1188 Care Volume 41, June 2018

“ ” Seth A. Berkowitz,1,2,3 Andrew J. Karter,4 Food Insecurity, Food Deserts, Giselle Corbie-Smith,5,6 Hilary K. Seligman,7,8 Sarah A. Ackroyd,9 and Glycemic Control in Patients Lily S. Barnard,10 Steven J. Atlas,1,3 and With Diabetes: A Longitudinal Deborah J. Wexler2,3 Analysis Diabetes Care 2018;41:1188–1195 | https://doi.org/10.2337/dc17-1981

OBJECTIVE Both food insecurity (limited food access owing to cost) and living in areas with low physical access to nutritious foods are public concerns, but their relative contribution to diabetes management is poorly understood.

RESEARCH DESIGN AND METHODS This was a prospective cohort study. A random sample of patients with diabetes in a EPIDEMIOLOGY/HEALTH SERVICES RESEARCH primary care network completed food insecurity assessment in 2013. Low physical food access at the census tract level was defined as no within 1 mile in 1Division of General Internal Medicine, Massa- urban areas and 10 miles in rural areas. HbA measurements were obtained from chusetts General Hospital, Boston, MA 1c 2Diabetes Unit, Massachusetts General Hospital, electronic health records through November 2016. The relationship among food Boston, MA 3 insecurity, low physical food access, and glycemic control (as defined by HbA1c)was Harvard Medical School, Boston, MA 4 analyzed using hierarchical linear mixed models. Division of Research, Kaiser Permanente North- ern California, Oakland, CA 5Center for Health Equity Research, Department RESULTS of Social Medicine, University of North Carolina Three hundred and ninety-one participants were followed for a mean of 37 months. School of Medicine, Chapel Hill, NC 6 Twentypercentofrespondentsreportedfoodinsecurity,and31%residedinanareaof Department of Medicine, University of North Carolina School of Medicine, Chapel Hill, NC low physical food access. In adjusted models, food insecurity was associated with 7Division of General Internal Medicine, University higherHbA1c (differenceof0.6%[6.6mmol/mol],95%CI0.4–0.8[4.4–8.7],P<0.0001), of California, San Francisco, San Francisco, CA which did not improve over time (P = 0.50). Living in an area with low physical food 8Center for Vulnerable Populations at Zuckerberg San Francisco General Hospital and Trauma access wasnot associatedwithadifferenceinHbA1c (difference0.2%[2.2mmol/mol], 2 2 P P Center, San Francisco, CA 95% CI 0.2 to 0.5 [ 2.2 to 5.6], = 0.33) or with change over time ( = 0.07). 9Department of Obstetrics, Gynecology & Re- productive Sciences, Temple University Hospital, CONCLUSIONS Philadelphia, PA 10 Food insecurity is associated with higher HbA1c, but living in an area with low physical UniversityofCalifornia,SanFrancisco,Schoolof food access is not. Food insecurity screening and interventions may help improve Medicine, San Francisco, CA glycemic control for vulnerable patients. Corresponding author: Seth A. Berkowitz, seth_ [email protected]. Received 22 September 2017 and accepted 27 Diabetes affects more than 30 million Americans and is both more common and more February 2018. likelytoleadtocomplicationsinthosewithlowersocioeconomicstatus(1,2).Adherence This article contains Supplementary Data online at to a healthy is a cornerstone of diabetes management, but those who are socio- http://care.diabetesjournals.org/lookup/suppl/ economically vulnerable face several important barriers to dietary adherence. First, food doi:10.2337/dc17-1981/-/DC1. insecurity, the uncertain or limited availability of food owing to cost, is highly prevalent in © 2018 by the American Diabetes Association. theU.S.,affecting;13%ofAmericanhouseholdsoverall,20%ofAmericanswithdiabetes, Readers may use this article as long as the work . is properly cited, the use is educational and not and25%ofAmericanswithpoorglycemiccontrol(hemoglobinA1c [HbA1c] 9.0%)(3,4). for profit, and the work is not altered. More infor- In addition, area factors, such as having a place to purchase nutritious food, termed mation is available at http://www.diabetesjournals “physical food access” by the U.S. Department of Agriculture (USDA) (5), lack of .org/content/license. care.diabetesjournals.org Berkowitz and Associates 1189

