Content uploaded by Katharina Janke
Author content
All content in this area was uploaded by Katharina Janke on Apr 07, 2020
Content may be subject to copyright.
Do current levels of air pollution kill? The impact of air
pollution on population mortality in England
Katharina Jankea, Carol Propperb, and John Hendersonc
aCMPO, University of Bristol, UK
bUniversity of Bristol, UK; Imperial College, London, UK; CEPR, London, UK
cDepartment of Social Medicine, University of Bristol, UK
December 5, 2014
Abstract
The current air quality limit values for airborne pollutants in the UK are low by
historical standards and are at levels that are believed not to harm health. We assess
whether this view is correct. We examine the relationship between common sources of
airborne pollution and population mortality for England. We use data at local authority
level for 1998–2005 to examine whether current levels of airborne pollution, as measured
by annual mean concentrations of carbon monoxide, nitrogen dioxide, particulate matter
less than 1010 µm in diameter (PM10) and ozone, are associated with excess deaths.
We examine all-cause mortality and deaths from specific cardiovascular and respiratory
causes that are known to be exacerbated by air pollution. The panel nature of our data
allows us to control for any unobserved time-invariant associations at local authority
level between high levels of air pollution and poor population health and for common
time trends. We estimate multi-pollutant models to allow for the fact that three of
the pollutants are closely correlated. We find that higher levels of PM10 and ozone are
associated with higher mortality rates, and the effect sizes are considerably larger than
previously estimated from the primarily time series studies for England
KEY WORDS: air pollution, child health, asthma, avoidance behaviour, panel analysis
JEL I12, I18, Q53, Q58
1 Introduction
The current levels of airborne pollutants in many OECD countries are low by historical
standards. The limits on air pollution set by the regulatory authorities are also low by these
standards. Yet recent research from the USA has shown that there are adverse effects from
airborne pollution for infants even at these levels (Currie and Neidell 2005). This paper
focuses on the impact of airborne pollutants on population health in England. England is of
interest because it has average pollutant levels that are low (around 30% lower than those
examined in recent studies from the United States (Currie and Neidell 2005, Neidell 2004))
and the limit values allowed by the regulatory authorities are set reflecting a belief that there
is a safe threshold at which no significant health effects can be observed.1The aim of the
paper is to test this belief by examining whether the current levels of airborne pollutants
in England are associated with adverse health effects – as measured by mortality – for the
population.
Adults have been the main focus of most of the research on air pollution and excess
mortality. Previous studies of the impact of airborne pollutants on mortality rates are basically
of two kinds. The first exploit high frequency time series data on levels of air pollution and
number of deaths to examine the time series relationship. Such studies measure the acute
effects of air pollution and generally focus on a single pollutant. However, the focus on a
single pollutant may over-estimate its impact, as several of the common airborne pollutants
are correlated, because they are components of traffic emissions. In addition, if temporarily
elevated levels of pollution hasten the deaths of frail persons who would have died within days
or weeks, then the effects of pollution are over-estimated. The second type of study examines
the impact of living in cities with different levels of pollution. Whilst these studies capture
more than the short term effects of pollution, comparisons of cities suffer from potential
omitted variable bias, as it is likely that these cities are different in important ways other
than in their level of pollution. So observed cross-sectional differences in deaths may not be
causal (Chay and Greenstone 2003).
In this paper, we use the following design to deal with these problems. We take as the
unit of observation the primary unit of local government in the UK (the local authority) and
examine the relationship between annual mortality rates and annual mean concentrations of
four common air pollutants over time at this level. The use of a panel allows us to fully control
for time-varying determinants of death that are national in scope and factors that differ across
local areas that remain fixed over time, so we can isolate the impact of pollution from other
unobserved differences between local authorities. The use of a time period of a year means
this design will not detect the small changes in life expectancy (changes of a few days) that
may underlie the associations found in time series studies. Focussing on annual mortality
rates also reduces one aspect of model uncertainty found in time series studies (see Clyde
1Seehttp://uk-air.defra.gov.uk/.
2
(2000) and Koop and Tole (2004)). Additionally, annual mortality rates for local authorities
are readily available, whereas daily, weekly or monthly rates are not publicly available for
confidentiality reasons. Finally, the research design allows us to control for the correlation
between the levels of common airborne pollutants.
Despite its advantages this design has been little used to examine pollution and mortality.
In one of the few studies using this approach, Chay et al. (2003) examine the effect of par-
ticulate matter on adult mortality in the US during the 1970s. They find no impact of this
pollutant on adult mortality. However, the pollutant measure used during the period covered
by their study (total suspended particles) was possibly too imprecise to pick up mortality
effects.
Our panel begins in 1998 after Local Air Quality Management came into effect in the UK
in December 1997. It ends in 2005. Local Air Quality Management required local authorities
to assess the air quality in their areas and, as a result, local authorities installed additional
air pollution monitoring stations that supplement the existing national monitoring network.2
This provides a dense network of air pollution monitors that allows us, using spatial matching
methods, to assign air pollution measures for about 90% of local authorities and all of the local
authorities with large populations. Our analysis focuses on the pollutants carbon monoxide
(CO), nitrogen dioxide (NO2), particulate matter less than 10 µm in diameter (PM10), and
ozone (O3).3UK (and European) legislation sets limit values for these pollutants, because
they have deleterious effects on human health.4
We examine deaths from all causes and then focus on deaths from specific causes – dis-
eases of the cardiovascular and respiratory system – which have been shown by recent medical
literature to be associated with air pollution (Pope and Dockery 2006). We control for ob-
served factors that may be correlated with pollution but are independent causes of early
deaths, such as education, employment and lifestyle. We estimate multiple pollutant models
to isolate the impact of specific pollutants. We also subject our results to a large number of
specification tests, including ‘placebo’ tests for a spurious association between air pollution
at local authority level and death rates by examining the association of air pollution with two
causes of death which are unlikely to be driven by air pollution. Our findings suggest that the
relatively low levels of pollution currently permitted in the UK are associated with mortality
rates in the population. We find significant effects of both PM10 and O3on mortality. The
magnitudes of these effects are both statistically and economically significant.
2Note that the local authorities are not the same bodies that are responsible for providing health care or
meeting health targets.
3Earlier studies have examined sulphur dioxide pollution. We focus here on those pollutants whose levels
are widely measured, reflecting current concerns over their impacts on health.
4See Appendix Appendix A for sources and effects of of these pollutants and Appendix Appendix B for
the air quality standards in operation in England during our sample period.
3
2 An overview of the literature on air pollution and mortality
The literature on air pollution and mortality is dominated by two types of study: time series
studies of the association between short-term variations in air pollution and mortality and
cross-sectional studies of cohorts followed over time or of cities with long-term differences in
pollution. Time series studies regress daily counts of deaths for a geographical area onto daily
means of air pollutant concentrations, controlling for confounding factors such as temperature,
humidity and barometric pressure. Exploiting short-term variation to identify pollutant effects
eliminates the effects of lifestyle factors such as exercise and diet, because these factors do not
change on the short run. Systematic reviews of the numerous published time series studies
report significant associations between air pollutants and mortality, with mean estimates
suggesting that per 10 µg/m3increase in NO2, PM10 or O3or per 1 mg/m3increase in CO
mortality increases by less than 1% (see, inter alia, Stieb et al. (2002), Bell et al. (2005) and
Department of Health (2006)).
There are two problems interpreting the findings from time series studies. The daily time
series design can only identify the acute effect of pollution. Part of the increase in mortality
may be caused by deaths of individuals who would have died only a few days later from other
causes (an issue known as “harvesting”). So, such studies may over-estimate the impact of
air pollution on health. In addition, levels of different pollutants may be strongly correlated;
identifying which pollutant is causing the increased deaths is therefore difficult from studies
based on short-term fluctuations in one pollutant.
Ecological studies of associations between spatial variations in air pollution and spatial
variations in mortality compare mortality in highly polluted areas with mortality in less pol-
luted areas, using population average values to control for other risk factors such as smoking,
deprivation and education. Typically, they suggest that a pollutant increase of 10 µg/m3
increases mortality by about 3% (Wilson and Spengler 1996). But these studies face severe
omitted variables problems, as they typically do not control for many individual or community
level variables which may be correlated with pollution.5
Finally, cohort studies use pollutant concentrations averaged over a year or longer periods.
Few such studies exist and there are none for the UK. Two key U.S. studies estimate an increase
in mortality risk of between 4% and 14% per 10 µg/m3increase in PM2.5(Pope et al. 2002,
Dockery et al. 1993). Estimated effects on cardiopulmonary mortality are generally larger.
Estimates of the effects of CO, NO2and O3tend to be insignificant (Krewski et al. 2000). The
only long-term studies for Europe are one for Norway, which finds a mortality risk increase
of 8% per 10 µg/m3increase in nitrogen oxides (NO2+ NO) for men (Nafstad et al. 2004)
and one for the Netherlands, which finds positive but insignificant effect estimates for NO2
5A very small number of studies uses exogenous changes in air pollution. Clancy et al. (2002) used the
ban on coal sales in Dublin in 1990, which reduced average black smoke concentrations. Studies of extreme
pollution episodes use one large fluctuation in air pollutant concentrations to identify short-term effects. A
classic example is the Great Smog of London in 1954 that caused 4,000 excess deaths (Wilkins 1954).
4
(Hoek et al. 2002). Because of their design, cohort studies are expensive and take long time
to complete. In addition, cohort studies may suffer from omitted variable bias, as the cities
or zip codes which are compared may differ from each other in important ways other than
just their levels of air pollution.
Within the economics literature, there have been several studies for the US which show
that current levels of pollution are associated with poor health outcomes. Currie and Neidell
(2005) examine the impact of CO, PM10 and O3on infant deaths in California over the 1990s.
Using individual-level weekly data, they find a significant effect of CO on infant mortality.
Aggregating up their data to zip code-quarter level, however, they find no effect for CO, but a
significant effect for PM10. The pollution levels in California during the 1990s are higher than
the pollutant concentrations in England during the period we examine: the sample mean of
PM10 in Currie and Neidell (2005) is 39.4 µg/m3, whereas our sample mean is 24.7 µg/m3.
