log(Yit) = α + β×Xit + γ×Zi + ftime + (PM2.5)×fspat, Spatial vulnerability of fine particulate matter on U.S. diabetes prevalence Hasanat Alamgir1, Lung-Chang Chien1,2, Hwa-Lung Yu3 1) University of Texas School of Public Health at San Antonio Regional Campus, San Antonio, Texas, USA; 2) University of Texas Health Science Center at San Antonio, Research to Advance Community Health Center, San Antonio, Texas, USA; 3) National Taiwan University, Department of Bioenvironmental Systems Engineering, Taipei, Taiwan We conducted a Bayesian structured additive regression modeling approach to determine whether long-term exposure to PM2.5 is spatially associated with diabetes prevalence after adjusting for the socioeconomic status of the county residents by using the following data sources from 2004-2010: BRFSS, ACS, and the EPA. We also conducted spatial comparisons with low, median-low, median-high, and high levels of PM2.5 concentrations. Vulnerable counties were in the Central, Southeast, and South Regions of the U.S. A similar spatial vulnerable pattern was also presented in the same three regions for low PM2.5 level. A clear cluster of vulnerable counties at median-high PM2.5 level was found in Michigan. This study identifies spatial vulnerability of diabetes prevalence associated with PM2.5, and thereby provides the evidence to establish enhanced surveillance to raise alertness to diabetes vulnerability in areas with low PM2.5 pollution. Recent research supports a link between diabetes and fine particulate matter (≤ 2.5 µg in diameter; PM2.5) in both laboratory and epidemiology studies. Preliminary evidence reveals that areas in the U.S. that have higher PM2.5 concentrations also have higher prevalence of diabetes (Pearson et al., 2010). The purpose of this study was to investigate the spatial variations and geographic disparities of diabetes prevalence due to county-level PM2.5 concentrations in the U.S. Three research questions were addressed: Whether PM2.5 concentration is spatially associated with diabetes prevalence Whether geographic disparities of diabetes prevalence can be quantified and vary with increased level of PM2.5, Whether counties with higher PM2.5 have a higher vulnerability to diabetes than counties with lower PM2.5. Data was collected from the Behavioral Risk Factor Surveillance System (BRFSS), the American Community Survey (ACS), the Small Area Health Insurance Estimates from the U.S. Census, the Environmental Protection Agency’s (EPA) Air Quality System monitoring stations that are located throughout the U.S. This study included 3,109 counties in the 48 contiguous states. Bayesian entropy method (BME) was used to spatially interpolate data in each county as less than 700 counties had data. Moran’s I statistics were applied to measure and test the spatial autocorrelation of diabetes prevalence and PM2.5 concentrations at the county level for each year (see Table 1). The structured additive regression (STAR) model was applied to evaluate the spatiotemporal influence of PM2.5 concentrations and to calculate geographic disparities on the county-level age-adjusted diabetes prevalence.1 A log-linear framework in which the PM2.5 concentration is an interactive term with locations controlled by other linear and nonlinear terms constructed the 1st equation: log(Yit) = α + β×Xit + γ×Zi + ftime + (PM2.5)×fspat, Yit =age-adjusted diabetes prevalence at county i (i = 1, 2, …, 3109) and calendar time t (t = 1, 2, …, 7) Xit =three diabetes risk factors (age-adjusted obesity prevalence, physical inactivity prevalence, and smoking prevalence) Zi =SES factors (male %, at least high school education %, non-Hispanic White %, non-Hispanic Black %, Latino/Hispanic %, median income, health insurance %, five occupational %) ftime = B-spline function with a second-order random walk to consider temporal autoregressive correlations over time fspat =Markov random fields with a conditional autoregressive prior to take into account for spatial autocorrelations PM2.5 was categorized into four groups (low, low-median, high-median, and high levels) by three quartiles (Q1, Q2, and Q3) in each year, and extended Eq.(1) to the 2nd equation: log(Yit) = α + β×Xit + γ×Zi + ftime + fspace.1 + IML×fspat.2 + IMH×fspat.3 + IH×fspat.4, IML coded 1 if median-low level and 0 otherwise IMH coded 1 if median-high level and 0 otherwise IH coded 1 if high level and 0 otherwise Figure 1 shows the average adjusted annual diabetes prevalence and the annual PM2.5 concentrations. Both maps showed a certain of similar variation. The increased percentage of RR calculated by the spatial function in Eq.