INTRODUCTION
Type 2 diabetes is a preventable chronic disease. A structured lifestyle modification program for individuals with prediabetes can reduce their risk of developing type 2 diabetes by as much as 58% [
1]. However, the Healthy People 2020 data indicated that in 2015, only 63.8% of adults aged 18 years or older engaged in leisure-time physical activity [
2]. In addition, 36.7% of adolescents and 40.2% of adults stated that they consumed fruit less than once daily [
3]. Physical activity and diet are among the risk factors associated with type 2 diabetes [
4,
5]. Therefore, preventive measures of diabetes should focus on modifying these factors.
Diabetes and obesity are influenced by American culture, which is characterized by the increasing consumption of higher-caloric food, sugar-sweetened drinks, and sedentary behavior [
6]. Instead of focusing on adults, who may be less inclined to change these behaviors, there may be a better chance to influence behaviors in the younger population [
6], and impact not only their current health, but also their future health [
7,
8]. Children spend most of their time at schools and supervised by teachers; therefore, schools provide an ideal setting for conducting lifestyle modifications, especially for the younger population [
9]. Many studies have been conducted on diabetes prevention interventions in schools, suggesting that overall, school-based diabetes prevention programs offer promising outcomes for reducing risk factors for type 2 diabetes [
10–
14]. For instance, in study performed by the HEALTHY Study group, which utilized a school-based program to assess risk factors for diabetes among children, found that their school-based interventions significantly reduced children’s body mass index (BMI), which—as the authors pointed out—could reduce the risk of childhood-onset type 2 diabetes. Diabetes acquired during childhood is known to usher in a host of medical complications, some of which may not manifest until adulthood, including complications such as cardiovascular disease, which is the leading cause of death globally [
10,
15].
Seeing a grave need to address the risk of childhood diabetes, the Centers for Disease Control and Prevention (CDC) established a State Public Health Actions grant, which funds statewide programs that aim to prevent, manage, and reduce the risk factors associated with chronic diseases such as diabetes. The funded programs target childhood obesity, diabetes, heart disease, and stroke in 32 states, including Michigan [
16]. Among other initiatives, the funding in Michigan is allocated to programs such as the Nutrition and Physical Activity Self-Assessment for Child Care, which seeks to increase consumption of healthy food and participation in physical activity among children from birth to the age of 5 years old [
17]. Another Michigan-funded program aiming to promote healthy behaviors among children is led by the non-profit organization Blue Cross Blue Shield of Michigan, which oversees the school-based Building Healthy Communities (BHC) program. The BHC program aims to help children adopt healthy behaviors at a young age, targeting school children from kindergarten to 12th grade [
18]. These programs all offer increasing access to healthy food and promote physical activity and seek to improve physical education through the school curriculum by providing training and equipment in a healthy and supportive environment [
19].
While school-based diabetes prevention programs are being successfully implemented in Michigan, public health resources are not limitless; therefore, it is important that the programs are targeted in the areas corresponding to the most need. New ways to prioritize school-based diabetes prevention for targeted program implementation are critical for the efficient distribution of limited resources [
20]. The aforementioned programs, which are currently being implemented in Michigan, may utilize information from studies that use diabetes-related mortality rate as a proxy for examining the diabetes burden by school district to prioritize areas of need, since data for diabetes risk are not widely available at the school district level. Diabetes mortality has been used as a measure to estimate the diabetes burden because it affects individuals’ functional capacities and quality of life, leading to significant premature mortality [
21]. Previous studies have used diabetes death records to describe the diabetes burden at the school district level [
22,
23]. However, unlike previous studies, this study explored the diabetes burden by examining clusters of diabetes-related mortality rates and the trends exhibited by these clusters over an 8-year period across space and time, through a method known as spatiotemporal analysis.
