ArticlePDF Available

Individual Mobility and Uncertain Geographic Context: Real-time Versus Neighborhood Approximated Exposure to Retail Tobacco Outlets Across the US

Authors:

Abstract and Figures

There is growing interest in the way exposure to neighborhood risk and protective factors affects the health of residents. Although multiple approaches have been reported, empirical methods for contrasting the spatial uncertainty of exposure estimates are not well established. The objective of this paper was to contrast real-time versus neighborhood approximated exposure to the landscape of tobacco outlets across the contiguous US. A nationwide density surface of tobacco retail outlet locations was generated using kernel density estimation (KDE). This surface was linked to participants’ (Np = 363) inferred residential location, as well as to their real-time geographic locations, recorded every 10 min over 180 days. Real-time exposure was estimated as the hourly product of radius of gyration and average tobacco outlet density (Nhour = 304, 164 h). Ordinal logit modeling was used to assess the distribution of real-time exposure estimates as a function of each participant’s residential exposure. Overall, 61.3% of real-time, hourly exposures were of relatively low intensity, and after controlling for temporal and seasonal variation, 72.8% of the variance among these low-level exposures was accounted for by residence in one of the two lowest residential exposure quintiles. Most moderate to high intensity exposures (38.7% of all real-time, hourly exposures) were no more likely to have been contributed by subjects from any single residential exposure cluster than another. Altogether, 55.2% of the variance in real-time exposures was not explained by participants’ residential exposure cluster. Calculating hourly exposure estimates made it possible to directly contrast real-time observations with static residential exposure estimates. Results document the substantial degree that real-time exposures can be misclassified by residential approximations, especially in residential areas characterized by moderate to high retail density levels.
This content is subject to copyright. Terms and conditions apply.
RESEARCH ARTICLE
Individual Mobility and Uncertain Geographic Context:
Real-time Versus Neighborhood Approximated Exposure
to Retail Tobacco Outlets Across the US
Thomas R. Kirchner, et al. [full author details at the end of the article]
Received: 31 August 2017 /Revised: 7 September 2018 / Accepted: 13 September 2018
#Springer Nature Switzerland AG 2018
Abstract
There is growing interest in the way exposure to neighborhood risk and protective
factors affects the health of residents. Although multiple approaches have been report-
ed, empirical methods for contrasting the spatial uncertainty of exposure estimates are
not well established. The objective of this paper was to contrast real-time versus
neighborhood approximated exposure to the landscape of tobacco outlets across the
contiguous US. A nationwide density surface of tobacco retail outlet locations was
generated using kernel density estimation (KDE). This surface was linked to partici-
pants(Np= 363) inferred residential location, as well as to their real-time geographic
locations, recorded every 10 min over 180 days. Real-time exposure was estimated as
the hourly product of radius of gyration and average tobacco outlet density (Nhour =304,
164 h). Ordinal logit modeling was used to assess the distribution of real-time exposure
estimates as a function of each participants residential exposure. Overall, 61.3% of
real-time, hourly exposures were of relatively low intensity, and after controlling for
temporal and seasonal variation, 72.8% of the variance among these low-level expo-
sures was accounted for by residence in one of the two lowest residential exposure
quintiles. Most moderate to high intensity exposures (38.7% of all real-time, hourly
exposures) were no more likely to have been contributed by subjects from any single
residential exposure cluster than another. Altogether, 55.2% of the variance in real-time
exposures was not explained by participantsresidential exposure cluster. Calculating
hourly exposure estimates made it possible to directly contrast real-time observations
with static residential exposure estimates. Results document the substantial degree that
real-time exposures can be misclassified by residential approximations, especially in
residential areas characterized by moderate to high retail density levels.
Keywords Spatial uncertainty.Exposure science .Human mobility.Urban computing .
Retail environment
1 Introduction
Geographic location provides a relational connection between individuals, their health
behaviors, and a rapidly expanding array of data on neighborhood structure and
Journal of Healthcare Informatics Research
https://doi.org/10.1007/s41666-018-0035-8
composition [13]. Once hierarchically integrated, it becomes possible to systemati-
cally investigate the impact of a wide range of neighborhood socio-ecological covar-
iates on indicators of individual health and well-being [47]. The geographic landscape
of retail tobacco products provides an excellent example [819]. A considerable body
of empirical evidence demonstrates the ways point-of-sale tobacco marketing influence
tobacco usersproduct preferences as well as decisions to initiate or refrain from use [8,
9,11,2026], and that targeted point-of-sale tobacco (POST) marketing tactics sustain
tobacco-related health disparities [12,2732]. Studies have also focused on the lowest
displayed pack price, promotions, placement for leading brands, and flavor descriptors.
More recent work has found that cigarette packs cost less at retailers located near public
schools than those near private schools [11], and that outlet proximity to schools and
parks is linked to advertising practices and illicit youth sales [12].
Despite a proliferation of studies that seek to identify and study mechanisms linking
neighborhood environments to the health-related decision-making and behavior of
residents, [2,3336] a number of unresolved conceptual issues persist, not least the
definition of the ecological units of analysis themselves. Because there are no naturally
occurring units of neighborhood, the neighborhood areal units employed by researchers
are often arbitrary or linked to confounding geographic circumstances, [37]
undermining their utility for geographically-explicit health research. [2] The ideal
neighborhood unit would represent the true causally relevant spatial context,[1]
however, so long as the true spatial context remains undefined, we are left with what
Kwan has termed the uncertain geographic context problem(UGCoP). [38]Inferen-
tial errors brought about by the use of arbitrary methods for delineating neighborhoods
likely contribute to the misclassification of effects and inconsistent evidence for socio-
ecological relationships and health behavior across the research literature. [3841]
Residential location remains the most readily available geographic linkage for
integrating individual and neighborhood data. Conventionally, researchers join residen-
tial addresses directly to the administrative zones they fall within (e.g., US Census
polygons) or draw new ego-centricbuffer zones around those addresses [4]. These
residential areal units act as spatially and temporally invariant containers for various
neighborhood risk and protective factors, each of which can then be linked to residents
themselves. Data on individual mobility patterns improve upon the assumptions that
administrative zones or ego-centric buffers make about the propensity for individuals to
come into contact with environmental risk factors. Mobility data capture actual travel
patterns, including each persons ever-accumulating set of origins and destinations,
allowing for precise estimation of exposure to environmental risk factors over time.
Dynamic conceptualizations of neighborhood exposures may enrich our understanding
of health and place by better reflecting the true causally relevant spatial context of
exposures to environmental risk in the population [6,7,4244].
This paper presents a direct comparison of real time versus neighborhood approx-
imated (residential address; [4]) exposure to a well-known environmental risk factor,
the spatial distribution of tobacco outlets across the contiguous United States [19,
4547]. This comparison provides an empirical framework for contrasting relative
uncertainty in spatial exposures and estimating the likelihood of their misclassification.
Residential and real-time exposures were computed for a sample of 363 people, each of
whom voluntarily recorded their location every 10 min for 180 days using their cellular
phone. Analyses use a common density-based exposure metric to examine the extent to
Journal of Healthcare Informatics Research
which variation in exposure levels observed in real-time could be explained by
exposure estimates approximated solely from residential location. By partitioning
variation in exposure due to routine patterns of mobility, we specifically tackle the
first of Chen and Kwans (2015) three key dimensions for research on retail access:
contextual uncertainty due to real-time mobility patterns.