(1) is visualized in Figure 2, which reveals a non-uniformly distributed influence of PM2.5 concentrations on diabetes in the U.S. As PM2.5 concentrations increased per unit, the increased percentage of RR for diabetes ranged from -5.47% (95% CI = -6.14, -4.77) to 2.34% (95% CI = 2.01, 2.70). A total of 1,323 counties (42.55%) were identified as spatial vulnerability areas. The spatial patterns along with the four varying levels of PM2.5 in conjunction with diabetes prevalence are shown in Figure 3. A clear cluster of significantly elevated RR for diabetes in Michigan state. Eq.(1) suggests county level age-adjusted diabetes prevalence associated with PM2.5 concentrations is about 1.04% (95% CI = 1.01, 1.07) greater or lower than the overall national average. In Eq.(2), the geographic disparities percentage of low PM2.5 level rose to 11.03% (95% CI = 10.64, 11.41) and decreased to 4.12% (95% CI = 3.51, 4.77), 1.31% (95% CI = 0.79, 1.85), and 4.76% (95% CI = 4.13%, 5.40%) in the median-low, the median-high, and the high PM2.5 concentration levels, respectively. The data passed the sensitivity analysis when including ozone measurements. We identified the following: Spatial vulnerability areas were primarily clustered in the Southeast (465 counties), Central (388 counties), and South (354 counties) Regions (see Figure 2) Different spatial patterns were determined along with the increase in PM2.5 levels, suggesting that a low level of PM2.5 concentration was enough to trigger an increased diabetes prevalence Counties in Michigan state conducted a vulnerable clustering to diabetes at the median-high level of PM 2.5 concentration The largest geographic disparities percentage occurred at the low level of PM2.5 concentrations. By knowing that air pollution is one of the major environmental risks associated with a high prevalence of diabetes, communities can reduce the risk and burden of diabetes by making efforts to reduce air pollution levels. Limitations: Type I and Type II diabetes were not distinguished in BRFSS County-level diabetes prevalence data were calculated and published each year by the CDC utilizing a random effect model, and no GIS and time data were used in the computations 2 BRFSS database only interviews people older than 18 years Due to the STAR model, we are unable to include data from land that does not share boundaries with the 48 contiguous states Some estimates of SES factors (i.e. male percentage, non-Hispanic White percentage) are not explainable References 1. Fahrmeir L, Lang S. Bayesian inference for generalized additive mixed models cased on markov random field priors. Journal of the Royal Statistical Society. Series C (Applied Statistics). 2001; 50: 201-20. 2. Caldwell B, Thompson TJ, Boyle JP, Baker L. Bayesian small area estimates of diabetes prevalence by US county, 2005. Journal of Data Science 2010; 8: 171-88. Results Abstract Table 1. Spatial autocorrelations of the annual age-adjusted diabetes prevalence and the annual average of PM2.5 measurement at the county level in the U.S., 2004-2010. Diabetes prevalence PM2.5 Year Moran’s I P-value 2004 0.181 <.0001 0.249 2005 0.186 0.258 2006 0.204 0.253 2007 0.205 0.257 2008 0.188 2009 0.184 0.157 2010 0.174 0.203 Conclusions Figure 1. Maps of spatial distribution of the average adjusted diabetes prevalence per 100,000 population (left) and the average imputed PM2.5 concentration (right) in the U.S. Introduction Methods Figure 3. Spatial maps (upper) shown with the corresponding significance maps (lower) for (a) low level, (b) median-low level, (c) median-high level, and (d) high level of PM2.5 concentration and for diabetes prevalence among 3,109 counties. Counties shaded by white color had a significantly elevated relative risk percentage, black color had a significantly decreased relative risk percentage, and grey counties had a non-significant relative risk percentage. Figure 2. Spatial map (left) of the relative risk percentage due to PM2.5 concentrations shown in correlation with the significance map (right) for diabetes prevalence among 3,109 U.S. counties. Counties shaded by white color had a significantly elevated relative risk percentage, black color had a significantly decreased relative risk percentage, and grey counties had a non-significant relative risk percentage.