Chay and Greenstone (2003) exploit variation across US counties in the depth of a sudden
economic recession in 1980 to 1982 to identify the effect of a medium-term reduction in total
suspended particles (TSP - particles with diameter ≤40 µm) on infant mortality. Again,
pollution levels are higher than currently in England.6They find a significant effect of TSP
reductions on decreases in infant mortality rates.
In one section of their paper, Chay et al. (2003) use the same approach as we adopt here,
using US counties as the unit of observation. They exploit within-county time-series variation
in TSP levels to study the effect of air pollution on mortality in adults over 50 years and adults
aged 65 to 84 years in 1969 to 1974. The average pollution level in their data is twice the
level we examine.7However, they find no association between their measure of air pollution
and mortality, perhaps because TSP are a rather crude measure of air pollution.
3 Our empirical approach
Our unit of analysis is a local authority, which is the main unit of political administration
below the national level in the UK. There are 354 local authorities in England, with an
average population of around 140,000 people, ranging from just over 2,000 to just over 1
million.8Local authorities are aggregated into 9 Government Office regions. Figure 1 shows
the location and size of local authorities and the Government Office regions.
6TSP is not measured in England during our sample period. To compare pollution levels, we convert TSP
levels using 0.55 as PM10/TSP ratio. In a review of studies of the acute effects of particles, Dockery and
Pope (1994) use this ratio, based on guidelines from the US Environmental Protection Agency. Chay and
Greenstone (2003) report TSP levels between 56.4 and 71.1 µg/m3, which is equivalent to PM10 levels between
31 and 39 µg/m3, one and a half times our sample mean of 24.7 µg/m3.
7Using 0.55 as PM10/TSP ratio their TSP sample mean of 93 µg/m3is equivalent to a PM10 level of 51
µg/m3. Our PM10 sample mean is 24.7 µg/m3.
8The smallest local authority used in the analysis here contains 34,000 people (Rutland) and the largest
one million people (Birmingham).
5
Figure 1: English local authorities and Government Office regions
We estimate equations of the following form:
mj
it =α+P0
itγj+Z0
itβj+Tj+Tj
r+µj
i+εj
it (1)
where iindexes the local authority, tindexes the year, rthe region and jthe cause of death.
mj
it is the logarithm of one of six mortality rates (all-cause; all circulatory diseases; coronary
heart disease; acute myocardial infarction; stroke; bronchitis, emphysema and other chronic
obstructive pulmonary diseases), Pit is a vector of air pollutants (CO, NO2, PM10, O3), Zit is
a vector of time-varying controls at local authority (or regional) level. Tjis a time trend, Tj
r
is a region-specific time trend (regions are Government Office regions), µj
iis a local authority
fixed effect, and εit is the error term for cause of death j. The coefficients of interest are the
γj.
We first estimate the impact of each pollutant separately, but our main specifications
6
include all pollutants together to allow for correlation between them. Identification comes
from the time series variation in pollutant concentrations at local authority level. As our
panel is short, within-group estimates may be biased, so we also estimate OLS models (in
which the local authority fixed effect is replaced by a set of regional dummies) and three-year
long-difference models (Griliches and Hausman 1986). In all our analyses we estimate robust
standard errors and weight by the size of the local authority population.
4 Data
Data on air pollution comes from the UK Air Quality Archive,9supplemented with data
from four regional air quality networks managed by the same operator and from another four
regional networks managed by the Environmental Research Group at King’s College London.
These sources provide data on a total of 192 automatic monitoring stations, of which 90, 174,
111 and 105 record concentrations of CO, NO2, PM10 and O3, respectively. Figure 2 shows
the positions of these monitors.10 The figure also shows the population densities of local
authorities; the darker the shading, the more densely populated the area. It is clear from the
figure that monitors are located in more densely populated areas, so that, while there is not
equal coverage across areas, those areas with few monitoring stations are also areas of small
populations.
We convert measurements given in volume ratios into mass units and compute daily pollu-
tant concentrations if only hourly readings are available (see also Appendix C-1). We use the
daily mean of NO2and PM10 and the daily maximum 8-h running mean of CO and O3(the
choice of unit is determined by the relevant pollution standard) to calculate annual means.
We assign these annual pollutant concentrations to local authorities using a procedure similar
to Currie and Neidell (2005). Using the geographical coordinates of the headquarters of a lo-
cal authority, we calculate the distance between the headquarters and all monitoring stations.
Then we use all monitoring stations whose distance to the headquarters is less than 30 miles
(less than 10 miles for the London boroughs where there are many monitoring stations within
relatively small distances) to calculate a weighted mean of the annual pollutant concentrations
measured by these stations. The weight assigned to a monitor is the inverse of the distance
between the headquarters and the monitor. Our measure is thus the distance-weighted mean
of the annual mean pollutant concentrations at monitors in a 30 (10) mile radius of the head-
quarters of a local authority. We assign a measure of CO, NO2, PM10 and O3for at least
two years to 312 out of 354 local authorities. The local authorities with missing air pollution
measures are less populated areas.
To assess the accuracy of our pollution measure, we use our method to predict pollutant
9Prepared by AEA Energy & Environment on behalf of the Department for Environment, Food & Rural
Affairs, www.http://uk-air.defra.gov.uk/.
10The map does not show two monitoring stations in Wales close to the English border, which we use for
computing air pollution measures for local authorities in the West Midlands and in the North West.
7
Figure 2: Positions of monitoring stations in England
concentrations at monitor locations and compare the predicted with the actual pollutant
concentrations. For the underlying daily data the correlations are relatively high (0.59, 0.61,
0.75 and 0.84 for CO, NO2, PM10 and O3, respectively), indicating this approach will predict
pollution at a location relatively well. The correlation coefficients for the annual data across
all observations are lower at 0.44, 0.45, 0.40 and 0.50 for CO, NO2, PM10 and O3, respectively,
due to the averaging induced by moving from daily to annual measures. However, the time
series correlation between the predicted and actual annual values within monitoring stations
8
is higher – 0.72, 0.47, 0.53 and 0.73 – for CO, NO2, PM10 and O3respectively.11 Since our
identification strategy relies on time series variation within local authorities, the accuracy of
our pollution measure seems reasonable.
Using measurements taken by stationary monitors at outside locations to calculate expo-
sure to air pollution, there may be an issue of the extent to which measures of ambient air
pollution predict personal exposure, as most people spend over 80% of their time indoors.
Indoor air quality is often worse than outdoor air quality, because of cigarette smoke, paints,
vinyl flooring, gas stoves, dust mites etc. However, empirical studies have shown that ambient
levels of air pollutants and personal exposure to air pollutants are significantly correlated.12
Personal exposure is determined by outdoor concentrations, indoor concentrations and activ-
ity patterns. But as factors determining indoor concentrations, e.g. gas stoves and tobacco
smoke, are not measured in our data we are not able to control for these. So we make the
assumption that the major part of the variation in personal exposure to air pollutants is deter-
mined by changes in ambient levels of pollutants. We do, however, control for smoking rates
and allow for separate regional time trends, which will pick up, inter alia, regional differences
in changes in indoor pollution.13
Figure 3 presents quantile plots of our pollution measures, showing the time series variation
in the annual pollutant levels. CO clearly declines over the years of our sample. There is also
a reduction in the variation: the distance between the top two quantiles and the other three
quantiles of the distribution falls over time. Measured at an annual level, no local authority
exceeds the limit value, which is defined in terms of the daily maximum 8-h running mean.
The annual mean level of NO2initially declines before it peaks in 2003. The variation across
local authorities remains pretty constant across the sample period. NO2exceeds the limit
value of 40 µg/m3in many local authorities. Even in the year in which there were fewest
instances of exceedances (2002), average annual levels of NO2were higher than the limit
value in 17% of local authorities. Annual means of PM10 fall until 2000, remaining relatively
constant since then, apart from a peak in 2003. The distribution is pretty constant over the
period. PM10 does not exceed 40 µg/m3, which is the limit value in force towards the end of
our sample period, but it does exceed 20 µg/m3, the limit value which will come into effect at
11Figures are mean within station correlations. The median within station correlations are higher: 0.87,
0.56, 0.64 and 0.79.
12Georgoulis et al. (2002) use measurements of personal exposure to CO for 401 individuals in five European
cities during a 48 hour period and find that ambient levels of CO are a significant determinant of personal
exposure to CO. Kousa et al. (2001) use the same data and find that ambient levels of NO2explain 11 to 19% of
personal NO2exposure variation. However, cross-sectional correlation coefficients between personal exposure
and ambient pollutant concentrations can be misleading. For example, Janssen et al. (2000) study the time-
series correlation between ambient levels of PM2.5and personal exposure to PM2.5for elderly subjects with
cardiovascular disease in two European cities. They find that personal exposure and ambient concentrations
are highly correlated within subjects over time.
13O3has considerably lower indoor concentrations (Committee on the Medical Effects of Air Pollution 1997).
Thus, for people who spend little time outdoors, personal exposure to O3and ambient levels of O3are not
correlated. O3concentrations, however, are elevated in summer, and people tend to spend more time outdoors
in summer. Hence, our measure of O3should explain at least part of the variation in personal exposure to O3.
9
Figure 3: Quantile plots of annual pollutant concentrations in English local authorities.
Grey lines indicate annual limit values: the annual mean of NO2must not exceed 40 µg/m3
by 31 December 2005, the annual mean of PM10 must not exceed 40µg/m3by 31 Decmeber
2004 and 20 µg/m3by 31 December 2010
10
the end of 2010. In contrast to the three other pollutants, annual means of O3rise over the
sample period. The variance of the distribution is fairly constant. There are two clear peaks
in the series which affect all local authorities, one in 1999 and another one in 2003. Both
years had above average sunshine, illustrating the potential difficulty of isolating the impact
of O3from that of weather conditions.
