The Khon Kaen University Ethics Committee approved the present study for Human Research (Reference No. HE642237).

All 77 provinces of Thailand were selected for this study. Thailand is an upper-middle-income country in Southeast Asia with a population of around 66 million and a land area of 514,000 km². The country has a land area of 511,770 km² and a water area of 2,230 km². There are 77 provinces, 878 districts, 7,225 sub-districts and 74,965 villages in the spatially administrative hierarchy.

These data were collected by the Health Data Center (HDC) of Thailand's MOPH between 2016-2021were used. The outcome is defined as having a stroke diagnosed previously by a physician. This study comprised 276,855 participants aged 35 years and above who matched the HDC inclusion criteria and were diagnosed with DM and/or HT. In addition, study participants were currently having insulin therapy or oral medications to control blood glucose levels and blood pressure, along with the participants being treated on a long-term basis.

The independent variables include provincial densities of alcohol stores available in the Center of Alcohol Studies (http://cas.or.th/cas/), provincial densities of tobacco stores available in the Center of Tobacco Studies (https://cts.rutgers.edu/) were collected. Similarly, the provincial concentration of elderly population data was acquired from the National Statistical Office of Thailand. In addition, the provincial concentration of Primary Care data was acquired from GIS Health, Ministry of Public Health (http://gishealth.moph.go.th/). The Land Surface Temperature (LST) data were detected by MODIS sensors, which use band frequencies of 20–23 and 30–31. LST is available from Google Earth Engine (https://developers.google.com/earth-engine/ datasets/catalog/MODIS _006_MOD11A1). The night time light (NTL) data for Thailand were extracted from Suomi National Polar-orbiting Partnership (SNPP) Visible Infrared Imaging Radiometer Suite (VIIRS) satellite available from Google Earth Engine (https://earthengine.google.com).

The GeoDa program was used in analyzing spatial autocorrelation and determining the spatial regression of geographical factors, economy, society, and prevalence of stroke among DM and HT patients in Thailand. The distance was used as a criterion for the weight matrix, and spatial correlation was analyzed. One of the most common approaches for calculating the degree of geographical correlation is to use the spatial autocorrelation statistic (Moran's I) [24]. Moran's I was the slope of the linear fit to the scatter plot. The product-moment correlation coefficient's interpretation was identical to that of the product-moment correlation coefficient. A value of +1 shows strong positive spatial autocorrelation (i.e., clustering of similar values), a value of 0 suggests random spatial ordering, and a value of -1 indicates strong negative spatial autocorrelation. The mathematical specification of the Equation quantifies the strength of a relationship between x_i and its neighbor x_j. The spatial weight matrix w_ij defines the geographical relationship between x_i and x_j. In particular, x_i-x and x_j-x represent the deviation of x_i and x_j from the average of x, respectively. (i.e., clustering of dissimilar values) (i.e., a checkerboard pattern). The formula for calculating Moran’s I statistic is as follows [19, 25].

(1)

Moran’s I have a flaw in determining the correlation's position. As a result, the mathematical core of Moran's I was extended to create local Moran I or LISA. In each location, it calculates spatial association. The generated maps (LISA significance maps) show sites with significant local Moran statistics and categories them by association type (LISA cluster maps). LISA's cluster maps feature four cluster categories: High-High, Low-Low, Low-High, and High-Low, and one pattern representing unpredictability. While Low-High and High-Low are not outliers, the statistically significant spatial correlation areas either imply a positive bivariate relationship with high incidence (High-High) or a positive bivariate link with low incidence (Low-Low) [22, 23].

The relation between sociodemographic factors and stroke rate in 77 provinces was examined with the regression models (i.e., SLM and SEM) using GeoDa version 1.6.6 to identify the presence of statistically significant stroke spatial clusters and multivariate associations [22].

Spatial regression models were used in analyzing the associations among socioeconomic factors, environment and prevalence of stroke among diabetes and hypertension patients in Thailand. The three main specifications of spatial regression models are (1) traditional OLS regression, (2) spatial lag model (SLM), and (3) spatial error model (SEM). SLM is mathematically defined, as shown in the following equation.

(2)

Alternatively, the spillover effect can be integrated into regression using the SEM specification. This framework, in particular, permits the spatial influence of a location j to affect the area ivia the disturbance of cross-boundary relation. SEM is mathematically defined, as shown in the equation.

