- Email: [email protected]
Document not found! Please try again
Street Connectivity and Obesity Risk: Evidence From Electronic Health Records
RESEARCH ARTICLE
Street Connectivity and Obesity Risk: Evidence From Electronic Health Records Claudia Leonardi, PhD, Neal R. Simonsen, PhD, Qingzhao Yu, PhD, Chi Park, MS, Richard A. Scribner, MD, MPH Introduction: This study aimed to determine the feasibility of using electronic health record (EHR) data from a federally qualified health center (FQHC) to assess the association between street connectivity, a measure of walkability for the local environment, and BMI obtained from EHRs.
Methods: The study included patients who visited Daughters of Charity clinics in 2012–2013. A total of 31,297 patients were eligible, of which 28,307 were geocoded. BMI and sociodemographic information were compiled into a de-identified database. The street connectivity measure was intersection density, calculated as the number of three-way or greater intersections per unit area. Multilevel analyses of BMI, measured on 17,946 patients who were aged Z20 years, not pregnant, had complete sociodemographic information, and a BMI value that was not considered an outlier, were conducted using random intercept models. Results: Overall, on average, patients were aged 44.1 years, had a BMI of 30.2, and were mainly non-Hispanic black (59.4%). An inverse association between BMI and intersection density was observed in multilevel models controlling for age, gender, race, and marital status. Tests for multiple interactions were conducted and a significant interaction between race and intersection density indicated the decrease in BMI was strongest for non-Hispanic whites (decreased by 2) compared with blacks or Hispanics (decreased by 0.6) (p¼0.0121). Conclusions: EHRs were successfully used to assess the relationship between street connectivity and BMI in a multilevel framework. Increasing street connectivity levels measured as intersection density were inversely associated with directly measured BMI obtained from EHRs, demonstrating the feasibility of the approach. Am J Prev Med 2017;52(1S1):S40–S47. & 2016 American Journal of Preventive Medicine. Published by Elsevier Inc. All rights reserved.
INTRODUCTION
D
espite growing awareness of the negative health impact of poor diet, physical inactivity, and excess weight, the prevalence of obesity has increased dramatically in the U.S.1,2 During the last 3 decades, the U.S. obesity rate has doubled in adults and tripled in children and adolescents.1 The consensus among public health experts is that human genetic changes are not responsible for the rapid rise in obesity and the explanation must lie in social determinants resulting from environmental and policy changes.3–6 Although there is consensus on the general role of social, physical activity, and food environments in contributing to the epidemic, there is little in the way S40 Am J Prev Med 2017;52(1S1):S40–S47
of consensus on the relative contribution of any one social determinant (e.g., stressful environment, food deserts, concentrated disadvantage), let alone the relative contribution of any specific neighborhood or community context (e.g., fast food outlet density, walkability).7–10 It From the School of Public Health, Louisiana Cancer Research Center, Louisiana State University, New Orleans, Louisiana Address correspondence to: Richard A. Scribner, MD, MPH, School of Public Health, Louisiana Cancer Research Center, Louisiana State University, 1700 Tulane Avenue, Suite 907, New Orleans LA 70112. E-mail: [email protected]. This article is part of a supplement issue titled Social Determinants of Health: An Approach to Health Disparities Research. 0749-3797/$36.00 http://dx.doi.org/10.1016/j.amepre.2016.09.029
& 2016 American Journal of Preventive Medicine. Published by Elsevier Inc. All rights reserved.
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
has been recognized that the lack of data measured at a scale sufficient to assess the role of the neighborhood environment is a real limitation.11 For example, the two main U.S. public health surveillance systems, with a sufficient number of participants to assess the role of the environment, are limited in their capacity for assessing neighborhood risk. The Behavioral Risk Factor Surveillance Survey, despite covering the entire U.S., uses self-reported measures of diet and physical activity and is organized with the smallest areal unit being county.12,13 The National Health and Nutrition Examination Survey, on the other hand, directly measures diet and physical activity, but is organized to provide surveillance estimates at the national or regional level.14 By contrast, the Affordable Care Act and the Health Information Technology for Economic and Clinical Health Act have made possible the conversion to electronic health records (EHRs) for most federally qualified health centers (FQHCs).15 This has created a situation where large volumes of clinical data can be spatially organized to potentially study the role of the neighborhood environment on a variety of outcomes, including obesity risk.16 The purpose of this study is to assess the feasibility of obtaining and using EHR data from an FQHC in order to assess the association between a measure of the local environment and obesity-related outcomes in a multilevel framework. A review of the health literature found only two studies in which health providers’ patient data were used to characterize obesity risks associated with neighborhood context. Oreskovic and colleagues17 geocoded by residence children (i.e., those aged 2–18 years) enrolled in Partners HealthCare, a healthcare network in eastern Massachusetts. The study focused on transportation measures and, of the eight studied measures, only number of subway stations within a child’s neighborhood showed a statistically significant inverse association with BMI. Drewnowski et al.18 studied obesity in adults enrolled in Group Health, a nonprofit healthcare provider for King County, Washington. Neighborhood was defined in terms of Census tract of residence. Area-based sociodemographic measures were obtained from the U.S. Census. Median home value, percentage college educated, and median household income at the Census tract level were found to be inversely associated with obesity. One factor that has been consistently linked to obesity risk in studies of neighborhood context is neighborhood walkability. One of the measures used to assess walkability is street connectivity. However, the association with obesity in general and physical activity in particular have not always been consistent.19–23 With this limitation in mind, the current study attempted to assess the feasibility of using directly measured data assessing obesity risk (i.e., BMI) obtained from EHRs. The possibility of using EHR data was made possible through a community partnership established as part of January 2017
S41
the Mid-South Transdisciplinary Collaborative Center. The Social Determinants of Health core of the Mid-South Transdisciplinary Collaborative Center had been working with a community-based health provider in the New Orleans area, Daughters of Charity (DOC), when the possibility of spatially organizing their patient population was discussed. DOC clinics needed the information to better characterize their patient population in terms of residence in at-risk neighborhoods.
METHODS The present study included patients who resided in Louisiana and visited DOC clinics in calendar years 2012–2013. DOC clinics are FQHCs serving primarily poor and minority residents in the greater New Orleans area. DOC have provided care in the New Orleans area for 4175 years at five clinic sites (Figure 1). The clinics provide primary and preventive care, pediatrics, women’s health, behavioral health, Special Supplemental Nutrition Program for Women, Infants, and Children, and dental services. Moredetailed information is available at their website (http://dcsno.org). Geographic, demographic, and clinical data are available through EHRs required for FQHCs through the Affordable Care Act. The present study was approved by the IRB of Louisiana State University Health Sciences Center.
Geocoding The home address listed in the patient record at their last clinic visit within the 2012 and 2013 calendar years was used for geocoding. Geocoding was carried out using Esri ArcMap, version 10.2, using a U.S. street address locator. The road network data used to build the address locator were based on the StreetMap North America, which contains 2005 Tele Atlas street data and were enhanced by both Esri and Tele Atlas. Patients without a home address or with a PO box for a home address were excluded from the study. In addition, addresses included were restricted to those with a matching score Z80. A total of 31,297 patients were eligible to be included in the study. Within the eligible sample, 27,659 patients’ addresses were matched by the aforementioned software and 648 were manually matched, for a total of 28,307 patients who were geocoded and assigned to a Census tract (90.4% matching). The remaining 2,990 addresses did not match because the address was not found (n¼1,662 [55.6%]), was incomplete (n¼735 [24.6%]), consisted of a PO box only (n¼568 [19.0%]), or some other issue (n¼25 [0.8%]). Subsequently, the 28,307 geocoded patients were assigned a randomly generated Census tract ID, which could be linked to the Census tractlevel variables to preserve the multilevel structure of the data but simultaneously allowed for removal of the actual Census tract identifier. Furthermore, patients whose address was geocoded were also assigned a randomly generated personal ID. All data were managed and analyzed on DOC computers to preserve confidentiality.