transportation, and neighborhood depri- items regarding diabetes history and social foodinyourarea)fromotherlimitstofood vation, can also make following an appro- circumstances. Responses to this question- access,suchasfoodinsecurity(beingunable priate diet more challenging. naire were linked with electronic health record to afford food) or the inability to go out A growing body of evidence has found information from the participant’sprimary andpurchasefoodowingtodisability.There associations between area-level factors care network. Participants were then have been several definitions of low phys- and cardiometabolic health overall and di- followed,withoutrecontact,throughtheir ical food access used in research. The most abetes in particular (6–13). However, there electronic health records. The social cir- common, and the primary definition for this is a significant gap in our knowledge re- cumstances survey items queried partic- study, comes from the USDA’s Food Access garding how these factors may act to ipants regarding their experience over the Research Atlas (5). This defines an area of influence the management of diabetes. In 12 months prior to instrument adminis- low physical food access as “atractwithat particular, it is important to distinguish tration; accordingly, this study used med- least 500 people or 33% of the population compositional factors (the individual-level ical record data beginning on 1 June living .1 mile (urban areas) or .10 miles features of the people who make up a 2012.Participantswerepassivelyfollowed (rural areas) from the nearest supermar- given area) from contextual factors (fea- until 30 November 2016 or their last in- ket, supercenter, or large grocery store.” tures of the area itself). This distinction has teractionwiththeprimarycarenetworkd We also considered an alternative defini- significant implications for a period of up to 54 months. Study data tion contained in the Food Access Re- and health policy. The concept of food werecollectedandmanagedusingREDCap search Atlas that uses a half mile in urban “deserts,” areas with low income and low (Research Electronic Data Capture) tools areas and 10 miles in rural areas as the cut physical access to food, has spurred major hosted at Partners HealthCare (18). The point. Two other factors are important for policy initiatives involving both the gov- Human Subjects Research Committee further determining whether a low phys- ernment and the private sector (14). Even at Partners HealthCare approved this ical food access census tract is a food as initiatives designed to address contex- protocol. “desert”: low income level and low access tual factors have been implemented, many to motor vehicles (a census tract–level programs that target individual-level fac- Assessment of Food Insecurity and proxy indicator). To incorporate these tors have been cut. For example, the Sup- Physical Food Access other considerations, we included in our plemental Assistance Program Food insecurity was assessed using the study three additional census tract–level (SNAP) (formerly the Food Stamp Program), six-itemshortformoftheUSDAFoodSecurity variables: median family income, poverty which directly combats food insecurity, Survey Module (19). An example item is, rate, and an indicator of low vehicle hasseenreductionsinbenefitlevels,along “‘The food that I bought just didn’tlast, access (a tract where $100 households with proposals for restrictingeligibility and and I didn’t have money toget more.’ Was do not have vehicle access and live more additional deep financial cuts (15,16). that often, sometimes, or never true for than half a mile from a supermarket). In the Given the implications for intervention you in the last 12 months?” In accordance U.S. overall, of 28,541 low physical food programs, distinguishing the health impact with standard scoring, participants who access tracts, 8,959 are considered food of low physical food access from food answeredaffirmatively(“often”or“some- “deserts” (5). insecurityishighlyrelevantforpolicy.Inthis times” for items where these were re- We also examined an alternate meth- study, we sought to determine the relative sponse categories and “yes” for items odology for classifying food access: the contribution of individual- and area-level where “yes” or “no” were the response Modified Retail Food Environment Index factorstoglycemiccontrol,basedonHbA1c, categories) to two or more items were (mRFEI), created by the Division of Nutri- in patientswith diabetes.From prior focus recordedasreportingfoodinsecurity.Asis tion, Physical Activity, and at the groupsthatweconducted,whichdiscussed common in the food insecurity and di- Centers for Disease Control and Preven- barriers to glycemic control in patients with abetes literature, we individualized the tion (CDC) (23). This index is calculated as diabetes, we hypothesized that food in- questions and treated food insecurity as thenumberofhealthyfoodretailersinthe security would be associated with worse an individual-level factor for analysis pur- census tract, divided by the total number glycemic control but that physical food poses, even though it is technically a of food retailers in the census tract (23). access would not. household-level factor (4,17,20–22). Five The mRFEI ranges from 0 to 1 and helps participants (1% of total) did not have a one to understand both low physical complete food insecurity assessment. food access (absence of healthy food RESEARCH DESIGN AND METHODS Residence in an area of low physical retailers, mRFEI scores = 0) and high Setting and Study Sample food access was based on linkage of partic- levels of unhealthy food retail, such as a Data collection methods for the study ipants’ geocoded addresses to the census predominance of fast food restaurants have previously been described (17). In tract–level physical food access data from (mRFEI scores close to, but not exactly, brief, a randomly selected cohort of adult the USDA’s Food Access Research Atlas 0), a situation sometimes termed living (aged $21 years) patients with diabetes (5). Participants with addresses that could in a food “swamp.” (any type) was enrolled between June and not be geocoded, most commonly as a The datafor census tract characteristics October 2013 from four clinics in eastern result of having a post office box listed for for this study were taken from the USDA’s Massachusetts(one academicprimary care the address, were excluded. Addresses Food Access Research Atlas Data Download practice, two community health centers, were updated and regeocoded annually 2015, the most recent version available, and one specialty diabetes practice). Par- throughout the study period. It is impor- except for the mRFEI score, which came ticipants completed a questionnaire, in En- tant to distinguish lack of physical food from the CDC. The variables in the Food glish or Spanish, consisting of validated access (i.e., having a place to purchase Access Research Atlas were collected 1190 Food Insecurity, Food Deserts, and Glycemic Control Diabetes Care Volume 41, June 2018