(3)

The conventional form of the regression equation defines the combination of an independent variable X_i and disturbance u_i jointly determining the dependent variable y_i. Specifies a spillover influence originating from u_j and affecting u_i through the spillover coefficient λ and the spatial weight matrix W_ij. The statistical significance of λ verifies the existence of spatial influence identified in SEM form by using statistical inference. The error term is Ɛ_i, which is independent and identically distributed [21].

The OLS regression has significant limitations in that it frequently assumes that the relationship between dependent and explanatory variables is uniform in space and does not account for spatial autocorrelation, which is frequently regarded as an outright violation of the classical regression model's principle of observation independence. Given that an OLS regression fails to detect the spillover effect, the usual specification is adjusted to include the impacts of neighboring areas, resulting in the SLM and SEM formulations. SLM presupposes a direct spatial influence between a specific location and its surroundings and is based on a spatially lagged dependent variable. Meanwhile, by allowing the influence of neighboring areas to pass through the disturbance term, SEM integrates indirect spatial dependency into the regression model [26]. The Akaike Information Criterion (AIC) value was also utilized to compare the model to the spatial regression model as a model selection decision indicator. It denotes a model with a high degree of goodness of fit. Models with small AIC values indicate a good fit.

The univariate Moran’s I scatter of the annual prevalence of stroke in Thailand’s provinces showed positive spatial autocorrelation. The Moran’s I value was 0.454, with a statistical level of significance (p < 0.05) with permutations (999). The clusters HH of stroke were mostly observed in the central region of Thailand, which included Chai Nat, Sing Buri, Ang Thong, Lop Buri, Ayutthaya, Saraburi, Nakhon Pathom, Nonthaburi, Pathum Thani, Nakhon Nayok, Samut Sakhon, Samut Prakan, Prachin Buri, Chachoengsao, Chon Buri, Buri Ram, and Surin however cold spot (low-low) clusters were located in Sakon Nakhon, Mukdahan, Songkhla, and Pattani mentioned in Table 1 (Figs. 2-4).

Fig. (1). Spatial distribution of stroke.

Fig. (2). Moran’s I scatter plot of prevalence stroke.

Fig. (3). LISA (univariate) map of stroke.

Fig. (4). Significant (univariate) map of stroke.

Moran's I indicated significant statistical association patterns of an independent factor and stroke (p-value < 0.05). There was a spatial correlation between the distribution pattern of alcohol store density in the same direction as the stroke pattern. The outcomes of the bivariate LISA revealed a statistically significant positive correlation between alcohol store density and stroke (Moran's I= 0.184). LISA indicated areas with a concentration of alcohol stored and high prevalence of stroke with high values in the surrounding 3 provinces (Hot-spot or High-High cluster) in Chai Nat, Suphan Buri, Ang Thong, Ayutthaya, Saraburi, Nakhon Pathom, Nonthaburi, Pathum Thani, Nakhon Nayok, Samut Sakhon, Samut Prakan, Prachin Buri, Chachoengsao, Chon Buri, Buri Ram, and Surin. In contrast, LISA analysis showed clusters of a province with a low alcohol store density and stroke with low values of the surrounding 3 provinces (Cold-spot or low-low clusters). There were 4 low-low clusters in Sakon Nakhon, Mukdahan, Songkhla, and Pattani provinces, as shown in Table 2 (Fig. 5).

Fig. (5). LISA and Moran’s I scatter plot matrix of alcohol store density on stroke.

There was a spatial autocorrelation of tobacco store density and a distribution pattern in the same direction with stroke (Moran's I of 0.165). The LISA analysis indicated 17 Hot-spots or High-High clusters of the high concentration of tobacco store and stroke with high values in the surrounding 3 provinces in Suphan Buri, Chai Nat, Lop Buri, Ang Thong, Ayutthaya, Saraburi, Nakhon Nayok, Prachin Buri, Chachoengsao, Chon Buri, Nonthaburi, Pathum Thani, Samut Prakan, Samut Sakhon, Nakhon Pathom, Buri Ram, and Surin. LISA analysis showed clusters of a province with a low concentration of tobacco stores and stroke with the low values of the surrounding 3 provinces (Cold spot). 4 low-low clusters in Sakon Nakhon, Mukdahan, Songkhla, and Pattani provinces are demonstrated in Table 3 (Fig. 6).