Measures Available self-reported demographic information, such as gender (female or male), race/ethnicity (non-Hispanic black or nonHispanic white or Hispanic), age at visit, and marital status (married or not married), on patients whose addresses were
S42
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Figure 1. Location of Daughters of Charity System clinics across the greater New Orleans area overlaid with Census tracts population density. geocoded was extracted by DOC personnel and compiled into a database. Self-identified “Hispanic” ethnicity was categorized as Hispanic in the race/ethnicity variable. All other non-Hispanic patients were categorized based on their self-reported races: nonHispanic white or non-Hispanic black. Non-Hispanics that did not identify as either white or black accounted for o2% of the sample and were excluded from the analysis. The outcome of interest in the present study was BMI, defined as weight in kilograms (kg) divided by height in square meters (m2). Individual BMI was obtained by extracting all the encounters (N¼102,611) for the above 28,307 geocoded patients and only the value obtained during the last visit where BMI was measured was retained for the present analysis. This analysis included individuals who were aged Z20 years and not pregnant during their last encounter. Pregnancy status was established using ICD-9 codes. Census tract was used as the aggregate areal unit to permit nesting of individuals in groups permitting multilevel analysis. Consequently, the authors assumed each areal unit was independent. Individuals residing on a Census tract border may threaten the assumption of independence. It should be noted that Census tracts are defined as “designed to be relatively homogeneous units with respect to population characteristics, economic status, and living conditions at the time of establishment with an average of 4,000
persons.”24 A total of 451 Louisiana Census tracts were included in the final analysis. The average population across the 451 Census tracts was 2,753 people/mile2 (median, 2,322 people/mile2) with o5% of the patients residing in a Census tract with o798 people/ mile2. Furthermore, the average number of patients per Census tract was 39.8 (range, 1–255). The street connectivity variable utilized in the present analysis was intersection density, calculated as the number of three-way, four-way, or greater intersections per unit area (square mile).13 Street connectivity was calculated at the Census tract level utilizing ArcMap, version 10.2, based on 2000 Census topologically integrated geographic encoding and referencing/line files. Highways, interstates, on/off ramps, and private roads for service vehicles only were excluded from the road network because they do not contribute to walkability. Intersection density was not normally distributed (W¼0.93, po0.0001), was right skewed (skewness, 0.70), averaged 161.1 (median, 147.9) and ranged from 0.1 to 572.3 count/mile2. Census tracts were categorized as being in low, medium, or high intersection density tertiles by splitting them to have approximately equal number of patients within each tertile, with tied values been assigned to the same group. The following number of patients was included in each of the three tertiles: 6,119 for low, 6,058 for medium, and 5,769 for high tertile. www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Statistical Analysis Data from DOC clinics were provided with no guarantee of accuracy. As a result, some extreme values were observed, presumably due to recording errors. Therefore, a procedure was developed to identify unusual observations by identifying the 1st and the 99th percentile within the 2010–2012 National Health and Nutrition Examination Survey population and using those to exclude observations below the 1st and above the 99th percentile. In 2015 and 2016, data were analyzed using SAS/STAT, version 9.4. Overall summary statistics were produced using either the MEANS or the FREQUENCY procedure. At first, a simple association between intersection density tertile and BMI or age at visit was established using the MIXED procedure without considering the multilevel structure of the data. Next, analyses for BMI were conducted using random intercept models implemented through the MIXED procedure. An empty model and models including various combinations of individual-level and tract-level variables with their two- and three-way interactions were run. Individual-level variables consisted of age, gender, race, and marital status. The tract-level variable was intersection density. Models were simplified by removing non-significant (p40.05) three- or two-way interactions. When a significant interaction or main effect was observed, least square means were compared using TukeyKramer multiple comparison adjustment. Significance was declared at po0.05.
RESULTS Of the original 28,307 patients, 22,526 were aged Z20 years and not pregnant; of these, 2,696 were further excluded from the analysis because they either had no BMI value (n¼2,022) or their value was considered an outlier (n¼674). Furthermore, patients with incomplete sociodemographic information (age, gender, race/ethnicity, and marital status) were excluded from the analysis, such that a total of 17,946 (Appendix Figure 1, available
S43
online) patients were included in the final analytic sample. Descriptive statistics for patients who were not pregnant, aged Z20 years, had a BMI value that was not considered an outlier, and had complete sociodemographic information are reported in Table 1. Overall, patients on average were aged 44.1 years and had a BMI of 30.2. The largest portion of DOC patients were nonHispanic black (59.4%) (Table 1). A significant difference across the three intersection density tertiles was observed for BMI, obesity, and all the demographic variables reported in Table 1. The lowest intersection density tertile had a higher BMI and percentage obese, compared with the middle and high tertiles. Patients within the highest intersection density tertile tended to be older and included a higher proportion of non-Hispanic blacks and a lower proportion of non-Hispanic, Hispanic, and married patients than the remaining tertiles. Various multilevel models were run with BMI as the continuous outcome of interest. Tests of fixed effects for given factors with their interactions and covariance parameter estimates are reported in Appendix Table 1 (available online), and least square means (95% CIs) for categorical independent variables and parameter estimates (95% CIs) for continuous independent variables are reported in Table 2. From the empty model, the variance in BMI was partitioned between the individual and Census tract levels (Appendix Table 1, Model 1, available online). The intraclass correlation coefficient, or the amount of variance in BMI partitioned to the Census tract level, was 2.2% (Appendix Table 1, Model 1, available online). Intersection density was significantly associated with BMI and explained some of the BMI variability
Table 1. Descriptive Statistics for Analyzed Sample (n¼17,946) Intersection density tertile Variables Continuousb BMI Age, years Categorical, % Male Race/ethnicity Non-Hispanic black Non-Hispanic white Hispanic Married Obese
Low
Medium
High
p-valuea
30.2 (6.8) 44.1 (14.1)
30.7 (0.1) 44.0 (0.2)
30.0 (0.1) 43.5 (0.2)
30.0 (0.1) 44.7 (0.2)
o0.0001 o0.0001
31.2
30.5
30.3
32.9
0.0041
59.4 20.9 19.7 15.9 45.5
55.8 21.6 22.6 19.1 48.7
51.7 23.3 25.0 16.4 43.8
71.3 17.7 11.0 12.2 44.0
o0.0001
Overall
o0.0001 o0.0001
Note: Boldface indicates statistical significance for model fixed effects (po0.05). Data analyzed included non-pregnant, 20 years and older patients with a known marital status and either Hispanic or non-Hispanic black or white and having a non-outlier BMI value, within 451 Census tracts. a Association between the row and the column variable was determined by either χ2 test of independence or ANOVA. b Data for continuous variables is M (SD) for the overall sample, and M (SEM) for the intersection density tertiles.
