Journal of Exposure Analysis and Environmental Epidemiology (1999) 9, 245±260
# 1999 Stockton Press All rights reserved 1053-4245/99/$12.00
http://www.stockton-press.co.uk
Models of exposure to carbon monoxide inside a vehicle on a Honolulu
highway
PETER G. FLACHSBART
Department of Urban and Regional Planning, University of Hawaii at Manoa, Honolulu, HI 96822
This paper presents statistical models of passenger exposure to carbon monoxide (CO) inside a motor vehicle as it traveled a coastal highway in Honolulu,
Hawaii during morning periods between November, 1981 and May, 1982. The 3.85-mile study site was divided into three links. The models predict the average
CO concentration inside the vehicle's passenger cabin on the third link as a function of several variables: the average CO concentrations inside the cabin on
previous links; traffic, temporal, and meteorological variables; motor vehicle CO emission factors; and ambient CO concentrations. Based on data for 80 trips,
the three most powerful models (adjusted R2=0.69) were nonlinear combinations of four variables: the average CO concentration inside the cabin for the second
link; wind speed and direction; and either the travel time, vehicle speed or CO emission factor for the third link. Several nonlinear models were based on data
for 62 trips for which nonzero, ambient CO concentrations were available. For this database, the most practical models (adjusted R2=0.67) combined three
variables: the ambient CO concentration; the second-link travel time; and either the travel time, vehicle speed or CO emission factor for the third link. Two
factors of third-link CO exposure varied seasonally. Relatively lighter traffic flows and stronger winds lowered cabin exposures during late fall, while heavier
traffic flows and calmer winds elevated cabin exposures during winter and spring. This study confirms the importance of seasonal effects on cabin exposure, as
observed by a California study, and adds new insights about their effects.
Keywords: carbon monoxide, motor vehicle emissions, passenger cabin exposure, statistical models.
Introduction
Reductions in tailpipe and evaporative emissions of carbon
monoxide (CO) from motor vehicles have been achieved
through various emission control technologies and clean
fuels. These reductions can be justified if they lead to
attainment of ambient air quality standards for CO and
lower population exposure to automotive emissions. Previous study of urban populations has shown that high levels
of CO exposure have occurred in certain microenvironments (e.g., passenger cabins of motor vehicles and parking
garages) where concentrations can be high even though time
spent in them may be relatively low (Akland et al., 1985).
Statistical models of exposure to motor vehicle emissions
enable identification of factors that contribute to high
exposure levels (Sexton and Ryan, 1988).
1. Abbreviations: ADT, average daily traffic; CO, carbon monoxide; 8F,
degrees Fahrenheit; ft, feet; g/veh-mi, grams per vehicle-mile; HIA,
Honolulu International Airport; in., inches; log, logarithm; mph, miles per
hour; PEM, personal exposure monitor; ppm, parts per million; psi, pounds
per square inch; US EPA, US Environmental Protection Agency.
2. Address all correspondence to: Peter G. Flachsbart, Ph.D., Department
of Urban and Regional Planning, University of Hawaii at Manoa,
Honolulu, HI 96822. Tel.: (808)956-8684. Fax: (808)956-6870. E-mail:
[email protected]
In previous work, Flachsbart (1985) developed three
statistical models for predicting average concentrations of
CO inside a vehicle on a 1.55-mile segment of the
Kalaniana'ole Highway in Honolulu, Hawaii. These models
were called `prototypal' because they were based on only 12
home-to-work trips taken during conditions of neutral
atmospheric stability. Data for the models came from field
surveys in 1981±1982 of personal exposure to motor vehicle
exhaust in various Honolulu microenvironments (Flachsbart
and Brown, 1985). Of the three models, the most powerful
one (R2=0.77) was based on the superposition theory, which
added ambient CO concentrations to CO emissions from
motor vehicles in the highway microenvironment. These
emissions were estimated as the product of the roadway's
traffic count and a vehicular emission factor as determined
from the Mobile2 model of the US Environmental
Protection Agency (US EPA). While the models explicitly
linked in-vehicle exposure directly to automotive emission
factors, the models ignored other potential explanatory
factors.
More recently, Ott et al. (1994) developed several
statistical models of passenger cabin exposure to CO from
highway emissions, based on 88 trips taken during a 13.5month period in 1980±1981. All trips occurred in one
vehicle with windows set in a `standard position' as it
traveled an arterial highway (El Camino Real) in the San
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Francisco Bay Area of California. The models are
noteworthy because they examined the explanatory power
of nine variables. The best model predicted the average CO
exposure per trip as a function of just two variables: traffic
conditions as measured by the proportion of travel time
stopped, and a seasonal trend term expressed as a cosine
function of the day of the year on which the trip was taken.
This model was not only powerful (adjusted R2=0.67), but
elegant as it explained in-vehicle CO exposure in terms of
only two variables. A model that included ambient CO
concentrations from a fixed-site monitor slightly improved
the power of the model (adjusted R2=0.71).
The insights provided by the El Camino Real models
inspired the author to launch an analytical study in 1988 to
develop new statistical models of passenger cabin CO
exposure to highway emissions in Honolulu. The author
presented preliminary findings of this study at annual
meetings of the International Society of Exposure Analysis
in 1991, 1994, and 1997. The new models are based on CO
exposure data collected inside three vehicles used in
commuting on two Honolulu highways between November,
1981 and May, 1982. Supplementary data on several
explanatory factors of exposure were derived from archival
sources. More details about these models appear in a report
(Flachsbart, 1998). As compared to the older `prototypal'
models, the new models are more robust, because they use a
database that encompassed a wider variety of atmospheric
conditions and a larger number of trips. Of those trips that
provided useful data for model development, most of them
were made by the author while commuting in his personal
vehicle from home to work in morning traffic on the
Kalaniana'ole Highway in Honolulu.
Although the California and Hawaii studies were
different in several respects, they reported similar passenger
exposures to CO in highway traffic for the early 1980s. In
the California study, the median CO level was 9.3 ppm
(parts per million) for 93 trips, which included five
nonstandardized trips to test the effect of various window
positions. Trips occurred throughout the day in both
directions of a 5.9-mile route on an arterial highway, and
took from 31 to 61 min to complete (Ott et al., 1994). In the
Hawaii study, the median CO level was 10.6 ppm for 104
trips taken over a 12-mile route that included both highway
and freeway traffic. Trips took from 20 to 64 min to
complete and occurred between 6:30 and 8:30 a.m.
(Flachsbart and Brown, 1985). The slightly higher median
CO exposure of the Hawaii study could be attributed to its
focus on rush-hour traffic, which typically slows vehicle
speeds and increases CO emission rates per vehicle
(Papacostas, 1987).
Current CO exposures are much lower than the levels
reported by the California and Hawaii studies. Lawryk et al.
(1995) reported median, in-vehicle CO levels of only 1.9
ppm for 33 urban trips and 2.3 ppm for 113 suburban trips
246
taken in 1991±1992 in the New Jersey/New York
metropolitan area. These lower levels of CO exposure are
evidence of the progress made by the Federal Motor Vehicle
Emission Control Program under the Clean Air Act. Further
evidence was reported by Flachsbart (1995) for 16 studies of
passenger cabin CO exposure done in the United States
between 1965 and 1992.
Unlike Ott et al. (1994) who repeated their study of El
Camino Real in 1991±1992, it would be difficult to repeat
the study of the Kalaniana'ole Highway in the present to
determine whether reductions in automotive emissions have
lowered commuter CO exposure. First, the author no longer
owns the test vehicle used in the 1981±1982 study. Second,
in the mid-1990s, another lane was built on the Kalaniana'ole Highway to accommodate traffic generated by new
home construction in areas served by the highway.
Construction of the highway lane required substantial
modifications of the surrounding landscape that could have
altered local wind patterns. Third, ambient CO levels near
the study site were monitored at a station that is no longer in
operation, and appropriate substitute stations are not
available.
Although the models presented in this paper are new,
they represent conditions in Honolulu during the early
1980s. Consequently, the models are not designed to predict
current or future exposure levels inside motor vehicles in
traffic either in Honolulu or elsewhere. Nevertheless, the
paper offers two contributions to the science of commuter
exposure. First, it seeks to determine whether the insights of
the California exposure models developed by Ott et al.