Table 1: Descriptive statistics for estimation sample (n= 2,338, groups = 312)
Variable Mean SD Between Within Mean Mean
local author- local author- in 1998 in 2005
ities SD ities SD
Pollutants
CO (mg/m3) 0.80 0.34 0.26 0.23 1.13 0.55
NO2(µg/m3) 36.6 9.1 8.5 3.8 41.0 33.2
PM10 (µg/m3) 24.7 3.3 2.9 1.7 26.3 24.2
O3(µg/m3) 55.9 7.5 6.6 4.1 49.9 57.5
Mortality rates (per 100,000)
Mortality from all causes 660.1 86.2 78.2 39.9 712.7 605.0
Mortality from all circulatory
diseases
243.7 42.0 31.6 28.4 286.3 202.2
Mortality from coronary heart
diseases
124.9 28.6 22.1 18.5 153.0 99.7
Mortality from acute myocar-
dial infarction
54.8 17.1 13.1 11.3 71.3 41.1
Mortality from stroke 63.7 11.9 8.2 8.7 73.2 53.0
Mortality from bronchitis, em-
physema and other COPD
29.2 9.8 8.8 4.6 31.5 27.3
Control variables
Smoking rate (%) 25.7 2.4 1.7 1.7 27.4 23.6
Employment rate (%) 76.0 6.3 5.8 2.4 75.6 76.2
NVQ 4+ level rate (%) 24.4 7.9 7.6 3.2 22.2 26.6
Annual mean of summer daily
maximum temperature (◦C)
18.6 1.2 1.0 0.6 17.8 18.7
Annual mean of precipitation
(mm)
2.2 0.63 0.51 0.38 2.4 1.7
Other mortality rates for robustness tests (per 100,000)
Mortality from chronic liver
disease including cirrhosis
8.9 4.4 3.5 2.7 7.8 9.6
Mortality from infectious and
parasitic diseases
5.7 2.9 2.0 2.1 5.0 7.2
The top panel of Table 1 presents descriptive statistics for the pollution data. In addition
to the average fall in all pollutants other than O3, it shows that the values of the within-
local authority standard deviations range from 45% to 80% of the values of the between-local
11
authorities standard deviations. This provides support for identification of air pollution effects
by exploiting within-local authority variations. CO, NO2and PM10 are positively correlated,
with correlation coefficients between 0.4 and 0.6. They are negatively correlated with O3,
which tends to be higher in rural areas, with correlation coefficients between -0.2 and -0.5
(see Table C-2).
The second panel of Table 1 presents the mortality rates. Sources are given in Appendix
C-1. We examine deaths from all causes as well as deaths from specific causes for which
the medical literature suggests biologically plausible mechanisms that hypothetically link air
pollution and adverse effects on human health (see Pope and Dockery (2006) and Pope et al.
(2004)). Mortality from all circulatory diseases comprises the ICD-10 categories I00 to I99.
Mortality from coronary heart disease is a subset of mortality from all circulatory diseases
(ICD-10 categories I20 to I25). Mortality from acute myocardial infarction (heart attack), in
turn, is a subset of mortality from coronary heart disease (ICD-10 I21 to I22). Mortality from
stroke (ICD-10 I60-I69) is another subset of mortality from all circulatory diseases. Mortality
from bronchitis, emphysema and other chronic obstructive pulmonary diseases consist of the
categories J40 to J44, which are a subset of diseases of the respiratory system. The subset
J40 to J44 excludes asthma, pneumonia and – most important – influenza, thus avoiding
confounding of the pollutant effects by epidemics, which might coincide with increased air
pollution. We use directly age-standardised rates to control for different population age
structures across local authorities.
The time series plots of the standardised annual means of the six mortality rates in Figure
4 show a strong downward trend for the cardiovascular mortality rates. Many factors are
likely to cause this fall, including the National Service Framework for Coronary Heart Disease
(Department of Health 2000), a ten year plan initiated in 2000 with the aim of reducing
coronary heart disease in the community. On the other hand, respiratory mortality has only
a slight downward trend with peaks in 1999 and 2003. Consequently, the downward trend in
mortality from all causes, which encompasses both cardiovascular and respiratory mortality,
is less pronounced and levels off after 2001 before continuing in 2004.
The time-varying control variables in Zit in Equation 1 are the smoking rate, the em-
ployment rate, the percentage of working-age people who hold qualifications at degree level
and above, the annual mean of summer daily maximum temperature and the annual mean of
precipitation. Smoking is a strong predictor of premature mortality and an important source
of indoor pollution. It is therefore important to control for smoking rates. Smoking rates are
for 1998 and 2000 to 2005 for Government Office regions, which we match to the 354 local
authorities in England. We interpolate rates for 1999. Employment rates proxy economic
conditions, which may be correlated with health. In an analysis of US data, Ruhm (2000)
shows that mortality rates fall when the economy temporarily deteriorates (though Gerdtham
and Johannesson (2003) show that in Sweden unemployment increases the risk of dying). Ed-
ucation, in contrast, has a well established positive effect on health. We measure education
12
Figure 4: Standardised annual means of mortality rates. Standardisation by dividing annual
means of a mortality rate by its standard deviation. Mortality from AMI (acute myocardial
infarction) is a subset of mortality from coronary heart disease, which is a subset of
mortality from all circulatory diseases.
13
as the percentage of working-age people who hold qualifications at first degree level or higher.
The effects of air pollution could be confounded with weather conditions.14 To control
for these, we use surface observation data on daily maximum temperatures and daily rainfall
amounts, which we assign to the headquarters of the local authorities with the same procedure
we use for the pollutants. Firstly, we calculate for all weather stations the annual means of
precipitation and the annual means of the daily maximum temperature during the summer
months April to September. Then we determine the distance of all stations to the headquarters
of a local authority. Finally, we calculate weighted means of rainfall and temperature, using
the annual means of all stations within a 10 miles radius and a 20 miles radius, respectively.
The inverse of the distance between the headquarters and the weather station provides the
weight. These measures should capture the effects of heat waves (for example, the summer of
2003) and very wet years.
The third panel in Table 1 presents descriptive statistics for the controls. Mean smoking
rates fell from 27.4% in 1998 to 23.6% in 2005, possibly reflecting the government’s efforts to
reduce smoking prevalence (Department of Health 1998). Mean employment rates and mean
degree-level qualification rates increased between 1998 and 2005. Mean temperatures have
increased during the sample period, with peaks in 1999 and 2003. Precipitation seems to
have fallen, but the trend is less clear. As for the pollutants and the mortality rates, there is
significant within local authority variation.
5 Results
5.1 Cross-sectional associations
Figure 5 maps the cross-sectional spatial distributions of mean all-cause mortality and mean
local authority pollutant concentrations. Five different shades indicate the quintiles of the
respective distribution. The figure shows a similar spatial distribution for three of the pollu-
tants – CO, NO2and PM10 – which are higher in urban areas, while O3, is higher in rural
areas. There is no clear north-south divide in this rural-urban split of pollution. In contrast,
all-cause mortality shows a marked north-south split, death rates being higher in the north
and lower in the more affluent south. So in the raw data, averaged over the sample period,
there is little correspondence between the spatial distribution of mortality rates and of air
pollutant concentrations.
Table 2 examines this further by reducing the information on variation shown in the maps
to a split of the sample into terciles of the pollutant distributions and showing mean mortality
from all causes across these terciles. There is some indication that higher concentrations of
14For instance, during heat waves, O3levels rise because of the greater sunshine. Without controls for
temperature, it may appear that O3increases mortality, while in fact the heat caused excess deaths. On the
other hand, to the extent that weather is associated with the level of pollution but does not have an independent
effect on deaths, inclusion of weather variables will reduce the amount of variation in our pollution measures
and make it more difficult to detect their effects.
14
Figure 5: Cross-sectional distribution of mortality from all causes, CO, NO2, PM10 and O3
15
CO and NO2are associated with higher mortality rates. For example, the mean mortality
rate for observations in the highest third of the NO2distribution is 1.8% higher than the mean
rate for the lowest third. The relationship, however, is not linear, with the mean rate for the
middle third being greater than the mean rate for the highest third. In contrast, highest
concentrations of O3are associated with lower death rates. There is no clear relationship
between PM10 and mortality, with the mean mortality rate for the middle third being smallest
and the rates for the lowest and the highest third being similar.
Table 2: Means of pollutants and all-cause mortality by terciles of pollutant distributions for
the estimation sample (n= 2,338)
Ranked by Variable Lowest 1/3 Middle 1/3 Highest 1/3
CO CO (mg/m3) 0.5 0.7 1.2
Mortality from all causes (per 100,000) 656.9 662.5 661.0
NO2NO2(µg/m3) 27.5 35.1 47.3
Mortality from all causes (per 100,000) 649.4 670.1 661.0
PM10 PM10 (µg/m3) 21.2 24.5 28.4
Mortality from all causes (per 100,000) 662.4 654.1 663.9
O3O3(µ/m3) 48.1 55.2 64.4
Mortality from all causes (per 100,000) 686.6 656.8 634.7
5.2 The relationship between each pollutant and all-cause mortality
We start with an analysis of all-cause mortality to see if air pollution has any impact on this
aggregate measure. We then focus on the specific causes of deaths for which the medical
literature suggests they are causally related to air pollution (Pope et al. 2004). We begin by
examining the separate association between each pollutant and mortality. The first column of
Table 3 presents the raw correlations, estimated by an OLS regression of the log of all-cause
mortality on a constant and the pollutant. We then control for trend, region and region-
specific trends and present OLS, within-group and three-year long-difference estimates. We
then add the time-varying controls for weather and for lifestyle differences between local
authorities. We multiply the outcome variable by 100 and divide the NO2, PM10 and O3
levels by 10, so the coefficients are estimates of the percentage change in the mortality rate
per 10 µg/m3increase in NO2, PM10 or O3or per 1 mg/m3increase in CO.
The first block of Table 3 shows the estimates for CO. This shows no association between
CO and all-cause mortality, apart from a slightly significant positive coefficient in the OLS
equation with controls for time, region, weather and lifestyle, but this is not robust to the
inclusion of local authority fixed effects. The second block shows the results for NO2. The raw
association is positive but not significant. The coefficient estimates are significantly positive
after controlling for trend and region, but adding the controls for lifestyle and weather makes
16
Table 3: Estimates of the assocation between air pollutant concentrations and all-cause
mortality rates in single-pollutant models. Dependent variable: ln(all-cause mortality) ×100
OLS WG Long diff. OLS WG Long diff.
