Neighborhood built and food environment in relation to glycemic control in people with type 2 diabetes in the Moving to Health Study

Abstract

Objective:

To examine whether built environment and food metrics are associated with glycemic control in people with type 2 diabetes.

Research

Design and Methods: We included 14,985 patients with type 2 diabetes using electronic health records from Kaiser Permanente Washington. Patient addresses were geocoded with ArcGIS using King County and Esri reference data. Built environment exposures estimated from geocoded locations included residential unit density, transit threshold residential unit density, park access, and having supermarkets and fast food restaurants within 1600-meter Euclidean buffers. Linear mixed effects models compared mean changes of HbA1c from baseline at 1, 3 (primary) and 5 years by each built environment variable.

Results:

Patients (mean age = 59.4 SD = 13.2, 49.5% female, 16.6% Asian, 9.8% Black, 5.5% Latino/Hispanic, 57.1% White, 20% insulin dependent, mean BMI = 32.7 +/− 7.7) had an average of 6 HbA1c measures available. Participants in the 1^st tertile of residential density (lowest) had a greater decline in HbA1c (−0.42, −0.43, and −0.44 in years 1, 3, and 5 respectively) than those in the 3^rd tertile (HbA1c = −0.37 at 1- and 3-years and −0.36 at 5-years; all p-values < 0.05). Having any supermarkets within 1600 meters of home was associated with a greater decrease in HbA1c at 1-year and 3-years compared to having none (all p-values < 0.05).

Conclusions:

Lower residential density and better proximity to supermarkets may benefit HbA1c control in people with people with type 2 diabetes. However, effects were small and indicate limited clinical significance.

Introduction

Type 2 diabetes prevalence is increasing among U.S. adults [1] with 11.3% of the U.S. population estimated to have type 2 diabetes. Type 2 diabetes is the highest cost chronic condition in the U.S [1]. Glycemic control helps to prevent complications including eye disease, kidney disease, nerve damage, stroke, and heart disease as over 50% of lifetime medical costs due to type 2 diabetes are related to disease complications [2].

Healthy diet, physical activity and weight management are cornerstones of effective diabetes self-management in conjunction with pharmacologic interventions [3–8]. In addition, social determinants such as living environment may contribute to diabetes management [8]. An important part of living environments is the neighborhood built environment, defined as physical, contextual characteristics of where people live, work and play [9]. Indeed, the built environment has been associated with diet and physical activity [10] and body weight [11] as well as health outcomes including hypertension and type 2 diabetes [12–15]. The built environment could influence glycemic control [16] through its associations with physical activity and diet, as well as through other factors that affect type 2 diabetes management such as stress and sleep [8].

Limited evidence has associated the built environment to glycemic control measured with hemoglobin A1c (HbA1c) values (lower is better). A New York City diabetes registry study found that living in, or moving to, ZIP codes with more access to “neutral” or “healthy” food outlets and higher walkability was associated with lower HbA1c over 7 years [17]. A key limitation of the study was geocoding at the ZIP code level and the inability to disentangle which specific built environment variables were most associated with glycemic control. In a Kaiser Permanente Northern California diabetes registry, loss of supermarket availability within 1 mile of a residential census block was associated with higher HbA1c over 4 years, particularly for those with moderate or poor glycemic control at baseline [18]. Paradoxically, gaining a supermarket was also associated with worse HbA1c among people with poor glycemic control at baseline [18].

It remains unclear whether specific built environment variables, including residential density, proximity to supermarkets, and proximity to parks, are associated with glycemic control among people with type 2 diabetes. The goal of this study was to leverage access to historical electronic health record (EHR) and residential address data to examine whether exposure to different neighborhood built environment features was associated with longitudinal changes in HbA1c in people with type 2 diabetes over long-term follow-up independent of socioeconomic status, demographics, and severity of diabetes. Overcoming limitations in prior studies, we measured built environment exposures at the address level rather than ZIP code or census tract level and included examination of parks and residential density. We hypothesized that greater residential density, proximity to parks, and food access (e.g., lower fast food outlet density and higher supermarket density) would be associated with better HbA1c, after accounting for socioeconomic status (as measured by individual-level residential property values).

