Chien, Li, and Staudt: Physical inactivity displays a mediator role in the association of diabetes and poverty: A spatiotemporal analysis

Physical inactivity displays a mediator role in the association of diabetes and poverty: A spatiotemporal analysis


Physical inactivity is one of the risk factors of diabetes. In addition, physical inactivity is attributed to urbanization-related factors, such as poverty, which is also one of the risk factors of diabetes. We hypothesized that physical inactivity is a mediator in the association between diabetes and poverty, and that spatial heterogeneity exists in these relationships. This study adopted a spatiotemporal modelling approach to conduct this mediator analysis. From 2004-2011, data were collected at the county level in 48 contiguous states (with a total of 3,109 counties) from the Behavioral Risk Factor Surveillance System (BRFSS) and American Community Survey. Poverty percentage significantly affected physical inactivity prevalence and diabetes prevalence in two separate models. Using a model with both physical inactivity and poverty percentages as independent variables, we verified that physical inactivity prevalence is a significant mediator. In this model, physical inactivity prevalence resulted in a significant positive association with diabetes prevalence, and the influence of poverty percentage on diabetes prevalence was significantly reduced (P=0.0009). An advanced spatiotemporal analysis revealed that 32.65% of counties having a significant positive association between diabetes prevalence and physical inactivity prevalence also had a significant positive association between physical inactivity prevalence and poverty percentage. Those counties were also likely located in the South and Southeast of USA. In summary, the findings of this study demonstrate the mediating effect of physical inactivity between diabetes and poverty. When implementing diabetes prevention in communities with higher poverty, appropriate strategies to reduce the cost burden of physical activity programmes should be considered.


Diabetes prevalence has consistently increased in recent years, and physical inactivity has proven to be a significant risk factor for this disease. Physical inactivity should be seen as a health misbehaviour, which can seriously affect human health. According to a World Health Organization (WHO) report, insufficient physical activity is a key risk factor for non-communicable diseases, including diabetes (WHO, 2015a). More than 23% of the world’s adults (aged 18 and over) do not have a sufficient level of physical activity. The discouragement of physical activity participation can be attributed to several environmental factors, like high crime rate, high-density traffic, low air quality and the lack of sports facilities (WHO, 2015b). These urbanisation-related factors may stem from poverty, which has been named a leading cause of Type 2 diabetes. Risk factors for diabetes include both poverty and physical inactivity. Physical inactivity is a significant risk factor for diabetes (Booth and Hawley, 2015). In addition, lack of physical activity has become a global problem associated with diabetes (Oggioni et al., 2014). Promoting physical activity is usually regarded as the primary prevention of diabetes (Avery et al., 2012); however, there are challenges because physical inactivity has a complicated relationship with other health behaviours and policy and environmental change (Kohl et al., 2012). Furthermore, disparities of physical activity by sociodemographic characteristics also exist in the USA (Ham and Ainsworth, 2009).

Additionally, the association between physical inactivity and poverty has been demonstrated (Riste et al., 2001). Results from a study conducted in a city in southern California indicated that residents in high poverty areas had less exercise. Instead, they spent more time watching electronic media as compared to those from low poverty neighbourhoods (Cohen et al., 2012). High poverty neighbourhoods were also less pleasurable for outdoor physical activity because the areas were less safe and the sidewalks and streets not maintained well (Franzini et al., 2010). Another study found the prevalence of leisure time inactivity was highest among study participants from the lower social classes (Marshall et al., 2007). In USA, physical inactivity dropped from 42.6% among study participants from the lowest income category to 15.1% among study participants from the highest income category (Pratt et al., 1999).

