You are on page 1of 13

Science of the Total Environment 506507 (2015) 409421

Contents lists available at ScienceDirect

Science of the Total Environment


journal homepage: www.elsevier.com/locate/scitotenv

Health impact assessment of trafc-related air pollution at the urban


project scale: Inuence of variability and uncertainty
Chidsanuphong Chart-asa a, Jacqueline MacDonald Gibson b,
a
b

Institute for the Study of Natural Resources and Environmental Management, Mae Fah Luang University, Chiang Rai, Thailand
Department of Environmental Sciences and Engineering, Gillings School of Global Public Health, University of NC, Chapel Hill, USA

a r t i c l e

i n f o

Article history:
Received 2 December 2013
Received in revised form 30 October 2014
Accepted 5 November 2014
Available online 26 November 2014
Editor: Lidia Morawska
Keywords:
PM2.5
Trafc
Health impact assessment
Variability
Uncertainty

a b s t r a c t
This paper develops and then demonstrates a new approach for quantifying health impacts of trafc-related particulate matter air pollution at the urban project scale that includes variability and uncertainty in the analysis. We
focus on primary particulate matter having a diameter less than 2.5 m (PM2.5). The new approach accounts for
variability in vehicle emissions due to temperature, road grade, and trafc behavior variability; seasonal variability in concentrationresponse coefcients; demographic variability at a ne spatial scale; uncertainty in air quality model accuracy; and uncertainty in concentrationresponse coefcients. We demonstrate the approach for a
case study roadway corridor with a population of 16,000, where a new extension of the University of North
Carolina (UNC) at Chapel Hill campus is slated for construction. The results indicate that at this case study site,
health impact estimates increased by factors of 49, depending on the health impact considered, compared to
using a conventional health impact assessment approach that overlooks these variability and uncertainty
sources. In addition, we demonstrate how the method can be used to assess health disparities. For example, in
the case study corridor, our method demonstrates the existence of statistically signicant racial disparities in exposure to trafc-related PM2.5 under present-day trafc conditions: the correlation between percent black and annual attributable deaths in each census block is 0.37 (t(114) = 4.2, p b 0.0001). Overall, our results show that the
proposed new campus will cause only a small incremental increase in health risks (annual risk 6 1010; lifetime
risk 4 108), compared to if the campus is not built. Nonetheless, the approach we illustrate could be useful for
improving the quality of information to support decision-making for other urban development projects.
2014 Elsevier B.V. All rights reserved.

1. Introduction
In the United States, nonprot organizations and public health practitioners increasingly advocate for formal health impact assessments
(HIAs) to inform regional and local land-use and transportation planning decisions (Wernham, 2011; Bhatia and Corburn, 2011). Signaling
the heightened interest in HIAs, the U.S. National Academy of Sciences
in 2011 published a report, Improving Health in the United States: The
Role of Health Impact Assessment, concluding that HIA is a particularly
promising approach for integrating health implications into decisionmaking (National Research Council, 2011). The report offered the following formal denition of HIA:
HIA is a systematic process that uses an array of data sources and analytic methods and considers input from stakeholders to determine

Corresponding author at: Department of Environmental Sciences and Engineering,


Gillings School of Global Public Health, University of North Carolina, Campus Box 7431,
Chapel Hill, NC, 27599-7431, USA. Tel.: +1 919 966 7892.
E-mail addresses: chidsanuphong@gmail.com (C. Chart-asa), jackie.macdonald@unc.
edu (J.M. Gibson).

http://dx.doi.org/10.1016/j.scitotenv.2014.11.020
0048-9697/ 2014 Elsevier B.V. All rights reserved.

the potential effects of a proposed policy, plan, program, or project


on the health of a population and the distribution of those effects
within the population. HIA provides recommendations on monitoring and managing those effects.
As the National Academies' report explains, the increasing demand for
HIAs in the United States is driven by the growing recognition that reducing obesity and chronic disease rates will require substantial changes to
decision-making processes in arenas outside the traditional healthcare
sector. For example, decisions by transportation and municipal planning
organizations can promote or limit opportunities for physical activity
and can exacerbate or decrease exposure to ambient air pollution.
While HIAs of local and regional decisions have been used in Europe,
Australia, Canada, and Thailand for decades, the rst U.S. HIA of a local
project was completed in 1999 by the San Francisco Department of Public
Health (Bhatia and Corburn, 2011; National Research Council, 2011). By
the end of 2012, however, at least 115 HIAs of local or regional U.S. projects had been completed, and another 64 were under way (SingletonBaldrey, 2012). Of the completed HIAs, 70 (more than 60%) focused on
proposed local or regional changes to the built environment and/or transportation networks (Singleton-Baldrey, 2012; Dannenberg et al., 2008).

410

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

To facilitate comparison of alternatives and guide decision-making,


HIAs ideally would provide quantitative estimates of the health outcomes of the decision options under consideration. That is, they would
estimate the number of deaths and illnesses prevented or caused by
each alternative. This information could be used to quantify the health
costs (positive or negative) of each alternative. Quantication can present health impacts more concisely (as numerical summaries) than
lengthy qualitative discussions. In addition (rightly or wrongly), quantitative assessment can lend legitimacy to the analysis. Furthermore,
some federal and state regulations require quantitative costbenet
analyses (Federal Highway Administration, 2003). However, only 5 of
the 70 U.S. HIAs focusing on local or regional transportation projects
carried out prior to 2013 quantied the expected health impacts
(Singleton-Baldrey, 2012; Bhatia and Seto, 2011). Table 1 summarizes
these HIAs. The remaining HIAs expressed qualitative conclusions.
The Aerotropolis Atlanta Browneld Redevelopment HIA (Ross et al.,
2011) illustrates the qualitative approach used by most previous local
and regional U.S. HIAs. This HIA evaluated a plan to convert a former
Ford assembly plant near Atlanta, Georgia, to a new community called
Aerotropolis Atlanta. The HIA's analysis of air quality impacts was
based on a review of previous studies (not associated with this project)
of trafc impacts on air quality and health. It concluded, Aerotropolis
may lead to a change in trafc volume around the site , potentially
impacting people who live, work, or visit within the air-shed of the affected streets. The HIA recommended several mitigation measures, including congestion pricing, increased public transit, zoning of sensitive
uses away from roadways, and vegetation buffers around roadways.
However, the HIA did not quantify the air quality or health impacts of
the proposed new development or these mitigation alternatives.

While the above-mentioned ve previous quantitative HIAs estimate the magnitude of air quality and related health impacts, none considers the potential variability and uncertainty of the estimates. Rather,
these HIAs each provide a single, deterministic prediction of health impacts for each decision option (see Table 1). In so doing, these HIAs not
only convey a potentially misleading degree of certainty but also neglect to provide decision-makers with information about the plausible
range of impacts. U.S. Environmental Protection Agency guidance
documents indicate that health risk assessments of national and state
policies should include sensitivity and uncertainty analyses (U.S.
Environmental Protection Agency, 2001). Indeed, sensitivity and uncertainty analyses are cornerstones of health impact estimates the agency
prepares to inform national policy decisions, such as changes to air pollution standards (US Environmental Protection Agency, 2010). Nonetheless, current U.S. local-level HIAs do not report variability and
uncertainty in their health impact estimates.
The reliance of local HIA practitioners on deterministic estimates is a
major limitation for several reasons. First, it fails to consider the full
range of potential risksthat is, the potential for risks at the tails of
the risk distribution. For example, vulnerable populations are often at
the upper tails, not the centers, of the exposure and effect distributions
(Fann et al., 2011). Second, risk estimates relying only on central tendencies of each input variable may differ from those considering the
full distributions of each input variable. Except in special cases, the expected value of a function of random variables is not the same as the
function applied to the expected values of each variable. Third, deterministic approaches ignore the potential dependencies among model
input variables (for example, dependencies in meteorological characteristics used to estimate pollutant dispersion). Fourth, deterministic

