Article Text
PDF
Abstract
Background People living in highly walkable neighbourhoods tend to be more physically active and less likely to be obese. Whether walkable urban design reduces the future risk of diabetes is less clear.
Methods We used inverse probability of treatment weighting to compare 10-year diabetes incidence between residents living in high-walkability and low-walkability neighbourhoods within five urban regions in Ontario, Canada. Adults (aged 30–85 years) who were diabetes-free on 1 April 2002 were identified from administrative health databases and followed until 31 March 2012 (n=958 567). Within each region, weights reflecting the propensity to live in each neighbourhood type were created based on sociodemographic characteristics, comorbidities and healthcare utilisation and incorporated into region-specific Cox proportional hazards models.
Results Low-walkability areas were more affluent and had more South Asian residents (6.4%vs3.6%, p<0.001) but fewer residents from other minority groups (16.6%vs21.7%, p<0.001). Baseline characteristics were well balanced between low-walkability and high-walkability neighbourhoods after applying individual weights (standardised differences all <0.1). In each region, high walkability was associated with lower diabetes incidence among adults aged <65 years (overall weighted incidence: 8.2vs9.2 per 1000; HR 0.85, 95% CI 0.78 to 0.93), but not among adults aged ≥65 years (weighted incidence: 20.7vs19.5 per 1000; HR 1.01, 95% CI 0.91 to 1.12). Findings were consistent regardless of income and immigration status.
Conclusions Younger adults living in high-walkability neighbourhoods had a lower 10-year incidence of diabetes than similarly aged adults living in low-walkability neighbourhoods. Urban designs that support walking may have important benefits for diabetes prevention.
- cohort studies
- diabetes
- environmental health
- epidemiological methods
- neighborhood/place
Statistics from Altmetric.com
Introduction
Obesity rates have increased dramatically worldwide leading to a global epidemic of diabetes and other obesity-related diseases. According to WHO, an estimated 1.9 billion adults were overweight in 2014, with 600 million meeting the definition of obesity.1 While the causes of such trends are complex, population-level shifts in lifestyle have played a central role, fuelled by an environment that encourages sedentary behaviour and the overconsumption of high-calorie convenience foods.
One avenue that has garnered considerable interest in recent years is the capacity for urban design to influence walking, cycling and other forms of physical activity.2–4 An international study of 14 cities found that participants living in the most activity-friendly areas—characterised by high walkability and greater access to public transit and parks—spent more time engaged in physical activity and were more likely to achieve the minimum recommended amount of physical activity.5 Furthermore, neighbourhoods that are highly walkable appear to have lower levels of obesity.6 7
Several studies have also reported an association between neighbourhood walkability and diabetes.6–13 However, few studies have been prospective or had a sufficiently large sample size, and none adequately accounted for systematic differences in risk factor prevalence between individuals living in high-walkability versus low-walkability settings. To address these gaps, we employed a statistical method based on propensity scores, known as Inverse Probability of Treatment Weighting (IPTW), to examine the association between neighbourhood walkability and diabetes incidence. Propensity score methods, like IPTW, were designed to minimise confounding in observational studies when randomised trials are unfeasible or impossible.14 15 This research was conducted using population-level data from 15 municipalities in Southern Ontario. This region is one of the fastest growing urban areas in North America, with a current population of approximately 7 million, which is expected to reach nearly 10 million by 2040.
Methods
Setting and population
This study used administrative healthcare databases from Ontario, Canada to identify all adults (aged 30–85 years) living in our study area who were free of diabetes on 1 April 2002. Hospital, physician and laboratory services are provided to all permanent residents through Ontario’s government-funded health system, without co-payments or user fees. Thus, these records can be used to track health outcomes for individual residents. For this study, we included individuals living in neighbourhoods with very high and very low walkability scores within five urban regions in Southern Ontario: Toronto, the Greater Toronto Area (communities adjacent to Toronto), Hamilton, London and Ottawa. We excluded individuals living in chronic care institutions. We also excluded those with healthcare coverage for <3 years prior to baseline to ensure sufficient healthcare data were available to ascertain subjects’ diabetes status at baseline.
