Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Contents lists available at ScienceDirect
Environment International
journal homepage: www.elsevier.com/locate/envint
Quantifying the impact of daily mobility on errors in air pollution exposure
estimation using mobile phone location data
Xiaonan Yu
a
, Cesunica Ivey
b
, Zhijiong Huang
c
, Sashikanth Gurram
d
, Vijayaraghavan Sivaraman
d
,
Huizhong Shen
e
, Naveen Eluru
a
, Samiul Hasan
a
, Lucas Henneman
f
, Guoliang Shi
g
,
Hongliang Zhang
h
, Haofei Yu
a,⁎
, Junyu Zheng
c
a
Department of Civil, Environmental, and Construction Engineering, University of Central Florida, Orlando, FL, USA
b
Department of Chemical and Environmental Engineering, University of California Riverside, Riverside, CA, USA
c
Inisitute for Environmental and Climate Research, Jinan University, Guangzhou, China
d
AirSage, Atlanta, GA, USA
e
School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA
f
T.H. Chan School of Public Health, Harvard University, Cambridge, MA, USA
g
College of Environmental Science and Engineering, Nankai University, Tianjin, China
h
Department of Environmental Science and Engineering, Fudan University, Shanghai, China
ARTICLE INFO
Handling Editor: Xavier Querol
Keywords:
Air pollution exposure
Exposure misclassification
Human mobility
Cell phone location data
Call detail record
ABSTRACT
One major source of uncertainty in accurately estimating human exposure to air pollution is that human subjects
move spatiotemporally, and such mobility is usually not considered in exposure estimation. How such mobility
impacts exposure estimates at the population and individual level, particularly for subjects with different levels
of mobility, remains under-investigated. In addition, a wide range of methods have been used in the past to
develop air pollutant concentration fields for related health studies. How the choices of methods impact results
of exposure estimation, especially when detailed mobility information is considered, is still largely unknown. In
this study, by using a publicly available large cell phone location dataset containing over 35 million location
records collected from 310,989 subjects, we investigated the impact of individual subjects’ mobility on their
estimated exposures for five chosen ambient pollutants (CO, NO
2
, SO
2
, O
3
and PM
2.5
). We also estimated ex-
posures separately for 10 groups of subjects with different levels of mobility to explore how increased mobility
impacted their exposure estimates. Further, we applied and compared two methods to develop concentration
fields for exposure estimation, including one based on Community Multiscale Air Quality (CMAQ) model out-
puts, and the other based on the interpolated observed pollutant concentrations using the inverse distance
weighting (IDW) method. Our results suggest that detailed mobility information does not have a significant
influence on mean population exposure estimate in our sample population, although impacts can be substantial
at the individual level. Additionally, exposure classification error due to the use of home-location data increased
for subjects that exhibited higher levels of mobility. Omitting mobility could result in underestimation of ex-
posures to traffic-related pollutants particularly during afternoon rush-hour, and overestimate exposures to
ozone especially during mid-afternoon. Between CMAQ and IDW, we found that the IDW method generates
smooth concentration fields that were not suitable for exposure estimation with detailed mobility data.
Therefore, the method for developing air pollution concentration fields when detailed mobility data were to be
applied should be chosen carefully. Our findings have important implications for future air pollution health
studies.
1. Introduction
Exposure to air pollution is the second leading cause of non-com-
municable disease worldwide (Neira et al., 2018). It is also associated
with more than 4 million premature deaths annually (Burnett et al.,
2018; Cohen et al., 2017) and numerous other negative health con-
sequences (Gakidou et al., 2016; Kampa and Castanas, 2008; Pope and
Dockery, 2006; Bernstein et al., 2004; Kim, 2004; de Zwart et al., 2018;
https://doi.org/10.1016/j.envint.2020.105772
Received 4 September 2019; Received in revised form 24 March 2020; Accepted 26 April 2020
⁎
Corresponding author at: 12800 Pegasus Drive Suite 211, Orlando, FL 32816, USA.
E-mail address: haofei.yu@ucf.edu (H. Yu).
Environment International 141 (2020) 105772
0160-4120/ © 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/BY-NC-ND/4.0/).
T
Münzel et al., 2017). An accurate estimation of human exposure to air
pollution is critical for assessing the potential connections between air
pollution exposure and certain health outcomes, and for quantifying the
health impacts of air pollution (Zhang et al., 2018a; Fann et al., 2017;
Malley et al., 2017; Chen et al., 2018). In many prior air pollution
health studies, human exposure to air pollution was estimated using
concentration data collected or simulated at the location of subjects’
home addresses (Reis et al., 2018; Zhang et al., 2018b), or even at
further aggregated zones such as census tract (Gray et al., 2013) or ZIP
code level (Cao et al., 2011). Detailed spatiotemporal movements of
subjects, i.e. human mobility, were usually omitted due to lack of data.
This home-based exposure (herein referred to as HBE), could introduce
considerable amount of exposure classification errors (Gurram et al.,
2015; Shafran-Nathan et al., 2017; Park and Kwan, 2017; Yoo et al.,
2015; Yu et al., 2018a; Gurram et al., 2019), which could potentially
bias subsequent statistical analyses (Setton et al., 2011; Pennington
et al., 2017).
To address this issue, a variety of methods have been adopted, in-
cluding utilizing travel surveys and diaries (Gurram et al., 2015; Klepeis
et al., 2001); personal measurements (Dons et al., 2011; Buonanno
et al., 2014), accounting for multiple addresses (e.g., residential or
work address) or full-day travel data (Gurram et al., 2015; Gurram
et al., 2019) during the temporal window of exposure (Reis et al., 2018;
Setton et al., 2011; Bell et al., 2018; Chen et al., 2010), tracking subjects
using GPS-enabled surveys (Yoo et al., 2015; Nieuwenhuijsen et al.,
2015), and employing a variety of modeling tools and techniques to
account for mobility (Park and Kwan, 2017; Tang et al., 2018). Though
prior results suggest exposure estimation errors due to the omission of
mobility could differ among individuals with different mobility patterns
(Gurram et al., 2015; Gurram et al., 2019), the direction and magnitude
of such errors remains under-investigated. Further, numerous methods
have been used in the past to develop pollutant concentration fields for
air pollution health studies, and the developed fields vary substantially
spatially and temporally (Yu et al., 2018b; Ivey et al., 2015; Bates et al.,
2018). How the choices of method impact exposure estimates when
human mobility is considered is still largely unknown.
In our exploratory study (Yu et al., 2018a), we demonstrated the
feasibility of using cell phone location dataset in air pollution exposure
estimation using a relatively small sample population (n = 9,886) with
different mobility levels. Here, building upon our previous work, we: 1)
applied two methods to develop pollution concentration fields, and
investigated the impact of different methods on exposure estimates
when detailed mobility information were considered; 2) included a
substantially larger sample population (n = 310,989), divided the en-
tire population into 10 groups with varying mobility levels, and in-
vestigated how different mobility impact exposure estimates; 3) in-
vestigated the temporal variability of exposure estimates among groups
with different mobility levels; 4) investigated how exposure classifica-
tion errors change due to mobility; and 5) quantified the impact of
exposure classification errors on subsequent health effect estimations.
