epidemiologically, the combined term () serves as a measure of residual risk, representing the influence of unmeasured geographic, environmental, or healthcare-system disparities not captured by the included covariates.

Based on this framework, three hierarchical models were constructed: Model 1, including only province-level spatial random effects without covariates. Model 2, incorporating fixed effects of covariates and spatial random effects but excluding province-specific random slopes. Model 3, extending Model 2 by additionally allowing province-specific random slopes for risk factors to capture heterogeneous provincial effects. Model fit was compared using the Deviance Information Criterion (DIC), a Bayesian analogue of the Akaike Information Criterion that balances goodness of fit against model complexity (the effective number of parameters) and is well suited to comparing hierarchical models estimated within the INLA framework; lower DIC values indicate a better trade-off between fit and complexity [31].

In addition, Moran’s I statistic was used to determine the global and local autocorrelation, geographic clustering of stroke prevalence. Global Moran's I was used to assess the overall spatial dependence (autocorrelation) of stroke prevalence across provinces [32], whereas local Moran’s I was used to identify localized spatial clusters and spatial outliers [33]. The spatial patterns of stroke prevalence were classified into high–high, low–low, high–low, and low–high clusters. A high–high cluster denotes a province with high stroke prevalence whose neighboring provinces also have high prevalence, indicating a regional concentration of elevated risk. A low–low cluster denotes a province with low prevalence surrounded by low-prevalence neighbors, indicating a concentration of lower risk. High–low and low–high patterns identify spatial outliers: a high-prevalence province surrounded by low-prevalence neighbors, or a low-prevalence province surrounded by high-prevalence neighbors, respectively.

Bayesian inference was conducted using the Integrated Nested Laplace Approximation (INLA) approach implemented in the R-INLA package [34]. The posterior distributions were summarized using posterior means and 95% Bayesian credible intervals (BCIs). Weakly informative prior distributions were assigned to the model hyperparameters [35], so that the posterior inferences were driven primarily by the observed data rather than by subjective prior assumptions, while providing mild regularization that stabilized estimation in the spatially varying coefficient model.

log () ~ logGamma (1, 0.1), log (),

log (), log () ~ logGamma (1, 0.01).

To obtain provincially representative prevalence estimates, we constructed analytical weights through a transparent, multi-step standardization procedure. First, we used the individual-level sampling weights () provided by the CHARLS dataset, which adjust for age, sex, and individual non-response. For participants with a missing sampling weight, we imputed the value using the mean of available participants within the same province [36]. So that no eligible participant was dropped solely due to missing weight.

Next, to align the sample structure with the target demographic population, we constructed provincial post-stratification weights () based on the Seventh National Population Census (2020). The post-stratification weight was calculated as:

where and represent the population counts from the Seventh National Population Census (2020) and the CHARLS 2018 sample, respectively, stratified by age group j, sex k, and residence type l in province i [37].

The final analytical weight for each individual m (m = 1,…,19,713) was defined as the product of the two weights:

Finally, the individual posterior probabilities of stroke derived from the Bayesian spatial model were integrated with these final weights to calculate the standardized provincial-level stroke prevalence estimates.

Missingness in the covariates was handled separately from the missing sampling weights described above. For the primary analysis, we employed a complete-case approach. Of the 19,816 individuals surveyed across the 28 provincial-level regions, 103 (0.5%) with missing data on one or more covariates were excluded, leaving 19,713 participants. To assess the robustness of these findings, a sensitivity analysis was conducted on the full sample (n = 19,816), in which the missing covariate values were multiply imputed using random-forest imputation [38]. The imputation procedure was iterated until convergence, with a maximum of five iterations [39]. Each incomplete variable was imputed using demographic characteristics and risk factors as predictors, and the procedure was iterated until convergence (up to 5 iterations). Estimates derived from the multiply imputed datasets were combined using Rubin’s rules [36], which account for both within- and between-imputation variability. To evaluate the imputation performance, a masking experiment was conducted to estimate the imputation error rate; lower values indicate higher predictive accuracy.

