index.text

---
author:
- 'James D Smith, MSc, BA, PgCert, PgDip\'
bibliography:
- 'bibtex/PhD-KCL.bib'
title: |
    Using a dynamic exposure model to improve understanding of exposure to
    urban air pollution
...

Abstract {#abstract .unnumbered}
========

Epidemiological studies of the negative effects on health of poor air
quality are typically based on subjects’ residential address. These
’static’ methods may be assigning exposure to subjects/populations
incorrectly. Possible sources of error include the coarse spatial and
temporal scale of the pollutant data, failing to account for lack of
movement of the subjects, and not adequately modelling the effects of
microenvironments. This PhD takes a large Transport for London (TfL)
survey (the ’LTDS’) of Londoners daily activities and uses geographical
information science (GIS) techniques to create a detailed model (the
’LTDS-X’) of Londoners typical movements including time of day, location
and microenvironment. This model is then combined with the King’s
version of the Community Multiscale Air Quality model (CMAQ-Urban),
which is a multi-pollutant and multi-source high resolution spatial and
temporal model of UK air quality. By combining the LTDS-X with
CMAQ-Urban and then undertaking further micro-environmental modelling on
top of this (in-car, in-train, indoors, the London Underground) detailed
exposure estimates to a range of pollutants for the population of London
are calculated and then compared to the ’static’ exposure method.
Results show that exposure indoors, and whether or not subjects use the
London Underground, were important determinants of Londoners daily
exposure. The exposure modelling for when subjects were on the London
Underground was therefore investigated further with a measurement
campaign across the network, resulting in a GIS routing model of the
network (’TubeAir’). As a stand-alone model this will be useful for
future exposure studies in London, and it’s use in the LHEM was
demonstrated on a sample journey. This research concludes by exploring
the difficulty of evaluating hybrid exposure models in terms of the
representativeness of any exposure calculated, by measuring the
PM$_{2.5}$ exposure of a repeated number of cycling journeys and
comparing these to modelled exposures.

Dedication {#dedication .unnumbered}
==========

For Aimee.\
\
\
\
\
\
\
\
\
\
\
\
\
\
\

![image](images/nffc_logo.png) \[fig:nffc\_logo\]

Acknowledgements {#acknowledgements .unnumbered}
================

I would like to express gratitude to my supervisors, Dr Benjamin Barratt
and Dr Heather Walton. Their guidance, ideas and support have been
crucial. I would also like to thank Professor Frank Kelly. who gave me
the opportunity to start this part-time PhD programee in the first
place. In may ’day job’, current and previous work colleagues in our
group have helped me at various times, mainly Dr Sean Beevers, Dr
Christina Mitsakou, Dr Nutthida Kitwiroon, Dr Andrew Beddows, Dr Gregor
Stewart, Dr David Dajnak, Dr David Green, Dr Ian Mudway and Dr Gary
Fuller. University aside, special thanks go to my family, friends and in
particular my wife Aimee for putting up with my elevated stress levels
for so long. Also to Frankie Manning, for giving me a way to express
myself in a totally different environment.

Finally, a little thank you to the various online communities that I
have interacted with and sought advice from over the year, who were able
to offer advice and suggestions when I ran into practical problems that
nobody else could help with.

Declaration {#declaration .unnumbered}
===========

I, James David Smith declare that all the work submitted in this thesis
is my own and that all references are cited accordingly.\
\
\
Signed:\
\
Date:

Introduction
============

Air pollution, and it’s related health impacts, is an issue that urban
areas have been struggling with for many years. In Roman times the
philosopher Seneca commented (after modern translation) that his
disposition improved after leaving the ‘heavy air of Rome’ (@seneca1969)
and after large cities in the UK started to use coal as a fuel in the
13th century, the wife of Henry VIII complained of coal smoke in the air
when she visited Nottingham Castle (@Brimblecombe1999). The links
between air pollution and negative effects on human health are still
with us today, although the sources have changed and the visibility of
the pollution has in many cases become less apparent. The Lancet’s
Global Burden of Disease 2013 study ranked outdoor air pollution as the
7^th^ highest contributing risk factor to deaths globally in 2010
(Figure \[fig:gbd\_air\_pollution\]) (@Lim2012).

![20 Leading Risk Factors Contributing to Deaths Globally in 2010<span
data-label="fig:gbd_air_pollution"></span>](gbd_air_pollution)

However the methods by which exposure to air pollution is estimated are
under constant revision. Due to limitations in these methods,
epidemiologists may therefore be making misleading or incorrect
conclusions.

The matter of presenting and interpreting data in this field in the
opinion of the author, can be confusing and hard to interpret by
lay-people. Given better information and understanding of air pollution,
would the public change their behaviour and therefore lower their
exposure? Possibly leading to health benefits?

This transfer report begins by giving a general introduction to the
subject of air pollution, it’s health effects, and how health studies of
air pollution have estimated population exposure in the past. It then
conducts a more detailed review of dynamic exposure-health studies and
the ways in which geographic information systems and science (GIS) have
been used in this field. This sets the scene for novel research into
personal exposure of the population of London and the ways in which GIS
can be used to further this field, and the understanding of the
implications.

\[chap:intro\]

Background
==========

Air pollution - sources and behaviour {#sec:whatisairpollution}
-------------------------------------

### What is air pollution? {#subsec:whatisairpollution}

Air pollution is defined by @colls1997 as “material emitted into the air
from stationary or mobile sources, moving subsequently through an aerial
path and perhaps being involved in chemical or physical transformation
before eventually being returned to the surface”. This thesis focuses on
the places in which humans, particularly in urban environments, are
exposed to this pollution.

### Particle types and sizes {#subsec:particletypesandsizes}

Air pollution is a summary term for many sub-categories of pollutants.
Pollutants can be solid particles, liquid particles, or gaseous
material. They can also be classified as either a primary or secondary
pollutant. For example fumes emitted from the stack of a power station
are classified as primary pollutants, whereas low-level (ground level)
ozone formed by chemical reactions between primary pollutants catalysed
by sunlight are classed as secondary pollutants. The UK Department for
Environment, Food and Rural Affairs (DEFRA) summarises the main
constituents of air pollution and their typical sources as follows
(@DEFRA2011):

-   Particulate matter

    -   Combustion (traffic or stationary sources), sea-spray,
        construction, quarrying.

-   Oxides of nitrogen (NO$_{x}$)

    -   Combustion. Road transport, electrical supply industry,
        other industry.

-   Ozone (O$_{3}$)

    -   A secondary pollutant, not emitted directly from human-made
        sources, but formed as a result of reactions between other
        pollutants (NOx, VOCs) in sunlight.

-   Sulphur dioxide (SO2$_{2}$)

    -   Combustion of fuels such as coal and heavy oils by
        power stations.

-   Polycyclic aromatic hydrocarbons (PAHs)

    -   Many different sources. DEFRA uses Benzo\[a\]pyrene as a marker.
        Main sources are coal and wood burning, fires,
        industrial processes. Traffic combustion (diesel inparticular)
        is a major contributor.

-   Benzene

    -   Domestic and industrial combustion and road transport.

-   1,3-butadiene

    -   Combustion of petrol i.e. motor vehicles that use petrol as a
        fuel source

-   Carbon monoxide (CO)

    -   Occurs from incomplete combustion of fuels that contain carbon.
        Road transport, residential combustion and industrial combustion
        are the main sources.

-   Lead (Pb)

    -   Combustion of coal and nonferrous metals

-   Ammonia

    -   Mainly from argiculture such as manure, fertilisers and slurry.

When discussing the amount of pollutants in the air, either volumetric
or gravimetric units are used. Volumetric units quantify the ratio of
volume of pollutants to clean dry air ( itself a mixture of nitrogen,
oxygen, argon etc. ), whereas gravimetric units quantify the mass of the
material per volume of air. Most laws and guidelines, such as the EU Air
Quality Standards (see Section \[sec:healtheffects\]), use gravimetric
measurements. Table \[tab:pollution\_units\] summarises the
abbreviations for volumetric and gravimetric units which are used
throughout this research.

  **Volumetric**
  ---------------------------------------------------------------------------
  Parts per million of pollutant per parts of air volume (10^-6^ ppm)
  Parts per billion of air pollutant per parts of air volume (10^-9^ ppb)
  Parts per trillion of air pollutant per parts of air volume (10^-12^ ppt)
  **Gravimetric**
  milligrams of pollutant per cubic metre (mg/m^3^)
  micrograms of pollutant per cubic metre ($\mu \text{g m}^{-3}$)
  nanograms of pollutant per cubic metre (ng/m^3^)

  : Abbreviations for volumetric and gravimetric pollutant units<span
  data-label="tab:pollution_units"></span>

Of the ten pollutants listed above, over half list transport combustion
as being a source. When combined with proximity to humans in urban
environments it is easy to see why air quality in towns and cities, and
in particular the pollutants caused by vehicles in these environments,
receives such great interest in the field of air quality research and
environmental science.

### Urban environments {#subsec:urbanenvironments}

In many cities around the world hundreds of thousands of people now live
within metres of major pollution sources such as car-filled roads, power
stations or industrial plants. Due to increasing urbanisation this is
rising. According to the World Health Organisation (WHO), as of 2010,
more than 50% of the world’s population live in urban areas. This is up
from 40% in 1990. The prediction is that by 2050 this number will rise
to 70% (@GlobalHealthObservatory2012). As the numbers of people living
in cities has grown, so has the infrastructure required to support them;
much of which causes air pollution. High numbers of people are now being
exposed to air pollution above WHO guidelines during their normal day to
day activities. Given this close link between population density and
pollution, it is important to understand the complexity of air pollution
in urban environments.

In 2012 WHO published a review which summarised data on particulate
matter levels in major cities across the world. The data were split into
two categories, particulate matter of a diameter of less than 2.5
micrometres (referred to as PM$_{2.5}$) and particulate matter of a
diameter of less than 10 micrometres (referred to as PM$_{10}$)
(@WorldHealthOrganization2012). Although the health effects of poor air
quality will be discussed in Section \[sec:healtheffects\], to provide
immediate context, we can refer to WHO Factsheet no. 313 which gives the
following numbers as limits for ’acceptable and achievable PM levels to
minimise health effects in the context of local constraints’
(@WorldHealthOrganization2011). NO2 values are also included for future
reference.

               Annual mean ($\mu \text{g m}^{-3}$)   24 hour mean ($\mu \text{g m}^{-3}$)
  ------------ ------------------------------------- --------------------------------------
  PM$_{2.5}$   10                                    25
  PM$_{10}$    20                                    50
  NO$_{2}$     40                                    200

  : Table of ’acceptable’ WHO PM and NO$_{2}$ levels<span
  data-label="tab:whopmlevels"></span>

Figure \[fig:mapofpm10\] shows the annual average PM$_{10}$ levels for
each major world city for 2003-2010, weighted by population, from the
same WHO review.

![A map of PM$_{10}$ in major world cities<span
data-label="fig:mapofpm10"></span>](images/who_pm10_world_map)

As can be seen, there are many urban areas where the WHO annual mean for
PM$_{10}$ is exceeded. By way of example we can consider Beijing
(39.913889 N, 116.391667 W) which has a population of 15.59m and is
China’s second largest city (@TheUnitedNationsStatisticsDivision2013),
is subject to seasonal dust storms, hot humid summers and cold dry
winters - and suffers from severe air pollution problems. During the
1990s attempts were made at controlling air pollution in the city by
introducing the use of low-sulphur coal, using natural gas as an
alternative to coal, phasing out leaded petrol, and moving factories and
heavy industry outside of the city. However, due to increasing vehicle
numbers and the rapid growth of the industrial sector, particle levels
continued to remain higher than national standards (@Sun2004). The
nearby area of Hebei Province (which rings Beijing and is heavily
industrial) made the efforts of the government to control air quality in
Beijing even more difficult as the Hebei region has lower fuel quality
standards, and emissions frequently drift into Beijing (@Tuo2013) due to
prevailing winds (Meteorology is discussed in Section
\[subsec:meteorology\]).

During the early 21$^{st}$ century the use of air pollution monitoring
equipment became more commonplace and better data was collected to
enable scientists to understand the issues that Beijing (and similar
urban environments) faced. @Sun2004 found that during the years 1999 and
2000, PM$_{2.5}$ concentrations were in the range 37 – 357
$\mu \text{g m}^{-3}$ , and estimated a yearly average of 89.7
$\mu \text{g m}^{-3}$. The research concluded that coal burning and
traffic exhausts, along with dust from long range sources, were the
major pollution sources in the urban environment of Beijing (@Sun2004).

The study of pollution in urban environments is essential as these areas
are where humans are most readily exposed and they are where the sources
of emissions are most frequent. Although PM has been discussed here,
similar issues apply to other traffic linked pollutants such as NO$_{x}$
and PAHs. There are also other factors, other than the type of
pollutant, that complicate the understanding of pollution in urban
environments, such as weather and geography. Within these urban
environments there are hyper-local conditions that can raise and lower
levels. These notions are explored in sections \[subsec:meteorology\],
\[subsec:urbantopography\], \[subsec:microenvironments\] and
\[subsec:trafficpollution\].

### Meteorology {#subsec:meteorology}

Local weather conditions have a strong influence on air quality. Air
pollution can be removed from the air in the process of cloud formation,
and then deposited on the ground when the clouds turn to rain at a
future time and/or place. Falling rain can also remove pollutants from
the air by collecting the pollution and ’cleaning’ the air as it falls.
Both of these processes are grouped into the term ’wet deposition’. A 6
$\mu \text{g m}^{-3}$ difference in PM$_{10}$ was observed in Edinburgh
between days with no rainfall compared to those with more than 20mm of
rain (@DEFRA2007). In December 2013 it was even reported that China was
considering using ’cloud seeding’, i.e. the process of engineering the
weather to rain, as a method to lower air pollution in the most polluted
regions of China (@Slezak2013).

Wind can effect air quality by trapping or recirculating air pollution
(discussed in Section \[subsec:urbantopography\]), but can also disperse
the pollution or move it to entirely new areas. Pollutants emitted in
one area can often have their impact on the air quality in other areas
due to transportation by air currents. This notion was first proposed in
the 1960s when studying the acidification of lakes in Scandinavia
@Summers1976, where it was theorised that the high acid levels were due
to air quality elsewhere in Canada. Depending on the lifetime and
properties of a pollutant, it can be transported on scales ranging from
the street level to the global scale (@Monks2009). @Stohl2003 found
gases were being transported from North America to Europe. More locally,
an odour event in the South-East of England on 18 April 2008
(@TheGuardian2008) was found to have originated from agricultural
emissions in northern Germany (@Smethurst2012). The Geneva Convention on
Long-range Trans-boundary Air Pollution was established in 1979 to look
at ways to deal with this movement of air pollution between borders in
terms of national air quality guidelines and limits, and came into force
in 1983 (@UnitedNationsEconomicCommissionforEurope1983). Wind can also
dilute air pollution, though this varies by pollutant. NOx
concentrations were found to have halved at a monitoring station in
Hillingdon, UK when wind speed rose from 5 to 10 m/s-1, while PM$_{2.5}$
also decreased, however the coarse PM (PM$_{2.5 - 10}$) increased due to
re-suspension of particles that had previously settled (@DEFRA2007).

Direct sunlight (UV radiation) and higher temperatures, on hot summer
days, can initiate reactions with nitrogen dioxide which can lead to the
formation of ozone. The ozone and ozone forming chemicals remain in the
atmosphere and can be transported over regional and national borders.
This layer can then settle over cities such as London and lead to what
is often referred to as summertime ’smog’. The South-East of England
often has high concentrations during spring and summer as, amongst other
sources, it is close to European pollution sources (@LondonAir).
Although vehicle emissions of nitrogen oxide can have the effect of
reacting with the ozone to lower the level of ozone in areas where
vehicle emissions are particularly high.
(@EnvironmentalProtectionAgency2012).

### Urban topography {#subsec:urbantopography}

In the urban environment, streets are often bordered by tall buildings
which can influence pollution levels for people at ground level.
Topography of this nature is often referred to as a ’street canyon’. In
extreme circumstances such as on the streets of Hong Kong, skyscrapers
are littered throughout the city, but on smaller scales such as Oxford
Street (London, UK) similar issues occur but with modest building
heights. Being bordered by tall buildings creates a sheltering effect
from the wind, stopping particles being moved elsewhere. The typical
street-canyon effect occurs when there is a steady flow over the top of
tall buildings (see figure \[fig:street\_canyon\]), where the mean flow
is perpendicular to the direction of the street (@Britter2003). With
roof-level wind-speeds of 1.5-2 m/s, the air is recirculated within this
’box’ and air quality deteriorates as sources (usually traffic) emit
more fumes.

![Pollutant dispersion in a regular street canyon<span
data-label="fig:street_canyon"></span>](street_canyon)

Mexico City is an example of a city that is subject to this effect on a
large scale. Situated in central Mexico, North America (19.4328° N,
99.1333° W), Mexico City is one of the most populated cities in the
world and has an estimated population of about 21m (2011) living within
the Mexico City Metropolitan Area (MCMA) of 1,485 km^2^
(@TheUnitedNationsStatisticsDivision2013). Mexico City is located in the
crater of a large extinct volcano, which means that the entire city
suffers from the aforementioned canyon effect. Almost like it is
surrounded by skyscrapers. This is exacerbated by a fleet of older
vehicles with poor engines (the effects of which are discussed more in
Section \[subsec:trafficpollution\]), low levels of oxygen (due to the
high altitude of the city), and wind patterns that concentrate
pollutants in the western and southern parts of the city (@Garza1996)
where the population is most dense.

To summarise, air pollution in the urban environment can be effected by
the meteorology and topography of the area. On the smaller scale than
’area’ there are also micro-environments that people spend time in, for
example inside a bus, which can exhibit very different characteristics
than the rest of the city or even street.

### Micro-environments {#subsec:microenvironments}

Micro-environments are defined as the immediate small-scale environment
of an organism, especially as a distinct part of a larger environment.
Examples in the context of this thesis include the air quality inside a
vehicle, a house or underground train carriage, in the context of the
environment outside. Understanding how pollutant levels change in these
micro-environments is key, as much of our time is spent within these
environments, thus our exposure level to air pollutants whilst in them
could have important impacts on our health.

#### Indoors {#subsubsec:indoors}

The most common micro-environment is within buildings, the air we are
exposed to when we are at home, in the office or at school etc. WHO
calculated in 2005 that people spend 89% of their time indoors
(@WorldHealthOrganization2005). In these environments, people are
exposed to pollutants generated outdoors that penetrate to the indoor
environment, as well as to pollutants produced indoors.

The EXPOLIS study examined how much time people spend in these
environments by asking 1427 people across their European project
partners to complete activity diaries. They found that the amount of
time people spend indoors varied by whether the people were employed or
not, in what type of job, whether they lived alone and/or whether they
had children. Gender and season of year were also found to be factors.
By bringing together some of the figures from their study, it can be
calculated that the mean number of hours that people spend indoors, not
taking account of any other adjustment factors, was 20.66 hours per day
(@Schweizer2007). So understanding indoor air quality, which would
include filtration of outdoor air into the building as well as
pollutants whose sources are inside, is important in understanding
personal exposure.

However research on indoor pollution has not had the same focus as
outdoor pollution for a number of reasons. Firstly the perceived need to
deal with coal and traffic emissions, the ease of monitoring outdoor
pollution using fixed monitoring sites (compared to monitoring every
home). Secondly epidemiologists have traditionally only linked outdoor
pollutant concentrations to health health issues, furthermore,
legislating the air that people can breathe in their own homes can be
seen as intrusive to people’s private lives, finally the funding and
policy initiatives around air pollution research has mainly come from
developed countries, which who do not have such an issue with indoor
pollution as developing countries (@WorldHealthOrganization2010) (due to
developing countries using solid fuel for cooking and heating inside
their homes more).

The pollutants that are emitted indoors, and which the US Environmental
Protection Agency (EPA) focus on in their guide to indoor air pollution,
include Volatile Organic Compounds (VOCs), carbon monoxide and nitrogen
dioxide (@UnitedStatesEnvironmentalProtectionAgency2008). VOCs in indoor
air come mostly from products used around the house such as paint,
varnish, cleaning sprays, air fresheners and pesticides but can also be
emitted by building materials and furnishings. Carbon monoxide and
nitrogen dioxide on the other hand are more commonly associated with the
use of indoor furnaces, gas cookers, gas heaters, leaking chimneys and
people smoking tobacco indoors.

The EPA report however, only focuses on indoor air pollution relevant to
buildings in the US. In different parts of the world indoor air quality
varies in terms of the pollutants and the impact, especially in Asian
countries where less clean combustion fuels are often used for cooking
in the home and the numbers of people that smoke while indoors is
greater (@Lee2010). @Baumgartner2011 sampled PM$_{2.5}$ in 44 kitchens
of the Yunnan area of China in 2010, where 95% of the kitchens used wood
or crop residue for cooking, and 96% used a mix of wood-charcoal and
wood or crop reside for heating. During the summer months, when the
sampling was done, mean concentrations were found to be around 107
$\mu \text{g m}^{-3}$. So presumably would be higher during the winter
when the windows are likely to be closed and air cannot escape as
easily. Similarly, @Li2011 compared pollutant concentrations in kitchens
in relation to different types of stoves in Peru. Means of 181
$\mu \text{g m}^{-3}$ and 3.5 ppm were found for the open-pit stoves for
PM$_{2.5}$ and CO respectively. In a larger study across 168 venues in
China, Japan, Korea, Malaysia, Pakistan and Sri-Lanka, PM$_{2.5}$
measurements were made and an average indoor level of 137
$\mu \text{g m}^{-3}$ was found (smoking venues were 156
$\mu \text{g m}^{-3}$, non-smoking venues were 34
$\mu \text{g m}^{-3}$).

Poor indoor air quality is not always due to indoor sources. The
pollutant levels outside an indoor environment has been found to have an
impact, although this is dependant on many mitigating factors such as
the buildings air filtration units, proximity to outdoor sources, and
wind speed/direction. In North America, homes close to Ambassador Bridge
(Detroit) were measured over five 24 hour periods and they found that
ambient black carbon concentrations significantly contributed to indoor
concentrations regardless of wind speed @Baxter2008. In Osaka, Japan
fine PM (PM$_{2.5}$) was significantly correlated with fine PM outside
the properties and it was estimated that about 30% of indoor PM$_{10}$
particles were from diesel exhausts from nearby roads (@Funasaka2000).
In Europe (Prague, Czech Republic) PM$_{2.5}$ was sampled in a school
gym during 2005 and 2006, and levels were found to be similar to a
nearby fixed-site monitor (24.03 compared to 25.47) (@Branis2009).

#### In-vehicles {#subsubsec:invehicle}

The air quality people are exposed to between these indoor
micro-environments, when travelling, can differ greatly from general
ambient concentrations. Similarly, when space inside vehicles has its
own air quality micro-environment. Conditions can be affected by having
windows open or closed, the type of vehicle, and the vehicle’s location
amongst other factors. @Adams2001 measured PM$_{2.5}$ during 465
journeys in London over a three week period, at peak and off-peak times
of the day during the summer of 1999 and winter of 2000.

  ------------------------------------------------------------------------------ --
  **<span>Transport mode</span> & **<span>Mean ($\mu \text{g m}^{-3}$)</span>\   
  Bus & 39.0\                                                                    
  Car & 37.7\                                                                    
  Underground tube & 247.2\                                                      
  Overground tube & 29.3\                                                        
  ****                                                                           
  ------------------------------------------------------------------------------ --

  : PM$_{2.5}$ by transport mode from @Adams2001<span
  data-label="tab:adams_transport_means"></span>

They observed a great deal of variability between travel modes (see
table \[tab:adams\_transport\_means\]), and against typical ambient
PM$_{2.5}$ levels (around 10-30 $\mu \text{g m}^{-3}$) recorded for
central London. Outside of London, also in 1999/2000, @Gulliver2004
conducted in-vehicle monitoring along a stretch of road in Northampton
(80 km North/North West of London). They also observed elevated levels
of particulates inside the vehicle (table
\[tab:gulliver\_vehicle\_means\]).