between 2010 and 2015 and were con- physical food access area was associated All statistical analyses were conducted sidered up to date at the time they were with greater HbA1c,wefit hierarchical inSAS,version9.4(SASInstitute,Cary,NC). released. Data on neighborhood eco- linear mixed models (SAS PROC MIXED). nomic circumstances contained in the To do this, we conceptualized HbA1c ob- RESULTS Food Access Research Atlas were collected servations as being nested within individ- Four hundred and eleven participants by the U.S. Census. The mRFEI score data ual participants and those participants completed the survey; 391 of these (95%) were released by the CDC in 2011. beingnestedwithincensustracts.Thecen- had baseline addresses that could be geo- sus tract information was time varying coded and were thus included in the study. Primary Outcome to correspond with the census tract the The demographic and clinical features of The primary outcome for this study was participant was living in at the time of those participants excluded because they HbA level, which was measured as part fi 1c HbA1c measurement. We rst calculated could not be geocoded were very similar of routine care. Multiple measurements intraclass correlations, which quantify the to the features of those who could be took place during the study period, and all variation explained at each level of the geocoded (Supplementary Table 1). The measurements were included in analysis of model: individual and census tract. We did overall mean and median durations of the outcome as described below. Data were this by fitting null (intercept only) models, follow-up were 37 and 41 months, respec- obtained from the electronic health record. following the method proposed by Bell tively. The mean age of study participants All HbA assays in this care network are 1c et al. (26), and calculating the percentage was 62 years, and the mean duration of conducted at the same central laboratory, of the variance explained at each level (by diabetes was 12 years. fi which is certi ed by the National Glyco- dividing the variance for a given level by Twenty percent of respondents re- hemoglobin Standardization Program. All the total variance). For our main analysis, ported food insecurity. The duration of available HbA1c values measured during the fi we t three-level models with random follow-up and number of HbA1c assess- study period were included in the analysis. effects terms for the individualandcensus ments for respondents with food insecu- tract. We included both random intercept rity did not differ from those for participants Covariates terms as well as random slope terms for without food insecurity (Table 1). Thirty- We considered several covariates that food insecurity and time (at the individual one percent of respondents lived in an may confound the relationship among level) and physical food access indicator area with low physical food access at food insecurity, food access, and HbA1c at (at the census tract level). This allows the baseline. Fifteen percent of participants both the individual and census tract level. HbA1c trajectory to vary between groups. who reported food insecurity were living With the exception of the census tract We used an autoregressive covariance in an area with low physical food access, variables discussed above, all data came structure to model HbA1c levels, as mea- and 10% of participants living in an area from the completed questionnaire or the surements taken closer together in time with low physical food access were food electronic health record. We considered aremorelikelytobecorrelated.Wetested insecure (P for comparison = 0.0009) the sociodemographic factors of baseline whether the HbA1c trajectory changed (Fig. 1). Participants who reported food age (in years), sex, race/ethnicity (catego- over time by specifying interaction terms. insecurity were more likely to be racial/ rized as non-Hispanic white, non-Hispanic We used interaction terms to test whether ethnic minorities, have Medicaid insur- black, Hispanic, and Asian/multiple ethni- the temporal trend in HbA1c varied by the ance, and have low educational attain- cities [multi]/other), educational attain- level of food insecurity or physical food ment compared with those who were ment (less than high school diploma, high access. In our modeling, variables related food secure (Table 1). Participants living in school diploma, or greater than high school to census tract of residence (low food an area with low physical food access were diploma), and health insurance (private, access, median family income, poverty more likely to be non-Hispanic white. Only Medicare, Medicaid [including Medicare- rate,andvehicleaccess)weretimevarying 2% of participants lived in a rural area. “ ” Medicaid dual-eligibles ], or self-pay/ (based on what census tract an individual Participants with food insecurity had a uninsured). We also considered health was living in at the time of the HbA1c greater number of outpatient visits during literacy (17,24) and whether the question- assessment) and updated annually, while the follow-up period. naire wascompletedin EnglishorSpanish. individual-level variables were time fixed Intraclasscorrelationsshowedthat62% Clinical characteristics considered in- fi and collected at baseline. After tting of the variation in HbA1c was explained by cluded age at diabetesdiagnosis, Charlson regression models, we used least squares individual-levelfactors(suchasageorfood comorbidity score (25), the use of insulin meanstocalculatepredictedHbA1c values insecurity), while only 4% of the variation in (of any kind), and 3-hydroxy-3-methyl- adjusted for the factors in the model. HbA1c was explained by census tract–level glutaryl-CoA reductase inhibitor (“statin”) As sensitivity analyses, we fitthesame factors(suchaspovertyratesorlivinginan use. To account for access to and engage- models as our main analysis but using al- area with low physical food access). ment in care, we also considered the ternative indicators for low physical In unadjusted results, mean HbA1c frequency of outpatient care visits during food access. First, we used the CDC’s was 7.86% (62 mmol/mol) in those who the study period. mRFEI as a continuous variable. Next, we reported food insecurity versus 7.30% categorized low physical food access us- (56 mmol/mol) in those who did not Statistical Analysis ing a different distance metric: no sup- (P = 0.001), and 7.21% (55 mmol/mol) in We made unadjusted comparisons using ermarketswithinhalfamileinurbanareas those who lived in a low physical food t tests for continuous variables and x2 and 10 miles in rural areas. Finally, we fit access area versus 7.49% (58 mmol/mol) tests for categorical variables. To analyze a model using a combined food desert in those who did not (P = 0.05). In whether food insecurity or living in a low indicator. hierarchical linear models, adjusted for care.diabetesjournals.org Berkowitz and Associates 1191