January 2017
S44
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Table 2. Parameter Estimates or Least Square Means (95% CI) and Random Effect Estimates for BMI From Multilevel Models (n¼17,946) Fixed effects Age, years Age2, years2 Marital status Married Not married Gender Female Male Race/ethnicity Non-Hispanic black Non-Hispanic white Hispanic Gender race/ethnicity Female, non-Hispanic black Female, non-Hispanic white Female, Hispanic Male, non-Hispanic black Male, non-Hispanic white Male, Hispanic Intersection density tertile Low intersection density tertile Medium intersection density tertile High intersection density tertile Race/ethnicity intersection density tertile Non-Hispanic black, low intersection density tertile Non-Hispanic black, medium intersection density tertile Non-Hispanic black, high intersection density tertile Non-Hispanic white, low intersection density tertile Non-Hispanic white, medium intersection density tertile Non-Hispanic white, high intersection density tertile Hispanic, low intersection density tertile Hispanic, medium intersection density tertile Hispanic, high intersection density tertile Random effects, covariance parameter estimates Census level estimate Individual level estimate Fit statistics Bayesian Information Criterion
Model 2
Model 3
Model 4
Model 5
— —
0.42 (0.38, 0.46) 0.0043 (0.0047, 0.0038)
0.42 (0.38, 0.46) 0.0042 (0.0047, 0.0038)
0.42 (0.38, 0.46) 0.0042 (0.0047, 0.0038)
— —
30.1 (29.9, 30.2) 29.5 (29.3, 29.6)
30.1 (29.8, 30.3) 29.4 (29.3, 29.6)
30.1 (29.8, 30.3) 29.4 (29.3, 29.6)
— —
30.0 (29.8, 30.2) 29.6 (29.4, 29.8)
30.0 (29.8, 30.2) 29.5 (29.3, 29.8)
30.0 (29.8, 30.2) 29.5 (29.3, 29.8)
— — —
32.0 (28.9, 29.5) 28.9 (28.6, 29.2) 29.2 (28.9, 29.5)
30.9 (30.8, 31.1) 29.0 (28.7, 29.2) 29.3 (29.1, 29.6)
30.9 (30.7, 31.1) 29.0 (28.7, 29.2) 29.4 (29.1, 29.6)
— — — — — —
32.0 (31.7, 28.9 (28.6, 29.2 (28.9, 29.9 (29.6, 29.2 (28.8, 29.6 (29.2,
32.0 28.8 29.2 29.9 29.2 29.5
32.0 (31.7, 28.8 (28.5, 29.2 (28.9, 29.9 (29.6, 29.1 (28.8, 29.6 (29.2,
32.2) 29.2) 29.5) 30.2) 29.6) 30.0)
(31.7, (28.5, (28.9, (29.7, (28.8, (29.2,
32.2) 29.1) 29.4) 30.2) 29.5) 29.9)
32.2) 29.1) 29.5) 30.1) 29.5) 30.0)
30.6 (30.4, 30.9) 30.0 (29.7, 30.2) 29.9 (29.6, 30.1)
—
30.2 (30.0, 30.5)
30.3 (30.1, 30.5)
—
29.7 (29.5, 30.0)
29.7 (29.5, 30.0)
—
29.3 (29.1, 29.6)
29.2 (29.0, 29.5)
—
—
—
31.2 (31.0, 31.5)
—
—
—
30.9 (30.6, 31.2)
—
—
—
30.6 (30.4, 30.9)
—
—
—
30.0 (29.6, 30.3)
—
—
—
28.9 (28.5, 29.3)
—
—
—
28.0 (27.6, 28.5)
—
—
—
29.7 (29.3, 30.1)
—
—
—
29.4 (29.0, 29.8)
—
—
—
29.1 (28.5, 29.6)
0.94 45.8
0.35 43.7
0.27 43.7
0.26 43.7
119,810
118,883
118,850
118,836
Note: Values are least square means (95% CI) unless otherwise noted. Boldface indicates statistical significance for model fixed effects (po0.05). Data analyzed included non-pregnant, 20 years and older patients with a known marital status and either Hispanic or non-Hispanic black or white and having a non-outlier BMI value, within 451 Census tracts.
www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
partitioned to the Census tract level (Appendix Table 1, Model 2, available online). Adding the sociodemographic covariates to the model revealed that age, old age (i.e., age squared), gender, marital status, and race were all associated with BMI (Appendix Table 1, Model 3, available online). Similar results were observed when both intersection density and sociodemographic covariates were included into the model (Appendix Table 1, Model 4, available online). Furthermore, when interactions between intersection density and sociodemographic covariates were included in the model, the only interaction that was significant was race/ethnicity by intersection density (Appendix Table 1, Model 5, available online). To explore these associations, parameter estimates and least square means are reported in Table 2. When only intersection density was included in the model, patients within the low intersection density tertile had a higher BMI (least square mean¼30.6; 95% CI¼30.4, 30.9) than patients in the medium (least square mean¼30.0; 95% CI¼29.7, 30.2) or high (least square mean¼29.9; 95% CI¼29.6, 30.1) intersection density tertile (Table 2, Model 2). The inclusion of sociodemographic covariates into the model with intersection density resulted in the estimated BMI being lower by 0.4–0.5 for each increase in intersection density (Table 2, Model 4). When the relationship between BMI and intersection density was allowed to vary across the various race/ethnicity categories, a significant race by intersection density interaction was also observed (Table 2, Model 5). As intersection density increased the estimated average BMI decreased to a different extent across the various race/ethnicity categories (Figure 2). The estimated average BMI change from low to high intersection density was more
Figure 2. BMI (calculated as kg/m2) least square means (⫾95% CI) by intersection density tertile presented by race category (dot with solid line¼non-Hispanic black, circle with dashed line¼non-Hispanic white, triangle with dotted line¼Hispanic). January 2017
S45
pronounced for non-Hispanic whites (decrease of 2) than for non-Hispanic blacks or Hispanics, who decreased by approximately 0.6 from low intersection density tertile to high intersection density tertile.
DISCUSSION A study of the association between obesity risk and a measure of walkability (i.e., street connectivity) was performed using directly measured anthropometric data obtained from EHRs of an FQHC. The study was successful in obtaining the data, preserving confidentiality, and structuring the data in order to conduct multilevel analysis using random coefficient models. Though only a feasibility study, the analysis did find an inverse association between increasing a measure of street connectivity and BMI. In addition, a strong interaction was observed, indicating the association was stronger in non-Hispanic whites than for non-Hispanic blacks. These finding are similar to what has been found in the literature.13,20,25 Typically, studies of the role of the neighborhood environment on health outcomes like obesity risk have relied on large national databases (e.g., the Behavioral Risk Factor Surveillance Survey) or smaller studies sampling a relatively small number of residents in a particular city.13,20,22,26 In either case, these studies have a different set of limitations compared with the current study. Although the use of EHR data from an FQHC has its own limitations, it adds a new dimension to the study of social determinants in general and neighborhood walkability in particular.
Limitations The feasibility of using EHR data obtained from an FQHC has a number of limitations. First, clinical data are not measured and recorded with the exacting precision of a formal study. For example, nearly 3.3% of the geocoded patients aged Z20 years and not pregnant who had a recorded BMI value were excluded because the values recorded were believed to be too extreme (i.e., exceeding the 1st and 99th percentile for BMI in the National Health and Nutrition Examination Survey). It is unclear whether these exclusions introduce any systematic biases. In addition, a relatively low number of patient addresses were geocoded (i.e., 90.4%) compared with formal studies where home address is obtained despite relaxing the matching score for ArcMap to 80. Next, there were technical limitations that arose from the use of clinic-derived EHR data. The difficulty in characterizing pre-existing conditions or comorbidities limited the analysis. Identifying comorbid conditions is a major challenge in outcomes research with EHR data.
S46
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Although it is a rather simple determination in surveys where patients are asked to list their comorbidities, in an EHR, the determination is made with ICD-9 or ICD-10 codes. The determination of just one comorbidity requires extracting dozens of these codes. Creating data sets that include each of these codes is not feasible and asking FQHC information technology staff to write the code to summarize comorbidities was beyond the scope of this study. Finally, a logistic limitation involved the procedures to preserve confidentiality. As noted in the Methods section, the study data set was only available for analysis onsite at the FQHC. This created a number of logistic obstacles in working with the information technology staff of the FQHC.