In USA, diabetes has found to show spatial variation in incidence and prevalence at the county-level. For example, in Medicaid recipient adults in South Carolina, living in a county with high levels of poverty was positively associated with diabetes prevalence (Stewart et al., 2011). A spatial analysis study showed that physical inactivity prevalence has a significant inverse relationship with diabetes prevalence in north-western USA, and a significant positive relationship between poverty percentage and diabetes prevalence was more likely observed in the Southwest and Northeast of the country (Hipp and Chalise, 2015). The positive association between poverty and diabetes is normally analysed using a linear model; however, after taking geographical location into account, the diabetes-poverty association can fluctuate spatially. For example, a spatial analysis applied a geographic weighted regression model to explore the macro-level spatial non-stationary of the diabetes-poverty relationship, and this study provided evidence that poverty may not always be positively associated with diabetes, especially in the states of Texas, Kansas and Washington (Siordia et al., 2012). Hence, it is plausible that geographic heterogeneity may exist, but further investigation is needed to determine the association between diabetes and poverty.

Certain variables may mediate the association between diabetes and the other risk factors. For instance, psychosocial factors, unhealthy behaviours and obesity may mediate the association between socioeconomic status and diabetes incidence (Demakakos et al., 2012). Usual source of care may also mediate the association between health insurance and diabetes (Hastings and Hawkins, 2009). Since studies have shown physical inactivity is a risk factor of diabetes, and poverty is a risk factor of both diabetes and physical inactivity, we hypothesized whether physical inactivity is actually a mediator in the association between poverty and diabetes rather than a purely risk factor of diabetes. This relationship has not yet been fully investigated.

This study retrospectively collected prevalence and poverty data in USA at the county level from a nationwide survey database and aimed to investigate whether physical inactivity mediates the association between diabetes and poverty. We applied an advanced spatiotemporal analysis to consider spatial variation and geographic heterogeneity among our data. Additionally, we aimed to identify physical inactivity as a significant modifier in counties vulnerable to diabetes and poverty. By using geographic information visualization, this study distinguished spatial clusters in the study areas.

Materials and Methods

Study area

We used county as the geographic unit in this research because it is the smallest geographic data collected by the Behavioral Risk Factor Surveillance System (BRFSS). A total of 3,109 counties in the 48 contiguous states in USA were included in the study area, and each county had at least one adjacent county, which was defined by sharing any part of a boundary with another county.

Data sources

From 2004-2011, age-race-gender adjusted diabetes prevalence and physical inactivity prevalence were gathered at the county level from official website of the U.S. Centers for Disease Control and Prevention ( We also used adjusted obesity prevalence for the sensitivity analysis. Poverty percentage and other important socioeconomic predictors were selected from the database of American Community Survey (ACS) maintained by the U.S. Census Bureau since 2005. Currently, this decennial survey provides the most detailed information related to population distribution, social structure, economic condition, and housing status for each county (US Census Bureau, 2014). Every year the U.S. Census Bureau publishes 1-year, 3-year and 5-year ACS data. This study used 5- year ACS estimates to guarantee complete data. Two 5-year estimates were adopted, while the first 5-year estimate (2005-2009) was applied to the data before 2009, and the second (2008-2012) was applied to the data after 2009. All ACS data can be downloaded from American Fact Finder (


The diabetes prevalence were calculated from respondents who were diagnosed with diabetes. A respondent was considered diabetic if he/she answered yes to the question: Has a doctor, nurse, or other health professional ever told you that you have diabetes? Note that females who had gestational diabetes were not included in the calculation of diabetes prevalence. The physical inactivity prevalence was calculated using respondents who reported doing no physical activity or exercise in the past 30 days. These respondents were identified as those who answered no to the question: During the past month, other than your regular job, did you participate in any physical activities or exercises such as running, calisthenics, golf, gardening, or walking for exercise? Those county- level prevalence data were estimated by applying Bayesian multilevel modelling techniques on individual data from the largest nationwide telephone survey - the BRFSS (CDC, 2016). Poverty percentage data was calculated using the percentage of families and people whose income during the past 12 months was below the poverty level. We selected several confounding variables, which included gender (male percentage), age (elder percentage and median age), race (Hispanic percentage, non-Hispanic White percentage and non-Hispanic Black percentage), percentage of at least high school education, health insurance coverage percentage and occupation percentages (management/business/science/art, service, sales/office, natural resources/construction/maintenance and production/transportation/material moving). Smoking prevalence was also considered a confounding variable. Smoking prevalence was calculated as the prevalence of current total cigarette smoking in nondaily and daily current smokers, which were determined by the two questions in the BRFSS: Have you smoked at least 100 cigarettes in your entire life and Do you now smoke cigarettes every day, some days or not at all?, respectively. The smoking prevalence data at the county level were estimated by the small area estimation approach, and they were published for public-use (Dwyer-Lindgren et al., 2014). Finally, for sensitivity analysis, we included obesity prevalence, which was identified from the same data source as diabetes and physical inactivity, and respondents were considered obese if their body mass index was 30 or greater (CDC, 2016).

