Post on 15-Aug-2020
EVALUATION OF INTERACTIVE EFFECTS
BETWEEN TEMPERATURE AND AIR POLLUTION
ON HEALTH OUTCOMES
BY
CIZAO REN
Bachelor of Medicine, Master of Medicine, Master of Science (Epidemiology)
A thesis submitted for the Degree of Doctor of Philosophy in the School
of Public Health, Queensland University of Technology
April 2007
I
SUMMARY
A large number of studies have shown that both temperature and air pollution (eg, particulate
matter and ozone) are associated with health outcomes. So far, it has received limited
attention whether air pollution and temperature interact to affect health outcomes. A few
studies have examined interactive effects between temperature and air pollution, but produced
conflicting results. This thesis aimed to examine whether air pollution (including ozone and
particulate matter) and temperature interacted to affect health outcomes in Brisbane, Australia
and 95 large US communities.
In order to examine the consistency across different cities and different countries, we used
two datasets to examine interactive effects of temperature and air pollution. One dataset was
collected in Brisbane City, Australia, during 1996-2000. The dataset included air pollution
(PM10, ozone and nitrogen dioxide), weather conditions (minimum temperature, maximum
temperature, relative humidity and rainfall) and different health outcomes. Another dataset
was collected from the 95 large US communities, which included air pollution (ozone was
used in the thesis), weather conditions (maximum temperature and dew point temperature)
and mortality (all non-external cause mortality and cardiorespiratory mortality).
Firstly, we used three parallel time-series models to examine whether maximum temperature
modified PM10 effects on cardiovascular hospital admissions (CHA), respiratory hospital
admissions (RHA), cardiovascular emergency visits (CEV), respiratory emergency visits
(REV), cardiovascular mortality (CM) and non-external cause mortality (NECM), at lags of
0-2 days in Brisbane. We used a Poisson generalized additive model (GAM) to fit a bivariate
model to explore joint response surfaces of both maximum temperature and particulate matter
II
less than 10 µm in diameter (PM10) on individual health outcomes at each lag. Results show
that temperature and PM10 interacted to affect different health outcomes at various lags. Then,
we separately fitted non-stratification and stratification GAM models to quantify the
interactive effects. In the non-stratification model, we examined the interactive effects by
including a pointwise product for both temperature and the pollutant. In the stratification
model, we categorized temperature into two levels using different cut-offs and then included
an interactive term for both pollutant and temperature. Results show that maximum
temperature significantly and positively modified the associations of PM10 with RHA, CEV,
REV, CM and NECM at various lags, but not for CHA.
Then, we used the above Poisson regression models to examine whether PM10 modified the
associations of minimum temperature with CHA, RHA, CEV, REV, CM and NECM at lags
of 0-2 days. In this part, we categorized PM10 into two levels using the mean as cut-off to fit
the stratification model. The results show that PM10 significantly modified the effects of
temperature on CHA, RHA, CM and NECM at various lags. The enhanced adverse
temperature effects were found at higher levels of PM10, but there was no clear evidence for
synergistic effects on CEV and REV at various lags. Three parallel models produced similar
results, which strengthened the validity of these findings.
Thirdly, we examined whether there were the interactive effects between maximum
temperature and ozone on NECM in individual communities between April and October,
1987-2000, using the data of 60 eastern US communities from the National Morbidity,
Mortality, and Air Pollution Study (NMMAPS). We divided these communities into two
regions (northeast and southeast) according to the NMMAPS study. We first used the
bivariate model to examine the joint effects between temperature and ozone on NECM in
III
each community, and then fit a stratification model in each community by categorizing
temperature into three levels. After that, we used Bayesian meta-analysis to estimate overall
effects across regions and temperature levels from the stratification model. The bivariate
model shows that temperature obviously modified ozone effects in most of the northeast
communities, but the trend was not obviously in the southeast region. Bayesian meta-analysis
shows that in the northeast region, a 10-ppb increment in ozone was associated with 2.2%
(95% posterior interval [PI]: 1.2%, 3.1 %), 3.1% (95% PI: 2.2%, 3.8 %) and 6.2 % (95% PI:
4.8%, 7.6 %) increase in mortality for low, moderate and high temperature levels, respectively,
while in the southeast region, a 10-ppb increment in ozone was associated with 1.1% (95% PI:
-1.1%, 3.2 %), 1.5% (95% PI: 0.2%, 2.8%) and 1.3% (95% PI: -0.3%, 3.0 %) increase in
mortality.
In addition, we examined whether temperature modified ozone effects on cardiovascular
mortality in 95 large US communities between May and October, 1987-2000 using the same
models as the above. We divided the communities into 7 regions according to the NMMAPS
study (Northeast, Industrial Midwest, Upper Midwest, Northwest, Southeast, Southwest and
Southern California). The bivariate model shows that temperature modified ozone effects in
most of the communities in the northern regions (Northeast, Industrial Midwest, Upper
Midwest, Northwest), but such modification was not obvious in the southern regions
(Southeast, Southwest and Southern California). Bayesian meta-analysis shows that
temperature significantly modified ozone effects in the Northeast, Industrial Midwest and
Northwest regions, but not significant in Upper Midwest, Southeast, Southwest and Southern
California. Nationally, temperature marginally positively modified ozone effects on
cardiovascular mortality. A 10-ppb increment in ozone was associated with 0.4% (95%
posterior interval [PI]: -0.2, 0.9 %), 0.3% (95% PI: -0.3%, 1.0%) and 1.6% (95% PI: 4.8%,
IV
7.6%) increase in mortality for low, moderate and high temperature levels, respectively. The
difference of overall effects between high and low temperature levels was 1.3% (95% PI: -
0.4%, 2.9%) in the 95 communities.
Finally, we examined whether ozone modified the association between maximum temperature
and cardiovascular mortality in 60 large eastern US communities during the warmer days,
1987-2000. The communities were divided into the northeast and southeast regions. We
restricted the analyses to the warmer days when temperature was equal to or higher than the
median in each community throughout the study period. We fitted a bivariate model to
explore the joint effects between temperature and ozone on cardiovascular mortality in
individual communities and results show that in general, ozone positively modified the
association between temperature and mortality in the northeast region, but such modification
was not obvious in the southeast region. Because temperature effects on mortality might
partly intermediate by ozone, we divided the dataset into four equal subsets using quartiles as
cut-offs. Then, we fitted a parametric model to examine the associations between temperature
and mortality across different levels of ozone using the subsets. Results show that the higher
the ozone concentrations, the stronger the temperature-mortality associations in the northeast
region. However, such a trend was not obvious in the southeast region.
Overall, this study found strong evidence that temperature and air pollution interacted to
affect health outcomes. PM10 and temperature interacted to affect different health outcomes at
various lags in Brisbane, Australia. Temperature and ozone also interacted to affect NECM
and CM in US communities and such modification varied considerably across different
regions. The symmetric modification between temperature and air pollution was observed in
the study. This implies that it is considerably important to evaluate the interactive effect while
V
estimating temperature or air pollution effects and further investigate reasons behind the
regional variability.
VI
Publications by the candidate on matters relevant to the thesis
Journal Articles
Ren C and Tong S. (2006) Temperature modifies the health effects of particulate matter in
Brisbane, Australia. Int J Biometerol. 51:87-96.
Ren C, Williams GM, Tong S. (2006) Does Particulate Matter Modify the Association
between Temperature and Cardiorespiratory Diseases? Environ Health Perspect 2006;
114:1690-1696.
Ren C, Tong S, Williams GM, Mengersen K. (2006) Does Temperature Modify Short-Term
Effects of Ozone on Total Mortality in 60 large Eastern US Communities? ( To be submitted)
Ren C, Williams GM, Tong S. (2006) Ozone, Temperature, and Cardiovascular Mortality in
95 Large US Communities, 1987-2000 -- Assessment Using the NMMAPS data. (To be
submitted)
Ren C, Williams GM, Morawska L, Mengersen K, Tong S. (2006). Ozone Modifies the
Associations between Temperature and Cardiorespiratory Mortality in 60 US Eastern Cities.
(To be submitted).
VII
Conference Presentation
Ren C, Williams GM, Tong S. Ozone modifies the association between temperature and
mortality in 12 US cities, 1987-2000.
Oral Presentation. The International Conference on Environmental Epidemiology and
Exposure. Paris. 2-6 September 2006
Ren C, Tong S. Temperature modifies the short-term effects of particulate matter on
cardiorespiratory diseases in Brisbane, Australia.
Poster Discussion. The International Conference on Environmental Epidemiology and
Exposure. Paris. 2-6 September 2006
Ren C. Williams GM, Tong S. The role of temperature in estimating the acute ozone effect
on cariorespiratory mortality in 95 large US communities during 1987-2000.
Oral Presentation. The 15th annual scientific meeting of the Australiasian
Epidemiological Association. Melbourne, Australia, 18-19 September 2006-10-06
VIII
STATEMENT OF AUTHORSHIP
The work contained in this thesis has not been previously submitted for a degree or diploma at
any other higher education institute. To the best of my knowledge and belief, the thesis
contains no materials previously published or written by another person except where
reference is made.