(1994) can be supported. This may be possible because the
surveys of Kalaniana'ole Highway were done about a year
after the surveys of El Camino Real. Comparisons between
the two studies may be informative given their similar time
frames albeit dissimilar locations and data collection
protocols. Since the models presented here cannot predict
future exposure levels, they should not be compared to a
model recently developed by Yu et al. (1996) to predict
exposure levels on El Camino Real in the near future.
Second, the paper seeks to identify factors (other than
smoking) that affect passenger cabin CO exposure, including factors that have not been explored in previous study,
and to determine whether these factors are interrelated and
to what extent.
Study design
The sections below describe the study site, data collection
methods, and quality assurance procedures of the field
surveys of passenger exposure to motor vehicle exhaust on a
Honolulu highway. More details about the study design
appear in Flachsbart and Brown (1985).
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Study Site
The study site was the Kalaniana'ole Highway, a coastal
artery on level terrain that connects suburban East Honolulu
and the city's urban core. In 1979, this artery had an average
daily traffic (ADT) count of nearly 65,000 vehicles in both
directions. By comparison, Ott et al. (1994) reported that the
ADT of El Camino Real in California ranged from 30,500 to
45,000 vehicles per day. The Honolulu study focused on
morning traffic which flowed primarily along the westbound portion of the highway toward the city's downtown
area. Since the highway served affluent residential and light
commercial areas, about 98% of the vehicle mix consisted
of light-duty vehicles (Department of General Planning,
1982).
The study site extended for 3.85 miles along an east±west
alignment, which was bounded by Kawaihae St. on the east
and by Ainakoa Ave. on the west. The site was divided into
three links and each link had intersections and traffic
signals. Link 1 extended 1.90 miles from Kawaihae St. to
Kirkwood St.; Link 2 stretched from there to W. Hind Dr., a
distance of only 0.40 miles; and Link 3 extended 1.55 miles
from W. Hind Dr. to Ainakoa Ave., which marks the start of
the H-1 Freeway. Although Link 1 had three lanes in the
westbound direction, one lane was designated as a contraflow lane for carpools and express buses. The contraflow
lane became a withflow, carpool lane at Kirkwood St. Link
3 also had three westbound lanes, all of which could be used
by any type of vehicle and passenger loading. A landscaped
medial strip divided the west- and eastbound lanes of Link
3. Link end points were chosen to enable a study of the
effectiveness of the highway's contraflow lane for highoccupancy vehicles in reducing commuter travel time and
CO exposure (Flachsbart, 1989). Although the statistical
models developed by Flachsbart (1985) focused only on
Link 3, the models presented below were developed from
data for all three links.
Data Collection Methods
The driver of the test vehicle tried to maintain a speed
representative of the speed of surrounding traffic using the
`floating car technique' (Baerwald, 1976). This technique
required the driver to pass the same number of vehicles as
passed his vehicle. Average test vehicle speeds for each link
were determined from a one-time distance measurement of
each link (measured in miles) using the test vehicle's
odometer and daily records of travel time on each link
(measured in minutes and seconds).
The US EPA provided portable, personal exposure
monitors (PEMs) to measure CO exposure. The PEM
consisted of a CO detector made by General Electric (Model
15ECS3CO3) attached by cable to an integrator developed
for EPA by Mark Stoelting of Custom Instrumentation in
Santa Monica, California. The CO detector performed with
minimal interference from other gases, needed no reagent
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Flachsbart
except water, delivered linear response with concentration,
and had a relatively rapid response time of less than 1 min
(Laconti et al., 1981). The accuracy of the CO detector was
‹2 ppm at zero concentration, and ‹10% for concentrations
ranging from 0 to 500 ppm. The CO detector sent a directcurrent signal (0±250 mV) of the CO measurement to the
integrator, which converted the signal into a millivoltminute integral. The integrator displayed the integral as
parts per million-minutes of CO exposure.
Using a Casio F-7 watch built into the integrator's cover,
the data recorder took simultaneous readings of both CO
exposure and time at the end points of each link of the study
site. After the trip, the average in-vehicle CO concentration
for the link was computed by dividing the accumulated CO
exposure (measured in parts per million-minutes) by the
time spent on the link (measured in seconds and converted
to minutes assuming accuracy of two decimal places, e.g.,
1.23 min).
By comparison, Ott et al. (1994) recorded CO exposure
data on a strip chart in their El Camino Real study, and later
digitized the data at 12-s intervals. Five consecutive
digitized values were then averaged to obtain a 1-min CO
average. Thus, the California and Hawaii studies used
different methods of collecting exposure data. More
advanced personal monitors are currently capable of
acquiring CO exposure measurements frequently (i.e.,
every few seconds) and storing data automatically for later
retrieval and downloading to personal computers. Such
monitors were not yet available during the California and
Hawaii studies of the early 1980s.
The study by Flachsbart and Brown (1985) offered data
for 142 morning trips, including weekday and weekend
travel, by three passenger cars on two highways in
Honolulu. These trips were screened to find the largest
number of trips made by only one vehicle on a single
highway. Of 142 trips, 125 occurred on the Kalaniana'ole
Highway and the remaining 17 were made on the Pali
Highway which served the windward side of the Island of
Oahu. Of the 125 trips on Kalaniana'ole Highway, 97 were
made in a 1975 Toyota Celica of which 17 occurred during
calm winds (i.e., zero wind speed). Since CO concentrations
on the highway were assumed to be inversely related to
wind speed, exposure models based on zero wind speed data
were mathematically unfeasible. This restriction further
reduced the Kalaniana'ole database from 97 to 80 trips. For
these 80 trips, travel on the study site occurred sometime
between 6:45 a.m. and 9:45 a.m. from November 2, 1981,
through May 4, 1982.
Since the test vehicle did not have air conditioning, the
positions of the front windows were adjusted daily for
passenger comfort. The rear windows were always closed.
Given Honolulu's tropical climate, the front windows were
typically open full or part way, except during rainfall when
all windows were usually closed. As a result, window
247
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
one-time, dual monitors precision test. The results of these
tests and procedures were satisfactory as reported elsewhere
(Flachsbart and Brown, 1985).
positions of the test vehicle in the Honolulu study were
similar, but not identical to the `standard window position'
of the test vehicle in the California study. In that study, the
driver's window was entirely open, the passenger's window
was open 3 in. (inches), and all other windows were closed
(Ott et al., 1994).
Variables
Quality Assurance
Several procedures were followed to assure collection of
high-quality data. To eliminate nontraffic sources of CO
inside the test vehicle, all passengers refrained from
smoking during each trip. The passenger cabin of the test
vehicle was inspected for exhaust leaks from the engine and
exhaust pipe and was found free of CO intrusion from these
sources. Thus, any CO measured inside the vehicle was
expected to come primarily from other motor vehicles on
the roadway and/or from background sources in the ambient
environment. Other quality assurance tests included: (1)
monitor calibration (zero, span) procedures twice a week;
(2) an EPA audit of the commercial gases used to calibrate
the CO monitors; (3) another EPA audit to test the monitor's
ability to identify unmarked CO sample mixtures; and (4) a
Table 1 lists variables for which data had been collected in
1981±1982 by Flachsbart and Brown (1985) as part of a
larger study with a different purpose. Of the 15 variables
listed in Table 1, the variable representing motor vehicle
emissions required additional data inputs and assumptions,
as described later, before CO emission factors could be
estimated. The dependent variable was the average CO
concentration inside the test vehicle on each link of the
study site. The independent variables fell into several
categories: traffic, temporal, and meteorological variables;
CO emission factors; and the ambient CO concentration.
While window positions were not explicitly treated as a
variable, they were affected by rainfall which was
represented by two variables.
Table 1. Variables used in model development.
Passenger cabin exposure
average CO concentration inside test vehicle on link i (ppm)
CEi
Traffic variables
TF3
traffic flow while test vehicle is on Link 3 (veh/15 min)
TTi
test vehicle's travel time on link i (min)
VS3
test vehicle's average speed on Link 3 based on TT3 (mph)
Temporal variables
time when test vehicle enters link i (min past 6 a.m.)