Statistical analysis

In order to test whether physical inactivity is a mediator in the association between diabetes and poverty among counties, the following three models were built:

Model I: DM=α1 + β1.POV + γ1.X + f(time) + fspat(c) Model II: PHY=α2 + β2.POV + γ2.X + f(time) + fspat(c) Model III: DM=α3 + β31.PHY + β32.POV + γ3.X + f(time) + fspat(c)

where variables DM, PHY and POV are adjusted diabetes prevalence, adjusted physical inactivity prevalence and poverty percentage at the county level, respectively. All 14 confounding variables were denoted by a vector X. Fixed intercepts and slopes were denoted by (α1, α2, α3) and (β1, β2, β31, β32, γ1, γ2, γ3), which were estimated by the structured additive regression (STAR) modelling approach with a fully Bayesian method using a Markov chain Monte Carlo simulation technique (Brezger and Lang, 2006). Variances of unknown smoothing parameters were estimated with priors in terms of an inverse Gamma distribution with two hyperparameters (a, b)=(0.001, 0.001) by default. Each model took temporal autoregressive correlation and spatial autocorrelation (i.e. the correlation of a variable with itself through space) into account by adding f(time) with a B-spline function for calendar time and fspat(c) with a spatial function for locations c. In particular, the spatial function was the Markov random fields (Kindermann and Snell, 1980) to analyse boundary data and to emphasize the relationship of adjacent areas. In detail, we defined a neighbourhood as an adjacent county c’ sharing the same boundaries of another county c. Each county had a neighbourhood set Ωc with a total number of neighbourhoods Nc≥1, which means that each county should have at least one neighbourhood. The Markov random fields follow a normal distribution with mean gh-12-2-528-e001.jpg and variance σ2 c/Nc, where fspat (C ’) is the spatial effect of an adjacent county c’, and σ2 c the original variance of county c. We interpreted the coefficient β32 as the direct effect, and the product of β1 and β32 as the indirect effect of a unit increase in poverty prevalence with respect to the mediator (i.e., physical inactivity prevalence). The identified criterion of regarding physical inactivity as a mediator relies on the four linear estimates (β1, β2, β31, β32), where β1, β2, and β31 are statistically significant, and β32 is smaller in absolute value than β1. The significance of βs was determined by the 95% credible interval (CI) of an estimate strictly greater than 0. We eventually applied the Sobel test to compute a P value to identify the significance of the mediating effect (Sobel, 1982; Baron and Kenny, 1986).

For the purpose of investigating spatial clusters of significant association between diabetes and physical inactivity caused by poverty, an advanced spatiotemporal analysis was carried out by applying the following two models:

Model IV: DM=α4 + PHY.fspat(c) + γ4.X + f(time) Model V: PHY=α5 + POV.fspat(c) + γ5.X + f(time)

where physical inactivity prevalence and poverty percentage were no longer linear terms but interacted with a spatial function to represent county-specific influence on dependent variables. The two interaction terms are equivalent to spatially weighted county-specific random slopes for physical inactivity and poverty percentage. Counties having a 95% CI of spatial estimate strictly greater than 0 in the spatial function fspat(c) in Model IV and Model V were regarded as clustering areas vulnerable to diabetes caused by physical inactivity, which can be attributed to a high-poverty percentage. We also accomplished a model diagnostic in autocorrelation and sampling path. Finally, a sensitivity analysis was conducted to test whether our findings were robust by choosing different hyperparameters.