CONCLUSIONS This study accessed EHR data from an FQHC to assess the feasibility of geocoding clinical data for studies examining the relationship between the neighborhood environment and a directly observed health outcome. Given the volume of data available on clinic populations with the move to EHRs, the promise of this type of analysis for conducting multilevel studies assessing the roles of social determinants of health in the local environment on health outcomes is great. Despite the success in conducting the multilevel analyses—which demonstrated obesity risk measured as BMI was associated with a measure of walkability (i.e., street connectivity) and the association was stronger for nonHispanic whites—the study had a number of limitations. The use of EHRs from an FQHC potentially represents a new method for assessing the role of environmental exposures in social determinants of health research if the limitations can be addressed in future studies.
ACKNOWLEDGMENTS Publication of this article was supported by the National Institutes of Health. The findings and conclusions in this article are those of the authors and do not necessarily represent the official position of the National Institutes of Health. Research reported in this paper was supported by the MidSouth Transdisciplinary Collaborative Center for Health Disparities Research through the National Institute on Minority Health and Health Disparities (U54MD008176). No financial disclosures were reported by the authors of this paper.
SUPPLEMENTAL MATERIAL Supplemental materials associated with this article can be found in the online version at http://dx.doi.org/10.1016/j. amepre.2016.09.029.
REFERENCES 1. Ogden CL, Carroll MD, Kit BK, Flegal KM. Prevalence of childhood and adult obesity in the United States, 2011-2012. JAMA. 2014;311 (8):806–814. http://dx.doi.org/10.1001/jama.2014.732. 2. Flegal KM, Carroll MD, Ogden CL, Curtin LR. Prevalence and trends in obesity among U.S. adults, 1999–2008. JAMA. 2010;303(3):235–241. http://dx.doi.org/10.1001/jama.2009.2014. 3. Swinburn BA, Sacks G, Hall KD, et al. The global obesity pandemic: shaped by global drivers and local environments. Lancet. 2011;378 (9793):804–814. http://dx.doi.org/10.1016/S0140-6736(11)60813-1. 4. Hill JO, Wyatt HR, Reed GW, Peters JC. Obesity and the environment: where do we go from here? Science. 2003;299(5608):853–855. http://dx. doi.org/10.1126/science.1079857. 5. Popkin BM, Duffey K, Gordon-Larsen P. Environmental influences on food choice, physical activity and energy balance. Physiol Behav. 2005;86(5):603–613. http://dx.doi.org/10.1016/j.physbeh.2005.08.051. 6. Sallis JF, Glanz K. Physical activity and food environments: solutions to the obesity epidemic. Milbank Q. 2009;87(1):123–154. http://dx.doi. org/10.1111/j.1468-0009.2009.00550.x. 7. Diez Roux AV, Mair C. Neighborhoods and health. Ann N Y Acad Sci. 2010;1186:125–145. http://dx.doi.org/10.1111/j.1749-6632.2009. 05333.x. 8. Ding D, Gebel K. Built environment, physical activity, and obesity: what have we learned from reviewing the literature? Health Place. 2012;18(1):100–105. http://dx.doi.org/10.1016/j.healthplace.2011.08.021. 9. Oakes JM. The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology. Soc Sci Med. 2004;58 (10):1929–1952. http://dx.doi.org/10.1016/j.socscimed.2003.08.004. 10. Oakes JM. Invited commentary: repeated measures, selection bias, and effect identification in neighborhood effect studies. Am J Epidemiol. 2014;180(8):785–787. http://dx.doi.org/10.1093/aje/kwu231. 11. Krieger N, Chen JT, Waterman PD, Rehkopf DH, Subramanian SV. Painting a truer picture of U.S. socioeconomic and racial/ethnic health inequalities: the Public Health Disparities Geocoding Project. Am J Public Health. 2005;95(2):312–323. http://dx.doi.org/10.2105/AJPH. 2003.032482. 12. Ewing R, Schmid T, Killingsworth R, Zlot A, Raudenbush S. Relationship between urban sprawl and physical activity, obesity, and morbidity. Am J Health Promot. 2003;18(1):47–57. http://dx.doi.org/10.4278/ 0890-1171-18.1.47. 13. Wang F, Wen M, Xu Y. Population-adjusted street connectivity, urbanicity and risk of obesity in the U.S. Appl Geogr. 2013;41:1–14. http://dx.doi.org/10.1016/j.apgeog.2013.03.006. 14. Krieger N, Chen JT, Waterman PD, et al. Geocoding and monitoring of U.S. socioeconomic inequalities in mortality and cancer incidence: does the choice of area-based measure and geographic level matter?: the Public Health Disparities Geocoding Project. Am J Epidemiol. 2002;156(5):471–482. http://dx.doi.org/10.1093/aje/kwf068. 15. Adler-Milstein J, DesRoches CM, Furukawa MF, et al. More than half of U.S. hospitals have at least a basic EHR, but stage 2 criteria remain challenging for most. Health Aff (Millwood). 2014;33(9):1664–1671. http://dx.doi.org/10.1377/hlthaff.2014.0453. 16. Hayman LL, Himmelfarb CD. Preventing obesity and promoting cardiometabolic health: the promise and potential of policies and the Affordable Care Act. J Cardiovasc Nurs. 2013;28(6):503–504. http://dx. doi.org/10.1097/JCN.0b013e31829eae4f. 17. Oreskovic NM, Winickoff JP, Kuhlthau KA, Romm D, Perrin JM. Obesity and the built environment among Massachusetts’s children. Clin Pediatr (Phila). 2009;48(9):904–912. http://dx.doi.org/10.1177/ 0009922809336073. 18. Drewnowski A, Rehm CD, Arterburn D. The geographic distribution of obesity by census tract among 59767 insured adults in King County, WA. Int J Obes (Lond). 2014;38:833–839. http://dx.doi.org/10.1038/ ijo.2013.179.
www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47 19. Berrigan D, Pickle LW, Dill J. Associations between street connectivity and active transportation. Int J Health Geogr. 2010;9:20. http://dx.doi. org/10.1186/1476-072X-9-20. 20. Heinrich KM, Lee RE, Regan GR, et al. How does the built environment relate to body mass index and obesity prevalence among public housing residents? Am J Health Promot. 2008;22(3):187–194. http://dx. doi.org/10.4278/ajhp.22.3.187. 21. Li F, Harmer PA, Cardinal BJ, et al. Built environment, adiposity, and physical activity in adults aged 50-75. Am J Prev Med. 2008;35(1):38–46. http://dx.doi.org/10.1016/j.amepre.2008.03.021. 22. McDonald KN, Oakes JM, Forsyth A. Effect of street connectivity and density on adult BMI: results from the Twin Cities Walking Study. J Epidemiol Community Health. 2012;66(7):636–640. http://dx.doi.org/ 10.1136/jech.2010.122556.
January 2017
S47
23. Oakes JM, Forsyth A, Schmitz KH. The effects of neighborhood density and street connectivity on walking behavior: the Twin Cities walking study. Epidemiol Perspect Innov. 2007;4:16. 24. Krieger N, Waterman P, Chen JT, et al. Zip code caveat: bias due to spatiotemporal mismatches between zip codes and U.S. census-defined geographic areas—the Public Health Disparities Geocoding Project. Am J Public Health. 2002;92(7):1100–1102. http://dx.doi.org/10.2105/ AJPH.92.7.1100. 25. Wen M, Kowaleski-Jones L. The built environment and risk of obesity in the United States: racial-ethnic disparities. Health Place. 2012;18 (6):1314–1322. http://dx.doi.org/10.1016/j.healthplace.2012.09.002. 26. Li F, Harmer P, Cardinal BJ, et al. Built environment and 1-year change in weight and waist circumference in middle-aged and older adults: Portland Neighborhood Environment and Health Study. Am J Epidemiol. 2009;169 (4):401–408. http://dx.doi.org/10.1093/aje/kwn398.