ETi
SD
serial day of field survey starting with November 1, 1981=Day 1, including weekends and holidays (days)
Meteorological variables
AP3
atmospheric pressure at sea level at HIA while test vehicle is on Link 3 (mbar)
AT3
ambient temperature at HIA while test vehicle is on Link 3 (8F)
IN
average depth to base of inversion layer at HIA for date (ft)
MP
RF
measured precipitation at HIA for date (in)
presence of rainfall on the study site (RF = 0 for no rainfall and RF = 1 for rainfall)
WD3
wind direction at HIA while test vehicle is on Link 3; if $=an azimuth from north, then WD3=0 for 08< $ < 808 and 2808< $ < 3608,
and WD3 =1 for 808 $ 2808
WS3
hourly wind speed at HIA while test vehicle is on Link 3 (mph)
Motor vehicle emissions
EF3
Mobile4.1 exhaust CO emission factor while test vehicle is on Link 3 (g/veh-mi); and
Ambient concentration
AC3
248
hourly ambient CO concentration recorded at Leahi Hospital while test vehicle is on Link 3 (ppm)
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Traffic Variables
The study used three variables to represent traffic conditions. These were traffic flow and the test vehicle's travel
time and average speed on each link. The relationship
between vehicle speed and traffic volume is most apparent
under conditions of uninterrupted or uniform traffic flow. In
such cases, traffic speed will fall as traffic flow increases
until a maximum flow is reached. The addition of more
vehicles to the roadway beyond its maximum capacity
causes both flow and speed to fall (Papacostas, 1987). In
reality, traffic flow on the study site was interrupted by a
series of traffic signals which meant that flow was not
uniform.
Traffic flows, measured in vehicles per 15-min periods,
were available from a pneumatic tube counter operated by
the Hawaii state Department of Transportation. The tube
stretched across all three westbound lanes of Link 3 just
before its terminus at Ainakoa Ave. Traffic counts revealed
a wide variation in traffic flows ranging from a low of 573
vehicles between 8:45 a.m. and 9:00 a.m. on November 11,
1981, to a high of 1349 vehicles between 7:45 a.m. and 8:00
a.m. on February 22, 1982. The test vehicle's speed, which
ranged from 7.1 to 37.3 mph (miles per hour) on Link 3,
reflected the wide variation in traffic counts. This variation
suggested that commuter exposure models were feasible
given that variation in vehicle speed affects variation in
tailpipe emissions of CO from a motor vehicle in traffic
(Papacostas, 1987).
Unfortunately, there were two problems with traffic
counts. First, they existed for only 21 of the 80 trips. These
21 trips spanned the period from November 2, 1981,
through April 5, 1982. Second, there were an unrecorded
number of vehicles that turned onto Link 3 from intersecting
streets located near the traffic counter. Although these
vehicles were included in traffic counts, it was plausible to
assume that tailpipe emissions from these vehicles probably
contributed little to the exposure of passengers inside the
test vehicle while it was on Link 3.
Given these problems with a direct measure of traffic
flow on Link 3, it was assumed that the test vehicle's
average speed and travel time on Link 3 were indirect
measures of traffic conditions. To verify this assumption,
models to predict travel time (TT3) and average vehicle
speed (VS3) on Link 3 as functions of traffic flow on Link 3
(TF3) were developed. First, VS3 (expressed in miles per
hour) was determined by dividing the length of Link 3 (1.55
miles) by TT3 (measured in minutes and converted to hours
for this calculation) using Equation 1 below. Then, separate
models of TT3 and VS3 based on TF3 were developed using
the least squares method of regression analysis. These
models are shown by Equation 2 for TT3 and Equation 3 for
VS3. Both models were statistically significant (F=7.6,
p=0.004 for Equation 2; F=11.1, p<0.001 for Equation 3),
and both models had respectable explanatory power
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Flachsbart
(multiple R2=0.46 for Equation 2; multiple R2=0.55 for
Equation 3). These two models supported the assumption
that the vehicle's average speed and travel time were
indirect measures of traffic conditions on the highway.
VS3 ˆ 1:55=…TT3 =60†
TT3 ˆ
27:92 ‡ 0:0691…TF3 †
…1†
0:0000317…TF3 †2
VS3 ˆ 0:2116…TF3 † 0:0003623…TF3 †2
‡ 0:00000016…TF3 †3 :
…2†
…3†
Temporal Variables
Studies have shown that highway traffic volumes in urban
areas vary by hour of the day and by season of the year
(Edwards, 1992). Two temporal variables were defined to
account for these factors: the serial day of the field survey
and the time when the vehicle entered link i. Plots (not
shown) of Link 3's average traffic flow by survey date and
by hour of the day supported these expectations. The plots
indicated that mean traffic flows on Link 3 were relatively
lower in November and December and higher in February
and March, and that average weekday traffic flows generally
reached a morning peak between 7:15 and 7:30 a.m.
Meteorological Variables
Data on meteorological variables were obtained from the
National Weather Service for its Honolulu International
Airport (HIA) station approximately 12 miles west of the
study site. The variables included atmospheric pressure at
sea level, ambient temperature, average depth to the
atmospheric inversion layer, two measures of precipitation,
and wind direction and speed. Certain aspects of Honolulu's
tropical climate make it different from most other cities.
During surveys of passenger cabin exposure, ambient
temperatures ranged from 58 to 818F (degrees Fahrenheit),
and wind speeds varied seasonally. Average wind speeds
were 12.40 mph for 22 trips taken in November and
December, 1981, but were only 7.82 mph for 58 trips taken
between January and May, 1982. This difference in seasonal
wind speeds was statistically significant (t = 4.56, p <0.001).
In Hawaii, northerly winds are known as `trade' winds
and southerly winds are called `Kona' winds. Trade winds
were more prevalent during the study period, occurring for
77.5% of the 80 trips, but could not be linked to one season
over another. However, the average speed of trade winds
(8.30 mph for 62 trips) was substantially weaker than the
average of Kona winds (11.76 mph for 18 trips). This
difference was statistically significant (t = ±3.03, p = 0.003).
Also, wind speed and atmospheric pressure were positively
249
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Emission factors were generated only for traffic flowing
westbound on the study site, because average vehicle speeds
were not available for eastbound traffic. This was not
considered a major loss, because northerly trade winds
typically kept eastbound traffic emissions from reaching the
test vehicle as it traveled in westbound lanes. However,
occasional southerly winds could have carried eastbound
traffic emissions onto the westbound lanes where passenger
exposures were measured. Hence, a dummy variable for
wind direction was created as indicated in Table 1.
correlated with each other as shown by the Spearman rank
correlation coefficient (rS=0.23, p=0.04). This meant that
higher wind speeds occurred during periods of higher
atmospheric pressures.
Emission Factors
Although CO emission factors can be estimated from
mobile source models, they are seldom estimated in
exposure studies because the estimates require data inputs
that are difficult to obtain. In this study, CO emission factors
were estimated for each of the 80 trips. Table 2 shows the
assumptions and inputs used to generate emission factors
while the test vehicle traveled Link 3 of the study site. These
factors were generated on a personal computer using
Mobile4.1 software (US Environmental Protection Agency,
1991). Mobile4.1 was selected over Mobile5 based on
results of observed CO emission rates from studies in 1992
of the Tuscarora Mountain Tunnel in Pennsylvania and the
Fort McHenry Tunnel under Baltimore Harbor. In the
Tuscarora Tunnel study, both emission factor models
overpredicted observed CO emission rates of light-duty
vehicles; however, the Mobile4.1 predictions were much
closer to observed emission rates than were predictions from
Mobile5 (Robinson et al., 1996). Although results of the
Fort McHenry Tunnel study favored Mobile5, the results of
the Tuscarora Tunnel were considered more appropriate to
this study, because the highway through the Tuscarora
Tunnel was relatively flat like the Kalaniana'ole Highway.