Details on the methods used in this study are presented in the next
section, followed by the results of the study and a discussion of the
potential of the methods and data, as well as associated limitations.
2. Material and methods
2.1. Data description and study area
The cell phone location data applied here are Call Detail Record
(CDR) data collected by mobile network operators. CDR data are col-
lected from cellphones when the phone communicates with a nearby
cell towers, specifically, when a network subscriber’s cell phone com-
municates with a nearby cell tower (such as phone call, text messaging,
or mobile data request), a suite of information is generated and ar-
chived for billing purposes (Zhao et al., 2016; Zhang et al., 2015; Zhang
et al., 2014). The archived information contains the identities of cell
towers that handle the communication, and the tower locations are
already known. CDR data contains tremendous amount of digital
footprints for virtually all subscribers of the network, and it has been
extensively used in criminal investigation (McMillan et al., 2013;
Kumar et al., 2017), the study of human mobility (Zhang et al., 2014;
Becker et al., 2013; Gonzalez et al., 2008), and urban and transporta-
tion planning (Becker et al., 2011; Wang et al., 2010; Iqbal et al, 2014).
It’s worth noting that location information contained in CDR data are
not the locations of cellphone users, rather they are the locations of
nearby cellphone tower that handled the user’s wireless communica-
tion.
In this study, we obtained a publicly available CDR dataset for
Shenzhen, China (Zhang et al., 2015; Zhang, 2020). Shenzhen is a
major city located in the Guangdong Province (Fig. 1). It has an area of
Fig. 1. The study area of Shenzhen, China.
X. Yu, et al. Environment International 141 (2020) 105772
2
1,991 km
2
and over 12 million residents, making it one of the most
populated cities worldwide. The original CDR dataset contains over 38
million location records collected from 414,271 anonymized Subscriber
Identification Module (SIM) cards on one typical weekday in October
2013. We excluded SIM cards with no location data available at night
(here defined as after 8 pm and before 7 am), which is required to infer
potential home addresses. The filtered CDR dataset applied here has
35.6 million location records for 310,989 unique SIM cards (herein
referred to as subjects), with an average of approximately 115 records
per subject per day. All identifiers contained in the original CDR data
were removed from this database, leaving only a randomized SIM card
ID, a time stamp, and latitude and longitude. This information was used
to construct daily mobility patterns for each subject.
2.2. Exposure estimation
Five pollutants were selected for this study, including carbon
monoxide (CO), nitrogen dioxide (NO
2
), sulfur dioxide (SO
2
), ground-
level ozone (O
3
), and particulate matter with aerodynamic diameter
equal or less than 2.5 µm (PM
2.5
). All of these pollutants are important
air pollutants regulated in both the United States (National Ambient Air
Quality Standards) and China (GB3095-2012), and they are considered
to pose harmful effects to human health and the environment, not only
for the US and China, but also worldwide.
Similar to our previous study (Yu et al., 2018a), we estimated all
subjects’ exposures to the five chosen pollutants using two methods: a
static, home-based exposure (HBE) calculated by assuming all subjects
stay at their corresponding home locations throughout the entire day;
and a dynamic, CDR-based exposure (CDRE) calculated by matching
detailed CDR location data with modeled pollutant concentrations at
the corresponding locations. Specifically, HBE and CDRE are estimated
as:
=
=
HBE C
n
h
n
h g
1,
=
= =
CDRE
C
k
n
h
n
m
k h m
1 1
,
where C
h,g
is pollutant concentration in hour hat the grid cell gwhere
the corresponding subject’ home is located; nis the total amount of
hours in the study period (n = 24); C
h,m
is pollutant concentration in
hour hat grid cell mwhere the subject is located within the corre-
sponding hour. The subject may be located in k(k > = 1) grid cells in
hour h. In the static method, each subject’s home location was assumed
to be their most frequent location at night (between 8 pm and 7 am),
and we used modeled pollutant concentration data at their corre-
sponding home location to estimate their exposures. In the dynamic
method, the CDRE was estimated by arithmetically weighting con-
centrations at different locations where the subject visited based on the
time (in hours) the subject spent at each location. If no location data
was available for one specific hour, we assumed the subject stayed at
the same location as in the previous hour. If location data was missing
for the first hour (12 am – 1 am), the subject was assumed to be at their
estimated home locations. For hours with multiple location records
available, we used averaged concentration from all locations in the
corresponding hour. We estimated HBE and CDRE for each subject se-
parately.
Different from our previous study (Yu et al., 2018a), we applied two
approaches to develop spatiotemporal concentration fields of the five
chosen pollutants: one based on outputs from the Community Multi-
scale Air Quality (CMAQ) model (Byun and Schere, 2006) for the cor-
responding day, and the other using the Inverse Distance Weighting
(IDW) method. Detailed information on CMAQ model configurations is
available elsewhere (Che et al., 2011). To correct for potential model
biases and errors, we fused hourly measurement data collected from 12
monitoring stations inside the CMAQ modeling domain (Fig. 1) into
CMAQ output by multiplying gridded hourly CMAQ fields with ad-
justment factors. The factors were calculated as the ratio between
measured and modeled concentrations at the locations of each mon-
itoring station, and then spatially interpolated to the center points of all
CMAQ grid cells using kriging (Yu et al., 2018b). For the IDW method,
we spatially interpolated hourly measurements from all monitoring
stations inside the study area using inversed and squared distance as the
weight. The spatial and temporal resolution of the concentration fields
for both methods are 3 km and 1 hour, respectively. We acknowledge
that an individual’s exposure to air pollution occur at finer scales, we
nonetheless still applied the aforementioned CMAQ and IDW fields
mainly for two reasons: 1) Developing higher resolution pollution fields
are not feasible in this study due to the limited availability of mea-
surement data in the study area (Fig. 1), and computational burden
involved in running higher resolution CMAQ simulations; and 2) the
location information in CDR are the locations of cellphone towers close
to the corresponding cellphone user. In addition, it’s important to note
that the aforementioned CMAQ and IDW methods are fundamentally
different, and the results of exposure assessment are expected to be
impacted substantially by the choice of methods.
To understand how different degrees of mobility impact exposure
estimation, we further subdivided all subjects into 10 groups based on
the number of unique CMAQ grid cells each individual subject visited
during the day. The number of grid cells each subject visited in group 1
through 9 correspond to their respective group number, while all sub-
jects that visited 10 or more unique grid cells were collectively assigned
into group 10. Subjects in groups with larger group numbers are ex-
pected to have a high degree of mobility. We estimated HBE and CDRE
separately for all 10 groups. While metrics, such as distance between
home and work location (Setton et al., 2011), have been used in past
studies, however, such information is not available in this study.