It has been revealed that the provinces have a high number of elderly populations high prevalence of stroke with a Moran's I of 0.202. The LISA analysis indicated 17 Hot-spots or High-High clusters of the high concentration of elderly population and patients with stroke with high values in the surrounding 3 provinces in Suphan Buri, Chai Nat, Sing Buri, Ang Thong, Lop Buri, Ayutthaya, Nakhon Pathom, Nonthaburi and Nakhon Nayok. In contrast, LISA analysis showed clusters of a province with a concentration of elderly population and prevalence of stroke was low values of surrounding 3 provinces (Cold-spot or low-low clusters). There were 9 low-low clusters found in Saraburi, Pathum Thani, Samut Sakhon, Samut Prakan, Chachoengsao, Prachin Buri, Chon Buri, Buri Ram, and Surin provinces illustrated in Table 4 (Fig. 7).

Fig. (6). LISA and Moran’s I scatter plot matrix of tobacco store density among stroke patients.

Fig. (7). LISA and Moran’s I scatter plot matrix of concentration of elderly population on stroke.

In addition, the bivariate LISA revealed a statistically significant negative correlation between primary care per population density and the prevalence of stroke (Moran's I= -0.109). LISA indicated areas with a concentration of primary care and a high prevalence of stroke with high values in the surrounding 3 provinces (Hot-spot or High-High cluster) in Chai Nat, Sing Buri, Ayutthaya, Ang Thong, and Saraburi. In contrast, LISA analysis showed clusters of a province with a low primary care per population density and stroke patients with low values in the surrounding 3 provinces (Cold-spot or low-low clusters). There were 3 low-low clusters found in Udon Thani, Sakon Nakhon, and Songkhla provinces. (Table 5) (Fig. 8)

Fig. (8). LISA and Moran’s I scatter plot matrix of the density of primary care per population with stroke.

The temperature of LST (day) mean per square kilometer has a spatial autocorrelation of concentration of LST (day) and a distribution pattern prevalence of stroke in the same direction with a Moran's I of 0.250. The LISA analysis indicated 16 Hot-spots or High-High clusters of the high temperature of LST (day) and patients having a stroke with high values in the surrounding 3 provinces in Suphan Buri, Chai Nat, Sing Buri, Ang Thong, Ayutthaya, Lop Buri, Saraburi, Nakhon Pathom, Nonthaburi, Pathum Thani, Samut Sakhon, Samut Prakan, Chachoengsao, Chon Buri, Buri Ram, and Surin. LISA analysis showed clusters of a province with a concentration of LST (day) and stroke with low values of surrounding 3 provinces (Cold-spot or low-low clusters). There were 3 low-low clusters found in Mukdahan, Songkhla, and Pattani provinces, as can be seen in Table 6 (Fig. 9)

Moreover, there was spatial autocorrelation of concentration of NTL and a distribution pattern in the same direction with stroke with a Moran's I of 0.120. The LISA analysis indicated 10 Hot-spots or High-High clusters of the high NTL index and stroke with high values in the surrounding 3 provinces in Sing Buri, Ang Thong, Ayutthaya, Saraburi, Nakhon Pathom, Nonthaburi, Pathum Thani, Samut Sakhon, Samut Prakan, and Chon Buri. In contrast, LISA analysis showed clusters of a province with a concentration of NTL and stroke with low values of the surrounding 3 provinces (Cold-spot or low-low clusters). There were 5 low-low clusters found in Udon Thani, Sakon Nakhon, Mukdahan, Yasothon, and Songkhla provinces, respectively summarized in Table 7 (Fig. 10).

Fig. (9). LISA and Moran’s I scatter plot matrix of LST (day) on stroke.

Fig. (10). LISA and Moran’s I scatter plot matrix of NTL and prevalence of stroke.

The spatial modelling results are summarized in Table 8. The results of the OLS regression and SLM model estimated concentration of the elderly population was likely to be associated with having a stroke. The OLS and SLM models explained approximately 30.40% and 46.20% of stroke prevalence (R² = 0.304) and (R²=0.462), respectively. In addition, the last SEM showed that 4 factors (tobacco store density, the concentration of the elderly population, land surface temperature (day), and nighttime light) were significant predictors and explained approximately 50.80% of stroke prevalence (R² = 0.508).

Therefore, SEM was the best regression model. The parameter estimation shows that the tobacco store density, concentration of the elderly population, land surface temperature (day)and nighttime light density were positively autocorrelated with the prevalence of stroke. The R2 value indicated that SEM accounted for 50.80% of the variation in stroke. In the AIC test, SEM slightly outperformed the SLM with AIC157.71 versus 161.92, respectively. Therefore, SEM showed better performance in explaining the geographical distribution of the prevalence of stroke among diabetes and hypertension patients in Thailand.