Table 1—Demographics and clinical characteristics Food access status Adequate Low Overall Food secure Food insecure P access access P N 391 308 78 271 120 Food insecure (at baseline) 20.21 n/a n/a 24.72 10.08 0.001 Baseline age, years 61.94 (11.90) 63.20 (11.88) 56.72 (10.55) ,0.0001 61.21 (11.63) 63.58 (12.38) 0.08 Female sex 48.08 47.08 51.28 0.51 47.97 48.33 0.95 Race/ethnicity 0.008 0.002 Non-Hispanic white 79.03 81.82 66.67 74.91 88.33 Non-Hispanic black 7.67 5.84 15.38 9.96 2.50 Hispanic 8.95 7.79 14.10 11.44 3.33 Asian/other/multi 4.35 4.55 3.85 3.69 5.83 Education 0.025 0.002 ,High school diploma 14.36 12.01 21.79 17.41 7.50 High school diploma 27.18 29.55 17.95 29.63 21.67 .High school diploma 58.46 58.44 60.26 52.96 70.83 Insurance at baseline 0.02 0.18 Private 50.94 54.83 36.84 48.66 56.36 Medicare 26.42 25.17 30.26 25.67 28.18 Medicaid 13.48 11.03 22.37 15.71 8.18 Uninsured/self-pay 9.16 8.97 10.53 9.96 7.27 Born outside U.S. 20.00 17.86 29.49 0.02 22.96 13.33 0.03 Low 27.58 24.51 38.46 0.01 31.60 18.49 0.008 Took survey in Spanish 5.12 4.22 8.97 0.09 7.01 0.83 0.01 Age at diabetes diagnosis, years 49.74 (14.05) 50.56 (13.92) 46.19 (14.00) 0.02 48.90 (14.06) 51.62 (13.90) 0.08 Baseline Charlson score 4.85 (2.98) 4.74 (2.90) 5.22 (3.34) 0.25 4.89 (2.90) 4.78 (3.17) 0.74 Baseline insulin use 53.96 51.30 64.10 0.04 56.09 49.17 0.21 Baseline statin use 83.38 86.36 70.51 0.001 83.76 82.50 0.76 Baseline ACEi/ARB use 81.33 83.44 73.08 0.04 81.55 80.83 0.87 Number of outpatient visits during study 47.43 (37.54) 44.60 (35.10) 59.08 (45.22) 0.01 50.13 (37.81) 41.26 (36.31) 0.03

Number of HbA1c assessments during study 10.00 (3.83) 9.84 (3.78) 10.48 (3.98) 0.20 10.19 (3.87) 9.45 (3.53) 0.08 Low food access, 1 and 10 miles 30.69 34.74 15.38 0.001 n/a n/a Low food access, half a mile and 10 miles 75.19 78.57 61.54 0.002 n/a n/a Low vehicle access 33.50 32.79 35.90 0.60 27.68 46.67 0.0002 mRFEI 8.30 (7.21) 8.17 (6.78) 8.68 (8.55) 0.69 7.70 (6.91) 9.27 (7.62) 0.09 Median family income of 91,921 96,189 76,104 0.0002 85,808 105,727 ,0.0001 census tract, $ (42,945) (42,746) (40,576) (45,420) (32,958) Census tract poverty rate 12.54 (9.14) 11.34 (8.65) 17.22 (9.69) ,0.0001 14.89 (9.34) 7.24 (5.87) ,0.0001 Data are % or mean (SD) unless otherwise indicated. Food access defined using the 1 mile in urban areas and 10 miles in rural areas definition. Five (1% of total) participants did not have a complete food insecurity assessment. ACEi, ACE inhibitor; ARB, angiotensin receptor blocker; n/a, not applicable. age, sex, race/ethnicity, education, in- full model in Supplementary Table 2). In which adjusted for the same covariates as surance, health literacy, survey language, models adjusted for the same factors, the main analyses, yielded similar results. age at diabetes diagnosis, Charlson living in an area with low physical food Whenphysical foodaccesswas defined by comorbidity score, insulin use, statin access was not associated with an overall the mRFEI, this contextual factor was still use, outpatient visits, food access, vehicle difference in HbA1c compared with living not associated with HbA1c (P =0.56),while access, median family income, and cen- in areas of adequate food access (differ- food insecurity was associated with 0.7% sus tract poverty rate, food insecurity ence0.2%[2.2mmol/mol],95%CI20.2to [7.7mmol/mol]higherHbA1c (P,0.0001) was associated with higher HbA1c (differ- 0.5 [22.2 to 5.5], P = 0.33) and was not compared with food-secure participants ence 0.6% [6.6 mmol/mol], 95% CI 0.4–0.8 associated with change inHbA1c overtime (full model in Supplementary Table 4). [4.4–8.7], P = ,0.0001) (Table 2). There (change in difference per month 20.007% Whenwe used the half milein urban areas was no evidence that the difference [less than 20.1 mmol/mol], P = 0.07) (full and 10 miles in rural area definition, improved over time (estimated change model in Supplementary Table 3). low physical food access was also not in difference per month 20.003% [less Sensitivity analyses using alternative associated with HbA1c (P = 0.91) and than 20.1mmol/mol],P=0.50)(Fig.2and definitions of neighborhood food access, food insecurity was still associated with 1192 Food Insecurity, Food Deserts, and Glycemic Control Diabetes Care Volume 41, June 2018