In epidemiological studies related to air pollution, subjects are fre-
quently assigned to different groups based on their exposure levels
(such as quartiles) (Chen et al., 2010; Clark et al., 2009; Dugandzic
et al., 2006; Mitchell and Popham, 2008; Gauderman et al., 2007).
Statistical comparisons are then performed among these groups to in-
vestigate whether higher exposure levels are associated with a higher
incidence of certain health outcomes. The statistical analysis could be
biased or confounded if subjects were misclassified into the wrong ex-
posure group. To explore the impact of including detailed mobility data
on exposure misclassification, we compared how subjects were assigned
to four quartiles based on their CDRE and HBE. We define “mis-
classification” as the assignment of one subject, based on HBE, into a
quartile that is different from CDRE-based quartile.
We performed the Wilcoxon rank sum test to examine whether the
medians of CDRE and HBE exposure estimates are statistically different.
We chose this test because the samples in this study are not normally
distributed. Furthermore, we also calculated the expected bias factors
to quantify potential biases in relative risk estimates when HBE was
used (Setton et al., 2011; Nyhan et al., 2018). According to the classical
error theory, exposure estimated using the home-based method may be
expressed as:
= +Z X E
(1)
In Eq. (1),Zis exposure estimated using HBE; Xis the true exposure
value; and Eis the error associated with the corresponding HBE. In this
study, we use CDRE to represent X, and, based on our previous results, E
is correlated with X(Yu et al., 2018a). Therefore, the following equa-
tion can be applied to calculate a bias factor (Setton et al., 2011; Nyhan
et al., 2018; Wacholder, 1995):
=
+
+ +
B
2
2
2 2
(2)
In Eq. (2),Bis the calculated bias factor; σ
2
is the variance of CDRE
of all subjects; φis the covariance between CDRE and errors in exposure
X. Yu, et al. Environment International 141 (2020) 105772
3
estimation (calculated based on HBE-CDRE); and ω
2
is the variance of
the errors in exposure estimation. The factor Brepresents the expected
bias in relative risk estimates when the home-based method is applied.
For example, a B factor of 0.75 suggests that applying the home-based
method would lead to the relative risk being underestimated by 25%.
It’s also worth noting that the Wilcoxon rank sum test is a different
statistical measure compared to the coefficient of determination (R
2
).
The former intends to test equality, while the latter quantifies the
proportion of variance contained in the dependent variable that can be
predicted by the independent variable.
3. Results
3.1. Concentration fields
The spatial concentration fields of the five chosen pollutants simu-
lated by the CMAQ and IDW methods differ considerably (Fig. 2),
especially for O
3
, NO
2
, and PM
2.5
, where the latter two pollutants are
known to have substantial primary contributions from transportation
sectors. Due to the sparseness of monitor network, the IDW method
generally results in smoother fields that lack spatial variabilities
compared with the CMAQ method. The locations of monitoring stations
can also be observed on the concentration fields as simulated by the
IDW method (Fig. S1).
3.2. Overall correlations between HBE and CDRE
Mean CMAQ-based HBE and CDRE estimates for all subjects were
highly correlated with each other (Fig. 3). The coefficient of determi-
nation (R
2
) ranged from 0.95 (NO
2
) to 0.98 (SO
2
), with the slopes of
linear regression close to 1, and intercepts were close to 0 for all pol-
lutants. The estimated regression parameters are considerably different
comparing with our previous study (Yu et al., 2018a) (e.g: R
2
ranged
between 0.65 and 0.76 in the previous study). We also observed many
vertically aligned data points, suggesting many subjects had identical
HBE but their CDRE was considerably different when individual mo-
bility was considered. Additionally, a large number of data points were
clustered near the 1:1 line, suggesting that a substantial portion of the
subjects had similar HBE and CDRE.
Similar findings were also observed for IDW-based exposures
(Fig. 3), including the clustered data points along the 1:1 line, the high
overall correlations between HBE and CDRE, and the varying CDRE
Fig. 2. Spatial fields of concentrations of the five chosen pollutants as simulated by the CMAQ (a-e) and IDW (f-j) methods.
X. Yu, et al. Environment International 141 (2020) 105772
4
estimates for many subjects with identical HBE estimates. However, the
range of estimates for both HBE and CDRE were much smaller for the
IDW exposures, particularly for NO
2
, O
3
and PM
2.5
, where the vast
majority of data points were clustered within small concentration
ranges. It’s also worth noting that results of Wilcoxon rank sum tests
show HBE and CDRE are overall statistically different for all pollutants.
3.3. The impact of mobility on exposure estimates
We found that the correlations between HBE and CDRE estimates
shrink with an increased degree of mobility (NO
2
presented in Table 1,
other pollutants in Tables S2 through S5). Compared with CMAQ, the
decreasing correlations between CDRE and HBE were smaller when
IDW fields were used, with considerably smaller RMSE, MNB and MNE.
For PM
2.5
, as shown by the numbers presented in Table S5, the RMSE,
MNB and MNE for the group with the highest degree of mobility (group
10) was only 5.4%, 6.7%, and 4.6%, respectively, of those when CMAQ
fields were used. For example, the MNE for group 10 is 3.23% when
CMAQ fields were used, but only 0.15% when IDW fields were used.
The only exception is SO
2
(Table S3), for which the RMSE and MNE
changed similarly between the CMAQ and IDW methods, though MNB
is only 0.9% when the IDW method was applied.
In this dataset, over half (54%) of all subjects stayed in the same
3 km grid cell throughout the entire day, and the majority (94%) of all
subjects visited 4 or fewer grid cells (Table 1). Although subjects that
were highly mobile (especially those who visited 6 and more grid cells)
accounted for a relatively small fraction of the entire population, the
sample sizes of all groups were still considerable due to the large overall
sample population (sample size = 916 for the smallest group, group 9).
The impacts of mobility on exposure estimates differ by pollutant
and by concentration fields used. Between CMAQ and IDW methods,
the range of variability was considerably smaller when the IDW method
was applied, particularly for NO
2
, O
3
and PM
2.5
. SO
2
again was the
exception where exposure variability was similar between the two
methods. Mobility had the greatest impact for NO
2
and O
3
. When
CMAQ concentration fields were applied, the observed differences were
more negative (higher CDRE than HBE) for CO, NO
2
and PM
2.5
, but
were more positive (lower CDRE than HBE) for O
3
. Such observations
are not clearly visible when the IDW concentration fields were applied.