Spatiotemporal analysis has been used to address a myriad of public health problems, such as traffic accidents among the elderly population in Seoul, Korea [
24]. It has also been used to examine opioid-involved fatalities in Connecticut, United States [
25]. Spatial analyses of diabetes-related mortality have been conducted in several studies in the United States [
26,
27] and China [
28]. A study conducted in the United States by Dwyer-Lindgren et al. [
26] analyzed spatial patterns in the mortality rate of various diseases at the county level, including diabetes, using death records from 1980 to 2014. Similarly, another study conducted in the United States by Kedir & Grigsby-Toussaint [
27] used county-level data to examine clusters of diabetes-related mortality in 7 years of aggregated data (2003 to 2010); however, no temporal analysis was conducted in either study. While these studies have helped to inform our understanding of diabetes-related mortality and sound methodological techniques to examine it, more information is needed on the distribution of diabetes mortality across space and time, especially using emerging hotspot analysis, which enables the examination of different types of clusters such as intensifying, persistent, sporadic and new hot/cold spots to better understand trends over time at the school district level. Therefore, this study was designed to address this gap in analysis by assessing whether diabetes-related deaths occurred at random or clustered in terms of space and time in Michigan for the years spanning from 2007 to 2014. Through spatiotemporal analysis, the findings can better inform state government and local organizations about the diabetes burden by school district, thereby helping direct resources to high-need areas.
RESULTS
During 2007–2014 (
Figure 1), the age-adjusted mortality rates with diabetes as the associated cause slightly declined among school districts from 43.1 to 40.6 per 100,000 population, and those of diabetes as the underlying cause remained steady at approximately 20 per 100,000 population, resulting in a slight decline of diabetes-related mortality rates from 64.0 to 61.6 per 100,000 population. Overall, the rate of diabetes as the associated cause of death among school districts was 2 times higher than that for diabetes as the underlying cause of death.
The mortality data from 2007 to 2014 identified clustering of diabetes mortality by school district using the local Moran’s I analysis. Spatial autocorrelation was noted for diabetes-related mortality rate in all years studied, with Moran’s I values ranging from 0.02 in 2009 to 0.19 in 2014, all of which were statistically significant (p<0.01 for each year).
The local Moran’s I analysis identified spatial cluster and spatial outlier locations [
38]. The spatial clusters were referred to as high-high or low-low clusters, where school districts with high diabetes-related mortality rates were surrounded by school districts with similar high rates, and school districts with low rates were surrounded by school districts with rates that were similarly low [
38]. Spatial outliers were defined as high-low locations, where high rates were surrounded by low rates, or low-high locations, where low rates were surrounded by high rates [
38]. As the aim of the study was to provide information for targeting school-based diabetes prevention programs, the results focused on high-high locations.
The series of maps displayed in
Figure 2, created for the years 2007–2014 and analyzed in the cluster analysis, demonstrated spatial clusters of high-high locations (in red on the map) of diabetes-related mortality rate by school district in the West and East Central, Southwest, and Southeastern Lower Peninsula of Michigan. No high-high clusters were identified in the Upper Peninsula of Michigan in all years of the study period. The number of school districts in the high-high clusters decreased from 40 school districts in 2007 to 21 school districts in 2014. Not all school districts were consistently in high-high clusters from 2007 to 2014. A visual examination indicated that there were temporal changes of high-high clusters school districts from 2007 to 2014. For example, the Flint City and Beecher Community school districts were consistently in high-high clusters without any single interruption within the 8 years of the study, while the Detroit City, Westwood Community, and Kalamazoo public school districts were in high-high clusters with at least 2 years of interruption. To be able to examine further the changes in diabetes mortality over time, the Space Time Pattern Mining tool in ArcGIS Pro was used.
Spatiotemporal results
The results from the Space Time Pattern Mining analysis included all death data across the 8-year study period for school districts in Michigan. The space time cube results showed that the overall data trend was similar for both data for all-cause deaths and diabetes-related deaths, indicating a decreasing trend over time with negative cluster statistic values of −1.39 and −0.99, respectively. However, both p-values indicated that the results were not statistically significant, 0.16 and 0.32, respectively. Although the overall trends were not statistically significant, the trend for each grid indicated that it may be significant and was further examined in the emerging hotspot analysis.
The results from the emerging hotspot analysis did not detect any cold spots (
Figures 3 and
4). This is consistent with the mortality count dataset used, which had a high number of grids with zero values, indicating no mortality during the study period. The non-zero bins were lower in diabetes-related deaths (23.28%) than in all-cause deaths (46.32%), because the proportion of diabetes-related deaths was only 13.58% of total matched-addresses of all-cause deaths included in the study (
Table 1). Both the datasets for all-cause deaths and diabetes-related deaths showed that the hotspots were mostly located in the West, East Central, Southwest, and Southeastern region of the Lower Peninsula school districts of Michigan, which is consistent with the result from the cluster analysis using local Moran’s I, and a few in the East Upper Peninsula (
Figures 3 and
4).