Table 1
Previous quantitative transportation-related HIAs in the United States.
Study area
population

Title

Project scenario
analyzed

Trafc-related
air pollutants
considered

Pittsburg Railroad Avenue specic plan


HIA (Human Impact Partners, 2008)

Construction of new Bay Area Rapid


Transit (BART) station and mixed-use
village in Pittsburg, CA, including 1600
housing units and 450,000 sq. ft. of retail,
commercial, and public service spaces

PM2.5

Evaluating the healthfulness of


affordable housing opportunity
sites along the San
Pablo Avenue Corridor using
HIA (Human Impact Partners, 2009)
Oak to Ninth Avenue HIA (UC Berkeley
Health Impact Group, 2006)

Construction of affordable housing sites


in El Cerrito and Richmond, CA

PM2.5

1,000,000

Development of new waterfront community


in Oakland, CA, including 3100 housing units
and 200,000 sq. ft. of retail, commercial,
and public service spaces

PM10

10,000

MacArthur BART Transit Village HIA


(UC Berkeley Health Impact Group,
2007)

Redevelopment of parking lot into a


mixed-use village in Oakland, CA,
including 625 housing units and
30,000 sq. ft. of retail, commercial,
and public service spaces

PM2.5

100,000

Health impact assessment of the Port of


Oakland (UC Berkeley Health Impact
Group, 2010)

Ongoing growth of port operations in West


Oakland, CA

PM2.5

22,000

4770

Estimated annual health impactsa

6 deaths (age 30) from long-term exposures,


= 0.0046 (0.0034, 0.0058)
5 hospital admissions for asthma (age b 65)
from short-term exposures, = 0.0025
(0.0015, 0.0036)
12 lower respiratory symptom days (ages 714)
from short-term exposures, = 0.0182
(0.0124, 0.0241)
3341 deaths (all ages) from long-term exposures,
RR = 1.014 (no report of 95% condence interval)

0.8 deaths (age 30) from long-term exposures,


= 0.0046 (0.0034, 0.0058)
0.4 chronic bronchitis cases (age 27) from
long-term exposures, = 0.0132 (0.0064, 0.0200)
10.6 emergency room visits for asthma (age b 65)
from short-term exposures, = 0.0037
(0.0024, 0.0049)
2.7 deaths (age 30) from long-term exposures,
= 0.0046 (0.0034, 0.0058)
1.0 chronic bronchitis cases (age 27) from
long-term exposures, = 0.0132 (0.0064, 0.0200)
34.2 acute bronchitis cases (ages 8-12) from
short-term exposures, = 0.0272 (0.0101, 0.0443)
0.1 hospital admissions for asthma (age b 65)
from short-term exposures, = 0.0025
(0.0015, 0.0036)
26.9 lower respiratory symptom days (ages 714)
from short-term exposures, = 0.0182 (0.0124,
0.0241)
1.3 deaths (age 30) from long-term exposures,
= 0.0046 (0.0034, 0.0058)

= concentrationresponse coefcient used to estimate health impacts; RR = relative risk used to estimate health impacts.

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

estimates fail to provide information to decision-makers about the degree of certainty in the estimated risks. For example, decision-makers
may be more concerned about a risk factor with a relatively low central
risk estimate (for example, 1 in 10,000) if there is a good chance that the
risk could be much higher than the central estimate (for example, a 10%
chance of the risk exceeding 1 in 100) than they would be if presented
only with the central estimate of risk.
Variability and uncertainty in estimated risks of trafc-related air
pollution can arise from multiple sources. Variability arises naturally
due to differences in members of a population, weather patterns, trafc,
geographic features, and so on; it is a property of nature, usually not reducible through further measurement or study (Frey and Burmaster,
1999). On the other hand, uncertainty arises due to the lack of information or knowledge, including limited data on a population, partial ignorance of phenomena inuencing a particular risk, and disagreements
between models and the reality they are intended to represent (Frey
and Burmaster, 1999). Example sources of uncertainty include the
mathematical form used to predict the effects of changes in pollution
exposure on public health, the parameters in such mathematical equations, and the accuracy of models predicting air pollution levels under
different trafc scenarios. Theoretically, uncertainty can be decreased
through further studies.
This paper aims to strengthen the knowledge base and tool set available to HIA practitioners wishing to incorporate variability and uncertainty in quantitative, transportation-related HIAs. Like the HIAs listed
in Table 1, this analysis focuses on a potential new land development
expected to increase future trafc on a major municipal road corridor
(see the Case study site section). The potential for increased trafc
has raised concerns about increases in air pollution and its associated
adverse health effects, including increased risks of cardiovascular and
respiratory diseases. Like four of the ve HIAs in Table 1, the analysis focuses on airborne particulate matter having a diameter less than 2.5 m
(denoted as PM2.5) as an indicator of trafc-related air pollution. Like
the other HIAs, this analysis is restricted to primary PM2.5 (that is,
PM2.5 emitted directly by vehicle operations rather than that formed
by chemical reactions in the atmosphere). This study considers the effects of short-term exposure to trafc-related PM2.5 on cardiovascular
and respiratory mortality (all ages) and unscheduled hospital admissions (age 65 and over). These health outcomes were previously selected for the core analysis in the U.S. Environmental Protection Agency's
(EPA's) quantitative health risk assessment for supporting the review
of the U.S. National Ambient Air Quality Standards for PM (U.S.
Environmental Protection Agency, 2010).
We use the case study road corridor to explore the effects on health
impact estimates of PM2.5 from roadway trafc when including or excluding various sources of variability and uncertainty. We rst use a
portion of the road corridor to explore the question, Which variability
and uncertainty sources have the greatest effects on the mean values
and upper condence limits of estimated health risks? Then, we demonstrate a method for incorporating the key variability and uncertainty
sources in a comprehensive assessment of potential air pollutionrelated health risks for the entire case study roadway corridor under
current conditions and future conditions with and without the proposed new development.
2. Case study site
We demonstrate the suggested new assessment process to explore
some of the potential health impacts arising from a planned new campus extension for the University of North Carolina (UNC) at Chapel
Hill. The new campus, called Carolina North, is intended to increase
the university's capability to translate research into applications. It will
be located about 3 km (2 miles) north of the existing campus (Fig. 1).
If constructed, it is expected to increase the number of trips to the
area by 10,000 per day by 2015, with most of the increases expected
to occur along Martin Luther King, Jr., Blvd., the main link to the existing