The impacts of mobility on exposures also differed by time of the
day (Fig. 4), with larger differences found during daytime for all
groups, though the biggest difference occurred at different hours for
different pollutants. When CMAQ concentration fields were applied,
CO, NO
2
and PM
2.5
exhibited the largest differences near the afternoon
rush hour, though these differences dissipates quickly thereafter. For
O
3
, the largest differences occurred around mid-afternoon at 4 pm
around when the highest ambient O
3
concentrations are expected. For
SO
2
, we observed a slight peak in differences between HBE and CDRE at
around 10 am. Additionally, the observed differences were mostly ne-
gative during daytime for CO, NO
2
and PM
2.5
, suggesting the home-
based method resulted in lower exposure estimates, although the dif-
ferences changed to slightly positive toward mid-night. However, the
exposure differences are mostly positive for O
3
, indicating higher ex-
posure estimates when the home-based method is used. When CMAQ
Fig. 3. Linear correlations between HBE and CDRE estimates of the five chosen pollutants for all subjects based on CMAQ (a,c,e,g,i) and IDW (b,d,f,h,j) concentration
fields. Pixels are color coded by sample size. The solid black line shown is the 1:1 line.
X. Yu, et al. Environment International 141 (2020) 105772
5
concentration fields were applied, the biggest exposure differences
were not observed for the group with the highest mobility (group 10),
rather it was observed for subjects with moderate to high degree of
mobility (group 7 for SO
2
, and group 5 and 6 for other pollutants).
The temporal variations of exposure differences, however, were
mostly not observed when IDW concentration fields were applied
(Fig. 4). We still observed generally larger differences during daytime
(though smaller magnitude), but the consistent patterns of fluctuations
as seen among CO, NO
2
and PM
2.5
in Fig. 4 were not observed when
IDW fields were applied. The biggest differences were observed at dif-
ferent hours for different pollutants and with no consistent directions.
Exposure differences generally showed a consistent increasing trend
with increased mobility.
We performed Wilcoxon rank sum tests to evaluate the differences
between HBE and CDRE estimates for each mobility group. When
CMAQ concentration fields were applied, most differences in HBE and
CDRE estimates were statistically significant (p < 0.05) during normal
business hours (9 am to 5 pm). The only exception is SO
2
, for which
HBE and CDRE estimates are statistically different between 1 pm and
10 pm. When IDW concentration fields were applied, HBE and CDRE
estimates are still generally statistically different between 10 am to
5 pm, although with considerably greater variability.
3.4. The impact of mobility on exposure classifications and effect estimates
To investigate potential exposure misclassifications associated with
Table 1
Comparison between HBE and CDRE estimate of NO
2
for all ten groups with different mobility.
Group number
1 2 3 4 5 6 7 8 9 10
CMAQ CDRE mean (ppbv) 16.1 16.6 16.7 16.8 16.7 16.3 15.9 15.9 15.6 15.6
HBE mean (ppbv) 16.1 16.5 16.3 16.2 15.8 15.5 15.2 15.2 15.0 15.1
a
RMSE (ppbv) 0.00 1.16 1.79 2.16 2.50 2.60 2.62 2.74 2.78 3.02
b
MNB (%) 0.0% −0.8% −2.3% −3.8% −5.0% −4.9% −4.3% −4.1% −3.5% −2.8%
c
MNE (%) 0.0% 3.6% 6.2% 8.1% 9.8% 10.5% 10.6% 10.8% 11.2% 11.9%
d
R
2
1.00 0.95 0.88 0.83 0.76 0.72 0.70 0.67 0.66 0.64
IDW CDRE mean (ppbv) 19.4 19.2 19.3 19.3 19.3 19.2 19.1 19.1 19.0 19.0
HBE mean (ppbv) 19.4 19.2 19.3 19.3 19.3 19.2 19.1 19.1 19.0 19.0
a
RMSE 0.00 0.23 0.35 0.43 0.49 0.56 0.62 0.62 0.67 0.72
b
MNB (%) 0.0% 0.0% −0.1% −0.1% −0.2% −0.1% 0.0% 0.0% 0.2% 0.4%
c
MNE (%) 0.0% 0.4% 0.8% 1.1% 1.4% 1.7% 1.9% 2.0% 2.3% 2.4%
d
R
2
1.00 0.98 0.94 0.92 0.88 0.85 0.81 0.81 0.78 0.75
Sample size 167,570 75,313 32,177 16,350 8354 4617 2700 1562 916 1430
a
RMSE: root mean squared error. Calculated as
=
HBE CDRE[ ( ) ]
Ni
N
i i
1
1
2 1/2
, where CDRE and HBE is the estimated exposures based on CDR and home-based
method for the ith subject.
b
MNB: mean normalized bias. Calculated as
=( )
Ni
NHBEiCDREi
CDREi
1
1
.
c
MNE: mean normalized error. Calculated as
=
Ni
NHBEiCDREi
CDREi
1
1
.
d
R
2
: coefficient of determination between HBE and CDRE estimates in the corresponding group.
Fig. 4. Temporal variations of exposure differences for all 10 mobility groups between HBE and CDRE when CMAQ and IDW concentration field were applied.
Exposure differences were calculated as HBE-CDRE.
X. Yu, et al. Environment International 141 (2020) 105772
6
omitting subject mobility, we investigated how subjects were assigned
to different quartiles based on their HBE and CDRE estimates. Results
for PM
2.5
are presented in Figs. 5 and 6, and results for other pollutants
are presented in Figs. S2-S9.
We observed that a high percentage of the sample population was
potentially misclassified into other quartiles, especially for groups with
higher degrees of mobility. When CMAQ concentration fields were
applied for PM
2.5
(Fig. 5), more than half of the sample population in
the middle quartiles (Q2 and Q3) were classified into different quartiles
for groups 4 through 10 when individual mobility was omitted. The
misclassification is especially prominent for the 2nd quartile of group 6
(Fig. 5), for which 71% of subjects were misclassified into other quar-
tiles when the home-based method was used. This finding was also
observed when IDW fields were used, although the potential mis-
classifications were less severe, but still substantial (Fig. 6). Similar
findings can be observed for both CMAQ and IDW concentration fields
for all other pollutants (Figs. S2-S9). For subjects with moderate ex-
posure levels (Q2 and Q3), generally more subjects were assigned to
quartiles with higher exposures when the home-based method was used
for CO (Figs. S2, S6) and NO
2
(Figs. S3, S7). This result was less con-
sistent for SO
2
(Figs. S4, S8) and somewhat reversed for O
3
(Figs. S5,
S9).
The estimated bias factors for groups with different mobility levels
are presented in Fig. 7. With increased mobility, the estimated bias
factors generally decrease regardless of concentration fields used. The
smaller bias factor, a value of 0.67, is observed for NO
2
and for group
10. This value suggests that the estimated relative risk for NO
2
will be
underestimated by 33% when mobility was ignored during exposure
estimation. Between CMAQ and IDW, the estimated bias factors are
relatively similar for NO
2
, but are considerably different for other
pollutants, especially for PM
2.5
. For group 10, the bias for PM
2.5
is 0.70
when CMAQ fields are used, and 0.94 when IDW fields are used.
Fig. 5. The directions of potential PM
2.5
exposure misclassifications when the home-based exposure estimation method was used and when CMAQ fields were used.
For simplification purposes only results for groups 2, 6 and 10 are presented. Subjects in quartile 1 has the lowest exposures, and subjects in quartile 4 has the highest
exposures.
