Neighborhood Greenspace and Socioeconomic Risk are Associated with Diabetes Risk at the Sub-neighborhood Scale: Results from the Prospective Urban and Rural Epidemiology (PURE) Study

The current cross-sectional study focusses on 5125 participants recruited between 2006 and 2009 in Hamilton and Vancouver, Canada, for whom we had robust address and community-level characteristics. These sites are part of the PURE study, which is an ongoing global cohort study investigating risk factor for chronic disease in 27 countries [19]. All participants provided informed consent and the study was approved by the local institution research ethics boards.

Individual variables collected using surveys included residential address, age, sex, household income range, tobacco and alcohol consumption, history of diagnosed diabetes, and history of diabetes medication use. Additionally, each participant’s diet was determined using the alternative healthy eating index (AHEI), a nine-component index on dietary choices and nutrient intake designed to assess dietary-based chronic disease risk [20], and daily physical activity metabolic equivalent of task (MET) scores were calculated from robust questionnaire instruments as described by Lear et al. [21]. Trained project staff conducted in-person measurements to derive body mass index (BMI, measured as a continuous variable) and waist-to-hip ratio (WHR).

Participants were classified as having diabetes (either type 1 or type 2) if they fulfilled any of the following criteria: self-reported a diagnosis of diabetes; current or past use of diabetes-specific medication; a fasting plasma glucose level greater than or equal to 7.0 mmol/l.

Data for Canadian census dissemination areas (DA, the smallest available census area with an average population of 400–800 residents) for the census year 2006 were acquired from Statistics Canada. The SES variables included in this analysis are listed in Table 1. Each DA was categorised as urban/suburban/rural using a previously validated, Canada-specific classification method based on active transportation and population density [7, 22, 23].

The Normalized Difference Vegetation Index (NDVI) provides a greenspace metric often used in health research [24]. Derived from satellite imagery, higher NDVI values indicate a greater intensity of greenness. We acquired cloud-free LANDSAT 5 satellite images for Hamilton (09.05.2006) and Vancouver (23.07.2006) through the United States Geological Survey’s EarthExplorer platform. Both images were preprocessed and NDVI scores were calculated and mapped as 30-m pixels (native resolution) in the study areas.

Data preparation and analysis were conducted using R (v.4), and mapping was completed using QGIS (v.3.10) on a Linux system (AMD Ryzen 9 3900 × CPU, 64 GB DDR4). All code including detailed documentation are available at https://​github.​com/​CHEST-Lab/​DRIGLUCoSE.

In order to estimate each participant’s potential exposures to greenspace and local SES, we mapped age- and sex-specific walkable buffer zones around each individual’s residential address using street and path networks derived from OpenStreetMap data (Fig. 1). Using age- and sex-specific walking speeds (average male–female difference = 0.13 km/h; [25]), we then mapped each participant’s walkable areas in 2-min increments from their residential address, with a maximum walking time of 20 min. The 20-min maximum parameter was derived through sensitivity analysis, in which walking areas were calculated in 2-min intervals and iteratively entered into the successive statistical models described below. The 20-min parameter was selected as it (i) featured the highest predictive performance, (ii) is a heuristically realistic representation of movement patterns in the city, in that it adequately captures activity spaces of persons in the study (this was ascertained through informal discussion with residents of the study area, in which several of the authors reside), and (iii) the 20-min zone also approximates the radius of structurally homogenous spaces/neighborhoods in the study area and may therefore be the most suitable proxy for composition of the built environment. Each participant was thereby assigned ten concentric walking zones from 0 to 20 min walking time (Fig. 1a). The concentric walking zones were then overlaid on the SES and greenspace data described above (Fig. 1b). A logit weighting function was applied to each buffer zone’s mean distance to derive distance-based variable weights, such that the estimated effect of an SES or greenspace variable decreases as distance from the home increases (Fig. 1c). A logit function was selected as it heuristically approximates a suitable distance-decay function [26] and various parameterisations were assessed through sensitivity analysis. The zone-distance-weighted mean of each of the 15 SES variables and 4 greenspace variables was then assigned to each participant for index derivation. Equations and documented code are provided on our GitHub page, linked above.

In order to increase the amount of built environment information contained in the model, we calculated 4 separate metrics from the NDVI data: median NDVI score; standard deviation; and 95^th and 5^th percentiles. These were selected as they respectively represent overall local greenspace levels, variability in the amount and intensity of greenspace within each walking zone, the intensity of the most green areas (P₉₅) that may serve as local attractants or forested areas within a walking zone, and non-green areas (P₅) that are characteristic of a dense urban or industrial built environment (greyspace) and bare land, hypothesised to exhibit negative effects.