2 Methods
For clarity of presentation the Methods are organized into three sections: 2.1 Data
Preparation, 2.2 Tobacco Outlet Exposure Estimation, and 2.3 Statistical Modeling.
Section 2.2 describes the calculation of home and real-time (hourly) exposure esti-
mates, and Section 2.3 describes the statistical analysis framework, which is a special
case of multinomial logistic regression, providing a framework for interpretation of the
model estimation process. Agresti [48] shows the direct correspondence between the
log-linear and logit modeling frameworks, both in mathematical and inferential terms,
and we present an application of this approach here.
2.1 Data Preparation
2.1.1 Longitudinal Human Mobility Data
Longitudinal human mobility data was continuously recorded via the OpenPaths
(https://openpaths.cc) platform, launched by the New York Times Company Research
and Development Lab in May 2011 [49]. Participants who provided consent to
participate in this research and downloaded the OpenPaths application for either iOS
or Android were then able to use the application to continuously capture their current
geographic locationi.e., latitude and longitude, wirelessly uploaded according to a
10-min sampling rate. When participants remained stationary (less than 5 m location
change between successive coordinates), the OpenPaths application suspended data
collection to save battery, ensuring that OpenPaths data represents periods of at least
minimal mobile activity rather than extended periods of rest, such as while people sleep
at night or sit for long periods at their desks. While stationary, the application continued
to listenfor a greater than 5-m location change and reinitiated continuous tracking
whenever such a change was detected.
The raw longitudinal mobility dataset used for the present analysis contained
8,458,902 observations collected from 859 individuals worldwide from 01/01/2012
to 06/06/2015. Each observation included a unique user ID, date, time, latitude, and
longitude. The raw mobility data were initially clipped to the United States using state
outline polygons published by United States Census Bureau, yielding a US cohort of
744 individuals with 4,647,152 observations. Data from the lower 48 states and
Washington, DC were then clipped to correspond with the tobacco density layer
(Section 2.2), yielding 729 individuals with 4,556,619 observations. Participants
tracked their location over a mean of 267 days (Median = 178; SD = 278), recording
an average of 6251 location coordinates (Median = 3192; SD = 10,289). To standardize
the amount of data contributed by each individual while maximizing the amount
available for analysis, each participants daily data records were left-truncated, retaining
Journal of Healthcare Informatics Research
their most recent 180 days of tracking data. Individuals with fewer than 180 days of
data were not included in the present analyses, ensuring all between-person contrasts
were based on an equivalent 180-day sampling frame. The final longitudinal mobility
dataset thus included Np= 363 individuals and 1,418,151 real-time observations over
the 180-day study period.
2.1.2 Estimating Tobacco Outlet Density
A nationwide density surface of tobacco retail outlet locations was generated using
kernel density estimation (KDE). This non-parametric method extrapolates from spa-
tially distributed point data by estimating their continuous density using spatial density
functions known as kernels, each of which has a specified circular radius size known as
the bandwidth [50]. Gaussian kernels with a fixed 8047 m (5 mile) bandwidth were
used to generate the final density surface, from which density estimates could be
extracted with a resolution of 250 m.
The empirical basis for this probability density surface was a national dataset of
tobacco retail outlets, identified by North American Industry Classification Systems
(NAICS) codes. Developed by the Office of Management and Budget, NAICS is the
standard used by Federal statistical agencies to classify businesses based on their
primary activity [51]. In 2012, geocoded data was obtained from D&B Hoovers. The
following retail categories and corresponding NAICS codes were included: beer, wine,
and liquor stores (NAICS: 445310); supermarkets and other grocery stores (NAICS:
44511); convenience stores (NAICS: 44512); pharmacies and drug stores (NAICS:
446110); gasoline stations with convenience stores (NAICS: 44711); other gasoline
stations (NAICS: 44719); department stores (NAICS: 452111); discount department
stores (NAICS: 452112); and tobacco stores (NAICS: 453991). For pharmacies and
department stores, we individually reviewed all chains with 50 or more locations to
determine if they sold tobacco and excluded them accordingly [52]. Based on this
analysis, we also excluded all other department stores and pharmacies as they likely do
not sell tobacco. In addition, we excluded major department chains and grocery stores
that, based on their stores policy, do not sell tobacco products (i.e., Target, Whole
Foods, Trader Joes, Wegmans). We also excluded pharmacies and drug stores in the 55
Massachusetts and 2 California municipalities that have banned the sale of tobacco
products within these establishments. The final dataset comprised N= 269,781 retail
outlets.
2.2 Tobacco Outlet Exposure Estimation
2.2.1 Residential Exposure
Residential locations were extracted from participantsOpenPaths mobility data using a
two-step procedure based on Toole et al. [53], and originally given by Zheng and Xie
[54]. In the first step, meaningful locationsstay eventswere extracted from mobility
data for each participant for each nighttime period (8 pm7 am). A second step then
combines each participants set of nighttime stay events covering the entire 180 day
analysis period to a set of aggregate stay points. The aggregate stay point that comprises
the most nighttime stay events for each participant was selected as the maximally
Journal of Healthcare Informatics Research
unbiased estimator of residential location. Residential retail exposures were then
generated by spatial joining each participants inferred residential location to the density
surface of tobacco retail outlet locations (Mean = 9.50, Median = 3.81, SD = 12.25).
Exposures were square root transformed owing to right skew, and classified into five
groups based on exposure intensity.
2.2.2 Real-Time Exposure
ParticipantsOpenPaths mobility data was used to compute hourly radius of gyration
(Rg), measured in meters [28]. Rgestimates the size and spread of a participants
personal activity space for a given hour. Rgis defined by the standard deviation between
locations and their center of mass:
Rg¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
N
N
k¼1
rkrmean
ðÞ
2
s
where Nis the total number of location coordinates collected from each individual per
hour, and rmean is the individual center of mass, or the mean longitude and latitude of all
Nlocations. The great circle distance in meters between a specific location and the
center of mass (rkrmean) was calculated using Vincentysformulae[55]. The dataset
was then aggregated over 465,279 hourly observations of Rg. Given the present focus
on routine day-to-day mobility patterns, long-range travel was excluded by dropping
observations with Rglarger than 160 km (Ndrop = 945), the maximum distance a car can
travel in 1 h. Hours with zero movement (Rg= 0) were then excluded, yielding 363
individuals with a total 304,164 hourly observations (65.4% of the hourly data).
Real-time exposure was conceptualized as the product of a participantshourly
movement (Rg) and their hourly aggregate exposure to retail outlets (Mean = 22,620,
Median = 7.01, SD = 144,191). Each real-time mobility coordinate contributed by the
participants was joined to a tobacco outlet density value extracted from the KDE
surface. Hourly exposure levels are, thus, the product of each participantsRgwithin
each hour under observation and the average tobacco outlet density value across the set
of mobility coordinates recorded within the same hour. This exposure variable approx-
imates the number of tobacco outlets that surrounded each participant within each hour
of the study and, as expected from a count variable of this kind, the observed
distribution of exposure values was heavily skewed to the right. To improve corre-
spondence with the standard assumptions of a categorical count-based data analysis
framework, both residential and hourly exposure values were square root transformed
and determined to closely follow a negative binomial distribution. This produced real
valued outputs in the range 027, which were divided into 27 clusters by binning with a
one unit spacing. Each cluster effectively labels all participant hours within a particular
range of exposure intensity, with subsequent clusters representing the increasing
intensity of participantshourly exposures in this sample.
