This study examines a subsample of patients with T2D from the Veteran’s Administration Diabetes Risk (VADR) cohort. VADR is a dynamic retrospective cohort established from the electronic health records of all patients nationally who received primary care services in the Veterans Health Administration from 2008 to 2018 and were diabetes-free upon enrollment (Avramovic et al., 2020). Eligibility criteria included having at least two diabetes-free primary care visits at least 30 days apart in the preceding five years. Patients were excluded if they did not have an address in the contiguous United States on record prior to or within two years after cohort entry, or if during the study period they changed states for clinical care, were deceased, or were lost to follow-up after no healthcare encounters for two years. From VADR, we identified 543,437 patients who received a new T2D diagnosis, defined as: 1) at least two separate healthcare encounters with T2D-specific International Classification of Diseases 9/10 codes, or 2) at least one encounter with T2D-specific codes and at least two elevated Hemoglobin A1C measures (≥6.5 %), or 3) at least one prescription for T2D medication other than Metformin or Acarbose alone (Avramovic et al., 2020). We followed patients from new T2D diagnosis until October 6, 2023 and retained those with at least two A1C tests during follow-up in the analytic sample. (See Fig. 1). All research procedures were approved by the Veterans Administration Institutional Review Board (study number 1667) and the NYU Grossman School of Medicine Institutional Review Board (study number s17–01428).

Patient addresses were geocoded using Esri ArcGIS Application Programming Interface for Python^™. Then, two LTPA environmental exposure variables were assigned to the patient’s census tract: park proximity and physical activity facility density. We drew from definitions established in the Retail Environment and Cardiovascular Disease (RECVD) dataset. Park proximity was defined as the distance to the nearest seven parks from a census tract’s population-weighted centroid. Parks included national (2 %), state (7 %), and local (91 %) parks obtained from the Centers for Disease Control and Prevention and originally sourced from the Homeland Security Infrastructure Program Gold 2011 database and 2010 Esri ArcGIS 10.1 data (Wen et al., 2013). The measure considers a census block’s spatial access probability to its nearest seven parks, weighted by block population, and aggregated to the census-tract level (Zhang et al., 2011). The measure is strongly correlated with Euclidean distance to the closest one, five, and nine parks in a given tract and avoids arbitrarily drawing fixed boundaries of exposure (Zhang et al., 2011). Physical activity facility density was defined as the five-year average density (count per km²) of facilities within a network buffer around the population-weighted centroid of a census tract. Facilities were sourced from the National Establishment Time Series database of multi-use establishments and included fitness facilities, gyms, athletics and sports clubs, instructional studios, and recreation centers. We assigned facilities data that were collected from 2007 to 2015 and averaged within the five years immediately preceding a patient’s T2D diagnosis. Walking and driving network buffers were defined using Esri ArcGIS StreetMap Premium^™ 2019 street network data with standard precision and default settings. We applied different street network distances to walking and driving network buffers based on four community types: higher density urban, low-density urban, suburban/small town and rural. These classifications were based on a modified version of Rural-Urban Commuting Area codes applied at the census-tract level (McAlexander et al., 2022). We assigned network buffers of exposure to physical activity facility density reflecting travel distances appropriate to each community type: one-mile walking buffer for high-density urban, two-mile driving buffer for low-density urban, six-mile driving buffer for suburban/small town, and 10-mile driving buffer for rural communities. These buffers reflect an average of 10–15 min that adults may be willing to travel to access recreation (Hurvitz et al., 2014; Yang and Diez-Roux, 2012). Physical activity facility counts were divided by the corresponding area in km² within the relevant buffer.

We defined baseline A1C as the closest recorded A1C value to the date of new T2D diagnosis that was within 30 days prior to or after this date. We considered severity of baseline A1C low if less than 7.0 %, moderate if 7.0 % to 8.9 %, and severe if 9.0 % or greater (ADA, 2021).

The primary outcome was a repeated-measures binary variable indicating whether the results of A1C tests administered more than 30 days after T2D diagnosis were considered to fall within a clinically therapeutic range, where 1 = in-range and 0 = not-in-range. Based on clinical guidelines (which differ slightly based on age and other chronic illness) (ADA, 2021; Conlin et al., n.d.; Qaseem et al., n.d.), we defined “in-range” as A1C test values of 5.0 % to 6.9 % for patients less than 80 years old and 5.0 % to 8.5 % for patients 80 years or older. We included all available, non-duplicate A1C tests during the follow-up period. We defined a secondary measure of T2D control as the proportion of tests considered in-range (tests in-range/total follow-up tests). We operationalized this outcome as all tests in-range = 1 and anything fewer = 0.

