Background
On average, every 34 and 40 seconds, myocardial infarction (MI) and stroke events occur in the US, respectively [
1]. Stroke ranks third in causes of death and is the leading cause of debilitation among Americans [
2]. It is estimated that approximately 15% of those who have an MI will die of it [
1]. These health conditions are serious economic burdens to the US health system with annual costs estimated at $73.7 billion for stroke and $177.1 billion for MI [
1].
Place of residence is an important determinant of cardiovascular health and disparities in the burdens of stroke and MI have been observed for different geographic areas [
1‐
3]. The highest risks of mortality have been reported in the southeastern US [
1,
4‐
6] and in populations living in rural areas [
7‐
9], particularly in the Appalachian region [
10,
11]. Many areas of the Appalachian region, including parts of Tennessee, form a portion of the US "stroke belt". Tennessee ranks 3
rd highest in the US for stroke [
1], and had an annual age-adjusted stroke mortality risk for the period 2000-2006 of 67.5 deaths per 100,000 persons compared to the national risk of 53.5 deaths per 100,000 persons [
12]. For coronary heart disease including MI, Tennessee ranks 4
th highest in the US [
1] with an annual age-adjusted mortality risk for the period 2000-2006 of 85.5 deaths per 100,000 persons compared to the national risk of 58.9 death per 100,000 persons [
12]
The geographic distributions of stroke and MI mortality have been investigated at state and county levels [
1,
5,
11]. However, geographic disparities have been shown to exist even after adjusting for variations in common risk factors like demographic factors (race, age), socioeconomic measures (income, education), behaviors (smoking, physical activity), and other conditions (diabetes, hypertension) [
4,
10,
11,
13]. These findings suggest that geographic variation in stroke and MI mortality could be due to more localized distributions of neighborhood risk factors. The clustering of determinants of stroke and MI at the neighborhood level can greatly affect the planning, implementation, and focus of health initiatives that seek to reduce disparities. Therefore, research should focus on identifying disparities at the neighborhood level to better understand health needs and thus, provide needs-based health services [
3,
14]. While many studies have defined neighborhoods as census tracts or smaller geographic units, the neighborhoods have not been used as the unit of analysis for many past studies investigating cardiovascular disease and stroke [
15‐
21]. Rather, these studies have investigated neighborhood characteristics as contextual effects in multilevel models that seek to explain individual level risk. Thus, ecological studies are needed to investigate the spatial patterns and clustering of high mortality risk with the neighborhood as the unit of analysis since this is important in identifying high risk communities and targeting resources to address health disparities and improve population health at the local level.
When investigating disease patterns in small geographic areas like neighborhoods, however, there are some challenges that must be addressed. Due to population heterogeneity, mortality risks from areas of low population will likely have higher variances and therefore be more unstable than those from areas of high population [
22]. This variance instability of small geographic areas is referred to as the small number problem [
23]. Spatial smoothing of risks is used to mitigate this issue by reducing the "noise" from areas with low population and therefore high variances [
24].
With these issues in mind, the objective of this study was to investigate spatial patterns and detect local neighborhood clusters of high risk of stroke and MI mortality in the East Tennessee Appalachian Region. The identification of neighborhoods with high risks is expected to aid local health planners in understanding the specific neighborhood health needs to guide health planning and provision of health services. Thus, identified clusters of high risks of stroke and MI mortality will be useful in guiding resource allocation, service provision, and policy decisions at the local/neighborhood level that are crucial for addressing neighborhood health disparities.
Discussion
The results show that spatial patterns of high risk of stroke and MI exist in the study area. These findings are consistent with those from other studies that have reported that southern states like Tennessee [
1,
6,
9,
34,
44], and specifically Appalachian counties [
10,
11,
57], have excess risk of stroke and MI. The excess risk has mostly been attributed to variations in the distribution of stroke and MI risk factors such as race, socioeconomic status, geography (urban vs. rural), and prevalence of other chronic diseases, such as diabetes and hypertension [
3,
6,
9,
58]. However, other studies have reported that geographic disparities exist even after adjusting for variations in these risk factors [
4,
10,
11,
13]. The apparent inconsistency in the association between high risks of stroke/MI and risk factors at the state and county levels suggests that disparities may be due to more localized distributions of risk factors.