Data management was done by SAS v9.3 (SAS Institute Inc., Cary, NC, USA). Spatiotemporal analysis was accomplished by BayesX v3.0.2 (Brezger et al., 2005). Results of spatial functions were visualized on maps by using ArcGIS v10 (ESRI Inc., Redlands, CA, USA).


On average, each county had 5.64 adjacent counties [standard deviation (SD)=1.36], and Washoe County in Nevada had the most adjacent counties (n=13). Figure 1 shows a clear pattern of the geographic distribution of the average adjusted diabetes prevalence, which ranged from 3.90% to 16.03%. No county had an average of adjusted physical inactivity prevalence less than 10%. Higher averages of adjusted diabetes prevalence were more likely clustered in the southeastern counties, where there were also higher averages of adjusted physical inactivity prevalence. The maximum average of poverty percentage was 43.70%, and a few high poverty clusters were found in southeastern and southern USA.

The mediator analysis shows that physical inactivity prevalence might be a mediator to diabetes prevalence. In Model I, the poverty percentage was significantly associated with diabetes prevalence. For each 1% increase of poverty percentage, there was a 0.013% (95% CI=0.008, 0.020) increase in diabetes prevalence. In Model II, the poverty percentage was also significantly related to physical inactivity prevalence. For each 1% increase in poverty percentage, there was a 0.039% (95% CI=0.018, 0.060) increase in physical inactivity prevalence. In Model III, the prevalence of physical inactivity still had a significant impact on diabetes prevalence (estimated coefficient=0.018; 95% CI=0.014, 0.022), while the estimated coefficient of poverty percentage reduced to 0.012 (95% CI=0.006, 0.018) from the original estimate in Model I. The Sobel test verified a significant mediating effect (P=0.0009) in terms of physical inactivity prevalence.

The advanced spatiotemporal analysis from Model IV demonstrates a significant spatial variation of diabetes prevalence caused by physical inactivity prevalence. Table 1 shows that 62.11% (1,931/3,109) of the counties had a significant increase in diabetes prevalence when physical inactivity prevalence increased. Harris County, Georgia had the greatest increase in diabetes prevalence as every 1% increase in physical inactivity prevalence elevated diabetes prevalence by 0.109% (95% CI=0.094, 0.125). The geographic pattern shown in Figure 2A demonstrates that three-fifths of counties had high diabetes prevalence caused by high physical inactivity prevalence, especially in counties located in the Southeast and Northeast of the country. Table 1 also reveals that 39.56% (1,230/3,109) of the counties had a significant positive association between physical inactivity prevalence and poverty percentage. The greatest impact of poverty on physical inactivity was found in Fairfox City, Virginia (estimated coefficient=0.937; 95% CI=0.596, 1.282). Figure 2B demonstrates the counties with a significant increase in physical inactivity prevalence caused by high poverty percentage were mostly located in southern, southeastern, and mid-northern USA. There is a large overlap in significantly spatial positive between Figure 2A and B. A total of 1,015 counties (32.65%) were significantly vulnerable to diabetes due to physical inactivity as well as vulnerable to physical inactivity due to poverty. Figure 2C shows that most of those overlapping highrisk counties were located in the south and southeast USA. The significance of the two spatial functions in the two models was moderately correlated (Spearman’s correlation=0.39).

Model diagnostics in Figure 3 presents that the autocorrelations along with all lags were within ±0.1, which suggests a good control of autocorrelation in our models. In addition, the sampling path plots verified a stable trace over iterations for main predictors. In the sensitivity analysis, the linear coefficients of physical inactivity prevalence and poverty percentage in Model I, II & III were robust (Table 2). We also confirm that different hyper-parameters did not affect our main findings in the spatial functions; shown in Figure 4.


