Spatial analysis of gender variation in the prevalence of hypertension among the middle-aged and elderly population in Zhejiang Province, China

Standardized prevalence, defined as the ratio of observed to expected count in the region under study, is commonly used for mapping the geographic dispersion of disease prevalence. The modeling procedure for SCM can be found in Knorr and Best, Held et al., and Ibanez-Beroiz et al. [31, 32, 38]. However, the models used in these studies most often did not include covariates or potential fixed effects between different diseases (or gender). Furthermore, these investigators mainly focused on rare diseases and adopted the Poisson distribution. Because the prevalence of hypertension is quite common, a binomial distribution was more appropriate for our model settings [34]. Therefore, our study broadens the existing SCM by pinpointing fixed effects according to disease or gender specific regional covariates and extends its application to hypertension, a non-rare disease. To account for regional information, the covariates selected were self-reported smoking, alcohol consumption and obesity. Based on the data collection methods of the China Health and Retirement Longitudinal Study, our study assumed that the sample was a representative sample of the Chinese population, aged 45 years and over. Thus, the extended SCM of our study can be presented as follows:where i denotes region (i = 1,2,…,11), j represents gender with j = 1 for males and 2 for females, and x _j,i was the regional level independent variables, and implications for other variables included:

O[j,i] and n[j,i]: the observed counts of cases and total number of observations, respectively;

eta[j,i] and ν[j,i]: overall and gender specific spatial variation in hypertension relative risk(RR);

δ and 1/δ: weights of shared component for males and females respectively with their ratio equals to δ ² ; and

η _j : proportion variance shared for males and females, respectively;

Of note, the weighting method mentioned above set the sum of the logarithm weight equal to 0, ensuring identifiability of the model.

Because Bayesian inference was used for model fitting and requires the specification of prior distribution for model’s parameters and hyper-parameters, we referred to the work of Arul Earnest et al. [37] and incorporated the structure of our dataset.

Convolution prior was adopted for φ[i] and ν[j,i] [47], denoting they both may be decomposed into the sum of unstructured (e.g. ush, ssh) and spatial structured random effects (e.g. bind, bspat):

Then ush and bind were all assigned Normal prior with their mean zero and α ₀ [j] respectively, while that of ssh, bspat and β[i] were Conditional Autoregressive (CAR) Normal prior and vunstr, vind[j], τ _ssh ,τ _bspat and τ _β are precision parameters for the corresponding prior distributions.

And the fixed effect α[j] and hyper-parameters α ₀ [j] were both assumed to be a non-informative prior. The logarithmic of weights δ was assumed N(0.0, 0.169) [37], which means that δ ² is between 1/5 and 5 with 95 % probability, regardless of whichever disease is labeled 1 or 2. As suggested by Mollie [48], in the absence of prior information, random effects of unstructured and spatially structured were assumed equally important, making all prior distributions of precision parameters the same. For example, the weakly information of gamma distribution Gamma (1, 1.0E-4) (priors 1) was assumed.

The Bayesian SCM models were fitted using Markov Chain Monte Carlo (MCMC) techniques. In order to address reliability, two parallel MCMC chains with widely different starting values were run for each model with a total of 50,000 iterations, keeping every 10th, after a 50,000 iteration burn-in period for each chain. Results were based on thinned sample sizes of 10,000. Brooks–Gelman–Rubin diagnostics [49], as well as graphical checks of chains and their autocorrelations were performed to assess convergence. After convergence, the commonly used deviance information criterion (DIC) [50] was calculated to guide model selection. Finally, to assess the sensitivity of the selected model with respect to different priors for the precision parameters, other alternatives such as Gamma (1000, 5.0E-7) (priors 2) and Gamma (5.0, 5.0E-5) (priors 3) were also considered [51].

Data extraction, management and logistic regression analysis were performed in Stata (version 11, Stata Corp, College Station, USA). The SCM models were run in the free software WinBUGS version 1.4.3 with the GEOBUGS version 1.2 add on. All weight matrices were created using GeoBUGS. Maps of Zhejiang province were first produced in ArcMap version 10.1 (ESRI, USA) and then imported into GeoBUGS.

Combining with results from the multiple logistic regression models, the proportion of the elderly (aged 60 years and over), alcohol consumption, smoking status, obesity and WHtR were used as candidate predictors to assess the correlation between hypertension and age, alcohol consumption, smoking status, obesity and WHtR in middle-aged (45–59 years of age) and elderly (60+ years and older) at the regional level. Hypertension relative risk (RR) of each city for both genders was obtained by MCMC techniques. According to DIC criterion, autocorrelation and convergence of model parameters, a model that included the proportion of the elderly, obesity and those with a WHtR > 0.5 was the most appropriate (Table 3). Further, the observations and expected cases of hypertension for both genders in Zhoushan City were substituted by the average value, and the missing covariates were assumed to have a normal distribution with the mean and variance obtained by the corresponding covariates respectively.