To our knowledge, this is the first study to investigate spatial patterns and clusters of stroke and MI risk to better understand observed disparities and identify specific health needs at the neighborhood level to aid population health planning. The results of the current study provide evidence that the risk of stroke and MI can be highly variable within a county and therefore studies that perform analyses at the county level fail to identify these disparities at lower (neighborhood) levels. For example, Knox and Hamblen counties are often reported to have lower risks of stroke and MI and are not considered economically distressed/disadvantaged when compared to other counties in the area [
10,
11]. However, it is evident from the findings here that a few neighborhoods in these counties have very high risks and are part of significant spatial clusters for stroke and MI. If analyses, research, and planning activities to address disparities in risk are conducted at county or higher levels as is often done, these spatial disparities within the counties would be missed. Therefore, neighborhoods would likely be erroneously ignored in programs geared towards addressing disparities in MI and stroke risk. The implication is that for health research and planning activities to be most effective, the focus must be on neighborhood level characteristics and specific needs to alleviate the variation seen at higher geographic levels.
Other studies have used multilevel analyses, including both neighborhood and individual characteristics, to describe disparities in MI risk for individuals [
15‐
21]. One study, using data from the Atherosclerosis Risk in Communities Study, categorized neighborhoods (CTs) into tertiles by neighborhood median household income and found that greater incidence risk of MI was associated with living in lower income neighborhoods [
38]. Diez Rouz, et al. (2001) also found that living in a disadvantaged neighborhood was associated with increased incidence of coronary heart disease, including MI, while adjusting for individual income, education, and occupation and defining neighborhoods as census block groups [
18]. However, some differences in incidence remained between neighborhoods after adjusting for common socioeconomic factors. The failure of individual level risk factors to substantially explain risk at aggregated levels is a common finding in multilevel studies [
45]. Some authors have suggested that neighborhood level socioeconomic variables capture information above and beyond the individual level, and so do not serve only as proxies for individual risk factors [
21]. Similar to reports from other studies [
16,
21], we found that neighborhoods with a high proportion of the population with low education had higher stroke and MI risks. However, we did not find significant association between median household income and risk of MI or stroke. This is contrary to findings from previous studies [
15,
18,
38,
43] and is likely because these were individual level studies while ours is a population/group (neighborhood) level study. In addition to the level of education, the confounding identified between the geography (urban versus rural), race, and gender distribution of each neighborhood is potentially important to understanding how geographic disparities arise in the study area. The influence of neighborhood socioeconomic and social conditions on health may be related, in part, to availability and accessibility to health care services, the built environment and infrastructure (i.e. quality schools, recreational facilities, stores and restaurants with healthy foods), neighborhood based attitudes towards health and related behaviors (i.e. smoking, physical activity, and diet), and the degree of social support [
14,
20,
59,
60]. Since health planning is performed at the population level, identifying geographic disparities for neighborhoods can provide insight into the social conditions, structures, and mechanisms that influence health outcomes in the population to better provide effective population based education campaigns and prevention strategies. Thus, studies, such as this one, that investigate neighborhood level patterns in risk should be considered in addition to those multilevel studies that assess risk of individuals in neighborhoods to ensure community health resources, services, and other efforts are best targeted to the populations at greatest risk.
Although mortality data are useful and commonly used in epidemiological studies to assess health and its patterns, they are not without limitations. First, the accuracy of the cause of death given on a death certificate can be affected by errors made by physicians or in coding, differences in diagnostic criteria, issues arising when there are multiple causes of death, or errors in data entry [
61]. Lloyd-Jones et al. (1998) reported that death certificates overrepresented coronary heart disease as cause of death, particularly for older populations, and cautioned that its use in etiologic studies could potentially lead to a bias towards the null value [
62]. There is also concern that mortality data reflects past, rather than current, health needs. However, mortality is often the most commonly available data for observational, population-based studies since (in the US) it is freely available through organizations, like health departments and the Centers for Disease Control and Prevention [
61]. Unfortunately, the mortality data in this study contained only decedent's residential address for geo-coding to the census tract level and gave no information on whether the address was a place other than a private home, such as nursing homes or prisons, thus limiting the ability to assess any effect such issues would have on the results of the study. However, we did identify to the best of our ability, the addresses known to be nursing homes and found that no more than 15 deaths occurred at any given address. Thus, we do not believe these issues would significantly affect the spatial patterns observed.