__________________
__________________
IX
Acknowledgements
I would like to thank the following people who made this thesis possible.
Thanks to my supervisors, A/Prof. Shilu Tong, Prof. Gail M Williams and Prof. Lidia
Morawska for their capable and experienced professional guidance. As a student from non-
English speaking background, I have been very lucky to be generously instructed by my
supervisor team, not only academically, but also linguistically and culturally. I would like to
thank A/Prof. Shilu Tong, the principal supervisor for his significant amount of time spent on
my professional guidance throughout my PhD study. I would also like to thank Prof. Gail M
Williams, industrial supervisor for her insightful statistical advice and guidance. I am grateful
to Prof. Lidia Morawska for her thoughtful instructions to my thesis as well.
In addition to my supervisors, I would like to thank Prof. Beth Newman and Prof. Kerrie
Mengersen for their time and energy to help and guide me for this thesis. Their contributions
made the thesis go more smoothly. I’d like to extend my appreciation to Genny Carter for
generous assistance during the last three years.
I am also grateful to Profs. Jonathan M Samet, Johns Hopkins University, H. Ross Anderson,
University of London, for their insightful comments on some of our manuscripts, and to Dr.
Roger Peng and his colleagues, Johns Hopkins University for their time and efforts in making
the NMMAPS data publicly available.
Besides the professional helpers, I am extremely grateful to my wife, parents and children, for
their patience, encouragement and emotional support throughout my PhD study.
X
I would also like to acknowledge all my colleagues and folks in the school, faculty and IHBI
for their advice and assistance with my research and personal friendship. I remember all of
you in my deep heart.
Finally, I would like to especially acknowledge the examination panel of my PhD study for
their contributions to improving this thesis.
XI
CONTENTS
CHAPTER 1: INTRODUCTION ..............................………………… 1
1.1 BACKGROUND ....……………………………………………………………...………..1
1.2 AIMS AND HYPOTHESIS …………………………………………………………….. 4
1.3 SIGNIFICANCE OF THE STUDY …………………………………………………….. 5
1.4 CONTENTS AND STRUCTURE OF THIS THESIS …………………………………. 6
CHAPTER 2: HEALTH EFFECTS OF AIR POLLUTION:
A LITERATURE VIEW …………………………………………………….. 8
2.1 INTRODUCTION .............................................................................................................. 8
2.2 MAIN AIR POLLUTANTS AND BIOLOGICAL MECHEMISMS ................................ 9
2.3 MAIN RESEARCH DESIGNS IN AIR POLLUTION EPIDEMIOLOGICAL STDUIES
............................................................................................................................................ 12
2.4 HEALTH EFFECTS OF AIR POLLUTION .................................................................... 25
2.5 CURRENT ISSUES .......................................................................................................... 36
2.6 SUMMARY ......................................................................................................................40
CHPATER 3: STUDY DESIGN AND METHODOLOGY ........................ 41
3.1 STUDY POPULATION ................................................................................................... 41
3.2 STUDY DESIGN ............................................................................................................ 42
3.3 DATA COLLECTION ................................................................................................. 42
3.4 DATA MANAGEMENT ................................................................................................. 45
3.5 ANALYTICAL PROTOCOL .......................................................................................... 45
XII
CHAPTER 4: TEMPERATURE MODIFIES THE HEALTH EFFECTS
OF PARTICULATE MATTER IN BRISBANE, AUSTRALIA ............. 53
ABSTRACT ...................................................................................................................... 54
4.1 INTRODUCTION ...................................................................................................... 55
4.2 MATERIALS AND METHODS ................................................................................. 57
4.3 RESULTS ................................................................................................................... 62
4.4 DISCUSSION ............................................................................................................. 73
ACKNOWLEDGEMENTS ................................................................................................ 77
REFERENCES ..................................................................................................................... 78
CHAPTER 5: DOES PARTICULATE MATTER MODIFY THE
ASSOCIATION BETWEEN TEMPERATURE AND CARDIO-
RESPIRATORY DISEASES? ....................................................................... 83
ABSTRACT ...................................................................................................................... 84
5.1 INTRODUCTION ...................................................................................................... 85
5.2 MATERIALS AND METHODS ................................................................................. 86
5.3 RESULTS ................................................................................................................... 92
5.4 DISCUSSION ............................................................................................................. 103
ACKNOWLEDGEMENTS ................................................................................................ 108
REFERENCES ..................................................................................................................... 109
CHAPTER 6: DOES TEMPERATURE MODIFY SHORT-TERM
EFFECTS OF OZONE ON TOTAL MORTALITY IN 60 LARGE US
COMMUNITIES? AN ASSESSMENT USING THE NMMAPS DATA
................................................................................................................. 114
XIII
ABSTRACT ...................................................................................................................... 115
6.1 INTRODUCTION ...................................................................................................... 116
6.2 MATERIALS AND METHODS ................................................................................. 117
6.3 RESULTS ................................................................................................................... 123
6.4 DISCUSSION ............................................................................................................. 128
ACKNOWLEDGEMENTS ................................................................................................ 134
REFERENCES ..................................................................................................................... 135
CHAPTER 7: OZONE, TEMPERATURE AND CARDIOVASCULAR
MORTALITY IN 95 US COMMUNITIES, 1987-2000 – ASSESSMENT
USING NMMAPS DATA ............................................................................. 139
ABSTRACT ...................................................................................................................... 140
7.1 INTRODUCTION ...................................................................................................... 142
7.2 MATERIALS AND METHODS .............................................................................. 143
7.3 RESULTS .................................................................................................................. 147
7.4 DISCUSSION ............................................................................................................ 152
ACKNOWLEDGEMENTS ................................................................................................ 157
REFERENCES ..................................................................................................................... 158
CHAPTER 8: OZONE MODIFIED ASSOCIATION BETWEEN
TEMPERATURE AND CARDIOVASCULAR MORTALITY IN 60
EASTERN US COMMUNITIES IN WARM DAYS, 1978-2000
......................................................................................................................... 162
ABSTRACT ...................................................................................................................... 163
8.1 INTRODUCTION ...................................................................................................... 165
XIV
8.2 MATERIALS AND METHODS ................................................................................. 166
8.3 RESULTS ................................................................................................................... 170
8.4 DISCUSSION ............................................................................................................. 175
ACKNOWLEDGEMENTS ................................................................................................ 179
REFERENCES ..................................................................................................................... 180
CHAPTER 9: GENERAL DISCUSSION ................................................... 184
9.1 OVERVIEW ................................................................................................................... 184
9.2 RESEARCH HYPOTHESIS .......................................................................................... 184