Ambient Concentrations
Ambient CO concentrations were available from the state
Department of Health for a fixed-site monitor located at
Leahi Hospital, located about 1.5 miles west of the
highway's Ainakoa Ave. intersection which marked the
end of Link 3. The hospital site provided hourly CO
concentrations for 64 of the 80 trips. Missing ambient data
for 16 trips were attributed to nonperforming monitors. The
ambient data were considered background concentrations,
because the hospital was in a residential area with light
traffic. For the 64 trips, the median ambient CO concentration was only 0.8 ppm and hourly concentrations ranged
from 0 to 3.5 ppm during the study period. This suggested
that passenger exposure on the study site could be attributed
primarily to emissions on the highway. Even so, exposure
models that included ambient data were tested and
developed.
Table 2. Assumptions for Mobile4.1 CO emission factors.
Input characteristic
Measured values or assumptions
Region
Low
Altitude
Year
Near sea level
1981±1982
Vehicle typesa
Light duty gas vehicles (89.4%)
Light duty gas trucks #1 (5.3%)
Light duty gas trucks #2 (3.1%)
Heavy duty gas vehicles (0.9%)
Light duty diesel vehicles (0.7%)
Heavy duty diesel vehicles (0.6%)
Vehicle registration
Ambient temperatures
Island of Oahu, Hawaii
58±818F
Traffic speeds
Approximately 7.1±37.3 mph
Cold starts
Default: 20.6% by noncatalyst and 20.6% by catalyst-equipped vehicles
Hot starts
Default: 27.3% by catalyst-equipped vehicles
Tampering rates
Default
Reid vapor pressure
11.5 psi (pounds per square inch)
Other modifiers
No inspection and maintenance or anti-tampering programs;
no air conditioning or trailer loads
a
June, 1982 survey by Department of Transportation, State of Hawaii.
250
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Air-Exchange Rates
Passenger cabin exposure models typically include a
variable to represent the test vehicle's air exchange rate.
Unfortunately, that rate was not determined by Flachsbart
and Brown (1985), and could not be determined during
model development as the author no longer owned the test
vehicle. Hence, average CO concentrations inside the
vehicle for a given link were assumed to equal roadway
concentrations. This is a plausible assumption when
concentrations inside the vehicle are averaged over a period
of time much greater than the vehicle's `time constant'. This
constant is the time required for the concentration inside the
vehicle to approach the concentration on the roadway.
The vehicle's `time constant' was assumed to fall
between 30 and 54 s. This assumption was based on a
study by Ott and Willits (1981) of a passenger car moving at
20 mph with partially open windows. In the present study,
travel times generally exceeded 30 s on each link with one
exception. The exception occurred on Link 2, because it was
only 0.40 miles long. Since Link 2 was short, some models
were tested using combined data for Links 1 and 2. When
Flachsbart
data were combined, the link subscript appears as i=20.
Otherwise, the assumption that average interior and exterior
CO concentrations were about equal for a given link
appeared to be satisfied.
Models
This section presents statistical models of in-vehicle
exposure to CO during morning commutes on a Honolulu
highway, based on linear and nonlinear combinations of
several explanatory variables, as well as models of
explanatory variables to show how some variables affect
others. Each model was fitted to the data using the least
squares method of regression analysis (Draper and Smith,
1981). This method minimized the sum of the squared
differences between predicted and observed values of the
dependent variable. Regression analysis was performed
using StatWorks2 1.2 on a Macintosh SE personal
computer. Nonlinear models included both polynomial
Figure 1. Factors affecting CO exposure inside a vehicle on the Kalaniana'ole Highway in Honolulu, Hawaii.
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
251
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
and power models. The power models required logarithmic
transformations of the data.
In evaluating the results of the regression analysis, the
first step was to identify candidate models that satisfied four
criteria: (1) the F-statistics for total regression had a
probability ( p) 0.05; (2) the Student's t statistic for each
variable coefficient had a probability ( p) 0.05; (3) the sign
of the variable coefficient could be explained by mathematical and scientific reasoning; and (4) predicted values of
dependent variables were not negative over the range of
values observed for independent variables. In most cases,
these statistics are not reported below for each model;
however, the multiple R2 and the adjusted R2 statistics are
reported.
In evaluating models of commuter exposure, the second
step is usually to select the `one best model' from among the
alternatives that satisfy specified criteria such as those
mentioned above. This step is appropriate if the main
purpose of the `one best model' is to predict precise levels of
exposure at a given point in time, i.e., developing the `best'
model that could predict exposure to CO emissions during
morning commutes on Honolulu's Kalaniana'ole Highway
in 1981±1982. That purpose seemed irrelevant for this
study, given that commuter CO exposure levels have
declined substantially since the early 1980s (Flachsbart,
1995). Given the age of the survey data, it seemed more
appropriate to advance the science of how certain factors
affect commuter exposure, including factors not previously
studied, and how these factors affect each other.
To that end, the following sections describe all of the
models that could be developed, and Figure 1 shows factor
interrelationships based on a selection of statistically
Table 3. Univariate models of in-vehicle exposure.
Multiple R2
Adjusted R2
Independent variable
17
0.343
0.334
log (vehicle's travel time on Link 1)
22
0.121
0.110
vehicle's entry time on Link 1
14
0.585
0.580
log (vehicle's travel time on Link 2)
8
9
0.551
0.521
0.546
0.515
log (average in-vehicle CO concentration for Link 1)
average in-vehicle CO concentration for Link 1
16
0.442
0.435
log (vehicle's travel time on Links 1 and 2 combined)
15
0.208
0.198
log (vehicle's travel time on Link 1)
20
0.178
0.156
vehicle's entry time on Link 2
21
0.171
0.150
vehicle's entry time on Link 1
4
5
0.566
0.548
0.560
0.542
log (average in-vehicle CO concentration for Links 1 and 2 combined)
log (average in-vehicle CO concentration for Link 2)
average in-vehicle CO concentration for Links 1 and 2 combined
Equation
Link 1
Link 2
Link 3
6
0.510
0.504
11
0.507
0.500
log (vehicle's average speed on Link 3)
10
0.506
0.500
log (vehicle's travel time on Link 3)
7
0.502
0.496
average in-vehicle CO concentration for Link 2
19
0.492
0.486
log (exhaust CO emission factor for Link 3)
12
0.467
0.460
log (vehicle's travel time on Link 2)
18a
13
0.512
0.368
0.390
0.360
traffic flow on Link 3
log (vehicle's travel time on Links 1 and 2 combined)
29b
0.251
0.238
log (ambient CO concentration while vehicle was on Link 3)
23
0.235
0.205
vehicle's entry time on Link 2
24
0.234
0.204
vehicle's entry time on Link 3
26
0.149
0.127
atmospheric pressure at sea level while vehicle was on Link 3
27
0.132
0.121
wind speed while vehicle was on Link 3
28c
0.116
0.091
measured precipitation for date
25
0.075
0.039
serial day of field survey
a
n = 21 trips.
n = 62 trips.
c
n = 37 trips.
b
252
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
superior models. To improve clarity, the factors in Figure 1
were arranged to minimize crossing the arrows connecting
the factors; however, crossing arrows could not be avoided
completely. In Figure 1, the arrow points from the
independent variable to the dependent variable, and the
number next to the arrow indicates the equation in the paper
that expresses their mathematical relationship. Although
Figure 1 applies specifically to this study, it may shed light
on a general theory of commuter exposure to motor vehicle
emissions on a highway and enable more intelligent field
survey designs of similar studies in the future.
Univariate Models of Cabin Exposure
This section presents several univariate statistical models of
passenger cabin CO exposure for the study site. Most of
these models predict cabin exposure on Link 3 (CE3) of the
study site, but use data collected for explanatory variables
from all three links, including the average CO concentration
inside the test vehicle while it was on previous links. The
models are discussed below by category of predictor
variable. Table 3 provides a summary of the univariate
models of in-vehicle exposure. For each link, the models are
listed in order of descending predictive power as indicated
by the adjusted R2 values.