From a methodological standpoint, while neighborhood level analyses provide the advantage of better insight and understanding of health disparities and needs, they are not without limitations. Due to the small number problem, visualization of raw risks from areas with low population or small number of deaths can be misleading. In this study, this problem was overcome using SEB smoothing of risks that reduces noise associated with population heterogeneity and variance instability by borrowing strength from neighbors. While the removal of noise from low populations with unstable risks eases visual interpretation, it may possibly introduce artifacts into the map [
24,
63] and therefore these risks should only be used for visualization and not statistical analyses [
64,
65]. Additionally, many smoothing techniques, including the SEB used in this study, are prone to edge effects such that neighborhoods on the edges of the study area have fewer neighbors than those in the interior, so there is less information to borrow from neighbors in smoothing [
23]. Thus the risks are shrunk toward a global instead of the local mean. Despite these disadvantages, spatial smoothing of risks minimizes erroneous visual interpretations associated with raw risks by reducing noise, making spatial patterns more evident, and reducing attention to outliers by focusing on the overall geographic pattern of the study area [
23]. In this study, the smoothed risks did not change the raw pattern very much, except to make localized patterns more visually obvious for both stroke and MI. This result indicates that extreme values (very high and low risks) in the wide mortality risk range were composed of neighborhoods with stable risks, i.e. risks with low variance. Since the SEB has a larger impact on unstable risks and little to no impact on stable risks (i.e. those with low variances) [
23,
64], it is not unexpected that there were minimal differences between the raw (unsmoothed) and SEB risks.
The visual interpretation of spatial patterns can be strongly affected by the number and width of class intervals used to represent risk values [
23,
66]. To reduce this potential bias, it has been suggested that intervals should be based on the overall shape of the distribution and not statistical frequency [
66]. Thus, this study employed the Jenks, or natural breaks, classification method which defines intervals based on the natural distribution of breaks or groupings in the data [
67]. The visualization of spatial patterns of disease is an important component in identifying geographic disparities. However, it is standard epidemiology practice not to rely on one's visual interpretation of a map of disease risks to differentiate significant spatial clusters from what may seem to be a cluster visually but is not statistically significant [
24,
65]. Furthermore, interpretations of spatial patterns from visual investigations become even more difficult when the population is heterogeneously distributed throughout the study area, resulting in differences in variances of disease risks across different areas in the map. Thus, statistical comparisons are needed to identify areas where statistically significant clusters of stroke and MI mortality exist, while taking into account population distribution, to better understand disease disparities. This explains the need to use SEB risk maps as well as spatial scan statistics to identify significant high risk spatial clusters. Moreover, other studies have also indicated that interpreting the results of cluster detection along with the spatial distribution of risk, especially with Bayesian smoothing, can strengthen findings of spatial analysis [
68‐
70].
Spatial scan statistics were used to identify and assess the statistical significance of areas with high risk of stroke and MI clusters. This methodology, implemented in SaTScan 8.0 [
71], has many advantages over other cluster detection methods: it corrects for multiple comparisons, adjusts for population heterogeneity in the study area, identifies clusters without
a priori specification of their suspected location or size and thus limits pre-selection bias, and allows for adjustment for covariates [
54,
72]. Using visualization of spatial patterns of SEB smoothed risk in conjunction with the results of spatial scan statistics in this study, the neighborhoods with the highest risks were consistent and easy to identify. Detection of spatial clusters of disease allows health planners to effectively identify and plan for the specific characteristics and health needs of the populations with the highest risks of disease [
68,
69]. For instance, median levels of stroke and MI mortality risk were observed for Knox County in the smoothed risk maps, but cluster detection highlighted just a few neighborhoods with statistically significant higher risk than surrounding neighborhoods in the county. The implication is that health planning and programs can be focused to specific neighborhoods of high risk to better meet their health needs instead of using a one-size-fits-all strategy for all neighborhoods within a county. Thus, neighborhood level analysis allows limited resources and efforts to be targeted to the highest risk communities [
68].
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
AP was involved in data acquisition, analyses, and interpretation, as well as preparation of the manuscript. TA was involved in data acquisition and review of the manuscript. AO conceived the research idea and was involved in data acquisition, study design, interpretation of results, as well as extensive editing of the manuscript. All authors certify that they have participated sufficiently in the research to believe in its overall validity and have read and approved the final manuscript.