Methods

Moving to Health Study Overview

The Moving to Health (M2H) Study was a retrospective cohort of Kaiser Permanente Washington (formerly Group Health) members. Data were captured through access to EHRs from 1/1/2005 to 4/30/2017 (study period). The Kaiser Permanente Washington (KPWA) institutional review board approved the study which included waiver of informed consent. Details on the cohort were previously published [11, 19, 20].

Inclusion and exclusion criteria

All included patients were KPWA members residing in King County, WA, ages 18 – 89 years at baseline, and had at least 365 days of continuous enrollment allowing for an enrollment gap of no more than 92 days during the study period. Only people with type 2 diabetes were included, defined as having an HbA1c value of ≥ 6.5% on any single occasion or at least one outpatient prescription fill of an oral diabetes medication or insulin (and they could not be pregnant at that time). If the only prescription fill was for metformin, a diagnosis of type 2 diabetes was required in the 6 months prior to, or after, the fill date. To be included, patients had to have at least 2 HbA1c values between 1/1/2005 and before 4/30/2017. Baseline was considered as the date of their first eligible HbA1c measure after entry into the cohort. Eligible HbA1c included a measure that occurred while residing at a geocodable King County address (described below), during a period of continuous enrollment, with a valid and non-missing BMI, with a valid and non-missing property value, and with non-missing built environment variables. People were excluded from analysis if, at their baseline HbA1c measure, they had cancer other than non-melanomatous skin cancer within the year prior, bariatric surgery within the prior year, were pregnant or within 3-months of delivery, had implausible (<15 or > 100) or missing BMI values, or did not have sex recorded in their EHR (see definition below in the covariates section). The cohort was limited to KPWA members with addresses that could be geocoded and for which residential property values could be computed. Censoring events included measurements taken 1 year prior to death, change in residential location, bariatric surgery, cancer diagnosis, gap in enrollment of 13 months or more, turning 90 years of age, or end of the study period.

Built Environment Exposures

Geocoding.

Addresses were initially geocoded in ArcGIS using King County rooftop-level precision address point reference data [21]. With this approach, 93.0% of addresses in the original cohort were successfully geocoded [11]. For addresses that did not match using the King County address point data, Esri Business Analyst reference data were used with a match threshold of 85%, resulting in an additional 1.5% of matched addresses. We excluded the 5.6% of members whose addresses were incorrectly recorded, were post office boxes, or could not be geocoded for other reasons. SmartMaps [22] were used to measure built environment exposures for each geocoded address point. SmartMaps are geographic information system (GIS) grid layers that represent spatially continuous neighborhood-level built environment measures. SmartMaps essentially produce GIS Euclidean buffer measures for specified radii around every grid cell location within a study area produced using moving-window procedures. SmartMaps’ front-loading of the GIS measurement process allows for rapid estimation of built environment characteristics for large point data sets, such as the current cohort’s geocoded addresses. Additionally, SmartMaps can be created in non-secure environments and then applied to sensitive data, such as EHRs within a health organization’s firewall. For residential density and park exposure variables, SmartMaps with buffer radii of 800- and 1600-meters (primary) were created. We selected these buffer sizes as they represent roughly a 10-minute and 20-minute walk respectively from home. Given that evidence suggests pedestrians only walk within a kilometer from home [23], we selected buffers that were slightly below and above this threshold. For food environment exposure variables, SmartMaps with buffer radii of 1600-meters (primary) as well as 5000-meters were created. We used these buffers to represent an area people could walk to (1600 meters) and a buffer in easy driving distance (5000 meters) [20]. We used driving distance for food environment exposures because research suggests the majority of the population drives to purchase food [24].

Density measures.

Population density, residential unit density, and street intersection density were initially considered for inclusion as proxy measures of walkability as more dense areas contain access to services, destinations, and transportation [23] and are predictive of walking [25]. However, the three metrics were highly correlated (Spearman r’s > 0.76) for the primary buffer of 1600 meters. Therefore, we only included residential unit density (henceforth referred to as residential density). Residential density was calculated as units per hectare. A hectare is equivalent to 10,000 square meters or 2.471 acres [26]. Residential density was categorized by tertiles for each buffer such that the first tertile indicated the lowest residential density. Residential density was also dichotomized by whether it exceeded the conventional cutoff of 18 residential units per hectare [27–29] to support efficient transit (referred to as transit threshold residential density henceforth). This cutoff has been associated with lower BMI and waist circumference [29]. Within King County, areas with higher than 18 residential units per hectare tend to be within the City of Seattle or denser inner suburbs.