Because spatial analyses based on a small number of areal units (28 provinces) can yield imprecise estimates of the spatial variance hyperparameters, we addressed this in three ways. First, the model was fitted to large-scale individual-level data nested within provinces, which stabilizes the fixed effects and the spatially varying coefficients [40]. Second, the BYM specification borrows strength across neighboring provinces through the spatial adjacency structure, stabilizing the province-level random effects when units are few [29]. Third, the resulting uncertainty in the spatial variance was propagated into the Bayesian credible intervals, which we report in full rather than relying on point estimates alone.

The entire process of data management and statistical analysis was conducted using R software (version 4.5.2).

Results

Of the 19,713 participants, 1,424 reported a history of stroke, corresponding to a crude prevalence of 7.2%. The demographic characteristics of the participants across the seven geographic regions are summarized in Table 1, and the spatial distribution of the eight risk factors is shown in Appendix F. Comparison of the three candidate models (Appendix G) showed that Model 1, which included no covariates, had the poorest fit but the lowest complexity, whereas Model 3, which incorporated both fixed effects and spatial non-stationarity, achieved the best absolute fit and the lowest DIC despite its larger effective number of parameters. Beyond model fit, the three models differed in how much of the between-province variation they accounted for. The between-province median odds ratio (MOR), which summarizes the residual structured variation in stroke prevalence on the odds scale, was 1.45 under Model 1. Adding geographically uniform covariate effects left it essentially unchanged (Model 2, MOR 1.46), whereas allowing the covariate associations to vary across provinces reduced it to 1.12 (Model 3), an approximately 90% reduction in the structured between-province variation. Accordingly, all maps and graphical results presented below are based on Model 3.

Table 1.

Characteristics of study participants in CHARLS 2018

Characteristics	Total	Northeast	North	Central	East	South	Northwest	Southwest	P value
	(n = 19,713)	(n = 1363)	(n = 2618)	(n = 3013)	(n = 6074)	(n = 1655)	(n = 1512)	(n = 3478)
Age (Mean ± SD)	61.8 ± 10.4	61.4 ± 9.9	61.0 ± 9.9	61.4 ± 10.1	62.3 ± 10.6	62.4 ± 10.7	60.7 ± 10.0	62.3 ± 10.8	< 0.001
Female	10,414 (52.8%)	724 (53.1%)	1358 (51.9%)	1575 (52.3%)	3212 (52.9%)	902 (54.5%)	811 (53.6%)	1832 (52.7%)	0.457
Residence									< 0.001
Urban	7947 (40.3%)	661 (48.5%)	1101 (42.1%)	1124 (37.3%)	2507 (41.3%)	828 (50.0%)	553 (36.6%)	1173 (33.7%)
Rural	11,766 (59.7%)	702 (51.5%)	1517 (57.9%)	1889 (62.7%)	3567 (58.7%)	827 (50.0%)	959 (63.4%)	2305 (66.3%)
Stroke	1424 (7.2%)	149 (10.9%)	256 (9.8%)	252 (8.4%)	359 (5.9%)	97 (5.9%)	122 (8.1%)	189 (5.4%)	< 0.001
Risk Factors
Smoking status									< 0.001
Never smoking	11,429 (58.0%)	692 (50.8%)	1411 (53.9%)	1805 (59.9%)	3572 (58.8%)	913 (55.2%)	881 (58.3%)	2155 (62.0%)
Former smoking	3024 (15.3%)	210 (15.4%)	452 (17.3%)	451 (15.0%)	976 (16.1%)	240 (14.5%)	226 (14.9%)	469 (13.5%)
Current smoking	5260 (26.7%)	461 (33.8%)	755 (28.8%)	757 (25.1%)	1526 (25.1%)	502 (30.3%)	405 (26.8%)	854 (24.6%)
Drinking status									< 0.001
Never drinking	10,379 (52.6%)	712 (52.2%)	1388 (53.0%)	1647 (54.7%)	3099 (51.0%)	952 (57.5%)	862 (57.0%)	1719 (49.4%)
Former drinking	2700 (13.7%)	152 (11.2%)	336 (12.8%)	383 (12.7%)	747 (12.3%)	195 (11.8%)	233 (15.4%)	654 (18.8%)
Current drinking	6634 (33.7%)	499 (36.6%)	894 (34.1%)	983 (32.6%)	2228 (36.7%)	508 (30.7%)	417 (27.6%)	1105 (31.8%)
Low education	13,734 (69.7%)	845 (62.0%)	1489 (56.9%)	2128 (70.6%)	4375 (72.0%)	1185 (71.6%)	1065 (70.4%)	2647 (76.1%)	< 0.001
Hypertension	7554 (38.3%)	538 (39.5%)	1211 (46.3%)	1077 (35.7%)	2405 (39.6%)	477 (28.8%)	595 (39.4%)	1251 (36.0%)	< 0.001
Diabetes	2532 (12.8%)	180 (13.2%)	407 (15.5%)	401 (13.3%)	868 (14.3%)	159 (9.6%)	186 (12.3%)	331 (9.5%)	< 0.001
Dyslipidemia	4335 (22.0%)	346 (25.4%)	814 (31.1%)	733 (24.3%)	1316 (21.7%)	279 (16.9%)	367 (24.3%)	480 (13.8%)	< 0.001
Heart disease	3892 (19.7%)	507 (37.2%)	750 (28.6%)	542 (18.0%)	921 (15.2%)	100 (6.0%)	435 (28.8%)	637 (18.3%)	< 0.001
Physical inactivity	1966 (10.0%)	138 (10.1%)	303 (11.6%)	251 (8.3%)	742 (12.2%)	130 (7.9%)	128 (8.5%)	274 (7.9%)	< 0.001

