Private garden exposure and public park access in relation to type-2 diabetes incidence: a UK Biobank cohort study

Abstract

Background

Existing longitudinal studies examining the relationship between greenspace exposure and incidence of type-2 diabetes (T2D) have primarily operationalized greenspace using vegetation indices. Little is known about the effect of greenspace types (e.g., private gardens) and public park access.

Methods

We investigated the associations between residential greenspace exposure, including private gardens (determined using Ordnance Survey MasterMap™ Greenspace) and public park access, with the incidence of T2D using the UK Biobank (UKBB) data. Public park access, such as nearest distances (i.e., walkable road network and Euclidean), and the number of parks were calculated for each participant. The incidence of T2D was ascertained through linkage of hospital admissions data. Cox proportional hazard models, adjusting for covariates, were used to estimate the hazard ratios (HR) and 95% confidence intervals (CI). We also performed stratified analyses by age, sex, neighbourhood deprivation, and family history of diabetes.

Results

Of the 423,282 UKBB participants (mean age:56.36 years) included in the study, 19,648 developed T2D over a median follow-up of 15.41 years. Compared to the first quartile, participants in the highest quartile of private garden cover (%) had a reduced risk of T2D (HR: 0.932; 95%CI: 0.88, 0.983). For park access, nearest distance (whether walkable or Euclidean) was not associated with the incidence of T2D. However, having a higher number of parks, particularly three or more parks within an 800-m buffer of the home location, was found to lower the incidence of T2D (HR: 0.943; 95% CI: 0.906, 0.981). Stratified analyses revealed that the beneficial effects of private gardens were stronger among participants in deprived areas and those without a family history of diabetes.

Conclusion

Private residential gardens exposure (often overlooked in greenspace-health research) and a higher number of parks around homes were found to lower the incidence of T2D. This has implications for urban planning and public health, particularly in the prevention and management of diabetes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-28277-1.

Introduction

The incidence and prevalence of diabetes has risen over the past three decades, placing a substantial burden on healthcare systems [1], with associated global health expenditures reaching US$966 billion [2]. In 2021, an estimated 529 million people of all ages were living with diabetes worldwide, and this number is projected to more than double to 1.31 billion by 2050 [3]. The global burden of disease (GBD) attributed to diabetes was 37.8 million years of life lost (YLLs) and 41.4 million years lived with disability, yielding a total of 79.2 million disability-adjusted life years (DALYs) [3]. In the UK, an individual is diagnosed with diabetes every two minutes, and one in 15 people now live with diabetes – approximately 7% of the UK population [4]. Diabetes is also a significant risk factor for coronary heart disease and stroke [5–7] which have been identified as the first and third leading causes of mortality in 2021, respectively [8]. Thus, diabetes is a significant public health challenge.

Type 2 diabetes (T2D) is the most common type of diabetes, accounting for 96% of all diabetes cases in adults [3]. The disease is characterized by low insulin secretion, insulin resistance, and high glycated haemoglobin (HbA_1c) [9, 10]. Several modifiable lifestyle factors, including physical inactivity, poor diet, and being overweight, play a crucial role in the development and progression of T2D [10]. The incidence of diabetes can be reduced via lifestyle interventions [11] or modifications [12]. Exposure to residential greenspaces offers a potential avenue for lifestyle interventions by facilitating behaviours such as increased physical activity, greater time outdoors for vitamin D synthesis, reduced sedentary behaviour, and enhanced social interaction, all of which are associated with improved metabolic health [13–17]. For example, Yuan and colleagues reported that increases in moderate-to-vigorous physical activity and decreases in sedentary behaviours (e.g., leisure screen time) were significantly associated with a reduced risk of T2D, in part due to reduced obesity and chronic inflammation, as well as higher lean mass [18]. Residential greenspace can also reduce human exposure to environmental air pollutants [17] by absorbing particulate matter and nitrogen dioxide (through processes such as deposition) [14], which are important risk factors for metabolic dysfunction [19].

Prospective [20–22] and retrospective [23, 24] cohort studies have examined the effects of residential greenspace exposure on the incidence of diabetes, but several research gaps remain. First, existing studies mostly relied on satellite imagery, such as the normalized difference vegetation index (NDVI), to quantify greenspace exposure [13]. While NDVI tends to homogenize greenspace exposure [25], it is limited in capturing the high spatial heterogeneity within urban settings [14, 26], and does not account for variation in greenspace types [27, 28], e.g., exposure to private residential gardens and public parks. As an alternative, greenspace typologies derived from land-use datasets (e.g., Ordnance Survey MasterMap™) can be used to capture more functionally distinct greenspace exposures. Second, there is limited evidence on how residential proximity to accessible types of greenspaces (e.g., distance to public parks or community gardens) influences the incidence of T2D. In a systematic review of 13 studies on greenspace exposure and the incidence of T2D, none of the studies measured or operationalized greenspace exposure in terms of proximity or accessibility [13]. This research gap was also present in a recent prospective cohort study among US women, which lacked information on residential greenspace accessibility or proximity [29]. Even in broader greenspace proximity-health literature, little is known about the actual health effect of walkable road network distance, with most studies using Euclidean (straight-line) distance as a proximity measures [30, 31], which may lead to exposure misclassification and biased results [32]. We have previously shown that exposure to specific greenspace types, such as private residential gardens, varies by socioeconomic status (SES) and may confer greater benefits among less deprived populations, partly due to their greater ability to afford housing with private gardens [17, 27]. Therefore, providing access to community gardens or public parks to deprived individuals could help address socioeconomic health disparities [27]. Third, while access to or visit to public parks has been linked with physical and community well-being [33, 34], evidence on how the density of parks in close proximity to residential homes impacts the risk of being diagnosed with T2D remains limited [13]. For example, do participants living in areas with ‘one or no parks’ versus those with ‘at least two public parks’ within an 800 m buffer (approximately 10–15 minutes’ walk) from home locations experience the same health benefits, such as a reduced risk of developing T2D? Exploring these questions could help inform urban planning policies on whether, and to what extent, access to public greenspaces can be leveraged to reduce the incidence of T2D, particularly in underserved communities.