Parks.

Total area of parks with a slope below a threshold of 5% (which indicates the portion of the park that is accessible or usable) was examined. Parks data were developed by the UW Urban Form Lab based on 2008 data from GIS data on parks from all jurisdictions in King County, WA [30], and combined with slope data from the National Elevation Dataset [31]. Tertiles of accessible park area were used as the exposure variable such that the lowest tertile indicated the lowest accessible park area. Tertiles of park counts were used as the exposure variable such that the lowest tertile indicated the fewest parks.

Food environment.

Fast food and supermarket counts were obtained from Public-Health Seattle King County food permit data [32]. Fast food included restaurants where food is ordered and paid for at the same time. Counts of fast food and supermarkets were dichotomized at the 1600-meter buffer to any or none; for 5000 meter buffers, both food sources were categorized by tertiles with the first tertile indicating the least fast food and fewest supermarkets.

Outcome Measures of Glycemic Control

We used HbA1c as the marker of glycemic control. HbA1c measures are taken regularly as part of the clinical care of KP members with diabetes. All KP laboratories, where the majority of HbA1c values are assessed, are CLIA certified. The primary outcome was change in HbA1c from baseline to 3-years of follow-up. We also examined change in HbA1c over 1- and 5-years of follow-up.

Covariates

Covariates included individual sociodemographic factors (age, sex, race and ethnicity, parcel-level home property value, insurance type) and health factors (e.g., insulin use, sleep apnea) that have been associated with glycemic control in other research [35, 36]. Baseline age from the EHR was measured as natural cubic splines with 10 degrees of freedom. The EHR defined sex as male or female according to self-report. No other measures of sex or gender were available during our study period. We measured race and ethnicity with a self-report in the EHR of whether a person considers themselves non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, Hispanic, Hawai’ian/Pacific Islander, Native American/Alaska Native, multiracial, or other racial or ethnic identities according to Census Bureau definitions. Race and ethnicity was used as a partial proxy for exposure to environmental and systemic racism which can affect access to health care, physical activity, and diet, all of which can impact weight and glycemic control. Insurance type was measured categorically (Medicaid, Medicare, Commercial, other). Individual parcel-level home property value [37] was measured in deciles by year (merging 2016 and 2017). We included the number of non-insulin diabetic medication classes participants were taking at baseline (0, 1, 2, or 3+) including sulfonylurea, metformin, and thiazolidinediones; we also included whether someone was taking insulin or not. Baseline BMI was used as a covariate and modelled using splines. For those missing baseline BMI (measured on the same date as the HbA1c), we imputed BMI based on whether there was a measure available in the year prior to the baseline measure or, if not, up to 1-year forward (13% of the sample had an imputed BMI based on this procedure). ICD-9 or 10 diagnosis codes were used to measure depression, anxiety, and sleep apnea diagnoses in the year prior to baseline HbA1c values. Smoking was self-reported at baseline as current, former, or never smoker. Finally, an indicator of whether someone had incident or prevalent diabetes at baseline was included.

Statistical analysis

We compared adjusted mean changes of HbA1c from baseline at 1, 3 (primary) and 5 years by each built environment variable applying linear mixed effects models [38]. We considered mean models of the form:

$$E\left(Y_{it}\right)=f_0\left(t\right)+{\sum }_{k=1}^{K}B{E}_{ik}\times f_k\left(t\right)+\gamma Z_i,$$

where $Y_{\mathrm{i}\mathrm{t}}$ is the change in HbA1c from baseline to time $t$ for person $i$, and $Z_i$ denotes a vector of person-specific baseline covariates. $B{E}_{ik}$ represents the built environment variable at baseline for person $i$ at the $k^{\text{th}}$ level of the categorical variable ($K$ is equal to 2 for built environment variables categorized into tertiles and 1 for binary indicators). The functions $f_0\left(t\right)$ and $f_k\left(t\right)$ denote longitudinal trajectories of the change in HbA1c for each of the $K+1$ levels of built environment. We flexibly modeled $f_0\left(t\right)$ and $f_k\left(t\right)$ using natural cubic splines [39, 40] with 5 degrees of freedom and interacted these terms with the built environment categories to allow the HbA1c change trajectories to vary across built environment categories. To account for within-person correlation arising from repeated measures at unequally spaced observations, we assumed an exponential correlation structure for the random effects. This correlation structure permits the correlation between follow-up HbA1c measures to decrease as the time between measures increases. All models adjusted for baseline HbA1c (as natural cubic splines with 5 degrees of freedom) and the covariates listed above. The models for the food environment variables also adjusted for residential density (as tertiles) at the same buffer size. We conducted sensitivity analyses that additionally adjusted for self-reported smoking on the subset of individuals with recorded smoking information.