Data are presented as mean ± standard deviation (SD) for continuous variables and as number (percentage) for categorical variables. Abbreviations: CHARLS China Health and Retirement Longitudinal Study. P values were calculated using one-way analysis of variance (ANOVA) for continuous variables and the Pearson chi-square test for categorical variables to assess differences among the seven geographical regions

The estimated provincial-level stroke prevalence derived from the final model is illustrated in Fig. 2. In 2018, stroke prevalence across the 28 provinces ranged from 1.89% (95% Bayesian credible interval [BCI]: 0.92%–3.61%) to 8.64% (95% BCI: 6.98%–10.63%) (Appendix H). Based on mean posterior estimates, the top 10% of provinces with the highest stroke prevalence included Heilongjiang, Inner Mongolia, and Jilin, whereas Chongqing, Beijing, and Guizhou ranked among those with the lowest prevalence. From a geographic perspective, provinces with higher stroke prevalence were primarily located in Northeast, North, and Northwest China, while lower prevalence was observed in South and Southwest China, as well as in major municipalities.

Fig. 2.

Spatial distribution of stroke prevalence among adults in China. The map illustrates the estimated provincial-level prevalence of stroke based on the CHARLS 2018 data

Global spatial autocorrelation analysis revealed a significant Moran’s I of 0.428 (Z = 3.697, P < 0.001), indicating strong positive spatial clustering of provincial stroke prevalence. Local indicators of spatial association further revealed significantly high–high clusters primarily in Northeast and North China, whereas low–low clusters were mainly concentrated in South, East, and Southwest China (Fig. 3). In addition, Beijing and Shandong were identified as low–high spatial outliers, representing low-prevalence provinces surrounded by neighboring provinces with higher stroke prevalence. In contrast, a distinct high–low outlier was observed in Hunan, where stroke prevalence was significantly elevated relative to its lower-prevalence surrounding areas.

Fig. 3.

Spatial autocorrelation analysis of Model- based weighted stroke prevalence across provinces. High-High clusters indicate provinces with high prevalence surrounded by high-prevalence neighbors. Low-Low cluster indicates provinces with low prevalence surrounded by low-prevalence neighbors

The results of the Bayesian spatially varying coefficient regression analysis (Model 3), examining the associations between the risk factors and stroke prevalence, are presented in Figs. 4 and 5. Figure 4 summarizes the distributions of province-specific odds ratios across the 28 provinces, whereas Fig. 5 illustrates the spatial variation in the estimated associations of the seven significant risk factors identified in the final model. Former smoking, current drinking, physical inactivity, hypertension, diabetes, dyslipidemia, and heart disease are significantly associated with stroke prevalence. The overall magnitude and provincial heterogeneity of these associations identified hypertension and dyslipidemia as the most influential risk factors, while current alcohol consumption observed a negative association with stroke prevalence, as shown in Fig. 4. The spatial distributions of these estimated associations are further illustrated in Fig. 5, with detailed numerical estimates provided in Appendix I.