  --------------------------------------------------------------------------- --
  **<span>PM Fraction</span> & **<span>Mean ($\mu \text{g m}^{-3}$)</span>\   
  PM$_{10}$ & 43.16\                                                          
  PM$_{2.5}$ & 15.54\                                                         
  ****                                                                        
  --------------------------------------------------------------------------- --

  : In-vehicle PM from @Gulliver2004<span
  data-label="tab:gulliver_vehicle_means"></span>

Neither of these authors comment on air quality inside vehicles or the
resultant exposure. They only conclude that in-vehicle pollutant
concentrations cannot be taken to be the same as outdoor values. The
situation is further complicated by additional variables such as whether
windows are open or closed, the speed of the vehicle, or the number of
people inside the vehicle.

### Traffic pollution {#subsec:trafficpollution}

Between the years of 1950 and 1994, there was a dramatic increase in
vehicle traffic on the worlds roads. Vehicle numbers increased from 53
million, to 460 million in the space of 44 years. However, this rate of
growth was not uniform around the world (@schwela2002). The rate slowed
considerably in industrialised countries, but population growth and
increased urbanisation and industrialisation accelerated the use of
vehicles elsewhere (see figures \[fig:europe\_vehicles\_per\_1000\] and
\[fig:china\_malaysia\_brazil\]) (@DepartmentforTransport2012).

![Vehicles per 1000 people from @TheWorldBank2013<span
data-label="fig:europe_vehicles_per_1000"></span>](europe_vehicles_per_1000)

![Vehicles per 1000 people from @TheWorldBank2013<span
data-label="fig:china_malaysia_brazil"></span>](china_malaysia_brazil)

Against this backdrop of increasing numbers of vehicles, there is
emerging evidence that traffic emissions are harmful to human health. In
2010 the US Health Effects Institute (HEI) published the findings of a
systematic review of evidence about traffic pollution, and whilst noting
that there were many areas still needing further research, there was
evidence to support a casual relationship between exposure to
traffic-related air pollution and asthma (@HPotHEoT-RA2010).
Toxicological research has now also started to link not only traffic
emission pollutants, but also non-exhaust pollutants such as road
abrasion, tyre wear and brake wear to adverse health effects
(@WorldHealthOrganization2013). The latter being particularly
significant given that there are no laws which consider this element of
traffic pollution and therefore no guidelines or limit values. Despite
this, measuring and defining pollution solely from traffic sources in
the urban environment is technically difficult, which makes linkages
with exposure estimates and health effects problematic. Different
studies have therefore used different pollutants as markers for traffic
emissions. Epidemiological studies have often used NO$_{2}$ as a marker
for combustion-related pollutants, in particular those emitted by road
traffic or indoor combustion sources (@WorldHealthOrganization2010).
More recently black carbon has started to become the standard way of
measuring diesel emissions due to it’s relative ease of measurement
using optical techniques such as micro-aethalometers.

Taking the city of Beijing as example again, in 2008 the Olympic Games
were held there and this heightened the world’s interest in Beijing’s
air quality and put the issue under national scrutiny. Global newspapers
focused on the effects that poor air quality might have on the
performance of the athletes. The reporter James Reynolds of the BBC
wrote “*China is spending billions of pounds on new roads, new venues
and on perfect celebratory shows but all that may come to nothing unless
this city cleans up its air*” (@BBC2007). Under this pressure, to try
and bring air quality problems under control (at least in the short term
while the Olympics were taking place) the Beijing Government implemented
a number of measures in the run-up to the games. Stricter vehicle
emission standards were adopted, better public transport infrastructure
was developed, and from July 2008 to September 2008 a traffic demand
management scheme was introduced whereby odd/even vehicle registrations
took it in turns to be used on the roads on alternate days (@Wang2009).
This provided an ideal real-world experiment for the local scientists to
quantify how much of the areas pollution was due to vehicle emissions.
Data on black carbon levels was collected by on-road, fixed background
and fixed road-side monitors, and then analysed to investigate whether
the scheme had achieved the desired effect.

![Black carbon concentrations during 2008 in Beijing from @Wang2009<span
data-label="fig:blackcarbonolympic"></span>](black_carbon_olympic)

The results from this study are shown in figure
\[fig:blackcarbonolympic\] and demonstrate how mean black carbon
concentrations dropped to around 5 $\mu \text{g m}^{-3}$ during the
Olympics (second boxplot) compared to around 14 $\mu \text{g m}^{-3}$
after the Olympics (third boxplot). In addition, during the Olympics on
days when the main sporting events were happening, there was a further
reduction to around 4 $\mu \text{g m}^{-3}$ (first boxplot). The traffic
in Beijing, at least within the limits of this small subset of data,
seemed to be contributing to around 10 $\mu \text{g m}^{-3}$ of black
carbon pollution in the air. The authors of this study go on to conclude
that the main source of emissions in Beijing at the time are from
traffic, and that the traffic demand management scheme was effective at
bringing emissions down to within acceptable (WHO) limits. However this
seems to be a simplification of the issue, especially given the impact
of factories in the vicinity (discussed in
\[subsec:urbanenvironments\]). Nonetheless, the exposure of the
residents of Beijing to pollution, a debatable proportion of which is
from tailpipe emissions, is high.

In Europe, where factories and heavy industry tend not to be based
within urban centres, the proportion of the populations exposure related
to traffic emissions is high. Often, due to meteorology (see
\[subsec:meteorology\]), the emissions may also be from other urban
centres. For example emissions from outside London are estimated to
account for around 40% of NO$_{2}$ concentrations (with the other 60%
being generated locally) (@GreaterLondonAuthorityGLA2010. This ratio
changes depending on different spatial resolutions. In areas close to
roads, the contribution of traffic emissions to overall pollution levels
is much higher due to the proximity to sources (vehicles). The effect of
traffic emissions as a percentage of airborne pollutants is well
demonstrated by a study by @Mayer1999. Figure \[fig:stuttgarttraffic\]
identifies clear trends related to rises in NO and NO$_{2}$ during
morning and evening rush hours on working days.

![Average weekly and diurnal cycles of NO, NO$_{2}$ ,O$_{3}$ and O$_{x}$
at the urban air-quality station Stuttgart-Bad Cannstatta for the period
1981-1993 from @Mayer1999<span
data-label="fig:stuttgarttraffic"></span>](stuttgarttraffic)

As air pollution from vehicles is harmful to health, and traffic
pollution is so prevalent in urban environments, more permanent
long-term attempts to reduce traffic pollution are ongoing in most major
cities and countries around the world. Though different countries have
sought to achieve this in different ways. Professor Williams recent
(December 2013) article for the website ’The Conversation’ explains how
the EU have attempted to legislate to reduce vehicle emissions
(@Williams2013), prompted by the Kyoto Protocol of 1997 which was linked
to the United Nations Framework Convention on Climate Change
(@UnitedNations1998). Regarding vehicle emissions the protocol sought to
specifically reduce CO$_{2}$ emissions and they did this by insisting
that car dealers in new passenger cars must provide potential buyers
with useful information on these vehicles’ fuel consumption and CO$_{2}$
emissions and this information must be clearly displayed. Limits were
placed on CO$_{2}$ emissions, as a ratio of kilometres travelled, to
encourage more efficient use of fuel and therefore lower emissions. This
legislation encouraged companies making vehicles for the EU markets to
invest in the production and marketing of diesel vehicles, as they
provide better fuel consumption than petrol vehicles. Europe now has a
vehicle fleet which is predominately diesel. The problem with this is
that diesel cars emit significantly higher levels of air pollutants than
petrol cars fitted with three-way catalytic converters (@Williams2013).
The difference between the two emission levels are even greater when
considered in the real-world rather than measured in laboratory
conditions (@Carslaw2011). A study in 2007 estimated that the health
effect of favouring diesel vehicles over petrol vehicles had the
combined effect of contributing to approximately 1850 additional
premature deaths over the period 2001-2020, or around 90 premature
deaths per year (@Mazzi2007).

The alternative to both diesel and petrol vehicles are vehicles that
produce different types of emissions, or low/no emissions. These are
often referred too as ultra-low emission vehicles, such as hybrid or
electric. Unfortunately for air quality in urban environments they are
currently a very small percentage of new vehicles. In 2013 only 1% of
new registrations in the EU (with some notable exceptions such as 4.5%
in the Netherlands where generous financial incentives have been offered
since 2007) were low or no emission vehicles (@Transportation2013).

### Summary {#subsec:whatissummary}

Section \[sec:whatisairpollution\] introduced the subject of air
pollution. It was considered as non-naturally occurring material in the
air, or material which has had it’s composition or levels altered by
non-natural sources. It can take many different forms and be categorised
in different ways, for example particulate matter, nitrogen oxides,
ozone or sulphur dioxide. It was discussed how many of the causes of air
pollution are linked to vehicle combustion engines and that in urban
environments, where people are increasingly living, this places the
sources and public in close proximity to each other. This can be
affected (both positively and negatively) by different meteorological
conditions and the topology of the region, city, and even individual
streets and buildings. Within these environments, it was explained that
there are micro-environments such as inside vehicles and buildings which
can also raise or lower pollution levels. As this research intends to
focus on urban environments, traffic emissions were then considered in a
little more detail. The Beijing Olympics 2008 was used as a case-study
to understand the impact that traffic emissions can have on air quality
in a major city, and the diesel dominated vehicle fleet of Europe was
then explained (touching on the impact compared to petrol that this has
had on air quality).

Now that the subject of air pollution has been introduced, the next
section of this background will give an overview of the impact on human
health from air pollution i.e. why Section \[sec:whatisairpollution\]
actually matters to us.

Health effects of air pollution {#sec:healtheffects}
-------------------------------

> “Clean air is considered to be a basic requirement of human health and
> well-being. However, air pollution continues to pose a significant
> threat to health worldwide” (@WorldHealthOrganisation2006).

### An overview {#subsec:anoverview}

From the 1600’s onwards coal was the main source of heat and energy in
major UK cities. Concern from the scientific community and coherent
programs of research about the possible negative effects of this fuel
source were limited. When undertaken the research often focused on poor
visibility or damage to buildings rather than human health. It was the
early 1900’s when coherent and robust studies began to investigate
mortality and links to ‘fog’, as it was known at the time. Notably with
Russell’s paper *’The Influence of fog on mortality from respiratory
diseases’* being published in The Lancet in 1926 (@Russell1926) . This
publication preceded London’s ’Great Smog’ of 1952, which was one of the
UK’s most important air pollution events in history in terms of
realisation of the links between pollution and health. Research
conducted since this event has had a great impact on the study of air
pollution, public perception and government regulation to combat it.
Data at the time showed that a rise in fog (pollution) levels was
closely followed by rises in mortality and morbidity (@Bell2003). At the
time it was estimated that between 3,500 and 4,000 more people died than
would have normally been the case in this period (See figure
\[fig:greatsmogdeaths\] from @GreaterLondonAuthorityGLA2002). The rise
in mortality was originally attributed to influenza, however sensitivity
analysis by @Bell2003 revealed that only an extremely severe influenza
epidemic could have accounted for the excessive deaths recorded for that
period. Subsequent reanalysis of the data estimated that between
December 1952 and March 1953 there were actually 13,500 more deaths than
during the same time period the previous year, attributable to
(controlling for temperature and influenza) rather than the 3000––4000
generally reported for the episode.

![Recorded deaths comparison during ’Great smog’ period from
@GreaterLondonAuthorityGLA2002<span
data-label="fig:greatsmogdeaths"></span>](great_smog_deaths)

Once this explicit link between air pollution and health became
apparent, laws and regulations began to be written and passed. For
example the Clean Air Act in 1956 (with various revisions over time,
notably in 1968), the 1970 EC Directive (70/220/EEC), the 1974 Control
of Pollution Act and the 1979 International Convention on Long Range
Transboundary Pollution. It is now widely accepted that air pollution
has harmful effects on human health (@WorldHealthOrganization2013.
Although in the Western world, the sources of pollution have shifted
from using coal for heating and cooking, to be dominated by combustion
engines in vehicles and similar (as discussed in Section
\[subsec:trafficpollution\]).

When considering the effects of air pollution on public health, studies
on large groups of people (tens of thousands plus) often use
epidemiological methods. Studies using epidemiological methods will be
discussed many times during this thesis, therefore a brief definition of
epidemiology and the key terms are now given.

Epidemiology is the the study of how often disease/poor health occurs in
a group of people, and the factors that lead to it. Or, more technically
defined by Bonita in a WHO publication as *’the study of the
distribution and determinants of health-related states or events in
specified populations, and the application of this study to the
prevention and control of health problems’* (@Bonita2006). Some key
terms include (adapted from ):

-   Incidence: The number of new ill people in the population over a
    specified time period

-   Prevalence: The existing number of ill people in the population over
    a specified time period.

-   Burden of disease: The total significance of the disease or illness
    to wider society. For mortality this is often measured in years of
    life lost.

-   DALY (Disability-Adjusted Life Year): A statistic to represent the
    health of a population. One DALY represents one lost year of healthy
    life and is used to estimate the gap between the current health of a
    population and an ideal situation in which everyone in that
    population would live into old age in full health.

Epidemiological studies have ’*For decades \[ ... \] been a cornerstone
of our approach to investigating the health effects of air pollution and
have been a principal basis for setting regulations to protect the
public against adverse health effects*’ @Zeger2000. Recent high profile
examples that look at air pollution include, but are not limited to;
respiratory problems (@Peacock2011), cardiovascular issues (@Brook2010)
and cancer (@Iii2012, @loomis2013). Indeed, recently (17 October 2013),
the International Agency for Research on Cancer (IARC) classified
outdoor air pollution as carcinogenic to humans (@loomis2013). In an
IARC press-release, Dr Kurt Straif, Head of the Monographs Section
stated *“The air we breathe has become polluted with a mixture of
cancer-causing substances. We now know that outdoor air pollution is not
only a major risk to health in general, but also a leading environmental
cause of cancer deaths”*.

However as discussed in Section \[subsec:particletypesandsizes\], the
term ’air pollution’ covers a myriad of different pollutants. Thus it is
important to untangle which pollutants are more or less harmful to
health. This will potentially help people avoid areas with pollution
that is most harmful, and help politicians and public sector workers to
develop policies that are most effective at reducing the most harmful
types of pollution. At present research in this area is underway to
fully answer which pollutants are most harmful to health. Evidence
points towards short-term exposure to PM$_{10}$ as having negative
health effects, but long-term exposure to PM$_{2.5}$ is a stronger risk
factor for mortality (@WorldHealthOrganization2013a, @Dockery1993).
Hence, a broad overview of epidemiological studies of the health effects
of air pollution that consider PM$_{2.5}$ will now be discussed.

Worldwide, WHO estimate that PM$_{2.5}$ causes about 9% of lung cancer
deaths, 5% of cardiopulmonary deaths, and 1% of respiratory infection
deaths (@WorldHealthOrganization2012). In 2013 the Global Burden of
Disease publication ranked exposure to air pollution and particulate
matter as one of the top ten risk factors for health globally,
estimating that over 430,000 premature deaths and around 7 million years
of healthy life were lost in Western, Central and Eastern Europe in 2010
from exposure to fine particulate matter (@Brauer2012) n.b. fine refers
to particulate matter smaller than 1 micron in diameter, ultrafine
smaller than 2.5 microns (includes the 1 micron particles), and coarse
smaller than 10 microns (includes the 1 and 2.5 micron particles).

In a study looking at PM$_{2.5}$ and anthropogenic ozone, @Silva2013
recently modelled ozone and PM$_{2.5}$ surface concentrations for the
entire world, then used concentration-response functions for long-term
exposure and mortality (from an American Cancer Society publication) to
estimate that annually and globally there are about 2.1 million
premature deaths from respiratory problems linked to PM$_{2.5}$, with
these being split 93:7 between cardiopulmonary disease and lung cancer
(@Silva2013).

We can therefore see that the health effects of PM$_{2.5}$, when taken
in context of large populations are significant. However, these figures
are likely to be the tip of a much larger concern as most do not include
morbidity i.e. poor health and detriments to the populations quality of
life that do not result in death. This is illustrated by figure
\[fig:airpollutionpyramid\] (@Mannino2000) showing a greater number of
the population have less severe health effects that still have a burden
on public well-being.

![Air pollution health effects pyramid<span
data-label="fig:airpollutionpyramid"></span>](air_pollution_pyramid)

In the UK, COMEAP (the Committee on Medical Effects of Air Pollutants)
has been set-up to provide advice to the government and related agencies
via the Department of Health’s Chief Medical Officer on the harmful
effects of air pollution. COMEAP regularly publish reports summarising
findings into the health effects of pollutants. In their 2010 report on
mortality, they estimated that around 29,000 deaths in the UK in 2008
were attributable to PM$_{2.5}$
(@CommitteeontheMedicalEffectsofAirPollution2010). Expressing this
differently, they estimated that air pollution may have contributed to
the earlier deaths of about 200,000 people in 2008, with an average loss
of life of about two years per death affected. Also on a UK scale,
@Yim2012 estimated that PM$_{2.5}$ causes about 19,000 premature deaths
in the UK and 3,200 in London. Note that the reason for these figures
being lower than that of COMEAP is that this study focused solely on the
deaths attributable to traffic emissions rather than all sources (and
gives at least in this context a rough idea of the impact of traffic
emissions compared to non-traffic emissions).

On a more regional level, @Miller2010 used life-tables and
concentration-response coefficients to estimate deaths attributable to
PM$_{2.5}$ to be 4,267 in Greater London for 2008.

While epidemiological studies find strong links between air pollution
and poor health, particularly pulmonary and cardiovascular disease, the
mechanism of **how** air pollution causes mortality is not yet fully
understood.

Chamber studies are often used to identify the toxicological effects of
air pollutants. Human subjects will be medically examined before entry
to the chamber, then sealed inside for a set period of time, and
re-examined after exposure to pollutants. The atmosphere within can be
carefully controlled by investigators to simulate the environment of
their choice, in the case of air pollution normally this is heavily
polluted air. Lung function tests, haematology and fiberoptic
bronchoscopy are undertaken to investigate the pathological pathways by
which pollutants may cause disease and poor health.

@Salvi1999 exposed 15 healthy human volunteers in an environmental
chamber to clean air and then diluted diesel exhaust, for one hour at a
time. Significant increases in neutrophils and platelets, markers of
stressed airways and lungs, were observed in the subjects peripheral
blood after the diesel exposure, however lung function measured before
and after exposure revealed no decline. The study demonstrated that at
high concentrations of diesel exhaust, at least in the short term, there
is a systemic and pulmonary inflammatory response in healthy human
volunteers, which is not detected by standard lung function measurements
alone. A further study by @Salvi2000 exposed healthy human volunteers in
chambers to a range of particles/diesel exhaust and fiberoptic
bronchoscopy was performed six hours after each exposure, the results
suggest airway leukocyte infiltration as the underlying mechanism for
diesel exhaust-induced respiratory health outcomes. In a study by
@Ghio2000 no immediate symptoms were observed, however 18 hours after
exposure inflammation was identified in the lower respiratory tract,
particularly in those with the highest particulate exposure compared to
clean filtered air.

The afore-mentioned studies suggest that pulmonary inflammation in the
airways is responsible for damage to the lungs. Although more studies
are needed to pin-point the specific pathways. Chamber studies have
helped to understand the inflammation pathways following short-term,
however further issues also need to be addressed, for example, the lack
of clarity between the length of exposures and the subsequent responses
and how long the lag is between exposure and underlying biological
mechanism responses (such as in the studies just mentioned where
different effects were noted after 6 hour and 18 hours)
(@EnvironmentalProtectionAgency2009).

To summarise, in a recent update to the American Heart Association’s
position on air pollution and cardiovascular disease @Brook2010
described the probable mechanisms into three general pathways following
inhalation of particulate matter:

1.  Release of proinflammatory mediators (eg, cytokines) from activated
    immune cells, or platelets or vasculoactive molecules (eg,
    Endothelin, histamine), or microparticles endothelium of blood
    vessels in the lung

2.  Perturbation of systemic autonomic nervous system balance or heart
    rhythm by particle interactions with lung receptors or nerves

3.  Translocation of PM (ie, ultrafine particles) or particle
    constituents (organic compounds, metals) into the systemic
    circulation

A simplified diagram from the same reference (@Brook2010) has been
adapted and is shown below (fig \[fig:biological\_pathways\]) for
illustrative purposes.

![The biological pathways linking PM exposure with cardio–vascular
disease from @Brook2010<span
data-label="fig:biological_pathways"></span>](biological_pathways)

A detailed review of the biological mechanisms underlying disease is not
the scope of this thesis. Rather it is concerned with the methods of
estimating levels of exposure to air pollution. The presumption of this
research is that even low levels of exposure to air pollution is
harmful. This is justified by studies such as the European ESCAPE
project who found that long-term exposure to fine particulate pollution
is associated with mortality, even at concentration ranges well below
the present European annual mean limit values (@Beelen2013). The work of
@Brauer2002 and his group looked at the use of minimum threshold values
for exposure to PM$_{2.5}$ and found that due to exposure
miss-classification, population-level thresholds were apparent at lower
ambient concentrations than common personal thresholds (such as the EU
limit values discussed in table \[tab:whopmlevels\] of Section
\[subsec:urbanenvironments\]).

In summary, epidemiological and toxicological studies have shown air
pollution is a major environmental risk to health. By reducing air
pollution levels, and exposure to air pollution, governments should be
able to reduce respiratory symptoms, heart disease, and lung cancer in
their population. In order to calculate the estimates of the detrimental
effects on people’s health, exposure and dose must be calculated as
accurately as possible. That is the amount and types of air pollution
that people are exposed to, and breathe in,every day of their lifetime.
This can be done in many different ways, and methods are undergoing
regular revision as scientists attempt to improve their accuracy.

One of the key issues to address to further understand the links between
pollutant exposure and poor health is the duration of exposure. Does
long-term exposure to low levels of pollution have the same effect as
short-term exposure to high levels of pollution in terms of disease
prevalence or DALY? The importance of short-term versus long-term
exposure on health is explored in the next section.

### Long term exposure v. short term exposure {#subsec:longtermvshortterm}

There are few studies that compare the effects of long and short-term
exposure to air pollution on a large population, at adequate spatial and
temporal scales, for a range of pollutants, and with appropriate health
information to enable us to determine which has the most impact on
health.

This section therefore presents a selection of the short term and long
term studies conducted to date, as well as the small number of studies
that have attempted to reconcile the two. This is an issue that is being
wrestled with by the community at present.

The most widely cited large-scale long-term exposure/health studies are
on the effects of nitrogen dioxide in New Zealand @Scoggins2004, fine
particles globally in the Global Burden of Disease 2010 ( @Brauer2012)
and a meta-analysis review of both by @Faustini2014. Scoggins modelled
annual average NO$_{2}$ concentrations over 3 km x 3 km grid squares in
Auckland for the years 1996-1999 and linked these to mortality data for
the same region provided by the Health Information Service. By doing
this they were able to estimate that there was a 1.3% increase in
mortality for a 1 $\mu \text{g m}^{-3}$ rise in NO$_{2}$ annual average
values. Regional estimates of deaths and DALYs for the years 1990, 2005
and 2010 were estimated by Brauer using global estimates of long-term
average ambient concentrations of fine particles (PM$_{2.5}$) and ozone
at 0.1^o^ x 0.1^o^ spatial resolution for 1990 and 2005. Thirdly,
Faustini’s review brought together studies that looked at NO$_{2}$ and
PM$_{2.5}$ on population mortality and found that there was rise of
1.04% and 1.05% per years of life per 10 $\mu \text{g m}^{-3}$ for
NO$_{2}$ and PM$_{2.5}$, respectively. Meta-analysis by @Hoek2013
estimates an excess risk of 6% per 10 $\mu \text{g m}^{-3}$ increase in
PM$_{2.5}$ exposure (see \[fig:long\_term\_meta\_analysis\] below).