From these models we estimated adjusted mean changes and 95% confidence intervals at a given time point (1, 3, and 5 years) in each built environment category assuming mean levels of all other covariates. We conducted omnibus tests at each time point comparing the HbA1c change at the 3^rd vs. 1^st tertile (or none vs. any or below or above the transit threshold of residential density) of built environment and obtained corresponding p-values. The omnibus tests are exploratory. Accordingly, we did not implement formal multiple comparisons corrections and recommend caution when interpreting results.

Effect Modification

We examined whether relationships between HbA1c change and built environment variables differed by age or use of insulin or not. For these effect modification analyses, we used linear mixed models of the same form as the main analysis model, but further included the following interactions of the effect modification variable (e.g. age categorized into 3 categories or insulin vs. not) with: 1) the $f_k\left(t\right)$, 2) the built environment variable, and 3) the interactions of $f_k\left(t\right)$ and the built environment variable. This allows the HbA1c change trajectories across baseline built environment categories to vary by the effect modifier. We also conducted omnibus tests that assessed if there were any differences in HbA1c change at the 3^rd vs. 1^st tertile (or none vs. any or below vs. above the residential density transit threshold) of a given built environment variable across effect modifier levels at a given time point. As an example, for insulin, we compared if the difference in the estimated adjusted mean change in HbA1c between the 3rd and 1st tertile was statistically different among those taking insulin versus those not taking insulin 3 years after baseline. Results

Participants

There were 25,100 Kaiser Permanente Washington patients identified with type 2 diabetes (see Figure 1). After excluding HbA1c values that did not meet inclusion criteria, 14,985 patients were identified with eligible data for the analyses, including 129,686 HbA1c measures. The median number of HbA1c measures available per participant was 6 (IQR = 3, 12). Almost one-third of the sample had HbA1c of 8% or greater at baseline. The sample was 49.5% female, 16.6% Asian, 9.8% Black, 5.5% Latino/a/Hispanic, 57.1% White, 36% had Medicare, 8.4% were current smokers, and nearly 20% of the sample used insulin (see Table 1). Mean BMI was 32.7 (SD = 7.7, IQR = 27.2–36.8). Table 2 shows the distribution of baseline built environment variables by mean baseline HbA1c and among those with worse glycemic control without adjustment. There were no clear patterns in the distributions.

Figure 1.

Patient data flow diagram

^a Initial sample of 25,110 KPW patient members with diabetes, aged 18–89 at baseline, having at least 270 days of continuous enrollment between January 2005 and April 2017 and residing in King County, Washington.

^b Apparent data entry error

^c An ‘eligible baseline HbA1c’ is defined as meeting all of the following criteria at the time of the HbA1c: • HbA1c occurring while patient lives in KC with a geocodable address

• HbA1c has continuous enrollment in the prior year 1 with allowable gap of 92 days

• HbA1c has valid and non-missing BMI measure associated with it

• HbA1c has valid and non-missing baseline property value

• HbA1c has baseline residential density

• HbA1c has baseline park variables

^d After baseline means on a date after the date of the baseline HbA1c measure. One person out of the 3,439 people excluded had an additional HbA1c measure on the date of their baseline HbA1c measure and this person was also excluded.

Table 1.