Fig. 6. The directions of potential PM
2.5
exposure misclassifications when the home-based exposure estimation method was used and when IDW fields were used. For
simplification purposes only results for groups 2, 6 and 10 are presented. Subjects in quartile 1 has the lowest exposures, and subjects in quartile 4 has the highest
exposures.
X. Yu, et al. Environment International 141 (2020) 105772
7
4. Discussion
4.1. The impact of method choices on exposure estimation
An appropriate characterization of spatial concentration distribu-
tions of air pollutants is fundamental for air pollution exposure esti-
mation. In this study, we applied two methods to develop air pollutant
concentration fields: one based on outputs from the CMAQ model, and
the other based on the IDW interpolation method. Spatial concentration
fields developed using the two methods were considerably different
from each other (Fig. 2). This is expected because, as described pre-
viously, the two methods are fundamentally different, and both
methods have their own strengths and weaknesses (Yu et al., 2018b).
Consequently, the estimated population average exposures (Table 1),
the distributions of individual exposure estimates (Fig. 3), particularly
among groups with different degrees of mobility (Fig. 4), and the im-
pact of neglecting mobility on exposure estimates (Figs. 5–6), was dif-
ferent between the two methods. Such results were expected due to the
different nature of the two methods. CMAQ is a mechanistic model that
calculates ambient concentrations of air pollutants based on input
emissions and meteorological data. IDW is an empirical spatial inter-
polation method that relies solely on available pollutant concentrations
measured at discrete locations (Yu et al., 2018b). Pollution hotspots
that are not captured by monitoring networks cannot be captured by
the IDW method but may possibly be captured by the CMAQ model if
appropriate emissions data are supplied. In this study, the monitoring
network is sparse, and only 1 out of 12 monitor is located inside
Shenzhen area (Fig. 1). As a result, pollutant concentration fields de-
veloped using the IDW method were smooth and lacked the spatial
concentration variabilities as observed in the CMAQ fields. Therefore,
it’s important to carefully select an appropriate method for developing
pollutant concentration fields, particularly when the monitoring net-
work is sparse.
When detailed mobility data were included, naturally, the appro-
priate characterization of spatial pollutant variability became even
more important. In such applications, purely spatial interpolation
methods, e.g., IDW, tessellation, or kriging, are also not ideal choices
for developing pollutant concentration fields for study regions without
an extensive monitoring network available (Yu et al., 2018b). These
results highlighted the importance of choosing an appropriate method
for developing pollutant concentration fields for exposure estimation
purposes, particularly when detailed mobility data were included.
Without an appropriate characterization of spatial pollutant
concentration variations, exposure assessment may not significantly
benefit from the inclusion of detailed mobility data at urban scale.
Subsequently, we will focus our discussion on results as obtained
using the CMAQ concentration fields.
4.2. The impact of mobility on exposure estimation
In this study, the estimated regression parameters are considerably
different from our previous study (Yu et al., 2018a). For example, the
estimated R
2
ranged between 0.95 and 0.98 vs 0.65 to 0.76 in the
previous study; and the slope ranged between 0.97 and 1.02 vs 0.60 to
0.72 previously. The seemingly contradictory findings can be explained
by the difference in sample population. In our previous study, 9,886
subjects with the most amount of CDR data available were selected to
explore the potential benefits of using CDR data in exposure estimation.
The subjects were not randomly sampled, and with an average of ap-
proximately 463 records per subject per day (vs 115 records per subject
per day in this study). The sample population in our previous study are
relatively more mobile, and the subjects visited on average 2.3 grid
cells over the study period (vs 1.9 grid cells in this study).
At the population level, we did not find substantial differences be-
tween HBE and CDRE exposures, consistent with our previous study (Yu
et al., 2018a) and other studies (Nyhan et al., 2018; Dewulf et al., 2016;
Picornell et al., 2018; Nyhan et al., 2016; Gariazzo et al., 2016). The
finding maybe partially explained by the fact that most subjects spent
most of their time within the same grid, as indicated by the large
number of data points clustered near the 1:1 line (Fig. 3). Our results
suggested that the home-based method for exposure estimation is still
informative in the study region when only average exposure estimates
for a sufficiently large population are of interest (Nikkilä et al., 2018).
However, it’s worth noting that several studies conducted in other cities
(Singh et al., 2019; Smith et al., 2016) have found that the population
level exposure estimates are lower when individual mobility data were
included in exposure estimation. The differences in findings may be
partially due to the potentially different population mobility patterns
among cities. Further studies are needed to investigate how our findings
may vary among cities.
One of the main focus of this manuscript is on how different levels
of mobility impact air pollution exposures. We found that the impact of
mobility on exposure estimates differed by time of day and by pollu-
tants (such analyses were not performed in our previous study, Yu et al.,
2018a). Generally, the differences between HBE and CDRE were the
smallest during early morning and midnight, a time when many
Fig. 7. The impact of mobility on bias factors when CMAQ and IDW concentration fields were applied.
X. Yu, et al. Environment International 141 (2020) 105772
8
subjects are expected to be at home. For traffic-related pollutants in-
cluding CO, NO
2
, and PM
2.5
, we found that the home-based method
likely underestimated subject exposures during daytime, especially near
afternoon rush hour, when CMAQ concentration fields were used
(Fig. 4). Meanwhile, subject exposures to ozone may be over-estimated
during daytime using HBE, with the highest error observed at around
4 pm, near the time when the highest ambient ozone concentrations are
expected (Fig. 4). The temporal differences in impacts of mobility on
exposure have also been noted previously (Picornell et al., 2018). In-
terestingly, during peak hours, the most significant differences between
HBE and CDRE were not observed for the group with the highest degree
of mobility, rather the largest differences were observed on subjects
with moderate to high degree of mobility (groups 5–7).
Our results showed that the impact of mobility on exposure could be
substantial at the individual level, particularly for subjects that are
highly mobile. Applying the home-based method yielded similar esti-
mates for those who live close to where they travel throughout the day,
although their actual exposure could be drastically different when in-
dividual mobility is considered. With an increased degree of mobility,
we found that the correlations between HBE and CDRE decreased
monotonically (Table 1), suggesting that the home-based method cap-
tured less exposure variability among individuals with increased mo-
bility (Chen et al., 2010). Therefore, we expect larger exposure classi-
fication errors for subjects that are highly mobile, which is supported by
our analysis on the potential exposure misclassifications based on HBE
and CDRE (Figs. 5–6). It is also worth mentioning again that 71% of
subjects (Fig. 5) in the second quartile of group 6 were misclassified
into different quartiles using HBE. These results suggest that the impact
of traffic-related pollutants on human health may be larger than pre-
viously documented, and these findings may have significant implica-
tions for studies that rely on air pollution exposure estimation.