411

campus and the major northsouth road corridor in Chapel Hill (Vanasse
Hangen Brustlin Inc., 2009). By 2025, the number of additional daily trips
to the campus is expected to increase by as many as 40,000. We consider
the potential impacts of the expected additional trafc-related air pollution among residents living in census blocks within 500 m of Martin Luther King, Jr., Blvd. In all, this area encompasses 160 U.S. census blocks
(see Fig. 1) and has a total population of about 16,000more than onequarter of Chapel Hill's total population of 57,000.
We analyze the effects of primary emissions from trafc along Martin Luther King, Jr., Blvd. on ambient PM2.5 concentrations and population health under three different scenarios: (1) the year 2009,
(2) 2025, assuming the new campus is not built, and (3) 2025, with
the new campus. The baseline comparison year is 2009, because the
most comprehensive transportation analysis of the study corridor was
conducted using 2009 data (Vanasse Hangen Brustlin Inc., 2009).
Table 2 provides summary information about the population size and
trafc volumes under these three scenarios.
3. Methods and data sources
This analysis has two main parts:
1) Analyze the effects of including variability and uncertainty in the HIA:
First, we investigate in the effects on health impact estimates of including several different uncertainty and variability sources, as compared to results obtained using the conventional deterministic
approach. For computational efciency, we focus on the 12 census
blocks highlighted in Fig. 1B, which our prior air quality modeling indicated are more vulnerable to trafc-related PM2.5 than most other
census blocks in the corridor (Chart-asa et al., 2013). The total population in the 12 blocks is 1117 (about 7% of the total population in
the study corridor).
2) Quantify the health impacts of trafc from the proposed new campus in
the study corridor: Second, we quantify trafc-related air quality and
health outcomes along the entire study corridor for the three development scenarios in Table 2. This analysis includes the variability
and uncertainty sources identied in part 1 as having an inuence
on the central estimates or upper condence estimates of the
modeled risks.
Both analyses use the same modeling framework, described in detail
in the following sections. However, the rst analysis introduces variability and uncertainty sources one at a time, in order to explore their potential inuence on the computed health risks, while the second
analysis includes all key variability and uncertainty sources.
3.1. Modeling framework overview
Quantifying the health impacts of trafc-related air pollution requires three categories of information: (1) estimates of the excess
PM2.5 concentrations to which the population is exposed as a result of
primary emissions from trafc, (2) concentrationresponse functions
relating exposure concentrations to probabilities of adverse health outcomes, and (3) incidence rates of the health outcomes of concern (from
all causes) in the exposed population (Ostro, 2004; Ostro and Chestnut,
1998; Cohen et al., 2005; Li et al., 2010). Fig. 2 summarizes how this
analysis combines these three information categories (shows as shaded
boxes) to estimate health impacts. The unshaded boxes show variability
and uncertainty sources considered in this study. The subscript notation
indicates that the analysis is conducted at the census block scale, where i
represents an individual block. That is, health risks are characterized
separately for each census block, considering variability in trafcrelated PM2.5 exposure concentrations and population demographic
characteristics within each block. The subscripts j, k, and l indicate differences in baseline health status by age (j), gender (k), and race (l).
In addition, this analysis considers seasonal (subscript m) variability,
because epidemiologic evidence suggests seasonal differences in

412

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

Fig. 1. (A) The study corridor between the intersection of Martin Luther King, Jr., Blvd. and Whiteld Rd. and the intersection of South Columbia St. and Mt. Carmel Church Rd., Chapel Hill,
NC, and the census blocks located within 500 m from the study corridor. (B) The road segment and census blocks for simulations to demonstrate differences in health burden estimates
when including variability and the uncertainty in the modeling approach. Dots represent census block centroids.

doseresponse functions (Zanobetti and Schwartz, 2009; Bell et al.,


2008). The following sections provide details on our methods for estimating PM2.5 exposure concentrations (left-most shaded diamond in
Fig. 2), selecting concentrationresponse functions (central diamond),
estimating baseline incidence rates of adverse health outcomes in the
study population (right-most diamond), and incorporating variability
and uncertainty (white rectangles) into the analysis.
3.2. PM2.5 concentrations attributable to primary emissions from trafc
The 24-hour exposures to PM2.5 arising from primary emissions
from trafc along the case study roadway corridor were estimated

using an integrated air quality modeling approach described in Chartasa et al. (2013). In brief, the approach employs standard trafc emissions and air quality dispersion modeling tools, but it adds a novel approach for modeling variability in vehicle emissions due to variability
in hourly temperature, road grade, and trafc behavior (including cruising speed and percent time spent idling, decelerating, and accelerating).
The exposure modeling approach links a novel application of MOVES
2010b, commonly used in the United States to estimate vehicle emissions factors (g/vehicle-mile), and CAL3QHCR, which characterizes
PM2.5 dispersion away from roadways. By linking these models and
employing a new approach for characterizing variability in emission factors, we simulated probability distributions of the average 24-hour

Table 2
Population size and trafc volumes under three scenarios considered.
Scenario

Trafc volumes of road segments on study corridor (veh/h)a Total population of 160 census blocks located within 500 m from study corridor

2009
41758
2025 without the new campus 52443
2025 with the new campus
52832
a
b

16,042
19,140b
19,140b

Ranges indicate variability in trafc ow by road segment, day of week, and time of day.
Computed from growth rates forecasted by the North Carolina Capital Area Metropolitan Planning Organization (2005).

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

413

Fig. 2. Overview of framework for incorporating variability and uncertainty into assessment of the health impacts of trafc-related PM2.5. The rectangles show sources of variability and
uncertainty. The shaded diamonds show the three major information categories needed for quantitative health impact assessment.

PM2.5 concentration in each season (winter, spring, summer, and fall) at


the centroid of each of the 160 census blocks in the study corridor.
Our analysis considers variability in vehicle emission factors by dividing the 8.2-km roadway corridor into 1200 links and estimating separate emission factors for each link for each hour of each simulation day
(Chart-asa et al., 2013). Unlike previous studies linking MOVES and
CAL3QHCR, our analysis considers hourly variability in temperature
and link-specic variability in road grade and vehicle behavior. Hourly
meteorological proles for 20062012 were obtained from the national
weather stations in Chapel Hill and Greensboro, North Carolina, and meteorological proles for input to CAL3QHCR were generated from EPA's
Meteorological Processor for Regulatory Models (NCDC, 2013; NOAA,
2013). The meteorological proles contained a total of 2100 days with
complete required data (525 days for winter, 560 days for spring,
532 days for summer, and 483 days for fall). For each census block, we
used CAL3QHCR to estimate the PM2.5 concentration (averaged over
24 h) attributable to primary trafc emissions from each of the 1200
roadway links for each of the 2100 days for which meteorological data
were available. Separate estimates were prepared for 2009 and 2025
using emission factors from MOVES modeling and simulated trafc
data for 2009 and 2025 scenarios with or without Carolina North from
the Transportation Impact Analysis (TIA) for the Carolina North Development (Vanasse Hangen Brustlin Inc., 2009). Then, for each development scenario, seven separate mean estimatesone for each of the
seven years for which meteorological data were availableof the seasonal mean value of the 24-hour average PM2.5 exposure concentrations
were computed for each season.
For each scenario (year 2009 and year 2025 with and without constructing Carolina North) and each season, we then computed bootstrap
estimates of the mean value and standard deviation of the seasonal daily

average PM2.5 exposure concentrations by randomly selecting one of


the seven years, assigning the associated seasonal mean concentrations
as computed using that year's data to each census block, and then repeating the process 1999 times. For each of the 160 census blocks, the
result was an estimated mean value and standard deviation of the seasonal 24-hour-average concentration of PM2.5 attributable to primary
emissions from trafc along the roadway corridor under each scenario.
Within each census block and for each scenario, the seasonal average
trafc-related 24-hour PM2.5 concentration then was represented as a
normal distribution (left-truncated at zero) with the mean and standard
deviation estimated from the corresponding 2000 bootstrap simulations. The TIA estimated hourly trafc counts for each scenario along
each roadway link only for weekdays; we assumed trafc counts on
weekends would be the same and hence may have slightly overestimated exposure concentrations.
In addition to considering variability in PM2.5 exposures arising from
primary trafc emissions, we assessed the effects of uncertainty in the
accuracy of the air quality model predictions. Our previous research
on the integrated air quality modeling approach, as well as previous
work by others, suggests that the combined MOVESCAL3QHCR
model generally predicts PM2.5 concentration within a factor of two of
measured concentrations (although accuracy varies with local conditions and the quality of data available to support the model) (Chartasa et al., 2013; Yura et al., 2007). Following Morgan and Henrion's guidance (Morgan et al., 1990), we represented model uncertainty with an
uncertainty factor (UF) parameterized by a triangular probability distribution with lower limit = 0.5, upper limit = 2.0, and mode = 1.0
(spanning the expected factor-of-two uncertainty in the model). According to Morgan and Henrion, the triangular distribution is especially
appropriate for situations in which the distributions of variables in a