Unlike diabetes risk factors, such as gender, family history and ethnicity, which cannot be changed, the risk factor of insufficient physical activity can be reversed. Other factors may contribute to a decrease in physical activity. This study hypothesised that poverty may be a plausible reason that people are not able to participate in physical activity. By collecting governmental data to compose large scale measurements of diabetes, physical inactivity, and poverty, this study found the following results: i) Physical inactivity is a significant mediator in the association between diabetes and poverty; ii) People living in south-eastern and north-eastern USA are most likely to have developed diabetes due to physical inactivity; iii) Almost two-fifths of all counties in USA had significantly higher physical inactivity prevalence due to high poverty percentage; iv) A large proportion of counties in most south or south-eastern states had significant associations with diabetes, physical inactivity and poverty. By using data with multiple years, our findings showed many counties to be at high risk for diabetes due to physical inactivity in USA. Within 15 southern states consisting of 644 counties, there is a concentrated area of high diabetes prevalence known as the diabetes belt. One study reported counties not located in the diabetes belt had a higher correlation between physical activity and diabetes (Barker et al., 2011). This conflicting evidence is possibly caused by the statistical analysis being applied in diabetes belt and non-diabetes belt counties separately. In 2007, an advanced investigation applied a spatial cluster analysis to divide USA counties into five clusters in terms of age-adjusted diabetes prevalence (Shrestha et al., 2012). This study showed that physical inactivity may also significantly impact some counties out of the diabetes belt. We explored the spatial association of diabetes prevalence, in terms of poverty percentage and physical inactivity percentage, and we identified more high-risk counties than the 644 included in diabetes belt.

Applying a spatial model in our study shows the geographic variation of influential areas. Even though this application is rarely used in diabetes and poverty research, we found a few previous studies adopting spatial methods or geographic information tools. For example, a study used a geographically weighted regression to explore the diabetes-poverty macro-level statistical relationship at the county level in USA (Siordia et al., 2012). The geographical association between poverty prevalence and diabetes prevalence was significant even though physical inactivity was not included in the study. Additionally, a study in 82 Mississippi counties adopted a spatial lag model to investigate disparities of a low-density lipoprotein cholesterol test, resulting in a negative association with poverty among the elderly with diabetes (Sharma, 2014). In recent years, using geographic information systems has been suggested to target diabetes-related public health efforts. A study with 83 counties in Michigan developed study maps to highlight areas with higher diabetes and poverty rates for interventions targeting lowincome minority populations with diabetes, and then used advanced statistical analysis to verify the significance (Curtis et al., 2013). The STAR model has been applied in a diabetes study previously (Chien et al., 2015), while this is the first time to have it in a mediator analysis. The advanced spatial analysis used in Model IV & V showed almost one-third counties were vulnerable to diabetes because of physical inactivity and also were vulnerable to physical inactivity because of poverty. More research is needed to establish a new research strategy for investigating whether physical inactivity displays a mediator role in the association between diabetes and poverty in those counties.

Traditional diabetes research usually uses both poverty and physical inactivity as confounding variables (Jiang and Pearlman, 2013; Adeniyi et al., 2015; Grundmann et al., 2014). While poverty is a proven risk factor for physical inactivity, this study applied a mediator analysis to evaluate the direct effect between poverty and diabetes prevalence from the parameter of poverty (i.e., β32) in Model III, and we found that an indirect (mediated) effect also existed from the two parameters of poverty (i.e., β1 and β2) in Model I & II. Note that the mediated effect should be explained conservatively because it was conducted from population data. Future research should investigate the mediated effect of physical inactivity between poverty and diabetes using individual data.