To address these knowledge gaps, this prospective cohort study aimed to examine the associations between multiple residential greenspace exposure metrics and the incidence of T2D in the UK Biobank. Specifically, we hypothesize that (i) higher exposure to residential greenspace, including private residential gardens, will be inversely associated with the incidence of T2D; (ii) Individuals living closer to public parks – particularly within an 800 m buffer – are less likely to develop T2D; and (iii) the reduced risk of T2D will be greater among participants with access to a higher number of public parks within an 800 m buffer of their home location compared to those with only one or no parks.

Methods

Study design and population

We conducted a prospective cohort study using data from the UK Biobank (UKBB), a large population-based cohort of over half a million participants aged between 37 and 73 years, recruited between 2006 and 2010 from 22 assessment centers across England, Scotland, and Wales [35]. Briefly, participants gave information on sociodemographic characteristics and health status through a self-administered questionnaire, participated in a nurse-led interview, underwent various physical assessments, and provided urine, saliva, and blood samples. Participants with self-reported diabetes (n = 20,665), hospital admissions for diabetes (n = 7,147), and use of diabetes medication (n = 96) at baseline were excluded. Missing covariates were handled using complete case analyses, leaving a total of 423,282 participants (Figure S1). The UKBB was approved by the North Multi-Centre Research Ethics Committee (REC reference: 21/NW/0157), and all participants provided their informed consent for the use of their data in research investigations.

Outcome assessment

Primary outcomes of the study included the incidence of T2D. Hospital inpatient admissions data were collected regularly through linkages to Health Episode Statistics (HES) for England, the Patient Episode Database for Wales (PEDW), and the Scottish Morbidity Records (SMR) for Scotland. The data contains information on participants’ diagnosis codes and dates recorded across all their hospital inpatient records in either the primary or secondary position and was used in the current analysis. Incident cases of T2D were identified using the International Classification of Diseases, tenth edition (ICD-10) code E11 (non-insulin-dependent diabetes mellitus) from hospital inpatient records.

Exposure assessment

We assessed residential greenspace exposure using total greenspace cover, private residential gardens, and access to public parks or gardens. Total greenspace cover was included to capture broader residential greenspace exposure and to provide contextual comparison with private residential gardens, a more specific greenspace type representing exposure to immediate, non-public environments.

Total greenspace cover and private residential gardens

Total greenspace cover and private residential gardens were previously ascertained by Roscoe et al., [28] using the Ordnance Survey MasterMap™ Greenspace Layer, which provides detailed, high-resolution geospatial data (1;1250 scale) of 18 functional greenspace categories, including private residential gardens, across the UK (Table S1). Briefly, the percentage of total greenspace cover was calculated as the aggregated percentage of all 18 OSMM greenspace categories within a 100 m buffer of UK Biobank participants’ home location. Private residential garden exposure was defined as the percentage of land classified as private gardens within the same buffer. We selected a 100 m buffer to capture the immediate residential environment, as private garden exposure is most accurately represented at this spatial scale compared with larger buffers [17, 27].

Access to public parks or gardens

Several measures for access to public parks have been proposed in previous studies, including distance to the nearest park, and the number of nearby parks within walking distance to home locations [31, 36, 37]. In the present study, we used Euclidean distance, walkable road network distance, and number of parks within or intersecting the 800 m buffer of the participant’s home location, as objective measures of residential access to public parks or gardens (hereinafter referred to as ‘parks’). Euclidean and walkable road network distances were included mainly to compare both measures and ascertain their individual effects on the risk of developing T2D, given evidence suggesting that Euclidean distance can substantially deviate from the actual distance [31].

Euclidean distance calculation

We used the UK Biobank participants’ home location (x, y coordinates), rounded to a 100-metre accuracy, as a proxy for each included residential address (starting points) and access points of parks (destination points) to calculate the Euclidean distance. Parks access points were obtained from the Ordnance Survey (OS) Open Greenspace dataset, which provides a comprehensive coverage of all publicly accessible greenspaces in the UK (scale 1:1250 to 1:10,000; Vector data first released in March 2017; Version date: October 2024). These publicly accessible greenspaces comprise ten classifications, including public parks or gardens, allotments, and playing fields, with detailed descriptions provided in Tables S1 and S3. Participants’ home locations and parks data were first converted to spatial features using the sf package in R and then transformed to the British National Grid coordinate reference system to ensure alignment. Given that most parks have more than one designated access point, a straight-line distance was calculated from each home location to the nearest access point (Fig. 1) using the st_nn function from the nnego package in R.

Fig. 1.

Spatial representation of park accessibility from a participant's home location using network-based and straight-line routes. Fig. 1 illustrates an example of how the nearest walkable road network distance and Euclidean distance were calculated from a participant’s home location () to the closest park access points () via the network-based route (), and staight-line route (), respectively. The figure also highlights the discrepancies between actual walkable access and theoretical proximity. These two calculation approaches were applied to all 423,282 UKBB participants

Network-constrained walking distance analysis

For walkable road network distances, we utilized the same data as above, along with the OpenStreetMap (OSM) highways dataset, which provides detailed information on road, path, and track networks categorized by type and usage (Table S2) (© OpenStreetMap; https://www.openstreetmap.org/copyright). OSM data (freely downloaded from https://www.geofabrik.de/) was used to construct the walkable road network. The road network reflects actual pedestrian access patterns (Fig. 1), unlike the Euclidean distance used in previous studies [30, 38], which does not account for real-world walking routes, including street networks, connectivity, and physical barriers (e.g., rivers, and fences).

Using ArcGIS Pro and R, we implemented several approaches to calculate the walkable road network distance while ensuring scalability and computational efficiency for over 420,000 participants. (1) Participants’ home location, park access points, and the OSM road data were spatially filtered to the UK regional boundary (one at a time) to constrain the analysis. (2) Participants’ home location (geocoded points), parks access points, and the road network were then transformed to WGS84 coordinate reference system (EPSG: 4326) to support network-based routing. (3) A 2,500 m buffer was applied around each participant’s home location (Fig. 1; Figure S2) to generate a localized walkable network graph (isodistance) using the dodgr package (version 0.4.2) in R [39]. This threshold was informed by preliminary analyses on a subsample of the cohort, which showed that all participants had their nearest park access point within approximately 2,400 m. The 2,500 m buffer was therefore selected as a conservative upper bound to capture all relevant nearest park access points while excluding distant road segments unlikely to contribute to the shortest-path calculation. The road segments within the buffer were assigned walkability weights based on their OSM functional classifications (e.g., footway, paths, and pedestrian, etc., Table S2). Participants’ home locations and park access points were matched to their nearest vertices (network nodes) on the walkable network graph. We then computed network-based distances from each participant’s location to all reachable park access points within the network graph, recording the shortest of these distances as the walkable road network distance for that participant. (4) Parallel batch processing using future and future.apply packages in R was employed to scale the analysis and allow multiple concurrent workers without exceeding memory limits.