414

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

model are not precisely known but in which values toward the middle
of the range of possible values are considered more likely to occur than
values near either extreme. Based on previous evaluations of the performance of near-roadway air pollutant dispersion models, the exact
form of the distribution representing model uncertainty is not known,
making the triangular distribution an appropriate choice for characterizing model uncertainty. Correspondingly, in each census block, the excess PM2.5 24-hour average exposure concentration attributable to
primary emissions from trafc on the case study roadway was estimated for each season as
PM exposurei;m U F  PM modeli;m

where PM exposurei;m represents the average 24-hour PM2.5 concentration


in census block i (i = 1160) and season m (m = winter, spring, summer, fall) attributable to primary trafc emissions on the case study
roadway; UF is the model uncertainty factor; and PM modeli;m is the corresponding model-predicted seasonal daily average PM2.5 concentration
arising from primary emissions from trafc.
3.2.1. Concentrationresponse functions
As recommended by the World Health Organization and others
(Ostro and Chestnut, 1998; Li et al., 2010; Aunan, 1996), we use
the following relationship to describe the link between seasonal
daily average PM2.5 concentrations and the relative risk of cardiovascular and respiratory health outcomes:
m;n PMexposure

RRi;m;n e

i;m

where m,n is the concentrationresponse coefcient describing the


effects of PM on health outcome n during season m and RRi,m,n is the
relative risk of health outcome n during season m in census block i.
The number of adverse health cases in the population attributable
to trafc-related PM2.5 then can be determined from the following
relationship:
0

yi; j;k;l;m;n yi; j;k;l;m;n  A F i; j;k;l;m;n

0
yi; j;k;l;m;n

0
yi; j;k;l;m;n

3a



RRi;m;n 1

3b

RRi;m;n


e m;n

PMexposure

i;m

m;n PMexposure

0
yi; j;k;l;m;n  1e m;n


3c

i;m

PMexposure


i;m

3d

where AFi,j,k,l,m,n and yi,j,k,l,m,n are the fraction and number of cases
of adverse health event n attributable to trafc-related PM2.5 in season m in census block i for age group j, gender k, and race l and
0
where yi,j,k,l,m,n
is the observed total number of cases in the same location and among the same population group. Eqs. (2), (3a), (3b),
(3c) and (3d) are the standard equations used in analyses by the
WHO and other organizations to attribute observed cases of adverse health events to specic risk factors (Ostro and Chestnut,
1998; Murray et al., 2003; Mathers et al., 2001; Prss-stn et al.,
2003).
The values in Eqs. (2), (3c) and (3d) (known as doseresponse coefcients) were drawn from the U.S. Environmental Protection Agency
guidance document, Quantitative Health Risk Assessment for Particulate
Matter (U.S. Environmental Protection Agency, 2010; Zanobetti and
Schwartz, 2009; Bell et al., 2008). Table 3 shows the coefcient values

used in this analysis. EPA retrieved these coefcients from peerreviewed epidemiologic studies that met certain quality-assurance
criteria, including, for example, the estimation of exposure from measured rather than modeled PM2.5 data. For mortality effects, the coefcients are specic to 15 U.S. metropolitan areas. For morbidity effects,
coefcients are specic to region (Northeast, Southeast, Northwest,
and Southwest). This study employed mortality coefcients developed
from studies in Atlanta, since Atlanta is climatologically the most similar
to Chapel Hill among the 15 cities studied. We used morbidity
coefcients for the Southeast region, in which Chapel Hill is located.
All concentrationresponse coefcients were represented as normal
distributions, with all negative values truncated at zero (to avoid associating PM exposure with positive health effects). Standard deviations for
each season and health outcome were estimated from the condence
intervals in Table 3.
3.3. Baseline incidence rates of adverse health outcomes
Data on baseline incidence rates of health outcomes were obtained
from North Carolina public health databases. Annual mortality rates
for each age group (Table 4) were calculated by dividing the total number of deaths in Orange County (where Chapel Hill is located) in 2010
(North Carolina State Center for Health Statistics, 2012) by the 2010 Orange County census population (Minnesota Population Center, 2011).
Annual unscheduled hospital admission rates (Table 5) were obtained
from 2009 emergency department visit data reported by the North Carolina Disease Event Tracking and Epidemiologic Collection Tool (NC DETECT) (University of North Carolina at Chapel Hill, 2011). We were
unable to obtain data on incidence rates by gender and race, so we assume that incidence rates are the same for both genders and all races
(which is a limitation of this analysis). It should be noted as well that
the ICD codes specic to the concentrationresponse coefcients
might not be entirely matched to the ICD codes specic to the incidence
rates used in this study, depending on reported data. Moreover, emergency department visits may not result in hospital admissions, and
some hospital admissions may occur without rst visiting the emergency department.
To reect seasonal variation, we adjusted the incidence rates for cardiovascular and respiratory mortality and unscheduled hospital admissions using data on temporal variability in cardiovascular and
respiratory deaths in Orange County during 19992010 from the CDC
WONDER database (Centers for Disease Control and Prevention,
2013). The fractions for cardiovascular events are 0.25, 0.31, 0.20, and
0.24 for winter, spring, summer, and fall respectively, while the fractions for respiratory events are 0.30, 0.26, 0.21, and 0.23 for winter,
spring, summer, and fall respectively.
To determine the total number of cases in any given season (i.e.,
0
yi,j,k,l,m,n
in Eqs. (3a), (3b), (3c) and (3d)), we multiplied the given incidence rate by the corresponding size of each demographic group in each
census block.
3.4. Testing the effects of variability and uncertainty on health impact
estimates
Five simulations of 2000 iterations each were run using Analytica
version 4.5 (Lumina Decision Systems, Los Gatos, California) to demonstrate differences in health burden estimates when including variability
and uncertainty. Table 6 lists the ve simulations and the variability and
uncertainty considered in each. The rst simulation (1a) follows the deterministic approach of previous HIAs, using average trafc volumes and
a constant trafc emission factor corresponding to trafc cruising at
35 mph on a at roadway under a constant ambient temperature of
70 F. Like previous HIAs, simulation 1a accounts for neither uncertainty
in the concentrationresponse coefcient (using the mean value as a deterministic estimate) nor seasonal variability. Simulation 1b is identical
to simulation 1a, except that it uses seasonal concentrationresponse

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

415

Table 3
Concentrationresponse coefcients used in this study.
Health outcome

Disease category

ICD-9 or ICD-10 codea

Age group

Season

Mean concentration-response
coefcient (95% CI), % per 10 g/m3b

Mortality

Cardiovascular

I01I59

All ages

Respiratory

J00J99

All ages

Cardiovascular

410414, 426429, 430438, and 440449

65 and over

Respiratory

464466, 480487, and 490492

65 and over

All-yearc
Winter
Spring
Summer
Fall
All-yearc
Winter
Spring
Summer
Fall
All-yearc
Winter
Spring
Summer
Fall
All-yearc
Winter
Spring
Summer
Fall

0.66 (0.66, 1.98)


1.35 (1.93, 4.62)
0.76 (2.73, 4.25)
0.62 (2.22, 3.47)
0.18 (2.93, 2.57)
1.21 (0.48, 2.90)
0.93 (1.44, 3.29)
0.35 (2.05, 2.75)
0.77 (1.55, 3.10)
0.96 (1.34, 3.25)
0.29 (0.19, 0.77)
1.05 (0.07, 2.19)
0.75 (0.26, 1.76)
0.67 (1.61, 0.26)
0.17 (0.72, 1.06)
0.35 (0.44, 1.13)
0.40 (1.46, 2.24)
0.75 (0.82, 2.31)
0.52 (2.09, 1.05)
0.14 (1.30, 1.58)

Unscheduled hospital admissions

a
b
c

ICD-10 for mortality, and ICD-9 for unscheduled hospital admissions.


Coefcients were originally from Zanobetti and Schwartz (2009) and Bell et al. (2008) respectively.
Used only in simulation 1.

coefcients (also deterministic). Simulations 24 systematically include


(one at a time) variability in vehicle emissions rates (simulations 24),
uncertainty in concentrationresponse coefcients (simulations 34),
and air quality model prediction error (simulation 4).
3.5. Comparing health impacts under alternative scenarios
As noted previously, we simulated health impacts for the full study
corridor (160 census blocks) for three different scenarios (Table 2):
year 2009 and year 2025 with and without the new campus. For each
scenario, 2000 simulations were run in Analytica. Trafc patterns (trafc volumes along each roadway link in the corridor) for each scenario
were taken from a previous trafc impact analysis conducted for the
Town of Chapel Hill (Vanasse Hangen Brustlin Inc., 2009). The 2025 census block populations were obtained from forecasts by the North
Carolina Capital Area Metropolitan Planning Organization (2005).
These growth rates account for demographic changes expected to
occur if the Carolina North campus is built.
4. Results and discussion
This analysis explored the effects of variability and uncertainty on
health impact estimates of near-roadway air pollution arising from trafc attracted by new suburban development projects. Most previous U.S.