We found that ignoring mobility in exposure assessment could lead
to up to 33% in underestimation of relative risk, though the magnitude
of underestimation differs among pollutants (Fig. 7). Between CMAQ
and IDW, the results are also different, especially for PM
2.5
, for which
the largest estimated bias factor is only 0.94 when the IDW fields were
applied (vs 0.70 for CMAQ field). These finding again demonstrated
that the benefit of including detailed mobility data in exposure as-
sessment may be reduced when the spatial variability of pollutant
concentrations were not captured, and the method for developing
pollution field need to be selected carefully when mobility data were to
be included. The finding also have implications for future air pollution
health studies.
4.3. Limitations
There are inherent limitations associated with this study. First, as
with many CDR databases, the location data used in this study are not
the exact location of the corresponding cell phone user, rather, they are
the locations of the cell phone tower that handled the wireless com-
munication, which are most likely the nearest tower to the cell phone
user. However, we do not expect this limitation to substantially impact
the findings for two reasons. 1) The study area is one of the most po-
pulated cities in the world with a well-known, densely distributed cell
tower network. The CDR dataset contains over 1,000 locations of cell
phone towers spread out across the study area. 2) We applied 3-km
resolution concentration fields in exposure estimation. The retrieved
concentration values are identical within one 3-km grid cell, and one
cell phone user in Shenzhen is highly likely to have at least one cell
tower within 3 km (see https://www.opencellid.org for more in-
formation on cell tower coverage in Shenzhen, China). Therefore, we do
not expect the findings to change considerably even when the exact
locations of all cell phone users are applied.
Second, CDR data comprise an “event-triggered” database. Location
data are only collected when a cell phone communicates with nearby
towers. Hence, CDR are temporally sparse in nature (Zhao et al., 2016),
and may not accurately capture the full spectrum of individual move-
ments, especially for individuals who only use cell phones occasionally.
Hence, exposures estimated using CDR may deviate from those esti-
mated using a more complete location dataset such as those collected
using dedicated applications (e.g. Dynamica (Fan et al., 2015), or other
momentarily collected data such as Google Maps Location History data
(Yu et al., 2019). However, in this study, our purpose is to compare the
differences between exposure estimates with and without detailed
mobility data applied. Given the large sample population in all 10
groups with different degrees of mobility, we do not expect the results
to change even with an ideally complete mobility database.
Third, despite the relatively large population (N = 310,989) and
number of location records (35.6 million), the CDR data used here are a
randomly sampled subset from all cell phone users within the entire city
of Shenzhen for one typical work day within a typical week. Therefore,
the spatiotemporal mobility patterns as represented in this CDR data-
base represent the unique characteristics of the study region. We do
expect the patterns of population mobility, the spatiotemporally
variability of air pollution concentrations, pollutant emissions, and
meteorology conditions to vary across different cities. Further studies
are needed to better understand how the findings from this study may
change in another city.
Fourth, as described previously, due to the nature of CDR data, the
availability of observations, and resources constrains, we applied air
pollution concentration fields with 3 km spatial resolution and 1 h
temporal resolution for estimating pollution exposures. We recognize
that such coarse resolution may introduce uncertainties into related
analyses and may also partially impact the findings, such as the impact
of mobility on population-level exposure estimates (Fig. 3) (Singh et al.,
2019; Smith et al., 2016). Here, we performed an additional analysis to
explore the impact of grid resolution on the classification of mobility
levels. We split all 3 km CMAQ grid cells into 1.5 km grid cells and
counted the number of unique grid cells each subject visited (Table 2).
With increased grid resolution, a considerably higher fraction of po-
pulation were assigned to higher mobility groups, especially for groups
with the highest mobility levels (Groups 6 through 10). Such result
exemplifies the need for fine-scale modeling, and further studies are
needed to investigate how grid resolution impacts the results of ex-
posure estimation with detailed mobility data. In addition, both CDR
data and pollution fields are expected to contain uncertainties. What
dataset contain greater amount of uncertainty remain unclear. Further
studies are also needed to determine the impact of uncertainties on
exposure outcomes.
Finally, the exposure estimates presented in this study are calcu-
lated using ambient pollutant concentrations. A subject’s exposure to
indoor pollution was not considered here. Estimating indoor pollution
exposure would require expanded datasets (e.g., type of micro-en-
vironments) and models of pollution infiltration to indoor). In addition,
due to the nature of CDR data, it is difficult to precisely determine the
location of micro-environment for each subject. For example, if one
subject’s CDR data is located in close proximity to a major roadway, the
Table 2
Subject population in each mobility group at 3 km and 1.5 km grid resolutions.
3 km grids 1.5 km grids Change (%)
Group 1 167,570 132,847 −20.7%
Group 2 75,313 72,821 −3.3%
Group 3 32,177 39,341 22.3%
Group 4 16,350 22,689 38.8%
Group 5 8354 13,918 66.6%
Group 6 4617 8845 91.6%
Group 7 2700 5886 118%
Group 8 1562 4105 163%
Group 9 916 2755 201%
Group 10 1430 7782 444%
X. Yu, et al. Environment International 141 (2020) 105772
9
investigator may not be able to determine whether the subject is driving
on the roadway, or walking along the roadway, or even sitting inside a
building next to the roadway.
5. Conclusion
In this study, we applied a large cell phone location database con-
sisting of over 35 million location records from 310,989 subjects to
investigate the impact of individual mobility on estimated ambient
exposures for five chosen pollutants (CO, NO
2
, SO
2
, O
3
, and PM
2.5
). We
further divided our sample population into ten groups with different
degrees of mobility and compared exposures estimates for each group.
We also applied and compared two methods to develop concentration
fields for exposure estimation, including one based on output from the
CMAQ model that was fused with observational data, and the other
based on the spatial interpolation of observations using the inverse
distance weighting method.
We found no substantial differences between population-averaged
exposures as estimated with and without detailed mobility data (e.g.:
exposure estimates differ by up to 5.4% for NO
2
,Table 1). Thus, the
traditional home-based exposure estimation method is still informative
when only averaged exposures on a large population are needed. We
observed generally increased variabilities in exposure estimates at the
individual level with increased mobility. Exposure classification errors
are also likely to increase with higher degrees of mobility, and could be
substantial for groups of individuals that are highly mobile. We also
examined the temporal variability of the differences between exposures
as estimated with and without mobility data. We found the home-based
method will likely under-estimate exposure to traffic-related pollutants
(CO, NO
2
and PM
2.5
) during day-time particularly during afternoon
rush-hour, but also will likely over-estimate exposures to ground level
ozone during mid-afternoon near the time when ambient ozone con-
centrations are expected to be the highest. These results suggest that
mobility could be important for air pollution health studies for which
obtaining accurate exposure estimates at individual level are critical,
such as case-control studies or studies with a small sample size.
We found that the concentration fields developed using the IDW
method failed to capture pollution hotspot as can be seen from the
CMAQ fields, due primarily to the sparse monitoring network, and
consequently limited measurement data available in the study domain.
Therefore, the IDW method may not suitable for air pollution exposure
estimations when detailed mobility data are considered, if a dense
measurement network is not available. When detailed mobility data
were to be applied in exposure estimation, the method for developing
air pollution concentration fields should be selected carefully.
