Conceptual framework

We start with a theoretical basis of the well-known public health triad of environment, behaviours and health outcomes. Health outcomes are influenced by health behaviours, which in turn are associated with the environment. We summarise this in figure 1. In Australia and elsewhere, a number of research papers have established the relationships between environment and behaviours (link A—see figure 1)14–18 or behaviours and health outcomes (link B—see figure 1).19 ,20 It logically follows that the environment is related to health outcomes through the individual lifestyle behavioural pathway. In addition, the built environment may directly influence health outcomes. For example, air pollution may be detrimental to respiratory and cardiovascular health,21 or perceptions on the environment may affect mental health.22 However, research on this relationship (link C—see figure 1) is limited, with most research, excepting a few,23 ,24 focusing on outcomes related to sedentary health behaviours such as obesity25 ,26 and conditions directly related to obesity.27 Our interest, therefore, was in investigating this relationship (link C—figure 1) between aspects of the physical environment and the four major NCDs in the ACT: circulatory system diseases, specific cancers, endocrine, nutritional and metabolic diseases (ENMDs) and respiratory diseases, using geocoded ACT hospitalisation data (from 2007 to 2013) and specific built environmental attributes. Note, however, that link C is mediated through multiple pathways, such as through health behaviours, and link C represents any relationship between environmental exposures and the chronic conditions described above, irrespective of mediating pathways.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Framework of relationships between environment, behaviours and health outcomes.

Investigating relationships

To investigate relationships between the built environment and NCD-related hospital admissions, we followed a combined exploratory-inferential approach. First, we asked ‘What are the spatial patterns of the four key chronic conditions in the ACT?’. This is addressed through exploratory mapping using spatial cluster analysis. Second, we investigated relationships between various individual and environmental predictors such as neighbourhood walkability, traffic volume and access to off-license alcohol outlets and the key NCD-related hospital admissions in the ACT. In the next section, we explain in detail the methods used to achieve this.

Data

Population data

In addition to the above data, population data were required for mapping rates of hospital admission. The smallest geography at which Australian demographic data (eg, age, gender, SES) are released is the SA1 (with an average of 500 people). SA1 is therefore a relatively small geographic area at which NCD-related hospital admission rates could be mapped. However, there were relatively smaller numbers of neoplasm and MI cases (table 1); hence, these conditions required a larger geography—the SA2 for mapping—because rates based on small numbers of expected cases are unstable and have large CIs. In this study, the term suburb is used to define the spatial boundary defined by the ABS in 2011 as SA2. Therefore, we aggregated up to the SA2 (suburb) level. In addition, while ENMDs and CSDs can be mapped at SA1s annually given their large annual numbers in the ACT (table 1), aggregate sums over multiple years were used for MI and neoplasms.

Australian census output geographies changed significantly between 2006 and 2011. While there are minimal differences between 2011 SA2 geographies and their 2006 counterpart Statistical Local Areas (SLAs) in the ACT,33 there was significant spatial mismatch between 2011 SA1s and their 2006 counterpart in the census hierarchy—Collection Districts (CDs). Thus, when mapping by SA1s or CDs (ENMDs, respiratory diseases and CSDs), we show separate maps for 2006 and 2011. Age-specific 2011 population counts at SA1s and 2006 counts at CDs were obtained from the ABS. For SA2 level maps of neoplasms and MI, counts of expected numbers of cases for the years 2007–2011 were required. Age-specific 2011 population counts and 2006 population counts were obtained at SA2s/SLAs. To obtain the age distribution for the intermediate years (2007–2011) at SA2s, we linearly interpolated the numbers in each SA2/age group between 2006 and 2011. This generated the fraction of people in each age group in a given year in an SA2. We then used an indirect age standardisation technique to calculate annual expected numbers of cases of an NCD using the annual age distributed ACT population as the standard population.34 Expected annual numbers were also calculated for the CD, SA1 and SA2 data. We used 2006 expected counts when mapping 2007 hospitalisation data since 2007 SA1 or CD population counts were not available.