Street Connectivity and Obesity Risk: Evidence From Electronic Health Records Claudia Leonardi, PhD, Neal R. Simonsen, PhD, Qingzhao Yu, PhD, Chi Park, MS, Richard A. Scribner, MD, MPH Introduction: This study aimed to determine the feasibility of using electronic health record (EHR) data from a federally qualified health center (FQHC) to assess the association between street connectivity, a measure of walkability for the local environment, and BMI obtained from EHRs.
Methods: The study included patients who visited Daughters of Charity clinics in 2012–2013. A total of 31,297 patients were eligible, of which 28,307 were geocoded. BMI and sociodemographic information were compiled into a de-identified database. The street connectivity measure was intersection density, calculated as the number of three-way or greater intersections per unit area. Multilevel analyses of BMI, measured on 17,946 patients who were aged Z20 years, not pregnant, had complete sociodemographic information, and a BMI value that was not considered an outlier, were conducted using random intercept models. Results: Overall, on average, patients were aged 44.1 years, had a BMI of 30.2, and were mainly non-Hispanic black (59.4%). An inverse association between BMI and intersection density was observed in multilevel models controlling for age, gender, race, and marital status. Tests for multiple interactions were conducted and a significant interaction between race and intersection density indicated the decrease in BMI was strongest for non-Hispanic whites (decreased by 2) compared with blacks or Hispanics (decreased by 0.6) (p¼0.0121). Conclusions: EHRs were successfully used to assess the relationship between street connectivity and BMI in a multilevel framework. Increasing street connectivity levels measured as intersection density were inversely associated with directly measured BMI obtained from EHRs, demonstrating the feasibility of the approach. Am J Prev Med 2017;52(1S1):S40–S47. & 2016 American Journal of Preventive Medicine. Published by Elsevier Inc. All rights reserved.
INTRODUCTION
D
espite growing awareness of the negative health impact of poor diet, physical inactivity, and excess weight, the prevalence of obesity has increased dramatically in the U.S.1,2 During the last 3 decades, the U.S. obesity rate has doubled in adults and tripled in children and adolescents.1 The consensus among public health experts is that human genetic changes are not responsible for the rapid rise in obesity and the explanation must lie in social determinants resulting from environmental and policy changes.3–6 Although there is consensus on the general role of social, physical activity, and food environments in contributing to the epidemic, there is little in the way S40 Am J Prev Med 2017;52(1S1):S40–S47
of consensus on the relative contribution of any one social determinant (e.g., stressful environment, food deserts, concentrated disadvantage), let alone the relative contribution of any specific neighborhood or community context (e.g., fast food outlet density, walkability).7–10 It From the School of Public Health, Louisiana Cancer Research Center, Louisiana State University, New Orleans, Louisiana Address correspondence to: Richard A. Scribner, MD, MPH, School of Public Health, Louisiana Cancer Research Center, Louisiana State University, 1700 Tulane Avenue, Suite 907, New Orleans LA 70112. E-mail: [email protected]. This article is part of a supplement issue titled Social Determinants of Health: An Approach to Health Disparities Research. 0749-3797/$36.00 http://dx.doi.org/10.1016/j.amepre.2016.09.029
& 2016 American Journal of Preventive Medicine. Published by Elsevier Inc. All rights reserved.
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
has been recognized that the lack of data measured at a scale sufficient to assess the role of the neighborhood environment is a real limitation.11 For example, the two main U.S. public health surveillance systems, with a sufficient number of participants to assess the role of the environment, are limited in their capacity for assessing neighborhood risk. The Behavioral Risk Factor Surveillance Survey, despite covering the entire U.S., uses self-reported measures of diet and physical activity and is organized with the smallest areal unit being county.12,13 The National Health and Nutrition Examination Survey, on the other hand, directly measures diet and physical activity, but is organized to provide surveillance estimates at the national or regional level.14 By contrast, the Affordable Care Act and the Health Information Technology for Economic and Clinical Health Act have made possible the conversion to electronic health records (EHRs) for most federally qualified health centers (FQHCs).15 This has created a situation where large volumes of clinical data can be spatially organized to potentially study the role of the neighborhood environment on a variety of outcomes, including obesity risk.16 The purpose of this study is to assess the feasibility of obtaining and using EHR data from an FQHC in order to assess the association between a measure of the local environment and obesity-related outcomes in a multilevel framework. A review of the health literature found only two studies in which health providers’ patient data were used to characterize obesity risks associated with neighborhood context. Oreskovic and colleagues17 geocoded by residence children (i.e., those aged 2–18 years) enrolled in Partners HealthCare, a healthcare network in eastern Massachusetts. The study focused on transportation measures and, of the eight studied measures, only number of subway stations within a child’s neighborhood showed a statistically significant inverse association with BMI. Drewnowski et al.18 studied obesity in adults enrolled in Group Health, a nonprofit healthcare provider for King County, Washington. Neighborhood was defined in terms of Census tract of residence. Area-based sociodemographic measures were obtained from the U.S. Census. Median home value, percentage college educated, and median household income at the Census tract level were found to be inversely associated with obesity. One factor that has been consistently linked to obesity risk in studies of neighborhood context is neighborhood walkability. One of the measures used to assess walkability is street connectivity. However, the association with obesity in general and physical activity in particular have not always been consistent.19–23 With this limitation in mind, the current study attempted to assess the feasibility of using directly measured data assessing obesity risk (i.e., BMI) obtained from EHRs. The possibility of using EHR data was made possible through a community partnership established as part of January 2017
S41
the Mid-South Transdisciplinary Collaborative Center. The Social Determinants of Health core of the Mid-South Transdisciplinary Collaborative Center had been working with a community-based health provider in the New Orleans area, Daughters of Charity (DOC), when the possibility of spatially organizing their patient population was discussed. DOC clinics needed the information to better characterize their patient population in terms of residence in at-risk neighborhoods.
METHODS The present study included patients who resided in Louisiana and visited DOC clinics in calendar years 2012–2013. DOC clinics are FQHCs serving primarily poor and minority residents in the greater New Orleans area. DOC have provided care in the New Orleans area for 4175 years at five clinic sites (Figure 1). The clinics provide primary and preventive care, pediatrics, women’s health, behavioral health, Special Supplemental Nutrition Program for Women, Infants, and Children, and dental services. Moredetailed information is available at their website (http://dcsno.org). Geographic, demographic, and clinical data are available through EHRs required for FQHCs through the Affordable Care Act. The present study was approved by the IRB of Louisiana State University Health Sciences Center.
Geocoding The home address listed in the patient record at their last clinic visit within the 2012 and 2013 calendar years was used for geocoding. Geocoding was carried out using Esri ArcMap, version 10.2, using a U.S. street address locator. The road network data used to build the address locator were based on the StreetMap North America, which contains 2005 Tele Atlas street data and were enhanced by both Esri and Tele Atlas. Patients without a home address or with a PO box for a home address were excluded from the study. In addition, addresses included were restricted to those with a matching score Z80. A total of 31,297 patients were eligible to be included in the study. Within the eligible sample, 27,659 patients’ addresses were matched by the aforementioned software and 648 were manually matched, for a total of 28,307 patients who were geocoded and assigned to a Census tract (90.4% matching). The remaining 2,990 addresses did not match because the address was not found (n¼1,662 [55.6%]), was incomplete (n¼735 [24.6%]), consisted of a PO box only (n¼568 [19.0%]), or some other issue (n¼25 [0.8%]). Subsequently, the 28,307 geocoded patients were assigned a randomly generated Census tract ID, which could be linked to the Census tractlevel variables to preserve the multilevel structure of the data but simultaneously allowed for removal of the actual Census tract identifier. Furthermore, patients whose address was geocoded were also assigned a randomly generated personal ID. All data were managed and analyzed on DOC computers to preserve confidentiality.