Baseline sample characteristics, follow-up time, and number of HbA1c measurements

	N	%	Follow-up Time in Years (IQR))	No. of HbA1c Measures (IQR)	% HbA1c >= 7	% HbA1c >= 8
Overall
	14985	100.0	2.9 (1.2, 5.9)	6 (3, 12)	58.2	31.5
Sex
Female	7414	49.5	2.8 (1.2, 5.8)	6 (3, 11)	55.2	28.7
Male	7571	50.5	2.9 (1.2, 6)	6 (3, 12)	61.1	34.3
Age categories (years)
18 to 29	284	1.9	1.7 (0.7, 3.4)	4 (2, 6)	75.4	54.6
30 to 44	1599	10.7	1.9 (0.8, 4.2)	4 (2, 7)	63.5	40.8
45 to 54	3123	20.8	2.7 (1.1, 5.9)	5 (3, 11)	63.8	38.4
55 to 64	4830	32.2	3.1 (1.2, 6.3)	6 (3, 12)	58.3	31.8
65 to 74	3162	21.1	3.6 (1.5, 7)	8 (4, 15)	51.7	24.1
75+	1987	13.3	2.7 (1.2, 5.2)	6 (3, 12)	52.4	21.3
Race and ethnicity
Asian	2489	16.6	3 (1.2, 6)	6 (3, 11)	55.5	28.0
Black	1470	9.8	2.9 (1.1, 5.9)	6 (3, 11)	57.3	32.8
Hawai’ian / Pacific Islander	215	1.4	2.1 (1.1, 4.7)	4 (3, 8)	71.2	44.2
Hispanic	824	5.5	2.6 (1.1, 5.5)	5 (3, 11)	60.8	35.1
Native American / Alaskan Native	228	1.5	3.2 (1.4, 6.1)	6 (4, 11)	65.8	39.9
Other	184	1.2	2.7 (1.1, 6.1)	6 (3, 11)	58.7	33.7
Unknown	1016	6.8	1.6 (0.8, 3.3)	4 (2, 7)	67.2	38.7
White	8559	57.1	3.1 (1.2, 6.2)	6 (3, 13)	57.2	30.6
Insurance type
Commercial	9172	61.2	2.8 (1.1, 5.9)	6 (3, 11)	61.6	36.2
Medicaid	413	2.8	1.3 (0.7, 2.7)	4 (2, 6)	55.7	32.7
Medicare	5371	35.8	3.2 (1.3, 6.1)	7 (4, 14)	52.5	23.6
Other	29	0.2	1 (0.4, 1.4)	3 (2, 4)	55.2	24.1
BMI categories
15 to 18.4	43	0.3	2.8 (1.3, 4.2)	5 (4, 10)	34.9	23.3
18.5 to 24.9	1943	13.0	2.7 (1.1, 5.7)	6 (3, 11)	56.7	30.4
25.0 to 29.9	4313	28.8	3 (1.2, 6.1)	6 (3, 12)	55.8	29.3
30 to 34.9	3900	26.0	3 (1.2, 6.1)	6 (3, 12)	59.0	32.4
35 to 39.9	2455	16.4	2.8 (1.1, 5.8)	6 (3, 12)	61.1	33.2
40.0 or more	2331	15.6	2.7 (1.1, 5.7)	5 (3, 11)	59.8	33.7
Property Value Deciles
1	1499	10.0	1.8 (0.8, 3.6)	4 (3, 8)	63.3	37.5
2	1496	10.0	2.2 (1, 4.8)	5 (3, 10)	61.4	34.7
3	1499	10.0	2.9 (1.2, 5.9)	6 (3, 13)	59.5	32.2
4	1502	10.0	3 (1.2, 6.1)	6 (3, 12)	60.4	33.7
5	1484	9.9	3 (1.2, 6.1)	6 (3, 12)	59.8	31.9
6	1498	10.0	3 (1.2, 6.3)	6 (3, 12)	57.4	30.6
7	1522	10.2	3.4 (1.4, 6.6)	7 (3, 13)	57.5	30.7
8	1487	9.9	3.3 (1.2, 6.5)	6 (3, 13)	56.1	30.3
9	1497	10.0	3.3 (1.4, 6.4)	6 (3, 13)	54.6	28.4
10	1501	10.0	3.4 (1.3, 6.6)	6 (3, 13)	51.6	25.3
Smoking status
Current, Self-Report	1259	8.4	2.1 (0.9, 4.7)	5 (3, 9)	61.0	37.0
Former, Self-Report	3165	21.1	2.5 (1.1, 4.9)	5 (3, 10)	54.5	28.2
Never, Self-Report	6466	43.1	2.7 (1.1, 5.4)	6 (3, 11)	56.7	30.2
Depression	266	1.8	1.3 (0.7, 2.8)	3 (2, 5)	52.6	32.3
Anxiety	299	2.0	1.7 (0.9, 3.7)	4 (3, 8)	48.8	27.1
Mood Disorder (depression and anxiety)	517	3.5	1.6 (0.8, 3.4)	4 (2, 7)	52.2	30.0
Sleep Apnea	80	0.5	1.8 (0.7, 3.8)	4 (2, 6)	42.5	27.5
Insulin use	2928	19.5	2.7 (1.1, 5.7)	6 (3, 13)	80.3	53.9
No. of Diabetes Meds (Excluding Insulin)
0	7719	51.5	2.8 (1.1, 5.7)	5 (3, 10)	49.7	23.7
1	5039	33.6	2.8 (1.2, 5.9)	6 (3, 13)	62.5	36.0
2	2170	14.5	3.4 (1.3, 7)	8 (4, 16)	77.9	48.6
3+	57	0.4	2.2 (0.8, 4.6)	6 (3, 10)	71.9	50.9