Environmental data

As summarised in figure 1, we wanted to investigate relationships between various built environmental attributes and health events (hospital admissions). A number of environmental covariates were collected, collated and/or created in-house by the authors. Our choices of environmental drivers were informed by previous research but also constrained by the available data. For example, we did not have geocoded data for food outlets and hence could not explore any relationships between hospital admissions and the food environment. The environmental indices that were available are described below:

1. Walkability: walking is the most prevalent form of physical activity in the population.35 ,36 The degree of neighbourhood walkability predicts the degree of walking.37 We measured the physical activity environment through suburb-level walkability. While other aspects of the physical activity environment such as access to parks and leisure/exercise centres are also important, the walking network remains one of the most important built environmental attributes for overall physical activity.13 Walk Score is a measure of walkability produced by a United States (US)-based company that has been validated37 and has been used in a number of public health studies in the US. In the Australian context, it has been found to have strong relationships with walking for transport in a recent study,14 though relationships with health outcomes have not previously been found.23 Walk Score is a composite measure of destination density. The scores are normalised to a 0–100 scale, with 0 being the lowest walkability and 100 being the highest. A five-scale categorisation is used: ‘walkers paradise’ (Walk Score 90–100), ‘very walkable’ (70–89), ‘somewhat walkable’ (50–69), ‘car-dependent’ (25–49) and ‘car-dependent’ (0–24) by the developers of Walk Score,38 and these categories have been used by other researchers.16 Walk Scores for ACT suburbs/SA2s were obtained from the Walk Score website.38 A map of Walk Scores at ACT suburbs is provided in figure 2.

2. Access to general practitioners (GPs): access to primary care is an important predictor of admittance to tertiary facilities.39 ,40 Access to GPs is related to better health management and lesser use of hospital services.39 ,41 We created an access measure by drawing a circular buffer around the Mesh Blocks of the patients in the hospitalisation data. The circular buffers around the Mesh Blocks adaptively grew to different sizes, with each buffer growing until a total of 1000 people were included in the circle. The numbers of GP clinics in the buffer circles were then summed to provide an approximate measure of access as the number of GP clinics per thousand persons. GP clinic data for 2010 were provided by the ACT Medicare Local, while underlying 2011 census population data were obtained from the ABS.

3. Neighbourhood SES: neighbourhood SES is a well-established marker of social environment including crime and social cohesion and a mature literature supports the relationship between neighbourhood SES and a range of health outcomes.42 The Socio-Economic Indexes for Areas (SEIFA) are indices of area level of SES in Australia developed by the ABS. The Index of Relative Socio-Economic Advantage and Disadvantage (IRSAD) is one such index that measures both advantage and disadvantage. The index was created by incorporating a number of measures including per cent unemployed, car ownership and per cent disabled. SA1 IRSAD scores for 2011, the finest resolution at which they are available, were incorporated into these analyses.

4. Alcohol outlets: along with the food environment, alcohol outlets are powerful predictors of lifestyle-related health outcomes.43 While the food environment is best represented by summary measures of access to a range of food outlets, we did not have access to an integrated, clean, geocoded data set of food outlet locations in the ACT for this study (see Discussion section). Easy access to alcohol has been related to a number of negative health and social outcomes,44 ,45 and we have used a measure of alcohol access in our analyses. A list of all licensed off-licence liquor outlets was obtained from the ACT Department of Regulatory Services46 and geocoded to the SA1. Off-licence outlets are licensed to sell alcohol, but alcohol cannot be consumed within premises, examples of which include supermarkets and bottle shops. The road network distance from each residential parcel within each SA1 to the nearest off-licence liquor establishment was calculated. The mean distance for all residential parcels per SA1 was then derived. Off-licence outlets were included if they were within the same ACT defined district as the SA1 of interest.