9.3 KEY FINDINGS ............................................................................................................ 185
9.4 METHODOLOGICAL DEVELOPMENT ................................................................... 187
9.5 INTERPRETATION OF FINDINGS ............................................................................. 189
9.6 STRENTHS AND LIMITATIONS................................................................................. 198
9.7 OPPORTUNITIES FOR FUTURE RESEARCH............................................................ 199
9.8 PUBLIC HEALTH IMPLICATIONS ............................................................................ 202
9.9 CONCLUSIONS ............................................................................................................. 204
REFERENCES .............................................................................................. 205
XV
LIST OF TABLES
TABLE 2.1: TIME-SERIES STUDIES OF SHORT-TERM HEALTH EFFECTS OF AIR
POLLUTION AFTER 2000 ........................................................................... 26
TABLE 4.1: SUMMARY STATISTICS OF HEALTH OUTCOMES, AIR POLLUTION
AND METEOROLOGICAL CONDITIONS IN BRISBANE, AUSTRALIA,
1996-2001 ..........................................................................................................62
TABLE 4.2: COEFFICIENTS OF MAIN AND INTERACTIVE EFFECTS OF MAXIMUM
TEMPERATURE AND PM10 ON MORBIDITY/MORTALITY .................. 69
TABLE 4.3: PERCENT INCREMENT OF MORBIDITY/MORTALITY PER 10µg/m³
INCREASE IN PM10 ACROSS TEMPERATURE LEVELS .......................... 72
TABLE 4.4: PERCENT INCREMENT OF MORBIDITY/MORTALITY PER 10µg/m³
INCREASE IN PM10 ACROSS TEMPERATURE LEVELS USING
DIFFERENT CUT-OFFS ON CURRENT DAY ............................................. 72
TABLE 5.1: SUMMARY STATISTICS FOR HEALTH OUTCOMES, AIR POLLUTANTS
AND METEOROLOGICAL CONDISITONS ................................................ 92
TABLE 5.2 COEFFICIENTS FOR MAIN AND INTERACTIVE EFFECTS OF
MINIMUM TEMPERATURE (ºC) AND PM10 (µg/m³)ON MORBIDITY/
MORTALITY .................................................................................................. 99
TABLE 5.3 PERCENT CHANGE IN CARDIO-RESPISTORY MORBIDITY/
MORTALITY PER 10°C INCREASE IN TEMPERATURE ACROSS THE
LEVELS OF PM10 .......................................................................................... 102
TABLE 6.1 MEANS OF DAILY OZONE, MAXIMUM TEMPERAUTRE AND THEIR
PEARSON CORRELATION COEFFICIENTS OF 60 EASTERN US
COMMUNITIES BETWEEN APRIL AND OCTOBER, 1987-
2000 ................................................................................................................ 123
XVI
TABLE 7.1 PEARSON CORRELATION COEFFICIENTS BETWEEN DAILY
MAXIMUM TEMPERATURE AND OZONE ACROSS DIFFERENT
REGIONS DURING THE SUMMER (MAY TO OCTOBER) ....................148
TABLE 7.2 PERCENTAGE CHANGES IN DAILY CARDIOVASCULAR MORTALITY
PER 10 PPB INCREASE IN OZONE ACROSS REGIONS AND
TEMPERATURE LEVELS DURING THE SUMMER USING BAYESIAN
META-ANALYSIS (LOG RELATIVE RATE) ........................................... 152
TABLE 8.1 DESCRIPTION SUMMARIES FOR MEANS OF DAILY OZONE,
MAXIMUM TEMPERATURE AND THEIR PEARSON CORRELATION
COEFFICIENTS OF 60 EASTERN US COMMUNITIES DURING WARM
DAYS, 1987-2000 .......................................................................................... 171
TABLE 8.2 PERCENT CHANGE IN DAILY CARDIORESPIRATORY MORTALITY
PER 10 ºC INCREASE IN MAXIMUM TEMPERATURE (THREE
PREVIOUS DAY AVERAGE) ACROSS REGIONS AND LEVELS OF
OZONE (THREE PREVIOUS DAY AVERAGE) DURING THE WARMER
DAY IN 60 EASTERN US COMMUNITIES (LOG RELATIVE RATE) .... 174
TABLE 9.1 SUMMARY OF INTERACTIVE EFFECTS BETWEEN PM10 AND
TEMPERATURE ON DIFFERENT HEALTH OUTCOMES IN BRISBANE,
AUSTRALIA, 1996-2001 ............................................................................... 186
TABLE 9.2 SUMMARY OF INTERACTIVE EFFECTS BETWEEN TEMPERATURE
AND OZONE ON DIFFERNET HEALTH OUTCOMES IN THE US
COMMUNITIES, 1996-2000 ........................................................................ 187
XVII
LIST OF FIGURES
FIGURE 2.1 HARVESTING PHENOMENON .................................................................. 32
FIGURE 3.1 REGIONAL GROUPS OF US OMMUNITIES ........................................... 44
FIGURE 4.1 JOINT PM10-TEMPERATURE RESPONSE SURFACES ON HEALTH
OUTCOMES AT LAG 2 ................................................................................ 64
FIGURE 4.2 DOSE-RESPONSE ASSOCIATION BETWEEN MZXIMUM
TEMPERATURE AND HEALTH OUTCOMES IN BRISBANE,
AUSTRALIA, 1996-2001 ................................................................................. 66
FIGURE 4.3 THE RELATIONSHIPS BETWEEN RESIDUALS AND DAYS OF YEAR
FOR RHA, CHA, REV, CEV, NECM AND CM AT ALG 2 ...........................67
FIGURE 4.4 THE SYNERGISTIC EFFECT BETWEEN MAXIMUM TEMPERATURE (°C)
AND PM10 ON RHA (TOP PANEL ) AND NECM (BOTTOM PANEL) ON
THE CURRENT DAY........................................................................................70
FIGURE 5.1 TIME SERIES DISTRIBUTION OF PM10, MINIMUM TEMPERATURE
AND HEALTH OUTCOMES DURING 1996-2001 IN BRISBANE ............. 93
FIGURE 5.2 TEMPERATURE-MORBIDITY/MORTALITY RELATIONSHIPS ............. 94
FIGURE 5.3 RELATIONSHIPS BETWEEN RESIDUALS AND DAYS OF YEAR FOR
DIFFERETN HEALTH OUTCOMES...............................................................96
FIGURE 5.4 BIVARIATE RESPONSE SURFACES OF MINIMUM TEMPERATURE
AND PM10 ON HEALTH OUTCOMES ON CURRENT DAY ..................... 98
FIGURE 5.5 THE SYNERGISTIC EFFECTS BETWEEN MINIMUM TEMPERATURE (ºC)
AND PM10 ON RHA (TOP PANEL) AND NECM (BOTTOM PANEL) ON
THE CURRENT DAY......................................................................................100
XVIII
FIGURE 6.1 PEARSON’S CORRELATION COEFFICIENTS BETWEEN DAILY OZONE
AND MAXIMUM TEMPERATURE WITH LATITUDE AND MEANS OF
MAXIMUM TEMPERATURE BETWEEN APRIL AND OCTOBER, 1987-
2000 IN 60 EASTERN US COMMUNITIES ................................................ 124
FIGURE 6.2 BIVARIATE RESPONSE SURFACES OF THREE-DAY MOVING
AVERAGES OF MAXIMUM TEMPERATURE AND OZONE ON TOTAL
NON-EXTERNAL DEATHS BETWEEN APRIL AND OCTOBER, 1987-
2000 ................................................................................................................. 125
FIGURE 6.3 PERCENT INCREASE IN DAILY MORTALITY PER 10 PPB INCREASE IN
THREE DAY AVERAGES OF OZONE ...................................................... 127
FIGURE 7.1 BIVARIATE RESPONSE SURFACES OF OZONE AND TEMPERATURE
ON CARDIOVASCULAR MORTALITY BETWEEN MAY AND OCTOBER,
1987-2000 ....................................................................................................... 149
FIGURE 7.2 COMMUNITY-SPECIFIC, REGIONAL AND NATIONAL BAYESIAN
ESTIMATES ACROSS TEMPERATURE LEVELS .................................... 151
FIGURE 8.1 BIVARIATE RESPONSE SURFACES OF THREE-DAY MOVING
AVERAGES OF MAXIMUM TEMPERATURE AND OZONE ON TOTAL
NON-EXTERNAL DEATHS DURING THE WARMER DAYS, 1987-
2000 ................................................................................................................. 172
XIX
LIST OF ABBREVIATION
CEV: Cardiovascular emergency visits
CHA: Cardiovascular hospital admissions
CI: Confidential interval
CM: Cardiovascular mortality
CVD: Cardiovascular deaths
EPA: Environmental protection agency
GAM: Generalized additive models
GLM: Generalized linear models
ICD: International Classification of Diseases
iHAPSS: Internet-based Health and Air Pollution Surveillance System
NECM: Non-external cause mortality
NMMAPS: The National Morbidity, Mortality, and Air Pollution Study
NO2: Nitrogen Dioxide
O3: Ozone
PI: Posterior interval
PM: Particulate matter
PM10: Particulate matter less than 10µm in aerodynamic diameter
PM2.5: Particulate matter less than 2.5µm in aerodynamic diameter
REV: Respiratory emergency visits
RHA: Respiratory hospital admissions
SO2: Sulphur dioxide
TSP: Total suspended particles
1
Chapter 1: Introduction
1.1 Background
In the past decades, the public awareness of possible health impacts of air pollution has risen
considerably. Many epidemiological studies have shown that ambient particulate matter (PM),
sulphur dioxide (SO2), nitrogen dioxide (NO2) and ozone (O3) were associated with the
increased morbidity and mortality from respiratory and cardiovascular diseases (Dominici et
al. 2003a; Katsouyanni et al. 1997; Pope et al. 1995). Recently, several large multi-site time-
series studies have shown that PM and ozone are consistently associated with various health
outcomes (Bell et al. 2004; Dominici et al. 2006; Gryparis et al. 2004; Ito et al. 2005; Levy et
al. 2005; Samet et al. 2000a, b, c).
Similarly, many studies have shown that episodes of heat waves have caused excess death in
the world (Basu & Samet 2002; Jones et al. 1982; Martens 1998; Patz et al. 2000; Tertre et al.
2006; Vanhemes & Gambotti 2003). For example, during a heat wave in St. Louis in 1980,
the maximum temperature was over 37.8°C for 16 days in July and resulted in a 56.7%
increment in mortality (ie, 850 excess deaths) (Jones et al. 1982). A large number of
epidemiological time-series studies have also shown that ambient temperature is associated
with health outcomes, and J-, U- or V-shaped patterns are usually observed between
temperature and mortality (Basu & Samet 2002; Patz et al. 2000). For example, Currie et al
(2002) examined the relationships between temperature and total mortality in 11 US cities and
found that the relationships between temperature and mortality showed J- or U-shaped
2
patterns, especially in northern areas, such as Boston, Chicago, New York, Philadelphia and
Baltimore, but such U- or J-shaped were much less obvious in southern areas.