To model temporal patterns of exposure, time-related variables based on social
convention were derived for each observation. These were time-of-day (24-h clock),
Journal of Healthcare Informatics Research
day-of-week (weekday versus weekend), and season. Time of day was defined by four
6-h windows: 3:009:00 as early,9:0015:00 as day,15:0021:00 as evening,
and 21:003:00 as late,treating observations recorded between midnight and 3 AM
as part of the preceding day [56]. Day-of-week was coded binary, indicating whether
each observation fell on a weekend, defined as falling after 17:00 Friday through 17:00
Sunday. Season was also defined as categorical, with DecemberFebruary as winter,
MarchMay as spring, JuneAugust as summer, and SeptemberNovember as fall.
2.3 Statistical Analyses
2.3.1 Log-Linear Model Selection
The interactive association between real-time exposures (27 clusters) and residential (5
clusters) exposures was stratified across time-of-day (4 categories), day-of-week (2
categories), and season (4 categories), using a set of multivariate contingency tables
that populated a 5-dimensional matrix with a total of 27×5×4×2×4 = 4320 cells.
Patterns of association within this large multivariate contingency table were analyzed
with generalized categorical data analysis techniques [48]. Specifically, we employed
an exponential, log-linear modeling framework. Log-linear models convert the multi-
plicative relations among joint and marginal counts in a contingency table to additive,
linear associations by transforming the counts to logarithms [48]. Hierarchically nested
model comparison techniques were used to iteratively identify the most parsimonious
combination of factors required to explain the observed data. Systematic comparison of
hierarchically nested log-linear models produced a likelihood ratio test statistic pre-
sented in the text. The saturated model represents the log frequencies for the cell index
(h,w,t,s,e) of all non-ordinal combinations of both real-time exposure and residential retail
exposure (home), time-of-day, weekend, and season:
ln μhwtse

¼λþλH
hþλW
wþλT
tþλS
sþλE
eþλHW
hw þþλHWT
hwt þþλHWTE
hwte
þþλHWTSE
hwtse ;h
¼1;;5;w¼1;2;t¼1;;4;s¼1;;4;e¼1;;27:
Where His home exposure, Tis time-of-day, Wis weekend, Sis season, and Eis real-
time exposure. h,t,w,s,andeare categories within H,T,W,S,andE.μhwtse represents
the expected cell frequencies in the five-dimensional contingency table. λis the
constant. λH
hdenotes the row effect of λH
i;ih.λW
w,λT
t,λS
s,andλE
ealso represent
row effects. λHW
hw is the interaction term λHW
ij between H and W, where ih,jw.
λHWT
hwt ,,λHWTE
hwte ,,λHWTSE
hwtse are higher dimensional interaction terms. For easier read,
letter symbols are used in Table 1to represent the highest interaction model terms.
2.3.2 Log-linear Model Interpretation: Ordinal Logit Modeling
Ordinal logit models were used to interpret observed associations within a log-linear
framework, utilizing well-established logistic model reporting and interpretation
Journal of Healthcare Informatics Research
standards [48]. In this paper, log-linear estimation and selection was used to identify the
most parsimonious model, best fitting the observed data, and then ordinal logit
modeling was used to examine our primary aim: examining the extent to which
variation in exposure to tobacco retail outlets observed in real-time could be explained
by exposure estimates approximated as a function of residential location and time
(time-of-day, day-of-week, season-of-year). Following Agresti (2012), this was accom-
plished by setting the residential exposure variable (home) from within the log-linear
model as an ordinal dependent variable. Within this framework, each ordinal residential
exposure level represented a clusterof participants, and model results estimate the
probabilityi.e., the log-odds ratio or logitthat each real-time exposure observation
was contributed by a participant from each of the residential density clusters.
3Results
Overall, mean hourly Rgwas 0.55 km with a standard deviation of 2.66 km. Minimum
and maximum were 0 and 152.94 km. Within each day, Rgwas the lowest in the middle
of the night and higher across the remainder of the day. Two spikes were observed in Rg
in early morning and late afternoon on weekdays, which is likely related to commuting
between home and workwhile there was generally more variation on weekends.
Figure 1illustrates the daily drop in mobility across the late-night hours, followed by a
steep rise across the early morning, and then divergence on Saturday and Sunday, with
early Sunday afternoon revealed as the window of greatest mobility on average.
Table 1 Step-down contrasts of best-in-classmodels
Model (Model Terms) Deviance df pvalue G2Δdf p(Δdf;
0.01)a
1.1 (W, T, H, S, E) 321,168.90 302,853 0.00*** 1057.6 54 < 0.001b
1.2 (WT, WH, WS, WE, TH, TS,
TE, HS, HE, SE)
320,111.30 302,907 0.00*** 108,897.9 495 < 0.001b
1.3 (HTE, WTE, HWT, HWE, HES,
HTS, TSE, WSE, WTS,
HWS)
429,009.20 303,402 0.00*** 169,139.32 286 < 0.001b
1.4 (WTSE, TSHE) 259,869.88 303,688 1.00 12,713.60 16 < 0.001b
1.5 (WTHE, TSHE) 272,583.48 303,704 1.00 158,014.6 127 < 0.001b
1.6 (WTHE, WSHE, TSHE) 114,568.88 303,831 1.00 50,677.26 77 < 0.001b
1.7 (WTHE, WTSE, WSHE, TSHE) 64,396.25 303,902 1.00 15,677.26 36 < 0.001b
1.8 (WTHE, WTSE, WTSH, WSHE,
TSHE)
48,718.99 303,938 1.00 48,718.99 303,938 1.00
1.9 (WTSHE) 0.00 0 1.00 ––
Wweekend, Ttime of day, Sseason, Hresidential exposure cluster, Ereal-time exposure level
***Significant lack of fit at 99% confidence level
apvalue of G2comparing to the model below
bHierarchically nested models are significantly different at 99.9% confidence level
Journal of Healthcare Informatics Research
Figure 2presents generalized additive model smoothed real-time exposure intensity
by time of day for each residential exposure cluster. On weekdays, real-time exposure
spikes in the early morning and late afternoon across all residential exposure clusters,
probably due to increased movement and exposure to retail outlets during the commute
to and from work. However, this relationship becomes more prominent as residential
exposure cluster increases from low to high. In contrast, weekend exposure was
elevated at noon and then remained high until early evening, and was broadly consis-
tent across residential exposure clusters, except in the high exposure cluster which
deviates from the patterns observed.
3.1 Model Fitting: Best-in-class Log-linear Model Selection
Tab le 1presents anoverview of the best-in-class model selection process used to identify
the most parsimoniousmodel, defined asthe minimal set of parameters required to provide
an adequate fit to the observed data. Following hierarchically nested model comparison
techniques, Model performance is evaluated by both degreeof freedom (df) and likelihood
ratio test statistic (G2). pvalue represents lack of fit of the models, which means models
withpvalueof 1 fit at 99% confidencelevel and pvalueof 0 indicates model not fitting.The
initial basis for comparison is Model 1.9, which is the saturated model that corresponds
perfectly to the raw data, having zero degrees of freedom, as the number of parameters is
equivalentto the total number of cells generated by all interactive combinations of the five
factors under study: retail exposure (the conceptual dependent variable), residential
exposure (home), weekend, time-of-day (time), and season. Model 1.4 fits while using
only 829 cells to model the total 1325 cells under study. It also accurately predicts a five-
dimensional contingency table with lower level four-way interactions. This is the most
Fig. 1 Radius of Gyration Measured Mobility Patterns by Time of Day across Day of Week
Journal of Healthcare Informatics Research
parsimonious model, effectively isolating an informative pattern in the data that then
becomes the basis for inference.