5. Road traffic exposure: the presence of road traffic can act as an impediment to physical activity in a neighbourhood environment.47 Road traffic exposure was based on a ratio of road hierarchy (as a proxy for traffic volume) by length of road segments within an SA1. Methods for this have been published previously.47

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Map of five categories of Walk Score by ACT suburbs. The five categories are ‘walkers paradise’ (Walk Score 90–100), ‘very walkable’ (70–89), ‘somewhat walkable’ (50–69), ‘car-dependent’ (25–49) and ‘car-dependent’ (0–24). ACT, Australian Capital Territory.

Analysis

Spatial patterning of hospital admissions related to NCDs was explored using a cluster detection tool, the Spatial Scan Statistic.48 Monte Carlo regression was then employed to investigate relationships between NCD-related hospitalisations and built environmental factors.29 ,49 Finally, a negative binomial model was also employed to test the relationship between NCDs and built environmental factors.

Exploratory spatial scan statistic

Exploratory methods allow us to generate hypotheses about relationships (Link C, figure 1) by visually correlating significant spatial patterns of NCD-related hospital admissions with spatial patterns of environmental variables. We used the well-validated and robust Spatial Scan Statistic to investigate significant spatial patterns.48 ,50 ,51 This method asks ‘What area or what combination of areas is most likely to have a statistically significantly ‘high’ or a significantly ‘low’ risk relative to areas outside the combination of areas?’ This would be framed as a ‘cluster detection problem’ in the spatial epidemiology literature.48

The Spatial Scan Statistic was implemented using the SaTScan software. This method implements a single maximum likelihood-based hypothesis test over geographic space to identify the regions where the distribution of cases relative to controls/population (or the expected number of cases) is most likely to be consistent with a significant excess risk. To implement this, SaTScan identified candidate clusters, which were circles of increasing radii, bound by a maximum population threshold radius (set here to 5% of the population), centred on prespecified locations such as SA1 centroids. The size of the cluster is sometimes sensitive to the threshold radius.52 The 5% threshold represents around a few hundred expected cases of most NCDs, and is sensitive enough to delineate small clusters, an early goal in our data exploration and analysis.

Over many candidate clusters SaTScan maximises the likelihood ratio, given by

where LLR represents the logarithm of the likelihood ratio, O refers to observed cases, E to expected cases, and n is the total number of cases in the entire region (ACT). The likelihood formula assumes that NCD cases are distributed as a Poisson random variable and the likelihood ratio is compared with simulated likelihood ratios generated from 999 Monte Carlo randomisations of the data to assess statistical significance. The area that has the highest likelihood value (or the lowest p value) is the primary cluster. If both low-risk and high-risk clusters are searched for, then the most likely (high and low) clusters will be identified and published by the software. Secondary or less likely clusters may also be reported. In our analyses, we restricted our results to primary or secondary clusters with a significant p value (p<0.05). Relative risks (RRs) at the significant clusters were reported as: (risk inside the cluster)/(risk outside the cluster).

SaTScan analyses were implemented for CSDs and respiratory diseases at the SA1 scale for 2011 and CD scale for 2007. Owing to an unexplained anomalously low number of hospitalisations for ENMDs in 2011 (table 1), we scanned 2012 SA1 and 2007 CD ENMD data. Owing to lower event rates, MI and selected cancers were analysed at the SA2 scale for the entire aggregated 2007–2011 period. Thus, SA2 observed and expected numbers were summed for the entire 5-year period 2007–2011. Results were mapped using ArcGIS V.10.1.