Trend, region, regional trend,
smoking rate, employment rate,
Controlling for trend, region degree level qualification rate,
OLS and regional trend temperature and precipitation
CO 0.39 1.14 −0.13 −0.06 3.40∗−0.42 −0.90
(1.58) (2.28) (0.61) (0.75) (1.74) (0.65) (0.77)
R20.00 0.49 0.94 0.02 0.71 0.95 0.10
NO2/10 0.43 1.58∗∗ 1.36∗∗∗ 1.40∗∗∗ 2.16∗∗∗ 0.42 0.21
(0.76) (0.75) (0.24) (0.24) (0.49) (0.27) (0.25)
R20.00 0.49 0.94 0.04 0.71 0.95 0.09
PM10/10 1.15 6.84∗∗∗ 4.07∗∗∗ 4.22∗∗∗ 3.96∗∗∗ 2.80∗∗∗ 2.38∗∗∗
(1.93) (1.50) (0.47) (0.51) (1.27) (0.51) (0.60)
R20.00 0.50 0.94 0.07 0.71 0.95 0.11
O3/10 −5.04∗∗∗ −0.86 1.92∗∗∗ 1.57∗∗∗ −1.47∗∗ 0.73∗∗ 0.12
(0.85) (0.85) (0.23) (0.26) (0.61) (0.29) (0.33)
R20.08 0.49 0.94 0.05 0.71 0.95 0.09
Obs. 2,338 2,338 2,338 1,404 2,338 2,338 1,404
Groups 312 312 312 301 312 312 301
WG = within groups. Coefficients are percentage changes in all-cause mortality rate per 1 mg/m3increase in CO
and per 10 µg/m3increase in NO2, PM10 and O3. Observations weighted by size of local authority population
(mid-year population estimates). Robust standard errors in brackets.*Significant at 10%, **significant at 5%,
***significant at 1%
17
the within-group and long-difference estimates insignificant. The results for PM10 in the third
block show a positive but insignificant raw association and significantly positive coefficients
for all other specifications. The within-group and long-difference estimates are similar. The
within-group estimate suggests that a 10 µg/m3increase in PM10 is associated with a 2.8%
increase in all-cause mortality.
The final block shows the results for O3. The raw correlation is negative and significant,
showing the association seen in Figure 5: rural areas, which have lower mortality rates, have
higher O3concentrations. Adding time varying controls does not change this negative sign,
though the point estimate is considerably smaller. Allowing for local authority fixed effects,
however, changes the direction of the association. Both the within-groups and the long-
difference estimates indicate a positive effect of O3on all-cause mortality. The within-groups
point estimate is a 0.7% increase in all-cause mortality for a 10 µg/m3increase in O3.
5.3 The relationship between all pollutants simultaneously and all-cause
mortality
Table 4 repeats the analyses of Table 3, but includes all pollutants simultaneously to allow
for correlation between the pollutant levels. It confirms that CO has no independent effect on
death rates. For NO2the within-group and long-difference estimates are again significantly
positive when controlling only for trend, region and regional trend, but become insignificant
when adding the controls for lifestyle and weather.
The coefficient on PM10 remains significantly positive in all specifications, though it falls
by 2 to 40%. The within-group estimate from the specification with all controls suggests
that per 10 µg/m3increase in PM10 the all-cause mortality rate increases by 2.7%. The
corresponding long-difference estimate suggests an impact of 2.4%, not significantly below
the within-group estimate. The associations between O3and mortality in the multi-pollutant
model are similar to those estimated by the single-pollutant model, though the negative
coefficient estimate in the OLS specification with all controls becomes insignificant. Using
the within-group specification with all controls, the estimated impact of O3is 0.8% per 10
µg/m3increase. The corresponding long-difference estimate is not significantly different from
zero. However, the long-difference sample is much smaller, so the effect of O3 might be
masked by the control for summer temperatures in this smaller sample. We therefore give
more credence to the within-group estimates.15
The association of the controls with all-cause mortality are shown in the final three columns
of Table 4. As expected, smoking rates are positively associated with higher death rates. The
estimate, however, is (marginally) significant only in the OLS specification. The employment
rate and the degree-level qualification rate are negatively associated with death rates, but
15Estimates using differences two periods apart support this argument, as they are closer to the within-group
estimates. The point estimates are, in fact, larger than the within-group estimates: NO20.59 (s.e. = 0.28),
PM10 2.50 (s.e. = 0.63), O30.81 (s.e. = 0.32), 1,701 observations, 304 groups.
18
Table 4: Estimates of the assocation between air pollutant concentrations and all-cause
mortality rates in multi-pollutant models. Dependent variable: ln(all-cause mortality) ×100
OLS WG Long diff. OLS WG Long diff.
Trend, region, regional trend,
smoking rate, employment rate,
Controlling for trend, region degree level qualification rate,
OLS and regional trend temperature and precipitation
CO 0.45 −2.27 −0.39 −0.79 0.27 −0.34 −0.94
(1.63) (2.30) (0.72) (0.80) (1.68) (0.69) (0.78)
NO2/10 −2.82∗∗∗ 0.49 0.64∗∗ 0.64∗∗∗ 1.50∗∗ 0.34 0.17
(0.94) (0.75) (0.27) (0.24) (0.62) (0.28) (0.26)
PM10/10 1.85 6.72∗∗∗ 3.03∗∗∗ 3.36∗∗∗ 2.33∗2.74∗∗∗ 2.37∗∗∗
(1.92) (1.55) (0.47) (0.52) (1.37) (0.51) (0.59)
O3/10 −6.60∗∗∗ −0.59 1.57∗∗∗ 1.06∗∗∗ −0.55 0.80∗∗∗ 0.18
(0.80) (0.80) (0.24) (0.27) (0.66) (0.29) (0.34)
Smoking 0.18∗0.07 0.20
rate (0.10) (0.08) (0.13)
Employm. −0.80∗∗∗ 0.00 0.05
rate (0.06) (0.04) (0.04)
Degree −0.46∗∗∗ −0.06∗−0.04
qual. rate (0.07) (0.03) (0.03)
Summer 0.96∗∗ 0.90∗∗∗ 0.94∗∗∗
temp. (0.39) (0.20) (0.23)
Precip- 1.30∗∗ 0.46 −0.12
itation (0.54) (0.35) (0.36)
R20.10 0.50 0.95 0.09 0.71 0.95 0.11
Obs. 2,338 2,338 2,338 1,404 2,338 2,338 1,404
Groups 312 312 312 301 312 312 301
WG = within groups. Coefficients are percentage changes in all-cause mortality rate per 1 mg/m3increase in CO
and per 10 µg/m3increase in NO2, PM10 and O3. Observations weighted by size of local authority population
(mid-year population estimates). Robust standard errors in brackets.*Significant at 10%, **significant at 5%,
***significant at 1%
19
the coefficients are significant only in the OLS specification, indicating that these variables
are capturing unobserved differences between local authorities rather than the effect of time
variation in employment and education on death rates.16
5.4 The relationship between pollutants and specific causes of mortality
The medical literature suggests that the association between air pollution and mortality is
driven by deaths from cardiovascular and respiratory causes (see, for example, Bell et al. (2005)
and Pope et al. (2002)). Several pathophysiological pathways that link particulate matter and
mortality from cardiovascular diseases have been suggested (see Pope and Dockery (2006) and
Department of Health (2006)). The two main hypotheses are the clotting hypothesis and the
neural hypothesis.17 From the first, we would expect to find positive associations between
PM10 and mortality from coronary heart disease in particular, but also stroke, heart failure
and atherosclerosis (Pope et al. 2004). Therefore, we examine mortality from all circulatory
diseases, coronary heart disease, acute myocardial infarction (heart attack) and stroke. Data
on mortality from heart failure and atherosclerosis on local authority level are not publicly
available. We are not able to examine the pathways suggested by the neural hypothesis.
Table 5 presents these estimates. The first column repeats the within-group estimates for
all-cause mortality from Table 4 for comparison. The results show that PM10 is positively
associated with all four cardiovascular mortality rates. We find a large and highly significant
positive effect on mortality from coronary heart disease (a subset of mortality from all circu-
latory diseases), for which we should find a strong effect according to the clotting hypothesis.
The estimates suggest that a 10 µg/m3increase in PM10 increases each of the four specific
mortality rates by around 4 to 5%. O3is positively associated with mortality from bronchitis,
emphysema and other chronic obstructive pulmonary diseases, suggesting that the association
between O3and mortality is driven by mortality from respiratory causes. The coefficient is
significant at the 10% level only, perhaps because the relatively small death rates (around 30
per 100,000 population) do not allow the effect to be estimated precisely enough.
Thus, we find that pollution levels are associated with those specific causes of death that
are indicated in the literature on the pathways by which pollution leads to death. Further,
16We also tested the robustness of our results to defining economic activity in terms of unemployment
instead of employment and to inclusion of an additional control for local pay rates (the log of the average
male pay). We found very similar results: no measures of economic conditions were significantly associated
with all-cause mortality in models which controlled for local authority fixed effects.
17The clotting hypothesis suggests that particles penetrating into the lungs cause an inflammatory response
in the lungs. The inflammation in turn might trigger changes in the control of blood clotting, causing,
for example, thrombosis. Alternatively, the inflammation might change chemical factors in the blood that
affect the stability of the atheromatous plaques in the arteries that supply blood to the heart muscle. The
atheromatous plaques might rupture, causing a blockage of the artery. The neural hypothesis proposes that
inhaled particles might trigger a reflex that leads to subtle changes in the heart rhythm, making the heart
more susceptible to dangerous changes in the rhythm that potentially cause sudden death. Therefore, we
would expect to find positive associations between particulate matter and mortality from dysrhythmias, heart
failure and cardiac arrest (Pope et al. 2004). Data on mortality from these causes on local authority level are
not publicly available, so we cannot examine this potential pathway.