BMI = body mass index

Table 2.

Baseline built environment characteristics by baseline HbA1c value and HbA1c of 7% or 8% or higher

	N	%	Mean HbA1c	% HbA1c >= 7	% HbA1c >= 8
Overall
	14985	100.0	7.7 (1.7)	58.2	31.5
Residential density, 1600m
[0,5.56] units per hectare	4995	33.3	7.7 (1.6)	58.6	31.2
(5.56,8.61] units per hectare	4995	33.3	7.8 (1.7)	58.8	32.9
(8.61,76.3] units per hectare	4995	33.3	7.7 (1.7)	57.1	30.6
Transit threshold residential density, 1600m
[ 0,18) units per hectare	14058	93.8	7.7 (1.7)	58.3	31.6
[18,85] units per hectare	927	6.2	7.7 (1.8)	55.7	31.1
Accessible park area, 1600m
[0,10.3] hectares	4995	33.3	7.7 (1.7)	58.3	31.2
(10.3,22.3] hectares	5000	33.4	7.7 (1.7)	58.2	31.9
(22.3,240] hectares	4990	33.3	7.7 (1.7)	58.0	31.5
Count of parks, 1600m
[0,7] hectares	5372	35.8	7.7 (1.7)	58.2	31.8
(7,11] hectares	4925	32.9	7.7 (1.7)	58.7	32.0
(11,44] hectares	4688	31.3	7.7 (1.7)	57.4	30.8
Fast food count, 1600m
None	5903	39.4	7.7 (1.6)	57.1	31.0
Any	9082	60.6	7.7 (1.7)	58.9	31.9
Supermarket count, 1600m
None	7119	47.5	7.7 (1.7)	57.5	30.9
Any	7866	52.5	7.7 (1.7)	58.8	32.1

Primary and secondary models

Mean HbA1c change in the population was −0.38 (95% CI: −0.41 to −0.35) over 1 year, −0.38 (95% CI: −0.41 to −0.35) over 3 years, and −0.39 (95% CI: −0.42 to −0.36) over 5 years. In the primary models (1600-meter buffers) for residential density, transit threshold, and parks, only differences between the 3^rd and 1^st tertile of residential density were significantly associated with HbA1c values over 1, 3, and 5 years (p’s < .05; see Table 3). Living in neighborhoods with lower residential density was associated with greater HbA1c declines (−0.42, −0.43, and −0.44 at years 1, 3, and 5 respectively) than living in the highest (HbA1c = −0.37 at 1- and 3-years and −0.36 at 5 years). For food environment variables, having any supermarkets within 1600 meters of home was associated with decreases in HbA1c of −0.40 at 1-year and 3-years compared to −0.36 at 1-year and −0.35 at 3-years among those with no supermarkets. In the secondary buffer size models, findings were attenuated and residential density comparisons became non-significant at 1-year while the findings for supermarkets became non-significant at 3-years (see Supplementary Table 1). In sensitivity analyses which additionally adjusted for smoking status, results were similar to the primary and secondary findings (Supplementary Table 2).

Table 3.