![Association between PM$_{2.5}$ and all-cause mortality (Relative risk
per 10 $\mu \text{g m}^{-3}$) from @Hoek2013<span
data-label="fig:long_term_meta_analysis"></span>](long_term_meta_analysis)

As stated, links between long-term exposure and poor health seem to be
fairly robust, however as long exposure time-frames that are used
(typically annual average concentrations), and are combined with large
spatial scales of hundreds of kilometres, it is possible that much of
the detail in the results is being missed - which will be discussed more
in sections \[sec:staticexposurehealth\] and
\[sec:dynamicexposurehealth\]. Furthermore, no short-term exposure on
health effect outcomes from poor air quality are considered in these
’long-term’ studies, or a comparison of the importance of long-term
versus short-term exposure effects.

On the flip-side, short-term epidemiological exposure health studies
tend to suffer from similar issues from the perspective of someone
trying to weigh-up short-term versus long-term exposure. The review by
@Brook2010 (briefly discussed earlier in \[subsec:anoverview\])
concluded that short-term epidemiological time-series studies tend to
find that around a 10 $\mu \text{g m}^{-3}$ increase in mean 24-hour
PM$_{2.5}$ concentrations increases the risk of cardiovascular mortality
by approximately 0.4% to 1.0% (though doesn’t comment on other
pollutants). But that importantly this risk is not distributed evenly
across the population i.e. people with existing medical conditions or
the elderly are more vulnerable to effects triggered by this short-term
exposure (which comes back to the point in the previous paragraph about
missing detail). Furthermore, a review of time series studies by
@Atkinson2014, looking at relationships between PM$_{2.5}$, daily
mortality and hospital admissions found similar percentage increases to
Brook. A 10 $\mu \text{g m}^{-3}$ in PM$_{2.5}$ was associated with a
1.04% increase in mortality. However a comparison in the same datasets
with long-term exposure is not available.

Two recent studies that have sought to answer the question of short-term
versus long-term exposure, or at least explore it, are those by
@Kloog2013 and @Beverland2012a. Kloog et al. geo-coded all deaths in
Massachusetts (USA) between the years 2000-2008, modelled short-term and
then long-term PM$_{2.5}$ concentrations for the area of their data-set,
and then used time-series analysis to try and examine the relationships.
They found for short-term relationships (day of death, and three days
prior) that every 10 $\mu \text{g m}^{-3}$ in PM$_{2.5}$ there was a
2.8% increase in PM mortality. Then for long-term exposure they found
that for every 10 $\mu \text{g m}^{-3}$ increase in PM$_{2.5}$ the odds
of death occurring rose to an odds-ratio of 1.6 (which simplistically
equates to around a 60% increase). Leading them to conclude that the
effects of long-term air quality appear much more pronounced than
short-term. Contention remains whether their definition of short-term is
the most appropriate. Since the investigators only looked at PM$_{2.5}$,
it might be that the effects of PM$_{2.5}$ are more pronounced in the
longer–term, but NO$_{x}$ may have the opposite relationship. More data
collection and analysis is needed.

@Beverland2012a take a similar approach to Kloog et al. by
retrospectively using mortality data from the ’Renfrew–Paisley’ and
’Collaborative’ cohort studies which began in the 1970s in Glasgow
(Scotland), linked to black-smoke data (the only pollutant routinely
measured at the time). They too considered short-term to be three days
lag, and found that there were short-term exposure-mortality
associations greater than those found in the general population, indeed
rises in black smoke levels were affecting mortality. Furthermore, there
were also long-term mortality associations, which were more strongly
associated than the short-term associations. They conclude, similarly to
Kloog et al., that long-term associations have more impact than
short-term. As this study was carried out retrospectively the collection
of data and population definitions were not ideal. In particular, the
way that the black smoke exposure was estimated using monitoring
stations for short term, but modelled for long-term, introduces
uncertainty. Also for comparing their associations, they used the
general population for short-term and the cohort for long-term,
therefore differences in these groups could cause bias in their results.

COMEAP published a report in 2010 titled *’The Mortality Effects of
Long-Term Exposure to Particulate Air Pollution in the United Kingdom’*.
From reviewing evidence of published studies they too concluded that
long-term exposure was a more important factor in contributing to
mortality than short-term exposure
(@CommitteeontheMedicalEffectsofAirPollution2010). Finally, bringing
together many data-sets and conclusions from other studies, @Stieb2002
produced a number of forest plots for different pollutants demonstrating
the effect of short-term exposure. The NO$_{x}$ plot is shown in figure
\[fig:short\_term\_no2\_meta\] which shows a consistent finding of an
increase in mortality as NO$_{2}$ increases.

![ Percent excess mortality from @Stieb2002<span
data-label="fig:short_term_no2_meta"></span>](short_term_no2_meta)

To summarise, the research that has been conducted so far, with the data
available, suggests that the effects of long-term pollution are a more
important determinant of poor health than that of short-term exposure.
However the studies are limited and lack detail sometimes in areas which
potentially alter their conclusions. Kloog et al. and COMEAP looked at
PM$_{2.5}$ (though perhaps this is sensible given the discussion in
Section \[subsec:anoverview\]), and Beverland et al. only considered
black smoke. There were limitations of spatial and temporal resolution
of the pollution data and subsequent linkage in the three of the
studies. Often large data-sets cannot consider all important details.
Definitions of how long ’long-term’ and how short ’short-term’ are could
also be an influence on conclusions drawn. Short-term is taken to be
around 3 days exposure, but this could be missing hyper-short term
exposure, such as a subject spending an hour cycling down a busy road
and perhaps triggering hospital admission for breathing difficulties.
With better air quality data and more in-depth better information on
where people spend their time (and therefore are exposed) it might be
possible to overcome these issues, however the data to do this does not
presently exist for epidemiologists to use.

It is important to consider both the rapid effects of air pollution
exposure e.g. pathways without hours of exposure **and** the chronic
effects of sustained exposure (@Brook2010).

It seems that short-term, hyper-short-term and long-term exposure are
all important in understanding the health effects of air pollution. Thus
studies that are able to consider large population exposures (but with
information on individuals available), and on these varying time–scales
are required. This issue of improving linkage between air quality and
exposure has evolved over time. The next section reviews the various
methods that have been used in the past to link air quality and exposure
time – what I term ’static exposure studies’. Studies that make use of
various different metrics for air quality, but do not take into account
the movement of people. Dynamic exposure studies are then described –
those that take this movement and other factors into account.

Static exposure & health studies {#sec:staticexposurehealth}
--------------------------------

The methods of estimating individual and population level exposure to
air pollutants (and from this data the impact on health) has evolved
over time; better data has become available, computer modelling has
become more complex (facilitated by more powerful computers), and more
accurate methods have been employed. This section of this transfer
report takes an overview of the methods of exposure assessment that I
term static exposure studies. That is, the subjects to which the air
pollution is being attributed do not move between environments and their
exposure is calculated at their residential address. In addition the
temporal and spatial granularity of the air quality data is often, in my
opinion, of low or insufficient quality.

I split this into categories as follows:

-   Large area exposure

-   Monitoring stations

-   Proximity to roads

-   Dispersion modelling

-   Land-use regression

Due to a large literature base, these sections draw on key texts for
each section as examples of the approach being described. By considering
the issues and problems with these studies, my proposal is that they
over-simplify exposure in different ways, and justify the use of dynamic
exposure models (which are explained and then thoroughly reviewed in
Section \[sec:dynamicexposurehealth\]).

### Large area exposure) {#subsec:largearea}

Large area exposure studies are studies which, as the title suggests,
attribute exposure to subjects over a large spatial area. This might be
at the level of a city, a county or even continent. The well known
Global Burden of Disease studies (discussed briefly in
\[subsec:longtermvshortterm\]) are classic examples of this approach.
For the global burden of PM$_{2.5}$ (chosen due to it’s strong links in
the literature with poor health) in 1990 and 2005 an annual average
layer of air pollution was modelled for the entire world at 0.1^o^ x
0.1^o^ spatial resolution (shown in figure \[fig:globalburdenpm25\]).

![Estimated 2005 annual average PM$_{2.5}$ concentrations (ug/m^3^)<span
data-label="fig:globalburdenpm25"></span>](global_burden_pm25)

To generate this worldwide layer of PM$_{2.5}$ satellite derived
observations of Aerosol Optical Depth (AOD) were used (@Brauer2012).
This is not a particularly common approach for exposure assessment as
using satellite data for pollutant estimation and exposure as a method
is still being refined, and the spatial resolution is low, however it
was used in this case due to it’s ability to generate data for such a
large area.

The resulting health conclusions arrived at are discussed in Section
\[subsec:longtermvshortterm\], however there are acknowledged
limitations with the methods employed. Firstly, ambient concentrations
are assigned to the entire population of the grid cells i.e. no
micro-environmental modelling is undertaken to allow for the time that
people spend indoors, or indeed in any other environment. Secondly,
modelling at 0.1^o^ x 0.1^o^ resolution means much of the spatial
variation in concentrations is lost, or perhaps it’s better to say it’s
not adequately modelled in the first place. To elucidate further, we
know from Section \[subsec:urbanenvironments\] that people are
concentrated in urban environments rather than distributed evenly across
countries and continents, and that within these environments levels of
pollution are higher than outside of them, primarily due to emissions
from combustion engines (Section \[subsec:trafficpollution\]). Taking
London as an example, the city is approximately 50 km wide, yet a degree
of longitude is approximately 113 km wide, meaning that (in this
oversimplification example) the exposure attributed to someone living in
the middle of the Kent downs where are much fewer sources of PM$_{2.5}$
is the same as someone living in the centre of London, where the sources
of PM$_{2.5}$ are more frequent. The authors argue that as the
population data is of a similar scale this does not matter so much,
however by not being able to take account of these urban environments,
or at least not at an adequate scale, exposure may be incorrectly
attributed. Though to what degree it is hard to know. Thirdly, the lack
of temporal resolution to the air quality layer is a problem. As was
briefly explored in the sections on traffic-generated pollution
(\[subsec:trafficpollution\]) and metereology (\[subsec:meteorology\]),
air quality varies by hour, week, days, months, seasons etc. Generating
annual average concentrations loses this detail and simply calculates
averages for the whole year.

Studies of similar scales include @Silva2013 who combined 14 atmospheric
chemistry and meteorological models to attribute annual average
PM$_{2.5}$ exposure to the worlds population on a scale of 0.5^o^ by
0.5^o^ degrees resolution, and @Boldo2011 who modelled PM$_{2.5}$ at a
resolution of 18 km by 18 km (figure \[fig:gspain\_pm25\_grid\]) to
cover the country of Spain (both approaches then used
concentration-response functions to estimate mortality).

![Grid squares used for PM$_{2.5}$ exposure from @Boldo2011<span
data-label="fig:gspain_pm25_grid"></span>](spain_pm25_grid)

It should be stated at this point, that the criticism levelled at these
papers is symptomatic of many studies discussed in the next sections of
this report. However doing so is often a little unfair. The researchers
were in most cases doing the best work that technology, available data,
or indeed time or known methods allowed them to do. As we move into
reviewing further literature in Section \[sec:dynamicexposurehealth\]
(Dynamic Exposure & Health), my premis is that methods can now be
improved beyond this position.

### Monitoring stations {#subsec:monitoringstation}

The term ’air quality monitoring station’ or simply ’monitoring
station’, within the context of this field of research, typically refers
to a static cabin or large metal box which houses a number of
instruments that measure various air pollutants and meteorological
conditions at specific time intervals. A typical station is shown in
Figure \[fig:monitoring\_station\].

![A typical monitoring station (image taken from
[www.londonair.org.uk](www.londonair.org.uk))<span
data-label="fig:monitoring_station"></span>](monitoring_station)

In the UK monitoring stations were first set-up after the introduction
of the Clean Air Act in 1956 but the numbers of stations, locations, the
accuracy, time resolution and numbers of pollutants that are monitored
has changed incrementally since then. The last change in organisation of
the sites came in 1998, when the then Department of Environment
established the ’Automatic Urban and Rural Network’ (AURN), bringing
together many sub-networks to form the most comprehensive automatic
national monitoring network in the country, made up of 127 sites
(@DEFRA2011a). The data collected from the sites is used for policy and
legislative purposes, including reporting to the European Union, however
as the data are routinely and automatically collected over long time
periods the outputs have been a popular data-set for researchers too.
@Mayer1999 presents a good example of the temporal variability of data
that can be collected such as NO, NO2, O3 and NOx at an urban air
quality station in Stuttgart, Germany (and then subsequently linked to
populations for estimating exposure).

One of the most highly cited research articles where data of this nature
was used is in the paper ’*An association between air pollution and
mortality in six US cities*’ by @Dockery1993. In this study air
pollution exposure was linked to 8111 adults over six cities in the U.S.
from six monitoring sites (one in each city) in the form of annual
average concentrations. Each station measured PM$_{2.5}$ and PM$_{10}$
(PM$_{15}$ before 1984). After adjusting for smoking and other risk
factors, significant associations between air pollution and mortality
were found. Studies that also used monitoring stations in a similar way
include @Atkinson2010 who did a time-series analysis of London hospital
admissions using data from a background monitoring site (PM$_{10}$,
PM$_{2.5}$, PM$_{10-2.5}$) and found associations between various
pollutants and health outcomes outcomes such as daily mortality
(particular for cardiovascular), and then @Samoli2005 who as part of the
APHEA Multicity Project took monitoring site data for 22 European cities
and found links with mortality from the air quality data of these
stations.

The main question with this type of approach however is that of temporal
and spatial variability and resolution - can the data from a monitoring
site be an appropriate measure of exposure for someone who, in many
examples, may live miles from the monitoring site. Certainly the studies
that have tried to address whether this is appropriate or not have found
poor correlation between the two. @Cyrys2008 found that ’*the use of a
single monitoring station in long-term epidemiological studies must be
insufficient to attribute accurate exposure levels of PNC to all study
subjects*’ i.e. it might work for some people but not for all. Although
it’s worth noting here that they believe that the monitoring stations do
an accurate job of reflecting temporal variability, particularly for
ultra-fine particles. Just not the spatial variability. @Goldstein197747
broadly agree, they found in New York that ’*the procedure of using one
aerometric station to represent the daily fluctuations of air pollution
throughout the large metropolitan area of New York City risks the use of
an unreliable or invalid measure of the short term variation in air
pollution’*. So their message is slightly different, in that their
research did not assesses how valid it would be as a measure of
long-term variation, but they did conclude it is poor for short-term
(whether long-term or short-term is more important a measure was
discussed in Section \[subsec:longtermvshortterm\]).

Whether this mis-classification of exposure to pollutants varies between
pollutants and height from the ground was considered by @Restrepo2004
who took data from three different monitoring stations (15m above
ground) and compared it to data from a van which contained similar
equipment but which was parked at three different locations and with the
equipment inlets at 4m above ground.The stations showed good agreement
between themselves, but not with the ground-level (van) data. PM$_{2.5}$
was closest matched, for ozone the ground level concentrations were
generally lower, and for NO$_{2}$ the concentrations at ground level
were over twice as high as those at the monitoring stations.

To conclude, using data from monitoring stations as a proxy for
estimating exposure for large populations over many kilometres does not
seem to accurately reflect individuals exposure. The studies above
tended to look at the correlation between a subjects residential address
and the monitoring station, meaning that there is then the additional
complicating factor that people do not spend all of their time at this
place. Whilst the data is certainly easily obtained and the temporal
resolution makes it attractive for time-series studies, the lack of
spatial variability and lack of any micro-environmental modelling or
understanding of where subjects actually spend their time mean that this
approach is simplistic.

### Proximity to roads {#subsec:proximitytoroads}

The use of subject’s address data, and then a calculation of the number
of roads (and sometimes traffic density on those roads), is another
proxy measure of exposure that has been used to consider exposure to air
quality and links between this and poor health. One of the most
highly-cited papers in this area is by @Gauderman2007 who specifically
looked at whether living near to major roadways in California had an
impact on lung-function growth of children between the ages of 10 and
18. In the study 3677 children were regularly monitored for 8 years and
yearly lung-function tests were completed, which were then considered
alongside their home address and the distance to the nearest freeway.
They found that proximity to freeway traffic is associated with
substantial deficits in lung function (which became less pronounced the
further the child lived from the freeway). Similar studies include those
by @Janssen2001 and @Rose2009, although both do not go as far as to make
conclusions of health outcomes, they focus on the method of exposure
estimation by road density. A wider-review of studies in this area was
also undertaken by @HPotHEoT-RA2010 in the HEI report ’*Traffic-related
air pollution: a critical review of the literature on emissions,
exposure, and health effects*’.

The presumption of this type of study is that the air quality the
subjects are exposed to during their normal day-to-day activities is
strongly correlated to the air quality at their home, and that the air
quality at their home is strongly correlated with the number of freeways
within certain buffer distances of their home. These studies also mostly
look at annual average traffic flows or similar, and therefore seem to
not take account of variation in road use and therefore pollution
levels. Indeed, combining these issues only exasperates the uncertainty,
for example road flows are mostly higher during the morning and evening
rush-hours, when children between 10 and 18 are likely on their way to
school or already at school i.e. not at home as the model presumes. They
also don’t consider the time a subject is not at their home, or any
other microenvironment.

### Dispersion modelling {#subsec:dispersionmodelling}

To better bridge this distance between where exposure is occurring
(often presumed to be the subject’s household) and where the exposure is
being estimated or measured (such as a monitoring site) modelling can be
used to create maps or layers of varying resolution for the area that
the study or subjects are based in.

Dispersion modelling is a way of simulating the movement of emissions
through the atmosphere using mathematical equations with input variables
such as wind speed, wind direction, air temperature and street geography
(@EnvironmentalProtectionAgency2008). By doing this at different scales
(country-wide, city-wide, individual streets) it is possible to create
pollution maps and to then associate the pollutant values at those
places with the people who live there - estimating their exposure - and
possibly linking this to health effects if data is available.

@Maroko2012, in studying environmental justice in New York City (USA)
compared the differences between using a proximity analysis technique
(similar to \[subsec:proximitytoroads\] - proximity to roads) and a
dispersion model of PM$_{2.5}$. Their hypothesis was that minority
populations were more likely to be located in areas of poor air quality,
and that proximity analysis may under-represent this problem. Figure
\[fig:maroko\_stack\_proximity\] shows tax-lots within 1/4 mile of the
point stacks upon which the initial exposure analysis (and subsequent
assessment of over-representation of minorities) was completed.

![Proximity analysis to PM$_{2.5}$ point sources from @Maroko2012<span
data-label="fig:maroko_stack_proximity"></span>](maroko_stack_proximity)

Figure \[fig:maroko\_dispersion\_map\] then shows the same area, but now
using dispersion modelling.

![PM$_{2.5}$ estimates from dispersion modelling allocated to tax–lots
from @Maroko2012<span
data-label="fig:maroko_dispersion_map"></span>](maroko_dispersion_map)

As is clear to see, the dispersion modelling approach provides a greater
degree of spatial detail and clarity. Using this second approach to
exposure assessment, they were able to identify that Latino groups in
the Bronx and Brooklyn were being disproportionally exposed to poor air
quality than other ethnic groups, which was not evident from the
proximity analysis approach. Also using a dispersion model but in
London, @Tonne2010 calculated the London annual average concentrations
for 2001 and 2005 (pre and post the Congestion Charging Scheme
(<http://www.tfl.gov.uk/modes/driving/congestion-charge>)) for NO$_{2}$
and PM$_{10}$ on a 20m x 20m grid, and then using this aggregated the
data to Ward level (shown in Figure \[fig:tonne\_wards\_dispersion\]).

![Map of NO$_{x}$ change at Ward level between 2001 and 2005, based on
dispersion modelling (@Tonne2010)<span
data-label="fig:tonne_wards_dispersion"></span>](tonne_wards_dispersion)

The association between this change in concentration and respiratory
hospital admissions was then calculated, although no conclusive relation
was found.

Both of these studies (@Maroko2012 and @Tonne2010) seem to be a step in
the right direction for improving estimates of exposure as they provide
air quality data at a higher resolution than some of the studies that
were considered earlier, however there are still concerns that they
mis-classify exposure by not adequately taking account of temporal
variation in air quality, where the subjects are actually spending their
time, and the micro-environmental aspects.

### Land-use regression {#subsec:landuseregression}

Land-use regression maps of air quality, and linking the concentrations
from these maps to subjects for health studies, was first undertaken as
part of the ’Small Area Variations In Air quality and Health’ (SAVIAH)
study by @Briggs1997. Land-use regression models combine monitoring of
air pollution at a small number of locations, and then develop models
using predictor variables normally obtained through GIS data. The model
is then applied to un-sampled locations in the study area, and
concentration values generated using the characteristics of the
location.

A review of the use of LUR for outdoor air concentration values was
published by @Hoek2008 in Atmospheric Environment. They found that the
models could be applied successfully to model annual mean concentrations
in various geographical locations, for various pollutants including
NO$_{2}$, NO$_{x}$, PM$_{2.5}$ and VOCs, and that the method was better
than other geo-statistical methods such as kriging and dispersion
methods. However the models were not as effective when a finer temporal
scale was required or desired. @Dons2013 in a study accross Flanders
(Belgium) used aetholometers with a 5-minute resolution to measure black
carbon at 63 locations continuously for seven days. When they compared
these values to a LUR model they concluded similarly to Hoek, that
existing LUR models were not ideal for considering exposure of their
local population due to the lack of temporal variability. Although they
did go on to develop hourly LUR models with some success (The R$^{2}$of
their models varied between 0.15 and 0.79 according to the time day and
the variables used as input).

In the recent paper by @Pedersen2013, which is an output from the ESCAPE
(European Study of Cohorts for Air Pollution Effects) project, a LUR
model was generated for 12 European countries, and then temporally
adjusted using levels from nearby monitoring sites (to try to introduce
temporal variability that has been lacking from this approach as noted
by Hoek). Links between the concentrations at the maternal address of
74,178 women, who had singleton deliveries between 1994 and 2011, were
then examined. They found that a 5 $\mu \text{g m}^{-3}$ increase in
concentration of PM$_{2.5}$ during pregnancy was associated with an
increased risk of low birthweight at term (adjusted odds ratio of 1.18).

To summarise, LUR models are being used for health studies, and
conclusions between health outcomes and air pollution are being drawn,
however there are similar problems with the exposure estimates/linkages
as in many of the other previous static exposure estimate methods,
including that there is no estimation of the amount of time that people
spend away from their home and the pollutants and concentrations that
they are exposed too i.e. how much time did the mothers in the ESCAPE
study actually spent at their maternal address? Studies using LUR also
don’t tend to conduct any micro-environmental modelling of the time that
subjects spend indoors or in transport, and the methods behind the
temporal variability might be considered further i.e. is using a nearby
monitoring stations daily variation sufficient to reflect the daily
variation at the address point.

A more generic problem with using LUR models for exposure assessment is
that LUR models need monitoring to be conducted in the area that the
model is proposed to be used in. Setting up a dense enough network of
monitors to provide accurate model results is expensive and
time–consuming.

In the review article *“Spatio-temporal epidemiology: principles and
opportunities”* @Meliker2011 discusses how exposure estimation is a
rapidly changing field, and how GIS, computer power and ’big data’ have
started to overcome issues that spatio-temporal epidemiology has often
struggled with (as discussed in the preceeding sections). One of the
areas that is still somewhat lacking however, despite exposure
assessments producing estimates of contaminants changing through time
and space (e.g. temporal air pollution maps), however, is human mobility
– it is seldom incorporated. They conclude their review with the
following quote, which sets the scene for the following section titled
’Dynamic exposure and health studies’, those being studies that
explicitly seek to take account of the movement of individuals through
different environments, and their exposure in those environments:

> “We expect exposure assessments to increasingly incorporate space–time
> dynamics in particular mobility and environmental contaminants, such
> that it becomes commonplace in the near future” @Meliker2011

\[chap:background\]

Dynamic exposure & health studies {#sec:dynamicexposurehealth}
---------------------------------

As we have seen from the studies reviewed thus far, effectively and
accurately quantifying human exposure to air pollution for
epidemiological studies is fraught with problems. Studies that look at
sufficiently large numbers of subjects find it difficult to accurately
assign exposure to them, as the subjects spend time in many different
environments, travel between these micro-environments, and often the air
quality data is of insufficient temporal or spatial quality. The
epidemiological studies that are based upon these estimations may
therefore be making incorrect health conclusions. @Brauer2002 measured
personal exposure compared to monitoring site exposure for PM$_{2.5}$
for 16 subjects in Canada and found that in 13 of the 16 subjects the
measured ambient concentrations were under-representing exposure.
Ashmore’s review of literature focusing on children’s exposure found
that their exposure is closely related to concentrations in the home, at
school, and in transport — differing significantly from adults
concentrations and general outdoor concentrations such as those at
monitoring sites (@Ashmore2009). Although the links between health and
air pollution are fairly clear as was discussed in section
\[sec:healtheffects\], it might be that this miss-classification error
is biasing calculated risk factors, regressions or coefficients towards
the null value (no association) i.e. once exposure classification is
improved the relative risks of increases in exposure to pollutants may
be greater than previously thought (@Armstrong1990).