20
Table 5: Within-group estimates of the associaton between air pollutant concentrations and
a range of mortality rates in a multi-pollutant model. Dependent variable: ln(mortality
rate) ×100
All cir- Coronary Acute Bronchitis,
culatory heart myocardial emphysema,
All causes diseases disease infarction Stroke other COPD
CO −0.34 −0.02 1.18 −1.51 −3.18 −3.82
(0.69) (1.53) (1.27) (2.36) (4.04) (3.51)
NO2/10 0.34 0.34 0.27 −1.91 0.82 1.91
(0.28) (0.49) (0.65) (1.27) (0.91) (1.27)
PM10/10 2.74∗∗∗ 4.38∗∗∗ 4.90∗∗∗ 5.03∗∗ 4.08∗∗ 1.80
(0.51) (0.78) (1.05) (2.09) (1.78) (2.49)
O3/10 0.80∗∗∗ −0.01 −0.39 −1.37 0.02 2.40∗
(0.29) (0.54) (0.69) (1.18) (0.90) (1.23)
R20.95 0.92 0.91 0.86 0.76 0.83
Obs. 2,338 2,338 2,338 2,338 2,338 2,338
Groups 312 312 312 312 312 312
COPD = chronic obstructive pulmonary diseases. Coefficients are percentage changes in all-cause mortality
rate per 1 mg/m3increase in CO and per 10 µg/m3increase in NO2, PM10 and O3. Controls are trend, region-
specific trends, smoking rate, employment rate, degree-level qualification rate, annual mean of daily maximum
temperature in summer and annual mean of precipitation. Observations are weighted by the size of the local
authority population (mid-year population estimates). Robust standard errors in brackets.*Significant at 10%,
**significant at 5%, ***significant at 1%
our estimates suggest that the effects of pollution on these specific causes of death account for
a high fraction of the estimated effect of pollution on all-cause mortality. Using the sample
mean for each specific mortality rate from Table 1 and applying our estimates from Table 5,
the overall estimated effect of PM10 on coronary heart disease and stroke accounts for 80%
of our estimated effect of PM10 on all circulatory diseases, while the effect on mortality from
circulatory disease accounts for 60% of our estimated effect on all deaths.
5.5 The relationship between all pollutants and all-cause mortality for
different age groups
The literature suggests that children and elderly persons are most likely to be susceptible to air
pollution (Pope and Dockery 2006). So if our results indicate some causal link, we should find
greater effects for these age groups. Table 6 presents within-group estimates of the association
between air pollutants and all-cause mortality by broad age groups: under 15 years, between
15 to 64 years, 65 to 74 years and older than 75 years. As directly age-standardised rates are
not publicly available for the older than 75 years group, we use non-age-standardised data
and control for population age structure by including controls for proportions of age groups
in 5-year age bands on the right-hand side.18
18This specification is on a per capita basis, so also acts as a robustness check of our use of age-standardised
mortality rates; see also Row 12 of Table 7.
21
Table 6: Within-group estimates of the association between air pollutant concentrations and
all-cause mortality for different age groups in a multi-pollutant model. Dependent variable:
ln(mortality rate) ×100
All ages <15 15-64 65-74 >75
CO −1.02∗5.08 2.65∗∗ −1.05 −1.60∗
(0.56) (5.57) (1.18) (1.13) (0.85)
NO2/10 0.72∗∗ −1.77 −0.25 0.88 0.82∗∗
(0.29) (2.68) (0.61) (0.58) (0.37)
PM10/10 2.46∗∗∗ 9.30∗∗ 2.27∗∗ 1.63 3.14∗∗∗
(0.53) (4.57) (1.05) (1.00) (0.64)
O3/10 0.69∗∗ 2.41 −0.12 0.51 0.90∗∗∗
(0.29) (2.34) (0.60) (0.54) (0.32)
R20.97 0.52 0.88 0.91 0.80
Observations 2,338 2,331 2,338 2,338 2,338
Groups 312 312 312 312 312
Coefficients are percentage changes in all-cause mortality rate per 1 mg/m3increase in CO and per
10 µg/m3increase in NO2, PM10 and O3. Mortality rates are not age-standardised. Controls are
proportions of age groups (5 year age bands), trend, region-specific trends, smoking rate, employ-
ment rate, degree-level qualification rate, annual mean of daily maximum temperature in summer
and annual mean of precipitation. Observations weighted by size of the local authority population
for respective age group. Robust standard errors in brackets.*Significant at 10%, **significant at
5%, ***significant at 1%
For comparison, the first column of Table 6 presents estimates for all ages using the
all-cause mortality rate that is not age-standardised. The coefficients are similar to those
obtained using age-standardised rates. Columns 2 to 5 of Table 6 show that the effects of
PM10 and O3are largest for the most vulnerable groups. The PM10 estimates are largest for
the youngest age group. The absolute impact is smaller, because of the very low death rates
in this age group. At the mean of the under 15 years old mortality rate, 44 per 100,000, a 10
µ/m3increase in PM10 increases the number of deaths by 4, whereas the coefficient estimate
for the over 75 years old suggests that at the mean mortality rate of 10,556 per 100,000 a
10 µg/m3increase in PM10 increases the number of deaths by 331. The coefficient on O3is
significant only for the over 75 years old, suggesting that the coefficient estimate for O3in the
all-ages specification is driven by this age group. Disaggregation by age also shows an effect
of NO2(again for the elderly) and for CO (in this case for those aged 15 to 64).
6 Robustness checks
Our method involves assignation of air pollution levels to local authorities and the estimation
of a linear relationship between pollution and death rates. We subject these assumptions to
robustness tests. We further explore whether our results are indicative of a causal relationship
by first undertaking ‘placebo tests’ and second by examining whether confounding factors
could account for the association we find between pollution and mortality. The results of
our robustness tests are summarized in Table 7. The baseline estimates in row 1 are the
22
within-group estimates from the specification with all four pollutants simultaneously and the
full set of controls in the last block of Table 4.
6.1 The assignation of air pollution to areas
Our air pollution measure is the distance-weighted mean of the annual mean pollutant con-
centrations at monitors within a 30 mile radius (10 miles for London) of the headquarters of
a local authority. Row 2 of Table 7 presents estimates using a 20 mile radius (5 miles for
London). The number of local authorities to which we can assign an air pollution measure
drops from 312 to 267. The coefficients on PM10 and O3fall by 20% and 5%, respectively,
but they are still significantly positive.
To calculate our air pollution measures we used monitoring stations that are situated in
different environments, for example in urban areas, at roadsides or in rural areas. If a local
authority has mainly roadside or kerbside monitoring stations, actual exposure might be lower
than our measures suggest. Row 3 of Table 7 examines robustness of our results when we
omit readings from kerbside and roadside stations. The number of local authorities to which
we can assign a pollution measure drops from 312 to 243. Though the coefficients on PM10
and O3drop by 20% and 5%, respectively, they are still significantly positive.
The summer of 2003 was unusually hot. This was also a year with higher death rates
and higher O3and PM10 levels. Row 4 examines the robustness of our results to omission of
this year. The estimated impact of both PM10 and O3falls by around a third, as might be
expected given this year is an outlier, but PM10 remains well defined.19 More generally, to
test that our results are not driven by areas with high levels of pollution which may not be
representative of England, we omit observations with one or more pollutants in the top 10%
of the pollutant distribution. Row 5 shows that the results are robust to this omission.
Our assignment of pollution measures to local authorities is based on distance to moni-
toring stations, without taking into account wind direction, which is predominantly from the
west in England. Local authorities located in the South West, in particular, will have mea-
sures predominantly based on stations to their east. To examine whether this is a problem,
Row 6 omits observations in the South West. Our results are little affected by omitting this
region.
Air pollution – at least from CO, NO2and PM10 – might be an urban phenomenon.
We therefore checked that our results were not solely due to London by omitting all London
observations. Row 7 shows that the estimates for PM10 and O3fall by around 20% but remain
19If we allow for a full set of year dummies the coefficient on PM10 falls to 1.12 (s.e. = 0.54) and the coefficient
on O3falls to -0.24 (s.e. = 0.32). However, both weather coefficients have incorrect signs (the coefficient on hot
weather is significantly negative and the coefficient on precipitation is positive and significant) and several of
the year dummies are not significantly different from each other. We conclude that we cannot identify separate
year, pollution and weather effects. A more parsimonious time specification that fits the time pattern in death
rates (a spline with knots in 1999 and 2003, both of which are years with higher death rates and higher
temperatures) gives significant positive coefficients for both pollutants and summer temperature (PM10 2.18
(s.e. = 0.55), O30.80 (s.e. = 0.29), summer temperature 0.66 (s.e. = 0.20)).