Serial Correlation in Exposure Table 4 shows six models
listed in order of descending predictive power that predict
the average CO concentration on link i based on the average
concentration measured on link i±1. The table shows that
the average CO concentration of the passenger cabin on one
link was a strong predictor of the cabin's average CO
concentration on the next link indicating serial correlation in
exposure. Table 4 also shows that the nonlinear models of
serial correlation (Equations 4, 5 and 8) were more powerful
than their counterpart linear models (Equations 6, 7 and 9),
respectively. Moreover, both the linear and nonlinear
models developed to predict the average concentration on
Link 3 based on the average concentration measured on
Links 1 and 2 combined (k =20) had slightly higher
Table 4. Exposure models based on previous link exposure.
Multiple R2
Adjusted R2
Model
4
0.566
0.560
CE3=3.96(CE20)0.567
5
0.548
0.542
CE3=3.64(CE2)0.584
Equation
Flachsbart
predictive power than comparable models based solely on
Link 2 exposure. Since time spent on Links 1 and 2
combined exceeded time spent solely on Link 2, this latter
result was contrary to an expectation that shorter time
intervals should be associated with higher serial correlation.
Traffic Variables Travel time on each link and the average
vehicle speed and traffic flow on Link 3 were denoted as the
traffic variables. As expected, travel time (TT3) and vehicle
speed (VS3) on Link 3 were able to predict CE3 fairly well,
and their predictive power was identical (multiple R2=0.51)
because vehicle speed was derived from travel time on Link
3 using Equation 1. Models with the most predictive power
of CE3 were based on the logarithms (logs) of TT3 and VS3.
By comparison, a linear model (not shown) to predict CE3
based on TT3 had less predictive power (multiple R2=0.35).
Equations 10 and 11 emerged as the best models of CE3
based on TT3 and VS3, respectively:
CE3 ˆ 1:40…TT3 †1:123 ;
…10†
CE3 ˆ 229:06=…VS3 †1:123 :
…11†
Equations 12 and 13 show that CE3 could also be
predicted from travel time on Link 2 (TT2) or travel time on
Links 1 and 2 combined (TT20), respectively. The predictive
power of Equation 12 (multiple R2=0.47) exceeded that of
Equation 13 (multiple R2=0.37):
CE3 ˆ 5:10…TT2 †0:891 ;
…12†
CE3 ˆ 1:37…TT20 †1:052 :
…13†
Both individual and combined travel times on the first
two links predicted cabin exposure on the second link (CE2)
as indicated by Equations 14, 15 and 16 below. However,
the predictive power of Equation 14 (multiple R2=0.59)
exceeded that of both Equation 15 (multiple R2=0.21) and
Equation 16 (multiple R2=0.44):
Link 3
6
0.510
0.504
CE3=5.9+0.92(CE20)
7
0.502
0.496
CE3=6.5+0.79(CE2)
0.551
0.521
0.546
0.515
CE2=2.43(CE1)0.663
CE2=3.5+0.86(CE1)
CE2 ˆ 2:35…TT2 †1:263 ;
…14†
CE2 ˆ 1:66…TT1 †0:992 ;
…15†
CE2 ˆ 0:39…TT20 †1:465 :
…16†
Link 2
8
9
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
253
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Equation 17 predicted cabin exposure on Link 1 (CE1)
based on travel time on Link 1 (TT1). Its predictive power
(multiple R2=0.34) fell midway between the power of the
three previous models:
CE1 ˆ 0:83…TT1 †1:277 :
…17†
Based on traffic flow on Link 3 (TF3), Equation 18 was
the best model that could be developed for predicting CE3.
While this model had good predictive power (multiple
R2=0.51), it was based on a small dataset (n=21 trips) and
gave erratic predictions of exposure for traffic flows outside
the range of 570 to 1350 vehicles/15 min:
CE3 ˆ 1231:30 5:8341…TF3 † ‡ 0:009995…TF3 †2
0:0000073…TF3 †3 ‡ 0:000000002…TF3 †4 :
254
CE1 ˆ 17:02
…22†
…19†
…20†
0:075…ET1 †:
Equations 23 and 24 predicted CE3 based on when the
vehicle entered Links 2 and 3, respectively. These two
models had identical predictive power (multiple R2=0.23).
Using differential calculus on Equation 24, one can
determine the entry times associated with the predicted
maximum and minimum CO exposures for Link 3. The
predicted maximum average concentration (20.3 ppm)
occurred at 7:34 a.m., and the predicted minimum average
concentration (6.1 ppm) occurred at 9:03 a.m.:
…18†
Temporal Variables The temporal variables included the
entry time on link i (ETi) and the serial day of the field
survey (SD). Of these two variables, link-entry time had
more predictive power, as shown in Table 3. Based on linkentry time, two models emerged for predicting passenger
cabin exposure on Link 2 (CE2). These were Equation 20
(multiple R2=0.18), based on Link 1 entry time, and
Equation 21 (multiple R2=0.17), based on Link 2 entry
time, neither of which had constants. Although alternative
models with constants had slightly more predictive power
than Equations 20 and 21, the alternatives predicted
negative cabin exposures for entry times that fell within
certain time periods. Equation 22 (multiple R2=0.12)
predicted cabin exposure on Link 1 (CE1) based on Link
1 entry time (ET1):
CE2 ˆ 0:511…ET1 † 0:00523…ET1 †2
‡ 0:00001354…ET1 †3 ;
…21†
CE3 ˆ
Motor Vehicle Emissions Equation 19 predicted cabin
exposure on Link 3 (CE3) based on the CO emission factor
as a predictor. This model had just slightly less power
(multiple R2=0.49) to predict CE3 than the power of
previously discussed models of cabin exposure based on the
vehicle's travel time (multiple R2=0.51) or speed (multiple
R2=0.51) on Link 3:
CE3 ˆ 0:106…EF3 †1:121 :
CE2 ˆ 0:4807…ET2 † 0:004658…ET2 †2
‡ 0:0000114…ET2 †3 ;
CE3 ˆ
48:53 ‡ 1:830…ET2 †
‡ 0:00003659…ET2 †3 ;
0:015224…ET2 †2
59:45 ‡ 2:055…ET3 †
‡ 0:00003917…ET3 †3 :
0:016594…ET3 †2
…23†
…24†
For the El Camino Real study, Ott et al. (1994) developed
a cosine model of passenger cabin exposure based on a
seasonal trend function. A similar model was attempted, but
could not be developed for the present study, which used the
serial day (SD) of the field survey period to represent
seasonal effects. Equation 25 was the best model that could
be developed to predict cabin exposure based on the survey
date; however, the predictive power of this model was very
low (multiple R2=0.08), and the statistics (F=2.06, p=0.11)
were not significant:
CE3 ˆ 20:48 0:4135…SD† ‡ 0:005797…SD†2
0:00002096…SD†3 :
…25†
Differential calculus was applied to Equation 25 to
determine the dates of the predicted maximum and
minimum exposures. The results indicated that the predicted
minimum average CO concentration inside the test vehicle
on Link 3 (11.7 ppm) occurred on the 48th day of the survey
period (December 18, 1981), and that the predicted
maximum average concentration (18.8 ppm) happened on
the 136th day (March 16, 1982).