food insecurity. We did not find evidence thatlivinginanareawithlowphysicalfood access was associated with worse glyce- mic control. These findings extend our understand- ing of food insecurity and food access among patients with diabetes. Prior cross- sectional studies have demonstrated an association between food insecurity and

higher HbA1c (4,21,27,28), but there have been few longitudinal studies. A recent study in Illinois also found an associa- tion between food insecurity and greater HbA1c at baseline, with a slightly greater decrease in HbA1c over time for food- insecure, compared with food-secure, participants (29). However, the Illinois study examined a more limited set of — Figure1 ProportionalVenn diagram showingoverlapbetweenthosewithfoodinsecurityand those potential confounders and did not exam- living in a census tract with low physical food access (1 mile in urban area and 10 miles in rural area definition). Seventy-eight (20%) participants were food insecure, 120 (10%) lived in an area of low ine area-level factors. There have been physical food access, and 12 (3%) were both food insecure and lived in an area of low physical food few previous studies of physical food access. access and HbA1c. One study in the San Francisco Bay Area found that losing or gaining a supermarket in a neighbor- increased HbA1c (0.5% [5.5 mmol/mol], education (an indicator of individual-level P = ,0.0001) (full model in Supple- socioeconomic status) interaction that hood was not associated with meaning- mentary Table 5). Results were essentially wouldconfoundtheresultsreported(Sup- fulchangeinHbA1c (30). Prior studies of unchanged when we fit a model with a plementary Table 8). Sequential model BMI and neighborhood food access in combined indicator of low census tract building revealed that the results were patients with diabetes also have gener- income and physical food access (food robust across several progressively more ally not found an important association – “desert”). The difference for those re- complex modeling specifications (Supple- (31 33). We also think it is important to fi porting food insecurity compared with mentary Table 9). note that our ndings regarding who food security remained significant (0.5% residesinareasoflowphysicalfoodaccess do not correlate well with common percep- [5.5mmol/mol],P=,0.0001),whilethere CONCLUSIONS tions of who lives in these areas, as we found was no significant difference in HbA for 1c In this longitudinal study of patients with fewer racial/ethnic minorities and higher those living in food deserts versus those diabetes, we found that individual-level educational attainment in these areas. This who did not (P = 0.35) (full model in factors, such as food insecurity, explai- is consistentwithnational data showingthat Supplementary Table 6). Examining the ned substantially more of the variation low physical food access and low income distribution of outpatient visit frequency in HbA1c than census tract–level factors, census tracts do not have a high degree of or number of HbA1c assessments did not such as low physical food access. Food overlap (5) and may challenge common suggest that differences in these factors insecurity was associated with ;0.6% perceptions of who lives in a low food between groups would explain the ob- (6.6 mmol/mol) higher HbA1c.Thisdiffer- access census tract. Further, while our served results (Supplementary Table 7). ence did not improve over time despite study took place in an urban context, prior We did not find evidence of a time-by- higheruseofclinicalservicesbythosewith work in rural areas has also found that

Table 2—HbA1c by food insecurity and food access Adjusted baseline Difference Average slope [change per month Difference in change per month

HbA1c, % (95% CI) (95% CI) P in HbA1c, % (95% CI)] in HbA1c, % (95% CI) P Food insecure 7.6 (7.3–7.9) 0.6 (0.4–0.8) ,0.0001 0.0002 (20.01 to 0.01) 20.003 (20.01 to 0.01) 0.50 Food secure 7.0 (6.8–7.3) 0.003 (20.001 to 0.007) Low food 7.4 (7.1–7.7) 0.2 (20.2 to 0.33 20.003 (20.02 to 0.001) 20.007 (20.02 to 0.001) 0.07 access 0.5) Adequate food access 7.3 (7.0–7.5) 0.005 (0.0004–0.009)

Data are means, and the difference in mean HbA1c was calculated using least squares means from the regression models presented in Supplementary Table 2 (for food security) and Supplementary Table 3 (for food access). Food access defined using the 1 mile in urban areas and 10 miles in rural areas definition. In addition to the exposures of interest, models were adjusted for age, age squared, sex, race/ethnicity, education, insurance, health literacy, language, age at diabetes diagnosis, baseline Charlson comorbidity score, insulin use, statin use, number of outpatient care visits, census tract median family income, census tract poverty rate, and census tract vehicle access. care.diabetesjournals.org Berkowitz and Associates 1193

