In this Glasgow study, we found that the association between locational access to supermarkets and individual-level fruit and vegetable consumption was highly sensitive to the food environment measure that was selected. The findings suggest that, for a number of the access measures we created, greater access to supermarkets was associated with higher fruit and vegetable consumption. However, this was not the case for all measures; for example, these results suggest that no association is apparent when we assessed counts of supermarkets within a 5 km road network buffer around an individual’s unit postal code location. This finding draws attention to the potential risk of committing a Type-II error.
Whilst measuring from a unit postal code provided a less aggregated and, arguably, more precise estimate for proximity measures than the measures from the geometric or population-weighted centroids of the larger spatial units (data zones), the proximity estimates did not vary greatly across the whole sample by point-of-origin and were not associated with our three outcomes. Hewko et al. have previously noted that aggregation error is a greater concern when examining more densely populated features because proximity estimates are more sensitive in this instance . Whilst supermarkets were reasonably densely populated in our study area, the analysis may have been more likely to detect an association if other more prominent food store types were examined (e.g. takeaway outlets) or a larger spatial unit was used rather than a data zone. However, given their size, using specific household addresses over postal codes would have only made minimal differences to these findings.
When the presence or absence of a feature is explored, the count data is dichotomised to examine whether any store is present within a set distance (e.g. is there a supermarket within 1 km). Again no significant association was observed using this approach amongst the full sample. However, when count data were investigated, some significant positive associations were detected, indicating a greater choice may be more important than access to a single store. Previously, greater choice in the form of different fast food chains has been linked to more frequent fast food use  suggesting that dietary behaviours are influenced by having access to a wider selection of options.
The strongest associations observed were when Kernel density estimates were used for the exposure measure. Chaix et al. previously posited that more often the use of a boundary (in this instance a buffer) implies a binary definition of access and that the use of a smooth transition between what is and is not accessible would more often reflect a truer representation of access . The adoption of Kernel density estimation enabled the application of a smoothing process by weighting areas more heavily when they were proximal to other stores. This weighting diminishes when the number of stores nearby is reduced and/or the distance to other stores is increased. Once these estimates are created, individuals are plotted to this map and assigned a density estimate based on their location. Whilst recent examples of studies using Kernel methods in food environment research exist [16–20], it remains a relatively underutilised technique compared to standard proximity or buffer approaches.
Another interesting finding to emerge from our analysis was the similarity between 1 km, 2 km, and 3 km Euclidean buffers with 2 km, 3 km, and 4 km road network buffers, respectively. This suggests such measures may be comparable. Sparks et al. previously reported similar associations between Euclidian and road network buffers and concluded that disparate measures of food access can often be compared . Consequently, they suggested that aggregated and Euclidean distance measurements offer the same outcomes as more sophisticated and potentially more resource intensive approaches (i.e. less aggregated data and road network measurements). However, contrasting findings are reported in a study using both Euclidean and network buffers to explore the role of land use on walking behaviours, with stronger associations found for network buffers , highlighting the need for clear conceptualisation of exposure measures prior to analysis.
The choice of distance for buffers varies considerably across studies assessing associations with the built environment [7, 21, 22]. However, the use of buffer distances that are too small can result in the lack of an adequate exposure gradient meaning the detection of an effect is unlikely . Further, using distances that are too large often overestimate the exposure by capturing features that individuals are unlikely to interact with and again may reduce the heterogeneity of the exposure measure. Our study demonstrates that the scale of the exposure measure can have a considerable bearing on the interpretation of the existence or otherwise of a relationship. Inconsistencies in scales are mainly driven by a lack of data that can be used to inform researchers as to what distance should be explored. It has previously been argued that understanding “true” environmental differences requires the identification of “true” environments . In this instance, defining a “true” environment would require us to know where people are being exposed to and buying food. Therefore, for an accurate assessment of the role of environmental influences on dietary behaviours, and for an improved conceptualisation of appropriate scales, it is essential that studies move from place-based to people-based measures of exposure .
It is also important that researchers begin to account for the wide variety of ways that individuals interact with their environment. Cummins noted that what constituted local and appropriate food access differs between individuals [48, 49]. In this instance, it may be because some individuals will travel further to food stores that meet their needs (e.g. product variety and quality, specific ethnic stores, cheaper prices) [48, 50]. More generally, socioeconomic factors are also likely to strongly influence an individual’s mobility and thus their ability to access particular food stores. For example, a low income can restrict motor vehicle ownership, potentially reducing an individual’s access to a wider variety of food stores [51, 52]. To date, most investigations on local residential food environments and dietary behaviours are limited by the assumption that all stores within the local area are equally accessible to all residents irrespective of potential mobility barriers. Chaix et al. has called on future studies to consider the use of an individual-specific rather than uniform definition of neighbourhood scale . Such analysis allows different scales to be applied based on individual characteristics. In our latter analysis, the exposure measure is further strengthened by considering an indicator of individual mobility through their vehicle ownership and use. When supermarkets within a walkable distance (0.4 km) were explored, an association was found with fruit consumption amongst those without a vehicle whereas when this was investigated amongst the full sample no relationship was detected. This suggests that considering personal mobility factors can strengthen our understanding of the links between the environment and health behaviours. Bader et al. previously explored the concept of “travel burden” whereby factors such as vehicle ownership, crime and public transit access are assessed to determine how these factors influence spatial access to healthy food. Conducted in New York City, their study found that adjustment for vehicle ownership and crime tended to increase the observed disparities between neighbourhood race and income and supermarket access. Whilst their study did not examine links to health behaviours, a prior US study found stronger associations between local healthy food resources and insulin resistance amongst those who did not own an automobile .
This study was strengthened by the comprehensive assessment of multiple access measures, and more importantly, how these influence associations with dietary outcomes. Whilst some prior studies have compared how different access measure effect exposure estimates, they have not explored a range of measures and buffer distances as comprehensive as that undertaken in this present study nor have they assessed the impact on dietary outcomes. It is important that the use of outcome data is acknowledged as it provides some indicative data on comparability with other food environment studies that have used varying measures of access.
The limitations of this study must be acknowledged. First, with regards to the food environment only a single source of fruits and vegetables (supermarkets) was examined. Further, this study did not have within-store data on these supermarkets to help inform the quantity, quality and price of the fruits and vegetables sold which are potentially important factors in determining purchasing and consumption behaviours . Additional factors at the area-level that may affect mobility (e.g. public transit option, crime, safety) were not considered nor were individual data related to perceptions of these factors. Our buffer estimates were all created based on distance metrics whereas additional data on speed limits may have allowed a more sophisticated approach that also included estimates of travel time. This is important to consider as it may be that time is more important than distance when considering how people interact with food stores or indeed there may be other factors such as individual’s preference for a particular area or store type that dictates where they shop. The cross-sectional nature of the data and the time lag between the individual survey data and the supermarket data may limit the applicability of the reported associations. However, first we reiterate that our primary aim was to demonstrate variations in associations based on different exposure measures rather than to establish causality between supermarket access and dietary outcomes. Second, with regards to the time lag, prior research demonstrates little change in the number of national ‘multiple-owned’ supermarkets in Glasgow between the years 1997 (n = 75) and 2007 (n = 78) and therefore the static nature of chain-brand supermarket locations would likely mean this is unlikely to significantly influence the results. Finally, the data analysed is restricted to a single urban area in one country and findings may firstly, not be applicable in rural areas and, secondly, would require confirmation in other urban contexts elsewhere.