### Personal Monitoring {#sec:personal_monitoring}

To gain further insight and a more realistic understanding of
individual-level exposure to pollutants, personal monitoring methods can
be used. Studies that are described as personal monitoring vary in their
use of equipment, and measuring of different pollutants, but generally
describe studies that attach portable pollutant monitoring devices to a
person or persons, who then go about their normal lives while the
devices collect data. The devices normally collect data at short time
intervals i.e. one minute, and are combined with location devices like a
GPS. By using these devices the researchers are able to collect data on
concentrations at the the places that the subjects are in, and see how
levels vary between them. Using such equipment, individual level
objective direct measures of exposure can therefore be collected. These
studies are considered the “gold standard” of exposure assessment
(@Ashworth2013 , @DeNazelle2008) .

@Steinle2013 conducted a review of personal exposure studies of this
style. They discussed the difference between traditional exposure
assessments using fixed air quality network sites, single
micro-environments and static populations, compared to the new
developments in sensor technology that have enabled researchers to
directly monitor pollutants while people move through varying
concentration fields and activity space. They summarised the personal
exposure approach with Figure \[fig:traditional\_personal\_monitoring\]
which visualises the joining of location data and concentration data

![Conceptual model illustrating the traditional approach for the
assessment of personal exposure to air pollution @Steinle2013<span
data-label="fig:traditional_personal_monitoring"></span>](traditional_personal_monitoring)

This area of air pollution exposure research has, as Steinle notes,
accelerated in popularity recently due to the advancements in related
technology. The devices have become cheaper, lighter and easier to
collect/process data from. The ubiquitous nature of smart phones has
also contributed in terms of it becoming the norm for people to carry
about technology and have elements of their lives tracked. Indeed,
studies such as @DeNazelle2013 have used data from smartphone
applications that are not designed for exposure assessment, but which
offer a rich data set (the application CalFit for estimating peoples
movements style i.e. running, cycling, walking). A search of the terms
“’personal exposure’ AND ’air pollution’” in PubMed, summarised by year
of publication in Figure \[fig:personal\_exposure\_publications\] nicely
illustrates the rise in these studies (whilst acknowledging that there
are mitigating circumstances such as the popularity of the field as a
whole, and numbers of journals in this area etc.):

![Numbers of publications about personal exposure to air pollution in
PubMed<span
data-label="fig:personal_exposure_publications"></span>](personal_exposure_publications)

@Dons2011 in Belgium completed personal monitoring campaigns for 16
couples. One member of each couple was identified as a full-time worker,
and one as a homemaker. Each couple’s personal exposure was measured
over 24 hours by them carrying a micro-aetholometer and a PDA device
which contained a GPS chip for location and time-activity diary for
contextual information (which was subsequently completed by the
individuals). Very different patterns of exposure were observed between
the couples, despite them living in the same place i.e. two individuals
living in the same location would have the same exposure according to
some of the approaches in the static studies section. Exposure differed
by upto 30% between couples, with exposure during transport being
identified by Dons et al. as the most important factor in this
discrepancy. In a similar study in Italy @Buonanno2014 conducted
personal monitoring on 24 non-smoking couples and measured their
exposure to UFPs using a Phillips NanoTracer which included a GPS for
tracking. In this study the couples were all male-female, the male being
a full-time worker and the female being a homemaker. Given the emphasis
that Dons found on transport, it was slightly surprising that this study
found women to have higher exposure than men. This is attributed to the
emissions from cooking stoves in the home and that the particles stay in
the environment for prolonged periods. Although the two studies actually
consider different metrics (black carbon compared to ultrafine
particles) so the comparison is hard to make directly – there are not
many sources of black carbon in the home, but transport is a major
source outside of the home. In a further study @Broich2011 used a GPS,
and GRIMM 1.09 monitor to measure particulate matter (in various sizes)
for sixteen people over 24 hours each in Italy.

![Average exposure over 24 hours to PM$_{10}$, comparing personal
monitoring, background monitoring stations and traffic monitoring
stations from @Broich2011<span
data-label="fig:broich_exposures"></span>](broich_exposures)

As can be seen from Figure \[fig:broich\_exposures\], there were large
variations between the personal exposure concentrations and
concentrations at nearby fixed-sites. The averages over 24 hour between
participants varied from 27 to 322 $\mu \text{g m}^{-3}$, and like Dons
they found that exposure levels were heavily dependant on the travel
participants undertook, and the micro-environments they spent time in.
Broich argues this is why direct personal exposure monitoring is needed
for accurate exposure assessments in the future, rather than proxy
methods. Although for epidemiological studies that might relate patterns
from monitoring sites to patterns in incidents of poor health, this is
not necessarily the case. If the personal exposure results for instance
are always say 40% than the monitoring site, then the monitoring sites
would still capture the trends and might be able to relate these trends
to hospital admissions.

While these types of study are excellent at providing hyper-local and
time resolved personal exposure data, they tend to struggle to provide
data that is useful for considering alongside the harmful health effects
of poor air quality. The studies give results for individuals or small
groups of people which are not able to be used alongside health-related
epidemiological data such as hospital admission records or incidences of
asthma in a certain region. As @Chaix2013 argues, improving measuring of
exposure to environmental conditions by accounting for the movement of
individuals is critical, however by using such small numbers I suggest
that there is a danger of over-generalisation between exposure and
health. Measuring 20 school-children’s exposure to PM$_{10}$ for
example, and then making wide-ranging assumptions about the health
effects of PM$_{10}$ on school children seems simplistic. It might be
that those 20 children are in the top 5% of exposures and not a good
representation of school children in general.

@Steinle2013 argues that the trend in the area of calculating accurate
exposure estimates for large populations is for the development of the
personal monitoring approach, i.e. distribution of low-cost, accurate
GPS devices combined with accurate low-cost unobtrusive portable air
quality measurement devices. In this way the amount of data collected
through this ’gold-standard’ method would be adequate to allow
extrapolation to large groups of people. A model of their theoretical
approach is shown in \[fig:steinle\_full\_assessment\].

![Conceptual model for the assessment of individual and population-wide
exposure to air pollution including effects from @Steinle2013<span
data-label="fig:steinle_full_assessment"></span>](steinle_full_assessment)

@Minguillon2012 agree with them after their study on pregnant women’s
exposure in Barcelona in 2009. They compared measurements from the
subjects household balcony, inside their house, and the data from
personal monitoring equipment. They found mean concentrations for
PM$_{2.5}$ of 20 ug/m^3^, 24 ug/m^3^, and 27 ug/m^3^, for outdoor,
indoor, and personal samples respectively. They concluded that it was
important to rely on personal exposure measurements for epidemiological
studies.

However I would argue that this is not currently practical. The
technology for this type of study is not yet available, as the authors
themselves note. In addition, even if it were, the distribution of the
devices and automated collection of the data on a large enough scale
seems difficult to achieve. Though this may change with the advent of
smaller and lower-cost measurement technology.

What the personal monitoring studies in this section do show the
researchers in this field is where subjects spend alot of their time
(indoors), and where they are exposed to the highest levels of air
pollutants (travel). They emphasise how important these factors are in
understanding the true exposure of individuals compared to current or
traditional methods. Developing a modelling approach, with focus on the
accuracy of the model in these places, to improve exposure estimates,
seems to be a more attainable goal. Understanding exposure in these
micro-environments for input to such models is therefore crucial.

### Infiltration {#sec:infiltration}

Infiltration refers to the diffusion of outdoor air into the environment
inside a building. The amount of outdoor air that moves inside depends
on ventilation, air conditioning and on the indoor––outdoor temperature
gradient. EU guidelines now state that buildings must be insulated to
save energy, however as buildings allow less air to be exchanged with
the outside environment, the indoor concentration of pollutants can
increase if there are significant indoor sources. Thus increasing the
air-tightness of buildings can have negative impacts on health
(@Gens2014). This is offset of course by meaning that harmful outdoor
pollutants will also not as easily infiltrate into the indoor
environment. The study of indoor air has become an inherent part of
modern exposure research (@Steinle2013). Figure \[fig:infiltration\]
below illustrates how outdoor air infiltrates buildings and mixes with
indoor air.

![Infiltration from @Chen2011<span
data-label="fig:infiltration"></span>](infiltration)

Indoor pollutants have many sources such as cigarette smoke, cooking,
heating and cleaning products. They can substantially modify a persons
exposure and characterisation of them is difficult and complex. To
consider the exposure of subjects to air pollution over
days/weeks/months, indoor air is an important consideration. A better
understanding of the factors influencing infiltration and indoor sources
will improve exposure assessment methods and contribute to reduced
exposure miss-classification in epidemiological studies (@Colbeck2010a,
@MacNeill2012). Although this of course depends on the metrics that are
being used for considering the health effects – to better understand the
effects of traffic pollution on health for example indoor sources would
not be needed and the infiltration to the indoor environment would be
key. But for a model which considered exposure to all pollutants then
all sources would be needed.

To try and better understand the differences between indoor and outdoor
pollution, as part of a large cohort exposure study in Ontario (Canada),
@MacNeill2012 measured concentrations of PM$_{2.5}$, UFPs, black carbon
and humidity both inside properties and in the back garden of properties
at the same time for a period of two weeks.

As is common with this type of study the results were discussed in terms
of indoor/outdoor ratios, otherwise known as I/O ratios. This is the
concentrations inside the property, divided by the concentrations
outside of the property. So if a property had a PM$_{2.5}$ I/O ratio of
0.5 at 10am and the concentrations outside were 10 ug/m^3^, then this
would mean that concentrations inside the property were 5 ug/m^3^ (10
multiplied by 0.5).

The median daily estimates found in the study ranged from 0.26 to 0.36
across seasons for PM$_{2.5}$, with the ranges typically related to
window-opening behaviours, air conditioning, meteorological variables,
home age, use of electrostatic precipitators and stand alone air
cleaners. The determinants of indoor source concentrations were related
to cooking, candle use, supplemental heating, cleaning, and number of
people in the home. Expanding on the last of the variables noted by
MacNeill, @Colbeck2010a comments on the close relationship between
indoor concentrations and human activities, stating that humans are
responsible for their own “personal cloud”, that is, exposure to
airborne particles resulting from their activities (e.g. occupation,
hobbies) or physical activities (e.g. jogging, vacuum cleaning).

@Challoner2014 looked at PM$_{2.5}$ and NO$_{2}$ concentrations in ten
different city centre buildings in Dublin. They found PM$_{2.5}$ I/O
ratios close to 1 (similar to outside) for the ten commercial buildings,
however after studying the temporal variation in these levels they
suggested that, in general, indoor sources and/or re-suspension of
PM$_{2.5}$ seemed to have a more significant impact compared to
variations in outside air quality. @Kearney2014 measured PM continuously
for seven consecutive days in 74 Edmonton (Canada) homes in 2010.
Simultaneous measurements of outdoor (near-home) and ambient (at a
central site) concentrations were also measured. As with the studies
above, they found considerable variability ranging from 0.10 to 0.92 in
winter and from 0.31 to 0.99 in summer.

Given the number of variables that seem to contribute to variation in
indoor air quality, making broad-sweeping assumptions to create inputs
to epidemiological models seems difficult. However WHO calculated in
2005 that people spend 89% of their time indoors, @Lai2004a found a
figure of (89.5%) from their research, and @Schweizer2007 found similar
- 20.66 hours (86%). Efforts to better quantify this exposure
relationship and create as accurate a picture of exposure are therefore
needed. The variability in I/O ratios within and between homes may cause
substantial exposure miss-classification compared to only using ambient
measurements (@Kearney2014).

So a dynamic exposure model, one that is going to take account of
different micro-environments where people spend their time, needs to
attempt to model concentrations indoors (given putting time resolved
monitors in every indoors environment is impractical). The following
paragraphs therefore consider a few recent studies that have attempted
to do that.

MESA-Air stands for the “Multi-Ethnic Study of Atherosclerosis and Air
Pollution” and is a large research project based in Washington State
(USA). The project is designed to examine the relationship between air
pollution exposures and the progression of cardiovascular disease
(@Allen2012). To do this they needed to be able to quantify exposure to
indoor air, and so they aimed to develop models to predict I/O ratios
for around 6,000 homes. They did this by collecting 526 two-week, paired
indoor–-outdoor PM$_{2.5}$ filter samples from a subset of homes in
their study. Taking account of specific weather and seasonal variables,
as well as using information from questionnaires, they made a regression
model which predicts I/O ratios (mean of 0.62, SD of 0.21) using easily
obtained variables. This is an important study in the cross–over area of
air pollution exposure assessment, indoor–outdoor concentrations and
epidemiology as it was the first study with a large number of subjects
to incorporate variation in residential exposure into exposure
assessment. The effect that this new information made to epidemiological
study of the prospective cohort is not yet published.

@Hussein2014 took a similar approach with their exposure model, however
they also consider dose (Figure \[fig:husseiniomodel\]). This model
calculates exposure by considering infiltration into the indoor
environment, indoor sources, building types, building sizes and the
time-activity pattern of the individual. A mass-balance model is used
for the indoor exposure which not only models infiltration but the
indoor sources.

![Exposure model incorporating indoor exposure by @Hussein2014<span
data-label="fig:husseiniomodel"></span>](hussein_io_model)

They illustrate the use of their model by calculating exposure and dose
for 24 hours for a test individual, showing the inputs needed for the
various model parameters. The advantage of this approach is that it
starts to show a way for indoor exposure to be properly included in an
exposure assessment model, however as the authors acknowledge the detail
and amount of data required for input to the model (i.e. volumes of
buildings, detail of indoor sources, time activity patterns) is not yet
available on a large enough scale and robust enough, to expand this
approach to large numbers of subjects. Lacking exposure modelling of the
journeys between the environments would also seem to be necessary for
this model to give a more holistic picture of exposure. However the
inclusion of the calculation of dose in such a dynamic exposure model is
clearly a positive step. Dose being a quantification of the amount of
pollutant(s) that the subject actually breathes in and makes its way to
the airways and lungs of the subject as per the discussion and diagram
of @Brook2010 in Section \[subsec:anoverview\].

@Johnson2012 also developed a dynamic exposure model that considered
pollutant levels in indoor micro-environments, specifically focusing on
NO$_{2}$. They had 60 school-children (ages 12–13) wear personal
monitors for two days and while doing so keep a detailed log of their
activity, location, and the activity of others around them that might
generate pollutants e.g. their parents cooking or smoking at home. Using
the time-activity patterns, modelled outdoor concentrations and I/O
ratios from literature (e.g. 2 for cars/buses, 0.5 for school) they
calculated exposure for the set of children and compared it to their
monitored data. The results of this “microenvironmental exposure model”
(MEEM) agreed well with the personal exposure measured by the children,
however a great deal of contextual information was required to come to
such close agreement and it is hard to see how model the could be
applied to a much larger cohort of subjects without the same level of
detail. Given this the transferability of the study to other subjects
and areas is not useful as a tool in and of itself, but it does show
that good agreement can be obtained between modelling and monitoring of
PE given sufficient data. An additional problem with using this approach
elsewhere is that the air quality data was based on a LUR model, which
as was discussed in Section \[subsec:landuseregression\], requires
extensive monitoring to be accurate and often lacks temporal variation
(though this study sought to overcome that issue with diurnal profiles
from a local monitoring station).

As discussed earlier in this research, the health effects of exposure to
air pollution are not yet fully understood. Studies have tended to
assign exposure based on outdoor concentrations at the subjects
residence using long term average concentrations or on monitoring
stations concentrations for studies of short-term exposure. Given this
it is no surprise that the negative effects of poor indoor air quality
on a populations health are similarly not yet understood (although the
approaches discussed in the preceding paragraphs are contributing to a
better of understanding of overall exposure i.e. they are starting to
contribute to understanding and quantifying exposure indoors, while
other studies that quantify exposure in other environments are
completing the picture).

@Gens2014 is one of very few studies that have attempted to consider how
indoor air is affecting health, specifically looking at the increase in
air-tightness of modern EU buildings. The assessment was based on
modelling exposure to fine particles originating from both outdoor and
indoor air, including environmental tobacco smoke. Exposure response
relationships were derived and the results showed an increase of adverse
health effects in all considered countries (ranging for health effects
from 0.4% in Czech Republic to 11.8% in Greece for 100% insulated
buildings) due to an accumulation of particles indoors. Unsurprisingly
considering only the effects of outdoor air led to a decrease of adverse
health effects. Although the conclusions drawn for this response
relationship seem in doubt as the odds-ratios have been applied to the
combination of indoor and outdoor air, when they are only designed and
suitable to be used for outdoor air.

@Chen2011 conducted a review of modelling approaches on the relationship
between indoor and outdoor particles and summarised that I/O ratios vary
considerably due to the difference in size-dependent indoor particle
emission rates, the geometry of the cracks in building envelopes, and
the air exchange rates. Concluding that due to this it is difficult to
draw uniform conclusions on I/O ratios, as they vary so much between
building types.

The studies discussed in this section show that indoor air varies
considerably from outdoor air, and that research is ongoing as to how to
effectively model this as one part of a holistic exposure model that
also incorporates other environments such as different transport modes.
Research so far has identified that indoor air quality is effected by
outdoor concentrations, the number and activity of people inside the
building (particularly apparent for particulates), the ventilation
systems (or lack of), indoor sources, meteorology and the air-tightness
of the building.

### Transport {#sec:transport}

As briefly discussed in Section \[subsubsec:invehicle\], air quality
when in transport can differ greatly from general ambient
concentrations. Table \[tab:adams\_transport\_means\] in the same
section showed measurements in different transport modes by @Adams2001,
which were different (mostly higher) than that of the ambient
concentrations at the time. The recent review of pollutant
concentrations in different transport modes by @Karanasiou2014 provides
an excellent resource for considering the breadth of the differences
between transport and ambient concentrations. Reflecting transport
exposure as part of a persons total exposure is important as it is
thought that individuals gain a significant contribution of their daily
exposure from it, despite the small percentage of time of the day that
is spent doing it. For black carbon exposure @Dons2011 found subjects
spent 6-8% of their day in transport, where they accumulated 21% of
their exposure, and approximately 30% of inhaled dose. Failing to
account for these important exposure events will lead (does lead) to
miss-classification in epidemiological studies.

As can be seen from the data compiled by the Office for National
Statistics in figure \[fig:transportprofiles\], the three most popular
modes of transport in England and Wales over the last half a century
have been car and van, followed by bus and coach, and finally rail.

![Transport profiling from ONS census data<span
data-label="fig:transportprofiles"></span>](transport_profiles)

Studies which have sought to measure and understand exposure in these
transport modes are now discussed, with the addition of cycling.
Cyclists are included as a special case due to their increased
inhalation rates and proximity to road traffic which may significantly
increases their exposure and dose of poor air quality.

#### Bus and Coach travel {#sec:bus_and_coach}

The review by @Karanasiou2014 found that pollutant concentrations inside
bus and coach vehicles vary greatly by country, reflecting the variation
in weather conditions, vehicle types, ambient concentrations, and fuel
sources of the surrounding fleet. In one study concentrations in the
Netherlands were found to be higher in diesel vehicles than on the same
routes with electric fuelled buses, suggesting that the own vehicles
exhaust impacts on the exposure inside the vehicle. Though this could
also be reflective of the permeability of the vehicle or the number of
times the doors were opened or closed. In comparison to ambient
concentrations, a Paris study found an I/O ratio of 1.3, and in the
Netherlands a similar ratio of 1.43 . Typical PM$_{2.5}$ concentrations
inside the vehicle across the eleven studies reviewed were in the range
35 – 69 ug/m^3^.

A study not covered in the review is by @Song2009 in York, U.K.
Measurements were made simultaneously by placing an optical particle
monitor in the middle of a bus in York over a period of three days in
May 2007, and comparing the data recorded to measurements from an
identical monitor on the bonnet of a car which followed the bus
(@Song2009). Measurements were completed during the morning and evening
rush hours over 24 hours. Figure \[fig:bus\_ratios\] shows how the
relationship between out-bus and in-bus concentrations was positively
linear, with a higher gradient as particle size increases (the column
labelled ’Dependent variable’ is the size fraction of PM). Also having
the windows closed on the bus actually led to increased concentrations
compared with having the windows open, suggested to be due to
re-suspension by passenger activity.

![Comparison of mean in-bus and out-bus particle concentrations<span
data-label="fig:bus_ratios"></span>](bus_ratios)

I/O ratios for PM in the size range of 0.75 to 2.0 are 1.53 and 1.55 for
when the windows are open, and 2.16 and 2.48 for when the windows are
closed. There are many other factors which complicate drawing
conclusions between studies, but it is interesting to note how closely
the former of these ratios agree with the studies in Paris and the
Netherlands for similar PM fractions discussed at the start of this
section.

Using data from this study, the @Song2009 paper concludes by compiling a
model for estimating the indoor/outdoor ratio of different PM fractions.
The model showed broad agreement with measured particle concentrations
inside buses and demonstrated that re-suspension by passenger activities
and deposition to the surface of the passengers had significant effects
on the concentrations.

The health implications of bus and coach exposure is not discussed by
the Karanasiou review or the Song paper as a separate issue or metric.
Being able to include exposure by this transport mode in a large-scale
exposure assessment exercise is required for exposure mis-classification
to be reduced. Though is further complicated, as in many of the sections
of this report, by the composition of particles i.e. particles from
different sources have more or less harmful effects on health.

#### Car travel {#sec:car}

Studies on exposure to subjects while travelling inside cars are more
frequent than those of other transport modes. From reviewing studies of
car exposure @Karanasiou2014 found that typical European PM
concentrations inside passenger cars were in the range of 36–-76 ug/m^3^
for PM$_{10}$ and 22-–85 ug/m^3^ for PM$_{2.5}$. Whether the car itself
was powered by diesel or gasoline was found in most studies to make
little difference to PM, PNC and BC. However @Jalava2012 found that
pollutant concentrations inside the car depended highly on the type of
fuel used.

A separate review by @Nasir2009 found that exposure to particulate
matter in the car micro-environment is largely dependent on traffic
congestion, road network layout, vehicle design/condition and ambient
concentrations.

@Dons2013 studied slightly different parameters, or ways of considering
the parameters, of BC exposure in vehicles in Figure
\[fig:road\_type\_exposure\]. It was noted that in urban areas
concentrations inside vehicles are higher compared to exposure in more
rural areas; with the same holding true for highways versus local roads
for motorists – suggesting that the surrounding fleet and traffic speeds
affect in-vehicle concentrations.

![External factors affecting in-car BC exposure (ng/m^3^) from
@Dons2013. Vehicles are the dark boxplots. Walking are light<span
data-label="fig:road_type_exposure"></span>](road_type_exposure)

By taking all the factors studied by Dons, a regression model of car I/O
ratios for black carbon was developed and is shown in Figure
\[fig:transport\_io\_ratios\].