Associations between built environment factors and hospital admission rates

We used two different models to investigate the relationships between the various NCD-related hospital events and built environment characteristics. The hospital admission data were complex, with multiple cross-classifications and nesting. For example, each person in the data could be hospitalised multiple times (nesting of hospitalisation episodes within people), people were nested in geographic neighbourhoods such as suburbs, and the temporal nature of the data implies likely temporal trends and seasonal patterns. In addition, the distributions of a number of predictors such as suburb-level Walk Score or GP density were not normal, which would render traditional linear models unusable, or require complex statistical transformations and/or models. To overcome this problem, we first modelled relationships using a robust method: Monte Carlo logistic regression.29 ,49 The approach was as follows:

1. Randomly sample 50% of the data;

2. Fit logistic regressions (or any other model to be tested) to estimate best explanatory model, store parameter estimates: intercept and slope values;

3. Repeat steps 1 and 2, N times (in our simulations, N=1000);

4. Calculate mean and 95% CIs for estimated model parameters from stored values in step 2.

We used logistic regressions as our explanatory model, with each hospitalisation event with a primary diagnosis of respiratory diseases as the control condition. The dependent variable was a hospitalisation event (1/0) with a primary diagnosis of each of the NCDs described in the Data section—cancers, CSDs, MI, ENMDs and comorbids being coded as 1. Separate models were run for each of the MIs, CSDs, specific neoplasms, ENMDs and comorbids. Respiratory diseases were chosen as the control condition, or coded as 0, because the drivers of respiratory diseases, with the exception of smoking, generally differ from the environmental drivers of the other three conditions. (While ideally we would have liked to use all hospitalisations as controls, these data were not available at the time of analysis.) When modelling neoplasms, since lung cancers have somewhat different environmental drivers than the remaining cancers, we ran the model with and without lung cancer. We also attempted to model hospitalisations with comorbid CSDs, specific neoplasms, ENMDs and respiratory disease conditions by coding hospitalisation with more than one condition as 1, and the rest as 0. The independent variables in these models included: sex, age, marital status, payment with private insurance (yes/no) of the person hospitalised. All these covariates, with private insurance as a marker of SES, are known to be associated with chronic conditions.53 In addition, ecological-level independent variables (described in the Data section) include the hospitalised person's access to GPs, neighbourhood walk score, IRSAD score, access to alcohol and logged traffic exposure.

We also report McFadden's pseudo R² for the Monte Carlo regression analyses. We understand that the use of pseudo R² is controversial,54 and publish these values for researchers who prefer to see them reported. These values were not used for model selection or for any other judgement on model quality.

Finally, for NCDs with significant environmental correlates in the Monte Carlo model, we also modelled the total number of hospitalisation events of a given condition in a given suburb as a function of counts of different predictors. The models can be written as:

where Y_j is the total count of a given condition in suburb j and x_jk is the count of the kth predictor in the jth suburb, for example, the total number of insured patient hospitalisations in a suburb or total number of female patient hospitalisations in a suburb. Y_j was considered to be negatively binomially distributed with mean μ_j and variance κ. A negative binomial model was used after it was found that the data were overdispersed, rendering a Poisson model unsuitable. The mean μ_j or suburb-level count of a given outcome was modelled as an exponential function of an intercept term β₀ and a slopes term β_k. These models require aggregate counts or summaries at the suburb level, and variables were recoded to satisfy this requirement. Thus, for example, discrete variables such as the marital status of a hospitalised person (1/0) translated to the total number of hospitalisations of married people in a given suburb. Continuous variables were similarly recoded, such as the number of hospitalisations of people in the topmost quartile of traffic exposure, number of hospitalisations of people in the lowest decile of IRSAD, number of hospitalisations of people with good GP access and so on. People with a GP density of one or more in their immediate buffer neighbourhood were considered to have good access.

We were interested in modelling counts of a hospitalisation outcome (eg, heart attack hospitalisations) in a small area as a function of counts of the characteristics of the hospitalised population in the negative binomial models. Note that the population size of a suburb does not necessarily predict the number of hospitalisations, which is a function of a number of neighbourhood compositional characteristics such as age, sex and SES. Counts of hospitalisations that capture these characteristics were included in the model. While modelling heart attacks as a fraction of all hospitalisations could be an alternative model, the results of the count negative binomial model, as described in the next section, converge with the results from the logistic Monte Carlo model, underscoring the strength of our analyses. The models were implemented using R and Stata.