HIAs of such projects have provided qualitative rather than quantitative


assessments of health impacts; the few quantitative HIAs have not
systematically represented variability and uncertainty in the variables
used to estimate health impacts or in the resulting health outcome predictions. We explored whether including variability and uncertainty
makes a difference in central estimates of health impacts, and we examined the magnitude of uncertainty in the resulting estimates. We then
employed an approach that accounts for variability and uncertainty to
model the expected health impacts in the year 2025 of new trafc generated by a new research campus development along a busy roadway
corridor in Chapel Hill, North Carolina.
4.1. Effect of including variability and uncertainty
Our results suggest that the conventional, deterministic HIA approach may systematically under-estimate potential health impacts of
trafc-related PM2.5 exposure (Fig. 3).
Incorporating trafc emission variability into the analysis (as in simulation 2) caused the mean value of estimated health impacts to increase by more than a factor of two, compared to estimates that
exclude such variability (simulation 1b). This increase occurred because
neglecting the effects on vehicle emissions of variability in temperature,
road grade, vehicle speed, and trafc behavior (idling, accelerating, decelerating, or cruising) resulted in under-estimates of PM2.5 exposure

Table 4
Annual mortality rates by race, gender, and age group for Orange County (per 1000 people).
Cause of death

ICD-10 code

Cardiovascular disease

I05I09, I10I15, I20I25, I26I28, and I30I52

Respiratory disease

J00J99

Age group

0 to 34
35 to 44
45 to 54
55 to 64
65 to 74
75 to 84
85+
0 to 54
55 to 64
65 to 74
75 to 84
85+

Race and gender


White male

Black male

Other male

White female

Black female

Other female

0.00
0.17
0.29
1.79
3.43
16.47
52.51
0.00
0.49
2.06
5.24
35.80

0.00
0.00
2.72
2.35
7.25
17.24
26.32
0.00
0.00
4.83
5.75
13.16

0.00
0.00
0.00
2.57
0.00
0.00
125.00
0.00
0.00
0.00
1.16
0.00

0.00
0.00
0.13
0.91
2.44
7.70
30.34
0.00
0.45
1.83
4.95
7.87

0.00
0.95
1.53
1.05
0.00
20.65
22.99
0.00
1.05
3.45
5.90
11.49

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

416

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

Table 5
Annual emergency department visits rates for North Carolina.
Cause of Visit

ICD 9 code

Age group

Annual rate

Cardiovascular disease
Respiratory disease

427.5, 428 and 518.4 (excluding failure due to fumes and vapors), 430435, and 437.0437.1
466 and 480486

65 and over
65 and over

0.0856
0.0355

concentrations, on average. For example, vehicle emission rates nearly


tripled when the road grade changed from 0% (as assumed in the
conventional modeling approach) to 10% (Chart-asa et al., 2013). Similarly, emissions doubled when the temperature decreased from 70 F
(the default assumption under the conventional assessment approach)
to 10 F (Chart-asa et al., 2013). For at roadways with trafc moving at
constant speeds in climates with minimal temperature uctuation, variability in emissions factors is expected to be small, but for most cases, it
is clear that emissions factor variability is an important consideration
when predicting health impacts.
When additionally including the uncertainty in concentrationresponse coefcients into the modeling approach (simulation 3), the
mean estimate of health impacts increased still further: the estimated
number of attributable CVD deaths and respiratory hospital admissions
more than doubled, while respiratory deaths and CVD hospital admissions increased by 69% and 11%, respectively. This result occurred because we represented concentrationresponse coefcients as rightskewed probability distributions (normal distributions left-truncated
at zero). This representation is appropriate because of the constraint
that the coefcients must be non-negative (since PM2.5 exposure does
not benet public health). The result is that the mean value of the concentrationresponse coefcients is greater than the median value,
which in turn increased the mean estimated health impacts, compared
to when such uncertainty was excluded.
When additionally including the uncertainty in model prediction
error (simulation 4), the mean estimates increased by another 1617%
compared to simulation 3. This result occurred because of the right
skew in the triangular distribution used to represent model uncertainty
and the interactions of this distribution with that used to represent the
concentrationresponse coefcient. As previously explained, the triangular distribution reects previous research on the performance of the
CAL3QHCR model (Chart-asa et al., 2013).
In summary, incorporating variability and uncertainty into the
model predictions increased the mean value of estimated health impacts, compared to predictions that excluded variability and uncertainty. The health impact estimates increased by factors of 7, 8, 4, and 9 for
CVD deaths, CVD hospital admissions, respiratory deaths, and respiratory hospital admissions, respectively. The estimates that excluded variability and uncertainty are biased so low that they are outside the 95%
Table 6
Sources of uncertainty and variability included in the ve simulations.
Uncertainty and variability sources

Simulation
number
1a 1b 2 3 4

Sources of uncertainty
PM2.5 exposure concentration
Air quality model prediction accuracy
Doseresponse function
Doseresponse coefcient
Sources of variability
PM2.5 exposure concentration
Vehicle emissions variability on each roadway link arising
from the following sources: temperature; road grade;
cruising speed; and percent time spent decelerating, idling,
accelerating, and cruising
Doseresponse function
Seasonal variability
Demographic characteristics of exposed population
Age, race, and gender (by census block)

x
x x

x x x

x x x

x x x

condence intervals of estimates including variability and uncertainty.


These biased predictions could have important implications for
decision-making. For example, it is possible that excluding variability
and uncertainty, and hence producing unrealistically low estimates of
health impacts, could result in a decision not to pursue mitigation measures that would have been determined cost-effective had the full impacts of variability and uncertainty been considered.
4.2. Overall population health impacts at the case study site
This analysis predicted that by 2025, the total number of adverse
health cases attributable to trafc-related PM2.5 on the case study roadway will decrease relative to 2009, with or without the Carolina North Development (although the decrease is lower with the development)
(Table 7). This decrease in the number of adverse health outcomes is
predicted to occur despite an expected 20% increase in the population
by 2025. Overall, the number of cases of CVD mortality, CVD hospital admissions, respiratory mortality, and respiratory hospital admissions are
expected to decrease by 42%, 38%, 47%, and 42%, respectively. The decreased risks arise from the built-in assumptions of MOVES that future
vehicles will be cleaner than today's eet, resulting in trafc emissions
that decline by about 50%, on average, compared to today's vehicles.
However, the increased trafc associated with the new campus will offset even greater decreases in near-roadway PM2.5 expected to occur in
2025 in the absence of the new campus; the number of adverse health
outcomes is expected to be about 30% lower if the new campus is not
built, compared to if it is built (results not shown).
The health risks of primary PM2.5 from trafc on the case study
roadway vary considerably by season and location (Fig. 4). For CVD
mortality, effects are highest in winter due to the inuences of high concentrationresponse coefcients, seasonal incidence variations, and
trafc emission factors during low temperatures. The spatial variability
in risk is especially pronounced in winter, as illustrated by the gradations by census block illustrated in Fig. 4. Similar seasonal and spatial effects are observed for the other three health outcomes (not shown).
To investigate the potential factors explaining the spatial distribution of risk, we calculated correlations between several potential explanatory variables and the total excess mortality and morbidity
attributable to PM2.5 from the roadway in each census block for the
year 2009. Variables included distance from the roadway to the census
block centroid, total census block population, population over age 64,
percent of the population identifying as black, and mean PM2.5 concentration attributable to the roadway across all seasons. For excess mortality, the correlations are largest for mean PM2.5 concentration (r = 0.42, t
(158) = 5.8, p = 1.4 108) and percentage of the population identifying as black (r = 0.37, t(114) = 4.2, p = 2.5 105). The correlations
are smaller for distance to the roadway (r = 0.22, t(158) = 2.8, p =
0.0028), total population (r = 0.15, t(158) = 2.0, p = 0.025), and population over age 64 (r = 0.16, t(158) = 2.0, p = 0.025). The results are
similar for excess morbidity. The spatial distribution in risk arises from
complex interactions among a variety of factors, including factors affecting population susceptibility (potentially including age and race) and
factors affecting exposure concentration. Factors that affect the spatial
distribution of exposure concentrations include not only distance from
the roadway but also road grade, vehicle types, vehicle speed, trafc volume, the presence of intersections, and wind speed and direction. The
effects of such factors are described in detail in Chart-asa et al. (2013).
The above-noted correlation between mortality risk associated with
trafc-related PM2.5 exposures and the percentage of the census block