Our study regarded poverty as a predictor of both diabetes and physical inactivity, and physical inactivity was identified as a modifier in this relationship. Health interventions should consider strategies to reduce financial burden when promoting physical activity programs for diabetes prevention, especially in lowincome communities. Poverty is associated with adverse health behaviours and insufficient medical treatment (Do and Finch, 2008). However, solving individual or household-level financial problems in a short time is usually difficult; therefore, managing health care for people living in poverty becomes a difficult task. For example, without enough stimulations or incentives, people living in poverty may not be willing or have time to participate in physical activity programs as they may have to spend their time trying to earn money to pay for living expenses (Borodulin et al., 2015). Poverty is also an obstacle of retrieving health-related information. People living in poverty may have less access to resources that encourage physical activity. Therefore, when establishing diabetes and physical activity interventions, providing access to facilities, parks, clubs, walking trials, and other environments that promote exercise is vital.

This study has some limitations. First, diabetes prevalence and physical inactivity prevalence data were calculated from the BRFSS database, where respondents are 18 years old and older. The lack of children’s diabetes data may cause an underestimation in our results. Second, the BRFSS survey does not distinguish between Type I and Type II diabetes mellitus. Third, only data from the 48 contiguous states were included. Data from Hawaii, Alaska, or Puerto Rico are not included in analyses because they are not adjacent to any of the 3,109 counties in the contiguous states, and Markov random fields cannot build up a neighbourhood matrix in those separate areas. However, we believe the exclusion of those areas will not significantly affect our findings because data from the excluded areas do not have a significant spatial correlation with the 48 contiguous states. Fourth, using a spatial interactive term in Model IV and Model V may not provide a better model fitting compared to the linear model, while it is a trade-off for investigating geographic disparities in a mediator analysis. Fifth, we proposed a spatiotemporal approach to carry out a mediator analysis, which originally makes all of the standard assumptions of the general linear model. The sample size of this study is large enough to exempt the normality assumption, but we still suggest future studies with a small sample size pay attention to model assumptions and use the STAR model if necessary.


In conclusion, this study investigated the role physical inactivity plays as mediator in the relationship between diabetes and poverty. National or regional health policies need to promote diabetes prevention by increasing the amount of physical activity in vulnerable populations. While the benefits of physical activity and diabetes prevention are known, people face barriers such as cost or inadequate infrastructure to support physical activity. Changing health behaviour requires personal commitment and reduction of barriers. Hence, when stakeholders take action to increase physical activity at global, regional and local levels, a priority in interventions will be providing low-cost access to physical activity, especially in lower income communities.


Funding for this project was provided by Dr. Lung- Chang Chien’s start-up funds from the University of Nevada, Las Vegas, and Dr. Amanda Staudt was supported by an Occupational Epidemiology Traineeship funded by Grant No. T42OH008421 from the National Institute for Occupational and Environmental Health (NIOSH)/Centers for Disease Control and Prevention (CDC) to the Southwest Center for Occupational and Environmental Health (SWCOEH), a NIOSH Education and Research Center at The University of Texas Health Science Center at Houston School of Public Health.



OV Adeniyi, B Longo-Mbenza, Goon D Ter, 2015. Female sex, poverty and globalization as determinants of obesity among rural South African type 2 diabetics: a cross-sectional study. BMC Public Health 15:298.


LE Barker, KA Kirtland, EW Gregg, LS Geiss, TJ Thompson, 2011. Geographic distribution of diagnosed diabetes in the U.S.: A diabetes Belt. Am J Prev Med 40:434-9.


RM Baron, DA Kenny, 1986. The moderator-mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J Person Soc Psychol 51:1173-82.


FW Booth, JA Hawley, 2015. The erosion of physical activity in Western societies: an economic death march. Diabetologia 58:1730-4.


K Borodulin, N Sipila, O Rahkonen, P Leino-Arjas, L Kestila, P Jousilahti, R Prattala, 2016. Socio-demographic and behavioral variation in barriers to leisure-time physical activity. Scand J Publ Health 44:62-9.


A Brezger, T Kneib, S Lang, 2005. BayesX: analysing Bayesian structured additive regression models. J Stat Software 14:1-22.


A Brezger, S Lang, 2006. Generalized structured additive regression based on Bayesian P-splines. Computat Stat Data Anal 50:967-91.