Over the last two decades, air pollutants have declined and their health effects of short
episodes became more difficult to ascertain in the developed countries (Nyber & Pershgen
2000). In recent air pollution and temperature epidemiological studies, ecological time-series
analysis designs have become dominant due to its advantages (Dominici et al. 2003a). The
major advantages of time-series study designs include: weather-driven variations in air
pollution concentrations produce large contrasts in exposure over time; population act as their
own controls; and studies can often use routinely collected monitoring data including health
outcomes, air pollution and weather conditions. As a result, the number of deaths or hospital
admissions in studies can easily be done in the hundreds of thousands, leading to sufficient
statistical power to detect small adverse health effects of air pollution and temperature
(Brunekreef & Holgate 2002).
In time-series studies, the associations of air pollution and temperature with health outcomes
can be confounded to some extent by many covariates, such as weather variables, seasonality,
short-term variation, long-term trend, and other co-existing factors (Samet et al. 2000c;
Zanobetti et al. 2000a, b; Zanobetti et al. 2002). In order to control for potential confounders,
generalized linear models (GLM) and generalized additive models (GAM) have been widely
adopted in the assessment of the relationship between exposure to air pollutants or ambient
temperature and health outcomes (Dominici et al. 2003a). As the GAM can include
nonparametric smooth functions to account for the potential nonlinear effects of confounding
factors on health outcomes, such as weather conditions and seasonal variables (Hastie &
Tibshirani 1990; Katsouyanni et al. 1997; Samet et al. 1998; Schwartz 2000a, b, c), it has
3
been widely used in the evaluation of the impacts of air pollution and temperature on health
outcomes. GLM with parametric smoothing splines (B-spline or natural cubic spline) is also
often adopted in the assessment of air pollution or temperature effects in time-series studies
(McCullagh & Nelder, 1989).
Although a great number of time-series studies have reported the association of air pollution
and temperature with health outcomes, it remains largely unclear whether there is a
synergistic effect of air pollution and temperature on human health. Some studies have
investigated the interaction between season and air pollutants (Katsouyanni et al. 1997; Levy
et al. 2005; Smith et al. 2000; Wong et al. 2001). For example, Levy et al. (2005) conducted a
meta-analysis for the health effects of ozone and found that the magnitude of the ozone-
mortality relationship differs substantially across seasons (higher in summer than in winter).
Several studies have examined whether temperature modifies the association of air pollution
with morbidity or mortality, but they produced contradictory results. Some authors found the
synergistic effects between weather conditions and air pollution on health outcomes (Choi et
al. 1997; Katsouyanni et al. 1993; Roberts 2004), but others did not (Samet et al. 1998). So far,
few studies have explored whether air pollution modifies the association of temperature with
health outcomes although some recent studies have adjusted for air pollution in the
assessment of temperature effects (O’Nell et al, 2003; Raimham & Smoyer-Tomic 2003).
Therefore, the primary aim of this thesis was to investigate whether there were interactive
effects between temperature and air pollution on morbidity/mortality.
4
1.2 Aims and Hypothesis
This study aimed to examine interactive effects between temperature and air pollution on
health outcomes in Brisbane, Australia, and in ninety-five large communities, in the United
States. It is biologically plausible that temperature and air pollution symmetrically modify
each other’s effects because a lot of studies show that air pollutants have adverse impacts on
human health (Holgate et al. 1999), and that marked changes in ambient temperature can
cause physiological stress and alter a person’s physiological response to toxic agents (Gordon
2003). This study used two datasets mainly because of two reasons: firstly, it is necessary to
examine the consistency of research findings across different cities and countries; secondly,
these datasets were readily available for this study. At the early stage of the thesis, we just
used the single-site dataset collected from Brisbane to examine whether particulate matter and
temperature interacted to affect various health outcomes. At the late stage, we used the multi-
site dataset from 95 large US communities to examine interactive effects of ozone and
temperature on mortality (Bell et al. 2004) because most of the US communities just
monitored particulate matter once six days but they recorded daily monitored data on ozone
during the peak season. Therefore, we used the US dataset to examine interactive effects of
ozone and temperature on mortality using the multi-site design.
1.2.1 Specific Objectives
To visualize the associations of temperature and particulate matter less than 10µm in
aerodynamic diameter (PM10) with morbidity and mortality in Brisbane, Australia.
To visualize the associations of temperature and ozone with morbidity and mortality in
95 large US communities.
5
To visualize the joint effects of both temperature and PM10 on morbidity and mortality
in Brisbane, Australia.
To visualize the joint effects of both temperature and ozone on mortality in 95 large
US communities.
To quantitatively estimate the interactive effects between temperature and PM10 on
morbidity and mortality in Brisbane, Australia.
To quantitatively estimate the interactive effects between temperature and ozone on
mortality in 95 large US communities.
1.2.2 Hypothesis to Be Tested
Joint effects exist between temperature and ozone or PM10 on morbidity and mortality in
Brisbane, Australia and 95 large US communities.
1.3 Significance of the Study
This study systematically explored interactive effects between temperature and air pollutants
on morbidity and mortality in different countries and multi-cities. It is the first investigation
of the symmetrical effect modification of air pollution and temperature across geographic
regions. This study used three parallel time-series models to investigate the interactive effects
between temperature and air pollution on health outcomes, and this methodological approach
is useful in exploring the effect modification of temperature and air pollution. This study also
developed multi-stage Bayesian meta-analyses to estimate overall effects of temperature and
air pollution (eg, ozone). Therefore, the findings of this study contribute to both the
6
estimation of the health effects of air pollution and temperature, and to methodological
advance in epidemiological research.
1.4 Contents and Structure of this Thesis
This thesis is presented in the publication style. As such, it consists of five manuscripts, each
designed to stand on its own. Chapter 2 provides a critical literature review to cover both
previous research findings and current knowledge gaps in this area. Chapter 3 describes the
study design, materials and data analytical protocol.
The five manuscripts are presented in Chapters 4-8. Each manuscript is written in the
conventional publication style according to a particular journal, including the reference styles.
Because each manuscript was designed to stand alone, there was an inevitable degree of
repetitiveness in their introduction, methods and discussion.
The first manuscript examined whether temperature modified the effects of particulate matter
on cardiorespiratory hospital admissions, emergency visits and mortality at each of 0-2 lags in
Brisbane, Australia, 1996-2000. The second manuscript examined whether particulate matter
modified the temperature effects on cardiorespiratory hospital admissions, emergency visits
and mortality at each of 0-2 lags in Brisbane, 1996-2000. The third manuscript investigated
whether temperature modified ozone effects on non-external mortality and whether the
modification was heterogeneous across different regions in 60 large US communities, 1987-
2000. The fourth manuscript examined whether temperature modified ozone effects on
cardiorespiratory mortality in 95 large US communities in summer season (April to
September) through 1987-2000. It also examined whether the modification was
7
heterogeneous across geographic regions. The last manuscript investigated whether ozone
modified the association between temperature and non-external mortality in 60 large US
communities.
Chapter 9 summarizes the key findings of Chapters 4-8 and then discusses issues raised in the
previous chapters. This chapter also discusses biological plausibility, strengths, limitations,
opportunities for future research, and public health implications. Finally, the conclusions are
made based on the findings observed in the five manuscripts.
Tables and figures are provided in the text to facilitate reading. The references for each of the
manuscripts are presented at the end of their corresponding chapters. A complete reference
list (including references cited in the Chapters 1, 2, 3 and 9) is provided at the end of the
thesis.
8
Chapter 2: Health Effects of Air Pollution
- A Literature Review
2.1 Introduction
It is well known that exposure to high levels of air pollution can adversely affect human
health. A number of air pollution catastrophes occurred in industrial countries between 1950s
and 1970s, such as the London smog of 1952 (Bell et al. 2001; Minister of Health 1954).
Subsequent legislation has effectively controlled the emission of air pollutants, and air quality
in Western countries has significantly improved since the 1970s. However, adverse health
effects of exposure to low level of air pollution remain a regulatory and public concern,
motivated largely by a number of the recent epidemiological studies which have shown
positive associations between low levels of air pollution and health outcomes using time-
series designs (Samet et al. 2000a, b, c; Schwartz 2000a, b, c). Health effects of temperature
changes have also been recognised for a long time. Several international and regional
assessments have been conducted to address this issue (Le Tertre et al. 2006; Vandentrren et
al. 2004; Watson et al. 1998; WHO 1996). The projections of future climate change have
compelled health scientists to re-examine weather/disease relationships: including the health
impacts of temperature rise, sea level rise, and extremes in the hydrologic cycle (McMichael
et al. 2006; WHO 1996).
9
Air pollution and temperature epidemiological studies are broad topics and similar
methodologies are used in both types of research. To provide a literature review and
background in this area, this chapter highlights the major epidemiological studies of ambient
air pollution on morbidity and mortality over the last two decades because the methodologies
for temperature epidemiological studies are similar. Firstly, we briefly discuss main air
pollutants and their biological mechanisms. Secondly, we focus on main epidemiological
research designs in current air pollution studies. Then, we review the main findings in the
epidemiological studies of air pollution. Finally, we attempt to identify research opportunities
for future air pollution epidemiological studies.
