A Time-Based Objective Measure of Exposure to the Food Environment

Thank   you for such kind words!

R
  1.2

The   introduction is a little pedestrian, but covers the main areas and it is   clear the importance/purpose of the paper. Methods used appear appropriate   and the data/analyses are novel and help answer the research gaps identified.   The topic is interesting to the journal readership and I can see become a   useful reference for future study.

Reviewer   comments on earlier drafts of this paper have indicated that this subject   matter is of interest to a wide range of disciplines. Finding the right   wording and language so that our manuscript is of interest and relevant to   that range has proven tricky. The end result is that our writing has indeed   become pedestrian, but we hope that it makes our work more accessible.

R
  1.3

I have one   issue. Four buffers are used - 21m, 100m, 500m and 1/2mile. The latter two do   not seem to fit with the narrative of the paper. If one of your key arguments   is that we need better data to understand how individuals are exposed in   their environments, why then select two options with large buffer sizes that   undermines the fact that you have their GPS location and therefore know their   exact exposures. There might a good reason for this (I doubt it tbh), however   it is never really made. While this might be overlooked, it is talked a lot   about in the discussion:
 
  "First, it appears that   exposure duration may reach a level of saturation where longer contact does   not lead to a different behavior." (lines 292-293)

In   the present manuscript we note that the larger buffers come from earlier GPS   food environment studies. Clarifying language was added to Section 2.4:   Exposure Measures in the manuscript and in Section 4: Discussion.   We do agree with the reviewer’s discomfort! The only reason for including the   two larger buffers is that they have been used in many papers over the course   of around 10 years of research on exposure to the BE. We think that the   previous use of large proximities came from confusing access to certain   features of the BE with exposure to the BE. Clearly people need to have   access to places that support healthy behaviors (e.g. fresh and affordable   fruit and vegetables), but also their behavior can be influenced negatively   or positively by what is near them. Two researchers have made a clear   difference between access and exposure, Basile Chaix and Mei-Po Kwan. Chaix   measures exposure based on travel patterns and trips characteristics. Kwan   measures exposure over the entire life course. For example, see

Chaix, B., 2018. Mobile Sensing in   Environmental Health and Neighborhood Research. Annual Review of Public   Health, 39, pp.367–384.

Perchoux, C. et al., 2014. Assessing   patterns of spatial behavior in health studies: Their socio-demographic   determinants and associations with transportation modes (the RECORD Cohort   Study). Social science & medicine (1982), 119, pp.64–73. Available at: http://www.ncbi.nlm.nih.gov/pubmed/25150652.

Thierry, B., Chaix, B. & Kestens, Y.,   2013. Detecting activity locations from raw GPS data: a novel kernel-based   algorithm. International journal of health geographics, 12, p.14. Available   at: http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=3637118&tool=pmcentrez&rendertype=abstract.

Kwan,   M.-P., 2002. Feminist Visualization: Re-envisioning GIS as a Method in   Feminist Geographic Research doi:10.1111/1467-8306.00309. Annals of the   Association of American Geographers, 92(4), pp.645–661. Available at: http://www.blackwell-synergy.com/doi/abs/10.1111/1467-8306.00309.

Kwan,   M.-P. & Lee, J., 2003. Geovisualization of Human Activity Patterns Using   3D GIS: A Time-Geographic Approach. In M. F. Goodchild & D. G. Janelle,   eds. Oxford: Oxford University Press, pp. 48–66.

R
  1.3

You do not   actually measure or test such saturation and your changing buffer sizes do   not really capture it as you suggest. The larger buffer sizes really do not   capture any exposure since an individual will not be aware of what is in   these larger zones (the justification of immediate exposure is lost here). It   also ends up reducing your spatial variation as the larger zones will produce   more similar results between individuals, hence maybe why results disappear.

We   modified Section 4: Discussion to take into account your comments.

R
  1.4

A weak   justification is given for the larger buffers:
  "The 500 m and 1/2 mile   proximities covered areas that participants could not immediately and   entirely experience, but might be aware of, based on cognitive processes and   memories. Exposure at these farther proximities did not appear to have an   effect on behavior." (lines 319-321)

No evidence   is given to support this assertion and I do not really agree with it at all.   The discussion spends too much time talking about the different buffer zones   that it becomes distracting and ignores the main point of their paper. I   would refocus the text back to the novelty and importance of the paper set up   at the start. That will make for a tighter paper and one that fairly reflects   on its results. You have a good paper, don't lose sight!