23
Table 7: Robustness tests in a multi-pollutant fixed effects model for all-cause mortality
Coefficient
on additonal
CO NO2/10 PM10/10 O3/10 Obs. Groups
1 Baseline −0.34 0.34 2.74∗∗∗ 0.80∗∗∗ 2,338 312
(0.69) (0.28) (0.51) (0.29)
2 Monitoring stations within −0.36 0.16 2.20∗∗∗ 0.76∗∗ 1,933 267
20 mile/5 mile radius (0.62) (0.31) (0.50) (0.31)
3 Drop kerbside and roadside −2.37∗∗ 0.50 2.16∗∗∗ 0.77∗∗ 1,778 243
monitoring stations (1.19) (0.39) (0.53) (0.34)
4 Drop observations for 2003 −0.29 0.28 1.88∗∗∗ 0.51 2,037 312
(0.73) (0.30) (0.53) (0.33)
5 Drop observations in top 10% −1.02 0.41 2.26∗∗∗ 0.63∗1,789 288
of pollutant distributions (1.01) (0.38) (0.75) (0.35)
6 Drop observations in South West −0.73 0.28 2.58∗∗∗ 0.93∗∗∗ 2,132 283
(0.82) (0.34) (0.52) (0.33)
7 Drop observations in London −0.47 0.65∗∗ 2.22∗∗∗ 0.63∗∗ 2,081 279
(0.83) (0.31) (0.51) (0.29)
8 Include lagged pollutants −0.37 0.67∗∗ 1.90∗∗∗ 0.72∗∗ 2,043 312
(0.98) (0.30) (0.68) (0.31)
9 Include annual maximum of 0.09 1.01∗∗∗ 1.77∗∗∗ 0.70∗∗ 2,338 312
weekly pollutant levels (0.77) (0.33) (0.61) (0.34)
Coefficient on annual maximum 0.03 −0.52∗∗∗ 0.36∗∗ −0.02
weekly pollutant levels (0.14) (0.2) (0.15) (0.06)
10 Dependent variable: ln(mortality −4.42 −0.36 0.10 4.63∗2,331 332
from chronic liver disease) ×100 (5.87) (2.73) (5.09) (2.80)
11 Dependent variable: ln(mortality 5.20 −2.99 −2.41 1.73 2,325 312
from infectious diseases) ×100 (6.71) (3.24) (6.60) (3.15)
12 Include population size/1000 −0.30 0.38 2.96∗∗∗ 0.83∗∗∗ −0.12∗∗ 2,338 312
as additional control (0.69) (0.29) (0.51) (0.29) (0.06)
Coefficients are percentage changes in all-cause mortality rate per 1 mg/m3increase in CO and per 10 µg/m3increase in NO2, PM10 and O3. Observations
weighted by size of the local authority population (mid-year population estimates). Baseline specification includes time trend, region-specific trends, smoking
rate, employment rate, degree-level qualification rate, annual mean of daily maximum temperature in summer and annual mean of precipitation. Robust
standard errors in brackets.*Significant at 10%, **significant at 5%, ***significant at 1%
24
significantly positive. The estimate for NO2increases by 30% and becomes significantly
positive.20
6.2 Dynamics and non-linearities
We were concerned that we might have mis-specified the dynamic structure of the model.
Row 8 therefore includes the lagged levels as well as the current levels of the pollutants. The
estimated effects of current PM10 and O3change slightly but remain statistically significant.21
We also conditioned on lagged mortality. Again, our results were robust to this, suggesting
that the local authority fixed effects do a good job of picking up unobserved heterogeneity
between local authorities.
Our model assumes that the impact of air pollution on mortality is linear. We investigate
non-linearities using splines in the levels of PM10 and O3in a within-group specification
controlling for CO, NO2, trend, region-specific trends and our full set of covariates. We place
two knots at the 33rd and 66th percentiles, dividing the pollutant data into terciles. Table 8
presents the results. For PM10 the coefficients for the middle tercile and the highest tercile
are larger than the coefficient for the lowest tercile, though the relationship is not linear, with
the largest estimate for the middle tertile. For O3there seems to be a negative relationship,
with the coefficient for the lowest tercile being larger than the coefficient for the middle tercile
and the coefficient for the middle tercile being larger then coefficient for the highest tercile.
The estimates for the middle and the highest tercile, however, are not significantly different
from zero.
We also tested whether the annual maxima of the weekly means of the pollutant concen-
trations have an impact on mortality to determine whether the long-term average level of
air pollution or short-term peaks drive the relationship between air pollution and mortality.
Row 9 in Table 7 presents the results. The first line shows the coefficients for the annual
mean pollutant concentrations, the second line the coefficients for the annual maxima of the
weekly mean pollutant concentrations. The coefficient for the annual mean level of PM10
drops by around one third, but is still significantly positive. The level of PM10 in the week
with the highest PM10 level is positively associated with the annual mortality rate, though
the size of the effect is only one fifth of the effect of the annual mean level of PM10. The
20We also checked if there are regional differences in the impact of PM10 on mortality. The region pattern is
quite complex, but there is a group of regions where the impact of PM10 is largest, both for all-cause mortality
and the specific mortality rates. These regions are South West, East Midlands and North East.
21The coefficients on the lagged pollutants are small and insignificant for three of the four pollutants. For
O3, however, the coefficient on the lag is similar and of opposite sign to that of current O3. This result might
indicate that the impact of O3is to bring mortality that would have otherwise occurred forward (harvesting).
Conditional on a positive association with the current level of pollution, a negative coefficient on the lagged
level could indicate harvesting, since individuals who died last year are not available to die this year. However,
the issue of harvesting has less force for annual data as – by definition – the mortality rates and the measures
of pollution average out short run increase and decreases. In our data, years with higher than average O3are
preceded by years with lower than average O3: it seems likely that in this short time series this is what the
lagged coefficient is picking up.
25
Table 8: Within-group estimates of the association between air pollutant concentrations and
all-cause mortality using a spline with knots at the 33rd and 66th percentile
Pollutant Lowest tercile Middle tercile Highest tercile
CO 2.24∗∗ 3.48∗∗∗ 2.69∗∗∗
(1.05) (1.13) (0.69)
NO20.97∗∗ 0.88 0.62
(0.49) (0.54) (0.45)
Coefficients are percentage changes in all-cause mortality rate per 1 mg/m3increase in CO and
per 10 µg/m3increase in NO2, PM10 and O3. Controls are CO, NO2, time trend, region-specific
trends, smoking rate, employment rate, degree-level qualification rate, annual mean of daily maxi-
mum temperature in summer and annual mean of precipitation. Observations weighted by size of
the local authority population (mid-year population estimates). Robust standard errors in brack-
ets.*Significant at 10%, **significant at 5%, ***significant at 1%
coefficient for the annual mean level of O3drops by less than 10%, and the coefficient for the
maximum weekly level of O3is not significantly different from zero. The coefficient for annual
mean NO2becomes significantly positive, but the coefficient for maximum weekly NO2has
an unexpected negative sign and largely offsets the effect of annual mean NO2, leaving the
joint effect similar to the baseline estimate.22 These tests show that there is some evidence of
non-linear effects, but they do not change our main finding that PM10 and O3are positively
associated with mortality.
6.3 Placebo tests
It is possible that the association of mortality with pollution does not result from pollution
effects, but that our pollution measures are proxies for some omitted factor which is correlated
with pollution but itself is the cause of deaths. To some extent, this is already dealt with
by using local authority fixed effects and region-specific time trends. Any non-time-varying
factors – such as poor health care services or the presence of health risks in urban settings –
will be controlled for by the fixed effects, and the region-specific trends will pick up changes
over time at regional level. However, it is possible that there are omitted time-varying factors
at local authority level that are correlated with changes in pollution and that are driving our
results.
One way of testing for this is to examine mortality from causes that are unlikely to be
affected by the within local authority time series variation in pollution. If the coefficients for
the air pollutants are similar to those found in the baseline specification, then this suggests
that some omitted factor may be driving the association we find between air pollution and
mortality rates. Two candidate causes are chronic liver disease (including cirrhosis) and
infectious and parasitic diseases. Rows 10 and 11 report the coefficients on the pollutants
22The somewhat odd results for NO2may be due to collinearity. In fact, annual mean NO2and maximum
weekly NO2are strongly correlated, with r = 0.86. The correlation coefficients for the other pollutants are
smaller: annual mean CO and maximum weekly CO: r = 0.65, annual mean PM10 and maximum weekly
PM10: r = 0.65, annual mean O3and maximum weekly O3: r = 0.57.
26
for the baseline specification with age-standardised mortality rates from liver disease and
infectious and parasitic diseases as the dependent variable. The baseline specification (as
all others in the table) includes the full set of controls to allow for the fact that mortality
from liver disease and from infectious and parasitic diseases may be associated with the
economic cycle and weather. The results show none of the coefficients on the pollutants are
statistically significant, apart from a marginally significant coefficient on O3for mortality
from liver disease.
6.4 Mitigating response to pollution: population mobility
Our estimates are weighted by the size of the local authority population, giving more impor-
tance to local authorities with bigger populations and consequently more reliable mortality
measures.23 The population size, however, might have an independent impact on mortality
other than affecting the precision of the mortality rate. For example, a population could
shrink because healthy people leave. Consequently, the proportion of frail people would in-
crease, causing an increase in mortality. If healthy people leave because of upward-trended
air pollution, the increase in mortality might wrongly be assigned to the rise in air pollution
rather than the fall in population. To test this, row 12 in Table 7 controls for the population
size. The coefficients on PM10 and O3are unaffected. The coefficient on population size is
significantly negative. Assuming that changes in the population size are mainly caused by
migration, this result supports the idea that healthy people are more mobile, leaving a more
frail population behind. These moves, however, do not appear to be a response to pollution
levels.
6.5 Magnitudes
Our results are statistically significant, but are they economically significant? The within-
group estimates from the penultimate column of Table 4 can be used to examine the effect of
a change in PM10 and O3on mortality. We focus on all-cause mortality.
Assuming no behavioural response (an issue we return to below) a 10 µg/m3increase in
PM10, holding all other pollutants fixed, is associated with a 2.7% increase in the all-cause
mortality rate. As the mean all-cause mortality rate is 660 per 100,000 population, this
increase equals around 18 more deaths per 100,000 persons. The 10th percentile of the PM10
distribution is 20.9 µg/m3, the 90th percentile is 29.0 µg/m3, and so a move from the 10th
percentile to the 90th percentile of the PM10 distribution would be associated with around 14
more deaths per 100,000 population. A 10 µg/m3increase in O3, holding all other pollutants
fixed, is associated with a 0.8% increase in the all-cause mortality rate. The 10th percentile
23The results are robust to not weighting. Estimates for the within-group multi-pollutant model with all
controls from Table 4 are: CO -0.37 (s.e. = 0.71), NO20.41 (s.e. = 0.29), PM10 3.02 (s.e. = 0.56), O30.55
(s.e. = 0.29).
27
of the O3distribution is 47.1 µg/m3, the 90th percentile is 66.4 µg/m3, so a move from the
10th to the 90th percentile would be associated with 10 more deaths per 100,000 population.
Alternatively, the difference between the 90th percentile and the 10th percentile of all-cause
mortality is 225 deaths per 100,000 population. So, a fall from the 90th to the 10th percentile
of PM10 would account for about 6% of the spread in all-cause mortality, while moving from
the 90th to the 10th percentile of the O3distribution would account for around 4% of the
spread in all-cause mortality.