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

417

Fig. 3. Effect on health impact estimates of including the variability and uncertainty sources shown in Table 6. Error bars represent 95% condence intervals.

population identifying as black suggests the possible presence of racial


disparities in exposure risks. The census block having the highest number of total deaths attributable to trafc on the study corridor under current conditions (block number 371350118002002, with a population of
201) also has a very high percentage of black residents, at 47%, compared to 9% in the study area as a whole. This census block is the
home of a public housing community, Airport Gardens, intended for
low-income families. The block has the second-highest PM2.5 exposure
concentration among all blocks in the study area. Of the 10 census
blocks with the highest number of attributable deaths, seven have
higher percentage black populations (1747%) than the average for
the study area. Nonetheless, even in the highest-risk of these census
blocks, the annual per-person risk of premature mortality due to
trafc-related PM2.5 exposure is vanishingly small: 5.8 108 (obtained by dividing the annual attributable deaths by the total population
of the census block). Over a 70-year lifetime, this equates to a risk of
4.1 10 6. Along other, busier roadways, however, the health signicance of such disparities could be much greater.
Overall, we predict that future risks of primary PM2.5 from increased
trafc associated with the Carolina North campus will be extremely low.
If the new campus is built, then 9 106 excess CVD deaths and 2
106 excess respiratory deaths are expected, compared to if the campus
is not built (Table 7). Summing these two estimates and dividing by the
future study corridor population of 19,140 yields a per-person annual
risk of about 6 1010. These risks are low even if one assumes a resident is exposed to such a risk level for a lifetime. For a 70-year lifetime,
the per-person lifetime risk is 4 108. Even in the most-exposed census block, lifetime risks attributable exclusively to the new campus are
relatively low (about 1 10 8 per year, or less than one-in-onemillion over a lifetime).

4.3. Sensitivity and uncertainty analysis


The 95% condence interval values of the risk estimates in Table 7
range over a factor of about 67. For example, the upper 95% condence
interval estimate of annual CVD deaths attributable to roadway trafc,
1.0 104, is about 7 times larger than the lower 95% condence interval estimate, 1.5 105. While from a policy standpoint the risks at
both ends of this condence interval are relatively low, at other sites
the optimal policy decision might change if the actual risk were close
to the upper or lower 95% condence interval value rather than the central estimate. Hence, in future applications of the HIA analysis approach
demonstrated in this article, identifying the variables with the biggest
inuence on the mean value of and uncertainty in the risk estimates
may be important, in order to guide additional data collection prior to
making a risk-informed decision.
In a future application, a decision-maker may wish to know the effects of changing each random variable in an HIA model to plausible
high and low values. Three key random variables underlie this analysis:
the PM2.5 concentration in each census block as predicted by the combined MOVES/CAL3QHCR model, the model uncertainty factor
(representing the departure of this combined model from actual PM2.5
concentrations), and the doseresponse coefcient. Fig. 5 shows the effects on the predicted number of CVD deaths of xing each of these variables at its lower and upper 95% condence interval value, while
keeping all other variables the same. The effects vary by census block
and hence are presented as cumulative distribution functions (CDFs).
For example, the doseresponse coefcient relating PM2.5 exposure
concentration to the risk of CVD mortality in winter is represented in
the base model as a truncated normal distribution with mean 1.35
103 and standard deviation 1.7 10 3; the lower 95% CI of this

Table 7
Comparison of HIA results by development scenario.
Scenario

2009
2025 without Carolina North
2025 with Carolina North
a
b

Number of census
blocks affecteda

Range of mean exposure


concentrations in affected
blocks (g/m3)b

Total cases 106


CVD
mortality

CVD hospital
admissions

Respiratory
mortality

Respiratory hospital
admissions

118148
75122
84137

0.00020.16
0.00020.10
0.00020.13

48 (15100)
19 (5.642)
28 (7.961)

140 (47280)
61 (19120)
87 (27170)

15 (530)
5.5 (1.712)
7.9 (2.417)

73 (21160)
30 (866)
42 (1293)

Number of census blocks with exposure concentrations greater than zero (varies by season).
Lowest and highest mean seasonal exposure concentration in affected census blocks (also varies by season).

418

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

Fig. 4. Spatial distribution of cardiovascular deaths (106) attributable to PM2.5 before and after Carolina North development.

distribution is 1.2 10 4, and the upper 95% CI is 4.7 103. The


DoseResponse Coefcient High curve in Fig. 5 shows a CDF of the
risk estimates for census blocks when this coefcient and those for the
other three seasons are xed at their upper 95% CI values (in the case
of winter, 4.7 10 3) rather than varying randomly, while leaving
the other model variables unchanged. Fig. 5 shows that for all census
blocks, the risk estimates are more sensitive to the concentration
response coefcient than to the other random variables in the risk
model (air quality model uncertainty factor and predicted PM2.5 exposure concentration). When the effects of xing each seasonal dose
response coefcient for CVD mortality at lower or upper 95% CI values
are summed across all census blocks, then the estimated number of
CVD deaths changes from the mean estimate of 47 10 6 to 2.5
106 and 120 106, respectively (Fig. 6). These results illustrate the
potential importance for future HIAs of strengthening the epidemiologic
basis for predicting the health effects of PM2.5 exposures, in order to decrease the potential for producing risk estimates that are either too high
or too low. (Note that results for other health outcomes, not shown
here, as similar to those illustrated in Figs. 56.)
A second question that decision-makers might ask is why the 95%
condence intervals in estimated risks are so wide. One approach to answering this question is to examine the rank-order correlation between
the estimated risks and each random variable in the model. A high rankorder correlation between an input variable and the risk estimate indicates that high values of the input variable drive the risk estimate

toward comparably high values. For this analysis, the rank-order correlations differ by census block, season, and health outcome. Fig. 7 shows
CDFs of the rank-order correlations between each random input variable and CVD mortality risks among the census blocks by season. In winter, the season in which PM2.5 exposure concentrations are highest,
uncertainty in the doseresponse coefcient drives uncertainty in the
risk estimates in all census blocks. In spring and summer, the air quality
model uncertainty factor drives the uncertainty in the risk estimates. In
fall, the model uncertainty factor drives uncertainty except for in about
20% of census blocks, where the doseresponse coefcient contributes
the most uncertainty. Hence, overall, to decrease uncertainty in the
risk predictions, both the strength of the epidemiologic evidence and
the performance of near-roadway air pollutant dispersion models
must be improved.
In summary, Figs. 57 illustrate the importance for future
transportation-related HIAs of decreasing uncertainty in epidemiologic
estimates of the concentrationresponse coefcient and improving the
ability to model near-roadway concentrations of PM2.5 from trafc
5. Limitations
Key limitations in this analysis arise from deciencies in the available epidemiologic evidence, the capabilities of the air quality model,
and future population data. In addition, the attributable fraction approach considers effects of PM2.5 exposure on the incidence of