Measures Available self-reported demographic information, such as gender (female or male), race/ethnicity (non-Hispanic black or nonHispanic white or Hispanic), age at visit, and marital status (married or not married), on patients whose addresses were
S42
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Figure 1. Location of Daughters of Charity System clinics across the greater New Orleans area overlaid with Census tracts population density. geocoded was extracted by DOC personnel and compiled into a database. Self-identified “Hispanic” ethnicity was categorized as Hispanic in the race/ethnicity variable. All other non-Hispanic patients were categorized based on their self-reported races: nonHispanic white or non-Hispanic black. Non-Hispanics that did not identify as either white or black accounted for o2% of the sample and were excluded from the analysis. The outcome of interest in the present study was BMI, defined as weight in kilograms (kg) divided by height in square meters (m2). Individual BMI was obtained by extracting all the encounters (N¼102,611) for the above 28,307 geocoded patients and only the value obtained during the last visit where BMI was measured was retained for the present analysis. This analysis included individuals who were aged Z20 years and not pregnant during their last encounter. Pregnancy status was established using ICD-9 codes. Census tract was used as the aggregate areal unit to permit nesting of individuals in groups permitting multilevel analysis. Consequently, the authors assumed each areal unit was independent. Individuals residing on a Census tract border may threaten the assumption of independence. It should be noted that Census tracts are defined as “designed to be relatively homogeneous units with respect to population characteristics, economic status, and living conditions at the time of establishment with an average of 4,000
persons.”24 A total of 451 Louisiana Census tracts were included in the final analysis. The average population across the 451 Census tracts was 2,753 people/mile2 (median, 2,322 people/mile2) with o5% of the patients residing in a Census tract with o798 people/ mile2. Furthermore, the average number of patients per Census tract was 39.8 (range, 1–255). The street connectivity variable utilized in the present analysis was intersection density, calculated as the number of three-way, four-way, or greater intersections per unit area (square mile).13 Street connectivity was calculated at the Census tract level utilizing ArcMap, version 10.2, based on 2000 Census topologically integrated geographic encoding and referencing/line files. Highways, interstates, on/off ramps, and private roads for service vehicles only were excluded from the road network because they do not contribute to walkability. Intersection density was not normally distributed (W¼0.93, po0.0001), was right skewed (skewness, 0.70), averaged 161.1 (median, 147.9) and ranged from 0.1 to 572.3 count/mile2. Census tracts were categorized as being in low, medium, or high intersection density tertiles by splitting them to have approximately equal number of patients within each tertile, with tied values been assigned to the same group. The following number of patients was included in each of the three tertiles: 6,119 for low, 6,058 for medium, and 5,769 for high tertile. www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Statistical Analysis Data from DOC clinics were provided with no guarantee of accuracy. As a result, some extreme values were observed, presumably due to recording errors. Therefore, a procedure was developed to identify unusual observations by identifying the 1st and the 99th percentile within the 2010–2012 National Health and Nutrition Examination Survey population and using those to exclude observations below the 1st and above the 99th percentile. In 2015 and 2016, data were analyzed using SAS/STAT, version 9.4. Overall summary statistics were produced using either the MEANS or the FREQUENCY procedure. At first, a simple association between intersection density tertile and BMI or age at visit was established using the MIXED procedure without considering the multilevel structure of the data. Next, analyses for BMI were conducted using random intercept models implemented through the MIXED procedure. An empty model and models including various combinations of individual-level and tract-level variables with their two- and three-way interactions were run. Individual-level variables consisted of age, gender, race, and marital status. The tract-level variable was intersection density. Models were simplified by removing non-significant (p40.05) three- or two-way interactions. When a significant interaction or main effect was observed, least square means were compared using TukeyKramer multiple comparison adjustment. Significance was declared at po0.05.
RESULTS Of the original 28,307 patients, 22,526 were aged Z20 years and not pregnant; of these, 2,696 were further excluded from the analysis because they either had no BMI value (n¼2,022) or their value was considered an outlier (n¼674). Furthermore, patients with incomplete sociodemographic information (age, gender, race/ethnicity, and marital status) were excluded from the analysis, such that a total of 17,946 (Appendix Figure 1, available
S43
online) patients were included in the final analytic sample. Descriptive statistics for patients who were not pregnant, aged Z20 years, had a BMI value that was not considered an outlier, and had complete sociodemographic information are reported in Table 1. Overall, patients on average were aged 44.1 years and had a BMI of 30.2. The largest portion of DOC patients were nonHispanic black (59.4%) (Table 1). A significant difference across the three intersection density tertiles was observed for BMI, obesity, and all the demographic variables reported in Table 1. The lowest intersection density tertile had a higher BMI and percentage obese, compared with the middle and high tertiles. Patients within the highest intersection density tertile tended to be older and included a higher proportion of non-Hispanic blacks and a lower proportion of non-Hispanic, Hispanic, and married patients than the remaining tertiles. Various multilevel models were run with BMI as the continuous outcome of interest. Tests of fixed effects for given factors with their interactions and covariance parameter estimates are reported in Appendix Table 1 (available online), and least square means (95% CIs) for categorical independent variables and parameter estimates (95% CIs) for continuous independent variables are reported in Table 2. From the empty model, the variance in BMI was partitioned between the individual and Census tract levels (Appendix Table 1, Model 1, available online). The intraclass correlation coefficient, or the amount of variance in BMI partitioned to the Census tract level, was 2.2% (Appendix Table 1, Model 1, available online). Intersection density was significantly associated with BMI and explained some of the BMI variability
Table 1. Descriptive Statistics for Analyzed Sample (n¼17,946) Intersection density tertile Variables Continuousb BMI Age, years Categorical, % Male Race/ethnicity Non-Hispanic black Non-Hispanic white Hispanic Married Obese
Low
Medium
High
p-valuea
30.2 (6.8) 44.1 (14.1)
30.7 (0.1) 44.0 (0.2)
30.0 (0.1) 43.5 (0.2)
30.0 (0.1) 44.7 (0.2)
o0.0001 o0.0001
31.2
30.5
30.3
32.9
0.0041
59.4 20.9 19.7 15.9 45.5
55.8 21.6 22.6 19.1 48.7
51.7 23.3 25.0 16.4 43.8
71.3 17.7 11.0 12.2 44.0
o0.0001
Overall
o0.0001 o0.0001
Note: Boldface indicates statistical significance for model fixed effects (po0.05). Data analyzed included non-pregnant, 20 years and older patients with a known marital status and either Hispanic or non-Hispanic black or white and having a non-outlier BMI value, within 451 Census tracts. a Association between the row and the column variable was determined by either χ2 test of independence or ANOVA. b Data for continuous variables is M (SD) for the overall sample, and M (SEM) for the intersection density tertiles.