To measure the separate strength and significance of each interaction in Model 1.4,
models excluding one of the two interaction terms were evaluated. Table 2shows the
likelihood ratio evaluation of the influence of the excluded term compared to Model 1.4
with all two-term four-way interactions. This method provides a specific test of
conditional independence between seasonal effects and real-time exposure levels.
These model fits indicate that while the interactions between weekend effects and other
factors are important, their associations do not contribute in Model 1.5 as much as the
interactions between residential (home) exposure, time, season, and real-time exposure.
3.2 Model Inference: Ordinal Logit Modeling of Residential Density Clusters
To examine real-time tobacco retail outlet exposures among clusters of residential
exposure, an ordinal logit model was constructed to be mathematically equivalent to
the most parsimonious log-linear model (see Section 2.3.2 Tabl e 1Model 1.4), itself
identified through the model selection process described in Section 3.1:
logit P H ¼hjT¼t;S¼s;E¼eðÞ½¼αþβT
tþβS
sþβE
eþβTS
ts þβTE
te þβSE
se
þβTSE
tse þε;h
¼1;;5;t¼1;;4;s¼1;;4;e¼1;;27:
Fig. 2 Weekday and Weekend Real-time Exposure by Time of Day for Each Residential Exposure Cluster
Journal of Healthcare Informatics Research
where the log-odds ratio (i.e., logit) of residential exposure cluster membership (home)
for each hour is modeled as a function of the real-time exposure level (H), time-of-day
(T), and season (S) associated with each hour. αis a constant. h,t,s,andeare
categories within H,T,S,andE,andεis the error term. βT
t,βS
s,andβE
erepresent the
effects of parameter T,S,andErespectively. βTS
ts ,,βTSE
tse represent the effects of
interactions between parameters. For easier read, letter symbols are used in Table 2to
represent the highest interaction model terms. Model predicted residential exposure
cluster membership and associated confidence intervals were generated via
bootstrapping with 500 random samples of 10,000 observations, drawn with replace-
ment from the empirical data distribution, effectively capturing the uncertainty associ-
ated with each parameter estimated by the model. This simulation-based resampling
approach allowed for precise discrimination between the different residential exposure
clusters (Fig. 3).
Figure 3presents the model predicted results of this process, illustrating the
predicted probability that each of the 304,164 h under study was contributed by a
participant from each of the residential exposure clusters. Separation between the
bootstrapped 95% confidence intervals corresponds to regions of the distribution of
real-time hourly exposures that were significantly explained by one or more of the
density clusters. Overall, 61.3% of real-time, hourly exposures were of relatively low
intensity, and after controlling for temporal and seasonal variation, 72.8% of the
variance among these low-level exposures was accounted for by residence in one of
the two lowest residential density quintiles. Residence in one of the two highest
residential density quintiles accounted for approximately 50% of the variance among
extreme exposure levels, but extreme levels of exposure were rare, constituting about
1% of the data. Altogether 55.2% of the variance in real-time exposures was not
explained by participantsresidential exposure cluster, and most moderate to high
intensity real-time exposures (38.7% of all hourly exposures) were no more likely to
have been contributed by subjects from any single residential density cluster than
another. In sum, OpenPaths participants experienced a heterogeneity in hourly
tobacco retail outlet exposures that is only partially explained by their static residen-
tial exposures.
Table 2 Likelihood ratio evaluation: removal of four-way terms from maximally parsimonious model
Model Hypothesis Log-likelihood df G2Δdf p(Δdf; 0.01)a
Maximally parsimonious model
1.4 (WTSE, TSHE)ψ1,363,431.41 303,688 ––
Conditional independence
2.1 WTSE = 0 2,116,700.547 303,123 1,506,538.272 565 < 0.001b
2.2 TSHE = 0 2,037,315.705 303,532 1,347,768.588 156 < 0.001b
W: weekend, Ttime-of-day, Sseason, Hresidential exposure cluster, Ereal-time exposure level
ψModel 1.4 from Table 1
apvalue of G2comparing to the Model 1.4
bSignificantly different at 99.9% confidence level compared with Model 1.4
Journal of Healthcare Informatics Research
4Discussion
While environmental exposuresare most commonly thought of as biological
internalcontact with toxic particles in the environment (e.g., air pollution and
infectious pathogens) there is growing recognition that monitoring of exposures to
the broader ecosphere or eco-exposomeis also important [5759]. Individual-level
geographic location data provide a spatial linkage that makes it possible to estimate the
multivariate impact of countervailing societal and environmental systems on individual
decision-making and behavior. This gives rise to the possibility of using such infor-
mation for disease prevention and intervention delivery. It is our position, however, that
continuous, real-time location data need not and should not be limited to use within
real-time, just-in-time adaptiveinterventions. In fact, we believe the present paper
provides an example of the way such micro real-time data can be better understood
when it is aggregated, because it is only then that we can properly account for the
relative significance of the various locations each participant frequents, at least as they
pertain to the tobacco point-of-sale landscape.
Traditional estimates of exposure to risk and protective factors in neighborhoods are
founded on the idea that the relative concentration of health-related factors around
peoples homes sufficiently captures and thus can be used to characterize aggregated
patterns of environmental exposure within and between neighborhood areas. This paper
evaluated the degree to which residential locations approximate actual exposures by
comparing empirical observations collected in real-time with static neighborhood
estimates that only used information about each participants residence. Results dem-
onstrate the utility of a continuous geolocation data smoother (i.e., Rg)thatmakesit
possible to generate dynamic, mobility weighted KDE exposure values that retain both
Fig. 3 Real-time Exposure for Each Residential Exposure Cluster
Journal of Healthcare Informatics Research
spatial and temporal resolution. Findings suggest that real-time exposures are
misclassified by person-level residential exposure estimates to a substantial degree,
especially among people residing in areas characterized by moderate to high levels of
residential density.
Essentially, results of this work indicate that exposures to moderate and high levels
of tobacco outlet density were systematically less-likely than lower density exposures,
and that subjects who resided in moderate to high density areas were less likely to
experience real-time exposures that were as high as estimated by the observed density
around their residential location. This finding suggests that residence-based neighbor-
hood approximations exhibited a tendency to over-estimate exposure levels experi-
enced by residents in the real-world, and somewhat counter-intuitively, that this may
have been particularly true within dense urban areas, where despite high-levels of
density overall, shorter travel distances among a smaller set of stores dampened
observed hourly exposure levels.
This paper advances the literature in a number of ways. Focusing on a reliable,
national source of tobacco outlet data allowed us to identify variation in urban
dynamics and behavior across different regions of the US. The use of continuous
real-time geo-location tracking provides excellent temporal and spatial resolution,
which improved sensitivity to detect dynamic patterns in the data. The analysis
framework developed here can be used to assess mobility patterns, exposure to points
of interest, and associated effects on health behavior. Nevertheless, this study also had
methodological limitations that should be considered. This sample is not nationally
representative, as participation in the mobility tracking was based on self-selection, and
required access to the Internet and a smartphone. Additionally, because no participant
demographic information was available, other factors that could potentially affect
participantsmobility, such as occupation or income level, could not be measured.