These datasets were linked using unique, encoded identifiers and analysed at the Institute for Clinical Evaluative Sciences (ICES). This protocol received ethical approval from ICES and the institutional review boards at St. Michael’s Hospital and Sunnybrook Health Sciences Centre in Toronto.
Neighbourhood walkability
A previously validated walkability index was used to identify residents living in the most and least walkable neighbourhoods within each study region on 1 April 2002. Creation of our walkability index has been described previously.6 7 Briefly, the index was created as a composite score based on four attributes: (1) population density (number of people per square kilometre); (2) residential density (number of occupied residential dwellings per square kilometre); (3) walkable destinations (number of retail stores, services (eg, libraries, banks, community centres) and schools within a 10-min walk) and (4) street connectivity (number of intersections with at least three converging roads or pathways). Scores for each of these attributes were created using ArcGIS software (V.10.2) and data from the 2001 Canadian Census (for population and residential density counts), 2003 DMTI Spatial Enhanced Points of Interest file (for street connectivity and retail/service locations) and data from each municipality and the Ontario Ministry of Education (for public recreational facility and school locations). Values were derived for small parcels of land defined by Statistics Canada (dissemination areas) comprising one or more adjacent city blocks, combined into a composite score, and divided into quintiles from lowest (Q1) to highest (Q5). Dissemination areas are fairly homogeneous in terms of social characteristics and have a population size of approximately 400–700 residents.
Primary outcome
Individuals were followed from baseline (1 April 2002) until 31 March 2012 for the development of diabetes. New diabetes cases were identified from the Ontario Diabetes Database (ODD)—a validated registry based on hospitalisation records and physicians’ services claims.16 The ODD’s algorithm has a high level of sensitivity (86%) and specificity (97%) for identifying physician-diagnosed cases of diabetes. The ODD is unable to distinguish cases of type 1 from type 2 diabetes; however, the vast majority of new cases in our cohort would be expected to be type 2 given the minimum age criteria of 30 years.
Baseline variables
Baseline sociodemographic and clinical characteristics were used to account for confounding, since those living in high-walkability neighbourhoods may differ systematically from those who live in low-walkability neighbourhoods. Baseline variables were chosen for inclusion in the propensity score models that were thought to be prognostically important (ie, factors shown previously to be related to diabetes incidence).17
Sociodemographic variables
Sociodemographic characteristics included age, sex, socioeconomic status (SES), ethnicity, immigration status and city/town of residence. Administrative healthcare databases contained age, sex and city/town of residence for all individuals. These databases were also linked to federal immigration data to identify members of our cohort who immigrated to Canada <10 years before baseline. Neighbourhood-level measures of SES and ethnicity derived from the Canadian Census were used in place of individual measures, as these are not available in administrative records. We included two multidimensional measures of SES from the Ontario Marginalisation Index: 1) material deprivation, based on neighbourhood characteristics (eg, poverty, education, unemployment, dwellings in disrepair) and 2) neighbourhood dependency, based on the relative population composition with respect to seniors, children and people not in the labour force.18 Because of the high number of South Asian immigrants in our study population and the heightened risk of diabetes in this group, we included the proportion of area residents of South Asian ethnicity based on census estimates as a covariate, as well as the proportion who were of non-South Asian, non-white ethnicity.
Clinical variables
Several clinical variables were included to account for differences in health status between populations living in high-walkability and low-walkability areas, including some (cardiovascular disease and hypertension) that are strongly associated with obesity and diabetes (see online supplementary eTable 1). Recent cardiovascular events were defined as an admission for acute myocardial infarction, stroke or revascularisation within the 3-year period prior to baseline. Individuals with a prior diagnosis of hypertension at baseline were identified using a validated algorithm, similar to the ODD. Comorbidities were captured using the Johns Hopkins Adjusted Clinical Groups case-mix system to create distinct categories (collapsed ambulatory diagnostic groups) based on diagnostic codes for conditions other than diabetes in the year prior to baseline.19 Health service utilisation variables (annual number of hospitalisations and primary care visits) were determined over the same time period.