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

419

Fig. 5. Effects of changing risk model input variables to their upper and lower 95% condence interval values. The cumulative distribution functions illustrate the variability in these effects
by census block in the case study roadway corridor.

cardiovascular and respiratory deaths but may overlook effects on the


population prevalence of CHD and respiratory diseases.
One limitation arises from the assumption that all PM2.5including
that generated by trafchas the same health effects as PM2.5 measured
at population-oriented central site monitors used as the basis for exposure estimates in the epidemiologic studies from which the concentrationresponse coefcients are drawn. This assumption is common to
nearly all air quality risk assessments to date (e.g., Cohen et al., 2005;
Li et al., 2010; Fann et al., 2012), because the understanding of differential health effects of PM2.5 from different sources is still emerging. According to a meta-analysis by Janssen et al., trafc-associated PM2.5
may have greater health risks than PM2.5 from non-combustion sources
(Janssen et al., 2011). Janssen et al. found that, theoretically, risk estimates using black carbon particles, which are associated with combustion from motor vehicle engines and other sources, as an indicator of
trafc-related pollution yielded risk estimates 49 times higher than estimates using overall PM2.5 as an indicator. However, our analysis required use of PM2.5, since MOVES and CAL3QHCR do not provide the
capability to estimate black carbon particle concentrations. Furthermore, the available epidemiologic evidence on the association between
black carbon particles and health risks is not nearly as extensive or thoroughly reviewed as that for PM2.5 (Janssen et al., 2011). Updating nearroadway dispersion models to predict black carbon particle

Fig. 6. Overall effect (across all census blocks) of changing random variables in the risk
model to the upper and lower ends of their 95% condence intervals. The chart is centered
on the mean value of the risk estimate, 48 106. The ends of each bar correspond to the
new risk estimate if the variable is changed to its low (left side) or high (right side) 95%
condence interval value.

concentrations and conducting further epidemiologic studies examining the effects of vehicle emissions on health are important areas of research. Nonetheless, for the case study site, the estimated risks would be
very low even assuming the risks are under-estimated by a factor of 9
(the upper bound of Janssen et al.'s predicted under-estimation when
using PM2.5 rather than black carbon particles as an air pollution indicator). In the baseline scenario (year 2009), the annual average CVD or respiratory mortality risk to an individual from trafc-related air pollution
predicted by our model is 3.6 10 9 (= 45 10 6 CVD deaths plus
13 10 6 respiratory deaths divided by a population of 16,000). Assuming a 70-year lifetime exposure period, the resulting lifetime risk
is 2.5 10 7. Increasing these risks by a factor of 9 results in an annual
risk of 3.3 108 and a lifetime risks of 2.3 106risks that are considered very low according to U.S. EPA guidelines, which in general have
long designated as acceptable risks of less than 10 4 to 106 (EPA,
1989).
A second limitation is that the concentrationresponse coefcients
assume that the exposure histories of current and future residents of
the case study area will be similar to those in the areas from which
the epidemiologic studies were drawn (Atlanta and the southeastern
United States). Once again, this limitation is inherent in current air quality risk assessments, due to the costs of conducting epidemiologic studies and the resulting lack of studies for each U.S. metropolitan area. This
limitation may bias the absolute results of the risk estimates, but it does
not affect the estimates of risks of one scenario relative to another.
Hence, the conclusion that the development of the Carolina North campus is unlikely to lead to substantial trafc-related air quality health impacts is valid even if exposure histories of the Chapel Hill population
differ from those of the populations from which relative risk estimates
were derived.
A third limitation is that Eqs. (3a), (3b), (3c) and (3d), which have
been used as the basis for assessing health impacts of air pollution
exposure by nearly all researchers to date, may neglect the effects of
air pollution exposure on the disease progression leading up to hospitalizations for respiratory illnesses and CVD (Perez et al., 2013). Perez et al.
recently found that including such effects in analyzing health impacts of
trafc-related road pollution increased estimated health impacts on average by a factor of about 10 in a study of 10 major European cities
(Perez et al., 2013). However, implementing the approach of Perez
et al. is not possible when attempting to predict changes in health effect
estimates in the distant future, because Perez's calculation relies on

420

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

Fig. 7. Cumulative distribution functions of rank-order correlations between model input variables and the predicted risk of CVD mortality by season for the census blocks in the case study
roadway corridor. A high rank-order correlation indicates that the variable has a strong inuence on the uncertainty in the estimated risk, so reducing uncertainty in the variable will substantially reduce uncertainty in the estimated risk. The cumulative distribution functions show the variability in these effects by spatial location (i.e., by census block) and season.

epidemiologic studies that use proximity to a busy roadway as the exposure metric. For estimating the effect of roadway emissions on coronary
heart disease (CHD) prevalence, for example, Perez relies on an epidemiologic study in Germany showing that living within 150 m of a
busy roadway (dened as an autobahn or federal highway) increased
the relative risk of CHD by 85%, compared to not living near such a roadway. Because per-vehicle emissions are expected to decrease substantially in the future, such studies cannot be used as the basis for
predicting the effects of road trafc pollution on populations in the distant future. We expect that future health impacts of living near busy
roads will decrease as vehicle emissions controls improve, so including
the effects on disease prevalence also would not change the conclusion
that the future risks will be less than today's risks, even if the new campus is built.

6. Conclusions
This study developed an improved modeling approach for estimating the health impacts of trafc-related PM2.5 air pollution under alternative future urban development scenarios. We then demonstrated the
approach by quantifying health impacts in a case study roadway corridor that could be affected by a new UNC campus extension in Chapel
Hill. The new approach accounts for the effects of variability in trafc
emissions factors and for seasonal variability in concentrationresponse
coefcients. It also accounts for uncertainty in concentrationresponse
coefcients and air quality model prediction error. The approach could

serve as a model for future health impact assessments considering


trafc-related PM2.5.
Comparisons to the conventional modeling approach used in other
quantitative HIAs revealed that those HIAs could under-estimate potential health burdens by failing to consider variability and uncertainty in
input variables used to generate the health impact estimates. Our analysis showed that in the case study corridor the conventional approach
under-predicted health impacts by a factor of 4 to 9, depending on the
health endpoint. As such, the conventional HIA approach could in
some circumstances lead to decisions that are not cost effective and/or
are not sufciently protective of public health.
This analysis also showed that a ne-scale approach that quanties
impacts over a small grid (in this case, U.S. census blocks), accounting
for demographic variability in each grid cell along with the previously
mentioned variability and uncertainty in model inputs, can be useful
for identifying health disparities. For example, this analysis reveals
that the neighborhood in the study area with the highest health burden
also has a very high minority population, compared to that in the rest of
the study area. In addition, this method of accounting for demographic
variability can be used to analyze differences in risks among age and
gender groups. It could be extended to analyze impacts among populations with pre-existing health conditions and/or among groups with different income levels and educational attainment levels, as suggested in
a recent analysis of distributional effects of air quality policies by Fann
et al. (Fann et al., 2011). An analysis of distributional effects would require data on health outcomes, educational attainment, and income at
the census block level; such data were not available for this analysis