These effects can be compared to those from the cohort and time series studies. We
would expect our estimates to lie between those of the cohort studies, which measure the
impact of air pollution over a long period (and cannot control for unobserved heterogeneity
across individuals), and the time series estimates, which measure the immediate response to
a change in air pollution. The American Cancer Society Cohort Study estimates that a 10
µg/m3increase in fine particles, PM2.5, would lead to a 6% increase in all-cause mortality
(Pope et al. 2002). The health effects from fine particles are worse than the effects from
coarser particles, which the PM10 measure includes but the PM2.5measure excludes. Thus,
we would expect our estimate to be lower. A meta-analysis of the time series studies (Stieb
et al. 2002) reports that multi-pollutant models estimate a 0.4% increase in mortality per 10
µg/m3increase in PM10. Our estimates indicate a 10 µg/m3increase in PM10 is associated
with a 2.7% increase in mortality. So our estimate is about half the size of that from the
cohort study – which has no UK counterpart – and nearly seven times as large as those from
times series studies that have been undertaken for the UK.24
There is no robust estimate of the effect of O3from the American Cancer Society Cohort
Study. Time series studies estimate a 0.3% death rate increase per 10 µg/m3increase in O3
in single-pollutant models and a 0.1% increase in multi-pollutant models (Stieb et al. 2002).
In our analysis, a 10 µg/m3increase in O3is associated with a 0.8% increase in mortality.
Again, our estimate is considerably higher than those from UK studies undertaken to date.
The extent to which we can use our estimates to quantify the effects of a change in pollution
depends on whether individuals are likely to take actions to protect themselves from increases
in pollution levels. Neidell (2004) finds that people in California respond to information about
air pollution (smog alerts) with avoidance behaviour. In England air pollution alerts have
to be issued when NO2levels exceed 400 µg/m3or when O3levels exceed 360 µg/m3(240
µg/m3since September 2003). Since these thresholds came into force in 2001 no alert has
been issued. And while air pollution forecasts are freely available via a variety of sources,25
anecdotal evidence shows that use of this information is limited. For example, in 2006 the
Sussex Air Quality Partnership piloted a service for respiratory sensitive people that sends
air quality forecasts to mobile phones. The study found that the service raised awareness of
24Our estimate of the impact of PM10 over a year is similar to the impact of a PM10 reduction caused by a
13-month strike at a steel mill in Utah (Pope 1996).
25Teletext, the World Wide Web, a Freephone telephone service and weather forecasts in newspapers, on
TV and radio.
28
pollution episodes and produced health behaviour modifications (Smallbone 2009). However,
the same information had been freely available before the service was introduced. Individuals
appeared to respond to air quality forecasts only when they received them as personalised
messages.
Assuming the extreme position of no behavioural response, our estimates can be used to
give a back-of-the-envelope calculation of the benefits of the recent UK policy to reduce the
limit value for PM10 to 20.0 µg/m3by 2010. We estimate a 10 µg/m3increase in PM10,
holding all other pollutants fixed, is associated with a 2.7% increase in all-cause mortality.
Therefore, reducing PM10 pollution from our sample mean of 24.7 µg/m3to 20.0 µg/m3(a
fall of just under 20%) would be associated with 8.4 fewer deaths per 100,000 population.
The population of England is just over 50 million, so this translates into around 4,200 fewer
deaths per annum over the whole population of England. Putting a monetary value on these
lives saved is less straightforward, because we do not know the life expectancy of those who
die prematurely. A value per year of life can be taken from the implicit figure used by the UK
body responsible for authorisation of the use of new drugs and therapies in the NHS, which
is around £30,000 (Devlin and Parkin 2004). If we assumed that those who died had another
10 years to live and were healthy, the value of the 42,000 life years gained is around £1,260
million.26 If those who died were less healthy, then our estimate is too high. But as we do
not take into account any of the non-mortality costs associated with air pollution, this figure
is more likely to be a lower bound.27
7 Conclusions
We identify the impact of airborne pollutants on mortality from time series variation in annual
average pollution levels in English local authorities. Our research is the first to use this design
for the UK and one of the few economic studies outside the USA. Our results suggest that
currently permitted levels of PM10 and O3are associated with population mortality, that these
pollutants are associated with higher death rates amongst those groups that are likely to be
affected by pollution, and that individuals die from those causes that the medical literature
indicates are most likely to be associated with pollution. In addition, we find no association
26This benefit figure is one and a half times the size of the £791 million expenditure on protec-
tion of ambient air and climate by the UK general government sector (£250 million) and UK industry
(around £541 million) in 2004 (http://www.statistics.gov.uk/downloads/theme_environment/EA_Jun08.pdf
and http://www.defra.gov.uk/environment/statistics/envsurvey/expn2004/eerp2004.pdf). It is in a similar
ballpark to estimates of the annualised cost of fitting all new cars and lorries with devices that reduce emis-
sions (Department for Environment, Food and Rural Affairs 2007), though it is estimated that this action will
decrease PM10 by only 0.8 and not the 4.7 needed to reach the new standard.
27If the short run effect of pollution is to kill the frail, our estimates are an upper bound. We repeat this
exercise for deaths in the age group 15 to 64 years old, as these individuals are least frail. We estimate that
a 10 µg/m3increase in PM10 increases mortality in this age group by 2.3%, so a 4.7 µg/m3drop in PM10
evaluated at the mean mortality rate for this age group, 247 per 100,000 population, would result in 2.7 fewer
deaths per 100,000 population. The population of 15 to 64 years old is around 32 million, so the drop in
mortality translates into 864 fewer deaths per annum. Assuming these individuals gain only 10 years of life –
so this estimate will give a lower bound – these 8,640 additional life years are worth £260 million.
29
between pollution and causes of death that are not affected by pollutants. This suggests
that although we cannot exploit a natural experiment and have to rely on annual time series
variation at the local authority level, we do identify a causal relationship.
The relationship that we find between pollution and mortality is for average levels of
pollution that are lower than studied in most of the existing research. We find an effect of
both PM10 and O3, with the effect being largest and most robust for PM10 . Recent economic
studies of infants and children in California tend to find lesser effects from PM10 and O3and
more from CO. On the other hand, the adult-focussed epidemiological literature finds short-
term associations between mortality and CO, NO2, PM10 and O3, but long-term associations
with PM10 only. Our yearly focussed approach has findings that accord with this broader
epidemiological literature. Finally, our estimates of the deaths arising from current levels of
airborne pollution are considerably higher than those which have been estimated previously
using UK data. They are, in fact, closer to those derived from the much less common – and
far more expensive – cohort studies. As none of these have been undertaken in the UK, our
results suggest that exploitation of the time series variation in annual data at small area level
may be used to provide evidence on the longer term impact of airborne pollution.
30
Appendix A Sources of CO, NO2, PM10 and O3and their effects
on human health
CO is a colourless, odourless, poisonous gas, which reduces the body’s ability to use oxygen.
CO results from combustion processes under insufficient oxygen supply. Burning fuel con-
taining carbon in idling or slow moving motor vehicles contributes the largest share of CO.
A smaller share results from processes involving combustion of organic matter, e.g. power
stations and waste incinerators. CO survives in the atmosphere for approximately one month
before it oxides to carbon dioxide.
NO2is a brown, reactive gas with a detectable smell, which is highly toxic in significant
concentrations. Relatively high concentrations of NO2cause inflammation of the airways and
can produce broncho-constriction in both asthmatics and non-asthmatics (Committee on the
Medical Effects of Air Pollution 1997). NO2occurs as a primary pollutant (emitted directly
from a source) and as a secondary pollutant (formed in the air by reactions of primary pol-
lutants). As a primary pollutant, NO2is mainly emitted from the tailpipe of diesel vehicles,
especially when they move slowly. As a secondary pollutant, NO2is mainly formed by oxida-
tion of nitric oxide, which is produced by burning fuel at high temperatures. Road transport
produces the largest share of NO2. Other important sources of NO2are power stations and
natural gas space heating (Air Quality Expert Group 2004). NO2converts to nitrates (e.g.
nitric acid), which rain or gravity return from the atmosphere to Earth.
Particulate matter has an unspecified chemical composition. Its most important charac-
teristic is the size of the particles. Coarse particles with a diameter of 2.5 to 100 µm consist
mainly of soil and sea salt elements and are produced by mechanical processes (e.g. suspension
of soil in farming and mining, construction, stone abrasion, and sea spray). Coarse particles
settle out quickly by gravity. Fine particles with a diameter of 0.1 to 2.5 µm consist of primary
particles that result from combustion processes and secondary particles that are, for instance,
formed by condensation of low volatile compounds and ammonia. Fine particles are too small
to settle out by gravity and too large to coagulate into larger particles, therefore they can
stay in the atmosphere over days to weeks and travel hundreds to thousands of kilometres
before rain returns them from the atmosphere to Earth. Ultra-fine particles with a diameter
of 0.01 to 0.1 µm have a short residence time in the atmosphere because of their Brownian
motion. Particles with a diameter less than 10 µm (PM10) are inhalable, but 60 to 80% of
particles with a diameter of 5 to 10 µm are trapped in the nose and pharynx (Wilson and
Spengler 1996). Smaller particles penetrate the trachea and the primary bronchi. Very small
particles penetrate deep into the lungs.
O3is a bluish, unstable gas with a pungent odour, which is toxic even at low concen-
trations. It is the “most potent (. . .) pro-inflammatory pollutant of the common range of
air pollutants” (Committee on the Medical Effects of Air Pollution 1997). O3is a secondary
pollutant that is formed by the action of sunlight on volatile organic compounds in presence
31
of NO2. It can travel large distances. Nitric oxide, which has high concentrations in urban
areas, scavenges O3, resulting in much higher O3levels in rural areas than in urban areas. As
the formation of O3requires sunlight, O3levels are highest in summer.