CDC, 2016. Methods and References for County-Level Estimates and Ranks and StateLevel Modeled Estimates. Available from:


LC Chien, H Alamgir, HL Yu, 2015. Spatial vulnerability of fine particulate matter relative to the prevalence of diabetes in the United States. Sci Total Environ 508:136-44.


DA Cohen, B Han, KP Derose, S Williamson, T Marsh, J Rudick, TL McKenzie, 2012. Neighborhood poverty, park use, and park-based physical activity in a Southern California city. Soc Sci Med 75:2317-25.


CJ Crespo, BE Ainsworth, SJ Keteyian, GW Heath, E Smit, 1999. Prevalence of physical inactivity and its relation to social class in U.S. adults: results from the Third National Health and Nutrition Examination Survey, 1988-1994. Med Sci Sports Exerc 31:1821-7.


AB Curtis, C Kothari, R Paul, E Connors, 2013. Using GIS and secondary data to target diabetes-related public health efforts. Public Health Rep 128:212-20.


P Demakakos, M Marmot, A Steptoe, 2012. Socioeconomic position and the incidence of type 2 diabetes: the ELSA study. Euro J Epidemiol 27:367-78.


DP Do, BK Finch, 2008. The link between neighborhood poverty and health: context or composition? Am J Epidemiol 168:611-9.


L Dwyer-Lindgren, AH Mokdad, T Srebotnjak, AD Flaxman, G Hansen, CJL Murray, 2014. Cigarette smoking prevalence in US counties: 1996-2012. Popul Health Metrics 12:5.


L Franzini, W Taylor, MN Elliot, P Cuccaro, SR Tortolero, MJ Gilliand, J Grunbaum, MA Schuster, 2010. Neighborhood characteristics favorable to outdoor physical activity: Disparities by socioeconomic and racial/ethnic composition. Health Place 16:267-74.


N Grundmann, A Mielck, M Siegel, W Maier, 2014. Area deprivation and the prevalence of type 2 diabetes and obesity: analysis at the municipality level in Germany. BMC Publ Health 14:1264.


SA Ham, BE Ainsworth, 2009. Disparities in data on Healthy People 2010 physical activity objectives collected by accelerometry and self-report. Am J Public Health 10(S1): S263-8.


JF Hastings, J Hawkins, 2009. Health insurance and diabetes among multiracial men: the mediation effects of usual source of care. Am J Men Health 4:207-17.


JA Hipp, N Chalise, 2015. Spatial analysis and correlates of county- level diabetes prevalence, 2009-2010. Prev Chronic Dis 12:E08.


Y Jiang, DN Pearlman, 2013. The link between poverty and type 2 diabetes in Rhode Island. Rhode Island Med J 96:43-7.


R Kindermann, JL Snell, 1980. Markov random fields and their applications. Providence, R.I.: American Mathematical Society.


HW Kohl, CL Craig, EV Lambert, S Inoue, JR Alkandari, G Leetongin, S Kahlmeier, 2012. The pandemic of physical inactivity: global action for public health. Lancet 380:294-305.


SJ Marshall, DA Jones, BE Ainsworth, JP Reis, SS Levy, 2007. Race/ethnicity, social class, and leisure-time physical inactivity. Med Sci Sports Exerc 39:44-51.


C Oggioni, J Lara, JC Wells, K Soroka, M Siervo, 2014. Shifts in population dietary patterns and physical inactivity as determinants of global trends in the prevalence of diabetes: an ecological analysis. Nutr Metab Cardiovasc Dis 24:1105-11.


M Pratt, CA Macera, C Blanton, 1999. Levels of physical activity and inactivity in children and adults in the United States: current evidence and research issues. Med Sci Sports Exerc 31:526-33.


L Riste, F Khan, K Cruickshank, 2001. High prevalence of type 2 diabetes in all ethnic groups, including Europeans, in a British inner city: relative poverty, history, inactivity, or 21st century Europe? Diabetes Care 24:1377-83.


A Sharma, 2014. Spatial analysis of disparities in LDL-C testing for older diabetic adults: A socio-environmental framework focusing on race, poverty, and health access in Mississippi. Appl Geogr 55:248-56.