We conducted descriptive analyses to summarize demographic and health-related characteristics and distributions of physical activity facility density and park distance variables across four community types (see Table 1). We also summarized the mean number of A1C follow-up tests per patient, stratified by severity of baseline A1C and proportion of A1C tests in-range (see Table 2). To examine associations between the LTPA environment and odds of glycemic control, we used generalized estimating equations (GEE) with a logit link function for repeated binary dependent variables (Ballinger, 2004; Parzen et al., 2011). When determining the need for a multilevel modeling approach to account for clustering at the county level, an intraclass correlation coefficient of 0.004 indicated that nesting of an average of 692 individuals per county across 392 counties accounted for a negligible amount of the total variance in the proportion of A1C tests in-range. We specified an exchangeable working correlation matrix to account for the dependency among within-individual repeated A1C follow-up measures over time. We preferred a GEE marginal model to maximum likelihood estimation in order to interpret parameter estimates in terms of population-averaged change in A1C due to differential, time-invariant access to physical activity facilities and parks across a large number of neighborhoods (Hubbard et al., 2010). GEE is also robust to unequal numbers of A1C follow-up measures across individuals, allowing for flexibility in the number of follow-up measures per participant. We included a time-varying covariate of years since T2D diagnosis to account for the timing of each test in relation to diabetes progression. We examined physical activity facility density and park proximity exposures in separate statistical models with and without adjusting for the other, and then exponentiated regression coefficients prior to interpretation. To demonstrate associations between LTPA environmental exposures and proportion of A1C tests in-range among Veterans with a new T2D diagnosis, we used binomial logistic regression to examine associations between each exposure and these outcomes.

At the individual-level, we also adjusted for baseline 1) A1C, 2) comorbid medical conditions assessed using the Elixhauser Comorbidity Index with Swiss weightings of −7 to 17 to yield possible index scores of −13 to 34 where lower scores indicate lower risk of mortality (Sharma et al., 2021), 3) age at VADR cohort entry, 4) sex (male or female), 5) non-Hispanic race and Hispanic ethnicity (Hispanic, non-Hispanic white, non-Hispanic black, non-Hispanic Asian, non-Hispanic American Indian/Alaska native, non-Hispanic native Hawaiian/other Pacific Islander, and missing), 6) low-income/disability flag (based on a patient’s priority group for Veteran benefits), and 7) marital status (married or living with partner, single, and missing). Indicator variables were created and included in statistical models to account for missing race/ethnicity (6.7 %), income/disability (0.8 %), and marital status (6.4 %).

At the census tract-level, we adjusted for percent Hispanic and percent non-Hispanic black residents and the following variables from the RECVD study: 1) neighborhood socioeconomic environment, a continuous z-score sum of the percent of adults with less than a high school education, unemployed, and households with <$30,000 annual income, in poverty, on public assistance, and without a car (higher score indicating higher disadvantage) (Yang et al., 2014), and 2) land use environment, created based on average block length, average block size, household density, intersection density, street connectivity, retail establishment density, and percent of developed land. Finally, we examined potential moderation by continuous neighborhood socioeconomic environment by including interaction terms in the GEE models with each physical activity facility density and park distance, which were mean-centered prior to analysis. We stratified all models by community type, using Python 3 with the “GENMOD” GEE package and SAS 9.4 to conduct the analyses.

The analytic sample included 274,463 Veterans who were newly diagnosed with T2D during the study period. The majority were male (95.0 %) and identified as non-Hispanic white (66.9 %). Patients were 64.0 ± 11.7 years of age upon receiving a new T2D diagnosis. About 36 % were disabled (irrespective of income), 38 % were not disabled but were low-income, and 25 % were neither. Patients’ mean baseline A1C was 7.2 % ± 1.8 %, descending from a high of 7.4 % ± 2.0 % in high-density urban to a low of 7.1 % ± 1.6 % in rural communities. Patients’ mean number of A1C follow-up tests (11.7 ± 7.7) and mean proportion of patients’ A1C follow-up tests that were considered in-range (0.69 ± 0.34) were also relatively uniform across community types. As shown in Table 2, patients with at least 75 % of their A1C follow-up tests in range had slightly fewer follow-up tests on average. The average density of physical activity facilities within network buffers was 2.1 ± 3.1 facilities per km² in high-density urban, 0.9 ± 0.7 in low-density urban, 0.3 ± 0.3 in suburban/small town, and 0.1 ± 0.1 in rural communities. The average distance to the nearest seven parks was 0.6 ± 0.6, 1.3 ± 1.4, 2.9 ± 2.9, and 7.3 ± 5.3 miles for each community type, respectively.

