In our sample, statistically significant differences in environmental exposures were found between GIS and GPS routes. This was particularly evident among pedestrians for whom GIS routes underestimated true route length, and overestimated exposure to busy roads, total food outlets and unhealthy food outlets. Our results suggest that while a GIS route may provide a reasonable proxy measure of route length, caution should be exercised in the assessment of environmental exposure.
GIS routes underestimated route length by an average of 21%. Underestimation was less severe for active travellers, but was still statistically significant. Living further from school, travelling by bus and living in rural locations were all associated with greater differences in length between GIS and GPS routes. GIS estimates of route length for children with these characteristics are therefore likely to be least reliable. These finding may have some impact on studies attempting to estimate physical activity accrued during travel to school. Although the mean difference of 97 m for those travelling on foot may represent only a small difference in potential physical activity, such differences may also be important for work attempting to identify distance thresholds for different modes, or for work such as that of Singleton (2014) attempting to estimate CO2 emissions from school commutes, where a 1 km difference in route length for car drivers may make a significant difference .
In terms of the specific environmental exposures we investigated, the general trend seemed to be that GIS routes overestimated exposure for active travellers, and underestimated for bus and car users. The impact of underestimation on environmental exposures in bus and car users is not necessarily clear, as their actual exposure will be dependent on their exiting their vehicles, and further research on this behaviour is required. In a finding similar to that of Duncan & Mummery , the study of GPS routes revealed a preference for quieter roads among walkers; the length of the route along a busy road was 17% lower on GPS routes compared to GIS routes. This trend was also apparent, although not statistically significant, among cyclists.
Given that walkers and cyclists potentially have greater opportunity to access the facilities they pass en route, accurate assessment of their exposure is important. Although the best fit model of percentage overlap indicated that certain characteristics of children and their environments (living closer to school, travelling by bike, living in an urban location, or attending a school in a more urban location) increased the likelihood that the GIS route more accurately represented that taken, the same factors were not associated with differences in the environmental exposure variable, food outlet exposure, as examined in regression models. Mean food outlet exposure ranged from 4–9 outlets on a route, according to travel mode, so it is possible that a relatively small deviation from the modelled route could result in a proportionally large difference in food outlet exposure, especially if outlets are clustered and a relatively large number may be passed in a short distance.
The only variable we found to be significantly associated with differences in food outlet exposure was whether the route was to or from school. Disparities in estimated exposure were greater by an average of 1.5 outlets for journeys home compared with those to school. It may be important to consider differing environmental exposures on routes to and from school in future work. Certainly, if GPS are being used to record routes, efforts should be made to include travel in both directions. It may be that during the period after school children have more time to deviate from a direct route, and therefore greater exposure to the school and route foodscapes can occur. Indeed, in this sample, mean food outlet exposure was 5.6 outlets across GPS routes to school, and 7.2 outlets across GPS routes from school.
While our results indicate that GIS modelled routes do not capture actual environmental exposures particularly well, the use of GPS data is also not without issue. Chaix et al.  argue that as GPS devices measure only where individuals have been, and not the environment they have the potential to use, the causality between environmental exposure and health behaviour is obscured. However, we believe that further use of GPS route measurement, coupled with GIS derived ‘potential environments’ and behavioural surveys and interviews may allow this issue to be unpicked, for example potentially examining how and why a child may deviate from the shortest route home to access food outlets, and thereby improving our understanding of how environments and behaviours interact.
In addition to this conceptual issue, the use of GPS data also raises questions about data representativeness. We modelled routes separately for each day and session (to or from school), giving up to 10 routes for each participant. Further research is needed to better understand how many routes may need to be recorded to assess habitual exposure. However, given the differences we found in variance partition when modelling percentage overlap (a general measure of path concordance) and food outlet differences (a specific environmental exposure measure), the number of routes required may vary according to the exposures being investigated.
This study has several strengths and weaknesses. In terms of strengths we included a large number of objectively measured GPS routes from participants living in a range of urban and rural settings. Participants travelled by different modes, and were recorded over multiple days. Secondary school-aged children such as those studied here are likely to travel independently to and from school , and therefore take routes of their own choosing.
While processing tools exist for the identification of trips within GPS data , it is not clear how successful the automated identification of routes to school may be, especially as they may be composed of multiple ‘trips’ if the individual has stopped along the way. To prevent potential errors as a result of trip identification automation, we manually identified routes between home and school from the GPS data, providing confidence in the routes derived. Additionally we were able to identify school entrances in an on-site audit improving the modelling of GIS routes.
However, limitations must also be acknowledged. Information on how each participant travelled to school on any given day were not generally available, so their self-reported usual mode of travel was used to determine GPS route mode and it is therefore likely that some routes were misclassified in terms of mode. Some data on actual route mode were available from the four-day food diary complete by SPEEDY participants, and which asked how the participant had travelled to and from school on two school days. In total 174 (99%) of the participants in these analyses completed the diary, and actual route mode was available for 464 of the 1191 GPS routes (39%). Of these 397 (86%) were made by the reported ‘usual’ mode of travel to school, as has been used in our analyses. This high agreement rate gives confidence to our findings, although the misclassification of route mode was not randomly distributed; of the 67 routes that were not made by the usual mode, 22 were journeys made on foot by children who reported usually travelling by car. This suggests that differences between car and walking routes may be underestimated in our models.
Only 8 of our participants reported usually travelling by bicycle. Although they provided 62 routes between them, numbers were still small, and so although differences between GIS and GPS routes for cyclists were detected, they were not statistically significant, possibly as a result of the small numbers.
To model routes in a GIS, defined start and end points are required, along with a network dataset. Home locations were derived from the address provided on the consent form (one address per participant), and we were therefore not able to account for instances where a child had more than one home. If a child had not travelled between school and the address on the consent form between the specified hours, the trip was not included in our analysis. This approach means that some legitimate routes to/from school may have been excluded.
The quality and completeness of the network used will impact the routes modelled. We were able to use a well-regarded, accurate road network for the modelling process, but this did not include footpaths or informal short-cuts. The overall median proportion of routes not on our road network was 0.3%, but was somewhat higher for pedestrians (4.8%). However, this may not give the complete picture of the impact the inclusion of footpaths may have on route modelling because the use of a small short-cut may only incur a small amount of travel ‘off-network’ but may enable a significantly different route to be taken, generating potentially large differences in environmental exposure.
The setting of the SPEEDY study within the county of Norfolk, UK may limit the transferability of our findings to other settings. Although we see no strong reason why the same factors would not impact GIS and GPS route differences in other similar settings (e.g. other rural counties in the UK or in other international settings), nor that some findings might have even wider transferability, care should be taken in assessing if and how the Norfolk situation may differ to other settings when attempting to apply these results in other contexts.
In conclusion, GIS modelled routes between home and school were not truly representative of accurate GPS measured exposure to obesogenic environments, particularly for pedestrians. While route length may be fairly well described, especially for urban populations, those living close to school, and those travelling by foot, the additional expense of acquiring GPS data, potentially coupled with behavioural surveys and interviews, seems important when assessing exposure to route environments.