SS Shrestha, KA Kirtland, TJ Thompson, L Barker, EW Gregg, 2012. Spatial clusters of county-level diagnosed diabetes and associated risk factors in the United States. Open Diabetes J 5:29-37.


C Siordia, J Saenz, SE Tom, 2012. An introduction to macro-level spatial nonstationarity: a geographically weighted regression analysis of diabetes and poverty. Human Geogr 6:5-13.


ME Sobel, 1982. Asymptotic confidence intervals for indirect effects in structural equation models. Sociol Method 13:290-312.


JE Stewart, SE Battersby, Fede A Lopez-De, KC Remington, JW Hardin, 2011. Diabetes and the socioeconomic and built environment: geovisualization of disease prevalence and potential contextual associations using ring maps. Int J Health Geogr 10:18.


Census Bureau US, 2014. American Community Survey Information Guide. Available from:


WHO, 2015a. Noncommunicable diseases. Available from:


WHO, 2015b. Physical inactivity: a global public health problem. Available from:

Figure 1.

Average adjusted diabetes prevalence (A), average adjusted physical inactivity prevalence (B), and average poverty percentage in the USA (C).

Figure 2.

Geographic distribution of spatial interacted estimates. A) Physical inactivity prevalence in upper plot with spatial significance maps (white=significant positive; black=significant negative; grey=non-significance) in lower plot. B) Poverty in upper plot with spatial significance maps (white=significant positive; black=significant negative; grey=non-significance) in lower plot. C) White areas show counties with significant positive findings in both significance maps in (A) and (B).

Figure 3.

Autocorrelation function and sampling path for main predictors in Models I to V.

Figure 4.

Sensitivity analysis for Model IV (A and B) and Model V (C and D) by using different hyper-parameters in spatial functions. By default, we used hyper-parameters a=0.001 and b=0.001.

Table 1.

Summary statistics of spatial estimates in terms of the results of significance derived in Models IV and V.

N % Spatial estimate
Min Q1 Median Q3 Max
Model IV
Significant positive 1931 62.11 0.011 0.029 0.042 0.056 0.109
Significant negative 26-May 16.47 -0.071 -0.042 -0.030 -0.022 -0.012
Non-significance 666 21.42 -0.022 -0.006 0.003 0.009 0.020
        Model V      
Significant positive 1230 39.56 0.047 0.153 0.223 0.316 0.937
Significant negative 738 23.74 -1.110 -0.515 -0.325 -0.199 -0.053
Non-significance 1141 36.7 -0.266 -0.044 0.014 0.062 0.303
Table 2.

Sensitivity analysis for Models I, II and III by using different hyper-parameters in poverty percentage and physical inactivity prevalence. By default, we used hyper-parameters a=0.001 and b=0.001.

(a, b)=(0.001, 0.001) Estimate (95% CI) (a, b)=(0.00001, 0.00001) Estimate (95% CI) (a, b)=(1, 0.05) Estimate (95% CI)
Model I      
POV 0.013 (0.008, 0.020) 0.013 (0.007, 0.018) 0.013 (0.007, 0.019)
Model II      
POV 0.039 (0.018, 0.060) 0.039 (0.019, 0.059) 0.039 (0.018, 0.059)
Model III      
POV 0.012 (0.006, 0.018) 0.012 (0.007, 0.018) 0.012 (0.007, 0.018)
PHY 0.018 (0.014, 0.022) 0.018 (0.014, 0.022) 0.018 (0.014, 0.022)

[i] CI, confidence interval; POV, poverty; PHY, physical (inactivity).

Abstract views:


Article Metrics

Metrics Loading ...

Metrics powered by PLOS ALM

Copyright (c) 2017 Lung-Chang Chien, Xiao Li, Amanda Staudt

Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
© PAGEPress 2008-2018     -     PAGEPress is a registered trademark property of PAGEPress srl, Italy.     -     VAT: IT02125780185