Spatial epidemiology of diabetes: Methods and insights

Table 1 Methods for geospatial analysis with examples of applications and findings

Methods	Geospatial techniques	Ref.	Geographic unit and location	Key findings
	Getis-Ord Gi	[92]	West Adelaide, Australia	Spatial distribution of dementia, depression, and type 2 diabetes varied across west Adelaide, respectively. Spatial convergence of the three diseases was identified in two large hot spot clusters
Spatial clustering	Local Moran's I and the Getis-Ord Gi	[93]	Individual level in west Adelaide, Australia	Spatial heterogeneity in type 2 diabetes risk was present across communities, with significant clusters in the central part of the study area
	Moran's I	[31]	District level in India	The prevalence of diagnosed diabetes was substantially higher than that of self-reported diabetes in southern India (7.64% vs 2.38%) Diagnosed diabetes prevalence had positive moderate autocorrelation, and it varied from 10.52% in Goa to 4.89% in Telangana. The diagnosed diabetes prevalence was associated with higher proportion of people with secondary education and above, wealthy and Christian populations
	Moran's index and spatial regression	[94]	District level in Southern India	Spatial variations of high blood glucose (HBG) and very high blood glucose (VHBG) were observed across districts for women aged 15-49 years. District-level prevalence of HBG and VHBG were clustering across southern Indian districts. The HBG and VHBG prevalence were associated with district-level proportion of tobacco use, overweight, obese, and general caste
	Spatial statistic scan	[72]	Individual level in India	Substantial geographic variation in diabetes prevalence in India was found, with a concentrated burden at the southern coastline; Regional tuberculosis endemicity and diabetes spatial distributions showed that there is a lack of consistent geographical overlap between these 2 diseases
	Getis-Ord Gi	[95]	Individual and statistical area level 1 regions in western Adelaide, South Australia	The spatial heterogeneity of obesity, cardiovascular diseases (CVD), and type 2 diabetes was present across communities. Hot spots of these conditions clustered in three locations across western Adelaide. Area-level prevalence of CVD, obesity, and type 2 diabetes were negatively associated with socioeconomic status (SES)
	Global Moran‘s I, Local Moran’s I and spatial regression	[32]	Individual and state level in Nigeria	Geographic clustering of diabetes mellitus (DM) and a DM pocket existed in the southeastern part of Nigeria. Obesity and education attainment were associated with the geographic variations of DM in the country
	Moran's I	[96]	Individual level in the city of Oslo, Norway	Diabetes prevalence clustered on the east side of Oslo. The diabetes prevalence was positively associated with neighborhoods with more fast foods and less healthy food shops and physical exercise facilities
	Spatial scan statistic and non-spatial linear regression	[47]	Administrative health area in the City of Winnipeg, Canada	Substantial clustering and small-area variations in DM prevalence existed in the city of Winnipeg. High rates of DM prevalence were associated with low SES, poor environmental quality and poor lifestyle
Spatial estimation models	Geographically weighted regression	[97]	Country-level of the continent United States	Significant spatial clustering of county-level diabetes prevalence was observed in the United State; the associations between diabetes prevalence and the percentage of poverty and percentage nonwhite population varied across regions in the United States.
	Geographically weighted regression	[98]	Country-level of the continent United States	The relationships between diabetes prevalence and poverty varied as a function of location
	Geographically weighted regression	[99]	Four-digit postal code level in Netherlands	Type 2 DM drug use is positively associated with population ageing, proportion of social welfare/benefits, proportion of low income, and proportion of pensioners. Spatial variabilities existed in these associations. Spatial analysis provided added value in predicting health care use at local level
	Tests of spatial autocorrelation and geographically weighted regression	[51]	Hospital referral regions in the United States	Lower-extremity amputation had spatial variations, with high rates clustered in southern states of the United States
	Spatial regression models	[100]	census-tract level in Chicago	Hypertension prevalence rates for patients were positively associated with areas with high rates of poverty, minority, and disability status. Neighboring tracts with high disease rates were the strongest predictor of cardiovascular-related chronic disease by several orders of magnitude. Diabetes had similar results
	Spatial autoregressive model	[101]	District-level in India	Spatial clustering was present in the burden of diabetes among women. The burden was relatively higher among women from the Southern and Eastern parts of the country. Diabetes was associated with obesity, hypertension, and living in urban areas
Multilevel models	Multilevel models	[102]	Individual level in Northern Netherlands	Individual risk factors at the neighborhood and municipality level explained 67.0% and 71.6% of the regional variations, respectively. Analysis on the smallest spatial scale best captured the regional variance. Individual and neighborhood body mass index (BMI) had significant interaction adjusting for the individual risk profile
	Multilevel negative binomial regression	[103]	Province level in China	Compared with the South, diabetes mortality was higher in the Northwest and Northeast. Diabetes mortality was higher in urbanized areas, with higher mean body mass index, and with higher average temperatures. Diabetes mortality was lower where consumption of alcohol was excessive