We   agree that the paper would be much tighter if we left out the section on   buffer sizes, however we think this links the paper to previous research and   provides some much needed speculation on the processes that link proximity to   exposure to behavior to health. If the reviewers disagree, we are more than   happy to cut this section from the paper.

As   for the quote, we have deleted it. 

R
  1.5

Why keep   age as a binary? And why split by 45 years? This is not clear and seems an   odd decision.

The dichotomization of the age variable   was due to a sample that skewed older as a result of our sampling goal   targeting households’ primary food shoppers.

R
  1.6

'Airline   distance' is cited but I would refer to this as euclidean distance. Seems odd   that you would calculate that and not actual road network distance given all   of the good spatial data work you do elsewhere - seems lazy.

We’ve changed “airline” to “Euclidean.” We   choose Euclidean buffers over network buffers so that we could link our work   to the previous research on the topic. Our research is in part an elaboration   on Burgoine & Monsivais (2013) and Burgoine et al. (2014). Both studies   used 100- and 500-m Euclidean buffers as their measure of exposure. Further,   Euclidean buffers allow for exposure during travel that is not occurring   along a street network, such as cutting through parking lots, or green parks.  

Burgoine, T., & Monsivais, P. (2013).   Characterising food environment exposure at home, at work, and along commuting   journeys using data on adults in the UK. The International Journal of   Behavioral Nutrition and Physical Activity, 10(1), 85. https://doi.org/10.1186/1479-5868-10-85

Burgoine, T., Forouhi, N. G., Griffin, S.   J., Wareham, N. J., & Monsivais, P. (2014). Associations between exposure   to takeaway food outlets, takeaway food consumption, and body weight in   Cambridgeshire, UK: population based, cross sectional study. BMJ (Clinical   Research Ed.), 348, g1464. https://doi.org/10.1136/bmj.g1464

R
  1.7

Line   segments were truncated at the boundary of Kings County - does this introduce   edge effects into your analysis?

In   our study edge effects are concerning because they may result in   underestimation of FFR exposure. Short of reviewing each of the line segments   comprising the GPS we cannot discount edge effects, though we did take steps   to minimize them. Namely, we removed participants from the analysis if their   workplace was outside the county. We also reviewed all of the instances in   which the GPS did not confirm a visit reported in reported in the travel   diary. Edge effects were assumed to be at play in these situations, however   that was not the case.

R
  1.8

Table 1   includes '# of cars' rather than number of cars which is a little lazy

Change   made

R
  1.9

"Table   46" (line 257) - I think this needs a dash in it

Per   comment R2.12, we've consolidated Tables 4 through 6 into one table.

R
  1.10

Tables 4-6   are reported as ‘odds’ but in text as odds ratios - I suggest you be clear   and consistent throughout

Changes   made

R
  1.11

All in all   a good paper, well done

Again   we greatly appreciate your kind words.

Thank   you for the very helpful review.

Abstract

R
  2.2

Line 19- include the age of   the participants.

Change made

R
  2.3

Line 30- My major concern is that   people who visit a fast food restaurant would have substantially higher   amounts of time spent in proximity to FFRs, thus the exposure and outcome are   the same?

We briefly mention this in Line 25, noting   that we removed self-reported exposures from these measures. In lines 73-77   we use Chaix’s concept, selective mobility bias, to frame this issue. We   revisit this issue again in the seventh paragraph of Section 4: Discussion   (lines 349-356). We subtracted from our exposure measures time spent in the buffers   when we were able to link that time to FFR trips reported in the Travel   Diary. However, this is a problem for FFR visits that were unreported in the   Travel Diary. See

Chaix, B. et al., 2013. GPS tracking in   neighborhood and health studies: A step forward for environmental exposure   assessment, a step backward for causal inference? Health & place, 21C,   pp.46–51. Available at: http://www.ncbi.nlm.nih.gov/pubmed/23425661

Introduction

R
  2.4

I think a larger discussion of   advantages and limitations of travel logs is warranted in the introduction as   your paper heavily relies on these. How many FFR visits do you feel were   missed by relying on travel logs? Other studies have used individual-level   GPS data to inform time-based measures of participants visits to destinations   (Brusilovskiy et al., 2016)  and children’s visits to food retails   outlets (Chambers et al 2017 also used wearable cameras) using the ST DBSCAN   algorithm (references provided at the bottom).