![Transport indoor/outdoor ratios from @Dons2013<span
data-label="fig:transport_io_ratios"></span>](transport_io_ratios)

However the use of these ratios are not well tested and do not always
agree with other ratios from the literature. It seems that given other
known influences such as ventilation systems that a deterministic model
based only on time of day and road type is too simplistic. For example
in the two studies by Gulliver and Briggs (@Gulliver2004 and
@Gulliver2007) PM concentrations inside the cars in Northampton and
Leicester respectively were roughly 30% higher than concentrations at
nearby monitoring stations (comparisons with immediately outside the
vehicles were not available). A table of mean concentrations from the
2004 paper are shown below in Table \[tab:gulliver\_vehicle\_means2\])
for reference.

  -------------------------------------------------------------- --
  **<span>PM Fraction</span> & **<span>Mean (ug/m^-3^)</span>\   
  PM$_{10}$ & 43.16\                                             
  PM$_{2.5}$ & 15.54\                                            
  ****                                                           
  -------------------------------------------------------------- --

  : In-vehicle PM from @Gulliver2004<span
  data-label="tab:gulliver_vehicle_means2"></span>

The main conclusions from studies on in-car expose were that exposure is
significantly influenced by traffic intensity and ambient air pollutant
concentrations (being strongly linked to each other) as well as the
choice of ventilation used inside the vehicles themselves. Is is a
complicated picture. Estimating the exposure of individuals from the
time they spend in the car micro-environment should take into account
the factors outlined to try and as accurately as possible reflect
reality.

#### Bicycle {#sec:bicycle}

Between 2001 and 2011 the number of people living in London that cycled
to work more than doubled from 77,000 in 2001 to 155,000 in 2011. There
were also increases in other large UK cities, for example Brighton
(increasing by 109% between 2001 and 2011), Bristol (94%), Manchester
(83%), Newcastle (81%) and Sheffield (80%)
(@OfficeforNationalStatistics2014).

This rise in cycling is due to a number of factors including local
authorities building more cycle-friendly infrastructures, the promotion
of cycling as a way to keep fit (supported by the Governments
cycle-to-work discount scheme), and the public being more
environmentally-aware.

In some cities bicycle sharing systems have also contributed to this
rise in cycling journeys. In London the Barclays Cycle Hire scheme was
launched in July 2010 with around 5,000 bikes distributed around London
at specially installed docking stations (@TransportforLondon2014). Users
were (and still are) able to hire the cycles using a pre-purchased
membership key, or by inserting their bank card. Similar systems operate
in many places around the World. Although there is no definitive list as
new schemes are opening all the time, in 2013 @larsen2013bike estimated
there were more than 500 cities in 49 countries hosting advanced
bike-sharing programs, with a combined fleet of over 500,000 bicycles
(compared to 213 schemes in 2008 (@Wikipedia2014)).

The @Karanasiou2014 review looked at cycling exposure in 20 European
studies, mostly in the UK, the Netherlands and Holland. They found mean
exposure values for PM$_{2.5}$ of 29–72 ug/m^3^ and for PM$_{10}$ of
37-62 ug/m^3^. The London studies (@Kaur2005 and @Adams2001) found that
cycling exposure to PM$_{2.5}$ depended on the route taken, finding
busier routes with more traffic increased cyclist’s exposure.
@Ragettli2013 found similar - bicycle travel along main streets between
home and work place contributed 21% and 5% to total daily UFP exposure
in winter and summer, respectively, and that exposure could be reduced
by half if main roads were avoided. At odds with this, the Netherlands
study (@Zuurbier2010) found no difference between routes. Though upon
reading their methods in more detail it seems that there was overlap
between the low-traffic and high-traffic routes, and also that their
low-traffic routes were only low on car traffic and they were frequently
used by mopeds.

The main variable affecting cycling exposure was found to be location
i.e. a cyclist riding in the middle of a busy road will be exposed to
higher concentrations than on a separate cycle-lane. Given cyclists are
not enclosed in any sort of vehicle this direct relationship between air
pollutant concentration and exposure is to be expected.

With regard to the negative health impacts of exposure to poor air
quality while cycling, unlike other transport modes such as car and bus,
a number of recent studies have tried to quantify this. @Woodcock2014
used Barclays Cycle Hire data from TfL to model the health impact for
578,607 people in London, mostly (78%) aged between 14 and 45. They used
the bike hire usage data, modelled the journeys between the docking
stations, and combined this with general travel data, data on physical
activity levels and road collisions, and finally 20 m by 20 m grids of
annual average PM$_{2.5}$ data. Cycling exposure was compared to
replacing those journeys with walking or public transport. They found
that the population benefits from the cycle hire scheme substantially
outweighed the negatives, with a net change of minus 72 DALYs among men
and minus 15 for women. London’s bicycle sharing system has positive
health impacts overall, but these benefits are clearer for men than for
women and for older users than for younger users. In a very similar
study in Barcelona @Rojas-Rueda2013 did a Health Impact Assessment (HIA)
for eight different scenarios where the the population of the
Metropolitan area of Barcelona (3.2 million) was presumed to change
transport mode from car to either cycling or public transport. Traffic
incidents, physical activity and air pollution exposure were then
estimated for each scenario and the relative health changes estimated in
DALYs. For the scenarios where 20% and 40% of car trips where replaced
by cycling trips, there were changes of minus 138 and minus 275 DALYs.
Rather than doing a health impact assessment @Nwokoro2012, back in
London, did more direct assessments of cycling exposure by looking at
the amount of black material in the airways of non-cyclists and
cyclists. They found that commuting to work by bicycle in London was
associated with increased long-term inhaled dosage of black carbon,
however the relationship is difficult to properly quantify. This was
because cycling typically takes longer for those people to cycle to work
than, for example, a train journey. Route choice further complicates
matters as although being away from highly polluted routes would lower
exposure, the increased journey time may increase total exposure over
the trip. Figure \[fig:airway\_cycling\_blackcarbon\] shows the
increased macrophage carbon measured in cyclists compared to
non-cyclists.

![Airway macrophage carbon in cyclists and non-cyclists from
@Nwokoro2012<span
data-label="fig:airway_cycling_blackcarbon"></span>](airway_cycling_blackcarbon)

In summary cycling exposure is very dependant on the routes taken by
cyclists, and their position on the road. The negative health effects of
cycling have been explored in comparative transport mode studies showing
that when taken alongside other variables such as risk of injury, there
are overall health benefits to cycling. However there is a lack of
studies that take cycling exposure as part of a larger daily exposure
health assessment alongside other transport modes and indoor
micro-environments. A number of the health studies are encouraging in
that they do model sufficiently large groups of people to draw
epidemiological conclusions (578,607 people in London and 1.3m in
Barcelona) but only the transport aspect of these people’s exposure is
explored. Also there are other parts of their models which could be
improved, for example both studies only used annual average air quality
concentrations, and the journeys were not temporally resolved.

#### Train travel {#sec:train}

In their 2014 review of commuter exposure in Europe (which has been
referred to extensively in this section on transport exposure),
Karanasiou did not review train travel – focusing on cycling, cars and
underground subway systems. However @Nasir2009, @Colbeck2010a,
@Dons2011, @Ragettli2013 and @Knibbs2011 all have elements of their
publications which take train travel exposure as a separate travel
microenvironment.

@Nasir2009 and @Colbeck2010a found concentrations during peak journey
times in air-conditioned carriages were 44 ug/m^3^, 14 ug/m^3^ and 12
ug/m^3^ for PM$_{10}$, PM$_{2.5}$ and PM$_{1}$ respectively, but that
during off-peak times, concentrations were about half this.
Concentrations were lower again in non-air conditioned carriages. They
draw the conclusions that particulate levels inside trains are strongly
influenced by the numbers of passengers and that their movement and
presence causes particle re-suspension in the air. Peak travel times
(when there are more passengers) coincide with high particulate
concentrations. Figure \[fig:nasirtrainstops\] shows the effect of train
stops on concentrations in the train cabin. It would have been
interesting to see simultaneous outdoor concentrations to see how they
varied to the indoor concentrations and more detailed information on
passenger numbers. Theoretically it might be that outdoor concentrations
contribute to a certain percentage of the indoor concentrations, which
are then varied positively or negatively by passenger numbers.

![Train stops and train exposure from @Nasir2009<span
data-label="fig:nasirtrainstops"></span>](nasir_train_stops.png)

@Knibbs2011 reviewed articles to look at the difference between diesel
and electric powered trains, and based on the limited data available
found that the power source of the rail vehicle appears to strongly
affect ultra-fine particle concentrations (diesel powered trains have
higher in-train particle concentrations than electric trains). High
concentrations of particles are not just confined to the trains
themselves either, @Ragettli2013 found ultrafine particle levels within
train stations twice as high as suburban background and nearby
residential streets in Basel, Switzerland.

In summary the variables that drive exposure levels on trains are the
fuel source, the number of stops that the train makes, the influx of
passengers at those stops, and the concentrations outside of trains. No
research on specifically how train travel exposure affects the health of
passengers was found. As with the other transport modes reviewed thus
far, a model which incorporates these variables alongside other
micro-environments should be able to develop a more accurate picture of
human exposure over days or weeks.

#### Underground subway systems {#sec:underground}

In many worldwide cities the population travel between locations using
metro systems. The London Underground was the first such system to open
in 1863, but other cities have followed suit including New York, Paris,
Seoul, Beijing, Berlin and Madrid. The size of each system varies in
length, as does the power source, depth and speed of the trains, however
the systems tend to be almost unanimously popular as a choice of travel
mode for commuters (not surprising given that they were built for this
purpose).

The @Karanasiou2014 review found 12 studies that had measured exposure
in such subway systems. Particulate levels were found to be generally
highest on the platforms of the system (compared to inside the vehicles
themselves). Exposure was also higher during peak travel times. Mean
PM$_{10}$ levels were in the range of 103 to 1030 ug/m^3^, and
PM$_{2.5}$ levels between 59 and 375 ug/m^3^. The highest levels were
found in London, Stockholm and Rome. Variation between and within
systems was thought to be explained by different brakes types, air-con
or natural ventilation systems, different types of rails, and the
numbers of trains. A summary table from the review is shown in Figure
\[fig:pm\_tube\_summary\].

Focusing specifically on London, the work of Adams in the late 1990’s
/early 2000’s is oft cited and well known in this area of transport
exposure due to it being the first comprehensive particulate matter
multi–mode transport user exposure assessment study in the UK. They
measured a large number of journeys (465) across many modes (bicycle,
car, bus, tube) over different periods in July 1999 (’summer’) and
February 2000 (’winter’) finding mean PM$_{2.5}$ values for the London
Underground of 247.2 ug//m^3^ in summer and 157.3 ug/m^3^ for the winter
(or an overall mean of 202 ug/m^3^ as shown in the Figure
\[fig:pm\_tube\_summary\]). Shortly after this publication another paper
was published by the same group which tried to apportion the PM to
sources (@Adams2001a). They concluded that most of the PM came from
metals from sources such as braking, tyre wear and the tracks. A high
iron content was found. They noted that the PM in the underground should
not be directly compared with PM above ground as it’s composition is
very different. In terms of the actual levels measured in the first
paper however extrapolation of these measures across the entire
underground network including within cabins and on the platforms is not
proven. @Seaton2005 found large variations between PM in cabins and on
the platforms themselves. In their research (funded by the London
Underground to look at at occupational health exposure to PM$_{2.5}$)
platform concentrations were 270–-480 ug/m^3^ and cab concentrations
were 130–-200 ug/m^3^. Similarly high to the other studies of the London
Underground. This paper also found a high iron content in the PM – 67%.
Seaton summarises however that although the concentrations are high,
they are unlikely to represent a health risk to workers. @Loxham2013
disagrees. Their research was not included in the Karanasiou review
(presumably due to it being published only shortly before the review).
They looked at PM$_{10}$, PM$_{2.5}$ and PM$_{0.1}$ composition finding
similar to the other research, that the particles are mostly from
interaction between wheels, rails, and brakes with a high iron content.
However they suggest that the potential health effects of exposure to
the ultrafine fraction of underground PM warrants further investigation,
as a consequence of its greater surface area/volume ratio and high metal
content.

To summarise, exposure and the related health effects of poor air
quality in subway systems has not been extensively studied in most
cities. Certainly not to the point where researchers can confidently say
what levels of particulate matter the public are being exposed to at
different times of the day, in different places of the underground, and
what composition that has (and how toxic to the body it is or is not).
What can be fairly confidently said however is that concentrations in
most systems, and certainly in London, are higher than street-level,
in-vehicle, cycling, buses or most other studied transport
micro-environments. Although the composition is very different and may
be more or less toxic. Due to the number of people that use this
transport mode daily (1.171 billion journeys annually in London
according to @TransportforLondon2014a) for extended periods of time,
understanding and being able to quantify exposure during it is necessary
to better estimate people’s daily exposure across a range of
micro-environments. Certainly they vary hugely from concentrations
measured at a subjects place of residence, as is used in the static
exposure studies considered earlier.

Studies of exposure while people are in-transit between locations using
transport modes such as cars, trains, subway systems and bicycles shows
a large range of concentration values. There are few exposure studies
which account for these variations (alongside outdoor indoor
infiltration) to provide exposure data on the daily lives of large
numbers of subjects suitable for epidemiological analysis. Until more
expansive exposure studies that follow large groups of people of varying
time-activity patterns are completed, the ability to discern the range
of commute-time’s specific contribution to total exposure is constrained
(@Knibbs2011).

### Dynamic and Hybrid exposure models {#sec:dynamic_hybrid_models}

Section \[sec:staticexposurehealth\] (Static exposure studies) concluded
that taking fixed-point or fixed-area measurements, often at one point
in space or time, was insufficient to accurately quantify human exposure
to air pollution. Exposures varies in space and time due to an
individual’s activities, different sources having different effects, the
time of day, and the location of the subject. For example the poor air
an individual is exposed too will vary depending on how long they spend
in public transport each day (and which mode of transport), how long
they spend at home or at work, and the concentrations within those
environments themselves (@Ozkaynak2013). Therefore models which look at
these different environments, and their effects on health, are needed.

Section \[sec:personal\_monitoring\] (Personal monitoring) showed how
personal monitoring devices could be used to better understand
individual-level exposures, but was concluded by discussing how
collecting enough accurate data using this method for large health
studies is very difficult to do. The following Sections of
\[sec:infiltration\] (Infiltration) and \[sec:transport\] (Transport)
looked at models which enable researchers to quantify exposure to
populations in those specific environments (and some discussed the
health implications of those environments), but they were not
’joined-up’ so as to enable estimates of total daily exposure alongside
other micro-environments and times.

This section looks at a the small number of research studies which use
models to attempt to be more holistic and include all micro-environments
as well as temporal and spatial variation in their inputs. This is a new
area of research which is not yet established, and the nomenclature is
thus still under development, but a trend towards discussion of ’hybrid’
or ’dynamic’ models has emerged. ’Hybrid’ being a description of a
combining of multiple approaches of exposure, such as time-activity
diaries combined with modelled air quality data and micro-environmental
modelling of transport modes, and ’Dynamic’ reflecting the improved
temporal scale of data available, for both the populations movement and
the variation in pollutant concentrations. Also discussed below are
publications by @Ozkaynak2013, @Meliker2011 and @Baxter2013 who have
attempted to advance this field by reviewing recent approaches and
suggesting key areas which need to be improved, or by identifying what
exactly these new types of model should do.

@Kousa2002 was a very early attempt at this sort of model. They
evaluated the temporal and spatial exposure of the population of the
Helsinki Metropolitan Area in different micro-environments (home,
workplace, traffic and other). For time-activity data (15 min
resolution) they used 435 diaries from the EXPOLIS study, and they then
linked this to modelled NO$_{2}$ data from a dispersion model which
outputted grids of 500 m by 500 m for the greater Helsinki area, and 50
m by 50 m for the city centre. For indoor environments they took an I/O
ratio of 0.76, including when the subjects were recorded as being in
transport. The GIS technology and techniques used in this model would
have been quite advanced at the time, and therefore modelling the
exposure of 8000 people was impressive. Specifically, modelling the air
quality with hourly variation and at 50 m resolution. However in-light
of contemporary work this model is simplistic in a number of ways
including: only describing four micro-environments (and using the same
I/O ratio for all of them), that the time-activity diaries are only at
15 minute intervals, that the data only includes people of ages 25-55,
and that the days modelled are only ’working days’ i.e. Monday to
Friday. It would also have been interesting to see what difference using
this model made on exposure assessment i.e. a comparison with other
methods - such as the research by @Dhondt2012. Based in Belgium they
compared ’static’ and ’dynamic’ models for exposure for NO$_{2}$ and
O$_{3}$, modelling their subject’s behaviours and location based on 8800
activity diaries into 12 km location ’zones’, then scaling this to give
results relevant for 5 million people. For their air quality inputs
hourly concentrations were modelled using a variety of nested models,
giving higher resolution nearer to roads and in urban areas, and lower
resolution outside of these areas. Time in transport in each zone was
also included. Maps showing the exposure over 24 hours accumulated by
each zone are shown in Figures \[fig:dhondt\_static\_exposure\] and
\[fig:dhondt\_dynamic\_exposure\].

![NO$_{2}$ static exposure results from @Dhondt2012<span
data-label="fig:dhondt_static_exposure"></span>](dhondt_static_exposure)

![NO$_{2}$ dynamic exposure results from @Dhondt2012<span
data-label="fig:dhondt_dynamic_exposure"></span>](dhondt_dynamic_exposure)

By using this approach to exposure modelling they found NO$_{2}$ was
being underestimated by on average 1.2%, and that O$_{3}$ was being
overestimated by 0.8%. These do not seem particularly large differences,
however as an input to a large cohort of people in a long time-series
analysis they could significantly change the health conclusions reached.
These regional averages also mask much of the variation in individuals,
where differences of upto 12% can be found depending on age group,
gender and place of residence. Having studied their methods however, the
reliability of the results must be questioned. This research is
certainly moving towards the type of dynamic exposure model described
earlier, as their modelled air quality and time-activity data is of a
high quality, however there is no micro-environmental modelling at all,
which given the time that we know people spend indoors must be a large
source of uncertainty in the results. As a final note on this
publication, the method of grouping the exposure results into
time-activity zones rather than by residence of the subject was
interesting and informative, showing the areas where greatest
miss-classification occurred in a simple, visual and effective manner.

The following year @DeNazelle2013 did a similar study for 7 days of 36
subjects in Barcelona, however in addition to using time-activity
diaries as an input, the subjects also had software installed on a
provided smartphone, which made use of the accelerometer and GPS
hardware to record their location and infer their type of activity i.e.
walking, sitting. After (substantial) processing of this data, the
subjects location and activity was linked with an hourly resolved 5 km x
5 km dispersion model grid of NO$_{2}$, and personal exposure was then
calculated – including adjustments for six different micro-environments
(indoor, bike, bus and tram, car and taxi, metro and train, motorcycle).
This study was truly a hybrid study, in that data was collected using
personal monitoring (smartphones) but then joined with modelled data and
further processing of the data. De Nazelle then calculated the time that
the subjects spent in different micro-environments, and compared this to
the percentage of their daily exposure in the same micro-environment
categories (Figure \[fig:time\_no2\_activity\_spaces\]).

![Time and NO$_{2}$ in activity spaces from @DeNazelle2013<span
data-label="fig:time_no2_activity_spaces"></span>](time_no2_activity_spaces)

From the results we can see that in terms of minutes, the subjects spent
most of their time at home and at work (left), supporting the use of
static exposure models where these fixed locations are used, however
when the actual exposure is considered the influence of the time spent
in those locations diminishes (right) and ’Others’ and ’Intransit’
become important (supporting the conclusions of @Dons2011 discussed in
Section \[sec:transport\]). The study concludes by discussing how these
techniques that have been used could be developed and used for much
larger populations (presumably as an input to health studies eventually)
, however I would suggest that the amount of post-processing of the
geographical data, the manual translation of activity diaries to
compliment the smartphone data, and the extra battery packs that needed
adding to the phones makes this seem less likely. To upscale this to
many hundreds or thousands of subjects would need a tremendous amount of
coordination and equipment which would need to be distributed, collected
and monitored. The data processing would also seem to be difficult
without large human resources, and linking to health outcomes would mean
the study needed to be designed so that the participants are a
statistical representation of the wider study area population, or the
numbers surveyed so close to the actual study population that they give
a good representation anyway.

In the same year @Gerharz2013 used similar methods to examine the
exposure of 10 people in Munster, Germany. They used activity diaries
alongside GPS data again, defined micro-environments using this
contextual information of ’home’, ’work’, ’other indoor’,
’transportation’, and ’outdoor’, joined this to an hourly average
PM$_{10}$ dispersion model for the air quality input, and then used I/O
ratios for the micro-environments e.g. 1.34 for cars and 2.49 for public
transport. The model is summarised in Figure
\[fig:gerharz\_exposure\_diagram\] below.

![The exposure process from @Gerharz2013 <span
data-label="fig:gerharz_exposure_diagram"></span>](gerharz_exposure_diagram)

The major difference between this and the work of De Nazelle and Dons is
that they also gave the subjects personal air quality monitoring data in
the form of a Grimm Aerosol Spectrometer. The spectrometer provided
information about particle numbers and mass in 32-size classes with a
high temporal resolution of 6 seconds. Using this data they were able to
evaluate the effectiveness of their modelled exposure by comparing it to
real data. The Pearson’s correlation coefficient between the mean of
modelled and measured data is shown in Figure \[fig:gerharz\_results\].

![Pearson’s correlation coefficient between the mean of modelled and
measured data from @Gerharz2013 <span
data-label="fig:gerharz_results"></span>](gerharz_results)

There is clearly some variation between the results, with P7 (person 7)
being particularly strong, and P10 being particularly weak. Gerharz
notes in their discussion that “Generally, the correlation between model
average and measurements is high” which I would dispute with the mean of
the correlations coming out as 0.44 i.e. Medium correlation. However
they are correct in saying that it does at least “clearly outperforms
the baseline approach of using urban background measurements as proxy
for the personal exposure”. Though in terms of using this modelled
approach for larger groups of people, the post-processing of GPS data
and activity-diaries, general data cleaning, and data linkage would seem
to make this an unlikely approach.

@Ozkaynak2013, @Meliker2011 and @Baxter2013 have written reviews /
position statements on this type of research. Ozkaynak’s diagram below
(Figure \[fig:ozkaynak\_exposure\_diagram\]) is useful for summarising
how the complexity of exposure modelling has advanced, and with the
complexity the inputs required.

![The evolution of exposure assessment from @Ozkaynak2013<span
data-label="fig:ozkaynak_exposure_diagram"></span>](ozkaynak_exposure_diagram)

They concluded that these new approaches can help refine the
significance of air pollution health outcomes, but that this will depend
on study-specific characteristics, including epidemiological study
design (e.g., time-series vs cohort), the form of of health outcome
being considered (e.g., long-term or short-term), which pollutants, and
the role of pollutant and building specific indoor infiltration and
human activity patterns. @Meliker2011 offers a slightly different but
useful perspective, by identifying what they consider are the five most
important domains to the development of improved spatio-temporal
epidemiological models, namely:

1.  spatio-temporal epidemiologic theory

2.  selection of appropriate spatial scale of analysis

3.  choice of spatial/spatio-temporal method for pattern identification

4.  individual-level exposure assessment in epidemiologic studies

5.  assessment and consideration of locational and attribute uncertainty

@Baxter2013 summarises this emerging area of research well by explaining
how, when compared with the use of central-site monitoring data (or
other fixed-location methods) the enhanced spatial (and temporal)
resolution of air quality or exposure models can impact on resultant
health effect estimates, especially for pollutants derived from local
sources such as traffic. They recommend that future research develops
pollutant-specific infiltration data, improves existing data on human
time-activity patterns and exposure to local sources, in order to
enhance human exposure modelling estimates. Also that these new
approaches are compared with existing approaches to exposure estimation
to better characterise estimates in chronic health studies.

Figure \[fig:james\_exposure\_diagram\] is proposed as a conceptual
model of a hybrid/dynamic exposure model. It should have highly temporal
and spatially resolved air quality inputs which consider both indoor and
outdoor sources (including regional and local source for the latter), it
should be able to model infiltration rates for different modes of
transport and building types, it should reflect the multiple
micro-environments that people spend their time in (and take account of
the temporal resolution of these) and finally it should (for linkage
through to epidemiological end-points) be able to consider different
breathing rates to quantify exposure and dose for multiple pollutants.