5Conclusions
Results of this work shed light on the nature of real-time exposures to a spatially
distributed environmental risk factor, as compared to a commonly used neighborhood
exposure estimate. Future work should leverage methods of this kind to advance our
understanding of individual decision-making and behavior change dynamics as a
function of environmental conditions. Natural extensions would incorporate other
policy and health-relevant risk and protective factors, such as the distribution of food,
alcohol, and cannabis products. Research that involves clinical populations attempting
to modify habitual health behaviors would be useful, including work with patients
working to adhere to dietary restrictions or quit cigarette smoking. It will also be
interesting to investigate the basic mechanisms underlying these associations, such as
memory and other cognitive processes affected by regular product exposures and
associated preferences. Variations across geographic areas and over time may provide
insight and identify targets of intervention for public health practitioners, urban plan-
ners, and policy makers.
Acknowledgements The authors are indebted to Michael Tacelosky, Morgane Bennett, Ollie Ganz, and
Seann Regan for their assistance.
Journal of Healthcare Informatics Research
Funding information This work was supported by the National Institute on Drug Abuse, National Cancer
Institute, and Office of Behavioral and Social Science Research; R01DA034734 & R01DA034734 (TRK).
Funding was also provided by the GeoSpatial Resource, part of the Norris Cotton Cancer Centers Biostatistics
Shared Resource [5P30CA023108, UL1TR001086].
Compliance with Ethical Standards
Conflict of Interest The authors declare that they have no real or potential conflicts of interest.
PublishersNote Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.
References
1. Diez Roux AV, Mair C (2010) Neighborhoods and health. Ann N YAcad Sci 1186:125145. https://doi.
org/ 10.1111 /j.1 749- 6632.2009.05333.x
2. Riva M, Gauvin L, Barnett T (2007) Toward the next generation of research into small area effects on
health: a synthesis of multilevel investigations published since July 1998. J Epidemiol Community
Health 61:853861
3. Goodchild M, Janelle D (2004) Spatially integrated social science. Oxford University Press, New York
4. Duncan D, Kawachi I, Subramanian S, Aldstadt J, Melly S, Williams D (2014) Examination of how
neighborhood definition influences measurements of youthsaccess to tobacco retailers: a methodolog-
ical note on spatial misclassification. Am J Epidemiol 179(3):373381
5. Kirchner TR, Shiffman S (2016) Spatio-temporal determinants of mental health and well-being: advances
in geographically-explicit ecological momentary assessment (GEMA). Soc Psychiatry Psychiatr
Epidemiol 51(9):12111223. https://doi.org/10.1007/s00127-016-1277-5
6. Matthews SA, Yang T-C (2013) Spatial polygamy and contextual exposures (SPACEs): promoting
activity space approaches in research on place and health. Am Behav Sci 57(8):10571081
7. Perchoux C, Chaix B, Cummins S, Kestens Y (2013) Conceptualization and measurement of environ-
mental exposure in epidemiology: accounting for activity space related to daily mobility. Health & Place
21:8693. https://doi.org/10.1016/j.healthplace.2013.01.005
8. Cantrell J, Kreslake JM, Ganz O, Pearson JL, Vallone D, Anesetti-Rothermel A, Xiao H, Kirchner TR
(2013) Marketing little cigars and cigarillos: advertising, price, and associations with neighborhood
demographics. Am J Public Health 103(10):19021909. https://doi.org/10.2105/AJPH.2013.301362
9. Kirchner TR, Cantrell J, Ansetti-Rothermel A, Ganz O, Vallone DM, Abrams DB (2013) Real-time
exposure to point-of-sale tobacco marketing predicts daily smoking status and smoking cessation
outcomes: a multilevel geo-spatial analysis. Am J Prev Med 45(4):379385
10. Kirchner TR, Anesetti-Rothermel A, Bennett M, Gao H, Carlos H, Scheuermann TS et al Tobacco outlet
density and converted versus native nondaily cigarette use in a national U.S. sample. Tob Control in press
11. Cantrell J, Ganz O, Anesetti-Rothermel A, Harrell P, Kreslake JM, Xiao H, Pearson JL, Vallone D,
Kirchner TR (2015) Cigarette price variation around high schools: evidence from Washington DC.
Health & Place 31:193198
12. Kirchner TR, Villanti AC, Cantrell J, Anesetti-Rothermel A, Ganz O, Conway KP, et al (2014) Tobacco
retail outlet advertising practices and proximity to schools, parks and public housing affect Synar
underage sales violations in Washington, DC. Tob Control, Published Online First: 25 February 2014.
https://doi.org/10.1136/tobaccocontrol-2013-051239
13. Ganz O, Cantrell J, Moon-Howard J, Aidala A, Kirchner TR, Vallone D (2015) Electronic cigarette
advertising at the point-of-sale: a gap in tobacco control research. Tob Control 24(e1):e110e1e2
14. Cantrell J, Anesetti-Rothermel A, Pearson JL, Xiao H, Vallone D, Kirchner TR (2015) The impact of the
tobacco retail outlet environment on adult cessation and differences by neighborhood poverty. Addiction
110(1):152161
15. Ilakkuvan V, Tacelosky M, Ivey KC, et al (2014) Cameras for public health surveillance: a methods protocol
for crowdsourced annotation of point-of-sale photographs. 3(2):e22. https://doi.org/10.2196/resprot.3277
Journal of Healthcare Informatics Research
16. Kirchner TR, Villanti A, Anesetti-Rothermel A, Tacelosky M, Pearson J, Cantrell J et al (2015) National
enforcement of the Family Smoking Prevention and Tobacco Control Act at point-of-sale. Tob Regul Sci
1(1):2435. https://doi.org/10.18001/TRS.1.1.3
17. Cantrell J, Pearson JL, Anesetti-Rothermel A, Xiao H, Kirchner TR, Vallone D (2016) Tobacco retail
outlet density and young adult tobacco initiation. Nicotine Tob Res 18(2):130137
18. Cantrell J, Ganz O, Ilakkuvan V, Tacelosky M, Kreslake J, Moon-Howard J, Aidala A, Vallone D,
Anesetti-Rothermel A, Kirchner TR (2015) Implementation of a multimodal mobile system for point-of-
sale surveillance: lessons learned from case studies in Washington, DC, and New York City. JMIR Public
Health and Surveill 1(2):e20
19. Kirchner TR, Anesetti-Rothermel A, Bennett M, Gao H, Carlos H, Scheuermann TS, Reitzel LR,
Ahluwalia JS (2016) Tobacco outlet density and converted versus native non-daily cigarette use in a
national US sample. Tob Control 26:8591. https://doi.org/10.1136/tobaccocontrol-2015-052487
20. Slater S, Chaloupka FJ, Wakefield M, Johnston LD, O'Malley PM (2007) The impact of retail cigarette
marketing practices on youth smoking uptake. Arch Pediatr Adolesc Med 161(5):440445. https://doi.