2.2 Main Air Pollutants and Biological Mechanisms
Air pollution is a complex mixture of chemicals. It has different impacts on the atmosphere
and human health and well-being depending on the source of the pollutants and levels of
human exposure. The following section provides an overview of the diverse array of air
pollutants and the relevant mechanisms of their health effects.
2.2.1 Carbon Monoxide (CO)
Carbon monoxide is a colourless and odourless gas with approximately the same density as
air. It is produced when substances containing carbon are combusted with an insufficient
supply of air. CO can have serious health impacts on humans and animals. When inhaled, CO
binds with haemoglobin in the blood and forms a very stable carboxyhaemoglobin complex.
This reduces the capacity of the haemoglobin to transport oxygen in the blood stream and
10
decreases the supply of oxygen to tissues and organs, especially the heart and brain (Maynard
& Waller 1999).
2.2.2 Ozone (O3)
Ozone is an indicator of photochemical smog and is a strong oxidising agent produced by
photochemical reactions involving other air pollutants, such as NOx (WHO 2000).
Concentrations in city centres tend to be lower than those in suburbs, mainly as a result of
scavenging of ozone by nitric oxide originating from traffic. It has been shown experimentally
that exposure to O3 can affect the human cardiac and respiratory systems, and irritate the eyes,
nose, throat, and lungs (Chan-Yeunf 2000; Schwela 2002; Thurston & Ito 1999).
2.2.3 Nitrogen Oxide (NOx)
Nitrogen oxide (NOx), a mixture of nitric oxide (NO) and nitrogen dioxide (NO2), is formed
from natural sources, motor vehicle emissions, and other fuel combustion processes (WHO
2000). Nitric oxide is colourless and odourless and is rapidly transformed into nitrogen
dioxide in the atmosphere in reaction with atmospheric oxidants such as ozone (Brunekreef &
Holgate 2002). Nitrogen dioxide is an odorous, brown, acidic, highly-corrosive gas and can
affect human health and the environment. Elevated levels of nitrogen dioxide cause damage to
the mechanisms of the human respiratory tract and can increase one’s susceptibility to
respiratory infection and asthma (Brunekreef & Holgate 2002). Long-term exposure to high
levels of NO2 can result in chronic lung diseases. It may also affect sensory perception by
11
reducing a person’s ability to smell an odour (Ackermann-Liebrich & Rapp 1999; CDHAC
2000; Queensland EPA 2007).
2.2.4 Particulate Matter (PM)
Ambient particulate matter is the term used to describe particles that suspends in the air.
Particulate matter is a mixture of solid, liquid, or solid and liquid particles. The main sources
of ambient particles are fossil fuel combustion, biomass burning, and the processing of metals.
Road transportation is the major source in urban areas (Holmam 1999; Pooley & Mille 1999).
Size is the main determinant of the behaviour of ambient particles. In practical terms, a
distinction is made in terms of the “aerodynamic diameter” which refers to unit density of
spherical particles with the same aerodynamic properties, such as falling speed. Recent
attention has focused on PM10 (“thoracic” particles smaller than 10 μm in aerodynamic
diameter), PM2.5 (“respiratory” particles smaller than 2.5μm in aerodynamic diameter) (WHO
2000). Their biological effects are related to the sizes of the particles. Particles with an
aerodynamic diameter of more than 5 μm are often deposited in the upper or larger airways
and smaller particles or PM2.5 are often deposited in the small airways (bronchioli) and
alveoli. Fine particles have the potential to penetrate the airway epithelium and vascular walls
(Jeffrey 1999). Therefore, small particles have stronger effect (Rom and Samet 2006).
12
2.2.5 Sulphur Dioxide (SO2)
Sulphur dioxide is a colourless gas with a sharp, irritating odour. It is produced from the
burning of fossil fuels (coal or oil) and the smelting of mineral ores that contain sulphur.
When SO2 combines with water, it forms sulphuric acid, which is the main component of acid
rain. Sulphur dioxide can affect the respiratory system, damaging the function of the lungs
and irritate eyes. When sulphur dioxide irritates the respiratory tract, it causes coughing and
mucus secretion, aggravates conditions such as asthma and chronic bronchitis, and makes
people more prone to respiratory tract infections. Sulphur dioxide can also attach itself to
particles which, if particles are inhaled, can cause serious damage (Schlesinger 1999).
2.3 Main Research Designs in Air Pollution Epidemiological
Studies
Observational studies are dominant in air pollution epidemiological studies, most of which are
“opportunistic”, combining data from different sources which have been collected for other
purposes (Dominici et al. 2003a). One of the main reasons to facilitate this type of research is
the availability of ambient monitoring data, meteorological records and regular health event
registrations in many parts of the world, particularly in developed countries. Monitors are
usually located at fixed sites and record the concentrations of local ambient pollutants and
meteorological conditions on a regular basis. These time-series data are widely used in
epidemiologic studies of air pollution (Bell et al. 2004; Brunekreef & Holgate 2002;
Dominici et al. 2006).
13
Health effects of air pollution can be either acute or chronic (American Thoracic Society
2000). Acute effects are due to short-term and transient exposure. Chronic effects are more
likely due to the cumulative effects of exposure, but could be associated with more complex
functions of lifetime exposure. Health outcomes can include major or minor life events (e.g.,
death or onset of symptoms), changes in function (e.g., vital capacity, lung growth and
symptom severity) or biomarkers (American Thoracic Society 2000). The nature of the
outcomes (e.g., binary or continuous) and the structure of the data decide the selection of an
appropriate model and the types of effects to be estimated. Regression models are generally
the methods of choice.
Most recent air pollution epidemiological studies belong to the following types: ecological
time-series, case-crossover or case-control, panel, and cohort study designs. Time series, case-
crossover, and panel designs are appropriate for estimating the acute effects of air pollution,
while the cohort study design is suited to estimate both acute and chronic effects (Dominici et
al. 2003a). A case-crossover study is a specialised case-control design, suited to the
examination of a transient effect of an intermittent exposure. Therefore, case-control and
case-crossover designs are discussed in the same section. Panel studies collect individual
temporally and spatially-varying exposures, confounders and outcomes, which are based on
spatially and /or temporally aggregated data (Dominici et al. 2003a; WHO 2004). In practice,
panel studies also depend on group-level data, and therefore, panel studies combine different
basic epidemiological study designs. Cross-sectional studies compare the prevalence of health
outcomes in different populations at the time of ascertainment while the exposure to air
pollutants is simultaneously measured. This study is generally the most economic and feasible
(Rothman and Greenland 1998). However, a major weakness is that the data are collected at
one point in time and therefore it fails to consider temporality, which is an important criterion
14
for causality. Because of its disadvantages it is infrequently employed in air pollution studies
(Samet and Jaakkola 1999), and this review therefore does not consider the cross-sectional
design. The following sections will discuss the different types of epidemiological designs
used in air pollution epidemiological studies.
2.3.1 Time-series Study Design
In western countries, concentrations of air pollutants are now generally much lower than 30
years ago and their short-term health effects have become quite small. Thus, traditional
methods may fail to detect their effects (Brunekreef & Holgate 2002). Time-series study
designs have dominated in air pollution research for the last two decades. Time-series studies
associate time-varying exposure with time-varying event counts. These are ecologic study
designs because they analyse daily population-averaged outcomes and exposure levels
without individual information. In time-series studies, the generalised linear model (GLM)
with parametric splines (e.g. B-spline or natural cubic spline) (McCullagh & Nelder 1989)
and generalised additive models (GAM) with non parametric splines (e.g. cubic smoothing
spline or LOESS smoother) (Hastie & Tibshirani 1990) are usually adopted to estimate effects
of exposure to air pollution on human health (morbidity or mortality). Recently, the GAM has
been widely applied in the assessment of air pollution effects because it allows for non-
parametric adjustment for non-linear confounding effects such as seasonality and weather
conditions, and is a more flexible approach than fully-parametric alternatives like the GLM
with natural cubic spline (Dominici et al 2003a). The following section will focus on the
application of a GAM approach. The formula for GLM is similar to that of GAM.
15
GAM assumes that the daily number of cases tY has an overdispersed distribution
( ttYE ( ), ttY ]var[ ) (Dominici et al. 2004). GAM model is presented as follows:
t
l
iltit DOWtimesconfsXYELog
),(),())((
0
(1)
where tX means daily levels of exposure at time t. DOW is the indicator variable for day of
the week, which is used to adjust for short-term fluctuation because the patients are reluctant
to attend hospital on weekends and air pollution also is weaker due to fewer cars on roads on
weekends (Barnett and Dobson, 2005). is a vector of coefficients. time is referred to as
calendar time or some function transfer to adjust for the seasonal fluctuation. Conf means
other potential confounders in the models, such as other air pollutants and meteorological
variations. l means the lags of the exposure. The smooth function )(.,s denotes a smooth
function of a covariate often constructed using smoothing splines, LOESS smoothers, or
natural cubic splines with a smoothing parameter . The smooth parameter represents the
number of degrees of freedom in the smooth spline, the span in loess smoother, -2 interior
knots in the natural cubic splines. The parameter of interest β represents the changes in the
logarithm of the population average mortality count per unit change in ltX .