Table 1 shows that the results identified 7 types of hotspots: new, consecutive, intensifying, persistent, diminishing, sporadic, and historical hotspots with persistent and intensifying hotspots as the 2 highest hotspots for both datasets.
After layering the grids of diabetes-related deaths over the grids of all deaths, the results showed that there were school district locations in diabetes-related deaths with trends that were not consistent with the trend in overall deaths (
Table 2), which is displayed in
Figure 5.
The map in
Figure 5 shows the differences in the hotspots between diabetes-related deaths and all-cause deaths in Michigan school districts. These hotspots are considered as the grids of concern where diabetes-related deaths increased while all-cause deaths did not. For example, the grids in green indicated that the overall deaths were either sporadic or had no cluster pattern existing over time, but diabetes-related deaths were statistically significant, increasing in 2014, making those locations new hotspots for diabetes-related deaths. Similarly, the grids in yellow indicated that all-cause deaths were in persistent, sporadic, or diminishing hotspots, but diabetes-related clusters were increasing over time.
The grids of concern were found in Genesee, Ingham, Jackson, Kent, Macomb, Saginaw, Oakland, Ottawa, and Wayne Counties. To determine which school districts to target for prevention efforts, the concerned grids were layered over the school district boundaries. The layering identified that the Lansing, Royal Oak, Flint City, Berkley, Detroit City, East Lansing, South Lake and Holt public school districts had at least 4 or more grids intersecting with the school district boundaries. These are the school districts that may be prioritized for school-based diabetes prevention program efforts.
DISCUSSION
Age-adjusted diabetes-related mortality rates slightly declined among Michigan school districts between 2007 and 2014 from 64.0 to 61.6 per 100,000 population. This slight decline was consistent with 2014 death reports released by the CDC indicating a significant decrease in diabetes mortality rates in Michigan [
42]. Although the report only included diabetes as the underlying cause of death, this may be related to the decrease in all diabetes deaths including diabetes as the associated cause. However, regardless of the year studied, diabetes-related mortality rates varied greatly by school district in this study, with similar statistically significant geographic cluster patterns in every year from 2007 to 2014 as determined by study results. That is, individuals residing in neighboring school districts were more similar in terms of their likelihood to die due to diabetes than individuals living farther away. This characteristic is consistent with Tobler [
43] first law of geography that neighboring locations are likely to have similar characteristics. This shows that diabetes-related mortality is not a random phenomenon. This analysis is particularly useful to identify which communities, school districts have remarkably high rates or low rates of diabetes-related mortality in the context of their neighbors; this method has been used in a variety of public health research studies. For example, these analyses were used for determining spatial clusters of diabetes prevalence and its associated risk factors [
44], clusters of BMI among adults with diabetes [
45], and geographic disparities in the diabetes-related mortality rate across counties in the United States [
27]. This information can be directly used for targeting communities with diabetes prevention and intervention efforts.
Although the local Moran’s I was used to analyze spatial patterns of diabetes-related mortality that existed in the data, showing high-high clusters in the East and West Central, Southeast, and Southwest Lower Peninsula school districts for yearly assessments from 2007 to 2014, the analysis did not indicate whether the mortality rates increased or decreased in those clusters. Thus, it is important to examine both spatial and temporal patterns in diabetes-related deaths in Michigan school districts by examining temporal changes in the 8-year data using the Space Time Pattern Mining tool that take advantage of not only the spatial aspect of the data, but also the temporal aspect of the data.