Fig. 4.

Provincial-level variation in the effects of covariates on stroke prevalence. Box plots display the distribution of posterior median odds ratios (ORs) across all provinces. The vertical dashed line at OR = 1.0 represents no association. Colors indicate the nature and statistical significance of the factors: Red: significant modifiable risk factors (95% CI > 1); Blue: significant protective factors (95% CI < 1); Purple: significant demographic factors (age and gender); Gray: non-significant factors (95% CI crosses 1). The length of the box indicates the degree of geographical heterogeneity for each risk factor

Fig. 5.

Geographical variation in the effects of modifiable risk factors on stroke prevalence across China. The maps display the provincial-level posterior mean odds ratios (ORs) for seven significant covariates: a Current drinking, b Former smoking, c Physical inactivity, d Diabetes, e Heart disease, f Dyslipidemia, and g Hypertension

Compared with never drinking, current drinking showed an inverse association with stroke across provinces, strongest in Shandong (odds ratio [OR] = 0.69, 95% BCI: 0.58–0.83), and weakest in Inner Mongolia (OR = 0.76, 95% BCI: 0.64–0.91). For smoking status, former smoking was found to be associated with an increased risk of stroke relative to never smoking, with significant spatial variation. The highest risk was observed in Shandong (OR = 1.52, 95% BCI: 1.23–1.89), while the lowest was observed in Henan (OR = 1.28, 95% BCI: 1.03–1.58). Provinces with increased risk estimates of former smoking were mainly located in northern and eastern China. Physical inactivity was significantly associated with increased stroke risk across all provinces, while diabetes exhibited moderate spatial variation, with ORs ranging from 1.29 (95% BCI: 1.09–1.53) in Shandong to 1.40 (95% BCI: 1.18–1.67) in Anhui. Heart disease was also significantly associated with increased risk, ranging from an OR of 1.29 (95% BCI: 1.08–1.54) in Hebei to 1.46 (95% BCI: 1.22–1.75) in Guangxi. Dyslipidemia and hypertension showed more pronounced geographic heterogeneity, with dyslipidemia ranging from 1.87 (95% BCI: 1.59–2.19) in Sichuan to 2.01 (95% BCI: 1.71–2.37) in Guangxi, and hypertension ranging from 2.39 (95% BCI: 1.83–3.13) in Anhui to 3.38 (95% BCI: 2.54–4.49) in Guangdong.

The combined spatially structured and unstructured random effects () represent the residual stroke risk in each province (Fig. 6). This is the variation in stroke prevalence not explained by the measured risk factors. A positive value indicates higher-than-expected prevalence, and a negative value lower-than-expected prevalence. The largest positive random effects were observed in Shandong, Inner Mongolia, and Jilin, indicating the highest residual risk in northern China, whereas the smallest were observed in Guangdong, Zhejiang, and Jiangxi.

Fig. 6.

Spatial distribution of provincial-level random effects for stroke prevalence. The map illustrates the combined spatially structured and unstructured random effects estimated from the Bayesian spatially varying coefficient model. Positive values (in red) indicate provinces with stroke risk higher than expected based on the included risk factors, while negative values (in blue) indicate lower-than-expected risk

The results of the sensitivity analysis are presented in Appendix J. Parameter estimates derived from the multiply imputed datasets were highly consistent with those from the above completed analysis, supporting the robustness of the findings. The imputation error rate for the stroke was 0.13.

Discussion

This study applied a Bayesian spatially varying coefficient model to a nationally representative sample. It examined both where stroke prevalence is concentrated in China and how its associations with established risk factors differ across provinces. Province-level prevalence varied more than fourfold, from 1.89% to 8.64%, and was strongly spatially clustered (global Moran's I = 0.428, p < 0.001), following a north-high, south-low gradient. Notably, the province in which a risk factor showed its strongest association was frequently not the province with the highest burden. The geography of stroke burden and the geography of risk-factor associations are therefore distinct. Distinguishing these two layers, rather than mapping prevalence alone, is the main contribution of this study.