org/10.1001/archpedi.161.5.440
21. Feighery EC, Henriksen L, Wang Y, Schleicher NC, Fortmann SP (2006) An evaluation of four measures
of adolescentsexposure to cigarette marketing in stores. Nicotine Tob Res 8(6):751759. doi:
G6730P56K6W02647 [pii]. https://doi.org/10.1080/14622200601004125
22. Carter OB, Mills BW, Donovan RJ (2009) The effect of retail cigarette pack displays on unplanned
purchases: results from immediate postpurchase interviews. Tob Control 18(3):218221
23. Institute PoPA (1992) The point-of-purchase advertising industry fact book. The Point of Purchase
Advertising Institute, Englewood
24. Henriksen L, Schleicher NC, Feighery EC, Fortmann SP (2010) A longitudinal study of exposure to retail
cigarette advertising and smoking initiation. Pediatrics 126(2):232238. https://doi.org/10.1542
/peds.2009-3021
25. Wakefield M, Germain D, Henriksen L (2008) The effect of retail cigarette pack displays on impulse
purchase. Addiction 103(2):322328
26. Cantrell J, Pearson J, Ansetti-Rothermel A, Haijun X, Vallone D, Kirchner TR Tobacco retail outlet
density and young adult tobacco initiation. Nicotine Tob Res in press
27. Siahpush M, Jones PR, Singh GK, Timsina LR, Martin J (2010) The association of tobacco marketing
with median income and racial/ethnic characteristics of neighbourhoods in Omaha, Nebraska. Tob
Control 19(3):256258. https://doi.org/10.1136/tc.2009.032185
28. Laws MB, Whitman J, Bowser DM, Krech L (2002) Tobacco availability and point of sale marketing in
demographically contrasting districts of Massachusetts. Tob Control 11(Suppl 2):ii71ii73
29. Paynter J, Edwards R (2009) The impact of tobacco promotion at the point of sale: a systematic review.
Nicotine Tob Res 11(1):2535
30. John R, Cheney MK, Azad MR (2009) Point-of-sale marketing of tobacco products: taking advantage of
the socially disadvantaged? J Health Care Poor Underserved 20(2):489506
31. Lovato C, Watts A, Stead LF (2011) Impact of tobacco advertising and promotion on increasing
adolescent smoking behaviours. London: John Wiley & Sons, Ltd. Cochrane Database of Systematic
Reviews (10) Art. No.: CD003439. https://doi.org/10.1002/14651858.CD003439.pub2
32. Feighery EC, Schleicher NC, Cruz TB, Unger JB (2008) An examination of trends in amount and type of
cigarette advertising and sales promotions in California stores, 20022005. Tob Control 17(2):9398
33. Black J, Macinko J (2014) Neighborhoods and obesity. Nutr Rev 66(1):220
34. Hartig T, Mitchell R, de Vries S, Frumkin H (2014) Nature and health. Annu Rev Public Health 35:207228
35. Larson N, Story M, Nelson M (2009) Neighborhood environments: disparities in access to healthy foods
in the U.S. Am J Prev Med 36(1):7481
36. Pickett K, Pearl M (2001) Multilevel analyses of neighborhood socioeconomic context and health
outcomes. J Epidemiol Community Health 55:111122
37. Openshaw S (1984) The modifiable areal unit problem. GeoBooks, Norwich
38. Kwan M-P (2012) The uncertain geographic context problem. Ann Assoc Am Geogr 102(5):958968
39. Chen X, Kwan MP (2015) Contextual uncertainties, human mobility, and perceived food environment:
the uncertain geographic context problem in food access research. Am J Public Health 105(9):1734
1737. https://doi.org/10.2105/AJPH.2015.302792
40. James P, Berrigan D, Hart JE, et al (2014) Effects of buffer size and shape on associations between the
built environment and energy balance. Health & place 27:162170. https://doi.org/10.1016/j.
healthplace.2014.02.003
41. Spielman SE, Yoo EH (2009) The spatial dimensions of neighborhood effects. Soc Sci Med 68(6):1098
1105. https://doi.org/10.1016/j.socscimed.2008.12.048
Journal of Healthcare Informatics Research
42. Kwan M-P (2013) Toward temporally integrated geographies of segregation, health, and accessibility.
Ann Assoc Am Geogr 103(5):10781086
43. Rainham D, McDowell I, Krewski D, Sawada M (2010) Conceptualizing the healthscape: contributions
of time geography, location technologies and spatial ecology to place and health research. Soc Sci Med
70(5):668676. https://doi.org/10.1016/j.socscimed.2009.10.035
44. Cummins S (2007) Commentary: investigating neighbourhood effects on healthavoiding the local
trap. Int J Epidemiol 36(2):355357. https://doi.org/10.1093/ije/dym033
45. Caspi CE, Sorensen G, Subramanian SV, Kawachi I (2014) The local food environment and diet: a
systematic review. Health & place 18(5):11721187. https://doi.org/10.1016/j.healthplace.2012.05.006
46. Kavanagh AM, Kelly MT, Krnjacki L, Thornton L, Jolley D, Subramanian SV, Turrell G, Bentley RJ
(2011) Access to alcohol outlets and harmful alcohol consumption: a multi-level study in Melbourne,
Australia. Addiction 106(10):17721779. https://doi.org/10.1111/j.13 60-0443.2011. 03510.x
47. Drewnowski A (2004) Obesity and the food environment: dietary energy density and diet costs. Am J
Prev Med 27(3 Suppl):154162. https://doi.org/10.1016/j.amepre.2004.06.011
48. Agresti A (1990) Categorical data analysis. Wiley, New York
49. House B (2013) OpenPaths: empowering personal geographic data. In: Proceeding of the ISEA; Sydney,
Australia
50. Carlos HA, Shi X, Sargent J, Tanski S, Berke EM (2010) Density estimation and adaptive bandwidths: a
primer for public health practitioners. Int J Health Geogr 9:39. https://doi.org/10.1186/1476-072X-9-39
51. U.S. Census Bureau (2014) Introduction to NAICS [cited 2014 June 9]. Available from: https://www.
census.gov/eos/www/naics/index.html
52. DAngelo H, Fleischhacker S, Rose SW, Ribisl KM (2014) Field validation of secondary data sources for
enumerating retail tobacco outlets in a state without tobacco outlet licensing. Health & place 28:3844.
https://doi.org/10.1016/j.healthplace.2014.03.006
53. Toole JL, Colak S, Sturt B, Alexander LP, Evsukoff A, Gonzalez MC (2015) The path most traveled:
travel demand estimation using big data resources. Transp Res C 58:162177
54. Zheng Y, Capra L, Wolfson O, Yang H (2014) Urban computing: concepts, methodologies, and
applications. In: proceedings of the ACM TIST. ACM Press, New York
55. Vincenty T (1975) Direct and inverse solutions of geodesics on the ellipsoid with application of nested
equations. Surv Rev 23:8893
56. Kirchner TR, Shiffman S (2013) Ecological momentary assessment. In The WileyBlackwell Handbook
of Addiction Psychopharmacology (eds J. MacKillop and H. de Wit). https://doi.org/10.1002