Using participant data and residential address locations, we first derived the index as follows: in a subsequent step, we calculated and mapped the index across both study areas.

In order to manage class imbalance (i.e., ratio of diabetes to non-diabetes participants) for the index derivation [27], we used a combination of undersampling and the SMOTE algorithm [28] to iteratively derive model calibration sets, as described in our code/documentation. To derive the DRI-GLUCoSE index, we used principal component analysis (PCA) [29‐32]. The 11 SES and 4 greenspace variables listed in Table 1 were normalised and tested for non-sphericity and suitability using Bartlett’s test and for sampling adequacy using Kaiser–Meyer–Olkin before being entered into a PCA model. The resulting loadings from the first component were then applied as variable weights for the 15 input variables, the sum of which was then rescaled to a range of − 1, 1 and taken as the participant DRI-GLUCoSE score. The 4 SES variables with negligible Eigenvalues or absent from the first component were removed, as these had no detectable influence on model accuracy, resulting in a final index with 11 SES variables and 4 greenspace variables. The index is scaled such that high index values correspond with socially deprived areas and low greenspace.

For model comparison, we derived three variants of the index: (a) the combined DRI-GLUCoSE index described above; (b) a variant using only the SES variables; (c) a variant only using the greenspace variables. These are described in more detail on our GitHub page (linked above).

In order to map index scores across both study areas, we first removed non-residential land use from the maps, then overlaid a 50-m pixel grid. The 50-m parameter was selected to achieve a balance between computational load and spatial precision following sensitivity analysis of variable pixel sizes. From the centre of each 50-m pixel, we assume a standard participant with a 20-min walking zone and a fixed walking speed of 4.7 km/h [25]. The DRI-GLUCoSE index scores for each pixel were then calculated using the weighting procedure described above. This resulted in a DRI-GLUCoSE index map for both study areas.

Of the total 5125 participants included in this analysis, 10% were classified as having diabetes (Table 1). The mean BMI and WHR were 27.8 kg/m² and 0.86, respectively, and participant median age was 53 years (IQR 46, 61) at the time of data collection. Approximately 28% of participants reported an average of more than one unit of alcohol consumed per day, and 58.8% of household annual incomes were above CAD $65,000. Participants were equally divided between the two cities, with 78% of participants residing in suburban or rural neighborhoods.

Bartlett’s test results confirmed strong non-sphericity (p < 0.001) of the predictor variables, and Kaiser–Meyer–Olkin returned an overall measure of sampling adequacy of 0.74, indicating a reasonable degree of suitability. The PCA results indicate that 49.3% of variance was explained in the first component, whose factor loadings for the derived index are shown in Table 2. The remaining components respectively explained 14.5%, 9.6%, 6.8%, and 5.2% of variance and were discarded. The factor loadings for the first component were then used as variable weights and entered a weighted linear combination to derive index scores.

The mapped index values (Fig. 2) highlight areas where the local SES and greenspace conditions predict a higher (dark) or lower (light) diabetes risk.

The results of the logistic regression models based on the testing subset are shown in Fig. 3. Participants’ local DRI-GLUCoSE score exhibited a consistent positive association across all models. A small, non-significant attenuation of this effect was observed in the fully adjusted models, where individual-level lifestyle factors were included (diet, physical activity, smoking, and alcohol consumption). Increased odds of diabetes were observed for obese participants, current/former smokers, and persons who consume an average of less than one unit of alcohol per day. A lower risk is observed for urban residence, higher household incomes, and physical activity scores, but healthy eating score was non-significant (odds ratios shown in online appendix: https://​github.​com/​CHEST-Lab/​DRIGLUCoSE).

The multivariable models’ accuracies were 75% for the semi-adjusted model (Youden Index = 0.41, sensitivity = 0.78, specificity = 0.62, AROC = 0.76) and 75% for the fully adjusted model (Youden Index = 0.41, sensitivity = 0.75, specificity = 0.71, AROC = 0.78). Model diagnostics indicated no overfitting. Experimental models based on alternative configurations of greenspace variables, univariate (non-PCA) predictor variables, and interactions between greenspace and SES predictors consistently featured lower accuracy and/or featured poor model fit.

The final models were compared to identically parameterised logistic models using two additional variants of the DRI-GLUCoSE index: SES only and greenspace only. The combined SES and greenspace model achieved the highest prediction accuracy of 75%, compared to 72% for SES only (see Appendix).