The north-high, south-low gradient is consistent with prior national surveys reporting higher stroke prevalence in northern China [15, 17] and with recent estimates of the 2020 burden [9]. This suggests that the model recovers an established macro-geographic pattern from independent survey data. The contiguous high-high belt across Northeast and North China (Fig. 3) coincides with the regions where hypertension, diabetes, dyslipidemia, and heart disease are most prevalent in our sample (Table 1). Several features of northern China may also contribute, including higher dietary sodium and a colder climate [41–43]. The low-low clusters in southern and southwestern China are correspondingly consistent with warmer conditions and dietary profiles associated with lower cardiovascular risk [42, 43].

The spatial outliers are particularly informative for planning. Beijing and Shandong were low–high outliers, with lower prevalence than their high-risk northern neighbors. In contrast, Hunan was a high-low outlier. The low–high pattern may partly reflect a stronger capacity for secondary prevention and acute stroke care. Regional audits have reported better adherence to stroke-care performance measures in more developed metropolitan regions [44]. The National Stroke Prevention Project has expanded and organized stroke management unevenly across the country [45, 46]. Therefore, clusters and outliers carry complementary planning value. Clusters identify where the underlying risk environment is most adverse. Outliers help distinguish provinces whose burden may be mitigated by health-system capacity (Beijing, Shandong) from those where elevated burden persists despite comparatively favorable surroundings (Hunan), and whose local care pathways may merit closer examination. Some of this variability should be interpreted cautiously. Provinces with few neighbors borrow less information under the spatial model. Their estimates are smoothed less and remain more uncertain.

The geographic disparity could reflect either where risk factors are more prevalent or where they are more strongly associated with stroke. The model comparison points clearly to the second. Adding geographically uniform covariate associations barely changed the structured between-province variation in stroke prevalence, whereas allowing the associations to vary across provinces reduced this variation by about 90% (median odds ratio 1.12). Therefore, the geographic disparity appears to arise less from where risk factors are more prevalent than from where they are more strongly associated with stroke. Hypertension illustrates this most clearly. It was both the strongest associated factor and the most spatially heterogeneous. Its association peaked in Guangdong (OR 3.38, 95% BCI 2.54–4.49) rather than in the high-burden north. One plausible reading concerns the background level of risk. In southern provinces, unmeasured contextual risk are low, as reflected in the small random effects in Guangdong, Zhejiang, and Jiangxi (Fig. 6). Where this background risk is low, the measured factors account for relatively more of the variation, which may allow hypertension to stand out as a stronger correlate of stroke. In the north, many risk factors co-occur and residual risk is higher (Inner Mongolia, Jilin), which may attenuate the relative association of any single factor. This carries a concrete prevention implication: the relative benefit of controlling hypertension may be greatest in the lower-burden southern provinces, even though absolute burden is highest in the north. The behavioral factors fit the same logic. The stronger former-smoking association in northern and eastern provinces likely reflects historically higher smoking prevalence and illness-prompted cessation [47]. The inverse association of current drinking is better interpreted as reverse causation or a sick-quitter effect than as protection [48]. Lower educational attainment showed no significant provincial association, suggesting any relationship operates through region-specific pathways.

The residual spatial variation remaining after adjustment was modest but concentrated in northern China (Fig. 6). It may be associated with unmeasured contextual factors, such as prolonged cold exposure, higher ambient PM2.5 across the North China Plain, and regional disparities in healthcare and socioeconomic conditions [12, 42, 45]. Related work on cardiovascular disease has combined spatial clustering statistics with machine-learning variable importance. A recent county-level analysis of CVD mortality across the continental United States using Moran's I, Getis–Ord Gi*, and Shapley values [49]. Such approaches typically estimate a single, nationally constant importance for each variable, whereas the present model additionally captures how associations vary geographically. A further limitation of the cross-sectional design is that it provides only a static snapshot. Spatiotemporal modeling has been advocated as a more robust framework for linking environmental risk factors to cardiovascular outcomes [50], and has been applied to other geographically structured diseases such as leptospirosis [51]; extending the present model across successive CHARLS waves would allow emerging clusters and temporal shifts in risk-factor associations to be detected.