The number of parks within or intersecting the 800 m buffer

As shown in Fig. 1, many parks in the OS Open Greenspace dataset have multiple access points. Therefore, we aggregated these points by their unique identifier to ensure that each park was counted only once, regardless of the number of access points. Using a similar approach from a previous study [36], number of parks was estimated by attributing park polygons geometry to a buffer if any part of the park was located inside or intersected with the 800 m buffer surrounding each participant’s home location. We chose an 800 m buffer because it represents participants’ neighbourhood that could be accessed within an approximate 10-minute walk [36, 40, 41].

Covariates

Individual sociodemographic and lifestyle factors were collected at baseline. Potential covariates were selected a priori and based on previous studies/reviews examining the relationship between greenspace and T2D [13, 21]. A directed acyclic graph (DAG), using DAGitty v3.1 software [42], was used to identify the minimal set of variables for inclusion in the model (Figure S3). Based on DAG, these variables include age, sex (male/female), education status (college degree or above/any other qualifications/no qualification), ethnicity (white/non-white), residential area (urban/others – including rural, and suburbs), and Townsend Deprivation Index (continuous). Lifestyle [e.g., smoking status (current, previous, never), alcohol intake frequency (three or more per week/twice or less per week/occasionally/never), and healthy diet score (ranging from 0 to 4)] and health-related factors [e.g., family history of diabetes (yes/no – based on diabetes status of first-degree relative, i.e., parent or sibling)] that contribute to variability of T2D were also identified.

Statistical analysis

The UKBB participants’ characteristics were presented based on diabetes status as mean (standard deviation, SD) or median (interquartile range, IQR) for continuous variables, and number (percentage) for categorical variables. We ran a Spearman correlation analysis for total greenspace cover, private residential gardens, walkable road network distance, Euclidean distance, and number of parks. A Bland-Altman plot [43] was used to visualize and assess limits of agree-/disagreement, and mean differences between Euclidean distance and walkable road network distance to parks for both raw and log-transformed values [due to skewness of the distance measures (Figure S4)].

Follow-up time was calculated from the date of recruitment to the date of T2D incidence, date of death, or end of follow-up (31st July 2024), whichever occurred first. We modeled the exposure-response relationship between residential greenspace (total greenspace cover and private residential gardens) and incident T2D using a restricted cubic spline (RCS) with 3 degrees of freedom (d.f.). The likelihood ratio test (LRT) was used to assess linearity and compare the RCS model with a corresponding linear model. A p-value of < 0.0001 for both total greenspace cover and private residential gardens indicates that the RCS model provided a significantly better fit than the linear model, suggesting strong evidence of nonlinear associations (Fig. 2). Multivariable Cox proportional hazard models were then used to estimate the hazard ratios (HR) and 95% confidence interval (CI) for the following. (1) The association between quartiles of total greenspace cover, and private residential gardens, and the incidence of T2D. (2) The association between Euclidean distance and walkable road network distance to the nearest park, and the incidence of T2D. We operationalize proximity to parks using three categories: 0–800 m, > 800–1500 m, and > 1500 m, with the middle distance as the reference category. The 800 m cut-off was selected as a commonly used proxy for a walkable distance to neighbourhood amenities, including public parks (approximately 10–15-minutes’ walk), as supported by previous studies [44–46]. The 800–1500 m category was included to capture intermediate accessibility beyond typical walking distance but still within a reasonable proximity, while distances greater than 1500 m represent lower accessibility to public parks. This allowed us to examine whether living ‘near parks (e.g., within ≤ 800 m) is associated with greater benefits compared to living farther away from a walkable distance to the closest park. To address outliers, we considered, a priori, 2500 m as the maximum walkable distance, as we did not expect participants to walk more than 30 min to the nearest public park. (3) The relationship between access to ‘two to three parks’ (moderate park availability) and ‘more than three parks’ (higher park availability) – compared to the reference category - one or no parks (limited park availability) – within an 800 m buffer of participants’ home location, and the incidence of T2D. All models were adjusted in a stepwise manner. Model 1 adjusted for a minimal set of confounders identified through a DAG, including age, sex, education status, ethnicity, residential area, and the Townsend Deprivation Index. Model 2 adjusted for Model 1 confounders, lifestyle factors (smoking status, alcohol intake frequency, and healthy diet score), and health-related factors (family history of diabetes). These models showed no evidence of multicollinearity based on tolerance (> 0.2) and Variance Inflation Factor (VIF) (< 5). The assumption of Cox proportional hazard models was tested using Schoenfeld residual plots. The scaled residual plots for most covariates were flat and primarily centered around zero, indicating the lack of any possible violations, except for smoking and alcohol intake frequency. A stratified Cox proportional model was therefore fitted, allowing for separate baseline hazard functions across strata of these variables while estimating the effects of the remaining covariates.

Fig. 2.

Exposure-response curve using restricted cubic spline (3 degrees of freedom (d.f.)) on the association between residential greenspace (total greenspace and private residential gardens) and the incidence of T2D. HR: hazard ratio; CI: confidence interval; IF: inflection points. Spline models were adjusted for age, sex, education status, ethnicity, residential area, Townsend Deprivation Index, smoking status, alcohol intake frequency, healthy diet score, and family history of diabetes. The solid blue line depicts the estimated HR, while the shaded blue area is the 95% CI. The vertical red dotted lines indicate the inflection point, while the horizontal black dotted lines indicate the point at which the exposure has no association with the outcome

Multiplicative interaction terms were used to test for potential effect modification of the greenspace-T2D association by age (< 65 and ≥ 65 years), sex, Townsend Deprivation Index (≤ median value and > median value), and family history of diabetes, to identify subgroups that might benefit more from greenspace exposure.