Thanks   for drawing our attention to the Brusilovskiy et al. and Chambers et al.   papers. Both of these were mainly focused on automatically identifying   visited destinations from passively collected data, whereas our main aim was   to estimate time-weighted exposure to specific locations.

Unfortunately,   we have no way of knowing how many FFR visits were missed by relying on the   travel logs. However, we were able to do the opposite and measure how many of   the 273 visits to FFRs reported in the travel log had corresponding GPS data   suggesting a visit. Applying a tolerance of +/- 10 minutes to visit times   reported in the log resulted in 211 instances (79.5% of reported visits) in   which a set of GPS points were inside the boundaries of a tax assessor parcel   containing an FFR with a brand name that matching that of the FFR reported in   the log. These findings were reported in

Scully, J. Y., Vernez Moudon, A., Hurvitz,   P. M., Aggarwal, A., Drewnowski, A., & Jacobs, D. (2017). GPS or travel   diary: Comparing spatial and temporal characteristics of visits to fast food   restaurants and supermarkets. PLOS ONE, 12(4), e0174859.   https://doi.org/10.1371/journal.pone.0174859

It would be interesting as a separate   study to “validate” FFR visits (and/or quantify visits that were missing from   the travel diary using a method such as ST_DBSCAN.

Methods:

R
  2.5

Lines 94 & 134 – What was   the data completeness over those seven days? How many Erroneous GPS points   were removed, how much of a problem was this?
  Given data loss is more likely when   participants are free-moving, does the data fairly reflect participants   normal activities?

Our   threshold of data completeness was 3 observation days. We removed 16   participants from the sample who did not meet this threshold due to   improperly filled out travel logs or obvious errors in the GPS such as the   total number of GPS points in the data totaling less than 100. We removed   less than 2% of the GPS points from the sample using the PALMS, and Tsui and   Shalaby data cleaning procedures. At the line segment level we only removed   0.14% of the line segments with durations longer than 30 seconds.

R
  2.6

Line 95 – Was there any effect   of seasonality in participants’ mobility patterns or FFR visits? Also, it   says 2102 when it is meant to be 2012.

2102   was changed to 2012—our apologies for this oversight. We did not check for   seasonality in this study. Unfortunately, with a sample size of 412 we needed   to limit our comparisons.

R
  2.7

Line 146-149 – I understand   the reason for excluding time around home and work due to participants not   turning off their devices but why was 125m decided upon? Was this based on   some formative analyses? Average lot size?

We   followed the precedent set by Hurvitz et al. 2014. Section 2.4 Exposure   Measures modified to reflect this. Due to GPS drift, any given point has some   error; although GPS manufacturers typically state 3-5 m, we found that   >90% of GPS points were within 125 m of the parcel centroid (roughly a   1-block face radius in pre-WWII US neighborhoods) for a data set collected   from a GPS unit at a stationary location within a typical Seattle area   wood-frame house. Any exclusionary buffer will unfortunately include both   false positives and false negatives; we thought that a larger buffer would be   erring on the side of caution.

See

Hurvitz, P. M., Moudon, A. V., Kang, B.,   Fesinmeyer, M. D., & Saelens, B. E. (2014). How far from home? The   locations of physical activity in an urban U.S. setting. Preventive Medicine,   69, 181–186. https://doi.org/10.1016/j.ypmed.2014.08.034

R
  2.8

Line 171 – were individuals   observation time used to offset the duration estimate?

We   did not apply a statistical offset in the logistic regression, but rather   prepared the data so that the FFR exposure values were normalized to the   person-day, and the dependent variable was binary (zero versus one or more   FFR visits over the study period). Our understanding is that if we used raw   counts of FFR visits or minutes of exposure, we would need to use the offset,   which was not the case here.