/9781118384404.ch20
57. Lioy PJ, Smith KR (2013) A discussion of exposure science in the 21st century: a vision and a strategy.
Environ Health Perspect 121:405409
58. Chokshi DA, Farley TA (2014) Health. changing behaviors to prevent noncommunicable diseases.
Science 345(6202):12431244. https://doi.org/10.1126/science.1259809
59. Cromley EK, McLafferty SL (2012) GIS and public health, 2nd edn. Guilford Press, New York
Affiliations
Thomas R. Kirchner
1,2,3
&Hong Gao
1
&Daniel J. Lewis
4
&Andrew
Anesetti-Rothermel
5
&Heather A. Carlos
6
&Brian House
7
*Thomas R. Kirchner
tom.kirchner@nyu.edu
1
College of Global Public Health, New York University, 715 Broadway, 12th Floor, New York,
NY 10003, USA
2
Center for Urban Science and Progress, New York University, New York, NY, USA
3
Department of Population Health, New York University Medical Center, New York, NY, USA
Journal of Healthcare Informatics Research
4
Department of Social and Environmental Health Research, London School of Hygiene and Tropical
Medicine, London, UK
5
Schroeder Institute at Truth Initiative, Washington, DC, USA
6
Norris Cotton Cancer Center, Dartmouth College, Hanover, NH, USA
7
Brown University, Providence, RI, USA
Journal of Healthcare Informatics Research
... The tobacco control and health geography literatures have sought to address various aspects of tobacco use in context. The bulk of these studies can be categorized into four areas: (1) emerging tobacco geographies theory (Thompson et al., 2007;Barnett et al., 2016;Pearce et al., 2012;Frohlich et al., 2002;Haines-Saah et al., 2013;McQuoid et al., 2020); (2) tobacco point-of-sale (POS) studies, including those focused on both exposure and retail density (Abdel Magid et al., 2020;Cantrell et al., 2016;Henriksen et al., 2008;Novak et al., 2006;Ogneva--Himmelberger et al., 2010;Tucker-Seeley et al., 2016;Anesetti-Rothermel et al., 2020;Golden et al., 2012;Yu et al., 2010;Kirchner et al., 2013Kirchner et al., , 2019Pearce et al., 2009Pearce et al., , 2016Watkins et al., 2013); (3) spatial concentrations of urban advertising and marketing in targeted communities (Giovenco et al., 2018;Hackbarth et al., 1995;Lee et al., 2015a;Ribisl et al., 2017;Robertson et al., 2014;Stoddard et al., 1998;Vardavas et al., 2009;Yerger et al., 2007;Henriksen et al., 2020); and (4) disparities in neighborhood and area-level characteristics as they relate to individual tobacco use behavior, or tobacco retail clustering (Anesetti-Rothermel et al., 2020;Yerger et al., 2007;Duncan et al., 2016;Fakunle et al., 2019;Farley et al., 2019;Holmes et al., 2020;Galea et al., 2004Galea et al., , 2007Patterson et al., 2012;Wilson et al., 2005;Hatzenbuehler et al., 2014). Collectively, this literature has confirmed the necessity of understanding local tobacco use dynamics for developing effective policy and public health interventions. ...
Article
The tobacco retail environment is where most advertising dollars are spent. However, most research on the retail environment has not methodologically situated tobacco retailers as part of a larger community, and few studies have incorporated community member perspectives of their own tobacco use in relation to their local environments. The purpose of this study is to describe and evaluate a multilevel, multimodal, mixed methods approach for understanding tobacco use in context. We combine quantitative data collected from tobacco retailer audits and geographically-explicit interviews with neighborhood residents to tell a more complete story of tobacco use behavior among adults in San Francisco's Marina district, and the Oakland Coliseum neighborhood in Alameda County, California. We find that while area-level and retail data provide a broad snapshot of two distinct communities with respect to sociodemographic characteristics and tobacco availability, interviews with community residents who use tobacco add important perspectives regarding how tobacco retailers are viewed and how residents interact with their neighborhood landscapes on a daily basis. The method we describe and critique has the potential to be scaled to incorporate a broader set of geographies, or tailored to address a multitude of health-related questions. Our approach further demonstrates the utility of including geolocated participant narratives as a means of understanding where researcher interpretations of urban environments diverge from those of community residents.
... The tobacco control and health geography literatures have sought to address various aspects of tobacco use in context. The bulk of these studies can be categorized into four areas: (1) emerging tobacco geographies theory (Thompson et al., 2007;Barnett et al., 2016;Pearce et al., 2012;Frohlich et al., 2002;Haines-Saah et al., 2013;McQuoid et al., 2020); (2) tobacco point-of-sale (POS) studies, including those focused on both exposure and retail density (Abdel Magid et al., 2020;Cantrell et al., 2016;Henriksen et al., 2008;Novak et al., 2006;Ogneva--Himmelberger et al., 2010;Tucker-Seeley et al., 2016;Anesetti-Rothermel et al., 2020;Golden et al., 2012;Yu et al., 2010;Kirchner et al., 2013Kirchner et al., , 2019Pearce et al., 2009Pearce et al., , 2016Watkins et al., 2013); (3) spatial concentrations of urban advertising and marketing in targeted communities (Giovenco et al., 2018;Hackbarth et al., 1995;Lee et al., 2015a;Ribisl et al., 2017;Robertson et al., 2014;Stoddard et al., 1998;Vardavas et al., 2009;Yerger et al., 2007;Henriksen et al., 2020); and (4) disparities in neighborhood and area-level characteristics as they relate to individual tobacco use behavior, or tobacco retail clustering (Anesetti-Rothermel et al., 2020;Yerger et al., 2007;Duncan et al., 2016;Fakunle et al., 2019;Farley et al., 2019;Holmes et al., 2020;Galea et al., 2004Galea et al., , 2007Patterson et al., 2012;Wilson et al., 2005;Hatzenbuehler et al., 2014). Collectively, this literature has confirmed the necessity of understanding local tobacco use dynamics for developing effective policy and public health interventions. ...
Preprint
Piloting a spatial mixed method for understanding neighborhood tobacco use disparities Abstract The tobacco retail environment is where most advertising dollars are spent. However, most research on the retail environment has not methodologically situated tobacco retailers as part of a larger community, and few studies have incorporated community member perspectives of their own tobacco use in relation to their local environments. The purpose of this study is to describe and evaluate a multilevel, multimodal, mixed methods approach for understanding tobacco use in context. We combine quantitative data collected from tobacco retailer audits and geographically-explicit interviews with neighborhood residents to tell a more complete story of tobacco use behavior among adults in San Francisco's Marina district, and the Oakland Coliseum neighborhood in Alameda County, California. We find that while area-level and retail data provide a broad snapshot of two distinct communities with respect to sociodemographic characteristics and tobacco availability, interviews with community residents who use tobacco add important perspectives regarding how tobacco retailers are viewed and how residents interact with their neighborhood landscapes on a daily basis. The method we describe and critique has the potential to be scaled to incorporate a broader set of geographies, or tailored to address a multitude of health-related questions. Our approach further demonstrates the utility of including geolocated participant narratives as a means of understanding where researcher interpretations of urban environments diverge from those of community residents.
Article
As activity space measures are increasingly used to estimate exposure to environmental determinants of health, little is known about the stability of these measures over time. To test the stability of GPS-derived measures of activity-space footprint, size, and environmental features over time, we compared 14-day measures at baseline and six months later for 35 adults in a large city. Activity-space measures were based on convex hulls and 500 m route buffers, and included the geographic footprint (i.e. location of the activity space), size (i.e., area in square miles; (Cummins, 2007)), and environmental features including supermarket, fast-food restaurant, and parkland density. The proportion of the participants' smaller geographic footprint covered by the larger was, on average, 0.64 (SD 0.17) for the 500 m route buffer and 0.84 (SD 0.18) for the convex hull. Mean percent change in activity space size ranged from 36.3% (mean daily 500 m route buffer) to 221.3% (cumulative convex hull). Mean percent change in the density of environmental features ranged from 28.8 to 66.5%. Forty-one percent to 92.4% of the variance at one timepoint was predicted by environmental features measured within approximately six months. Activity-space size and environmental features were moderately to highly stable over six months, although there was considerable variation in stability between measures. Strategies for addressing measurement error in studies of activity space-health associations are discussed.