Sensitivity testing

Several sensitivity analyses were conducted. First, analyses were repeated after excluding participants diagnosed with T2D within the first two years of follow-up. This approach minimizes the likelihood of delayed diagnosis while ensuring a sufficient latency period for exposure to influence the outcome. Second, we adjusted waist circumference as a covariate in the main model, as it is a marker of metabolic risk associated with visceral fat [47]. Third, the temporality and risk of reverse causation were addressed by restricting the analysis to participants who had lived at their residential address for ≥ 10 years at recruitment. Fourth, analyses were repeated after excluding potential undiagnosed cases of diabetes based on glycated hemoglobin levels (HbA1c ≥ 48 mmol/mol) [48]. Fifth, we re-ran analyses for total greenspace cover and private residential gardens using alternative buffers (300 m, 500 m, and 1000 m) based on methodological guidance for assessing optimal greenspace exposure [14]. Sixth, we tested whether the size of a park (measured in acres) near participants’ home location is linked with a reduction in the incidence of T2D, as a recent study has reported that larger urban parks are associated with a lower prevalence of poor physical health [49]. In line with a previous study [49], the continuous measure of park size was rescaled to 2.47 acres (0.01 km²) to encourage data interpretation. We followed the guidelines of the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement [50].

Results

Study population

The baseline characteristics of UKBB participants are presented in Table 1. Of the 423,282 participants, a total of 19,648 cases of T2D occurred during a median follow-up of 15.41 years (6,855,847 person-years). Participants had a mean age (SD) of 56.36 (8.08) years, and were predominantly female (55.36%), white (95.09%), never smokers (55.43%), and living in urban areas (89.23%). Individuals with incident T2D were older, mostly males, current smokers, socioeconomically deprived, less likely to have at least a college education, and had a higher prevalence of family history of diabetes compared to those without.

Table 1.

Baseline characteristics of the 423,282 UK Biobank participants based on T2D status

Variable	Incident T2D
	Total N = 423,282	with T2D N = 19,648	without T2D N = 403,634
Age (years), mean (sd)	56.36 (8.08)	58.90 (7.41)	56.20 (8.09)
Sex [n (%)]
Male	188,969 (44.64)	11,056 (56.30)	177,913 (44.10)
Female	234,313 (55.36)	8,592 (43.70)	225,721 (55.90)
Education status [n (%)]
College degree or above	202,389 (47.81)	6,838 (34.80)	195,551 (48.40)
Any other qualification	149,806 (35.39)	6,963 (35.40)	142,843 (35.40)
No qualification	71,087 (16.79)	5,847 (29.80)	65,240 (16.20)
Alcohol intake frequency [n (%)]
Three or more/week	186,530 (44.07)	6,426 (32.70)	180,104 (44.60)
Twice or less/week	110,451 (26.09)	4,752 (24.20)	105,699 (26.20)
Occasionally	94,347 (22.29)	5,805 (29.50)	88,542 (21.90)
Never	31,954 (7.55)	2,665 (13.60)	29,289 (7.26)
Smoking [n (%)]
Current	44,137 (10.42)	2,981 (15.20)	41,156 (10.20)
Previous	144,521 (34.14)	7,767 (39.50)	136,754 (33.90)
Never	234,624 (55.43)	8,900 (45.30)	22,724 (55.90)
Healthy diet score [n (%)]
0–1 (unhealthy)	62,535 (14.77)	3,406 (17.30)	59,129 (14.60)
2	89,721 (21.20)	4,024 (20.50)	85,697 (21.20)
3	152,884 (36.12)	6,959 (35.40)	145,925 (36.20)
4 (healthy)	118,142 (27.91)	5,259 (26.80)	112,883 (28.00)
Residential area [n (%)]
Urban	377,680 (89.23)	17,981 (91.52)	359,699 (89.10)
Others	45,602 (10.77)	1,667 (8.48)	43,935 (10.90)
Ethnicity [n (%)]
White	402,499 (95.09)	17,544 (89.30)	384,955 (95.40)
Non-white	20,783 (4.91)	2,104 (10.70)	18,679 (4.63)
Townsend deprivation index, mean (sd)	-1.33 (3.06)	-0.46 (3.40)	-1.37 (3.04)
Family history of diabetes [n (%)]
Yes	70,394 (16.63)	5,231 (26.60)	65,163 (16.10)
No	352,888 (83.37)	14,417 (73.40)	338,471 (83.90)
Greenspace percentage cover in 100 m buffer, mean (sd)
Total greenspace cover	57.52 (12.92)	56.10 (13.10)	57.60 (12.90)
Private residential gardens	44.57 (15.17)	42.30 (15.40)	44.70 (15.10)
Walkable road network distance to the nearest park (meters), [n (%)]
0–800 m	268,540 (63.44)	12,855 (65.40)	255,685 (63.30)
> 800–1500 m	105,229 (24.86)	4,789 (24.40)	100,440 (24.90)
> 1500 m	49,513 (11.70)	2,004 (10.20)	47,509 (11.80)
Euclidean distance to the nearest park (meters), [n (%)]
0–800 m	336,008 (79.38)	15,913 (81.00)	320,095 (79.30)
> 800–1500 m	65,385 (15.45)	2,924 (14.90)	62,461 (15.50)
> 1500 m	21,889 (5.17)	811 (4.13)	21,078 (5.22)
Number of parks within or intersected by an 800 m buffer, [n (%)]
0–1	195,638 (46.22)	8,648 (44.00)	186,990 (46.30)
> 1–3	148,457 (35.07)	6,997 (35.60)	141,460 (35.00)
> 3	79,187 (18.71)	4,003 (20.40)	75,184 (18.60)

T2D type-2 diabetes, sd standard deviation

Exposure-response relationship for total greenspace cover and private residential gardens

We observed a non-linear (p for non-linearity < 0.0001), reverse U-shaped, dose-response relationship for total greenspace cover and private residential gardens with incidence of T2D (Fig. 2). Visual inspection of spline curves suggests inflection points (IF) of around 27% and 59% for private residential gardens and total greenspace cover exposure, respectively. These points indicate where the initially non-significant increasing hazard ratio for lower greenspace exposure begins to decline – slightly for private residential gardens and more steeply for total greenspace cover – signaling the onset of protective effects against the incidence of T2D. A marked, significant decline was observed beyond 53% of private residential greenspace exposure.