R
  2.9

It would also be good to get   an idea of how the travel logs worked. Did participants record the trip   before they left or after the trip was concluded? Did they report every   destination they visited or the primary reason?

A   sentence responding to these questions was added to the first paragraph of   Section 2: Material and Methods.

R
  2.10

Line 219 – A flow chart of   participant exclusion with reasons may help bring the methods together – 712   potential participants 516 enrolled 18 excluded for missing demographic   responses, 477 had concurrent GPS and travel log data. Sixteen removed for   incomplete data etc.

We   have included a figure with the exclusion criteria (Figure 1)

R
  2.11

Line 227 – perhaps a point for   discussion that 30% of the visits could not be verified, given the existing   limitations of self-report data, I think this needs to be discussed in more   detail in the discussion with options for alternatives to self-reported   measures of outcomes (visits). Line 231 – do you know why the visits were not   recorded by GPS? Was the data erroneous? Or the participant forgot to use the   GPS?

We   discuss this issue in last two paragraphs of Section 4: Discussion. We are   not sure how much more can be said on the topic. Unverified visits are black   boxes. They could have been the result of self-report error or a failure to   carry the GPS device. We have no way of assessing the reasons a self-reported   visit would go unverified by GPS data. At this time, we think that research   must use self-report and GPS in tandem. GPS could be used to imply that a   visit had taken place simply because they measure XY locations, but   self-report provides descriptive information about the location and the uses   at that site. GPS points in proximity to an FFR may indicate a visit or a   driver who has pulled over to take a phone call. Together, the two sources of   locational data can be used to create a more complete picture.

Results

R
  2.12

Could you combine tables 4-6   into a table like table 3 to make it a bit more readable.

Thank   you, the consolidated table makes much more sense.

R
  2.13

Line 257- tables 46 is meant   to be tables 4 to 6.

Fixed   by consolidating Tables 4 through 6 into one table.

R
  2.14

Line 262 In addition to   comment about line 30, is it possible that the duration is more effective   because it is more likely a reflection of participants walking (thus higher   accessibility), while counts could reflect a large number of exposures by   participants driving by, thus the results are bias by mode of transport?

Line 285 – Again I think   duration may be associated with exposure because of points on line 30 and   line 262. It would be good to discuss this in more detail.

A   paragraph was added to Section 4: Discussion to address your concerns.

Discussion:

R
  2.15

Line 318-322 – Perhaps be more   cautious and avoid using terms like to ‘affect their behaviour’ as you are   only testing associations between your exposure measure and behaviour which   are not casual and have methodological limitations.

Noted   and changed.

R
  2.16

Line 331 – again I feel   comparing counts to duration may be an unfair comparison, given duration is   likely reflect difference transport modes, leisure activity (shopping v   commuting) and the outcome (visits to FFRs). I think this needs to be   discussed further in order to justify that duration is a superior/alternative   to counts.

We   hope that we managed to follow this advice in the new paragraph added to Section   4: Discussion.

R
  2.17

Line 335 – This is were other   measures could be used to validate unrecorded visits to FFRs such as ST   DBSCAN, wearable cameras, amenities data or some combination of more   objective measures.

References I referred to   earlier are below:
 
  Brusilovskiy, E., Klein, L.A.,   Salzer, M.S., 2016. Using global positioning systems to study health-related mobility   and participation. Soc. Sci. Med. 161, 134e142. https://doi.org/10.1016/j.socscimed.2016.06.001.
 
  Chambers, T., Pearson, A.L., Kawachi,   I., Rzotkiewicz, Z., Stanley, J., Smith, M., Signal, L., 2017. Kids in space: measuring   children's residential neighborhoods and other destinations using activity space GPS and wearable camera   data. Soc. Sci. Med. 193, 41–50. https://doi.org/10.1016/j.socscimed.2016.06.001

These suggestions are   valuable. Future studies would benefit by focusing on the validity of travel   diaries and FFR utilization to include usage of wearable cameras, or using   ST_DBSCAN as a way of estimating FFR visits that were not recorded in the   diary. We have added three references supporting this (Chambers et al. 2017;   Brusilovskiy et al. 2016; Carlson, J.A. et al., 2014).