![A conceptual dynamic exposure model<span
data-label="fig:james_exposure_diagram"></span>](james_exposure_diagram)

\[chap:dynamic\_exposure\]

My research aims
================

The hypothesis of this research is that “*Individual and population
level air pollutant exposure can be estimated using time-activity
surveys, GIS and routing tools, coupled with high resolution
spatio-temporal air quality models, facilitating a greater understanding
of the health impacts of air pollution and how public health risks can
be reduced*”. The research objectives that will need to be met to
achieve this goal are as follows:

1.  Reconstruct the time-space activity of London’s population

2.  Link modelled air quality, estimate exposure, and compare with
    traditional methods

3.  Refine the models estimates of London Underground exposure

4.  Evaluate the results

In more detail I will model the movements of the subject’s in the TfL
dataset, reconstruct their days on a fine temporal and spatial scale,
and then use this data in conjunction with a novel exposure model. The
air quality input will be a CMAQ-UK model (see Section
\[sec:cmaq\_urban\]). The results of the model will allow interrogation
of exposure by individual, or grouped by various demographics, which
will enable epidemiologists to increase their understanding of exposure
miss-classification. Results will be compared to static modelling
approaches of the type discussed in Section
\[sec:staticexposurehealth\]. The model will then be improved with
personal monitoring on the London Underground, before an introduction to
validation/testing of the results will be undertaken. The objectives
will be met by following the work-plan below.

Modelling Londoners movements
-----------------------------

**Aim:** Create a model of Londoners daily movements based upon the
London Transport Demand Survey\
**Objectives:**

1.  Explore the LTDS dataset and identify key data

2.  Move data from local Microsoft Access database into a PostgreSQL +
    PostGIS database

3.  Clean data

4.  Complete modal specific routing (interrogation, querying, storage)
    between origin and destinations

5.  Quality check the routing results and do further data cleaning

6.  Create and populate final dynamic location database table

7.  Analysis of results

Dynamic exposure modelling
--------------------------

**Aim:** Model exposure to PM$_{2.5}$ and NO$_{2}$ of the LTDS-X
subjects, and compare with traditional exposure models\
**Objectives:**

1.  Link data to CMAQ-Urban UK, incorporate IO ratios and
    micro-environmental factors (The ’LHEM’)

2.  Create a postcode comparison dataset

3.  Create a address-point comparison dataset

4.  Create a monitoring sites comparison dataset

5.  Analysis and comparison of results

Thus far, these aims cover the first half of the conceptual exposure
model shown in Figure \[fig:james\_exposure\_diagram\] (which is
repeated here for convenience).

![A conceptual dynamic exposure model<span
data-label="fig:james_exposure_diagram_v2"></span>](james_exposure_diagram_v2)

The following two sections will be on improving exposure estimates, and
evaluating them respectively.

Exposure to on the London Underground
-------------------------------------

**Aim:** Create an exposure model for PM$_{2.5}$ on the London
Underground\
**Objectives:**

1.  Measure PM$_{2.5}$ across the London Underground network

2.  Transcribe a time-location diary and link to pollutant data

3.  Import other relevant data

4.  Analyse data to characterise PM$_{2.5}$ on the London Underground

Evaluating dynamic exposure models
----------------------------------

**Aim:** Develop an understanding of methods to evaluate predictions of
exposure from hybrid models\
**Objectives:**

1.  Establish how to use mobile monitoring equipment to replicate static
    monitoring site datasets (which are used for air quality
    model evaluation)

2.  Collect mobile monitoring data representative of a journey from the
    LHEM

3.  Model exposure of this journey using LHEM methodology

4.  Compare the monitored and modelled exposures

Methods and tools
=================

In order to complete the work required to meet my aims, various pieces
of software, tools and methods are required. Appropriate
(spatially-aware) relational database software is chosen and developed
to store and interrogate data, geographic information systems (GIS) and
statistical programming languages are used for spatial data processing
and visualisation, statistics software for drawing conclusions from the
processed data, and other libraries and software such as CartoDB and D3
javacript are used for visualisation and presentation.

The tools and methods I have chose to work with are all ’Free and Open
Source Software’ (FOSS) rather than commercial or proprietary
alternatives. This choice is mostly made as it is the author’s belief
that community developed software leads to easier and readily available
customisation by the users, more frequent improvements and updates,
quicker fixing of bugs, and generally more choice of tools for specific
purposes. Given the software is free, this also negates any concerns
about licenses and installation on multiple computers which might be a
concern with other software.

The London Transport Demand Survey {#sec:the_ltds}
----------------------------------

The London Transport Demand Survey (The LTDS) is a survey of households
in the London area, covering all London Borough’s as well as the area
between those Borough’s and the M25. It is organised by the Strategic
Analysis section of the Planning department at Transport for London. The
survey is a continuous rolling–survey, started in 2005–2006, and still
happening today. Around 8,000 households per year are surveyed.

The LTDS survey captures information on households, the people that live
in those households, the trips that those people make in the day, and
the vehicles that they use/own. Everyone in the house that is surveyed
answers the questionnaires, except for children under the age of 5.
There are three questionnaire that each person completes. The household
questionnaire gives details on household structure and includes
demographic information such as income, housing tenure and vehicle
ownership. The individual questionnaire contains demographic information
about the individuals in the household, working status, frequency of
transport mode use, driving licenses, public transport tickets held and
similar. The third questionnaire is the trip sheet which is key to how I
will use this data. It captures data on all the trips made on the
designated travel day (which is the same day for all members of the
household). The details include the purpose of the trip, the transport
modes that were used, the time of the day the trip started, the time of
the day the trip ended, and the origin and destination of the trips.
This section of the database can effectively be used to interrogate
exactly where each person was during the 24 hours that the survey covers
(it covers 4am on the day of the survey, to 4am on the following day,
unless that person was in-transit at 4am in which case it continues
until the journey is complete).

The LTDS is designed to enable statistically-robust representative
estimates of travel patterns and demand in London. Results from a single
year are robust enough to analyse at the London-wide level, however
combining three or more years data for analysis is preferable. The
results can be considered at a London-wide level, or disaggregated to
Borough of residence.

The database contains 100 tables of data, however many of these are
’look-up’ tables for data in the main tables. For example the year that
a household was surveyed is stored in a column called ’hyearid’ which
simply contains a number between 5 and 9. These numbers allow linkage to
a table called YEARID\_T where the value 5 can be seen to correspond to
the description of ’2005/2006’. A summary of the key tables and numbers
of records in these tables of the LTDS that I use is shown below in
Table \[tab:ltds\_data\].

  **Year**    **Households**   **People**   **Trips**   **Stages**
  ----------- ---------------- ------------ ----------- ------------
  2005–2006   5,008            11,583       29,797      61,542
  2006–2007   8,005            18,241       47,029      95,930
  2007–2008   7,873            17,926       44,828      91,967
  2008–2009   8,134            18,975       43,076      89,701
  2009–2010   8,290            19,187       43,475      92,121

  : Data contained in the LTDS<span data-label="tab:ltds_data"></span>

Spatial databases {#sec:spatialdatabases}
-----------------

Relational database management systems (RDMS) are a common tool and are
widely used for effectively storing and querying datasets. However to
meet my research aims a RDMS that can store extensive spatial and
temporal data was needed, and these are much less common. Spatially
aware databases have been developed alongside GIS systems and are
traditionally seen as the back-end of a GIS i.e. the place where the
data is stored, while the GIS serves it to the user in a graphical
map-like interface where a user might do analysis. However as
spatially-aware databases have become more common, a disconnect has
occurred and now it is not uncommon to use spatial databases on their
own, and only connect them to a GIS when needed (if at all).

The specifications of what a spatial database should be were published
in the very highly cited (884 at time of writing) 1994 paper *’An
introduction to spatial database systems*’. This paper is so well cited
as it was the first paper to really define what a spatial database was,
namely ’*a database system that offers spatial data types in its data
model and query language and supports spatial data types in its
implementation, providing at least spatial indexing and spatial join
methods*’ (@Guting1994). More recently @ObeHsu201410 defined it as a *a
database that defines special data types for geometric objects and
allows you to store geometric data (usually of a geographic nature) in
regular database tables. It provides special functions and indexes for
querying and manipulating that data using something like Structured
Query Language (SQL)*.

There are a selection of spatial databases which could be used for the
storage of the data required for this work. @Ellul2012 identifies the
three most popular spatial-RDMS as:

-   PostgreSQL (@PostgreSQL2014)

-   Oracle Express (@Oracle2014)

-   MySQL (@MySQL2014)

Of these three, I chose PostgreSQL (with the spatial add-on PostGIS) for
this research due to it’s ease of installation on multiple operating
systems, but more importantly for the greater availability of spatial
SQL queries when compared to the other two systems. That is, more
functions and tools for manipulating spatial data are available in this
system than in the other two alternatives. The range of FOSS desktop GIS
that can easily be connected to PostGIS was also a consideration, as was
(to a lesser degree) my familiarity with their use. Relevant data from
the LTDS (described in Section \[sec:the\_ltds\]) is therefore exported
from the Access database from in which it was provided by TfL, and
uploaded to PostgreSQL instead.

Geographic Information Systems & Science (GIS) {#sec:gis}
----------------------------------------------

FOSS that is specifically used for geographic type data, such as a GIS,
is often referred to as FOSS4G (Free and Open Source Software for
GIS/Geospatial). Different categories of GIS are designed for different
purposes. Each GIS product, whether web-based, desktop client or
otherwise tends to perform a number of functions. @Steiniger2013
summarises the functions as follows:

1.  Data creation

2.  Data editing

3.  Data storage

4.  Data integration

5.  Data querying

6.  Data analysis

7.  Data transformation

8.  Map creation

9.  Data visualisation

It is typical to use a desktop-based client for the majority of the uses
listed above, but as I decided to use the PostgreSQL & PostGIS RDMS (see
section \[sec:spatialdatabases\]) for items 1-8 of the list, the full
functionality of a typical GIS is not required, as the GIS is mostly
just used for map creation after processing of the data within the
database itself. This approach is taken as keeping the data within the
database until it is ready for visualisation leads to quicker processing
and manipulation of the data.

Maps are normally defined either as static maps or dynamic/slippy maps.
A static map being a map that is produced using a GIS Desktop client and
then outputted as a PDF or PNG file or similar, and then perhaps
embedded in a document or presentation. Once produced the data cannot be
edited by the viewer, or the area of the map changed. For this type of
output, I use the Quantum GIS (QGIS) (@qgis2014) project. QGIS is a
user-friendly Open Source GIS, licensed under the GNU General Public
License. It is developed by the Open Source Geospatial Foundation
(OSGeo) and runs on Linux, Unix, Mac OSX, Windows and Android. It
supports numerous vector, raster, and database formats and
functionalities. QGIS is the most popular, comprehensive and stable
FOSS4G software and so was the obvious choice.

Routing {#sec:routing}
-------

One of the innovative aspects of this research is to attempt to quantify
exposure across space and time, rather than in static locations such as
homes and offices. In order for this to be possible, data on the
subjects locations on a fine temporal and spatial resolution is
required. The data provided by TfL and included in the LTDS database
gives geographical and temporal data (as well as mode of travel) on the
start and end locations of journeys that the subjects make, but not the
routes that the journey took. To calculate this missing information
routing databases and Application Programming Interfaces (APIs) can be
used. By doing so routes between two points that respect the
road/cycle/train etc networks will be generated and stored, rather than
perhaps drawing a straight line between the two locations. To use a
routing database network data on the choice of transport mode is
required and then requests need to be made to this network for routes
using SQL queries.

Loading of these network datasets can be accomplished by downloading the
data, for example from OpenStreetMap, and then processing and importing
it into PostgreSQL. Once this is done, the FOSS PostgreSQL add-on
pgRouting is installed and the network data is processed and a topology
of the network is created which is suitable for use with standard SQL
query language. The benefits of this approach are that the network is in
complete control of the user and can be customised at will, for example
road speed-limits can be adjusted (which will affect routes chosen by
the algorithms), a preference for certain road types factored in,
one-way restrictions either ignored or obeyed. There is also a certain
independence to this approach i.e. no reliance on an external provider
keeping the dataset up-to-date or needing a web connection for the
routing to take place (as is required for the API approach discussed
shortly). On the negative side however the network can be difficult and
time-consuming to set-up in the first place, the amount of data to
download and process is large and complex, and a new network is required
for each transport mode that routing will be required for. This is
particularly pertinent for the LTDS as routing on many different
transport modes is required including underground, bus, cycling,
driving, walking and overground/mainline trains - and data to create
networks for each of these modes is difficult to obtain and maintain,
particularly for bus routes and train lines as the data is held by the
relevant companies that operate these services and even when made
publicly available is often not in a suitable format or there is missing
information.

APIs offer an alternative approach to routing. Instead of building and
maintaining network datasets in a users own database/server, the
networks are hosted by other commercial and non-commercial organisations
who allow external users to interact with their data (but not to edit it
or use it in ways which they do not allow). A user, having read the
normally provided documentation, first forms a query string (similar to
a website address but longer and more complex) and then ’sends’ this to
the organisations API service. The API interprets the request that the
user has made, and ’returns’ the data required. In the case of routing,
the request from the user is typically a start-point, end-point,
transport mode and time of travel, and the return from the API is a
route made up of such information as a list of tube stations, or
coordinates, or text instructions. The amount of detail that can be
submitted as a request, and the amount of detail that is provided by the
reply, varies by service. A typical request string to the TfL directions
API is shown in figure \[fig:tfl\_api\_call\].

![A ’call’ to the TfL routing API<span
data-label="fig:tfl_api_call"></span>](tfl_api_call)

This is submitted to the API using software of the users choice, code is
written to parse the data that is returned (made into a format that the
software can read), and then the results are stored and used as the user
wishes.

A search of routing API services was therefore conducted, based upon the
list detailed in the Wikipedia page ’Online Routers’
(<http://wiki.openstreetmap.org/wiki/Routing/online_routers>) as well as
an internet search of the term ’Online routing API’. A review of those
considered suitable for this research is presented in table
\[tab:api\_summary\_table\]. A small handful of other APIs were
dismissed before this stage as they were unsuitable for reasons such as
geography (e.g. only covers Germany) or the API does not expose the
route to the user (e.g. Bing Maps Directions plots the line onto a map,
but does not give the actual data for storage).

<span>@| R<span>3.3cm</span> | R<span>3.2cm</span> | R<span>1.9cm</span>
| R<span>3.2cm</span> | R<span>3.4cm</span> | R<span>4.5cm</span>
|@</span>

**API Service (coverage)** & **Available transport modes** & **Usage
limits** & **Documentation** & **Simplicity of use** & **Customisation
and notes**\

Google Directions API (Worldwide) & Car, pedestrian, cycling, public
transport & 750 per day (without license) & Provided, with many examples
& Very simple to use & Limited. Parameters such as use of tolls,
language of content, shortest or quickest. Public transport modes cannot
be distinguished\

OpenRouteService (Worldwide, but dependant on OSM data quality) & Car,
pedestrian, cycling & Unlimited & In the form of a Wikipedia entry.
Covers some examples but is not comprehensive. & Very simple & Limited.
Road types, languages, zoom levels\

Project OSRM (Open Source Routing Machine) (Worldwide, but dependant on
OSM data quality) & Car & Unlimited & Basic instructions provided on
GitHub & Very simple to use & Limited. Zoom level and output file type\

Transport for London (Greater London) & Underground, overground, bus,
train, tram, boat, cable—car, Docklands Light Railway & Unlimited (once
registered) & Extremely comprehensive manual, but unclear in many places
and often seemingly contradictory & Very difficult. Only simple examples
are provided. So many parameters must be considered for a properly
formed request & Fully customisable in almost every way. However it is
limited to Greater London and does not alert when routes go outside.\

MapQuest (Worldwide) & Car & Unlimited & Basic instructions provided on
GitHub & Very simple to use & Limited. Zoom level and output file type\

Each routing tool has advantages and disadvantages and it is therefore
my expectation that a combination of these will need to be used
depending on the specific use-case e.g. routes using the London
Underground will need to use the TfL API.

CMAQ Urban {#sec:cmaq_urban}
----------

CMAQ (Community Multi-scale Air Quality model) is an ongoing open-source
project, coordinated/led by the U.S. EPA Atmospheric Science Modelling
Division. It is made up of a variety of software packages and processes
for simulating air quality. To quote from their website, “CMAQ combines
current knowledge in atmospheric science and air quality modeling with
multi-processor computing techniques in an open-source framework to
deliver fast, technically sound estimates of ozone, particulates,
toxics, and acid deposition”
(@UnitedStatesEnvironmentalProtectionAgency2014). The three main
components are as follows (@CMASCentre):

1.  A meteorological modelling system for the description of atmospheric
    states and motions

2.  Emission models for man-made and natural emissions that are injected
    into the atmosphere

3.  A chemistry-transport modelling system for simulation of the
    chemical transformation and fate

ADMS-Urban is a separate air quality model which is distinctive as it is
able to model a range of scales, from street to city, taking into
account emissions sources such as traffic, industry and domestic
sources. It incorporates advanced algorithms for the height-dependence
of wind-speed, turbulence and stability to produce predictions. It also
includes information on street canyons and mixing introduced by road
traffic re-suspension
(@CambridgeEnvironmentalResearchCounsultantsCERC2014a).

The Environmental Research Group at King’s College London, where the
research within this report is taking place, have created a version of
ADMS-Urban (KCL-Urban) which gives annual mean air quality predictions
of NO, NO$_{2}$, O$_{3}$, PM$_{10}$ and PM$_{2.5}$ on a regular 20 m x
20 m grid. This is combined with a CMAQ regional scale model to create
CMAQ-Urban which provides predictions of the same pollutants at the same
scale on an hourly temporal resolution. The two models are equipped with
similar capabilities in that KCL-Urban is quick to run and can provide
details of the poor air quality sources at any location within it’s
domain. By combining these two ( CMAQ and KCL-Urban) the result is a
model which is partly deterministic (uses fundamental physics and
chemistry), but provides the same spatial detail as KCL-Urban.
Importantly, it is therefore capable of predicting hourly concentrations
(@Beevers2013). An example of the output of CMAQ-Urban is shown in
Figures \[fig:kcl\_urban\_no2\] and \[fig:cmaq\_uk\].

![Annual mean NO$_{2}$ concentrations in London for the year 2008
predicted onto a regular grid of 20 m x 20 m using the KCLurban
model.<span data-label="fig:kcl_urban_no2"></span>](kcl_urban_no2)

![An illustrative example of the domain covered by CMAQ-Urban<span
data-label="fig:cmaq_uk"></span>](cmaq_uk_v2)

CMAQ-urban was submitted to the UK Model Intercomparison exercise run by
the UK Government department DEFRA (@Carslaw2013), where it performed
well against other urban and regional models with r values of 0.9 for
NO$_{X}$ and NO$_{2}$, and 0.77 for PM$_{2.5}$.

For calculating exposure on a fine temporal (monthly-daily-hourly
averages) and spatial scale (20 m x 20 m grids) the locations and time
of individual subjects will be spatially joined with this model and then
concentrations extracted from it.

Interactive data visualisations and animations {#sec:interactivedata}
----------------------------------------------

When it is desirable to produce maps of a higher quality, perhaps with
interactive elements such as the user being able to click for more
details or zoom in on areas of interest, the static maps that QGIS can
produce as a PDF or similar are not suitable, and different tools are
used to produce slippy maps (maps that can be dragged around using a
mouse or other input device) with overlays of data. Figure
\[fig:clear\_streets\] shows the website
[www.clearstreet.org](www.clearstreet.org) which uses such a ’slippy’
map and allows the viewer to zoom in or out on their area and examine
the snow-plough activity.

![The clear streets website with a ’slippy’ map<span
data-label="fig:clear_streets"></span>](clear_streets)

Previously making maps of this type required specialist GIS and
programming skills, and were almost exclusively made using the Google
Maps API which allows embedding of the Google Maps data on your own
webpage and with your own data added, however there has been a huge
surge in mapping products over the last five to ten years as the
profitability and usage of geographical data has become clear and
commonplace respectively. As the field is rapidly evolving most books or
review articles are out-of-date within weeks of publication, and
therefore the following list of providers of global web-based slippy
maps is taken from Wikipedia (@wiki-maps-2014):

-   OpenStreetMap (<http://www.openstreetmap.org/>)

-   ArcGIS Online (Esri) (<http://www.arcgis.com/home/webmap>)

-   Google Maps <https://www.google.com/maps/>)

-   Bing Maps (<http://www.bing.com/maps/>)

-   MapQuest (<http://www.mapquest.com/>)

-   WikiMapia (<http://wikimapia.org/>)

-   HERE (<http://here.com/>)

-   Mappy (<http://en.mappy.com/>)

-   Yahoo! Maps (<https://maps.yahoo.com/>)

Of this list the most popular and widely used are Google Maps and
OpenStreetMap, however in following with the earlier stated aim of using
FOSS tools, I decided to use OpenStreetMap background data for making
online maps where needed in this research, due to it’s community-made
nature and lack of any restrictions on it’s use or frequency of use.

To be able to use OpenStreetMap data as a background map with overlaps
of data that has been produced during this research, further tools are
needed. The two introduced below are both FOSS4G implementations of the
javascript programming language and I intend to use them both as
required, particularly for the later in this research where I explore
novel visualisations and techniques for exposure and air quality data.

-   CartoDB.com (figure \[fig:cartodb\_screen\]): CartoDB is a
    cloud-based web-mapping, analysis and visualisation service. It
    allows storage of geographical data on their servers which can then
    be visualised on a map and published to a webpage. The advantages of
    CartoDB are the simplicity of producing professional looking maps
    with very little effort, however the amount of data that can be
    stored is limited without having to pay for additional storage.
    Despite this, for ad-hoc visualisations and/or ones that do not
    require a large amount of data to be displayed, this tool is
    very useful. CartoDB.com uses OpenStreetMap data as the background
    for the maps that it serves to the users and can be
    easily customised.

-   Leaflet.js: Leaflet prides itself as a lightweight javascript
    library which displays OpenStreetMap data. By lightweight the
    community developers emphasise the very small amount of coding and
    expertise that is required to produce a map. Adding data and
    creating visualisations on top of the map is then of
    moderate difficulty. The main difference between CartoDB and Leaflet
    is that leaflet does not store the data, the webpage is coded so
    that the data is fed to it from another source such as an API or
    indeed a CSV file or similar which is stored on the same server as
    the map-coding itself. This has the advantage of there being no
    data-storage limits as the developer is responsible for the data
    themselves, however the knowledge and skills required to put the
    data onto the map are more advanced than that of using CartoDB.

![The CartoDB data visualisation platform<span
data-label="fig:cartodb_screen"></span>](cartodb_screen)

When outputs of this research are required that are not suitable for a
map, then I will make graphs and charts using the plotting functionality
of the FOSS ’R’ project (@RFoundationforStatisticalComputing2014). One
of the benefits of using R is that it can be easily connected to the
PostgreSQL RDMS that was chosen earlier, and then statistical analysis
and publication-quality plots can be produced with relative ease and a
few lines of coding. For interactive or web-based charts and graphs the
D3 javascript library developed by Mike Bostock will instead be used
(@Bostock:2011:DDD:2068462.2068631) for similar reasons - it is free,
open-source, and there is a large community of developers actively using
it which gives a great deal of inspiration and examples to base future
work on e.g. <https://github.com/mbostock/d3/wiki/Gallery>.

Modelling Londoners movements {#chap:the_ltdsx}
=============================

Aim {#sec:1aim}
---

Create a model of Londoners daily movements based upon the London
Transport Demand Survey

Objectives {#sec:1objectives}
----------

-   Explore the LTDS dataset and identify key data

-   Move data from local Microsoft Access database into a PostgreSQL +
    PostGIS database

-   Clean data

-   Complete modal specific routing (interrogation, querying, storage)
    between origin and destinations

-   Quality check the routing results and do further data cleaning

-   Create and populate final dynamic location database table

-   Analysis of results

Background {#sec:1_background}
----------

As was heavily discussed in Section \[sec:staticexposurehealth\] many
air quality exposure studies do not consider the movements of the
subjects and/or to varying degrees other factors such as the temporal
fluctuations in pollutant concentrations and levels of infiltration into
micro-environments such as the home. The aim of this chapter was to
process and characterise the LTDS prior to application to the hybrid
exposure model.