Walkable road network and euclidean distance

The walkable road network distance was positively correlated with Euclidean distance [Spearman correlation coefficient (ρ) = 0.94] and negatively correlated with the number of parks within an 800 m buffer of participants’ home location (ρ = -0.58) (Figure S5). Based on the Bland-Altman method, walkable road network distances were consistently longer than Euclidean distances, with a mean difference of 216.24 m and discrepancies widening at greater distances (Fig. 3). Upon log-transforming the measures, a mean log difference of 0.36 was observed, indicating that walkable road network distances were, on average, 43% longer than Euclidean distances.

Fig. 3.

Comparison between walkable road network distance and Euclidean distance to parks using a Bland-Altman plot. A Bland-Alman plot using raw (actual) values; B Bland-Alman plot using log-transformed values due to skewness in the distribution of both the walkable road network and Euclidean distances. m: meters. The horizontal blue line indicates the mean difference and mean log difference between the walkable road network distance and Euclidean distance, while the shaded grey area depicts the limit of agreement between walkable distance and Euclidean distance

Greenspace and incidence of T2D

Associations between greenspace and incidence of T2D are presented in Table 2. After adjusting for a minimal set of confounders (Model 1), participants in the 2nd (HR: 1.082; 95% CI: 1.034, 1.133) and 3rd (HR: 1.098; 95% CI: 1.048, 1.152) quartiles of total greenspace exposure showed a slight increase in risk of T2D, while those in the highest quartile exhibited a non-significant risk reduction (HR: 0.978; 95% CI: 0.930, 1.028). Similar results were seen for higher quartiles of total greenspace cover when lifestyle and health-related factors were further adjusted (Model 2). Compared to the lowest quartile, we observed that participants in the highest (but not the intermediate) quartiles of private residential garden exposure had a lower risk of T2D across the two models – Model 1 (HR: 0.928; 95% CI: 0.881, 0.977) and Model 2 (HR: 0.932; 95% CI: 0.885, 0.983). In terms of greenspace accessibility, living either ≤ 800 m or > 1500 m from parks was not associated with T2D incidence when compared to living between > 800 and 1500 m, regardless of the proximity measure used. Access to more than three parks within an 800 m buffer of participants’ home location was linked to a lower risk of T2D in the fully adjusted model (HR: 0.943; 95% CI: 0.906, 0.981).

Table 2.

Association between greenspace types and accessibility and the incidence of T2D

Greenspace measure	T2D [HR (95% CI)]
	Model 1	Model 2 (main model)
Total number of incident cases of T2D for total greenspace cover and private residential gardens, n = 14,333
Q1 (low total greenspace cover)	ref	ref
Q2	1.082 (1.034, 1.133)	1.063 (1.014, 1.113)
Q3	1.098 (1.048, 1.152)	1.075 (1.024, 1.129)
Q4 (high total greenspace cover)	0.978 (0.930, 1.028)	0.972 (0.924, 1.024)
Q1 (low private residential gardens)	ref	ref
Q2	0.996 (0.952, 1.042)	0.988 (0.944, 1.034)
Q3	0.991 (0.946, 1.040)	0.984 (0.938, 1.033)
Q4 (high private residential gardens)	0.928 (0.881, 0.977)	0.932 (0.885, 0.983)
Total number of incident cases of T2D for access to parks,n = 19,648
Walkable road network distance to the nearest park
0–800 m	0.973 (0.940, 1.006)	0.980 (0.948, 1.014)
> 800–1500 m	ref	ref
> 1500 m	0.991 (0.939, 1.046)	0.992 (0.940, 1.047)
Euclidean distance to the nearest park
0–800 m	0.953 (0.915, 0.992)	0.962 (0.924, 1.001)
> 800–1500 m	ref	ref
> 1500 m	0.930 (0.858, 1.007)	0.938 (0.866, 1.016)
Number of parks within or intersected at an 800 m buffer (~ 10 minutes’ walk)
0–1	ref	ref
> 1–3	0.978 (0.947, 1.010)	0.981 (0.950, 1.013)
> 3	0.924 (0.888, 0.962)	0.943 (0.906, 0.981)

Model 1 adjusted for a minimal set of confounders identified through a DAG, including age, sex, education status, ethnicity, residential area, and the Townsend Deprivation Index. Model 2 adjusted for Model 1 confounders, lifestyle factors (smoking status, alcohol intake frequency, and healthy diet score), and health-related factors (family history of diabetes)

T2D type-2 diabetes, HR hazard ratio, CI confidence interval

Effect modification

The relationship between quartiles of private residential gardens and the incidence of T2D was modified by the Townsend derivation index, with stronger beneficial associations seen among individuals with higher scores compared to those with lower scores (Table 3). Family history of diabetes also modified the private garden-T2D association, but the beneficial effect was only observed among individuals in the highest quartiles of private gardens exposure who had no family history of diabetes. There were no potential effect modifiers identified in the two-way interactions for total greenspace cover (Table 3), walkable road network distance, Euclidean distance (Table S4) or number of parks (Table S5).

Table 3.