As shown in Table 3, in high- and low-density urban communities (but not in suburban/small town and rural), a one-facility per km² greater physical activity density within a patient’s network buffer was associated with 1.01 and 1.03 times greater odds of in-range A1C tests during follow-up, respectively. That is, high density urban-dwelling patients were 1 % more likely, and low density urban-dwelling patients were 3 % more likely, to have in-range A1C tests, controlling for time since new T2D diagnosis, if they had access to an additional physical activity facility. Across community types, farther distance to travel to the nearest seven parks was not associated with lesser odds of in-range A1C tests during follow-up. Unexpectedly, in low-density urban communities, odds of in-range A1C tests were diminutively greater at farther park distances. Tests of moderation by neighborhood socioeconomic environment were not significant in models for either physical activity facility density or park distance in any community type (all 95 % confidence intervals for the adjusted odds ratio for each interaction term were inclusive of 1.00).

Also shown in Table 3, in low-density urban communities only, patients with a one-facility per km² greater physical activity density within their network buffer were more likely to have all A1C follow-up tests in-range as compared to having at least one follow-up test that was not in-range. That is, low-density urban residents were 4 % more likely to maintain glycemic control if they had access to an additional physical activity facility. Similarly to the first modeling approach, low-density urban residents who had one mile farther to travel to parks were 2 % more likely to have all A1C follow-up tests in-range as compared to having at least one test not in-range.

To minimize temporality bias, environmental exposures were assigned at the time of T2D diagnosis; however, we could not account for potential changes in environmental attributes throughout the follow-up period that might have influenced the observed associations. Residential relocation during the study period could also change the exposure, which was partially accounted for by excluding those who changed states for clinical care during the study period. We could not address possible bias from residential self-selection. Buffer sizes may not reflect an individual’s most relevant neighborhood context. Physical activity and healthy eating data were not available and should be considered in future studies. Fast-food restaurant and supermarket densities have been associated with higher and lower T2D risk, respectively, in the VADR cohort (Kanchi et al., 2021). Medical records provide data on verified illnesses but may be inaccurate or incomplete. There may have been unmeasured illness and treatment effects during follow-up (e.g., undiagnosed T2D, medication use, or program participation). However, because this patient population had access to Veterans Affairs healthcare, we anticipated robust diagnosis, testing, and treatment rates that were not geographically distributed, and therefore unrelated to the exposure variables. Considering patients may begin new treatments during follow-up, treatment could be examined in future studies as an exposure of interest in relation to glycemic control, potentially modified by the LTPA environment.

Participants who remained in the study over time may differ characteristically (e.g., have better health or higher income) from participants who were deceased or discontinued care. Generalizability may also be limited beyond non-male, non-white populations whose monitoring of A1C is not well documented in an electronic health record. The exclusion of 229,904 patients with T2D without A1C data at baseline may not limit generalizability considerably because they were comparable in demographics and neighborhood-level characteristics to those in the final analytic sample, though they were on average three years older when diagnosed (67.2 ± 12.3), and slightly more were non-Hispanic white (74.4 %) and low-income (41.5 %). Strengths of this study include its large nationwide sample, a decade-long longitudinal design, multiple repeated A1C follow-up measures, utilization of geographic information systems-based measures, control for confounding at the individual- and neighborhood-level, and comparisons across the urban-rural continuum.

Associations between the LTPA environment, physical activity behavior, and chronic disease management are highly complex, and likely differ based on local contexts, cultures, and population demographics (Gidlow et al., 2019). Broader access to neighborhood physical activity facilities may contribute in small ways to maintaining glycemic control in urban populations with newly diagnosed T2D. Research is needed that examines the interplay between individual, socioeconomic and physical neighborhood environments and TD2 outcomes (Schüle and Bolte, 2015). Our understanding of the appropriate geographies to examine and their associations with physical activity behavior and glycemic control is particularly limited in suburban/small town and rural communities. Ultimately, future studies should identify optimal typologies of built, natural, physical activity, and food environmental supports that contribute to the effectiveness of chronic disease management interventions across diverse geographies and communities. Research is also needed to evaluate interventions that increase access to and promote the use of LTPA facilities and parks, while tailoring interventions to local communities.