This paper contributes to obesogenic foodscape research in two ways. Firstly, measures of density and proximity were compared to assess the extent to which they represent different facets of the foodscape. These measures are theoretically distinct and commonly differentiated in the literature, and it has been argued that to use just one metric is to do a disservice to the range of ways in which environments may affect behaviour . However, using both metrics may lead to multicollinearity issues in ensuing analyses, resulting from covariance between the two metrics. Further, opting for either of the two metrics may provide a ‘shortcut’ for researchers wishing to employ only a single, representative metric of foodscape exposure. Indeed, correlation analyses presented here provided tentative evidence that these measures represent the foodscape in very similar ways, suggesting that neighbourhoods with high food outlet density were also those where residents live within closer proximity to food, supporting previous findings [9, 10]. However, this study does not represent definitive evidence that this is always the case, and opposite results have been observed . Furthermore, although incidental, at the very least each measure may be susceptible to a different level of measurement error . For example, where food outlet location data is not totally comprehensive, measures of food outlet density may be more prone to error (systematic underestimation of exposure) than measures of proximity (which, by chance, may remain completely accurate).
Secondly, this paper evaluated the extent to which methodological heterogeneity in calculating an exposure metric (density or proximity) contributes to a lack of inter-study comparability in findings. To some extent, results here suggested that such criticisms may be unfounded: measures of proximity from population-weighted centroids were highly correlated when using Euclidean or street network distance, corroborating findings from the only one comparable previous US study . In addition, measures of ‘pseudo-individual’ density using Euclidean or street network distances were nearly equal at similar geographical scales, as has also been found elsewhere [51, 60], whilst becoming more similar at greater distances. Buffer distances beyond 1000m have been used in the literature [6, 11, 12, 25] and we could therefore expect even greater convergence. This said, there was evidence to suggest that Euclidean buffers were most comparable to marginally larger street network buffers: 400m Euclidean buffers most strongly correlated with 800m street network buffers and 800m Euclidean buffers most strongly correlated with 1000m street network buffers, confirming similar results from Thornton et al.. This is likely to be because in general, Euclidean buffers result in a larger footprint, thus encapsulating more food stores, as illustrated in Figure 1. Statistical investigations also reported density per km2 and per 1000 population to be similar, inviting comparability across this methodological binary.
Measures of density at the area or ‘pseudo-individual’ level were only moderately correlated with one another, suggesting a limited extent to which findings across this divide should be compared. Unfortunately, this divide across metrics at the area or ‘pseudo-individual’ level separates two large bodies of academic work (Table 1), where studies have either located individuals within administrative area neighbourhoods, or created bespoke neighbourhood buffers around said individuals. Therefore, it may be of benefit to the field if a single (area or ‘pseudo-individual’) density metric were to be used in future research to maximise inter-study comparability, wherever possible. At the very least, studies should consider providing a rationale for their preference of approach to density calculation, whilst better and more fully acknowledging the assumptions and limitations inherent to either choice.
Beyond implying comparability across studies then, results suggest that there may only be small gains made from using street network distances for measures of ‘pseudo-individual’ density or proximity. This may be important when time or resources are at a premium or where street network data is unavailable. However, we disagree with Sparks et al. who conclude that this equates to a reduced “computational burden on those wishing to use GIS methodology” – from which we infer that through not having to use more complex street network data, the usability of GIS methods is increased. It is argued here that there is a ‘tipping point’, where the value of using increasingly detailed metrics begins to diminish in relation to the computational effort required to create them. However, as we know little of this ‘tipping point’, and we should be mindful of scale differences (calculating street network availability may be more critical at 400m radii than 1000m radii, for example), we should always try to do the best that we can, even when confronted with technological challenges.
We argue here that we need to further consider how we can advance our methods and our metrics of exposure in objective studies of obesogenic food environments. Some studies have already sought to use measures of variety in relation to the food environment [6, 11, 32]; for example, the ratio of fast food to full service restaurants , designed to complement measures of density and proximity. Others have begun to use inverse distance weighted (IDW) measures of facility access, which to some extent ameliorate concerns arising over the relatively arbitrary definitions of ‘neighbourhood’ applied throughout the literature . In reality however, it may never be possible, or even appropriate, to reach a point where even a well-conceived measure of food access can become a universal standard, suitable for use across all studies, as has been tentatively suggested . Other methodological differences between studies – in statistical techniques, or in terms of study populations and their characteristics, which might vary between countries for example, and so on – ensure that two studies will rarely, if ever, be completely comparable to one another. This does not detract however from the importance of attempting to understand the implications of such diversity, which is what we have begun to address here. A valid and reliable research evidence base, where differences in study findings can be fully understood and appreciated with respect to the methods used, from which conclusions about neighbourhood level effects on diet can be accurately drawn, will be absolutely critical in justifying neighbourhood interventions or pilot interventions designed to promote health, such as restricting the clustering of unhealthy food retailers.
This paper has a number of limitations. In this study we did not compare all foodscape metrics employed in the field to date; instead, we focused on many of the most commonly used metrics in order to relate to as much of the field as possible. Also, we did not investigate the entire foodscape here; different relationships may have been found for ‘Food Consumed Out of the Home’ outlets, although work not presented suggests this is not the case. There is also little compelling logic to suggest that the relationships tested here might be systematically biased according to the type of food outlet selected for study. Furthermore, the category of ‘Food Bought Out of the Home’ includes outlet types such as ‘supermarkets’ and ‘convenience stores’, density and proximity of which are frequently assessed in the literature. We did not have access to data on the exact locations of participants in this study, hence measures of proximity and some measures of density were calculated from administrative area population-weighted centroids, where we assumed at least one individual to live. This is an approach adopted in the literature where participant location data has not been available [11, 20], however we acknowledge that this could constitute a type of ‘errors-in-variables’ bias and that exposure throughout any given administrative area will vary , and would be likely to decrease away from population-weighted centroids. Our approach of using population-weighted centroids was at least consistent between areas. External validity in findings cannot be assumed, and we cannot rule out that results here may be particular to the North East of England, despite the large and heterogeneous study area, and the similarities between the study area and many other regions of the UK, notably in terms of its diverse socio-economic profile, with which we know foodscape exposure varies [15, 35, 37]. Lastly, we acknowledge that a greater density or proximity of food does not necessarily equate to more utilisation of these facilities. Considerations such as transport preferences, motivation to walk, economic factors, neighbourhood perceptions and so on will all contribute broadly to ‘access’ beyond purely a geographic perspective. We know for example that in Newcastle upon Tyne, in the North East of England, adults conducting their main supermarket shop on foot travel a median distance one-way of 510 m, as compared to 2528 m for those with access to a car . It is also worth considering that the vast majority of food environment exposure studies have tended to focus exclusively on residential neighbourhood exposure, despite the apparent necessity of accounting for time spent in wider ‘activity spaces’ , too. This said, our study compared foodscape metrics that are widely used in the field, rendering this research highly relevant, regardless of whether these previous studies conceived of access in a purely spatial sense or otherwise.