Effect modification of the associations between total greenspace cover and private residential gardens and the incidence of T2D

Modifying factors	Private residential gardens		total greenspace cover
	T2D HR (95% CI)	P-int.	T2D HR (95% CI)	P-int.
Age
< 65 years
Q1 (low private residential gardens/total greenspace cover)	ref	ref	ref	ref
Q2	0.972 (0.923, 1.025)	0.193	1.096 (1.040, 1.155)	0.272
Q3	0.972 (0.920, 1.027)	0.205	1.091 (1.032, 1.153)	0.700
Q4 (high private residential gardens/total greenspace cover)	0.930 (0.876, 0.987)	0.110	0.979 (0.923, 1.039)	0.389
≥ 65 years
Q1 (low private residential gardens/total greenspace cover)	ref		ref
Q2	1.041 (0.951, 1.140)		1.034 (0.944, 1.133)
Q3	1.039 (0.948, 1.139)		1.069 (0.975, 1.172)
Q4 (high private residential gardens/total greenspace cover)	1.013 (0.923, 1.113)		1.027 (0.934, 1.128)
Sex
Female
Q1 (low private residential gardens/total greenspace cover)	ref	ref	ref	ref
Q2	0.964 (0.901, 1.032)	0.396	1.071 (1.008, 1.139)	0.891
Q3	0.994 (0.928, 1.066)	0.596	1.087 (1.020, 1.157)	0.521
Q4 (high private residential gardens/total greenspace cover)	0.937 (0.870, 1.009)	0.863	0.989 (0.926, 1.057)	0.430
Male
Q1 (low private residential gardens/total greenspace cover)	ref		ref
Q2	1.002 (0.943, 1.064)		1.065 (0.995, 1.139)
Q3	0.970 (0.911, 1.033)		1.055 (0.983, 1.132)
Q4 (high private residential gardens/total greenspace cover)	0.929 (0.869, 0.993)		0.952 (0.884, 1.025)
Townsend Deprivation Index
≤median value
Q1 (low private residential gardens/total greenspace cover)	ref	ref	ref	ref
Q2	1.030 (0.925, 1.148)	0.026	1.083 (0.966, 1.214)	0.119
Q3	0.996 (0.898, 1.105)	0.044	1.010 (0.903, 1.130)	0.979
Q4 (high private residential gardens/total greenspace cover)	0.887 (0.801, 0.982)	0.988	0.900 (0.804, 1.007)	0.639
> Median value
Q1 (low private residential gardens/total greenspace cover)	ref		ref
Q2	0.900 (0.856, 0.946)		0.980 (0.931, 1.031)
Q3	0.883 (0.836, 0.933)		1.012 (0.958, 1.069)
Q4 (high private residential gardens/total greenspace cover)	0.886 (0.834, 0.947)		0.928 (0.873, 0.986)
Family history of diabetes
Yes
Q1 (low private residential gardens/total greenspace cover)	ref	ref	ref	ref
Q2	1.042 (0.957, 1.135)	0.120	1.073 0.985, 1.169)	0.903
Q3	0.972 (0.890, 1.063)	0.821	1.104 (1.011, 1.206)	0.435
Q4 (high private residential gardens/total greenspace cover)	1.016 (0.926, 1.114)	0.029	0.999 (0.910, 1.1096)	0.499
No
Q1 (low private residential gardens/total greenspace cover)	ref		ref
Q2	0.964 (0.914, 1.017)		1.066 (1.010, 1.125)
Q3	0.984 (0.931, 1.040)		1.061 (1.003, 1.122)
Q4 (high private residential gardens/total greenspace cover)	0.904 (0.851, 0.959)		0.963 (0.908, 1.021)

All models were adjusted for age, sex, education status, smoking status, alcohol intake frequency, ethnicity, residential area, healthy diet score, Townsend deprivation index, and family history of diabetes, except for the stratifying variable

T2D type-2 diabetes, HR hazard ratio, CI confidence interval, ref reference category, p-int p value for interaction 

Sensitivity analysis

Sensitivity analyses generally yielded similar results after: (i) removing participants with T2D within the first two years of follow-up (Table S6); (ii) further adjustment of waist circumference (Table S7); (iii) restricting analyses to individuals who had lived at their baseline address for ten years or more before recruitment (Table S8); (iv) excluding potential undiagnosed cases of diabetes based on HbA1c ≥ 48 mmol/mol (Table S9). Compared to the lowest quartile, higher levels of private gardens exposure were not associated with the incidence of T2D when alternative buffer sizes were used (Table S10). There was no association between park size near participants’ home location and the incidence of T2D (Table S11).

Discussion

When examining the effect of a specific type of greenspace, we found a reduced risk of T2D incidence among participants in the highest quartile of private residential garden exposure compared to those in the lowest quartile. The observed beneficial effect of private gardens was notable in the deprived individuals (i.e., those with higher Townsend Deprivation scores) and participants with no family history of diabetes. No association was observed between living within close (e.g., ≤ 800 m) or far (e.g.,>1500 m) walking distance to parks and the incidence of T2D, irrespective of the proximity measure applied (e.g., walkable road network distance versus Euclidean or straight-line distance). Despite the lack of association, we showed that using Euclidean distance as a measure of residential proximity to greenspaces may lead to exposure misclassification, and potentially, over-/under-estimation of the effect estimates. Beyond greenspace proximity, the number of parks within an 800 m buffer of participants’ home location was associated with a lower incidence of T2D.

Accumulating evidence has demonstrated the protective effects of greenspace on the incidence of T2D [21, 22, 24, 29, 51]. However, most prior longitudinal studies have operationalized residential greenspace exposure using proportion of land-use types or overall greenness (e.g., NDVI) [13], often overlooking the impacts of private gardens – an important type of greenspace that provides direct exposure and interaction with nature [27, 52, 53]. For example, a previous study used a generalized land use database (GLUD) greenspace within a 300 m buffer of UKBB participants’ home locations [21]. They omitted GLUD domestic gardens in their greenspace definition, despite these making up the majority of residential greenspaces in the UK [27, 28]. Not only do private gardens offer some similar exposure benefits to publicly accessible greenspaces, they also provide additional advantages, such as serving as spaces for food cultivation [52], space for social functions, interaction with companion animals, and prolonged outdoor play for younger children. The beneficial effects of these may include improved diet quality and, in turn, better blood glucose regulation, increased outdoor physical activity, and better exposure to bacterial biodiversity and psychological effects [54]. Recently, a cross-sectional study in the Netherlands found that a greater amount of garden greenery (at least 50 m²) was associated with a lower prevalence of diabetes (OR: 0.914; 95% CI: 0.872, 0.957), with stronger effect estimates observed in the highest categories (≥ 120 m²) of garden greenery (OR: 0.817; 95% CI: 0.878, 0.962) [55]. Our study builds upon these findings by using a longitudinal design with 15.41 years of follow-up to show a 7% reduced risk of incident T2D among participants in the highest quartile of private garden exposure, in a similar range to the Dutch study.