	Multilevel logistic regression	[104]	Province level in China	Diabetes prevalence and detection had widespread geographic variations across provinces in China. Adjusted regional diabetes prevalence was higher in the north (12.7%) than in the northeast (8.3%). Adjusted regional diabetes prevalence was higher in urban high socioeconomic circumstances (SEC) (13.1%) than in rural low-SEC counties/districts (8.7%). Adjusted diabetes detection was higher in the north (40.4%) and in urban high-SEC counties (40.8%) than in the southwest (15.6%) and the rural low-SEC counties (20.5%)
	Multilevel poisson regression	[105]	Canton-level in Southeastern France	Prevalence of treated diabetes was significantly higher in the more deprived and population-dense cantons
	Multilevel logistic regression	[106]	Census blocks in Paris, France	Prevalence of type 2 diabetes was higher in neighborhoods with the lowest levels of education attainment. Meanwhile, accounting for geographic variations in participation led to an 18% decrease in the log prevalence for low versus high neighborhood educations
Spatial analysis and GIS mapping	Choropleth mapping and logistic regression	[29]	County-level United States	Identifying a diabetes belt consisting of 644 counties in 15 mostly southern states in the United States People in the diabetes belt were more likely being Non-Hispanic African American, leading a sedentary lifestyle, and being obese
	GIS methodology of spatial join	[107]	Census-tract level in greater Sacramento area United States	Neighborhood SES was a barrier to optimal glucose control, but not associated with low-density lipoprotein control. GIS analysis is useful for disease management programs
	Data aggregation to state-level and region-level	[108]	State-level United States	The spatial variations in the ratios of children with diabetes to pediatric endocrinologists were present: the ratios in Midwest (370: 1), South (335: 1), and West (367: 1) are twice as high as in the Northeast (144: 1). Across states, there is up to a 19-fold difference in the observed ratios of obese children to pediatric endocrinologists
	Data aggregation to district level and GIS mapping	[109]	Tower Hamlets, an inner city district of London, United Kingdom	Hot spots where up to 17.3% of all adults were at high risk of developing type 2 diabetes were identified. Small-area geospatial mapping is feasible for epidemiological and environmental data
	Data aggregation to electoral wards and Regression analysis	[110]	Electoral wards in England	The diabetes prevalence varied across different locations, ranging from 2.4% in Thames Valley to 4% in North East London. The methodology of prevalence estimates is applicable to developing small area prevalence estimates for a range of chronic diseases
	Data aggregation to electoral wards and GIS mapping	[111]	Electoral wards in Greater London	Environmental factors affected diabetes outcomes. The age-adjusted mortality rates in diabetic patients were higher in deprived areas than in prosperous areas
	Data aggregation to climato-geographic and administrative regions of the Ukraine	[112]	Administrative regions in Ukraine	Geographic variations in the insulin-dependent diabetes mellitus (IDDM) were present across various administrative regions of the Ukraine. The prevalence of IDDM varied from 1740 to 3813 patients per 1 million populations across Ukraine, with the west zone having lower prevalence than the average
Bayesian estimation approaches	Bayesian spatial analysis	[113]	Local administrative district level in Bangladesh	People of older age, higher education, better socio-economic condition, higher BMI were more likely to have hypertension and diabetes. Significant regional variations were observed with prevalence for hypertension ranges between 10% and 35% and for diabetes between 6% and 19% while their national prevalence were reported as 24% and 11%, respectively
	Bayesian hierarchical joint spatial analysis	[114]	Electoral wards in the Yorkshire, United Kingdom	Childhood lymphoblastic leukemia and type 1 diabetes varied across geographic locations, clustering in more rural areas
	Bayesian Small Area Estimates	[30]	County-level United States	Diabetes incidence was high in the southeastern United States, the Appalachian region, and in scattered counties throughout the western United States
	Bayesian estimation approach	[115]	Zip code census tract of United States	Significant spatial effects existed in the diabetes prevalence even after adjusting for age, education, ethnicity and known state predictors
Regression accounting for spatial variations	Regression-based β-convergence approach, accounting for spatial autocorrelation	[116]	County-level United States	County-level disparities in diagnosed diabetes prevalence in the United States broadened, while the disparities in diagnosed diabetes incidence narrowed. Demographic, socio-economic characteristics and risk factors of type 2 diabetes were associated with changes in disparities
	Sparse Poisson convolution; sparse Poisson missing-completely-at-random	[117]	County-level; Tract-level United States	The type 1 and type 2 DM incidences in young in United States varied across regions; the type 1 and type 2 DM incidences also differed across small areas within study region. The joint spatial correlation between type 1 DM and type 2 DM was present at the county level, but not at tract level

HBG: High blood glucose; VHBG: Very high blood glucose; SES: Socioeconomic status; DM: Diabetes mellitus; GIS: Geographic Information Systems; BMI: Body mass index; IDDM: Insulin-dependent diabetes mellitus.