32
Appendix B Current air quality standards
Appendix B.1 Annual
The annual mean of NO2must not exceed 40 µg/m3by 31 December 2005. The annual mean
of PM10 must not exceed 40 µg/m3by 31 December 2004 and 20 µg/m3by 31 December
2010.
Appendix B.2 Daily
The daily maximum of the running 8 hour mean of CO must not exceed 10 mg/m3by 31
December 2003. The 24-h mean of PM10 must not exceed 50 µg/m3more than 35 times per
year by 31 December 2004. The daily maximum of the running 8-h mean of O3must not
exceed 100 µg/m3more than 10 times per year by 31 December 2005.
Appendix C Data sources
Table C-1 provides the data sources for all variables except air pollution.
Air pollution data were downloaded from the web sites of the following networks:
•Automatic Urban and Rural Network (www.airquality.co.uk)
•London Air Quality Network (www.londonair.org.uk)
•Hertfordshire & Bedfordshire Air Pollution Monitoring Network (www.hertsbedsair.org.uk)
•Kent and Medway Air Quality Monitoring Network (www.kentair.org.uk)
•Sussex Air Quality (www.sussex-air.net)
•South Cambridgeshire District Council (http://scambs-airquality.aeat.co.uk)
•Oxford Airwatch (www.oxford-airwatch.aeat.co.uk)
•Newham Council (http://apps.newham.gov.uk/pollution/)
•Air Quality Monitoring in Slough (www.aeat.co.uk/netcen/aqarchive/slough/site_map.html)
We dropped provisional values, keeping only ratified values. Some data came in volume
ratios, which we converted into mass units, using the conversion factors used for reporting
data to the European Commission:
•CO: 1 ppm = 1.16 mg/m3
•NO2: 1 ppb = 1.91 µg/m3
•O3: 1 ppb = 2.00 µg/m3
We multiply data on PM10 from TEOM analysers by 1.3 and data from BAM analysers
by 0.83 to obtain gravimetric equivalent measures. Annual means of pollutant concentrations
at station level are based on at least 100 observations.
33
Table C-1: Data sources
Variable Source Years covered
Mortality rates (per 100,000)
Mortality from all causes
Mortality from all circulatory diseases
Mortality from coronary heart disease
Mortality from acute myocardial infarc-
tion
Mortality from stroke
Mortality from bronchitis, emphysema
and
other chronic obstructive pulmonary dis-
eases
Mortality from chronic liver disease
including cirrhosis
Mortality from infectious and
parasitic diseases
Directly age-standardised
rates from Clinical and
Health Outcomes Knowledge
Baes (www.nchod.nhs.uk),
calculated using data on reg-
istered deaths from Office for
National Statistics (ONS) and
2001 Census based mid-year
population estimates from
ONS
1998-2005
Covariates
Smoking rate, regional level Clinical and Health Outcomes
Knowledge Base
1998,
2000-2005
Employment rate Labour Force Survey 1998-2005
Percentage of working age people educated
to degree level or higher
(www.nomisweb.co.uk)
Annual mean of summer daily Met Office - 1998-2005
maximum temperature MIDAS Land and
Annual mean of precipitation Surface Station Data
Table C-2: Correlation between annual pollutant concentrations
Correlation CO NO2PM10 O3
CO 1
NO20.6 1
PM10 0.4 0.6 1
O3-0.3 -0.5 -0.2 1
34
References
Air Quality Expert Group: 2004, Nitrogen Dioxide in the United Kingdom, Department for
Environment, Food and Rural Affairs, London.
Bell, M. L., Dominici, F. and Samet, J. M.: 2005, A meta-analysis of time-series studies
of ozone and mortality with comparison to the National Morbidity, Mortality, and Air
Pollution Study, Epidemiology 16, 436–445.
Chay, K., Dobkin, C. and Greenstone, M.: 2003, The Clean Air Act of 1970 and adult
mortality, Journal of Risk and Uncertainty 27(3), 279 – 300.
Chay, K. Y. and Greenstone, M.: 2003, The impact of air pollution on infant mortality: Evi-
dence from geographic variation in pollution shocks induced by a recession, The Quarterly
Journal of Economics 118(3), 1121–1167.
Clancy, L., Goodman, P., Sinclair, H. and Dockery, D. W.: 2002, Effect of air-pollution control
on death rates in dublin, ireland: an intervention study, The Lancet 360, 1210–1214.
Clyde, M.: 2000, Model uncertainty and health effect studies for particulate matter, Environ-
metrics 11, 745–763.
Committee on the Medical Effects of Air Pollution: 1997, Handbook on air pollution and
health, The Stationery Office, London.
Currie, J. and Neidell, M.: 2005, Air pollution and infant health: What can we learn from
California’s recent experience?., Quarterly Journal of Economics 120(3), 1003 – 1030.
Department for Environment, Food and Rural Affairs: 2007, The air quality strategy for Eng-
land, Scotland, Wales and Northern Ireland - Volume 1, The Stationery Office, London.
Department of Health: 1998, Smoking kills. A White Paper on tobacco, The Stationery Office,
London.
Department of Health: 2000, National service framework for coronary heart disease, Depart-
ment of Health, London.
Department of Health: 2006, Cardiovascular disease and air pollution - A report by the Com-
mittee on the Medical Effects of Air Pollutants, Department of Health, London.
Devlin, N. and Parkin, D.: 2004, Does NICE have a cost-effectiveness threshold and what
other factors influence its decisions? a binary choice analysis, Health Economics 13, 437–
452.
Dockery, D. W. and Pope, C. A.: 1994, Acute respiratory effects of particulate air pollution,
Annual Review of Public Health 15, 107–132.
Dockery, D. W., Pope, C. A., Xu, X., Spengler, J. D., Ware, J. H., Fay, M. E., Ferris, B. G.
and Speizer, F. E.: 1993, An association between air pollution and mortality in six U. S.
cities, The New England Journal of Medicine 329, 1753–1759.
35
Georgoulis, L. B., Hänninen, O., Samoli, E., Katsouyanni, K., Künzli, N., Polanska, L.,
de Bruin, Y. B., Alm, S. and Jantunen, M.: 2002, Personal carbon monoxide exposure
in five European cities and its determinants, Atmospheric Environment 36, 963–974.
Gerdtham, U.-G. and Johannesson, M.: 2003, A note on the effect of unemployment on
mortality, Journal of Health Economics 22, 505–518.
Griliches, Z. and Hausman, J. A.: 1986, Errors in variables in panel data, Journal of Econo-
metrics 31, 93–118.
Hoek, G., Brunekreef, B., Goldbohm, S., Fischer, P. and van den Brandt, P. A.: 2002, Associ-
ation between mortality and indicators for traffic-related air pollution in the Netherlands:
a cohort study, The Lancet 360, 1203–1209.
Janssen, N. A., de Hartog, J. J., Hoek, G. and Brunekreef, B.: 2000, Personal exposure to
fine particulate matter in elderly subjects: relation between personal, indoor and outdoor
concentrations, Journal of the Air & Waste management Association 50, 1133–1143.
Koop, G. and Tole, L.: 2004, Measuring the health effects of air pollution: to what extent can
we really say that people are dying from bad air?, Journal of Environmental Economics
and Management 47, 30–54.
Kousa, A., Monn, C., Rotko, T., Alm, S., Oglesby, L. and Jantunen, M. J.: 2001, Per-
sonal exposures to NO_2in the EXPOLIS-study: relation to residential indoor, outdoor
and workplace concentrations in basel, helsinki and prague, Atmospheric Environmment
35, 3405–3412.
Krewski, D., Burnett, R. T., Goldberg, M. S., Hoover, K., Siemiatycki, J., Jerret, M., Abra-
hamowicz, M. and White, W. H.: 2000, Re-analysis of the Harvard Six Cities Study
and the American Cancer Society Study of Air Pollution and Mortality, Health Effects
Institute, Cambridge, MA.
Nafstad, P., Haheim, L. L., Wisloff, T., Gram, F., Oftedal, B., Holme, I., Hjermann, I.
and Leren, P.: 2004, Urban air pollution and mortality in a cohort of Norwegian men,
Environmental Medicine 112, 610–615.
Neidell, M.: 2004, Air pollution, health, and socio-economic status: the effect of outdoor air
quality on childhood asthma, Journal of Health Economics 23, 1209–1236.
Pope, III, C. A.: 1996, Particulate air pollution and health: a review of the Utah valley
experience, Journal of Exposure Analysis and Environmental Epidemiology 6, 23–34.
Pope, III, C. A., Burnett, R. T., Thun, M. J., Calle, E. E., Krewski, D., Ito, K. and Thurston,
G. D.: 2002, Lung cancer, cariopulmonary mortality, and long-term exposure to fine
particulate air pollution, Journal of the American Medical Association 287, 1132–1141.
Pope, III, C. A., Burnett, R. T., Thurston, G. D., Thun, M. J., Calle, E. E., Krewski, D. and
Godleski, J. J.: 2004, Cardiovascular mortality and long-term exposure to particulate
36
air pollution. Epidemiological evidence of general pathophysiological pathways of disease,
Circulation 109, 71–77.
Pope, III, C. A. and Dockery, D. W.: 2006, Health effects of fine particulate air pollution:
lines that connect, Journal of the Air & Waste Management Association 56, 709–742.
Ruhm, C.: 2000, Are recessions good for your health?, The Quarterly Journal of Economics
115, 617–650.
Smallbone, K.: 2009, Air quality forecasting and its value as a preventive measure for sensitive
people - Presentation at Investigation of Air Pollution Standing Conference.
URL: http: // www. iapsc. org. uk/ document/ 1209_ K_ Smallbone. pdf
Stieb, D. M., Judek, S. and Burnett, R. T.: 2002, Meta-analysis of time-series studies of air
pollution and mortality: effects of gases and particles and the influence of cause of death,
age, and season, Journal of the Air & Waste Management Association 52, 470–484.
Wilkins, E. T.: 1954, Air pollution and the London fog of December 1952, Journal of the
Royal Sanitary Institute 74, 1–15.
Wilson, R. and Spengler, J.: 1996, Particle in our air - concentrations and health effects,
Harvard School of Public Health, Cambridge, MA.
37