Meteorological Variables The meteorological variables
were not very powerful predictors of cabin exposure as
shown in Table 3. Even so, univariate models of exposure
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
were developed based on wind speed, measured precipitation, and atmospheric pressure. Exposure models could not
be developed for ambient temperature, the depth of the
inversion layer, presence of rainfall on the study site and
wind direction. Of the meteorological variables, atmospheric pressure for Link 3 (AP3) was the best predictor
(multiple R2= 0.15) of CE3, as shown by Equation 26. Wind
speed for Link 3 (WS3) was the next best predictor of CE3
(multiple R2= 0.13), as shown by Equation 27. The negative
sign of the wind speed coefficient in Equation 27 is
consistent with the expectation that wind dilutes CO
concentrations:
CE3 ˆ
10:82 ‡ 0:493…AP3 †
CE3 ˆ 22:81
0:75…WS3 †:
0:00197…AP3 †2 ;
…26†
…27†
The presence of rainfall on the study site was expected to
slow average vehicle speed compared to days without
rainfall. Slower vehicle speeds could increase commuter
exposure both directly, through an increase in vehicle
emissions, and indirectly as the test vehicle with closed
windows moved slowly in heavy traffic. As expected,
rainfall on the study site reduced average vehicle speeds on
Link 3, as VS3 averaged 15.5 mph for 60 trips without
rainfall and 11.3 mph for 20 trips with rainfall. This
difference in average speed was statistically significant
(t = ±2.00, p=0.025) based on a one-tailed test of the
hypothesis. However, rainfall on the study site had no effect
on cabin exposure on Link 3, as CE3 levels averaged 15.5
ppm for the 60 trips without rainfall and 16.6 ppm for the 20
trips with rainfall. This difference in cabin exposure was
statistically not significant (t = 0.334, p=0.37). On the other
hand, Equation 28 (multiple R2=0.12) shows that CE3 could
be predicted from precipitation measured daily at the airport
(MP), based on data (n=37 trips) for which precipitation
exceeded zero. This model indicated that cabin exposure
increased with greater amounts of measured precipitation:
CE3 ˆ 14:68 ‡ 6:14…MP†:
…28†
Ambient CO Concentration Data on simultaneously
measured ambient CO concentrations were available for
64 of the 80 trips. Equation 29 was the best model that could
be developed to predict CE3 based on the ambient CO
concentration while the vehicle was on Link 3 (AC3). This
nonlinear model was actually based on just 62 trips for
which ambient CO levels exceeded zero. Since the log of
zero is undefined, two of the 64 trips were dropped prior to
developing Equation 29. Relative to other variables, Table 3
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Flachsbart
shows that the ambient CO concentration had modest
explanatory power (multiple R2=0.25) to predict cabin
exposure on Link 3:
CE3 ˆ 13:80…AC3 †0:610 :
…29†
Multivariate Models of Cabin Exposure
Given the variety of independent variables, many multivariate models of CE3 were possible. For practical reasons,
only models that had less than five independent variables
were tested. For multivariate models, the multiple R2
increases as more variables enter the model. The adjusted
R2 was used to compare the predictive power of multivariate
models, because it compensates for the number of variables
in the model. Only models whose adjusted R2 values
exceeded the predictive power of the best univariate model
of CE3, which was Equation 4 (adjusted R2=0.56), are
discussed below.
There were 12 multivariate models of CE3 with
predictive power substantially better than that of Equation
4. The models were equally divided into two groups of six
models each: one group based on all 80 trips; and the other
based on 62 trips for which nonzero, ambient CO
concentrations were available. The models in each group
are listed in Table 5 in order of descending predictive power
as indicated by the adjusted R2 values. For both groups,
models that included CE2 among the independent variables
had higher predictive power than those that did not. Each
group could be divided into two categories of three models
each. Each category had identical predictive power due to
relationships among three predictor variables for Link 3: the
test vehicle's travel time and average speed, and the CO
emission factor which was a function of the vehicle's
average speed. Although models in the first category of each
group were more powerful than those in the second
category, those in the second category were considered
more practical because they were based on obtainable data.
Models Based on All 80 Trips For this group of models, the
first category (adjusted R2=0.68) had higher predictive
power than the second category (adjusted R2=0.61), because
CE2 was included as a predictor in the first category and
excluded from the second. Each model in the first category
was a nonlinear combination of four variables. The first
three variables were the wind direction while the vehicle
was on Link 3 (WD3), the log of the average CO
concentration inside the vehicle while it was on Link 2
(log CE2), and the log of wind speed while the vehicle was
on Link 3 (log WS3). For the fourth variable, the models
included the log of one of three mathematically related
variables for Link 3: either the vehicle's travel time (log
TT3), the vehicle's average speed (log VS3), or the CO
255
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
emission factor (log EF3). Equations 30 through 32 are the
three models:
log CE3 ˆ 1:449 ‡ 0:346…log CE2 †
‡ 0:328…log AC3 †;
0:590…log VS3 †
…37†
log CE3 ˆ 0:534 ‡ 0:292…log CE2 † ‡ 0:700…log TT3 †
‡ 0:104…WD3 † 0:356…log WS3 †;
…30†
log CE3 ˆ
0:308 ‡ 0:352…log CE2 † ‡ 0:589…log EF3 †
‡ 0:331…log AC3 †:
…38†
log CE3 ˆ 1:913 ‡ 0:292…log CE2 † 0:700…log VS3 †
‡ 0:104…WD3 † 0:356…log WS3 †;
…31†
log CE3 ˆ
0:157 ‡ 0:295…log CE2 † ‡ 0:696…log EF3 †
‡ 0:112…WD3 † 0:365…log WS3 †:
…32†
The second category had the same independent variables
as the first category except for log CE2. Equations 33
through 35 are the three models:
log CE3 ˆ 0:665 ‡ 1:036…log TT3 † ‡ 0:145…WD3 †
0:524…log WS3 †;
log CE3 ˆ 2:705 1:036…log VS3 † ‡ 0:145…WD3 †
0:524…log WS3 †;
log CE3 ˆ
…33†
In the second category, the log of travel time on Link 2
(log TT2) replaced cabin exposure on Link 2 (log CE2) as
the first independent variable of each model. The remaining
variables of the second category are identical to those of the
first category. Equations 39 through 41 are the three models:
log CE3 ˆ 0:428 ‡ 0:425…log TT2 † ‡ 0:593…log TT3 †
‡ 0:432…log AC3 †;
…39†
log CE3 ˆ 1:597 ‡ 0:425…log TT2 †
‡ 0:432…log AC3 †;
0:593…log VS3 †
…40†
…34†
log CE3 ˆ
0:366 ‡ 1:037…log EF3 † ‡ 0:158…WD3 †
0:539…log WS3 †:
0:158 ‡ 0:437…log TT2 † ‡ 0:585…log EF3 †
‡ 0:437…log AC3 †:
…41†
…35†
Models Based on 62 Trips For the second group, the first
category (adjusted R2=0.73) had higher predictive power
than the second category (adjusted R2=0.67), again because
CE2 was included as a predictor in the first category and
excluded from the second. Each model in the first category
was a nonlinear combination of three variables. Two of the
variables were the log of the average CO concentration
inside the cabin while the vehicle was on Link 2 (log CE2)
and the log of the ambient CO concentration for Link 3 (log
AC3). The third variable included the log of one of the three
related variables while the vehicle was on Link 3: either the
vehicle's travel time (log TT3), average speed (log VS3), or
the CO emission factor (log EF3). Equations 36 through 38
are the three models:
log CE3 ˆ 0:287 ‡ 0:346…log CE2 † ‡ 0:590…log TT3 †
‡ 0:328…log AC3 †;
…36†
256
Models of Explanatory Variables
Theory and previous empirical studies suggested that many
of the explanatory variables could be interrelated. As shown
above, the two temporal variables (i.e., link-entry time and
the serial day of the field survey) were relatively weak
predictors of passenger cabin CO exposure. However,
further analysis revealed that each was a stronger predictor
of certain other variables (e.g., travel time, wind speed) that
were directly related to exposure. This analysis enabled a
better understanding of the context of cabin exposure for the
study site. Hence, models of explanatory variables are
presented below.
Link-Entry Time As expected, the time when the test vehicle
entered one link predicted when it entered the next link of
the highway extremely well. Equation 42 (multiple
R2=0.99) predicted Link 2's entry time (ET2) based on
Link 1's entry time (ET1), and Equation 43 (multiple
R2=0.99) predicted Link 3's entry time (ET3) based on Link
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
Flachsbart
Table 5. Multivariate models of in-vehicle exposure.