Previous studies examining the association between residential greenspace accessibility or proximity and the incidence of T2D are currently lacking [13]. A recent cross-sectional study using the UK Biobank data found no association between proximity to parks and the risk of cardio-metabolic multimorbidity (OR: 1.07; 95% CI: 0.99, 1.16) [30]. However, the study was limited by several factors. First, residential accessibility or proximity to parks was operationalized solely based on the presence or absence of public parks within a 300 m and 1500 m Euclidean buffer of participants’ home location, without accounting for road network buffers (i.e., actual distance, or travel time) or the number of parks. In acknowledging their limitations, the study suggested that future evidence integrating measures such as road network distances could provide a better and more accurate representation of mobility challenges faced by participants while enhancing our understanding of how accessibility affects health outcomes. Second, over 80% of the initial 502,429 participants were excluded from their study due to the unavailability of Urban Atlas (UA) data, or part of the residential buffer area falling outside the UA boundary, which could introduce bias. These limitations were addressed in the present study by calculating walkable road network distance and the number of parks for over 420,000 UK Biobank participants (using OS Open Greenspace - a high-resolution dataset of all publicly available parks within the UK) and examining their associations with the incidence of T2D.

While distance is considered an important factor when examining greenspace accessibility [37], previous studies suggest that living close to a park does not immediately translate to improved health outcomes [56, 57], or indeed visiting the parks. The latter was the case in our study, with findings indicating that neither walkable road network distance nor Euclidean distance was significantly associated with a lower risk of developing T2D. This is partly because greenspace works via many avenues, and some of those involve being in the park. The salutogenic effects of parks may depend on the type of park, individuals’ perception, and the varying patterns of park use. For example, parks that are poorly maintained or perceived as unsafe (with poor walkability) may preclude participants’ access and usage, irrespective of how close the park is to residential settings [57]. Another possible explanation may lie in the size of the park, which was not captured by distance alone. Evidence suggests that larger parks may encourage use by providing spaces for physical activity, social connection, and relaxation, as well as reducing environmental stressors – factors that promote physical health and well-being [49, 58, 59]. However, our sensitivity analysis examining the effect of park size within the nearest distance of participants’ home location found no association with the incidence of T2D. Kaczynski and colleagues have shown that parks with more features were more likely to be used, whereas park size and distance to the nearest park were not significant predictors of park use [60]. Wang et al. also demonstrated that physical and locational features, including the availability of an adequate number of parks and a pleasant walking experience, are the key factors influencing perceived residential accessibility to urban parks [61]. Other studies have suggested that park usage might depend on its aesthetic beauty and quality [37, 62, 63], including the quality of street greenspaces linking residential settings to the park [57]. However, a recent study has reported no significant association between park quality and the prevalence of diabetes [64].

Despite the null findings for distance to parks, the present study highlights the methodological importance of incorporating walkable road network distance (where possible) alongside other proximity measures, rather than relying solely on Euclidean distance in greenspace-health research. Although we observed a strong correlation between walkable road network distance and Euclidean distance (Figure S5), the use of the latter as a measure of proximity to parks may lead to exposure misclassification, misrepresentation of accessibility, and biased results, in part due to its inability to reflect the actual distance between spatial objects [32]. For example, we showed that walkable road network distances for UK Biobank participants were, on average, 43% longer than Euclidean distances to the nearest park, and that an acceptable limit of agreement between the two measures (i.e., “a reference range within which 95% of all differences between measurements using the two methods are likely to lie” [43]) was not met (Fig. 3). Also, our examination of the distribution of participants across the categories of distance to the nearest parks supported this. Using walkable road network distance, we found that 63.4% of the participants had parks within an 800 m buffer of their home location, whereas this proportion increased to 79.38% when Euclidean distance was used (Table 1). This discrepancy indicates that reliance on Euclidean distance may result in systematic overestimation of park accessibility, potentially obscuring true associations between greenspace exposure and health outcomes.

The observation that distance to parks did not show an association, whereas a greater number of parks did, suggests that park count plays a role in reducing the incidence of T2D, and perhaps, serves as a better indicator of greenspace accessibility and utility than proximity to parks. Access to a variety of parks supports different activities and attracts diverse users, potentially enhancing the chance that individuals would find a park that aligns with their preferences and activity needs. Prior research has indicated that respondents from eight countries, including the UK, who reside in areas with a high number of parks spend, on average, 24 min more on moderate to vigorous physical activities (MVPA) than those living in areas with the least number of parks [36]. Recently, it was found that an increase in the duration of accelerometer-derived MVPA among UK Biobank adults was associated with a progressive decline in the incidence of T2D [65]. Also, having a higher number of parks nearby not only improves greenspace accessibility, but also their usage. Evidence indicates that the number of parks (but not distance to nearest park) was positively associated with park use [66] and that increased park use is linked with reduced sedentary time [67], a factor implicated in metabolic risk [17, 68] and T2D [18].

Prior studies investigating the effect modification of neighbourhood SES on the association between public greenspaces and diabetes found conflicting evidence. While some studies have reported no significant interaction [23, 51, 69, 70], others have found stronger effect estimates among participants in low and medium-income neighbourhoods [71], as well as higher SES neighbourhoods [29]. In our study, we observed that the effect of private residential gardens (but not proximity to, and number of, parks) was pronounced among participants with higher Townsend deprivation index scores, suggesting that individuals from lower socio-economic groups might benefit more from such exposure. Residents from deprived neighbourhoods have limited lifestyle choices, and usually face challenges in accessing quality greenspaces, including the inability to afford housing with access to private gardens [27, 56]. As such, access and usage of even a small-scale private garden may have a profound impact on their health and well-being, possibly through pathways such as increased gardening and relaxation [52]. Gardening has the potential to promote increased (alternative) activity in harder-to-reach groups, such as individuals from lower SES [52]. We also found evidence of effect modification by family history of diabetes, with participants in the highest quartile of private residential garden exposure who had no family history of diabetes showing a greater reduction in the risk of T2D. This finding is not unexpected, as previous studies have shown that individuals with a family history of diabetes are at increased risk of developing T2D, with an even higher risk observed among those with a biparental history [72]. Additionally, empirical evidence suggests that the association between residential greenspace and T2D may vary according to genetic predisposition, with more pronounced protective effects observed among individuals with lower genetic risk [73]. Together, these findings suggest that higher baseline genetic or familial risk can attenuate the relative contribution of residential greenspace exposure, and may help explain the more pronounced beneficial association of private residential gardens among individuals without a family history of diabetes. In contrast to our findings, an earlier study using the same cohort reported no evidence of effect modification by family history of diabetes [21]. This discrepancy may be due to differences in exposure assessment, as that study used GLUD (which excludes domestic garden measures), whereas the present study specifically examined exposure to private residential gardens.