Figure 2—Predicted mean HbA1c values with SE bars from the models described in Supplementary Table 2 (left) and Supplementary Table 3 (right), by 6-month increments, comparing those who reported food insecurity and those who are food secure (left) and those who lived in neighborhoods with adequate and poor food access (right). In addition to the exposures of interest, models were adjusted for age, age squared, sex, race/ethnicity, education, insurance, health literacy, language, age at diabetes diagnosis, baseline Charlson comorbidity score, insulin use, statin use, number of outpatient care visits, census tract median family income, census tract poverty rate, and census tract vehicle access. more deprived areas may, paradoxically, associations with glycemic control, this contextual factors is likely needed (for have closer food access (34). It is worth is only one element of diabetes self- example, both improving food access noting, however, that area poverty is management that physical food access and reducing food insecurity to optimize associated with decreases in supermarket interventions might try to address. Other an individual’s ability to manage diabe- availability over time (35). Finally, this health outcomes, such as weight and men- tes).CutstoFarm Bill–supported nutrition studymayhighlightalimitationofcurrent, tal health, may have stronger associations programs such as SNAP, as are currently geographic information system–based with area food access. In addition, improving proposed(16), would likelymake address- methodsforassessingthefoodenvironment. physical food access could indirectly lead to ing food insecurity more difficult for vul- Measures that incorporate an individual’s health benefits by driving economic growth, nerable patients. subjective view of their food environment generating tax revenue to support public The findings of this study should be may correlate better with health outcomes. services, and improving neighborhood interpreted in light of several limitations. Though the mechanism by which food livability. Food security assessment occurred at a insecurity may be related to higher HbA1c Given their potential broad impact on single time point, and so if the food security is not clear, proposed mechanisms sup- overall health and many other conditions status of participants changed during the portedbypriorstudiesincludelowerdietary besides diabetes per se, we should not study period, this could lead to misclassi- quality (36); increased trade-offs between eschew interventions to improve phys- fication, which may bias results toward the food and medications (37,38), resulting in ical food access. Rather, combining food null. Next, this study was conducted within reduced medication adherence; and in- access interventions with interventions a network in predominately creased psychological distress (39) related that directly address food insecurity may urban eastern Massachusetts. While the to food insecurity. Longitudinal studies be more likely to improve diabetes out- occurrence of both food insecurity and low to examine these relationships may offer comes specifically. Health care systems physical food access was similar to overall insight into how best to address food in- might consider routinely measuring food national averages (3–5), the results may security and improve management of insecurity in those with suboptimal glyce- not generalize to other areas. Next, be- diabetes. mic control and then using strategies to causethestudyreliedonself-report,there Although this study did not find an address food insecurity once it has been is the possibility of recall error. However, association between glycemic control identified. Such strategies might include food insecurity is an inherently subjective and physical food access, this should be assistance with enrollment in SNAP, in- concept, so there is no measurement al- considered within a larger context. Be- terventions that link patients to commu- ternative. Additionally, we used adminis- sides physical food access, there are other nity resources (40), work with local food trative records where possible (e.g., for elements of the food environment that banks (41), or prescriptions for healthy laboratory values, number of clinic visits, may impact self-management of diabetes food (42). Since these interventions rely and insurance information) to minimize and could be studied in future work. Other on having community resources for pa- reliance on self-reported information. considerations in access to healthy foods tients to use, however, there is a clear Area-level data used in the study were include food affordability, food quality, interrelationship between individual- and the most recent available, but changes in and individual food preferences. We also area-level factors that should be con- neighborhood characteristics that were think it is important to note that while sidered. In highly vulnerable neighbor- not captured in the data sets used may physical food access did not have strong hoods, addressing both compositional and have resulted in misclassification. Finally, 1194 Food Insecurity, Food Deserts, and Glycemic Control Diabetes Care Volume 41, June 2018