Supplemental material
Analysis
In a previous study, the relation between walkability and transportation behaviours (rates of walking, cycling and transit use vs driving) appeared to be non-linear, and significantly lower levels of obesity were seen only in the highest quintile of walkability.6 Therefore, our primary analysis focused on residents living in the top and bottom quintile of walkability only. Inverse probability of treatment weighting was used to account for observed systematic differences in baseline characteristics between those who lived in high-walkability (Q5) versus low-walkability (Q1) neighbourhoods.14 Inverse probability of treatment weighting and propensity score matching have been shown to remove the effects of confounding to a similar degree.15 However, a general requirement for propensity-score matching to perform well is that the number of potential control subjects (eg, those living in low-walkability neighbourhoods) is substantially higher than the number of exposed subjects (eg, those living in high-walkability neighbourhoods). However, this condition was not satisfied in our design, as the two groups each contain about 20% of the population. For these reasons, we elected to use IPTW. The first step in creating inverse probability of treatment weights involved the development of propensity scores that reflect the probability of living in a high-walkability neighbourhood, conditional on measured baseline covariates. To do so, a logistic regression model was fit, in which living in a high-walkability versus low-walkability neighbourhood was regressed on the baseline covariates described above. Stabilised Inverse Probability of Treatment Weights were then computed for each individual using the following equations: 
for those living in high-walkability areas and 
for those living in low-walkability areas, where e denotes the propensity score and Pr(Z=1) denotes the proportion of the sample that lived in a high-walkability neighbourhood. Because individuals living in different regions may be systematically different from one another, we derived the inverse probability of treatment weights separately for the sample population in each of the five regions: Toronto, the Greater Toronto Area, Ottawa, Hamilton and London.
Standardised differences were used to assess the performance of the propensity score models.20 To do so, standardised differences were generated to compare the baseline characteristics of individuals from high-walkability and low-walkability areas before and after applying inverse probability of treatment weights.21 Standardised differences reflect differences between group means and proportions (or weighted means and proportions in the case of inverse probability of treatment weighting) relative to the pooled SD; values >0.1 were interpreted as indicating a meaningful difference between groups.
Furthermore, we assessed the distribution of inverse probability of treatment weights to identify outliers. Truncating weights at the 99th percentile did not appreciably alter model outputs; therefore, only analyses which included all weights generated by the propensity score models (ie, non-truncated weights) are reported herein.
We estimated the effect of high-walkability versus low-walkability on the incidence of diabetes in each of the five regions separately. For the main analysis, a Cox proportional hazard model was fit to the study sample in each region—in which the outcome, the hazard of developing diabetes, was regressed on an indicator variable denoting neighbourhood walkability. The model incorporated the inverse probability of treatment weights, and used a robust variance estimator to account for within-subject homogeneity induced by the weighting, since the process of weighting results in some individuals’ attributes and outcomes to be counted more than once. For example, in the case where an individual is assigned a weight of 2, two copies of that individual would be included in the analysis. The inherent correlation of observations within weighted subjects needs to be taken into account.22 Because prior research suggested that the walkability effect is moderated in older age groups, analyses were conducted separately in individuals aged under 65 years and those aged 65 years and older.
We then conducted a meta-analysis using random effects models for each age group (<65 and≥65 years) to derive summary HRs from all five regional models. To account for any residual confounding, we also fitted age group-specific, weighted Cox proportional hazard models for each region that adjusted for all baseline variables in the model. Estimates derived from these fully adjusted models were also combined using a random effects model to derive an overall adjusted summary HR for each age group.
To test whether the observed effects were consistent across higher and lower SES neighbourhoods, we repeated our Cox proportional hazards models, stratifying by area SES (upper two vs bottom two quintiles of material deprivation). We repeated this process stratifying instead by immigration status to assess whether findings were consistent among recent immigrants (in Canada ≤10 years prior to baseline) and long-term residents (Canadian-born and those immigrating to Canada >10 years prior to baseline). The inverse probability of treatment weights generated in the steps above were used in these subgroup analyses.