It is unnecessary to include an over-dispersion parameter in GAMs. Over-dispersion is a
separate statistical matter to GAMs. Using over-dispersion makes GAMs more flexible than
the standard Poission model by breaking the assumption of the mean equalling the variance.
Also, it would have been interesting to see the actual over-dispersion parameter in the results
(Hastie and Tibshirani 1990).
16
Since the mid 1990s, a growing number of studies have used distributed lag models and time-
scale models to estimate the cumulative and longer time-scale health effects of air pollution.
Distributed lag models (Almon 1965; Zanobetti et al. 2000b, 2002) are used to estimate
associations between health outcomes on a given day t, and air pollution several days prior by
replacing
l
ilti X
1
in (1) with ltl
ll X
1
, ltl
ll X
1
=1 where θ measures the cumulative
effect, and l measures the contribution of the lagged exposure ltX to the estimation of θ.
Time-scale models (Dominici et al. 2003b; Schwartz 2001; Zeger et al 1999) are used to
estimate the relationship between smooth variations of air pollution and daily health outcomes
by replacing
l
ilti X
1
in (1) with tkk
kkW
1
, ttkk
kk XW
1
, where ,...,,...,1 ktt WW KtW is a set of
orthogonal predictors obtained by applying a Fourier decomposition to tX . The
parameter k represents the log relative rate of the health outcome for increment of air
pollution at time scale k. Time scales of interest may be short- (1 to 4 days) and longer-term
variations (1 to 2 months) of air pollution. Beyond two months, any effects are possibly
dominated by seasonal confounding (Zeger et al. 1999).
Some time-series studies have examined synergistic effects between air pollution and
temperature by including an interactive term in GAM models (Morris and Naumova 1998;
Roberts 2004). For example, Roberts (2004) used bivariate models to examine the joint effect
of PM10 and temperature on mortality in two US cities and found that temperature might
synergistically modify the effect of particulate matter. Morris and Naumova (1998) reported
that the association between carbon oxide and hospital admissions varied with temperature.
When temperature was high, the association was strong.
17
Several studies have discussed how to choose the parameters of smoothing (Cakmak et al.
1998; Kelshal et al. 1997). For seasonality, the parameter should be big enough to eliminate
fluctuation of seasonal changes so that a shorter-term variation in health outcomes and
exposure is less than 2 months (Kelshal et al. 1997; Samet at al. 1995). To control weather
variation, a suitable parameter of smoothing is also necessary. For example, in many US cities
mortality decreases smoothly with increased temperature before a certain temperature point
and then increases quite sharply with temperature above a certain temperature point. Here a
smoothness parameter greater than three is necessary to capture the highly non-linear bend in
mortality as a function of temperature (Dominici et al. 2003a).
There are several potential biases with gam function in S-Plus software. Two main sources of
bias have been identified when gam function in S-Plus software is used to fit a Poisson
regression model in air pollution time-series studies (Dominici et al. 2002; Rasmay et al.
2003). One arises from the default criteria of convergence and could be reduced by the use of
the stringent criteria, say 1.0×10-10 (Dominici et al. 2002). Another source of bias in the gam
function is due to concurvity, which has the potential to reduce underestimation of standard
errors of coefficients for parametric terms when a semi-parametric model is fitted (Ramsay et
al. 2003). The gam.exact function was developed to correct the concurvity problem in gam
function (Dominici et al. 2004). However, when we conducted a Poisson regression with the
stringent criteria for convergence using gam.exact function to calculate asymptotically exact
standard errors, the convergence criteria could not be obtained with some datasets even
though we used a large number of iterations, say 5000. In this situation, GLM is a good option
because the maximum-likelihood estimates of the parameter β could be obtained by
iteratively-reweighted least squares (IRLS) in GLM (McCullagh and Nelder 1989).
18
There are several advantages and disadvantages with time-series study designs. The first
advantage is the ability to use a large amount of available monitoring data to control for long-
or short-term variance with smoothing functions without rigid linear assumption between
outcome and predictors. The second advantage is that it is suited to estimate the short-term
effect of exposures. The third advantage is that the design is economic and statistically
powerful enough to estimate the weak effect of air pollution or weather conditions due to a
large amount of available monitoring data on air pollution, weather conditions, and regulatory
health event registrations. However, this study design has several disadvantages as well. First,
it is difficult to determine the best model, such as how to select the degrees of freedom, how
to determine lag effects, and how to control for other covariates with lag effects. Secondly, it
would be likely to produce biased estimates due to lack of individual-level exposures. In time-
series studies, air pollution and meteorological measures used are generally based on
monitoring records which may be poorly represent of individual exposures. Thirdly, it is
difficult to control for some potential confounders resulting from individual-level covariates
and human behaviour changes. Finally, it is unable to be used to estimate long-term or
cumulative effects of exposure.
2.3.2 Case-Crossover Study (or Case-Control Study) Design
In a case-control study, cases that develop a certain event are identified and their past
exposure to suspected aetiological factors is compared with that of controls or referents that
do not experience the event (Rotheman and Greenland, 1998). The case-crossover design is a
special case of a case-control design, suited to the study of a transient effect of an intermittent
exposure on the subsequent risk of a rare acute-onset disease hypothesised to occur a short
time after exposure (Maclure 1991). The principle of the case-crossover study design is that
19
exposures of cases just prior to the event are compared with the distribution of exposures of
the same cases from separate time periods. Therefore, it can be considered as a modification
of the matched case-control design (Schlesselman 1994). More specifically in a time-series
study, the exposure at the time just prior to the event (the case or index time) is compared to a
set of controls or referents. In this way, some measured and unmeasured time-invariant
characteristics of the subject (such as gender, age and smoking status) are matched, and
therefore the potential confounding originating from those unmeasured characteristics is
minimised (Maclure 1991).
The first decade of practice with case-crossover study design has shown that this design
applies best if the exposure is intermittent and transient, and the effect is immediate and
abrupt (Janes et al. 2005; Maclure & Mittleman 2000). This design has been found to be
effective for estimating the risk of a rare event associated with a short-term exposure because
the widespread availability of ambient monitoring data presents opportunities to further
analyse existing case series from case-control studies (Levy et al. 2001b).
Here a fixed number of referent days before and possibly after the case in a short time frame
are used, in order to restrict the referents in time to reduce seasonal confounding (Lumley &
Levy 2000). The data in the case-crossover design consist of exposure measures itZ for
subject i = 1,…, n at time t = 1,…,T. The referent sampling scheme determines the referent
sets iW of exposure that are used in the analysis. Estimates of the relative risks are obtained
by the following equation:
ii
is
it
iWt Ws
X
Xit
itie
eZZU
)( (2)
20
This is the “conditional likelihood” estimating function typically used in environmental and
other case-crossover studies (Lee & Schwartz 1999; Mittelman et al. 1995; Neas et al. 1999).
Standard computer softwares such as SAS PHREG (SAS Institute Inc 2004) can be used for
this analysis.
A key difficulty in case-crossover studies is how to properly define the referent sets iW .
Control for bias in the estimation of the relative risk β is the dominant concern in the choice
of the referent sampling strategy, although the size of the referent set also affects efficiency.
Two main sources of bias in case-crossover studies have been identified (Janes et al 2005;
Lumley & Levy 2000). The first source of bias arises from the trend and seasonality in the
time series analyses of air pollution health effects. Since case-crossover comparisons are
made within subjects at different points in time, the case-crossover analysis implicitly
depends on the assumption that the exposure distribution is stationary. The long-term time
trends and seasonal variation inherent in the time series violate this assumption (Baseson &
Schwartz 1999, 2001; Levy et al. 2001a; Navidi 1998). The second source of bias is called
overlap bias. If the referent windows iW , i = 1, …, n, are exactly determined by the case
period and are not disjoint, then the independent sampling inherent in the conditional
likelihood approach is invalidated (Austin 1989; Lumley & Levy 2000). Lumley and Levy
(2000) quantified the overlap bias analytically and simulations studies showed that the
direction is unpredictable and that the magnitude is a function of the size of the coefficient β
(Lumley & Levy, 2000). However, for the small effects seen in exposure to air pollution,
current experience suggests that overlap bias is similar to the small-sample bias, e.g. the bias
obtained by estimating β in (2) with a small number of referents (Dominici et al. 2003a).
21
In order to control for bias in case-crossover studies, many authors have proposed several
approaches to select referents, including ambidirectional, symmetric bidirectional, semi-
symmetric directional and time-stratified sampling (Janes et al. 2005; Navidi 1998; Navidi &
Winhandle 2002; Schwartz and Lee 1999).