unmeasured confounding is always a 1.InStandardsofMedicalCareinDiabetesd2017. 18. Harris PA, Taylor R, Thielke R, Payne J, threat to validity in observational studies Diabetes Care 2017;40(Suppl. 1):S6–S10 Gonzalez N, Conde JG. Research electronic data – such as this one. These limitations are 3. Coleman-Jensen A, Rabbitt MP, Gregory CA, capture (REDCap) a metadata-driven methodol- Singh A. Household food security in the United ogy and workflow process for providing trans- balanced by several strengths, however. States in 2015 [Internet], 2016. U.S. Department lational research informatics support. J Biomed This was a longitudinal study, few patients of Agriculture, Economic Research Service. Avail- Inform 2009;42:377–381 were lost to follow-up, and there was ablefromhttps://www.ers.usda.gov/webdocs/ 19. U.S. Department of Agriculture, Economic Re- detailed assessment of both individual- publications/79761/err-215.pdf?v=42636. Accessed searchService.FoodsecurityintheU.S.:Surveytools and area-level factors to permit adjust- 11 August 2017 [Internet]. Available from: https://www.ers.usda 4. Berkowitz SA, Baggett TP, Wexler DJ, Huskey .gov/topics/food-nutrition-assistance/food-security-in- ment for measured confounders. KW, Wee CC. Food insecurity and metabolic con- the-us/survey-tools/#adult. Accessed 11 August 2017 Among patients with diabetes, food trol among U.S. adults with diabetes. Diabetes 20. Seligman HK, Bindman AB, Vittinghoff E, insecurity, but not physical food access, Care 2013;36:3093–3099 Kanaya AM, Kushel MB. Food insecurity is asso- was associated with substantially worse 5. U.S. Department of Agriculture, Economic ciated with diabetes mellitus: results from the glycemic control that did not improve de- Research Service. Food Access Research Atlas: National Health Examination and Nutrition Exam- Documentation [Internet]. Available from https://www ination Survey (NHANES) 1999-2002. J Gen Intern spite higher frequency of outpatient visits. .ers.usda.gov/data-products/food-access-research- Med 2007;22:1018–1023 These findings suggest that interventions atlas/documentation/. Accessed 11 August 2017 21. Seligman HK, Jacobs EA, Lopez´ A, Tschann J, to improve glycemic control in vulnera- 6. Christine PJ, Auchincloss AH, Bertoni AG, et al. Fernandez A. Food insecurity and glycemic control ble patients with diabetes should consider Longitudinal associations between neighborhood among low-income patients with type 2 diabetes. – addressing food insecurity systematically physical and social environments and incident Diabetes Care 2012;35:233 238 type 2 diabetes mellitus: the Multi-Ethnic Study of 22. Essien UR, Shahid NN, Berkowitz SA. Food and highlight the importance of distinguish- Atherosclerosis (MESA). JAMA Intern Med 2015; insecurity and diabetes in developed societies. ing between compositional and contextual 175:1311–1320 Curr Diab Rep 2016;16:79 factors that may contribute to poor health. 7.ZenkSN,TarlovE,WingC,etal.Geographic 23. Centers for Disease Control and Prevention. accessibility of food outlets not associated with CensustractlevelstatemapsoftheModifiedRetail change among veterans, 2009- Food Environment Index (mRFEI) [Internet], 2011. 14. Health Aff (Millwood) 2017;36:1433–1442 Available from ftp://ftp.cdc.gov/pub/Publications/ Funding. The role of S.A.B. in the research 8. LarsonNI, StoryMT, NelsonMC.Neighborhood dnpao/census-tract-level-state-maps-mrfei_TAG508. reported in this publication was supported by environments: disparities in access to healthy pdf. Accessed 11 August 2017 the National Institute of Diabetes and Digestive foods in the U.S. Am J Prev Med 2009;36:74–81 24. Sarkar U, Schillinger D, Lopez´ A, Sudore R. and Kidney Diseases of the National Institutes of 9. Malambo P, Kengne AP, De Villiers A, Lambert Validation of self-reported health literacy ques- Health (K23-DK-109200). The role of S.A.A. was EV, Puoane T. Built environment, selected risk tions among diverse English and Spanish-speaking supportedbyT32DK007028-38/BAEDRC5P30DK057521. factors and major outcomes: populations. J Gen Intern Med 2011;26:265–271 The role of G.C.-S. was supported by the National a systematic review. PLoS One 2016;11:e0166846 25. Charlson ME, Charlson RE, Peterson JC, Heart, Lung, and Blood Institute (2K24 HL105493-06- 10. Smalls BL, Gregory CM, Zoller JS, Egede LE. Marinopoulos SS, Briggs WM, Hollenberg JP. The 07 and 1R01HL120690-03). The role of H.K.S. was Assessing the relationship between neighborhood Charlson comorbidity index is adapted to predict supported by the National Institute of Diabetes factors and diabetes related health outcomes and self- costs of chronic disease in primary care patients. J and Digestive and Kidney Diseases of the National care behaviors. BMC Health Serv Res 2015;15:445 Clin Epidemiol 2008;61:1234–1240 Institutes of Health (P30-DK-092924). 11. Unger E, Diez-Roux AV, Lloyd-Jones DM, et al. 26. BellBA,EneM,SmileyW,SchoenebergerJA.A The content is solely the responsibility of the Association of neighborhood characteristics with multilevel model primer using SAS PROC MIXED authors and does not necessarily represent the cardiovascular health in the multi-ethnic study of [Internet], 2013. Available from http://support official views of the National Institutes of Health. atherosclerosis. Circ Cardiovasc Qual Outcomes .sas.com/resources/papers/proceedings13/433- Duality of Interest. No potential conflicts of 2014;7:524–531 2013.pdf. Accessed 11 August 2017 interest relevant to this article were reported. 12. DubowitzT,Ghosh-DastidarM, EibnerC, etal. 27. Lyles CR, Wolf MS, Schillinger D, et al. Food Author Contributions. S.A.B. and D.J.W. con- The Women’s Health Initiative: the food environment, insecurity in relation to changes in hemoglobin ceived of the study and drafted the manuscript. neighborhood socioeconomic status, BMI, and blood A1c, self-efficacy, and fruit/vegetable intake dur- S.A.A. and L.S.B. assisted with collection and pressure. Obesity (Silver Spring) 2012;20:862–871 ing a diabetes educational intervention. Diabetes interpretation of data and revised the manuscript 13. Mujahid MS,DiezRoux AV, Morenoff JD, etal. Care 2013;36:1448–1453 critically for intellectual content. A.J.K., G.C.-S., Neighborhood characteristics and . 28. MayerVL,McDonoughK,SeligmanH,MitraN, H.K.S., and S.J.A. assisted with analysis and inter- Epidemiology 2008;19:590–598 Long JA. Food insecurity, coping strategies and pretation of data and revised the manuscript criti- 14. FleischhackerSE,FlournoyR,MooreLV.Mean- glucose control in low-income patients with dia- callyforintellectualcontent.Allauthorsgaveapproval ingful, measurable, and manageable approaches to betes. Public Health Nutr 2016;19:1103–1111 of the manuscript version to be submitted. S.A.B. is evaluating healthy food financing initiatives: an 29. Shalowitz MU, Eng JS, McKinney CO, et al. theguarantor of this work and, as such, had full access overview of resources and approaches. J Public Food security is related to adult type 2 diabetes to all the data in the study and takes responsibility Health Manag Pract 2013;19:541–549 control over time in a United States safety net for the integrityof the dataand the accuracy of the 15. Center on Budget and Policy Priorities. More primarycareclinicpopulation.NutrDiabetes2017; data analysis. than500,000adultswillloseSNAPbenefitsin2016 7:e277 Prior Presentation. Interim results of this study as waivers expire [Internet], 2015. Available from 30. Zhang YT, Mujahid MS, Laraia BA, et al. were presented in abstract form at the 75th https://www.cbpp.org/research/food-assistance/ Association between neighborhood supermarket Scientific Sessions of the American Diabetes more-than-500000-adults-will-lose-snap-benefits- presence and glycated hemoglobin levels among Association, Boston, MA, 5–9June2015. in-2016-as-waivers-expire. Accessed 1 February 2018 patients with type 2 diabetes mellitus. Am J 16. Center on Budget and Policy Priorities. Pres- Epidemiol 2017;185:1297–1303 ident’s budget would radically restructure, cut 31. Laraia BA, Downing JM, Zhang YT, et al. Food References SNAP [Internet], 2017. Available from https://www environment and weight change: does residential 1. Centers for Disease Control and Prevention. .cbpp.org/blog/presidents-budget-would-radically- mobility matter? The Diabetes Study of Northern National Diabetes Statistics Report, 2017 [Internet]. restructure-cut-snap. Accessed 15 February 2018 California (DISTANCE). Am J Epidemiol 2017;185: Availablefromhttps://www.cdc.gov/diabetes/pdfs/ 17. Berkowitz SA, Meigs JB, DeWalt D, et al. 743–750 data/statistics/national-diabetes-statistics-report Material need insecurities, control of diabetes 32. Zhang YT, Laraia BA, Mujahid MS, et al. Is a .pdf. Accessed 10 August 2017 mellitus, and use of health care resources: results reduction in distance to nearest supermarket 2. American Diabetes Association. Promoting of the Measuring Economic Insecurity in Diabetes associatedwithBMIchangeamongtype2diabetes health and reducing disparities in populations. Sec. study. JAMA Intern Med 2015;175:257–265 patients? Health Place 2016;40:15–20 care.diabetesjournals.org Berkowitz and Associates 1195