HbA1c change with 95% confidence intervals and p-values over 1−, 3−, and 5-years from baseline

	1-year Estimate	p-value	3-year Estimate	p-value	5-year Estimate	p-value
HbA1c	−0.38 (−0.41, −0.35)	n/a	−0.38 (−0.41, −0.35)	n/a	−0.39 (−0.42, −0.36)	n/a
Residential Density 1600m (units per hectare)
[0,5.56]	−0.42 (−0.46, −0.38)		−0.43 (−0.47, −0.39)		−0.44 (−0.48, −0.39)
(5.56,8.61]	−0.36 (−0.40, −0.32)		−0.33 (−0.38, − 0.29)		−0.36 (−0.41, − 0.32)
(8.61,76.3]	−0.37 (−0.41, −0.33)	0.049	−0.37 (−0.41, −0.32)	0.016	−0.36 (−0.41, −0.31)	0.013
Transit Threshold Residential Density 1600m (units per hectare)
[ 0,18)	−0.38 (−0.41, −0.36)		−0.38 (−0.41, −0.35)		−0.39 (−0.42, −0.36)
[18,85]	−0.34 (−0.42, −0.25)	0.25	−0.32 (−0.41, −0.23)	0.18	−0.33 (−0.43, −0.22)	0.24
Fast Food Count 1600m
None	−0.37 (−0.41, −0.33)		−0.37 (−0.41, −0.33)		−0.37 (−0.42, −0.33)
Any	−0.39 (−0.43, −0.36)	0.28	−0.38 (−0.42, −0.35)	0.63	−0.40 (−0.44, −0.36)	0.26
Supermarket Count 1600m
None	−0.36 (−0.39, −0.32)		−0.35 (−0.39, −0.32)		−0.37 (−0.41, −0.33)
Any	−0.40 (−0.44, −0.37)	0.033	−0.40 (−0.44, − 0.37)	0.034	−0.40 (−0.45, − 0.36)	0.16
Park Count 1600m
[0,7]	−0.40 (−0.44, −0.36)		−0.40 (−0.44, −0.36)		−0.41 (−0.45, −0.36)
HbA1c	−0.38 (−0.41, −0.35)	n/a	−0.38 (−0.41, −0.35)	n/a	−0.39 (−0.42, −0.36)	n/a
(11,44]	−0.38 (−0.43, −0.34)		−0.37 (−0.41, −0.33)		−0.40 (−0.45, −0.35)
(7,11]	−0.36 (−0.40, −0.32)	0.62	−0.36 (−0.40, −0.32)	0.27	−0.35 (−0.40, −0.31)	0.86
Accessible Park Area 1600m (hectares)
[0,10.3]	−0.39 (−0.43, −0.35)		−0.38 (−0.43, −0.34)		−0.37 (−0.41, −0.32)
(10.3,22.3]	−0.36 (−0.40, −0.32)		−0.35 (−0.39, −0.31)		−0.37 (−0.42, −0.33)
(22.3,240]	−0.39 (−0.43, −0.35)	0.84	−0.40 (−0.44, −0.36)	0.52	−0.42 (−0.47, −0.38)	0.055

Effect modification analyses

We examined whether insulin use at baseline modified associations between the built environment variables at 1600-meter buffers and change in HbA1c at 1, 3, and 5 years. Only 2 out of 18 comparisons (transit threshold residential density at 1-year and accessible park area at 5-years) showed statistically significant differences. If the observed differences were random, we would expect 1 out of the 18 comparisons to be significant. Therefore, we believe the significant results are likely due to chance and the effects are of limited clinical significance. Examining whether age category was an effect modifier of the built environment on HbA1c, we observed significant effects for only 1 out of 18 comparisons (1^st vs. 3^rd tertile of residential density at 1-year). These effects are of limited clinical significance, so we believe these results are likely due to chance. These results indicate that associations between the built environment and HbA1c do not vary by insulin use or age.

Discussion

Few studies have directly linked GIS-derived built environment variables to glycemic control in people with type 2 diabetes. Contrary to our hypothesis, we found that living in lower residential density neighborhoods was associated with greater declines in HbA1c. Given that residential density is a proxy for walkability, that might indicate that less walkable environments were more favorable for glycemic control among people with type 2 diabetes. Another interpretation might be that residential density does not capture all characteristics of walkability that may affect the probability of walking among people with type 2 diabetes (e.g. sidewalk presence and quality and traffic safety).

