In this study, there were no associations between the proportion of parkland and psychological distress where it was perceived that the neighbourhood was safe. However, there were significant statistically significant associations between the proportion of parkland and high or very high psychological distress when the proportion of parkland was ≥20% and if it was perceived that the neighbourhood was unsafe. As with many other studies, we controlled for important potential confounders such as age, gender, income, level of education and population density which, in our data, were associated with high or very high psychological distress. In addition, we also investigated interactions between proportion of parkland and perceptions of neighbourhood safety and area deprivation.
Although many studies have suggested that exposure to greenspace or parkland is associated with better health and well-being [16, 17, 19], we were not able to demonstrate that exposure to a higher proportion of parkland was associated with less psychological distress. However, we did find evidence indicating that perceptions of neighbourhood safety and area deprivation were statistically significant effect modifiers of the association between parkland and psychological distress.
No studies, as far as we are aware, have reported on effect modification of the association between parkland and mental health by area deprivation. Our findings show that increased psychological distress is associated with high proportion of parkland in both the most and the least socio-economically disadvantaged areas but only if the neighbourhood is perceived as unsafe. If the neighbourhood was perceived as safe, then there were only weak non-statistically significant associations between parkland and psychological distress regardless of area deprivation. In the United Kingdom, Mitchell and Popham  reported poorer self-rated health with increasing percentage of greenspace in suburban low income areas but not in urban and rural low income areas and suggest that this may be due to a larger proportion of poorer quality greenspace in low income suburban areas. Residents of deprived neighbourhoods also have poorer perceptions of access to greenspace (despite shorter mean distances to greenspace compared to less deprived neighbourhoods) [36, 41] and safety of greenspace . More disadvantaged areas are also more likely to have poorer lighting (Crawford et al.) which may influence safety from crime .
The importance of neighbourhood safety is also reflected in several papers where it is proposed that safe environments stimulate positive behaviours (e.g. amount of physical activity) and leads to reduction in stress [33, 42, 43]. Also, Agyemang et al. suggested feeling unsafe and dissatisfaction with greenspace was associated with poor self-rated health and discourages people engaging in outdoor activities . Parkland associated with poor neighbourhood safety is consistently negatively associated with physical activity [41, 45–47]. Jones et al.  reported statistically significant decreasing trends in levels of adequate physical activity with increasing perceptions of lack of safety of greenspace. Parklands may be a space for criminal activities and antisocial behaviours, and would explain why people may feel unsafe in neighbourhoods with large proportions of parkland . This in turn may impact on recreational physical activity levels which in turn may adversely impact on psychological health and well-being.
In our study, respondents of the middle disadvantaged group, compared to the most disadvantaged and least disadvantaged areas, were less likely to report increased psychological distress with increasing percentages of parkland regardless of perceptions of the safety of the neighbourhood. It is not clear to us why this is so. This may be due to anomalies in the spatial level of our parkland data or to some important unmeasured confounders. However, it is not unusual to observe similar non-linear relationships in the literature. For example, a non-linear relationship between socio-economic status and psychological distress has been previously reported  and Poulos at al.  found a non-linear relationship between socio-economic status and childhood injuries.
The strengths of our study were that our subjects were part of an ongoing population based health status and risk behaviour survey, we had a large sample size and we took into account the multi-level nature of the data. The relatively high response rates (63.6% in 2007, 63.4% in 2008 and 58.7% in 2009) and the representativeness of the NSW Population Health Survey weighted sample ensures that our results are generalisable to metropolitan Sydney and other similar major Australian cities .
There are a number of limitations to our study. We obtained information on parkland from the ABS as land use data at the mesh block level. However, the electronic data did not allow us to analyse the data by number of discrete parks as a number of parcels of parkland could have contributed to a defined park. Further, parkland included state forests and national parks. We could not categorise parkland into more usable categories, for example, sports fields, bushland, presence of picnic facilities, etc., nor could we assess the quality of the parkland. We were also not able to calculate alternate metrics for access to parklands such as travel distance or travel time to parks as we did not have access to addresses for respondents in the NSW Health Surveys (addresses are not collected as part of the surveys). We were only able to obtain health survey data at the postcode level and hence we were restricted to using postcode as the unit of analysis. We would have preferred individual level data or data at smaller spatial units for analyses so as to have more accurate measures of exposure to parkland and to minimise exposure misclassification which is likely to be non-differential. This limitation of our study may decrease the precision of the estimates but not bias the results. Further, although the NSW Health Surveys were designed to provide estimates for spatial units larger than those used in this study, we do not expect this will affect our findings. Other limitations were that by using data from an existing cross-sectional survey we are not able to make a causal link between parkland and the reported health outcomes and that although we adjusted for a number of important potential confounders, there may yet be some residual confounding. However, we share this limitation with most other published studies on neighbourhoods and health.