Results of model parameters from the SCM are shown in Table 3. WHtR > 0.5 was a risk factor for hypertension in males, whereas the impact of age and obesity (especially obesity) were uncertain (with confidence intervals containing one) at the regional level. Unlike males, the impact of these factors (especially obesity) was uncertain in females. Despite this uncertainty, the posterior mean of the relative risk of age and WHtR for females were both greater than one with the lower bound close to one. These findings suggest that these factors may significantly increase the risk of hypertension in females in most regions but not in Jinhua, ZS and Quzhou, ZS. Furthermore, obesity was a key factor for both genders in Lishui, ZS and females in the southeast coastal area. Overall, there appeared to be much more uncertainty regarding the impact of obesity on both genders which may indicate that city as a basic unit may not be the most parsimonious tool. Hence, it may be necessary to conduct analysis at finer scale (e.g., at the county level) to improve the precision of these estimations in the future research.

Comparing results of the SCM with that of the multiple logistic regression analysis to determine which model better identifies potential risk factors for hypertension indicated that the risk of hypertension was significantly increased for males whose WHtR > 0.5 at both the individual and regional levels. However, the results of the SCM also showed that a WHtR > 0.5 was also a risk factor for females in most regions (except Jinhua, ZS and Quzhou, ZS) which were not identified in the multiple logistic analysis. Furthermore, the multiple logistic regression analysis demonstrated that obesity was only significant for females. However, the SCM was able to reflect areas associated with obesity visually by mapping the spatial pattern for both genders even if the posterior credible intervals contained one. Finally, the risk of hypertension decreased slightly for males who were 60 years of age and over compared with those 45–59 years of age, contrary to the multiple logistic regression models.

Table 3 further shows that the variance of spatial structured was larger than that of unstructured for both components of shared and gender specific (eg.σ _str  > σ _unstr ), suggesting the existence of a significant spatial effect. The fraction of variance shared was about 49 % for both genders, which indicated the variation of shared random effects was slightly less than gender-specific effects (51 %). Thus gender variation for effects of unobserved covariates in hypertension was relatively important. The posterior mean of the weight of shared components δ being larger than one indicated that the effects of other shared risk factors, in addition to WHtR, such as smoking status, alcohol consumption, may be more important for males. It should be noted that the posterior confidence intervals of δ and η were relatively wide due to the imprecision inherent in estimating such ratio parameters [39], which may be overcome if analyzed at a finer scale (e.g. county).¹ The posterior estimation of parameters by the SCM identified gender variations of age and obesity in hypertension at the regional level and the risk of hypertension was significantly increased for those with a WHtR > 0.5 compared with those with a WHtR ≤ 0.5, especially among males.

In order to investigate the spatial distribution of hypertension and related risk factors in the middle-aged and elderly populations, we mapped the standardized prevalence ratio (SPR), the RR, and posterior mean of gender variations of hypertension by SCM modeling on the GEOBUG map of Zhejiang Province (Fig. 2). Figure 2 (a)- (i) displays the SPR, the RR, posterior mean of RR adjusted for WHtR and age and posterior mean of β of Zhejiang Province in 2012 by gender. The spatial distribution of the RR adjusted for obesity can be found in Figure S2 in the Additional file 3.

Figure 2 (a) and (b) show that the spatial distribution of the SPR for males and females in ZN was larger than that of ZS, although there were slight differences. The largest values occurred in Jiaxing, ZN and Ningbo, ZN for males and Hangzhou, ZN and Shaoxing, ZN for females, whereas the lowest value was in Quzhou, ZS for males. It is well known that the SPR is the maximum likelihood estimator (MLE) of RR [35], which is usually unbiased. However, it may be unstable for areas with small populations as the small changes in the denominator may bring about greater variations in estimates. Therefore, resulting maps may be misleading [32]. To avoid this problem, we assessed the cities with the highest SPR. According to our findings, the expected cases in Jiaxing, ZN were the lowest, implying that the extreme value of SPR for males in this region was most likely due to the occasional changes in the denominators. However, the extreme effects of SPR were smoothed out after spatially smoothing by the SCM (Fig. 2 (c)). Furthermore, the RR for hypertension for females in most areas was higher than those for males after spatially smoothing, and the former showed the spatial pattern of coastal to inland and the previous pattern of SPR for the latter was no longer obvious (Fig. 2 (c) and (d)). A potential reason for the relatively high RR in some regions may be that their corresponding WHtR were larger which in shown in Table 3 and Fig. 2 (e) and (f). By the same reasoning, obesity was also a risk factor for hypertension for both genders in Lishui, ZS and females in the southeast coastal regions (Figure S2 (a) and (b) in the Additional file 3).

Figure 2 (g) and (h) show that in most areas (except Jinhua, ZS and Quzhou, ZS), the RR for hypertension for females was significantly decreased after adjustment for age whereas it was opposite for males, thus implying that 60 years of age or older was a risk factor for hypertension for females yet protective for males to some extent. Finally, Fig. 2 (i) indicates that the RR for hypertension in the southeast coastal regions for males was a bit higher than females, with the opposite trend in other regions, though not obvious. These findings suggest that there were other risk factors resulting in the gender variation of the RR after adjustment for age and obesity, consistent with the implication of η in Table 3.