To examine whether the effects of walkability were non-linear, we repeated the above process to create new inverse probability of treatment weights in order to compare individuals living in each upper quintile (Q2, Q3, Q4 and Q5) with those in Q1. For this set of analyses, weighted Cox regression models were fit for each age group in the population overall, both with and without adjusting for baseline covariates. The models were not stratified by study region but region was included in the fully adjusted models.
Sensitivity analysis
As a sensitivity analysis, we repeated our models without inverse probability of treatment weights but instead adjusted for baseline variables in the model. Models were fit for each age group in the population overall and were not stratified by study region. These unweighted multivariable Cox models adjusted for all of the variables that were incorporated in the propensity score model in the primary analysis.
All analyses were performed using SAS software (V.9.3, SAS Institute, Cary, North Carolina, USA). All statistical tests were two-sided with a threshold for significance of p<0.05.
Results
Overall, the sample included 958 567 individuals living in high-walkability and low-walkability areas in Southern Ontario. Over 40% of the sample was from Toronto (n=411 878), with the rest from the Greater Toronto Area (298 532), Ottawa (n=128 237), Hamilton (n=85 117) and London (n=34 803). The majority of individuals in the sample were 30–64 years of age (n=821 296).
Many characteristics of individuals from high-walkability and low-walkability areas were comparable at baseline, even prior to applying inverse probability of treatment weights (see online supplementary eTable 2). However, high-walkability areas were more likely to be of low SES and had a somewhat lower percentage of residents who were South Asian, although a higher percentage of residents from another non-white ethnicity. After applying inverse probability weights, residents living in high-walkability and low-walkability neighbourhoods, overall and within each region, were extremely well balanced for all measured baseline attributes (standardised differences<0.1), except for area dependency (table 1). The balance between groups was apparent in both younger (<65 years) and older (≥65 years) individuals. Healthcare utilisation was high in all groups regardless of area walkability.
Weighted* baseline characteristics of study population by neighbourhood walkability
Primary analyses
After a median follow-up of 9.2 years, 90 922 new cases of diabetes were observed—68 801 among those aged under 65 years and 22 121 among those aged 65 years and older. Among younger individuals (<65 years), high walkability was associated with a significantly lower rate of developing diabetes overall (summary HR 0.88, 95% CI 0.81 to 0.97) with fairly consistent effects across study regions (figure 1A). Fully adjusting for baseline covariates to account for residual confounding resulted in similar findings (summary HR 0.85, 95% CI 0.78 to 0.93; table 2 and figure 1A). This association persisted when groups were stratified by immigration status and area SES (figure 2). Walkability appeared to have no significant impact on the rate of developing diabetes among older adults (figure 1B and table 2) (summary HR 0.99, 95% CI 0.88 to 1.13).
Diabetes incidence in high-walkability vs low-walkability areas, by region and age group. Models incorporated inverse probability of treatment weights. Unadjusted models included walkability; adjusted models included walkability and adjusted for all baseline covariates from which inverse probability of treatment weights were derived. GTA, Greater Toronto Area, excludes Toronto.
Diabetes incidence in high-walkability vs low-walkability neighbourhoods among adults aged 30–64 years, by length of time in Canada and neighbourhood deprivation. *Based on time of immigration from the Immigration, Refugees and Citizenship Canada Database. †Based on the material deprivation dimension of the Ontario Marginalisation Index. Models incorporated inverse probability of treatment weights and adjusted for all baseline covariates from which inverse probability of treatment weights were derived.
Weighted and non-weighted incidence of diabetes by neighbourhood walkability and age group
Among younger adults, the association between neighbourhood walkability and diabetes incidence appeared to be non-linear. Compared with individuals living in areas within the lowest walkability quintile (Q1), those living in neighbourhoods that were in the second highest walkability quintile (Q4) had a slightly reduced incidence of diabetes, while those in the middle and second lowest walkability quintile (Q3 and Q2) had a slightly increased incidence of diabetes (figure 3).
Diabetes incidence by walkability quintile (Q) and age group. Models incorporated inverse probability of treatment weights and adjusted for all baseline covariates from which inverse probability of treatment weights were derived.