This was achieved by using the methods outlined in Sections
\[sec:spatialdatabases\] (Spatial databases), \[sec:gis\] (GIS) and
\[sec:routing\] (Routing), and reconstructing the time-space location of
the population of London on a minute-by-minute basis, using the LTDS
dataset. This allowed interrogation of the data and such questions to be
answered as; how much time do people spend indoors each day, and is
there ethnic bias in the distance that people travel to work? To
demonstrate the capacity of the model/dataset visualisations/maps of
peoples routes were created and summary graphs of interesting
information.

Methods {#sec:1_methods}
-------

### Data Processing {#sec:reconstruction_data_processing}

The LTDS comprised of around 58 tables of data stored in an Microsoft
Access database. The majority of these tables were ’look-up’ tables that
store information that can be linked too from other tables e.g. the
Stages table described the stages of trips that the subjects made, and
had a column for transport mode which simply contains integers in the
range 1-26. These integers needed cross-referencing against the Mode
table to ascertain their meaning (1 = walk, 3 = car etc). This allows
for more efficient storage of data, but made gaining a full
understanding of the dataset and how tables linked together a lenghty
process. The four key tables that were required were (as briefly
discussed in Section \[sec:the\_ltds\]) the Household, Person, Trip and
Stage tables. Within each table there were columns that were more or
less useful than others. Table \[tab:key\_ltds\_fields\] below lists
some of the more important and useful ones:

  **Table**   **Key fields**
  ----------- --------------------------------------------------------------------------------------------------------------
              hhid: household id
              hhaboro: Household Borough
              htdate: Date house survey refers too
              hincomeei: Household income
              hhose and hhosn: easting and northing of household
              householddatatime: This column was created and populated with a full time-stamp
              phid: Person id
              phid: Household id of person
              ppiwt: Expansion factor of person
              psexi: Person gender
              pagei: Person age
              pegroup: Person ethnic group
              pegroup: Person ethnic group
              pseg: Person socio-economic group
              pnotrips: Number of trips person takes in survey 24 hours
              ttid: Trip id
              tpid: id of person doing the trip
              tstagesn: Number of stages in the trip
              tdpurp: Destination purpose of trip
              tstime: Trip start time
              tetime: Trip end time
              toose and toosn: Trip origin (OSGB36 Easting and Northing)
              tdose and tdosn: Trip destination (OSGB36 Easting and Northing)
              tdurn: Duration of trip
              departuretime: This column was created and populated with a full time–stamp of the trip departure time
              arrivaltime: This column was created and populated with a full time–stamp of the trip arrival time
              ssid: Stage id
              stid: Trip id
              smode5y: Stage mode of transport
              soose and soosn: Stage origin (OSGB36 Easting and Northing)
              sdose and sdosn: Stage destination (OSGB36 Easting and Northing)
              sdurn: Duration of trip (minutes)
              stagedeparturetime: This column was created and populated with a full time–stamp of the stage departure time
              stagearrivaltime: This column was created and populated with a full time–stamp of the stage arrival time

  : Key LTDS fields<span data-label="tab:key_ltds_fields"></span>

The tables listed above were exported as CSV files and then imported
into a PostgreSQL installation on a virtual Ubuntu machine. PostGIS was
then added to the PostgreSQL installation. SQL scripts were ran to
create full time-stamp columns in the Household, Trip and Stage tables
rather than use the format that the data was stored in e.g. dates were
stored in the Household table in the form ’20060221’ (meaning
2006-02-21) and times in the Trips table as ’730’ (meaning 07:30) which
made performing temporal queries on them difficult. Further
complications arose in the Trips and Stages tables as the survey period
begins and ends at 4am (or once the persons final trip was completed).
So times in the Trips table showed as, for instance ,’300’ which meant
3am, however that actually meant 3am on the morning of the day after the
survey date, rather than the survey date itself. Another factor was that
the stages table did not contain any stage departure times or stage
arrival times, only the duration of each stage. Therefore the stages
table needed cross-referencing against the trip table and then summing
incrementally by stage id in order to calculate the correct stage
departure and arrival times (and then extensively checking).

### Data cleaning {#sec:reconstruction_data_cleaning}

Data cleaning was required throughout the methods, however is briefly
mentioned here. A number of different processes were required, a few of
which are listed as bullet-points below. It should be noted that an
early decision was taken whereby if any data linked to a Person was
clearly incorrect, or not recorded properly, then that person was
removed from the dataset entirely (by creation of a column in the Person
table titled ’badflag’ and then entering the reason the person was
removed). For example if a person had a stage which started 15 minutes
before the next noted stage began, that person was removed rather than
just a correction of the data (in this case because it would be
difficult to know stage which to correct). Removal of records from the
dataset occurred in the following situations:

-   Trip start and end times were mis-aligned

-   Stage start and end times were mis-aligned

-   Stage transport mode was missing or refused

-   Key demographic data was missing or refused

-   Stage start and end locations were mis-aligned with the next or
    previous trip by greater than 80 metres

-   LTDS respondent did not live in London (the survey stretches just
    beyond the London Borough boundary)

-   Location of stage start or stage end were missing.

-   Transport mode was not available via a known API e.g. Plane or a bus
    journey outside of London

To ensure that removal of subjects from the dataset had not compromised
the overall structure and statistical strength of the data (and
therefore made using the weighting factors to scale the data
unsuitable), t-test and Wilcoxon–Mann–Whitney tests were performed to
compare the original data, and this subset of data. No statistical
significance was found when comparing age, Borough of residence,
ethnicity, income, sex, distances travelled each day, or the amount of
time in transport. The final dataset to be used was therefore judged as
suitable for purpose.

### Mode-specific routing {#sec:reconstruction_mode_routing}

To calculate the locations of the subjects while they were travelling
between locations the stages table of the LTDS was used, specifically
the start easting and northing, the destination easting and northing,
the transport mode, and the stage departure time (if allowed as an input
by the API). These columns were extracted from the database directly
into R using the RPostgreSQL extension and stored in an R dataframe.
These details were then formed into URL requests and sent to the routing
APIs by R. A small extract of the large XML file that is received in
response to a request to the TfL Journey Planner API is shown in Figure
\[fig:tfl\_response\_example\]. This was parsed using the RJSONIO and
XML packages, and the route between the two locations extracted and
stored (and then decoded in the case of Google). The route was then
formed into a linestring data type (which PostGIS can store as a spatial
object) and entered back into the database as the route taken between
the coordinates in an extra column called ’thegeom’.

![An example of an XML response from the TfL API<span
data-label="fig:tfl_response_example"></span>](tfl_response_example)

Due to restrictions on usage limits, and the modes of transport that
each API has available, different APIs were used for each transport mode
in the LTDS (see Table \[tab:api\_transport\_modes\]). Pre-processing
was required to simplify the transport modes before this, for example
the LTDS contains transport modes such as ’Car (passenger)’, ’Taxi’,
’Van (driver)’, ’Van (passenger)’ and ’Motorcycle (driver)’ which were
combined to the mode of ’car’ given that the likely routes would be the
same (and more importantly that the API did not have the ability to
distinguish between them as an input).

<span>@ | R<span>5cm</span> | R<span>8cm</span> |@</span> **<span>LTDS
Transport Mode</span> & **<span>API</span>\
Walking & OpenRouteService
([www.openrouteservice.org](www.openrouteservice.org))\
Cycling & Google Directions\
Train, Overground, Underground, Docklands Light Railway, Bus & Transport
for London Journey Planner\
Car & Project OSRM API ([www.project-osrm.org](www.project-osrm.org))\
****

The cleaned data contained 45,079 people, who took 98,770 trips during
the day of the survey. Each trip consisted of multiple stages that
needed routing independently, totally 340,754. These 340,000 routes were
stored as linestrings in the PostGIS database next to the id (ssid) of
the stage.

### Quality checking {#sec:reconstruction_quality_checking}

Spatial cleaning of the routes was now performed. This is discussed
earlier in Section \[sec:reconstruction\_data\_cleaning\]. It primarily
consisted of error-checking of the routes that had been obtained from
the APIs by making visualisation of the transport modes to manually
identify large errors e.g. were there any tube journeys recorded as
being in Scotland, or a taxi journey accross the Irish sea. The results
of the underground and Docklands Light Railway (DLR) routing are shown
in \[fig:underground\_routing\] below.

### Data manipulation {#sec:reconstruction_data_manipulation}

A ’base table’ for each subject in the cleaned and routed LTDS data
(45,079) was now created (and named the hybridlocation table). This
table contained four columns, Person ID (ppid), Time (pointtime), Mode
(mode) and Location (thegeom). A blank row for every minute of the
subjects 24 hours was created (45,079 subjects, multiplied by 24 hours,
multiplied by 60 minutes). The table was then populated with data from
the Stage table, by taking the line-strings of each route and splitting
them into minute-by-minute interval points using bespoke spatial
interpolation SQL scripts, and then matching those points to the correct
time in the new table. The mode and id of the stage were also copied
over for ease of future reference. A graphic which attempts to
demonstrate the premis behind the spatial interpolation method is shown
below in \[fig:spatial\_interpolation\_example\].

![The time and location between two known locations and times were
calculated using custom-made SQL scripts and the spatial functions of
PostGIS<span
data-label="fig:spatial_interpolation_example"></span>](spatial_interpolation_example)

Inbetween Trips subjects were presumed to be stationary and indoors at
the final point recorded of the previous trip. At the start and end of
days subjects locations were presumed to be the starting point of their
first trip, and the ending location of their last trip respectively (and
again, indoors). By doing this the location of the 45,079 people were
calculated and stored on a minute-by-minute basis for 24 hours. An
extract of the final table is shown below in
\[fig:hybrid\_location\_table\] (for reference mode 0 = indoors, 1 =
walking, 3 = car).

Results {#sec:reconstruction_results}
-------

The hybrid location table was used to examine the movements of the
subjects of the LTDS on a fine temporal and spatial scale. The data was
explored to better understand the detail available, to check that it was
suitable for use as the basis of a dynamic exposure model, and to see
whether there were any interesting results already present before air
quality was considered.

### Visual inspection of individuals {#sec:visual_inspection_of_individuals}

Before starting to summarise data and look for patterns in large groups
of people, Figures \[fig:routed\_results\_example\] and
\[fig:routed\_results\_example2\] were created and show the results of
two randomly selected individuals from the dataset. The first
(\[fig:routed\_results\_example\]) shows a fairly simple reconstructed
day. The Person made two car journeys around the local neighbourhood, as
well as one walking trip to a few streets away. Other than visiting
those three local locations, they spent their time at home. The second
(\[fig:routed\_results\_example2\]) shows a much more complicated
picture. The individual lives near Norbiton. In the morning they walked
to Norbiton station, took a train to Waterloo, and then caught a bus to
Clerkenwell (presumably for work). At lunchtime they walked around the
local neighbourhood (buying lunch?). At 20:13 they took a taxi back to
Waterloo, where they got a train back to Norbiton, and walked home from
the station to their house.

### Journey start and end times {#sec:journey_start_and_end_times}

By grouping the start times of the stages by hour, and then summarising
by the number of stages within that hour, a histogram of journey start
times was created (Figure \[fig:when\_people\_start\_trips\]).

![Histogram of when the population of London start trips<span
data-label="fig:when_people_start_trips"></span>](when_people_start_trips)

A similar method was also used for the ending times (Figure
\[fig:when\_people\_end\_trips\]).

![Histogram of when the population of London end trips<span
data-label="fig:when_people_end_trips"></span>](when_people_end_trips)

Figures \[fig:when\_people\_start\_trips\] and
\[fig:when\_people\_end\_trips\] show peaks of travel around peak travel
times that might be expected in the morning and evening rush hours,
however it is interesting to note that considerable travel occurs during
the day (although at this stage the data is not split by specific days,
so this may be influenced by travel on a Saturday and Sunday). It was
also interesting to see how trips during the evening (4pm-8pm) are much
more dispersed than in the morning, suggesting that people tend to leave
work over a longer time-frame than when they start work.

### Journey distances by sex {#sec:journey_distances_by_sex}

By summing the total distance travelled by each person over 24 hours,
and then grouping by gender, graph \[fig:mean\_distance\_by\_sex\] shows
differences in total travel distance.

![Boxplot of distances travelled, by gender (outliers &gt;100 km omitted
for clarity, red line links each mean)<span
data-label="fig:mean_distance_by_sex"></span>](mean_distance_by_sex)

Although the difference was not huge (Male mean of 18.28 km, Female mean
of 13.89 km), there was a significant difference (using a t-test)
between the kilometres that each gender travels in a day. Whether this
is because women stay at home more than me, or whether the journeys that
men are required to do during their day take them further was not clear
(but could be investigated further as the dataset contains a great deal
of contextual information).

### Journey distances by income {#sec:journey_distances_by_income}

A similar method was then used but the demographic of household income
was used as the variable to be considered rather than gender
(individual-level income variables were unfortunately not collected).
The result is shown in Figure \[fig:mean\_distance\_by\_income\].

![Boxplot of distances travelled by income group (outliers &gt;100 km
omitted for clarity, red line links each mean)<span
data-label="fig:mean_distance_by_income"></span>](mean_distance_by_income)

Interestingly the dataset showned that subjects that live in households
with a higher income actually travel further than households with a
lower income level (albeit with a levelling off and even slight dip
around the 75,000 mark).

### Journey distances by age group {#sec:journey_distances_by_age_group}

Distance of travel by age group was now plotted in Figure
\[fig:mean\_distance\_by\_age\].

![Mean distances travelled by age group (outliers &gt;100 km omitted for
clarity, red line links each mean)<span
data-label="fig:mean_distance_by_age"></span>](mean_distance_by_age)

Figure \[fig:mean\_distance\_by\_age\] showed a clear rise in the
distance that people travel each day as they get into their late teens,
which then becomes fairly steady until mid-50s at which point the
distances start to decline again. That said, a clear cut-off was not
visible. The gradient of the slops between 60 and 90 is much more steady
that at the other end of the age range, perhaps showing that the ages
that people retire are more dispersed than the ages at which people
start work.

### Journey distances by Borough of residence {#sec:journey_distances_by_borough_of_residence}

The distances that the people of London travel, depending on their
Borough of residence, was now considered. The centre of London was
defined (the monument outside of Charing Cross station) and then the
distance between this point and the centroid of each London Borough was
calculated (Figure \[fig:borough\_centroids\]). Figure
\[fig:mean\_distance\_by\_borough\] was then created, with the Boroughs
ordered by this distance metric i.e. the centroid of the City of London
is the closest Borough centroid to Charing Cross, so was plotted first.
This ordering was undertaken as I theorised that individuals living
nearer the centre of London would spend less time travelling than those
in outer London.

![Mean distances travelled by Borough of residence<span
data-label="fig:mean_distance_by_borough"></span>](mean_distance_by_borough)

There did not appear to be any clear pattern between where people live
and the distance that they travel each day. As the Boroughs were plotted
in order, if it was true that people living in outer London Boroughs
travel more, then the medians and means of the graph should rise from
left (closest to centre of London) to right (furthest from centre of
London). The Borough with the lowest mean was actually found to be
Greenwich, perhaps reflecting that the people who live in Greenwich tend
to work more locally (this could be investigated using the dataset). The
boroughs with the highest mean travelling distance were Havering,
Richmond and Merton.

### Transport mode choice by age group {#sec:transport_mode_choice_by_age_group}

Focusing now on transport mode choice, the percentage of time that each
person spends within each micro-environment, namely bus, car, cycling,
train, underground and walking, was calculated. The indoors environment
was found to be very dominant i.e. people spent most of their days
indoors. For clarity of visualisation, the indoor environment was
removed, and the remaining ones plotted by age bracket (Figure
\[fig:age\_barplots\_time\_transport\]).

![Percent of typical day using transport modes by income bracket<span
data-label="fig:age_barplots_time_transport"></span>](age_barplots_time_transport)

This graph agrees well with Figure \[fig:mean\_distance\_by\_age\]
(travel distances by age group) which showed a bell-curve like
distribution, with a leaning towards the upper age groups. Additionally
however Figure \[fig:age\_barplots\_time\_transport\] shows how use of
transport modes changes with age. Except for the 0–10 age group, bus
transport is used the same amount by all ages. Walking has a similar
pattern, being used frequently by people even up to the ages of 80 and
above. Cycling is popular with the 20 to 60 year old’s, but then hardly
used by people older than that. The underground results are interesting
in that it is almost never used by anyone in the 0-10 age bracket, or
the 80+ age bracket. Despite free travel being granted to people in
those age groups.

### Time near residence {#sec:time_near_residence}

One of the main motivations behind this research was to examine where
people spend time, and therefore accumulate their exposure, and whether
using address or postcodes (or dynamic methods) for exposure assessment
was valid. To examine this, a 1 km buffer was created around each
residents recorded residential address (example shown in Figure
\[fig:one\_km\_buffer\_house\]).

![Example of a 1 km buffer around residential address<span
data-label="fig:one_km_buffer_house"></span>](one_km_buffer_house)

Then the percent of time that each occupant of each house spends within
that buffer was calculated. The results are shown in Figure
\[fig:time\_with\_one\_km\_home\_by\_age\].

![Boxplot of percent of time within 1 km of home address, means shown by
red line<span
data-label="fig:time_with_one_km_home_by_age"></span>](time_with_one_km_home_by_age)

Figure \[fig:time\_with\_one\_km\_home\_by\_age\] shows an interesting
pattern of time spent within subjects local neighbourhoods varying by
age group. Considering the mean line (red), time spent at home hovers
between the 70% and 80% mark with some small variation up and down
between the ages of 5 and 60. From age 60 onwards the time spent within
the local neighbourhood steadily increases to 90% by age 70, and mid 90s
from there onwards (with a sudden dip in the 98 and 99 year olds,
however this seems likely to be due to small numbers of subjects of this
age who happened to be quite active on the day of the survey).

For illustrative purposes three simple animations of the data–set were
made. The first two using the QGIS TimeManager plug-in, and the third
using javascript. They can be viewed online at the following URLs:

-   A webpage with an interactive slider, whereby the user can view the
    time–activity points of the trips the subjects of the LTDS make, on
    a minute by minute basis, with the points colour-coded according to
    the mode at that point–time:

    -   <http://www.londonair.org.uk/research/dynamic_london/mode_animation.html>

-   An animated GIF file of the movements of the LTDS, clipped to show
    specifically London

    -   <http://www.londonair.org.uk/research/ltds_uk/traffic_london.gif>

-   An animated GIF file of the movements of the LTDS, showing the whole
    of the UK

    -   <http://www.londonair.org.uk/research/ltds_uk/traffic_uk.gif>

These illustrations allow a much better understanding of the data–set
than was possible before. Instead of rows of an database, the data can
be viewed and understood. When used in presentations to colleagues and
external stakeholders (such as the TfL staff who collect the LTDS data),
it was particularly useful in helping them become more familiar with the
results of the process. Particularly interesting to note in the
animations is the spatial distribution of the LTDS subjects, with their
movements on the day of the survey clearly not limited to the
geographical area of London (and thus justifying the need for a UK-wide
CMAQ model of air quality concentrations). Also the number of people who
travel from London to outside of London for work or otherwise.

Discussion {#sec:1Discussion}
----------

Generating or manipulating the LTDS with the methods and tools discussed
in this chapter proved complicated, particularly in writing the SQL and
R scripts to interrogate the APIs, and then the spatial and temporal
interpolation techniques to re-define the data in the resolution
required. However having initially attempted to build custom transport
networks using PostGIS and PgRouting (the PostgreSQL extensions), the
methods attempted were much more productive in that they allowed a far
higher percentage of routes to be solved, and with much higher accuracy.

The dataset that has been produced is, to my knowledge, unparalleled in
this area of study. Few studies have produced similar datasets, a
notable one being @Dhondt2012 and their 5 million subjects in the
’FEATHERS’ dataset, however the temporal resolution of that model was
only hourly and the spatial resolution down to 16 km by 16 km zones of
Flanders (Belgium) rather than exact coordinates. Conversely, it could
be argued that the hybridlocation table actually be more highly resolved
that it is. Time intervals of one-minute were chosen for convenience,
when 30 seconds or even 1 second could have been used. However this
seemed the most sensible choice on the basis that the dataset is already
very large. Further research to assess the impact of perhaps changing to
30 second intervals could be worthwhile. The linked demographic data to
the people and households of this data was also very useful in examining
travelling patterns and mobility by different characteristics, which has
been done in many studies before however normally by using crude
measures such as euclidean distances between households and recorded
workplaces, rather than by using routing solutions.

The limitations of this approach include the lack of finely-detailed
control over the routing process and solution. Whilst most of the APIs
allowed for different transport modes to be selected, and some of them
give options such as ’avoid major roads’, they do not allow cost-factors
to be dynamically assigned to roads or for the network to be manipulated
such as removing specific routes from the choices available. Whilst not
an issue at this stage of this research, a situation can be envisaged
whereby low exposure routing would be desirable, and this will prove
difficult. Similarly, there is currently no way of validating the routes
that were chosen by the subjects, as being representative of the actual
routes taken. For some transport modes this seems less of a worry than
for others. A subject taking a tube from London Bridge to Elephant and
Castle for example has very little choice of their route (without
significantly adding time to their journey), and much would be the same
for other journeys using public transport. Private modes such as driving
and walking would be less certain, however the time taken for the route
matches the time in the LTDS, so whilst small detours are possible,
large detours are unlikely.

The way that time inbetween trips is considered might be leading to an
over-estimation of the amount of time that subjects spend indoors each
day. This is because when subjects finish trips the model presumes that
they are then inside until the next trip begins. Time spent not
travelling but in outdoor spaces, is therefore neglected (for example in
a park). Intersecting an Ordnance Survey land-use map with the dataset
could lead to improvements in this area of the model. Further efforts
could also be made to ’recover’ discarded data which was removed in the
data-cleaning process, for example subjects with miss-aligned trip start
and end times were removed, but further investigation may have been able
to resolve these issues.

Going forward, this dataset will now be referred to for simplicity as
the LTDS-Expanded dataset, or the LTDS-X for short. Specifically, the
dataset created (expanded) by processing of the original LTDS data
through routing algorithms and GIS.

Conclusions {#sec:1conclusions}
-----------

The aim of this research was to process and characterise the LTDS to
create a new data–set that can be used as an input for a hybrid exposure
model. Spatially-enabled databases with custom scripts were wrote to do
this, R was used as an interface with various APIs, and the
’hybridlocation’ table was created successfully. Extensive quality
checking and data cleaning was undertaken. The final data-set allows
interrogation of the daily activities of 6.8 million people in London,
which can be interrogated on a fine spatial and temporal detail, aswell
as by various demographics. The main findings from this data–set were as
follows:

-   There are peaks in travel between 8am and 10am in the morning, and
    between 5pm and 6pm in the evening (As would be expected), however
    substantial travel also occurs between those hours.

-   Men travel more each day than women (mean of 18.28 km, compared to
    women with 13.89 km).

-   Households with a higher average annual income, are inhabited by
    people who travel more than households with a lower average
    annual income.

-   The distances that people travel each day increases as they get
    older, peaking around the ages of 38-42. It then steadily declines,
    however much more gradually than it rose. By late 80s, people travel
    very little.

-   There appears to be no pattern to where people live compared to the
    distances they travel each day

-   The amount of time that people spend in transport each day peaks in
    the 30-50 age categories.

-   Car use is the most popular form of transport amongst all age groups
    over 20 (until that point walking is the most popular). Walking is
    the second most popular, followed by the London Underground.

-   People over 80 rarely use the London Underground.