Our study has several strengths. We provided the first evidence of the effect of a specific type of greenspace (e.g., private residential gardens) and park accessibility in relation to the incidence of T2D using a longitudinal study design with a median follow-up of 15.41 years for over 420,000 UK Biobank participants. Accessibility to parks was determined using different objective measures, including walkable road network distance, number of parks, and size of the nearest park (in sensitivity analysis). Wang and colleagues reported that most studies on access to urban parks used park centroids as the destination points and that limiting a park polygon to a single point could lead to bias in the accessibility measures, particularly for large parks [31]. In our study, we considered all the access points where an individual could enter a park as destination points (Fig. 1) and matched them to their nearest vertices (network nodes) on the walking graph. This ensured a more accurate measure of the nearest walkable road network distance to parks. The walkable road network distance data, along with other greenspace accessibility measures, have been integrated into the UK Biobank and can be accessed by other researchers examining the impacts of park accessibility/proximity on the health and well-being of the UK Biobank adults.

Some limitations of the study should be taken into consideration. First, participants’ residential coordinates (eastings and northings) used for our analysis were restricted to a 100 m accuracy to the exact home location by the UKBB to protect privacy and preclude re-identification of participants. This restriction may have increased the likelihood of spatial inaccuracies, misclassification of proximity measures (e.g., 750 m vs. 850 m), and aggregation of some home location data into the same coordinates, thereby masking intra-neighbourhood variability, and potentially leading to under- or overestimation of our effects estimates. Second, we cannot rule out the uncertainty about whether the exposure preceded the outcome, given that the Ordnance Survey Greenspace dataset used in the present study was first released several years after the baseline assessment of the UKBB participants (2006–2010). Nevertheless, evidence has confirmed that greenspace cover in the UK remains relatively stable over time [28]. Third, although we found beneficial effects among participants in the highest quartile of private residential gardens exposure, we assume that these participants use these private spaces, even though they might hire a gardener or spend more time away from home [27]. Fourth, we did not have data on usage, perception of safety, and quality of parks, which may explain the lack of association found between distance to parks and incidence of T2D. As suggested by a previous study using the UK Biobank, assessing these measures imposes methodological constraints, particularly due to the large sample size [30]. Fifth, hospital inpatient data were used to define incident T2D, which may not capture some cases diagnosed and managed exclusively in primary care settings. As such, our approach may underestimate the true incidence of T2D and may preferentially identify more severe or clinically recognised cases. However, incident T2D has been ascertained in previous UK Biobank studies using hospital inpatient data [21, 74], which includes information not only on participants’ admissions but also on diagnoses (and underlying conditions) coded using ICD-9 and ICD-10. Finally, our analysis was based on participants’ home location at baseline (2006–2010), and we could not account for home addresses during the follow-up period, as some participants may have moved. We assessed the robustness of our findings by performing a sensitivity analysis for participants who had lived at their baseline address for ≥ 10 years before recruitment.

Conclusion

In this prospective cohort study, we demonstrated that participants whose domicile was in the highest quartiles of percentage of private residential gardens exposure had a lower risk of developing T2D compared to those in the lowest quartile. Disadvantaged groups (e.g., participants living in deprived neighbourhoods) and individuals with no family history of diabetes appear to benefit more from increased private garden exposure. We also revealed that having a greater number of parks closer to home, rather than the nearest distance to parks, may play a role in reducing the incidence of T2D. Taken together, our study highlights the need for integrating private residential gardens alongside accessible greenspaces (e.g., public parks or community gardens) in urban planning and public health. This could play a role in reducing and managing the burden of chronic diseases, such as diabetes, among adults.

Abbreviations

T2D

Type-2 diabetes

UKBB

United Kingdom Biobank

GBD

Global burden of disease

YLL

Years of life lost

DALY

Disability-adjusted life years

HbA_1c

Glycated haemoglobin

NDVI

Normalized difference vegetation index

SES

Socioeconomic status

HES

Health Episode Statistics

PEDW

Patient Episode Database for Wales

SMR

Scottish Morbidity Records

ICD-10

International classification of diseases, tenth edition

OS

Ordnance Survey

OSM

OpenStreetMap

HR

Hazard ratio

CI

Confidence interval

OR

Odds ratio

P-int

P value for interaction

IQR

Interquartile range

DAG

Directed acyclic graph

STROBE

Strengthening the reporting of observational studies in epidemiology

IF

Infection point

GLUD

Generalised land use database

UA

Urban Atlas

MVPA

Moderate-to-vigorous physical activity

Authors' contributions

CCO: conceptualized the study, led the investigation, developed the methodology, and was responsible for data curation, formal analysis, visualization, validation, and writing the original draft. DD: contributed to the conceptualization and methodology of the study and provided supervision. CR: contributed to conceptualization and methodology. SR: provided supervisory support. NJO: contributed to conceptualization, methodology, supervision, and project administration. All authors contributed to writing – review & editing of the manuscript and approved the final manuscript.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The data used in this study cannot be shared or transferred to any other person due to confidentiality restrictions outlined in the Material Transfer Agreement we signed with the UK Biobank. General inquiries regarding the UK Biobank data can be directed to [access@ukbiobank.ac.uk](mailto: access@ukbiobank.ac.uk) .

Declarations

Ethical approval and consent to participate

The current study was assessed as exempt from ethics review by the University of Queensland Health Research Ethics Committee (No: 2022/HE002005). All participants provided their informed consent for the use of their data in research investigations.

Consent for publication

Not applicable.

Competing interests

Dr Charlotte Roscoe is an Editor for the Collection (Greenspace, Biodiversity and Health) in BMC Public Health. The authors declare no other competing interests.