Sensitivity analysis
In analyses that did not incorporate inverse probability of treatment weights, the rate of developing diabetes was again lower among younger populations (aged <65 years) living in high-walkability compared with low-walkability areas (HR 0.86, 95% CI 0.85 to 0.88) after adjusting for baseline covariates (table 2). The magnitude of these effects was similar to those observed when inverse probability of treatment weighting was used, with some small differences in regional estimates (data not shown). In this analysis, seniors living in highly walkable areas had a small but significant increase in the rate of developing diabetes (HR 1.04, 95% CI 1.01 to 1.07) compared with those living in low-walkability areas (table 2).
Discussion
Young and middle-aged adults living in highly walkable urban neighbourhoods in Southern Ontario had a significantly lower 10-year incidence of diabetes compared with those living in low-walkability areas—after accounting for systematic differences in the health status of individuals who choose to live in one neighbourhood type over another, and other neighbourhood contextual effects influencing their risk of diabetes.
These findings are plausible given the observed association between walkability, physical activity and obesity. In the International Physical activity and Environment Network study, participants living in areas with higher levels of walkability, greater numbers of public transit options and more parks accrued on average 1–1.5 more hours of moderate to vigorous physical activity per week than those living in areas with low levels of these features.5 A similar phenomenon has been observed in Canada.23 In prior research, our neighbourhood walkability scores were directly associated with the number of daily walking, cycling and public transit trips per capita among local residents and inversely associated with the number of car trips per capita.6 7 Transportation patterns mirrored differences in obesity and diabetes incidence across high-walkability and low-walkability neighbourhoods, whereas information on diet and non-transportation-related physical activity levels did not. Highly walkable neighbourhoods may offer more opportunities to incorporate physical activity into daily life by promoting active transportation and reducing the time per day spent sitting in automobiles, which, itself has been linked to obesity.24 25
Epidemiological studies of the built environment and health are frequently criticised for their inability to control for the self-selection of individuals into their neighbourhood of residence. In the context of walkability and diabetes, our research is the first to address this criticism by applying a propensity score-based technique to control for systematic differences in known confounders between individuals living in high-walkability versus low-walkability areas. A prior Swedish study reported an increased incidence of diabetes among Stockholm residents living in low-walkability areas, but—unlike our findings—this association was no longer significant after adjusting for household income.13 Cultural attitudes towards walking, urban development practices and cycling and transit infrastructure vary considerably between European and North American settings, which may have contributed to this difference. A time series analysis conducted in the same setting as ours found more favourable rates of overweight and obesity and a lower incidence of diabetes in highly walkable neighbourhoods; however, given the ecologic nature of the study design, the impact of walkability on the future health of individual residents was not examined.6 However, rates of active transportation (walking, cycling and public transit use) were substantially increased in highly walkable communities. Several other studies have also noted a higher rate of type 2 diabetes among residents living in areas that were less walkable or were perceived to be.7–12 26 However, most of these studies were small or cross-sectional in nature, and none accounted for systematic differences in the health status of populations living in different communities or their tendency to seek healthcare. By addressing such confounding, our research provides strong additional evidence of an association between walkability and diabetes, as well as important differences in this association by age.
Prior studies on this topic have not specifically examined this relationship among older adults. Our study suggests walkability—as we measured it—is not associated with diabetes risk in older populations. Prior research shows that older individuals living in highly walkable areas walk more than those in less walkable areas; although the effects appear to be more modest than in younger populations.20 27 Studies examining the association between walkability and obesity in older populations have yielded mixed results.28 29 However, although such a relationship may exist, it may not translate to a lower risk of diabetes. Diabetes incidence peaks around age 70 years, primarily because of the effects of ageing on insulin secretion.30 Thus, while obesity is a major risk factor for type 2 diabetes, its influence wanes at older ages. There may also be a survivor effect, whereby obese individuals who are highly prone to develop diabetes will likely do so at an earlier age.