These patterns argue against a single national prevention template and suggest that strategies should be matched to each region's dominant risk structure. In the high-burden northern provinces, much of the risk appears to be contextual and only partly captured by measured factors. Population-level environmental and lifestyle measures may yield the greatest aggregate benefit, including air-quality action, dietary sodium reduction, and cold-season vascular protection. In the lower-burden southern provinces, hypertension shows its strongest individual-level association with stroke. Intensified clinical detection and management of cardiovascular comorbidities is likely to be more efficient per case. Aligning the type and intensity of intervention with the locally dominant risk profile offers a practical route to more equitable and efficient allocation of stroke-prevention resources.

Several limitations should be acknowledged. First, stroke status was based on self-reported physician diagnosis and is subject to recall bias. Because CHARLS is a cross-sectional household survey that captures only surviving cases, survivor (Neyman) bias may lead to underestimation, particularly where case-fatality is high. Second, limited sample sizes in some age groups and provinces increase uncertainty in the corresponding estimates and widen their credible intervals, although Bayesian smoothing and census-based standardization help stabilize these estimates. Third, the cross-sectional design precludes causal interpretation; the associations reported here reflect co-occurrence rather than effects. Extending the framework to successive CHARLS waves in a spatiotemporal model [50, 51] would be a valuable next step. Fourth, the spatial component was estimated from only 28 areal units, which can reduce the precision of the spatial variance hyperparameters. The CAR prior mitigates this by borrowing strength from neighboring provinces. However, provinces with few neighbors borrow less information; their estimates are smoothed less and remain more uncertain, which may inflate their apparent variability. Finally, the spatially varying coefficients and the contextual factors used to interpret them (climate, air pollution, health-system capacity) are summarized at the provincial level. Inferring individual risk from these aggregate patterns is subject to ecological bias. The coefficients should therefore be read as region-specific average associations, not individual causal effects, and residual confounding from unmeasured variables remains likely.

Conclusions

By jointly modeling the spatial clustering of stroke prevalence and the geographic non-stationarity of its risk-factor associations, this study shows that the relationship between established risk factors and stroke in China is context-dependent rather than uniform: the provinces with the heaviest burden are not necessarily those where a given risk factor is most strongly associated with stroke. A single national prevention strategy is therefore unlikely to be optimal; resources may be used more effectively by matching interventions to each region's dominant risk profile—emphasizing environmental and lifestyle measures in the high-burden north and clinical management of cardiovascular comorbidities in the south. Linking this spatial framework to longitudinal and registry data is the necessary next step to confirm these patterns and track their evolution over time.

Authors’ contributions

HF: Conceptualization, Methodology, Formal analysis, Writing—original draft, Writing—review & editing. JC: Methodology, Prepared the tables and figures, Writing—review & editing. TC: Data collection, Writing—original draft, Writing—review & editing. ZH: Prepared the tables and figures. YY: Methodology, Writing—review & editing. WJ: Methodology, Writing—review & editing, Funding acquisition. SW: Supervision, Conceptualization, Methodology, Writing—review, Funding acquisition. LY: Data processing, Modeling & analysis, Writing—review & editing, Funding acquisition. All authors read and approved the final manuscript.

Funding

This research was funded by Zhejiang Provincial Natural Science Foundation (Nos. LZYQ25H270001, LY24H270007), National Natural Science Foundation of China (No. 82505384), Zhejiang Province Traditional Chinese Medicine Science and Technology Plan Project (No. 2026ZF41).

Data availability

The datasets of the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study used publicly available data from CHARLS, which received ethical approval from the Ethical Review Committee of Peking University in June 2008 (IRB No. IRB00001052-11015). All methods in this study carried out in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to data collection. The data analyzed in this study were anonymized and aggregated at the areal level, with no identifiable information at the individual or household level. As this research involved a secondary analysis of de-identified data, no additional ethical approval was required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.