Article
Full-text available
OpenPaths, created by the New York Times Company R&D Lab, is a platform that demonstrates the collective value of personal data sovereignty. It was developed in response to public outrage regarding the location record generated by Apple iOS devices. OpenPaths participants store their encrypted geographic data online while maintaining ownership and programmatic control. Projects of many kinds, from mobility research to expressive artwork, petition individuals for access to their data. In the context of locative media practice, OpenPaths expands the notion of the tracing to address the components of an ethical implementation of crowd-sourced geographic systems in the age of "big data".
Article
Full-text available
PurposeOverview of geographically explicit momentary assessment research, applied to the study of mental health and well-being, which allows for cross-validation, extension, and enrichment of research on place and health. Methods Building on the historical foundations of both ecological momentary assessment and geographic momentary assessment research, this review explores their emerging synergy into a more generalized and powerful research framework. ResultsGeographically explicit momentary assessment methods are rapidly advancing across a number of complimentary literatures that intersect but have not yet converged. Key contributions from these areas reveal tremendous potential for transdisciplinary and translational science. Conclusions Mobile communication devices are revolutionizing research on mental health and well-being by physically linking momentary experience sampling to objective measures of socio-ecological context in time and place. Methodological standards are not well-established and will be required for transdisciplinary collaboration and scientific inference moving forward.
Article
Full-text available
With increasing regulation of tobacco industry marketing practices, point-of-sale advertising has become an important channel for promoting tobacco products. One hundred and ten convenience stores in Oklahoma County were surveyed for tobacco-related advertising. There were significantly more point-of-sale tobacco advertisements in low-income and minority neighborhoods than in better educated, higher-income, predominantly White neighborhoods. Storeowners or managers were also interviewed to determine who has decision-making power regarding store signage and placement, and to elicit perceptions of industry tactics. Contracts with tobacco companies leave storeowners with little or no control over promotion of tobacco product,; Within their store, and many are unaware of the implications of the tobacco industry point-of-scale practices. Local ordinances that regulated Outdoor signage reduced outdoor tobacco advertisements, as well as tobacco signage and promotions within the store. Policy change, rather than education targeting storeowners, is recommended as the most effective strategy for reducing point-of-sale tobacco advertising.
Article
Full-text available
Objective Investigate whether non-daily smokers’ (NDS) cigarette price and purchase preferences, recent cessation attempts, and current intentions to quit are associated with the density of the retail cigarette product landscape surrounding their residential address. Participants Cross-sectional assessment of N=904 converted NDS (CNDS). who previously smoked every day, and N=297 native NDS (NNDS) who only smoked non-daily, drawn from a national panel. Outcome measures Kernel density estimation was used to generate a nationwide probability surface of tobacco outlets linked to participants’ residential ZIP code. Hierarchically nested log-linear models were compared to evaluate associations between outlet density, non-daily use patterns, price sensitivity and quit intentions. Results Overall, NDS in ZIP codes with greater outlet density were less likely than NDS in ZIP codes with lower outlet density to hold 6-month quit intentions when they also reported that price affected use patterns (G2=66.1, p<0.001) and purchase locations (G2=85.2, p<0.001). CNDS were more likely than NNDS to reside in ZIP codes with higher outlet density (G2=322.0, p<0.001). Compared with CNDS in ZIP codes with lower outlet density, CNDS in high-density ZIP codes were more likely to report that price influenced the amount they smoke (G2=43.9, p<0.001), and were more likely to look for better prices (G2=59.3, p<0.001). NDS residing in high-density ZIP codes were not more likely to report that price affected their cigarette brand choice compared with those in ZIP codes with lower density. Conclusions This paper provides initial evidence that the point-of-sale cigarette environment may be differentially associated with the maintenance of CNDS versus NNDS patterns. Future research should investigate how tobacco control efforts can be optimised to both promote cessation and curb the rising tide of non-daily smoking in the USA.
Article
Background: In tobacco control and other fields, point-of-sale surveillance of the retail environment is critical for understanding industry marketing of products and informing public health practice. Innovations in mobile technology can improve existing, paper-based surveillance methods, yet few studies describe in detail how to operationalize the use of technology in public health surveillance. Objective: The aims of this paper are to share implementation strategies and lessons learned from 2 tobacco, point-of-sale surveillance projects to inform and prepare public health researchers and practitioners to implement new mobile technologies in retail point-of-sale surveillance systems. Methods: From 2011 to 2013, 2 point-of-sale surveillance pilot projects were conducted in Washington, DC, and New York, New York, to capture information about the tobacco retail environment and test the feasibility of a multimodal mobile data collection system, which included capabilities for audio or video recording data, electronic photographs, electronic location data, and a centralized back-end server and dashboard. We established a preimplementation field testing process for both projects, which involved a series of rapid and iterative tests to inform decisions and establish protocols around key components of the project. Results: Important components of field testing included choosing a mobile phone that met project criteria, establishing an efficient workflow and accessible user interfaces for each component of the system, training and providing technical support to fieldworkers, and developing processes to integrate data from multiple sources into back-end systems that can be utilized in real-time. Conclusions: A well-planned implementation process is critical for successful use and performance of multimodal mobile surveillance systems. Guidelines for implementation include (1) the need to establish and allow time for an iterative testing framework for resolving technical and logistical challenges; (2) developing a streamlined workflow and user-friendly interfaces for data collection; (3) allowing for ongoing communication, feedback, and technology-related skill-building among all staff; and (4) supporting infrastructure for back-end data systems. Although mobile technologies are evolving rapidly, lessons learned from these case studies are essential for ensuring that the many benefits of new mobile systems for rapid point-of-sale surveillance are fully realized.
Article
We examined the uncertainty of the contextual influences on food access through an analytic framework of the uncertain geographic context problem (UGCoP). We first examined the compounding effects of two kinds of spatiotemporal uncertainties on people's everyday efforts to procure food and then outlined three key dimensions (food access in real time, temporality of the food environment, and perceived nutrition environment) in which research on food access must improve to better represent the contributing environmental influences that operate at the individual level. Guidelines to address the UGCoP in future food access research are provided to account for the multidimensional influences of the food environment on dietary behaviors. (Am J Public Health. Published online ahead of print July 16, 2015: e1-e4. doi:10.2105/AJPH.2015.302792).
Article
Urbanization's rapid progress has modernized many people's lives but also engendered big issues, such as traffic congestion, energy consumption, and pollution. Urban computing aims to tackle these issues by using the data that has been generated in cities (e.g., traffic flow, human mobility, and geographical data). Urban computing connects urban sensing, data management, data analytics, and service providing into a recurrent process for an unobtrusive and continuous improvement of people's lives, city operation systems, and the environment. Urban computing is an interdisciplinary field where computer sciences meet conventional city-related fields, like transportation, civil engineering, environment, economy, ecology, and sociology in the context of urban spaces. This article first introduces the concept of urban computing, discussing its general framework and key challenges from the perspective of computer sciences. Second, we classify the applications of urban computing into seven categories, consisting of urban planning, transportation, the environment, energy, social, economy, and public safety and security, presenting representative scenarios in each category. Third, we summarize the typical technologies that are needed in urban computing into four folds, which are about urban sensing, urban data management, knowledge fusion across heterogeneous data, and urban data visualization. Finally, we give an outlook on the future of urban computing, suggesting a few research topics that are somehow missing in the community.