We also acknowledge that the CDR data applied in this study re-
present the unique characteristics of the study region, and further stu-
dies are needed to investigate how our findings could change among
regions with different spatiotemporal patterns of population and pol-
lution concentrations. Despite the limitation, overall, our results have
significant implications for future air pollution health studies in which
subject mobility is important.
CRediT authorship contribution statement
Xiaonan Yu: Formal analysis, Software, Writing - review & editing.
Cesunica Ivey: Methodology, Writing - review & editing. Zhijiong
Huang: Data curation, Writing - review & editing. Sashikanth
Gurram: Conceptualization, Writing - review & editing.
Vijayaraghavan Sivaraman: Conceptualization, Writing - review &
editing. Huizhong Shen: Conceptualization, Writing - review & editing.
Naveen Eluru: Writing - review & editing. Samiul Hasan: Writing -
review & editing. Lucas Henneman: Methodology, Writing - review &
editing. Guoliang Shi: Conceptualization, Writing - review & editing.
Hongliang Zhang: Conceptualization, Writing - review & editing.
Haofei Yu: Methodology, Software, Formal analysis, Writing - original
draft, Writing - review & editing. Junyu Zheng: Data curation.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influ-
ence the work reported in this paper.
Acknowledgements
We would like to acknowledge Dr. Desheng Zhang (Rutgers
University) for providing the CDR data. This research was partially
funded by the University of Central Florida startup grant. This material
is based upon work supported by the National Science Foundation
under Grant No. 1931871. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the author(s)
and do not necessarily reflect the views of the National Science
Foundation.
Appendix A. Supplementary material
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.envint.2020.105772.
References
Bates, J.T., et al., 2018. Source impact modeling of spatiotemporal trends in PM2. 5
oxidative potential across the eastern United States. Atmos. Environ. 193, 158–167.
Becker, R.A., et al., 2011. A tale of one city: Using cellular network data for urban
planning. IEEE Pervasive Comput. 10 (4), 18–26.
Becker, R., et al., 2013. Human mobility characterization from cellular network data.
Commun. ACM 56 (1), 74–82.
Bell, M.L., Banerjee, G., Pereira, G., 2018. Residential mobility of pregnant women and
implications for assessment of spatially-varying environmental exposures. J. Eposure
Sci. Environ. Epidemiol. 1.
Bernstein, J.A., et al., 2004. Health effects of air pollution. J. Allergy Clin. Immunol. 114
(5), 1116–1123.
Buonanno, G., Stabile, L., Morawska, L., 2014. Personal exposure to ultrafine particles:
the influence of time-activity patterns. Sci. Total Environ. 468, 903–907.
Burnett, R., et al., 2018. Global estimates of mortality associated with long-term exposure
to outdoor fine particulate matter. Proc. Natl. Acad. Sci. 115 (38), 9592–9597.
Byun, D., Schere, K.L., 2006. Review of the governing equations, computational algo-
rithms, and other components of the Models-3 Community Multiscale Air Quality
(CMAQ) modeling system. Appl. Mech. Rev. 59 (2), 51–77.
Cao, J., et al., 2011. Association between long-term exposure to outdoor air pollution and
mortality in China: a cohort study. J. Hazard. Mater. 186 (2–3), 1594–1600.
Che, W., et al., 2011. Assessment of motor vehicle emission control policies using Model-
3/CMAQ model for the Pearl River Delta region, China. Atmospheric Environ. 45 (9),
1740–1751.
Chen, L., et al., 2010. Residential mobility during pregnancy and the potential for am-
bient air pollution exposure misclassification. Environ. Res. 110 (2), 162–168.
Chen, R., et al., 2018. Fine particulate air pollution and the expression of micrornas and
circulating cytokines relevant to inflammation, coagulation, and vasoconstriction.
Environ. Health Perspect. 126 (1) 017007–017007.
Clark, N.A., et al., 2009. Effect of early life exposure to air pollution on development of
childhood asthma. Environ. Health Perspect. 118 (2), 284–290.
Cohen, A.J., et al., 2017. Estimates and 25-year trends of the global burden of disease
attributable to ambient air pollution: an analysis of data from the Global Burden of
Diseases Study 2015. The Lancet 389 (10082), 1907–1918.
de Zwart, F., et al., 2018. Air pollution and performance-based physical functioning in
Dutch older adults. Environ. Health Perspect. (Online) 126 (1 %@ 1552-9924).
Dewulf, B., et al., 2016. Dynamic assessment of exposure to air pollution using mobile
phone data. Int. J. Health Geographics 15 (1), 14.
Dons, E., et al., 2011. Impact of time–activity patterns on personal exposure to black
carbon. Atmos. Environ. 45 (21), 3594–3602.
Dugandzic, R., et al., 2006. The association between low level exposures to ambient air
pollution and term low birth weight: a retrospective cohort study. Environ. Health 5
(1), 3.
Fan Y., et al., 2015. SmarTrAC: A Smartphone Solution for Context-Aware Travel and
Activity Capturing. University of Minnesota: Minneapolis, MN.
Fann, N., et al., 2017. Estimated Changes in Life Expectancy and Adult Mortality
Resulting from Declining PM 2.5 Exposures in the Contiguous United States:
1980–2010. Environ. Health Perspect. 97003, 1.
Gakidou E., et al., 2016. Global, regional, and national comparative risk assessment of 84
behavioural, environmental and occupational, and metabolic risks or clusters of risks,
1990–2016: a systematic analysis for the Global Burden of Disease Study
2016. The Lancet. 390(10100), 1345–1422.
X. Yu, et al. Environment International 141 (2020) 105772
10
Gariazzo, C., Pelliccioni, A., Bolignano, A., 2016. A dynamic urban air pollution popu-
lation exposure assessment study using model and population density data derived by
mobile phone traffic. Atmos. Environ. 131, 289–300.
Gauderman, W.J., et al., 2007. Effect of exposure to traffic on lung development from 10
to 18 years of age: a cohort study. The Lancet 369 (9561), 571–577.
Gonzalez, M.C., Hidalgo, C.A., Barabasi, A.-L., 2008. Understanding individual human
mobility patterns. Nature 453 (7196), 779.
Gray, S.C., Edwards, S.E., Miranda, M.L., 2013. Race, socioeconomic status, and air
pollution exposure in North Carolina. Environ. Res. 126, 152–158.
Gurram, S., Stuart, A.L., Pinjari, A.R., 2015. Impacts of travel activity and urbanicity on
exposures to ambient oxides of nitrogen and on exposure disparities. Air Qual. Atmos.
Health 8 (1), 97–114.
Gurram, S., Stuart, A.L., Pinjari, A.R., 2019. Agent-based modeling to estimate exposures
to urban air pollution from transportation: Exposure disparities and impacts of high-
resolution data. Comput. Environ. Urban Syst. 75, 22–34.