Equation
Multiple R2
Adjusted R2
Independent variables
0.701
0.685
log (average in-vehicle CO concentration for Link 2)
n=80 trips
30
log (vehicle's travel time on Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
31
0.701
0.685
log (average in-vehicle CO concentration for Link 2)
log (vehicle's average speed on Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
32
0.698
0.682
log (average in-vehicle CO concentration for Link 2)
log (exhaust CO emission factor for Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
33
0.629
0.614
log (vehicle's travel time on Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
34
0.628
0.614
log (vehicle's average speed on Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
35
0.624
0.609
log (exhaust CO emission factor for Link 3)
wind direction while vehicle was on Link 3
log (wind speed while vehicle was on Link 3)
n=62 trips
36
0.740
0.726
log (average in-vehicle CO concentration for Link 2)
log (vehicle's travel time on Link 3)
log (ambient CO concentration for Link 3)
37
0.740
0.726
log (average in-vehicle CO concentration for Link 2)
log (vehicle's average speed on Link 3)
log (ambient CO concentration for Link 3)
38
0.740
0.726
log (average in-vehicle CO concentration for Link 2)
log (exhaust CO emission factor for Link 3)
39
0.685
0.669
log (vehicle's travel time on Link 2)
log (ambient CO concentration for Link 3)
log (vehicle's travel time on Link 3)
log (ambient CO concentration for Link 3)
40
0.685
0.669
log (vehicle's travel time on Link 2)
log (vehicle's average speed on Link 3)
41
0.684
0.667
log (ambient CO concentration for Link 3)
log (vehicle's travel time on Link 2)
log (exhaust CO emission factor for Link 3)
log (ambient CO concentration for Link 3)
2's entry time (ET2). These results are consistent with
common sense:
ET2 ˆ 8:57 ‡ 0:973…ET1 †;
…42†
ET3 ˆ 5:37 ‡ 0:979…ET2 †:
…43†
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Knowledge of the study site suggested that entry time on
link i might predict travel time on link i or even link i+1.
Hence, models of travel time on link i (TTi) based on time of
entry onto link i (ETi) were developed. These models are
shown by Equation 44 for TT1, by Equations 45 and 46 for
TT2, and by Equation 47 for TT3. Although the models were
statistically significant, their explanatory powers varied
257
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
greatly (multiple R2=0.16 for Equation 44; multiple R2=0.29
for Equation 45; multiple R2=0.32 for Equation 46; and
multiple R2=0.63 for Equation 47). The predictive powers
of alternative models with constants exceeded the predictive
powers of Equations 45 and 46, but the alternatives
predicted negative travel times for entry times that fell
within certain time intervals of the commuting period:
TT1 ˆ 0:235…ET1 † 0:00241…ET1 †2
‡ 0:00000659…ET1 †3 ;
…44†
Vehicle Speed Equation 50 shows that the inverse of test
vehicle speed on Link 3 (1/VS3) could be used to estimate
the CO emission factor for Link 3 (EF3). This model had
very high predictive power (multiple R2=0.99). In effect, the
CO emission factor for Link 3 was virtually a deterministic
function of average vehicle speed on Link 3. Hence, travel
time, vehicle speed, and the CO emission factor for Link 3
could be viewed as interchangeable variables in terms of
their ability to predict passenger cabin CO exposure on Link
3:
EF3 ˆ 936:57=VS3 :
TT2 ˆ 0:1157…ET2 † 0:0010386…ET2 †2
‡ 0:00000234…ET2 †3 ;
…45†
TT2 ˆ 0:126…ET1 † 0:0012216…ET1 †2
‡ 0:00000298…ET1 †3 ;
…46†
TT3 ˆ
29:7 ‡ 1:02…ET3 † 0:008258…ET3 †2
‡ 0:00001964…ET3 †3 :
…47†
ET3 was a more powerful predictor of TT3 than of CE3.
This can be seen by comparing the predictive power of
Equation 47 (multiple R2=0.63) with the power of Equation
24 (multiple R2=0.23). Using differential calculus on
Equation 47, one can determine the link-entry times
associated with the predicted maximum and minimum
travel times for Link 3. The predicted maximum travel time
(9.87 min) occurred at 7:34 a.m., and the predicted
minimum travel time (2.99 min) occurred at 9:02 a.m.
These link-entry times are virtually identical to the linkentry times associated with the predicted maximum and
minimum cabin exposures for Link 3 that were discussed
previously for Equation 24.
Travel Time Traffic flow theory suggested that link travel
times were correlated, and that travel time on one link could
predict travel time on the next link. Hence, linear models of
travel time on link 1 were developed to explain travel time
on link i+1. Equation 48 (multiple R2=0.22) predicted travel
time on Link 2 (TT2) based on travel time on Link 1 (TT1),
and Equation 49 (multiple R2=0.47) predicted travel time on
Link 3 (TT3) based on travel time on Link 2 (TT2):
TT2 ˆ 1:61 ‡ 0:28…TT1 †;
…48†
TT3 ˆ 4:16 ‡ 1:14…TT2 †:
…49†
258
…50†
Seasonal Variable As noted previously, ambient temperature, traffic flow, and wind speed all varied over the 6-month
survey period. These three variables could be predicted
from the serial day of the field survey (SD), as shown by
Equation 51 for ambient temperature (AT3), Equation 52 for
traffic flow (TF3), and by Equation 53 for wind speed
(WS3). Equation 52 was based on only 21 of the 80 trips due
to missing traffic flow data and gave unrealistic estimates of
traffic flow beyond the 179th day (April 28, 1982) of the
field survey. All three models had F values that were
statistically significant (p0.005), and each model had
some predictive power (multiple R2=0.15 for Equation 51;
multiple R2=0.52 for Equation 52; and multiple R2=0.34 for
Equation 53):
AT3 ˆ 74:36
0:108…SD† ‡ 0:00055…SD†2 ;
TF3 ˆ 913:9 9:94…SD† ‡ 0:24954…SD†2
0:0012413…SD†3 ;
WS3 ˆ 8:37 ‡ 0:271…SD†
‡ 0:000016…SD†3 :
…51†
…52†
0:00425…SD†2
…53†
Using differential calculus, the predicted maximum and
minimum traffic flows and wind speeds were determined for
Equations 52 and 53, respectively. For Equation 52, the
predicted traffic flow was highest (1188 vehicles/15 min) on
the 110th serial day (February 18, 1982), and lowest (802
vehicles/15 min) on the 24th day (November 24, 1981) of
the survey period. For Equation 53, the predicted wind
speed was highest (13.4 mph) on the 41st day (December
11, 1981), and was lowest (5.9 mph) on the 140th day
(March 17, 1982). Hence, the effects of traffic flow and
wind speed appeared to reinforce each other on a seasonal
basis. Lighter traffic flows and stronger winds contributed to
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
relatively lower cabin exposures during late fall; and heavier
traffic flows and weaker winds led to higher exposures
during winter and spring. These expectations corresponded
fairly well with the predictions made previously from
Equation 25, which was the model of cabin exposure based
directly on the serial day of the field survey.
Meteorological Variables Four of the meteorological
variables affected ambient CO concentrations at the fixedsite monitor while the test vehicle was on Link 3 (AC3).
Equations 54 through 57 show these relationships. Each
model was based on ambient data for 64 trips, except
Equation 55 which was based on data for only 29 trips. That
was because the definition of its independent variable was
restricted to measured precipitation greater than zero at the
airport for the survey date. Wind speed (WS3) had the most
predictive power (multiple R2=0.31), followed by the
measured precipitation (MP) variable (multiple R2=0.28).
The inverse of ambient temperature (1/AT3) and atmospheric pressure (AP3) had identical predictive power
(multiple R2=0.15):
AC3 ˆ 2:82
0:33…WS3 † ‡ 0:012…WS3 †2 ;
…54†
AC3 ˆ 0:89 ‡ 1:61…MP†;
…55†
AC3 ˆ
3:83 ‡ 344:42=…AT3 †;
…56†
AC3 ˆ
0:83 ‡ 0:037…AP3 †
0:00015…AP3 †2 :
…57†
While the test vehicle was on Link 3, it was also possible
to show that wind speed (WS3) varied as a function of
ambient temperature (AT3) as shown by Equation 58
(multiple R2=0.33), and wind direction (WD3) as shown
by Equation 59 (multiple R2=0.11). These models implied
that wind speeds were stronger during higher ambient
temperatures and during southerly winds:
WS3 ˆ 6:910 9 …AT3 †4:908 ;
…58†
WS3 ˆ 8:30 ‡ 3:47…WD3 †:
…59†
Although ambient temperatures are important inputs to
the Mobile4.1 model, which was used to generate CO
emission factors, a statistical model of emission factors
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)
Flachsbart
based solely on ambient temperatures could not be
developed for this study.