C. Chart-asa, J.M. Gibson / Science of the Total Environment 506507 (2015) 409421

but potentially could be obtained through special requests to the county


health department and U.S. Census Bureau.
Overall, the annual mortality risks of trafc-related PM2.5 from the
proposed new campus are very low (less than 1 108) even for the
most-exposed populations. Nonetheless, it is important to recognize
that we consider only one type of trafc-related pollutant and one roadway corridor. Risks would be higher if including all roadways affected
by trafc from the new campus and all trafc-related pollutants. Furthermore, it is important to keep in mind the many other sources of ambient air pollution exposure in the study area and the cumulative effects
of multiple exposures. Taking steps to reduce trafc from the new campus (e.g., increasing the frequency of public transit service, encouraging
carpooling, charging for parking, and other steps) will reduce air pollution exposures and produce benets beyond those along the single
roadway considered in this case study.
Overall, this work highlights the sensitivity of trafc-related health
impact assessments to uncertainty and variability in concentrationresponse coefcients, air quality model prediction accuracy, and trafc
emissions factors. Future HIAs should account for these inuential variability and uncertainty sources.
References
Aunan K. Exposureresponse functions for health effects of air pollutants based on epidemiological ndings. Risk Anal Oct. 1996;16(5):693709.
Bell ML, Ebisu K, Peng RD, Walker J, Samet JM, Zeger SL, Dominici F. Seasonal and regional
short-term effects of ne particles on hospital admissions in 202 US counties, 1999
2005. Am J Epidemiol Dec. 2008;168(11):130110.
Bhatia R, Corburn J. Lessons from San Francisco: health impact assessments have advanced political conditions for improving population health. Health Aff (Millwood)
Dec. 2011;30(12):24108.
Bhatia R, Seto E. Quantitative estimation in health impact assessment: opportunities and
challenges. Environ Impact Assess Rev Apr. 2011;31(3):3019.
Centers for Disease Control and Prevention. CDC WONDER; 2013.
Chart-asa C, Sexton KG, Macdonald Gibson J. Trafc impacts on ne particulate matter air
pollution at the urban project scale: a quantitative assessment. J Environ Prot (Irvine,
Calif) December, 2013;4.
Cohen AJ, Ross Anderson H, Ostro B, Pandey KD, Krzyzanowski M, Knzli N, Gutschmidt K,
Pope A, Romieu I, Samet JM, Smith K. The global burden of disease due to outdoor air
pollution. J Toxicol Environ Health A 2005;68(1314):13017.
Dannenberg AL, Bhatia R, Cole BL, Heaton SK, Feldman JD, Rutt CD. Use of health impact
assessment in the U.S.: 27 case studies, 19992007. Am J Prev Med Mar. 2008;34
(3):24156.
EPA. Risk Assessment Guidance for Superfund (part A); 1989 [Washington, D.C.].
Fann N, Lamson AD, Anenberg SC, Wesson K, Risley D, Hubbell BJ. Estimating the national
public health burden associated with exposure to ambient PM2.5 and ozone. Risk Anal
Jan. 2012;32(1):8195.
Fann N, Roman HA, Fulcher CM, Gentile MA, Hubbell BJ, Wesson K, Levy JI. Maximizing
health benets and minimizing inequality: incorporating local-scale data in the design and evaluation of air quality policies. Risk Anal Jun. 2011;31(6):90822.
Federal Highway Administration. Economic analysis primer: benetcost analysis; 2003
[Washington, DC].
Frey HC, Burmaster DE. Methods for characterizing variability and uncertainty: comparison of bootstrap simulation and likelihood-based approaches. Risk Anal Feb. 1999;19
(1):10930.
Human Impact Partners. Pittsburg Railroad Avenue specic plan health impact assessment; 2008 [Oakland, CA].
Human Impact Partners. Pathways to community health: evaluating the healthfulness of
affordable housing opportunity sites along the San Pablo Avenue Corridor using
health impact assessment; 2009 [Oakland, CA].
Janssen NAH, Hoek G, Simic-lawson M, Fischer P, Van Bree L, Brink H, Keuken M, Atkinson
RW, Anderson HR, Cassee FR, Van Bree L. Black carbon as an additional indicator of

421

the adverse health effects of airborne particles. Environ Health Perspect 2011;19
(12):16919.
Li Y, Gibson JM, Jat P, Puggioni G, Hasan M, West JJ, Vizuete W, Sexton K, Serre M. Burden
of disease attributed to anthropogenic air pollution in the United Arab Emirates: estimates based on observed air quality data. Sci Total Environ Nov. 2010;408(23):
578493.
Mathers C, Vos T, Lopez A, Salomon J, Ezzati M. National burden of disease studies: a practical guide. Edition 2.0. Global program on evidence for health policy; 2001 [Geneva,
Switzerland].
Minnesota Population Center. National Historical Geographic Information System: version 2.0. Minneapolis, MN: University of Minnesota; 2011.
Morgan MG, Henrion M, Small M. Uncertainty: a guide to dealing with uncertainty in
quantitative risk and policy analysis. Cambridge; New York: Cambridge University
Press; 1990. p. 332.
Murray CJ, Ezzati M, Lopez AD, Rodgers A, Vander Hoorn S. Comparative quantication of
health risks: conceptual framework and methodological issues. Popul Health Metr
2003;1(1):1.
National Research Council. Improving health in the United States: the role of health impact assessment. Washington, D.C: National Academy Press; 2011.
NCDC. Quality Controlled Local Climatological Data (QCLCD); 2013.
NOAA. NOAA/ESRL Radiosonde Database; 2013.
North Carolina Capital Area Metropolitan Planning Organization. Socio-economic demographic forecasts shapele. 2030 Long Range Transportation Plan; 2005.
North Carolina State Center for Health Statistics. Detailed mortality statistics, 2010; 2012.
Ostro B. Outdoor air pollution: assessing the environmental burden of disease at national
and local levels, no. 5. Geneva: WHO; 2004.
Ostro B, Chestnut L. Assessing the health benets of reducing particulate matter air pollution in the United States. Environ Res 1998;76(2):94106.
Perez L, Declercq C, Iiguez C, Aguilera I, Badaloni C, Ballester F, Bouland C, Chanel O,
Cirarda FB, Forastiere F, Forsberg B, Haluza D, Hedlund B, Cambra K, Lacasaa M,
Moshammer H, Otorepec P, Rodrguez-Barranco M, Medina S, Knzli N. Chronic burden of near-roadway trafc pollution in 10 European cities (APHEKOM network). Eur
Respir J Sep. 2013;42(3):594605.
Prss-stn A, Mathers CD, Corvalan C, Woodward A. Introduction and methods:
assessing the environmental burden of disease at national and local levels; 2003
[Geneva].
Ross CL, Elliott ML, Rushing MM, Barringer J, Cox S, Frackelton A, Kent J, Rao A.
Aerotropolis Atlanta Browneld redevelopment health impact assessment, Vol. I
Atlanta. : Center for Quality Growth and Regional Development at the Georgia
Institute of Technology; 2011. p. 169.
Singleton-Baldrey L. The impacts of health impact assessment: a review of 54 health impact assessments, 20072012. University of North Carolina at Chapel Hill; 2012.
University of North Carolina at Chapel Hill. The UNC Department of Emergency Medicine
Carolina Center for Health Informatics report, overview and analysis of NC DETECT
emergency department data: 2009; 2011 [Chapel Hill, NC].
UC Berkeley Health Impact Group. Oak to Ninth Avenue health impact assessment; 2006
[Berkeley, CA].
UC Berkeley Health Impact Group. MacArthur BART health impact assessment; 2007
[Berkeley, CA].
UC Berkeley Health Impact Group. Health impact assessment of the Port of Oakland; 2010
[Berkeley, CA].
U.S. Environmental Protection Agency. Risk Assessment Guidance for Superfund (RAGS)
volume III part A: process for conducting probabilistic risk assessment; 2001
[Washington, D.C.].
US Environmental Protection Agency. Quantitative health risk assessment for particulate
matter. EPA-452/R-10-005. NC: Research Triangle Park; 2010.
U.S. Environmental Protection Agency. Quantitative health risk assessment for particulate
matter. EPA-452/R-10-005. NC: Research Triangle Park; 2010.
Vanasse Hangen Brustlin Inc. Transportation impact analysis for the Carolina North development.; 2009 [Watertown, MA].
Wernham A. Health impact assessments are needed in decision making about environmental and land-use policy. Health Aff (Millwood) May 2011;30(5):94756.
Yura EA, Kear T, Niemeier D. Using CALINE dispersion to assess vehicular PM2.5 emissions.
Atmos Environ Dec. 2007;41(38):874757.
Zanobetti A, Schwartz J. The effect of ne and coarse particulate air pollution on mortality:
a national analysis. Environ Health Perspect 2009;117(6):898903.

You might also like