January 2017
S44
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Table 2. Parameter Estimates or Least Square Means (95% CI) and Random Effect Estimates for BMI From Multilevel Models (n¼17,946) Fixed effects Age, years Age2, years2 Marital status Married Not married Gender Female Male Race/ethnicity Non-Hispanic black Non-Hispanic white Hispanic Gender race/ethnicity Female, non-Hispanic black Female, non-Hispanic white Female, Hispanic Male, non-Hispanic black Male, non-Hispanic white Male, Hispanic Intersection density tertile Low intersection density tertile Medium intersection density tertile High intersection density tertile Race/ethnicity intersection density tertile Non-Hispanic black, low intersection density tertile Non-Hispanic black, medium intersection density tertile Non-Hispanic black, high intersection density tertile Non-Hispanic white, low intersection density tertile Non-Hispanic white, medium intersection density tertile Non-Hispanic white, high intersection density tertile Hispanic, low intersection density tertile Hispanic, medium intersection density tertile Hispanic, high intersection density tertile Random effects, covariance parameter estimates Census level estimate Individual level estimate Fit statistics Bayesian Information Criterion
Model 2
Model 3
Model 4
Model 5
— —
0.42 (0.38, 0.46) 0.0043 (0.0047, 0.0038)
0.42 (0.38, 0.46) 0.0042 (0.0047, 0.0038)
0.42 (0.38, 0.46) 0.0042 (0.0047, 0.0038)
— —
30.1 (29.9, 30.2) 29.5 (29.3, 29.6)
30.1 (29.8, 30.3) 29.4 (29.3, 29.6)
30.1 (29.8, 30.3) 29.4 (29.3, 29.6)
— —
30.0 (29.8, 30.2) 29.6 (29.4, 29.8)
30.0 (29.8, 30.2) 29.5 (29.3, 29.8)
30.0 (29.8, 30.2) 29.5 (29.3, 29.8)
— — —
32.0 (28.9, 29.5) 28.9 (28.6, 29.2) 29.2 (28.9, 29.5)
30.9 (30.8, 31.1) 29.0 (28.7, 29.2) 29.3 (29.1, 29.6)
30.9 (30.7, 31.1) 29.0 (28.7, 29.2) 29.4 (29.1, 29.6)
— — — — — —
32.0 (31.7, 28.9 (28.6, 29.2 (28.9, 29.9 (29.6, 29.2 (28.8, 29.6 (29.2,
32.0 28.8 29.2 29.9 29.2 29.5
32.0 (31.7, 28.8 (28.5, 29.2 (28.9, 29.9 (29.6, 29.1 (28.8, 29.6 (29.2,
32.2) 29.2) 29.5) 30.2) 29.6) 30.0)
(31.7, (28.5, (28.9, (29.7, (28.8, (29.2,
32.2) 29.1) 29.4) 30.2) 29.5) 29.9)
32.2) 29.1) 29.5) 30.1) 29.5) 30.0)
30.6 (30.4, 30.9) 30.0 (29.7, 30.2) 29.9 (29.6, 30.1)
—
30.2 (30.0, 30.5)
30.3 (30.1, 30.5)
—
29.7 (29.5, 30.0)
29.7 (29.5, 30.0)
—
29.3 (29.1, 29.6)
29.2 (29.0, 29.5)
—
—
—
31.2 (31.0, 31.5)
—
—
—
30.9 (30.6, 31.2)
—
—
—
30.6 (30.4, 30.9)
—
—
—
30.0 (29.6, 30.3)
—
—
—
28.9 (28.5, 29.3)
—
—
—
28.0 (27.6, 28.5)
—
—
—
29.7 (29.3, 30.1)
—
—
—
29.4 (29.0, 29.8)
—
—
—
29.1 (28.5, 29.6)
0.94 45.8
0.35 43.7
0.27 43.7
0.26 43.7
119,810
118,883
118,850
118,836
Note: Values are least square means (95% CI) unless otherwise noted. Boldface indicates statistical significance for model fixed effects (po0.05). Data analyzed included non-pregnant, 20 years and older patients with a known marital status and either Hispanic or non-Hispanic black or white and having a non-outlier BMI value, within 451 Census tracts.
www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
partitioned to the Census tract level (Appendix Table 1, Model 2, available online). Adding the sociodemographic covariates to the model revealed that age, old age (i.e., age squared), gender, marital status, and race were all associated with BMI (Appendix Table 1, Model 3, available online). Similar results were observed when both intersection density and sociodemographic covariates were included into the model (Appendix Table 1, Model 4, available online). Furthermore, when interactions between intersection density and sociodemographic covariates were included in the model, the only interaction that was significant was race/ethnicity by intersection density (Appendix Table 1, Model 5, available online). To explore these associations, parameter estimates and least square means are reported in Table 2. When only intersection density was included in the model, patients within the low intersection density tertile had a higher BMI (least square mean¼30.6; 95% CI¼30.4, 30.9) than patients in the medium (least square mean¼30.0; 95% CI¼29.7, 30.2) or high (least square mean¼29.9; 95% CI¼29.6, 30.1) intersection density tertile (Table 2, Model 2). The inclusion of sociodemographic covariates into the model with intersection density resulted in the estimated BMI being lower by 0.4–0.5 for each increase in intersection density (Table 2, Model 4). When the relationship between BMI and intersection density was allowed to vary across the various race/ethnicity categories, a significant race by intersection density interaction was also observed (Table 2, Model 5). As intersection density increased the estimated average BMI decreased to a different extent across the various race/ethnicity categories (Figure 2). The estimated average BMI change from low to high intersection density was more
Figure 2. BMI (calculated as kg/m2) least square means (⫾95% CI) by intersection density tertile presented by race category (dot with solid line¼non-Hispanic black, circle with dashed line¼non-Hispanic white, triangle with dotted line¼Hispanic). January 2017
S45
pronounced for non-Hispanic whites (decrease of 2) than for non-Hispanic blacks or Hispanics, who decreased by approximately 0.6 from low intersection density tertile to high intersection density tertile.
DISCUSSION A study of the association between obesity risk and a measure of walkability (i.e., street connectivity) was performed using directly measured anthropometric data obtained from EHRs of an FQHC. The study was successful in obtaining the data, preserving confidentiality, and structuring the data in order to conduct multilevel analysis using random coefficient models. Though only a feasibility study, the analysis did find an inverse association between increasing a measure of street connectivity and BMI. In addition, a strong interaction was observed, indicating the association was stronger in non-Hispanic whites than for non-Hispanic blacks. These finding are similar to what has been found in the literature.13,20,25 Typically, studies of the role of the neighborhood environment on health outcomes like obesity risk have relied on large national databases (e.g., the Behavioral Risk Factor Surveillance Survey) or smaller studies sampling a relatively small number of residents in a particular city.13,20,22,26 In either case, these studies have a different set of limitations compared with the current study. Although the use of EHR data from an FQHC has its own limitations, it adds a new dimension to the study of social determinants in general and neighborhood walkability in particular.
Limitations The feasibility of using EHR data obtained from an FQHC has a number of limitations. First, clinical data are not measured and recorded with the exacting precision of a formal study. For example, nearly 3.3% of the geocoded patients aged Z20 years and not pregnant who had a recorded BMI value were excluded because the values recorded were believed to be too extreme (i.e., exceeding the 1st and 99th percentile for BMI in the National Health and Nutrition Examination Survey). It is unclear whether these exclusions introduce any systematic biases. In addition, a relatively low number of patient addresses were geocoded (i.e., 90.4%) compared with formal studies where home address is obtained despite relaxing the matching score for ArcMap to 80. Next, there were technical limitations that arose from the use of clinic-derived EHR data. The difficulty in characterizing pre-existing conditions or comorbidities limited the analysis. Identifying comorbid conditions is a major challenge in outcomes research with EHR data.
S46
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47
Although it is a rather simple determination in surveys where patients are asked to list their comorbidities, in an EHR, the determination is made with ICD-9 or ICD-10 codes. The determination of just one comorbidity requires extracting dozens of these codes. Creating data sets that include each of these codes is not feasible and asking FQHC information technology staff to write the code to summarize comorbidities was beyond the scope of this study. Finally, a logistic limitation involved the procedures to preserve confidentiality. As noted in the Methods section, the study data set was only available for analysis onsite at the FQHC. This created a number of logistic obstacles in working with the information technology staff of the FQHC.