Conclusions
This paper presented statistical models of passenger
exposure to CO inside a vehicle during morning travel
periods on three links of a coastal artery in Honolulu,
Hawaii. Like other studies, the models showed that certain
factors, apart from passenger smoking which was not
studied, affect CO exposure inside a vehicle. Cabin
exposure on a given link was a powerful albeit impractical
predictor of cabin exposure on the next link. This serial
correlation of exposure measurements on three links most
likely was caused by traffic conditions affecting the entire
study site. This evidence of serial correlation supports a
similar finding made by Ott et al. (1994).
The models showed that cabin exposure was strongly
affected by travel time and average vehicle speed, which
were assumed to be indirect measures of traffic flow. These
measures in turn were affected by the time that the test
vehicle entered each link of the highway. This implied that
stochastic simulations of exposure (e.g., the SHAPE model
by Ott et al., 1988) should not assume that trip times and
commuter exposures are independent of trip-starting times,
as such assumptions could distort estimates of commuter
exposure.
Using the test vehicle's speed as an indicator of the
average speed of surrounding traffic, this study derived CO
exhaust emission factors to build models of passenger cabin
CO exposure based solely and directly on these factors. In
this study, travel speed rather than ambient temperature was
the dominant factor affecting CO emissions. The study
presented various univariate and multivariate models that
directly related cabin CO exposures on different links of the
study site to various predictor variables. Due to serial
correlation, multivariate models of cabin exposure on the
third link that included second-link exposure among the
independent variables had more explanatory power than
models that did not. When second-link exposure was
excluded as an independent variable, the most powerful
multivariate models were based on nonlinear combinations
of the ambient CO concentration, the second-link travel
time, and either the travel time, vehicle speed or CO
emission factor for the third link.
Ott et al. (1994) reported that passenger cabin CO
concentrations were higher during winter than in summer on
an arterial highway in California. This was attributed
primarily to seasonal temperature changes in the weather
rather than to seasonal variation in traffic conditions. In
California, cold winter temperatures increased `cold-start'
CO emissions, and warm summer temperatures reduced
259
Flachsbart
Models of exposure to carbon monoxide inside a vehicle on a Honolulu highway
those emissions. In Hawaii, temperatures were never cold
enough during the winter to increase CO emissions from
motor vehicles and passenger cabin exposures. Likewise,
mild temperatures contributed to low ambient CO levels
throughout the study period.
In this study, seasonal variation in cabin exposures
occurred primarily because of seasonal variations in traffic
flows and wind speeds. Relatively lighter traffic flows and
stronger winds lowered passenger cabin exposures during
late fall, and heavier traffic flows and weaker winds elevated
cabin exposures during winter months. Wind direction was
an important factor, also. Northerly winds were prevalent
during most of the study period. These winds reduced cabin
exposures by dispersing emissions on westbound lanes of
the study site where exposures were measured, and
southerly winds increased exposures by sending emissions
from eastbound vehicles to the westbound lanes of the study
site. Since these effects were unanticipated at the outset of
this study, future studies may wish to explore how these
factors combine to raise or lower exposure.
Acknowledgments
This paper is based on data that were collected through a
Cooperative Agreement CR 808541-01-3 between the US
Environmental Protection Agency and the University of
Hawaii at Manoa. Subsequent data analysis, model development, and paper preparation were at the author's personal
expense. The author wishes to thank two anonymous
reviewers who provided constructive comments on a draft
manuscript. Any remaining errors or misstatements in the
paper are the author's sole responsibility. Also, any
reference to brand name products should not be construed
as an official endorsement by the author or the US
Environmental Protection Agency.
References
Akland G.G., Hartwell T.D., Johnson T.R., and Whitmore R.W. Measuring
human exposure to carbon monoxide in Washington, DC, and Denver,
Colorado, during the winter of 1982±83. Environ. Sci. Technol. 1985:
19: 911±918.
Baerwald J. (Ed.). Transportation and Traffic Engineering Handbook.
Institute of Transportation Engineers. Prentice-Hall, Englewood Cliffs,
NJ, 1976, p. 427.
Department of General Planning. Preliminary report on standards and
260
controls related to conditions along major highways. City and County
of Honolulu, Honolulu, HI, 1982.
Draper N., and Smith H. Applied Regression Analysis, 2nd edn. John
Wiley and Sons, New York, NY, 1981.
Edwards J., Jr. (Ed.). Transportation Planning Handbook. Prentice-Hall,
Englewood Cliffs, NJ, 1992, 43±44.
Flachsbart P.G. Prototypal models of commuter exposure to CO from
motor vehicle exhaust. Paper No. 85-39.6 presented at the 78th Annual
Meeting of the Air Pollution Control Association, Detroit, MI, 1985.
Flachsbart P.G. Effectiveness of priority lanes in reducing travel time and
carbon monoxide exposure. Inst. Trans. Eng. J. 1989: 59: 41±45.
Flachsbart P.G. Long-term trends in United States highway emissions,
ambient concentrations, and in-vehicle exposure to carbon monoxide
in traffic. J. Exposure Anal. Environ. Epidemiol. 1995: 5: 473±495.
Flachsbart P.G. Statistical models of exposure to carbon monoxide inside
three vehicles on two Honolulu highways. Department of Urban and
Regional Planning, University of Hawaii at Manoa, Honolulu,
HI, 1998.
Flachsbart P.G., and Brown D.E. Surveys of personal exposure to vehicle
exhaust in Honolulu microenvironments. Department of Urban and
Regional Planning, University of Hawaii at Manoa, Honolulu,
HI, 1985.
Laconti A.B., Nolan M.E., Kosek J.A., and Sedlak J.M. Recent
developments in electrochemical solid polymer electrolyte sensor cells
for measuring carbon monoxide and oxides of nitrogen. In: Chemical
Hazards in the Workplace. American Chemical Society, Publication
No. 0097-6156-81-0149-0551, Washington, DC, 1981, 551±573.
Lawryk N.J., Lioy P.J., and Weisel C.P. Exposure to volatile organic
compounds in the passenger compartment of automobiles during
periods of normal and malfunctioning operation. J. Exposure Anal.
Environ. Epidemiol. 1995: 5: 511±531.
Ott W., and Willits N. CO exposures of occupants of motor vehicles:
modeling the dynamic response of the vehicle. SIMS Technical Report
No. 48. Department of Statistics, Stanford University, Stanford,
CA, 1981.
Ott W., Thomas J., Mage D., and Wallace L. Validation of the Simulation
of Human Activity and Pollutant Exposure (SHAPE) model using
paired days from the Denver, Colorado, carbon monoxide field study.
Atmos. Environ. 1988: 22: 2101±2213.
Ott W., Switzer P., and Willits N. Carbon monoxide exposures inside an
automobile traveling on an urban arterial highway. J. Air Waste
Manage. Assoc. 1994: 44: 1010±1018.
Papacostas C. Fundamentals of Transportation Engineering. Prentice-Hall,
Englewood Cliffs, NJ, 1987: 92±98, 346±347.
Robinson N.F., Pierson W.R., Gertler, A.W., and Sagebiel J.C. Comparison
of Mobile4.1 and Mobile5 predictions with measurements of vehicle
emission factors in Fort McHenry and Tuscarora mountain tunnels.
Atmos. Environ. 1996: 30: 2257±2267.
Sexton K., and Ryan P.B. Assessment of human exposure to air pollution:
methods, measurements, and models. In: Watson A.Y., Bates R.R.,
Kennedy D. (Eds.), Air Pollution, the Automobile, and Public Health.
National Academy Press, Washington, DC, 1988: 207±238.
US Environmental Protection Agency. User's Guide to Mobile4.1 (Mobile
Source Emission Factor Model), EPA AA-TEB-91-01. Office of Air
and Radiation, Office of Mobile Sources, Ann Arbor, MI, 1991.
Yu L.E., Hildemann L.M., and Ott W.R. A mathematical model for
predicting trends in carbon monoxide emissions and exposures on
urban arterial highways. J. Air Waste Manage. Assoc. 1996: 46:
430±440.
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(3)