33. Zhang YT, Laraia BA, Mujahid MS, et al. Does glycemic control in diabetes: results from the food insecurity and diabetes in Canada. Diabetes food vendor density mediate the association between Boston Puerto Rican Health study. Diabetes Care Care 2009;32:2218–2224 neighborhood deprivation and BMI? A G-computation 2014;37:2587–2592 40. Berkowitz SA, Hulberg AC, Standish S, Reznor mediation analysis. Epidemiology 2015;26:344–352 37. Berkowitz SA, Seligman HK, Choudhry NK. G,AtlasSJ.Addressingunmetbasicresourceneeds 34. Sharkey JR, Horel S. Neighborhood socioeco- Treat or eat: food insecurity, cost-related medication as part of chronic cardiometabolic disease man- nomic deprivation and minority composition are underuse, and unmet needs. Am J Med 2014;127: agement. JAMA Intern Med 2017;177:244–252 associated with better potential spatial access to 303–310.e3 41. SeligmanHK,LylesC,MarshallMB,etal.Apilot the ground-truthed food environment in a large 38. Billimek J, Sorkin DH. Food insecurity, pro- foodbank interventionfeaturing diabetes-appropriate rural area. J Nutr 2008;138:620–627 cesses of care, and self-reported medication under- food improved glycemic control among clients in three 35. Chen H-J, Wang Y. The changing food outlet use in patients with type 2 diabetes: results from the states. Health Aff (Millwood) 2015;34:1956–1963 distributions and local contextual factors in the California Health Interview Survey. Health Serv Res 42. Goddu AP, Roberson TS, Raffel KE, Chin MH, United States. BMC Public Health 2014;14:42 2012;47:2159–2168 Peek ME. FoodRx:acommunity-universitypartnership 36. BerkowitzSA, Gao X, TuckerKL. Food-insecure 39. Gucciardi E, Vogt JA, DeMelo M, Stewart DE. toprescribehealthyeatingontheSouthSideofChicago. dietary patterns are associated with poor longitudinal Exploration of the relationship between household J Prev Interv Community 2015;43:148–162