There are several major advantages to use a case-crossover study design in air pollution
epidemiology. Firstly, because the cases and controls in the design are the same subjects, it
can control for some unmeasured confounders of individual characteristics, which might
confound the estimates. Secondly, it is suited to estimate the short-term transient effects
(Levy et al. 2001). Thirdly, this design can control for the second variable although the
matching will consume a lot of time (Barnett et al. 2006; Schwartz 2005). For example, the
study design can match temperature highly correlated with ozone when estimating ozone
effects (Schwartz 2005). Additionally, bi-directional selection of control periods allows
individual adjustment for seasonal and secular trends (Jaakkola 2003). Finally, it can use a
large amount of available monitoring data to estimate effects of exposure. However, there are
also several disadvantages with a case-crossover design. Firstly, it is difficult to select
referents because an improper referent will result in a biased estimation (Levy et al. 2001a)
and matching will take a lot of time. Secondly, it usually makes a linear assumption between
the health outcome and the predictors. If the assumption is substantially violated, bias may be
induced. For example, some studies have shown that the relationship between temperature
and health outcomes is nonlinear (V or U pattern). Stratification for some variables could
solve the nonlinear problem (Barnett 2007), but it is difficult to determine cut-offs.
Additionally, compared with Poisson regression time-series analysis, the statistical power will
reduce approximately 50% (Bateson & Schwartz 1999; Jaakkola 2003). Finally, it is not
suitable for estimation of long-term or cumulative effects of exposure.
22
2.3.3 Panel Study
A panel study enrols a cohort of individuals at the commencement of the study and then
follows them over time to repeatedly measure changes in health outcomes (WHO 2004). The
exposure measurement could be from a fixed-site ambient monitor or personal monitors. The
panel study design is effective for assessing short-term health effects of air pollutants.
A variety of models have been used to estimate the effects of air pollutants on health in a
panel study setting depending on how to deal with the longitudinal data. Repeated
measurements of health outcomes and exposure in individuals results in the variation within
the individuals (Armitage et al. 2002). Modern approaches to accommodate the complex
variance induced by such longitudinal data include mixed, marginal, transition models and
Bayesian hierarchical models (Diggle et al. 1994) and the analysis can be accomplished by
several software packages such as SAS GENMOD, MIXED and GLIMMIX (SAS Institute
Int. 2006) and Bayesian software such as WinBugs package (Spiegelhalter et al. 2000). A
study that follows a panel of individuals over a longer time period, say multiple years, is
generally known as a cohort or longitudinal study rather than a panel study (Dominici et al.
2003a).
A panel study is often applied to estimate the short-term health effects of air pollution, but the
interpretation of these results is often unclear because it is difficult to observe all individuals
on the same days in a panel study. If all individuals are not observed at the same period, the
estimated effects of the exposure should be considered carefully (Dominici et al. 2003a).
Another concern in the interpretation of a panel study is the within-person variation which
23
may be induced by time-varying confounders such as indicators for day or week, function of
seasonality and weather, and time-varying personal behaviours (Dominici et al. 2003a).
2.3.4 Cohort Study
Cohort studies have been increasingly conducted to estimate long-term health effects of
exposure to air pollution over the last decade (Abbey et al. 1999; Dockery et al. 1993; Hoek et
al. 2002; Jalaludin et al. 2004; Klot et al. 2005; Pope et al. 2002). Both prospective (Filleul et
al. 2005) and retrospective (Hansen et al. 2006) cohort study designs have been used. In a
prospective cohort study design, participants are enrolled at the beginning of study to collect
information about age, sex, education and other subject-specific characteristics (Rothman and
Greenland 1998). They are followed up over time to collect information about health events,
exposure and confounders. A cumulative exposure is often used as the exposure variable
(Dockery et al. 1993). A key design consideration for air pollution cohort studies is to identify
a cohort with sufficient exposure variation in cumulative exposure, particularly when ambient
air pollution measurements are used (Pope et al. 2002). However, in maximizing the
geographical variability of exposure the relative risk estimates from cohort studies are likely
to be confounded by area-specific characteristics (Dominici et al. 2003a).
Survival analysis tools can be used to evaluate the association between air pollution and
mortality. Typically the Cox proportional-hazards model is used to estimate mortality rate
ratios while adjusting for potential confounding variables (McMichael et al. 1996; Watson et
al. 1998). Relative risk is estimated as the ratio of hazards for an exposed group relative to an
unexposed or reference group. The hazard for individual i, )(ti , is modelled as
24
)exp()()( 0 sConfounderZtt ijiij (3)
where )(0 ti represents the baseline hazard for the ith stratum at time t, ijZ is the long term
exposure for the individual. β is the coefficient or relative risk parameter, which is assumed to
be the same for all individuals across all strata. Time may be defined as calendar time or age.
Stratification of the baseline hazard function into disjoint groups makes the proportional
hazards assumption flexible and allows separate and not necessarily proportional hazards
across strata (Dominici et al. 2003a).
The cohort study design has some important advantages. Firstly, it can be used to estimate a
long-term or cumulative effect. Secondly, it is relatively easier to control confounders of
individual’s characteristics such as gender, age and occupational exposure. Additionally, it
provides the strongest evidence on causality in observational studies (Rothman and Greenland
1998). However, the disadvantages are also obvious. Firstly, it is sometimes unaffordably
expensive in air pollution studies. The reasons are that (1) the health effect of air pollution is
usually small and a large number of participants need to be followed up; (2) all the
participants need to be followed up for a relatively long time so that the long-term or
cumulative effects of exposure can be estimated. Secondly, it is still difficult to avoid
exposure misclassification because exposure to air pollution is still measured at fixed central
sites and air pollution levels vary across the geographic areas during the follow-up period.
Such the misclassification could be reduced to some degree by setting more monitoring
stations close to subjects or using individual-level monitors.
25
2.4 Health effects of Air Pollution
2.4.1 Time-series Studies
There have been numerous studies on the short-term health effects of air pollution, with
emphasis on mortality and hospital admissions. Time-series design is a major approach to
estimating short-term health effects of air pollution in epidemiological studies. Many studies
have found associations between daily changes in ambient air pollution and increased
cardiorespiratory hospital admissions (Anderson et al. 1996; Burnett et al. 1997; Linn et al.
2000; Moolgavkar et al.1997; Moolgavkar 2000; Morris et al. 1995; Schwartz 1999), along
with cardiorespiratory mortality (Hoek et al. 2001; Mar et al. 2000; Rossi et al. 1999) and all
cause mortality (Roemer and van Wijnen 2001). Table 2.1 summarises some recent time-
series studies on short-term health effects of PM10 and ozone around the world, omitting
earlier studies because early findings have already been reviewed (Brunekreef & Holgate
2002; Nyberg & Peshagen 2000).
Single-site time-series studies have been criticized because of substantial variation of the air
pollution effects and the heterogeneity of the statistical approaches used in different studies
(Dominici et al. 2006; Li and Roth 1995). Recently, several multi-site time-series studies have
been conducted in Europe and the United States. Two large collaborative air pollution
projects in Europe and U.S. are summarised below.
26
Table 2.1 Time-series studies of short-term health effects of air pollution after 2000
Study Pollutant Population Methodology Main findings
Czech Republic and rural
region in Germany (Peter et al.
2000)
TSPMortality 1982-
1994
Poisson regression
(GAM)
Czech Republic: 3.8% increase (95% CI: 0.8%, 9.6%) per 100
μg/m³; No evidence for association in the rural area in German at
the Czech border.
10 US cities (Schwartz 2000b) PM10Mortality 1986-
1993
Poisson regression
(GAM)
0.67% increase for a 10μg/m³ (95% CI: 0.52%, 0.81%). No
difference between summer and winter.
New Zealand (Hales et al.
2000)PM10
Mortality Jun
1988-Dec 1993
Poisson regression
(GAM)
1% increase for all-cause mortality (95% CI: 0.5%, 2.2%); 4%
increase for respiratory diseases (95% CI: 1.5%, 5.9%)
10 US cities (Schwartz 2000c) PM10Mortality 1986-
1993
Distributed lag
model (GAM)
1.4% (95% CI: 1.15%, 1.68%) increase for 10μg/m³ on a single
day using a quadratic distributed lag model; 1.3% increase (95%
CI: 1.04%, 1.56%) using an unstrained lag model
20 US cities (Samet et al.
2000c)
PM10, O3,
SO2, CO,
NO2
Mortality 1987-
1994
Poisson regression
(GAM)
PM10: 0.51% increase (95% CI: 0.07%, 0.93%) per 10 μg/m³ for
all causes; 0.68% increase per 10 μg/m³ for cardiovascular and
respiratory diseases (95% CI: 0.20%, 1.16%)
O3: weaker evidence during the summer;
Other pollutants: no evidence
27
Table 2.1 Time-series studies of short-term health effects of air pollution after 2000 (continued)
Study Pollutant Population Methodology Main findings
Hong Kong (Wong et al. 2002) PM10, SO2Morality 1995-
1998
Poisson regression
(GAM)
Significant associations were found between mortalities for all
respiratory diseases and ischaemic heart diseases (IDH). The
increases for all respiratory mortalities (for a 10 μg/m³ increase
in the concentration) are 0.8% (95% CI: 0.1%, 1.4%) for
PM10and 1.5% (95% CI: 0.1%, 2.9%) for SO2 ; the increases for
IDH are 0.9 % (95% CI: 0.0%, 1.8%) for O3 and 2.8% (95% CI:
1.2%, 4.4%) for SO2.