The emerging hotspot analysis produced new, consecutive, intensifying, persistent, diminishing, sporadic, and historical hotspots using the 2007–2014 dataset, although the space time cube did not contain significant trend results, presumably due to the length of time interval used in the analysis. If data were available for 10 or more years, a temporal bin size of 1 year could be utilized, which may reduce the time interval and improve the results. These hotspot locations are proximal to the location identified in the yearly local Moran’s I result in
Figure 2, and may suggest quite similar hotspots with the exception of a few locations in school districts in the Upper Peninsula of Michigan (
Figures 3 and
4). The hotspots may not exactly overlap because of the different measures and geographical level of analysis used between the local Moran’s I and emerging hotspot analyses. In the local Moran’s I analysis, the geographical level of the analysis of school district and mortality rate were used, while in the emerging hotspot analysis, cubes of 3 km
2 and the mortality count were used; this resulted in the modifiable areal unit problem (MAUP) phenomenon, introduced by Openshaw [
46], in which different results are obtained from the analysis of the same data that are aggregated into different groups of geographic levels. However, Manley et al. [
47] established that MAUP is not a problem; instead, it is a resource. Data aggregated at different geographic levels can help us identify processes working at different levels [
47]. It is obvious that it is not possible to define an ideal single geographic level that captures all the processes for all variables, especially for disease distribution. This study used the geographic level of the school district for local Moran’s I and identified high-high clusters of diabetes-related mortality rates, mostly in school districts located in highly dense, urbanized areas (no clusters were found in the Upper Peninsula of Michigan). When using the smaller geographic level (3 km
2) instead of the school district, the study identified the spatiotemporal change of a few areas in the Upper Peninsula (
Figures 3 and
4). This is a strategy that a study by Matisziw et al. [
48] suggested is useful, stating that downscaling of the spatial structure of polygonal units could offer valuable information on the spatial pattern of disease.
From all the hotspots identified in the emerging hotspot analysis, the focus was given to intensifying and new hotspots because these areas indicated that an increased number of diabetes-related deaths occurred from 2007 to 2014. After comparing the hotspots between all-cause deaths and diabetes-related deaths, grids were specified that were in intensifying and new diabetes hotspots, while all-cause deaths showed persistent or diminishing hotspots. By intersecting the school district boundaries and the grids of intensifying and new hotspots, the study revealed the school districts that had an increasing number of diabetes-related deaths compared to others. These school districts included Lansing, East Lansing, and Holt school districts, which are located in between Ingham, Eaton and Clinton Counties; the Flint City school district in Genesee County, as well as the Berkley, Detroit City, Royal Oak, and South Lake school districts in Oakland, Macomb, and Wayne Counties. The results of this study indicate that public health officials may want to prioritize these school districts for targeted school-based diabetes prevention programs.
The methodology used in this study was able to examine the spatial and temporal clustering of diabetes-related mortality in the state of Michigan. This analysis added upon previous work by investigating multiple years of diabetes-related mortality data at the county level [
26,
27], while combining spatial and temporal components, rather than studying them separately. The results provide information about the spatiotemporal trends of diabetes-related deaths that can be used to make better decisions in implementing diabetes prevention programs as well as to evaluate the effectiveness of the implemented programs over time.
The limitations of this study include the inability of the Space Time Pattern Mining tool to easily support the current results as it requires a longer time period than was available in the current dataset to be set in an annual bin size. Additionally, the geocoding process may exclude unmatched addresses, lowering the number of deaths included in the analysis. The study also did not take into account other potential confounders in determining variation in the diabetes mortality rate by school district, such as socioeconomic status, healthcare facilities, and educational level. Finally, the migration of individuals in and out of school districts was also not accounted for in this study.
The combined spatial and temporal analyses used in this study demonstrated a novel method for using available death record data to target school-based diabetes prevention programs at the statewide level, particularly in Michigan. Future work should explore the application of these techniques to a dataset with a longer time span. In addition, the methods used in this study could be adapted to other public health problems and target populations by using different secondary health databases.
In conclusion, this study demonstrated the presence of diabetes-related mortality hotspots within the state of Michigan at the school district level. The results indicated that diabetes-related mortality is not a continuous feature statewide, but rather varies across both space and time. Furthermore, the patterns of diabetes-related mortality can vary in quite different ways from that of the full data set encompassing all-cause deaths in the state of Michigan. Understanding spatial and temporal hotspots could further improve our ability to design future diabetes prevention programs that are targeted to high-risk communities to ensure that limited public health resources are allocated where they are needed most, and may even be utilized to evaluate the effectiveness of programs that have already been implemented.