Iqbal, M.S., et al., 2014. Development of origin–destination matrices using mobile phone
call data. Transport. Res. Part C: Emerg. Technol. 40, 63–74.
Ivey, C.E., et al., 2015. Development of PM < sub > 2.5 < /sub > source impact spatial
fields using a hybrid source apportionment air quality model. Geosci. Model Dev. 8
(7), 2153–2165.
Kampa, M., Castanas, E., 2008. Human health effects of air pollution. Environ. Pollut. 151
(2), 362–367.
Kim, J., 2004. Ambient air pollution: health hazards to children. Pediatrics 114 (6),
1699–1707.
Klepeis, N.E., et al., 2001. The National Human Activity Pattern Survey (NHAPS): a re-
source for assessing exposure to environmental pollutants. J. Eposure Sci. Environ.
Epidemiol. 11 (3), 231.
Kumar, M., Hanumanthappa, M., Kumar, T.S., 2017. Crime investigation and criminal
network analysis using archive call detail records. In: Advanced Computing (ICoAC),
2016 Eighth International Conference on. IEEE.
Malley, C.S., et al., 2017. Preterm birth associated with maternal fine particulate matter
exposure: a global, regional and national assessment. Environ. Int. 101, 173–182.
McMillan, J.E.R., W.B. Glisson, Bromby, M. 2013. Investigating the increase in mobile
phone evidence in criminal activities. In: System sciences (hicss), 2013 46th hawaii
international conference on. IEEE.
Mitchell, R., Popham, F., 2008. Effect of exposure to natural environment on health in-
equalities: an observational population study. The Lancet 372 (9650), 1655–1660.
Münzel, T., et al., 2017. Environmental stressors and cardio-metabolic disease: part
I–epidemiologic evidence supporting a role for noise and air pollution and effects of
mitigation strategies. Eur. Heart J. 38 (8), 550–556.
Neira, M., Prüss-Ustün, A., Mudu, P., 2018. Reduce air pollution to beat NCDs: from
recognition to action. The Lancet 392 (10154), 1178–1179.
Nieuwenhuijsen, M.J., et al., 2015. Variability in and agreement between modeled and
personal continuously measured black carbon levels using novel smartphone and
sensor technologies. Environ. Sci. Technol. 49 (5), 2977–2982.
Nikkilä, A., et al., 2018. Effects of incomplete residential histories on studies of en-
vironmental exposure with application to childhood leukaemia and background ra-
diation. Environ. Res. 166, 466–472.
Nyhan, M., et al., 2016. “Exposure Track” the impact of mobile-device-based mobility
patterns on quantifying population exposure to air pollution. Environ. Sci. Technol.
50 (17), 9671–9681.
Nyhan, M., et al., 2018. Quantifying population exposure to air pollution using individual
mobility patterns inferred from mobile phone data. J. Eposure Sci. Environ.
Epidemiol.
Park, Y.M., Kwan, M.-P., 2017. Individual exposure estimates may be erroneous when
spatiotemporal variability of air pollution and human mobility are ignored. Health &
Place 43, 85–94.
Pennington, A.F., et al., 2017. Measurement error in mobile source air pollution exposure
estimates due to residential mobility during pregnancy. J. Eposure Sci. Environ.
Epidemiol. 27 (5), 513.
Picornell, M., et al., 2018. Population dynamics based on mobile phone data to improve
air pollution exposure assessments. J. Eposure Sci. Environ. Epidemiol. 1.
Pope III, C.A., Dockery, D.W., 2006. Health effects of fine particulate air pollution: lines
that connect. J. Air Waste Manag. Assoc. 56 (6), 709–742.
Reis, S., et al., 2018. The influence of residential and workday population mobility on
exposure to air pollution in the UK. Environ. Int. 121, 803–813.
Setton, E., et al., 2011. The impact of daily mobility on exposure to traffic-related air
pollution and health effect estimates. J. Eposure Sci. Environ. Epidemiol. 21 (1), 42.
Shafran-Nathan, R., Levy, I., Broday, D.M., 2017. Exposure estimation errors to nitrogen
oxides on a population scale due to daytime activity away from home. Sci. Total
Environ. 580, 1401–1409.
Singh, V., Sokhi, R.S., Kukkonen, J., 2019. An approach to predict population exposure to
ambient air PM2.5 concentrations and its dependence on population activity for the
megacity London. Environ. Pollut. 113623.
Smith, J.D., et al., 2016. London hybrid exposure model: improving human exposure
estimates to NO2 and PM2.5 in an urban setting. Environ. Sci. Technol. 50 (21),
11760–11768.
Tang, R., et al., 2018. Integrating travel behavior with land use regression to estimate
dynamic air pollution exposure in Hong Kong. Environ. Int. 113, 100–108.
Wacholder, S., 1995. When measurement errors correlate with truth: surprising effects of
nondifferential misclassification. Epidemiology (Cambridge, Mass.) 6 (2), 157–161.
Wang, H., et al., 2010. Transportation mode inference from anonymized and aggregated
mobile phone call detail records. In: Intelligent Transportation Systems (ITSC), 2010
13th International IEEE Conference on. IEEE.
Yoo, E., et al., 2015. Geospatial estimation of individual exposure to air pollutants:
Moving from static monitoring to activity-based dynamic exposure assessment. Ann.
Assoc. Am. Geogr. 105 (5), 915–926.
Yu, H., et al., 2018b. Cross-comparison and evaluation of air pollution field estimation
methods. Atmos. Environ. 179, 49–60.
Yu, H., et al., 2018a. Using cell phone location to assess misclassification errors in air
pollution exposure estimation. Environ. Pollut. 233, 261–266.
Yu, X., et al., 2019. On the accuracy and potential of Google Maps location history data to
characterize individual mobility for air pollution health studies. Environ. Pollut.
Zhang, S., et al., 2018b. Long-term effects of air pollution on ankle-brachial index.
Environ. Int. 118, 17–25.
Zhang, Z., et al., 2018a. Long-Term Exposure to Fine Particulate Matter, Blood Pressure,
and Incident Hypertension in Taiwanese Adults. Environ. Health Perspect. 126 (1)
017008–017008.
Zhang, D., et al., 2014. Exploring human mobility with multi-source data at extremely
large metropolitan scales. In: Proceedings of the 20th annual international conference
on Mobile computing and networking. ACM.
Zhang, D., et al., 2015. coMobile: Real-time human mobility modeling at urban scale
using multi-view learning. In: Proceedings of the 23rd SIGSPATIAL International
Conference on Advances in Geographic Information Systems. ACM.
Zhang, D., 2020. Desheng Zhang, Rutgers University. [cited 2020 March 8]; Available
from: https://www.cs.rutgers.edu/~dz220/data.html.
Zhao, Z., et al., 2016. Understanding the bias of call detail records in human mobility
research. Int. J. Geograph. Informat. Sci. 30 (9), 1738–1762.
X. Yu, et al. Environment International 141 (2020) 105772
11