This study’s strengths include its large, population-level design and the use of inverse probability of treatment weighting to account for systematic differences in measured confounders. Using inverse probability of treatment weighting, it is possible to explicitly assess the degree to which weighting has made the two treatment groups comparable in terms of measured baseline covariates. It is much more difficult to assess the extent to which adjustment of baseline covariates accounts for systematic differences in treatment groups when using conventional multivariable regression analysis. Furthermore, the target estimated when using IPTW is the marginal treatment effect (the average effect of the treatment on the population), which is the same estimate as in randomised controlled trials, assuming all confounding has been measured, whereas multivariable adjustment estimates conditional effects.
However, this study has several limitations that merit discussion. While inverse probability of treatment weighting mitigated the role of measured confounders in mediating the relationship between walkability and diabetes, it is possible that there was some degree of imbalance in unmeasured confounders that contributed to these findings. For example, there were notable imbalances with respect to area SES, although these tended to favour low-walkability areas which were generally wealthier. We also relied on area-level measures of income and ethnicity in place of individual measures, which were lacking in our data sources. Furthermore, our models may have been susceptible to overadjustment since obesity-related conditions, such as hypertension and cardiovascular disease, were included in the propensity score models; if so, we may have underestimated the effect of walkability on diabetes. Our study did not incorporate other contextual factors such as the retail food environment, which may influence the risk of obesity and diabetes independently.31 32 Because we assigned treatment weights based on baseline residential location, it is possible that some participants moved during the follow-up period. However, prior research suggests that residents tend to move to neighbourhoods with similar social and physical characteristics.33 Lastly, although there is strong evidence from an international study linking walkability to physical activity,5 it is unclear whether the association between walkability and diabetes incidence observed in our study is generalisable to other settings.
In Southern Ontario, Canada, adults aged under 65 years living in high-walkability neighbourhoods had a significantly lower rate of developing diabetes than similarly aged individuals living in low-walkability areas. This study was the first to demonstrate the prospective impact of walkability on future diabetes risk using a propensity score-based method to control for systematic differences in the population characteristics of those living in neighbourhoods of high and low walkability. There is growing evidence that policies that promote walking, cycling and public transit and reduce driving could have tangible health benefits through a reduction in chronic disease.2–4 Achieving such outcomes will require a multifaceted strategy targeting multiple sectors and levels of government to support healthy community design, as well as the political and public will to sustain it. Further research is needed to evaluate natural policy experiments as they occur to understand the full population health impact of changes to the built environment.
What is already known on this subject
There is mounting evidence linking neighbourhood walkability with physical activity, overweight and obesity.
More recently, there is evidence to support a link between neighbourhood walkability and diabetes; however, few studies have compared the risk of developing diabetes among individual residents according to area walkability.
Furthermore, a major criticism of epidemiological studies examining the built environment and health is their inability to control for self-selection into residential neighbourhoods.
What this study adds
Our study addresses this gap by using Inverse Probability of Treatment Weighting analysis—a statistical technique based on propensity scores that was designed to minimise confounding in observational studies when randomised trials are unfeasible or impossible.
Even after accounting for systematic differences in a broad array of baseline variables (including sociodemographic characteristics, health status and health services use), the 10-year incidence of diabetes was significantly lower among younger adults (aged <65 years) living in highly walkable neighbourhoods compared with those living in low-walkability areas.
Moreover, ours is the first study to demonstrate a differential association between walkability and diabetes risk in younger versus older populations.
Supplemental material
References
Footnotes
Contributors GLB had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: GLB, MIC and PCA. Acquisition, analysis or interpretation of data: GLB, MIC, JL, GSF, AJ, LCR, RHG, RM, PG and PCA. Drafting of the manuscript: GLB and AJ. Critical revision of the manuscript for important intellectual content: GLB, MIC, JL, GSF, AJ, LCR, RHG, RM, PG and PCA. Statistical analysis: JL and PCA. Obtained funding: GLB. Administrative, technical or material support: GSF and AJ. Study supervision: GLB.
Funding This study was funded through an open operating grant from the Canadian Institutes of Health Research (CIHR).
Competing interests None declared.
Patient consent Not required.
Ethics approval This protocol received ethical approval from ICES and the institutional review boards at St. Michael’s Hospital and Sunnybrook Health Sciences Centre in Toronto.
Provenance and peer review Not commissioned; externally peer reviewed.