Seoul Korea (Kim et al. 2003) PM10Mortality 1995-
1999
Poisson regression
(GAM)
3.7% increase (95% CI: 2.1%, 5.4%) for non-accident causes,
13.9% increase (95% CI: 6.8%, 21.5%) for respiratory disease,
4.4 % increase (95% CI: -1.0%, 9.0%) for cardiovascular disease
and 6.3% increase (95% CI: 2.3%, 10.5%) for cerebrovascular
disease per interquartile increase of PM10 (43.12μg/m³)
Shanghai, China (Kan & Chen
2003)
PM10,
SO2, NO2
Mortality Jun 2000
to Dec 2001
Poisson regression
(GAM)
0.3% increase (95% CI: 0.1%, 0.5%) for PM10, 1.4% increase
(95% CI: 0.8%, 2.0%) for SO2 and 1.5% increase (95% CI:
0.8%, 2.2%) for NO2 per 10μg/m³
28
Table 2.1 Time-series studies of short-term health effects of air pollution after 2000 (continued)
Study Pollutant Population Methodology Main findings
Brisbane, Australia
(Petroeschevsky et al. 2001)
BSP, O3,
SO2, NO2
Hospital admission
1987-1994
Poisson regression
(GLM)
BSP: 1.5% increase (95% CI: 0.6%, 2.3%) for respiratory
diseases per 24-hr 10 5 /m increase.
O3: 2.3% increase (95% CI: 0.6%, 2.3%) for respiratory disease
per 8-hr unit increase (pphm).
SO2: 8.0% increase (95% CI: 3.0%, 13.1%) for respiratory
disease per 24-hr unit increase (pphm).
NO2: -0.1% increase (95% CI: -0.3%, 0.2%) for respiratory
disease per 1-hr-max unit increase (pphm).
29
Table 2.1 Time-series studies of short-term health effects of air pollution after 2000 (continued)
Study Pollutant Population Methodology Main findings
Brazil (Braga et al. 2001)
PM10, O3,
SO2, CO,
NO2
Respira-tory
disease Hospital
admission 1993-
1997
Distributed lag
model
9.4% increase (95% CI: 7.9%, 10.9%) for 2 or less years old
group and 7.0% (95% CI: 5.7%, 8.2%) for all age group per
interquartile PM10 increase (35μg/m³);
1.6% increase (95% CI: 0.1%, 3.0%) for 2 or less years old
group and 0.8% (95% CI: -7.5%, 9.2%) for all age group per
interquartile O3 increase (46μg/m³);
5.9% increase (95% CI: 4.5%, 7.4%) for 2 or less years old
group and 4.5% (95% CI: 3.3%, 5.8%) for all age group per
interquartile SO2 increase (14μg/m³);
5.0% increase (95% CI: 3.3%, 6.8%) for 2 or less years old
group and 4.9% (95% CI: 3.5%, 6.4%) for all age group per
interquartile CO increase (3ppm);
9.4% increase (95% CI: 6.2%, 12.6%) for 2 or less years old
group and 6.5% (95% CI: 3.3%, 9.7%) for all age group per
interquartile NO2 increase (80μg/m³);
30
In Europe, the APHEA (Air Pollution and Health: a European Approach) studies have
provided many new insights. Initial studies were based on older data (APHEA-1)
(Katsouyanni et al. 1996, 1997; Touloumi et al. 1996), and a new series of studies (APHEA-2)
used data of the PM10 fraction since the late 1990s (Atkinson et al. 2001; Katsouyanni et al.
2001; Tertre et al. 2002). The APHEA-2 mortality studies covered over 43 million people and
29 European cities, which were all studied for more than 5 years in the 1990s. The combined
effect estimate showed that all-cause daily mortality increased by 0.6% (95% CI 0.4%-0.8%)
for each 10 µg/m3 increase in PM10 from data involving 21 cities. It was found that there was
heterogeneity between cities with different levels of NO2. The estimated increase in daily
mortality for an increase of 10 μg/m³ in PM10 were 0.2% (95% CI: 0.0%, 0.4%), and 0.8%
(95% CI: 0.7%, 0.9%) in cities with low and high average NO2, respectively (Katsouyanni et
al. 2001). The APHEA-2 hospital admission study involved 38 million people living in eight
European cities. Hospital admissions for asthma and chronic obstructive pulmonary disease
(COPD) increased by 1.0% (95% CI: 0.4%, 1.5%) per 10 µg/m3 PM10 increment among
people older than 65 years (Atkinson et al. 2001). Tertre et al. (2002) reported that in 8
European cities, the pooled increase in cardiovascular admissions associated with a 10 μg/m³
increase in PM10 and black smoke was 0.7% (95% CI: 0.2%, 0.8%) and 1.1% (95% CI: 0.4%,
2.2%), respectively.
In the United States, the National Morbidity, Mortality and Air Pollution Studies (NMMAPS)
focused on the 20 largest metropolitan areas in the USA, involving 50 million inhabitants,
during 1987-94 (Samet et al. 2000a; 2000b; 2000c). All-cause mortality increased by 0.5 %
(95% CI: 0.1%, 0.9%) for each increase of 10 µg/m3 in PM10. The estimated increase in the
relative rate of death from cardiovascular and respiratory disease was 0.7 % (95% CI: 0.2%,
1.2%) (Samet et al. 2000c). Effects on hospital admissions were studied in ten cities with a
31
combined population of 1 843 000 individuals older than 65 years (Zanobetti et al. 2000a).
The model used considered simultaneously the effects of PM10 up to the lag of 5 days and
effects of PM10 on chronic obstructive pulmonary disease admissions to be 2.5% (95% CI:
1.8%, 3.3%) and on cardiovascular disease admissions to be 1.3% (95% CI: 1.0%, 1.5%) for
an increase of 10 µg/m3 in PM10. Bell et al. (2004) analysed the 95 NMMAPS community
data to examine the association between ozone concentrations and mortality, showing that a
10-ppb increase in the previous week’s ozone was associated with a 0.5% (95% posterior
interval [PI], 0.3% - 0.8%) increase in daily mortality and a 0.64 %(95% PI, 0.31% -0.98%)
increase in cardiovascular and respiratory mortality. The effect estimates of the exposure over
the previous week were larger than those considering only a single day’s exposure (Bell at el.
2004). Recently, Dominici et al. (2006) examined the short-term association between fine
particulate air pollution and hospital admissions and found that exposure to PM2.5 was
associated with different health outcomes. The largest association was observed for heart
failure, viz., and a 10 µg/m³ increase in PM2.5 was found to be associated with a 1.3% (95%
PI: 0.8%, 1.8%) increase in hospital admissions from heart failure on the same day.
Although time-series studies have shown that day-to-day variations in air pollutant
concentrations are associated with daily deaths and hospital admissions, it is still unclear how
many days, weeks or months air pollution has brought such events forward. Harvesting or
mortality/morbidity displacement means that some cases are occurring only in those to whom
it would have happened in a few days anyway (Schwartz 2000a; Zanobetti et al. 2002). If so,
the increase in cases during and immediately after exposure would be offset by a deficit in
daily deaths a few days later (Schwartz 2001; Zanobetti et al. 2002; Zeger et al. 1999) (Figure
2.1). If air pollution has harvesting or long term effects, normal time-series models are unable
to estimate the effects due to the issues of collinearity and statistical power (Schwartz 2000a,
32
c). The distributed polynomial model (Schwartz 2000c) and the time-scale model (Zanobetti
et al. 2002) have been adopted to explore whether air pollution has harvesting or displacement
effects on daily deaths or hospital admissions. While recent studies have not found obvious
evidence of harvesting, they have found that estimated effects increase when longer lags of air
pollution are included (Schwartz, 2001; Zanobetti et al. 2002).
Figure 2.1 Harvesting phenomenon (Schwartz, Epidemiology, 2001; 12: 56-61)
2.4.2 Case-crossover Studies
Case-crossover study design is another way to estimate short-term health effects of air
pollution in epidemiological studies. In the last decade, the case-crossover design has been
applied in many studies of air pollution and health (Barnett et al. 2005; 2006; Levy et al.
2001b; Neas et al.1999; Schwartz & Lee 1999; Schwartz 2005). For example, Neas et al.
(1999) used a case-crossover study design to estimate the association between air pollution
and mortality in Philadelphia and found a 100 μg/m³ increment in the 48 hours mean level of
TSP was associated with increased all-cause mortality [odd ratio (OR) = 1.06 (95% CI: 1.03,
1.09)]. A similar association was observed for deaths in individuals over 65 years of age (OR:
hallaThis figure is not available online. Please consult the hardcopy thesis available from the QUT Library
33
1.07 (95% CI: 1.04, 1.11)). Levy et al. (2001) estimated the effect of short-term changes in
exposure to particulate matter on the rate of sudden cardiac arrest. The cases were obtained
from a previously conducted population-based case-control study and were combined with
ambient air monitoring data. The results