CONCLUSIONS This study accessed EHR data from an FQHC to assess the feasibility of geocoding clinical data for studies examining the relationship between the neighborhood environment and a directly observed health outcome. Given the volume of data available on clinic populations with the move to EHRs, the promise of this type of analysis for conducting multilevel studies assessing the roles of social determinants of health in the local environment on health outcomes is great. Despite the success in conducting the multilevel analyses—which demonstrated obesity risk measured as BMI was associated with a measure of walkability (i.e., street connectivity) and the association was stronger for nonHispanic whites—the study had a number of limitations. The use of EHRs from an FQHC potentially represents a new method for assessing the role of environmental exposures in social determinants of health research if the limitations can be addressed in future studies.
ACKNOWLEDGMENTS Publication of this article was supported by the National Institutes of Health. The findings and conclusions in this article are those of the authors and do not necessarily represent the official position of the National Institutes of Health. Research reported in this paper was supported by the MidSouth Transdisciplinary Collaborative Center for Health Disparities Research through the National Institute on Minority Health and Health Disparities (U54MD008176). No financial disclosures were reported by the authors of this paper.
SUPPLEMENTAL MATERIAL Supplemental materials associated with this article can be found in the online version at http://dx.doi.org/10.1016/j. amepre.2016.09.029.
REFERENCES 1. Ogden CL, Carroll MD, Kit BK, Flegal KM. Prevalence of childhood and adult obesity in the United States, 2011-2012. JAMA. 2014;311 (8):806–814. http://dx.doi.org/10.1001/jama.2014.732. 2. Flegal KM, Carroll MD, Ogden CL, Curtin LR. Prevalence and trends in obesity among U.S. adults, 1999–2008. JAMA. 2010;303(3):235–241. http://dx.doi.org/10.1001/jama.2009.2014. 3. Swinburn BA, Sacks G, Hall KD, et al. The global obesity pandemic: shaped by global drivers and local environments. Lancet. 2011;378 (9793):804–814. http://dx.doi.org/10.1016/S0140-6736(11)60813-1. 4. Hill JO, Wyatt HR, Reed GW, Peters JC. Obesity and the environment: where do we go from here? Science. 2003;299(5608):853–855. http://dx. doi.org/10.1126/science.1079857. 5. Popkin BM, Duffey K, Gordon-Larsen P. Environmental influences on food choice, physical activity and energy balance. Physiol Behav. 2005;86(5):603–613. http://dx.doi.org/10.1016/j.physbeh.2005.08.051. 6. Sallis JF, Glanz K. Physical activity and food environments: solutions to the obesity epidemic. Milbank Q. 2009;87(1):123–154. http://dx.doi. org/10.1111/j.1468-0009.2009.00550.x. 7. Diez Roux AV, Mair C. Neighborhoods and health. Ann N Y Acad Sci. 2010;1186:125–145. http://dx.doi.org/10.1111/j.1749-6632.2009. 05333.x. 8. Ding D, Gebel K. Built environment, physical activity, and obesity: what have we learned from reviewing the literature? Health Place. 2012;18(1):100–105. http://dx.doi.org/10.1016/j.healthplace.2011.08.021. 9. Oakes JM. The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology. Soc Sci Med. 2004;58 (10):1929–1952. http://dx.doi.org/10.1016/j.socscimed.2003.08.004. 10. Oakes JM. Invited commentary: repeated measures, selection bias, and effect identification in neighborhood effect studies. Am J Epidemiol. 2014;180(8):785–787. http://dx.doi.org/10.1093/aje/kwu231. 11. Krieger N, Chen JT, Waterman PD, Rehkopf DH, Subramanian SV. Painting a truer picture of U.S. socioeconomic and racial/ethnic health inequalities: the Public Health Disparities Geocoding Project. Am J Public Health. 2005;95(2):312–323. http://dx.doi.org/10.2105/AJPH. 2003.032482. 12. Ewing R, Schmid T, Killingsworth R, Zlot A, Raudenbush S. Relationship between urban sprawl and physical activity, obesity, and morbidity. Am J Health Promot. 2003;18(1):47–57. http://dx.doi.org/10.4278/ 0890-1171-18.1.47. 13. Wang F, Wen M, Xu Y. Population-adjusted street connectivity, urbanicity and risk of obesity in the U.S. Appl Geogr. 2013;41:1–14. http://dx.doi.org/10.1016/j.apgeog.2013.03.006. 14. Krieger N, Chen JT, Waterman PD, et al. Geocoding and monitoring of U.S. socioeconomic inequalities in mortality and cancer incidence: does the choice of area-based measure and geographic level matter?: the Public Health Disparities Geocoding Project. Am J Epidemiol. 2002;156(5):471–482. http://dx.doi.org/10.1093/aje/kwf068. 15. Adler-Milstein J, DesRoches CM, Furukawa MF, et al. More than half of U.S. hospitals have at least a basic EHR, but stage 2 criteria remain challenging for most. Health Aff (Millwood). 2014;33(9):1664–1671. http://dx.doi.org/10.1377/hlthaff.2014.0453. 16. Hayman LL, Himmelfarb CD. Preventing obesity and promoting cardiometabolic health: the promise and potential of policies and the Affordable Care Act. J Cardiovasc Nurs. 2013;28(6):503–504. http://dx. doi.org/10.1097/JCN.0b013e31829eae4f. 17. Oreskovic NM, Winickoff JP, Kuhlthau KA, Romm D, Perrin JM. Obesity and the built environment among Massachusetts’s children. Clin Pediatr (Phila). 2009;48(9):904–912. http://dx.doi.org/10.1177/ 0009922809336073. 18. Drewnowski A, Rehm CD, Arterburn D. The geographic distribution of obesity by census tract among 59767 insured adults in King County, WA. Int J Obes (Lond). 2014;38:833–839. http://dx.doi.org/10.1038/ ijo.2013.179.
www.ajpmonline.org
Leonardi et al / Am J Prev Med 2017;52(1S1):S40–S47 19. Berrigan D, Pickle LW, Dill J. Associations between street connectivity and active transportation. Int J Health Geogr. 2010;9:20. http://dx.doi. org/10.1186/1476-072X-9-20. 20. Heinrich KM, Lee RE, Regan GR, et al. How does the built environment relate to body mass index and obesity prevalence among public housing residents? Am J Health Promot. 2008;22(3):187–194. http://dx. doi.org/10.4278/ajhp.22.3.187. 21. Li F, Harmer PA, Cardinal BJ, et al. Built environment, adiposity, and physical activity in adults aged 50-75. Am J Prev Med. 2008;35(1):38–46. http://dx.doi.org/10.1016/j.amepre.2008.03.021. 22. McDonald KN, Oakes JM, Forsyth A. Effect of street connectivity and density on adult BMI: results from the Twin Cities Walking Study. J Epidemiol Community Health. 2012;66(7):636–640. http://dx.doi.org/ 10.1136/jech.2010.122556.
January 2017
S47
23. Oakes JM, Forsyth A, Schmitz KH. The effects of neighborhood density and street connectivity on walking behavior: the Twin Cities walking study. Epidemiol Perspect Innov. 2007;4:16. 24. Krieger N, Waterman P, Chen JT, et al. Zip code caveat: bias due to spatiotemporal mismatches between zip codes and U.S. census-defined geographic areas—the Public Health Disparities Geocoding Project. Am J Public Health. 2002;92(7):1100–1102. http://dx.doi.org/10.2105/ AJPH.92.7.1100. 25. Wen M, Kowaleski-Jones L. The built environment and risk of obesity in the United States: racial-ethnic disparities. Health Place. 2012;18 (6):1314–1322. http://dx.doi.org/10.1016/j.healthplace.2012.09.002. 26. Li F, Harmer P, Cardinal BJ, et al. Built environment and 1-year change in weight and waist circumference in middle-aged and older adults: Portland Neighborhood Environment and Health Study. Am J Epidemiol. 2009;169 (4):401–408. http://dx.doi.org/10.1093/aje/kwn398.