Having any supermarkets was associated with a greater decline in HbA1c at 1- and 3- years but not at 5-years. This finding is somewhat novel as prior research has not examined the effect of potential access to supermarkets on HbA1c in the same way. Prior research [18] has found mixed associations in whether gaining or losing supermarkets is associated with glycemic control. One study at the ZIP code level found that having more “neutral” or “healthy food” outlets was associated with better glycemic control measured with HbA1c [17]. Our findings support the concept that greater physical accessibility to supermarkets (e.g. more supermarkets within shorter distances of home) could support the health of people with type 2 diabetes [41]. Our findings are consistent with some past findings that distance to full service supermarkets was associated with higher quality diets [42] though other studies have not shown an association [43]. The present data were also consistent with results of many previous studies showing that distance to fast food restaurants had no impact on body weight trajectories [11]. The present analyses failed to find any impact of fast food restaurants within 1600m or 5000m from home, on glycemic control however the variable we used also includes a wide variety of fast food outlets ranging from coffee shops to traditional fast food chains.

Overall, the observed associations were small (differences of 0.05–0.08 HbA1c percentage points) and of limited clinical significance, indicating that residential density (our proxy for walkability) and food environment had limited association with HbA1c control among people with type 2 diabetes after adjusting for property values as a measure of the socioeconomic environment. Medical treatment and socioeconomic variables may be more influential. To provide context for the magnitude of the associations in our study, medications for diabetes generally provide reductions in HbA1c of about 0.5 to 1.25% [44] in clinical trials. In our study, the strongest associations were consistent with reductions of HbA1c only up to 0.08% indicating that any role built environments could play in influencing HbA1c control are likely modest.

It could also be that socioeconomic status is a more important driver of fluctuations in HbA1c over time. For example, we found that the association between property values and HbA1c control was larger than the association between physical activity and food built environment variables and HbA1c control in our cohort. Prior studies have linked socioeconomic status to glycemic control [45–47]. One study showed the depressive symptoms mediated associations between poverty and education and HbA1c [45]. Future studies should examine the role of the social environment including measures of socioeconomic status and other social determinants of health [47].

Our finding that lower density residential environments were associated with greater declines in HbA1c could suggest that suburban and rural built environments better support physical activity in people with type 2 diabetes. However, there could be possible explanations for this observed association other than the pathway through physical activity. Other built environment indicators not measured here include access to stable housing or to transit, residential segregation, crime, and noise and air pollution, which can all impact stress, inflammation, and glycemic control [14, 47]. Further, lower residential density areas are typically easier to navigate in a car which could make access to medical care and treatment for diabetes easier.

Limitations of our study include the retrospective design, inability to account for individual confounders such as leisure-time physical activity or eating behaviors given that these variables were not available in the EHR. While our models adjusted analytically for sex, we were unable to fully account for potential confounding by gender. We also included a sample that has health insurance and access to care for the period examined. Most diabetes medications with Kaiser Permanente Washington are accessible to members without high-cost co-pays. We also may not have adequately accounted for changes in medications over time, which are inherently difficult to model. We measured the built environment exposures based on home addresses and are unable to account for the broader activity spaces that participants may use such as the neighborhoods around the workplace. We used different sized buffers to account for varying distances from home, however, people also use chose to use resources outside of the selected buffer areas. Furthermore, we used Euclidean buffers rather than street network buffers given data constraints.

Conclusions

Conventional measures of the built environment, including residential density, living nearer to parks and selected food sources, had limited associations with HbA1c over time in people with type 2 diabetes. There were small protective associations for patients residing in more suburban and rural areas and with a supermarket within 1600m of home. Other factors may be more influential in glycemic control including directly measured physical activity and diet, stress, sleep, and social determinants of health. Future studies should be conducted to confirm and extend these findings, such as by examining the role of food deserts, given that there are few published research studies explicitly examining the role of the built environment on glycemic control in people with type 2 diabetes.

Highlights.

• Among people with type 2 diabetes, those in the lowest residential density areas had better glycemic control over 5-years

• People with type 2 diabetes who had a supermarket within 1600 meters of home had better glycemic control over 3-years

• Associations were small indicating that other built and social environment factors, such as the socioeconomic environment, may have larger impacts on glycemic control