-   The amount of time that people spend within 1 km of their home, when
    analysed by age group, has a similar but slightly different pattern
    as the distance that people travel. Between the ages of 5 and 20,
    the mean is around 77%, which then drops to the low 70% mark for
    people aged 20 to late 50s. From 60 onwards, the time at home
    gradually increases up to high 80s.

This initial exploration and validation of the key characteristics of
the data were important to confirm that it was suitable for use in
further research. It will now be used as the main input for a hybrid
exposure model to estimate subjects exposure to air pollution.

Dynamic exposure modelling {#chap:the_lhem}
==========================

Aim {#sec:2aim}
---

Model exposure to PM$_{2.5}$ and NO$_{2}$ of the LTDS-X subjects, and
compare with traditional exposure models

Objectives {#sec:2objectives}
----------

-   Link the LTDS-X to CMAQ-Urban UK

-   Incorporate I/O ratios and micro-environmental factors

-   Create a postcode comparison dataset

-   Create a address-point comparison dataset

-   Create a monitoring sites comparison dataset

-   Analysis and compare the results

Background {#sec:2background}
----------

Chapter One focused on processing, checking and exploring the spatial
data that was created from the LTDS, leading to the LTDS-X. The daily
journeys of around 45,000 people were recreated on a fine spatial and
temporal scale from survey data. This data-set was created to allow
investigation of exposure in urban environments, and also
miss-classification, as described in Section
\[sec:dynamicexposurehealth\], namely the differences between assigning
someone a static exposure value based on their home location, postcode,
nearest monitoring site or otherwise, compared with a model which
considers all the micro-environments and varying concentrations during
the subjects daily movements.

The thrust of this chapter will therefore be to link the LTDS-X dataset
to the CMAQ-Urban dataset (described in \[sec:cmaq\_urban\]), and then
to undertake micro-environmental modelling for when the subjects are
indoors and for when they are in transport. The completed model of
exposure is henceforth referred to as the London Hybrid Exposure Model
or LHEM. The LHEM will then be used to explore exposure variations
within the subjects, to attempt to identify and understand any patterns,
and to consider exposure missclassification that may be occurring using
standard exposure methods. Potential policy uses of the LHEM, and ways
in which it might be useful to the general public, are then discussed.

Methods {#sec:2methods}
-------

In order to calculate the ’static’ exposure estimates which will be
compared to the LHEM estimates, a number of input data-sets and
methods/processes needed to be completed. They are now listed and
explained below, with the methods and input data-sets grouped by the
final dataset under construction.

### Running the London Hybrid Exposure Model {#sec:running_the_lhem}

#### Linking LTDS-X to outdoor concentrations {#sec:linking_ltdsx_to_cmaq}

During Methods section \[sec:cmaq\_urban\] the CMAQ-Urban air quality
model was introduced. It was explained how this model outputs daily
(weekday/Saturday/Sunday), hourly concentrations for a 20 metre by 20
metre grid covering the UK. The concentration of a range of pollutants
at the location of each individuals minute-by-minute location over 24
hours was therefore now extracted from this model output i.e. the LTDS-X
was linked to a CMAQ-Urban layer. Due to the memory and processing power
needed to run the CMAQ-Urban model, and the language that it is written
in (Fortran with SQL inputs), linking CMAQ-Urban directly to the
PostgreSQL database that the LTDS-X data is held in was not possible. A
CSV file of the LTDS-X in lat/long format along with a date-time-stamp
was thus exported from the database. To avoid duplication at this stage,
a SQL query was written to only export unique points. To explain, if
someone stayed in the same location between 07:00 and 07:45 on a
Saturday, only one point was exported as the temporal resolution of the
CMAQ-Urban model is monthly, daily, hourly and thus the concentration at
that point would not change during that time-frame. If however the same
person was in constant movement between 07:00 and 07:45, then 45 points
were outputted (as concentrations would be different in different
locations for each minute). By taking the temporal resolution of the
CMAQ-Urban model into account when doing the export query, the number of
points that needed concentrations extracting from the CMAQ-Urban model
was reduced from around 64 million (45,079 people multiplied by 24 hours
multiplied by 60 minutes) to just over 4 million. Dr Kitwiroon of the
Environmental Research Group at King’s College London processed this
dataset with the CMAQ-Urban model, and returned a CSV file with
additional columns for a range of pollutants (including crucially
NO$_{2}$ and PM$_{2.5}$). This data was then re-imported back to the
PostgreSQL database and linked to the LTDS-X data.

#### Modelling for in-building exposure {#sec:modelling_in_building}

While LTDS-X subjects are indoors, the air quality that they are exposed
to is different to outdoor air. A method to estimate exposure to outdoor
air when indoors was therefore required, particularly given that
subjects spend so much of their time indoors (See Figure
\[fig:age\_barplots\_time\_transport\] in Section
\[sec:reconstruction\_results\]). We decided to use an indoor/outdoor
(I/O) model to estimate concentrations inside buildings, by taking
ratios and applying them to the outdoor CMAQ-Urban concentrations. The
model that was chosen for this was developed by Dr J. Taylor of UCL, in
which he assumes 15 building types derived from the English Housing
Survey, and then creates building physic models using the location of
the dwelling, window opening and closing behaviour, occupant behaviour,
deposition rates and penetration factors. This model was chosen as it
was specifically developed for London (and therefore the building
archetypes are representative), was recent (2014), and due to our close
relationship with Dr Taylor meaning that we were able to ask for minor
customisations to the model to be completed, and for a bestoke number of
model runs to be udertaken to examine sensitivity of the model (not
covered here). The methods are more fully described in @Taylor2014, and
the data was provided by personal communication with Dr Taylor (with
minor customisations as mentioned to provide hourly results and the
addition of NO$_{2}$ instead of just PM). A map of the PM$_{2.5}$ ratios
with the London Underground overlaid to help with the readers
orientation is shown below in Figure \[fig:ratios\_underground\].

![Map of average indoor/outdoor (I/O) ratios used in the LHEM.
Superimposed on the map is the London Underground network to aid
orientation<span
data-label="fig:ratios_underground"></span>](ratios_underground)

##### Importing I/O ratios {#sec:importing_io_ratios}

The data-set was provided as a CSV file which gave hourly I/O averages
for each district level postcode in London. The data was re-organised
slightly, before being imported to the PostgreSQL database.

##### Linking postcodes boundaries to Indoor/outdoor ratios {#sec:linking_postcodes_to_ratios}

In order to link the provided I/O ratios to the locations in the LTDS-X,
the areas that each postcode covers was required (the I/O file provided
by Taylor contained a ratio, and a postcode i.e. SE173DA, 0.6 – but did
not contain geographical data defining the area that this postcode
covers). The geographical postcode data-set that was used is described
in Section \[sec:import\_postcode\_boundaries\], and linked to the
ratios dataset using the postcode field common in both files.

##### Take each point and link to the correct I/O on time/place {#sec:linking_points_to_postcode_ratios}

To then link the LTDS-X data to the postcode I/O ratios, a spatial SQL
query to join the two tables was written. The join was performed using
the location of the LTDS-X point, as well as the time of day (as the I/O
dataset had this level of temporal resolution). This process is
summarised in the bullet-points below:

1.  Take the LTDS-X location (Easting and Northing) and locate the
    postcode boundary that contains the point

2.  Now take the hour of the day that the LTDS-X is for (0-24)

3.  Extract the I/O ratio

##### Multiply I/O ratio by CMAQ-Urban concentration {#sec:multiple_io_by_cmaq}

The appropriate I/O ratio was then multiplied by the CMAQ unique point
value for that location, day and hour, the result being the indoor
exposure at that minute. This process was repeated for all the LTDS-X
data-points that are noted as being indoors in the dataset
(approximately 62 million points)

#### Modelling for in-vehicle exposure {#sec:in_vehicle_modelling}

Locations in the LTDS-X that were recorded as travelling inside vehicles
(buses, cars, trains etc.) needed further micro-environmental modelling
to take into account that the air quality is different from the outside
air. This concept was introduced in Section \[subsubsec:invehicle\] and
then exposure studies that considered this area of modelling and
exposure were examined in Section \[sec:transport\]. To calculate
in-vehicle exposure in this model, the pollutant concentration
(C$_{in}$) was derived by solving the following mass balance equation
below (Equation \[eq:mass\_balance\]).

$$\frac{dC_{in}}{dt} = \lambda _{win} (C_{out} - C_{in}) - n\lambda _{HVAC} C_{in} - V_{g} (A\textsuperscript{*}/V) C_{in} + Q/V
\label{eq:mass_balance}$$

-   C$_{out}$ is the outdoor CMAQ-Urban concentration linked in Section
    \[sec:linking\_ltdsx\_to\_cmaq\]

-   $\lambda _{win}$ is the air exchange rate from the windows

-   $\lambda _{HV AC}$ is the air exchange rate from the mechanical
    ventilation system

-   *n* is the filter removal efficiency taking values between 0 and 1

-   V$_{g}$ is the deposition velocity in m/h$^{-1}$

-   A^\*^ is the internal surface area available for deposition

-   V the volume of the vehicle

-   Q is the in-vehicle particle emission rate in ug/h$^{-1}$ (defined
    as the product of the re-suspension rate and the number of
    active passengers)

This was solved analytically and the general solution is shown below in
Equation \[eq:solved\_mass\_balance\].

$$C_{in} = (C_{in_{0}} - \frac{b_{0}}{a_{0}}) \cdot exp(-a \cdot t) + \frac{b}{a}
\label{eq:solved_mass_balance}$$

The parameters for this model change as the subjects move, and for
different vehicle types. For an example of how this module of the LHEM
operates, please find a standalone piece of R code in Appendix .

##### The London Underground {#sec:the_london_underground}

For subjects locations in the LTDS-X that are described as being on the
London Underground, fixed concentrations were used due to a lack of the
data required to create a model of adequate spatial and temporal
resolution. These fixed concentrations were derived from measurements
conducted at underground platforms and on trains, in a separate as yet
unpublished study within the Environmental Research Group at King’s
College London (Dr. Barratt, personal communication). For PM$_{2.5}$ the
values of 94 $\mu \text{g m}^{-3}$ (winter) and 68 $\mu \text{g m}^{-3}$
(summer) were used, and for NO$_{x}$ the value of 51
$\mu \text{g m}^{-3}$ was used. n.b Chapter
\[chap:monitoring\_on\_underground\] aims to refine the estimation of
exposure while on the London Underground. NEED TO ADD PARIS REFERENCE
HERE ABOUT NO2

#### Summary of LTDS-X to LHEM {#sec:summary_of_ltdsx_to_lhem}

The LTDS-X data is processed, as described above, using the appropriate
method for each micro-environment. Once complete a dataset of 1,440
records (24 hours x 60 minutes) including time, location and exposure is
output for each individual and then the 1-minute resolution data are
averaged into hour of the day or full 24 hour period to obtain the
typical exposure for each individual. These are then grouped or
disaggregated as appropriate depending on the analysis required.

### Creation of a postcode comparison dataset {#sec:creating_postcode_dataset}

#### Importing postcode boundaries {#sec:import_postcode_boundaries}

To be able to calculate annual average pollutant concentrations for each
London postcode, a dataset of postcodes was required, and specifically
one which contains the geographical information describing the boundary
of the postcode polygon. In the UK the Royal Mail is the organisation
with authority for maintaining a list of all postcodes, specifically the
dataset is called the Postcode Address File or PAF. However this dataset
does not contain the geographical information to link the postcodes to
any other spatial data. Therefore the Ordnance Survey dataset,
Code-Point Open (@OrdnanceSurvey2015a), was considered. This is a
dataset maintained by the Ordnance Survey, derived from the Postcode
Address File, which adds the Easting and Northing of the postcode
centroid to each of the Royal Mail postcodes. Although geographical
coordinates now allow plotting of this dataset as points, to calculate
postcode polygon averages a boundary is required, and therefore this
dataset was also not suitable. Fortunately the organisation Edina, part
of the “EDINA and Data Library” division of the Information Services
Department at the University of Edinburgh, and funded by the UK Higher
Education Authorities Joint Information and Systems Committee, has
derived postcode polygon boundaries from this dataset and makes this
dataset freely available to other UK Higher Education Institutions as a
file called ’Code-Point with Polygons’ (@OrdnanceSurvey2015). This
dataset was therefore downloaded as an ESRI shapefiles and then the
shp2pgsql tool used to load the shapefiles into a PostgreSQL/PostGIS
database. The area of Waterloo is shown in Figure
\[fig:postcode\_polygons\_example\] to illustrate the detail of the
final postcodes dataset.

![Postcode polygons from Edina Digimap<span
data-label="fig:postcode_polygons_example"></span>](postcode_polygons_example)

#### Importing CMAQ-urban annual average points {#importing_cmaq_annual_averages}

To calculate the mean annual pollutant concentration within each
postcode polygon, a CMAQ-Urban output file containing annual average
2011 concentrations covering a 20m x 20m grid of London was generated by
Dr Kitwiroon of the Environmental Research Group at King’s College
London. This was imported into the PostgreSQL/PostGIS database using the
raster2pgsql tool. To demonstrate this data,
\[fig:example\_cmaq\_annual\] and
\[fig:example\_cmaq\_annual\_with\_postcodes\] were produced by loading
the data into QGIS and a colour gradient applied (and are shown below).

![CMAQ-Urban annual mean concentration raster (2011)<span
data-label="fig:example_cmaq_annual"></span>](example_cmaq_annual)

![CMAQ-Urban annual mean concentration raster (2011) with postcode
layer<span
data-label="fig:example_cmaq_annual_with_postcodes"></span>](example_cmaq_annual_with_postcodes)

#### Calculating the mean concentration for each postcode {#calculating_mean_postcode_data}

To calculate the mean concentration for each postcode, the mean of the
concentration of all the 20m x 20m cells that intersected the postcode
were taken. Each of the LTDS subjects home address Easting and Northing
was then taken, spatially joined with the postcode dataset to establish
which postcode they lived in, and then the relevant annual mean
concentration taken from the previously mentioned join/method.

### Creation of address-point comparison dataset {#sec:creating_address_point_dataset}

To calculate the address-point comparison dataset, the home address
location (Easting/Northing) of each LTDS participant was first taken, as
well as the day of the week that the participant was surveyed
(translated into weekday/Saturday/Sunday as this was the temporal
resolution of CMAQ-Urban). This data was then extracted as a CSV for
similar reasons to the section \[sec:linking\_ltdsx\_to\_cmaq\] and
processed by Dr Kitwiroon who returned a CSV with additional columns for
a range of pollutants at each location. The 24 hour average for each of
the LTDS subjects addresses was then calculated for each pollutant.

### Creating of monitoring sites comparison dataset {#sec:creating_monitoring_site_dataset}

#### Monitoring site data {#sec:monitoring_site_data}

As discussed in Section \[subsec:monitoringstation\], monitoring
stations have often been used as a measure of exposure for a population,
particularly in time-series studies e.g. @Atkinson2010, but also in some
cohort studies e.g. @Dockery1993. Comparisons between the annual average
of a London ’roadside’ monitoring station and a London ’background’
monitoring station were therefore chosen to compare with the LHEM
results. The data for these sites was downloaded using OpenAir from the
London Air Quality Network (@LondonAir) for the sites of Marylebone Road
(roadside) and North Kensington (background) respectively (shown in
Figure \[fig:background\_roadside\_monitors\]).

![The monitoring stations and surrounding areas (North Kensington left,
Marylebone Road right<span
data-label="fig:background_roadside_monitors"></span>](background_roadside_monitors)

Specifically, the hourly means for 2011, for each site, for PM$_{2.5}$
and NO$_{2}$ were downloaded. Providing that there was at least a 75%
capture rate for the year for each pollutant (which there was) then the
mean of the hours was taken as per DEFRA guidance (@DEFRA2009). The
results are shown in Table
\[tab:mean\_monitoring\_site\_concentrations\]:

  --------------------------------------------------------------------------------------------------------------------------- -- --
  **<span>Site</span> & **<span>PM$_{2.5}$ ($\mu \text{g m}^{-3}$)</span> & **<span>NO$_{2}$ ($\mu \text{g m}^{-3}$</span>\      
  North Kensington & 16.33 & 35.96\                                                                                              
  Marylebone Road & 24.45 & 97.05\                                                                                               
  ******                                                                                                                         
  --------------------------------------------------------------------------------------------------------------------------- -- --

  : Mean pollutant concentrations from London monitoring sites for
  2011<span data-label="tab:mean_monitoring_site_concentrations"></span>

#### Summary {#subsubsec:methods_summary}

Now that the LHEM has been created, it allows detailed interrogation and
investigation of the exposure of \~45,000 Londoners (and the demographic
and geographic information linked to them). Calculations can be made to
answer many questions. To give an indication of it’s capabilities, some
examples are listed below:

-   What is the average NO$_{2}$ exposure of those under 18, compared to
    those over 18

-   What is the average PM$_{2.5}$ exposure of people of Indian ethnic
    origin living in Southwark

-   What is the difference between exposure taken from monitoring sites,
    compared to exposure using the LHEM

-   What percentage of Londoners daily PM$_{2.5}$ exposure comes from
    their morning commute

-   How much less (or more?) NO$_{2}$ is someone exposed to by working
    at home instead of in the office

-   Which Borough of London residents have the lowest exposure

-   Is household income and air quality exposure related

-   Do children get most of their daily exposure from within 1km of
    their house

-   What is the difference between exposure using address-point methods,
    compared to exposure using the LHEM

-   During which hour of the day, do people aged between 30-40 years old
    get most of their daily exposure

Results {#sec:2results}
-------

The results presented here are illustrations of potential uses, focused
on exploring exposure classification and missclassification in the study
subjects.

### The effect of microenvironments on exposure {#subsec:time_exposure_microenvironments}

Figure \[fig:age\_barplots\_time\_transport\] in Section
\[sec:reconstruction\_results\] showed the amount of time that people in
different age groups spend in various microenvironments during their
day. Although the results varied by age group, generally the time that
people spent indoors was around the 95% mark, perhaps adding weight to
the sort of exposure estimates discussed in Section
\[sec:staticexposurehealth\] whereby concentrations at the subjects home
or general area of residence are used to investigate the negative health
effects of air quality (with the caveat that they tend to take the
outdoors concentration at the area of residence, rather than attempt to
model or measure the indoors concentration). Using the LHEM model, we
are now able to investigate the actual exposure that occurs from each
microenvironment, and examine the contribution that all
microenvironments make to a subjects daily exposure. Table
\[tab:results\_microenvironments\_exposure\] therefore summarises the
time spent in each microenvironment for NO$_{2}$ and PM$_{2.5}$ by age
category for the 45,709 people in the dataset, and compares the figures
to the exposure they accrue in the same environment during their day (as
a percent of their total exposure).

  -- -------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --
     **<span>Child</span> (5-17) & **<span>Young adult</span> (18-29) & **<span>Adult (30-59)</span> & **<span>Elderly (&gt;=60)</span> & **<span>Overall</span>\            
     People & 10856 & 7474 & 18370 & 8379 & 45079\                                                                                                                           
     \                                                                                                                                                                       
     Driving & 0.77, 0-0 & 1.36, 0-1.32 & 2.24, 0-3.12 & 1.54, 0-2.08 & 1.63, 0-2.08\                                                                                        
     Indoor & 97.72, 96.39-100 & 94.94, 92.23-98.82 & 94.69, 92.16-98.47 & 96.41, 94.86-100 & 95.73, 93.55-100\                                                              
     Walking & 0.86, 0-1.53 & 1.66, 0-2.5 & 1.49, 0-2.22 & 1.15, 0-1.6 & 1.31, 0-2.01\                                                                                       
     Underground & DLR & 0.06, 0-0 & 0.73, 0-0 & 0.5, 0-0 & 0.16, 0-0 & 0.38, 0-0\                                                                                           
     Bus & 0.53, 0-0 & 0.94, 0-0.56 & 0.66, 0-0 & 0.63, 0-0 & 0.67, 0-0\                                                                                                     
     Cycle & 0.02, 0-0 & 0.07, 0-0 & 0.1, 0-0 & 0.01, 0-0 & 0.06, 0-0\                                                                                                       
     Train & 0.03, 0-0 & 0.24, 0-0 & 0.2, 0-0 & 0.06, 0-0 & 0.15, 0-0\                                                                                                       
     Motorcycle & 0, 0-0 & 0.02, 0-0 & 0.04, 0-0 & 0, 0-0 & 0.02, 0-0\                                                                                                       
     \                                                                                                                                                                       
     Driving & 3, 0-0 & 5.32, 0-4.37 & 8.52, 0-12.62 & 5.64, 0-6.94 & 6.21, 0-7.49\                                                                                          
     Indoor & 92.01, 87.35-100 & 82.04, 71.22-96.46 & 81.2, 70.58-94.99 & 87.97, 81.58-100 & 85.02, 75.21-100\                                                               
     Walking & 2.64, 0-4.36 & 5.42, 0-8.41 & 4.66, 0-7.24 & 3.3, 0-4.64 & 4.08, 0-6.28\                                                                                      
     Underground & DLR & 0.21, 0-0 & 2.49, 0-0 & 1.72, 0-0 & 0.57, 0-0 & 1.31, 0-0\                                                                                          
     Bus & 1.97, 0-0 & 3.61, 0-1.86 & 2.52, 0-0 & 2.2, 0-0 & 2.52, 0-0\                                                                                                      
     Cycle & 0.07, 0-0 & 0.26, 0-0 & 0.39, 0-0 & 0.05, 0-0 & 0.24, 0-0\                                                                                                      
     Train & 0.07, 0-0 & 0.64, 0-0 & 0.54, 0-0 & 0.15, 0-0 & 0.38, 0-0\                                                                                                      
     Motorcycle & 0, 0-0 & 0.08, 0-0 & 0.19, 0-0 & 0.02, 0-0 & 0.10, 0-0\                                                                                                    
     \                                                                                                                                                                       
     Driving & 1.3, 0-0 & 2.34, 0-2 & 3.81, 0-5.38 & 2.52, 0-3.15 & 2.77. 0-3.34\                                                                                            
     Indoor & 95.89, 93.97-100 & 87.77, 82.96-98.04 & 88.59, 84.75-97.54 & 93.4, 91.47-100 & 90.98, 87.88-100\                                                               
     Walking & 1.39, 0-2.36 & 2.51, 0-3.8 & 2.28, 0-3.4 & 1.74, 0-2.48 & 2.02, 0-3.06\                                                                                       
     Underground & DLR & 0.44, 0-0 & 5.22, 0-0 & 3.55, 0-0 & 1.19, 0-0 & 2.71, 0-0\                                                                                          
     Bus & 0.9, 0-0 & 1.56, 0-0.9 & 1.09, 0-0 & 0.99, 0-0 & 1.11, 0-0\                                                                                                       
     Cycle & 0.04, 0-0 & 0.12, 0-0 & 0.18, 0-0 & 0.02, 0-0 & 0.11, 0-0\                                                                                                      
     Train & 0.04, 0-0 & 0.35, 0-0 & 0.3, 0-0 & 0.08, 0-0 & 0.21, 0-0\                                                                                                       
     Motorcycle & 0, 0-0 & 0.03, 0-0 & 0.08, 0-0 & 0.01, 0-0 & 0.04, 0-0\                                                                                                    
     **********                                                                                                                                                              
  -- -------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --

  : Time and exposure ($\mu \text{g m}^{-3}$) in microenvironments by
  age category from the LHEM model<span
  data-label="tab:results_microenvironments_exposure"></span>

The results of comparing time in microenvironments to exposure in those
same microenvironments shows that the contribution of the indoor
environment is very important to people’s overall exposure – people
spend \~95% of their time indoors. However when taken as a percentage of
their daily exposure, it’s importance is slightly diminished, overall
the people only accumulated \~85% of their daily NO$_{2}$ and \~90% of
their daily PM$_{2.5}$ while in that microenvironment.

In balance to this, the time that people spend in transit is small, but
becomes more important when daily exposure to pollutants is considered.
For example time spent driving is less than 2% of time, but over 6% of
NO$_{2}$ exposure, and the underground accounts for less than 0.5% of
time, but contributes almost 3% of PM$_{2.5}$ exposure.

The variation of time and exposure between age groups varies. For
NO$_{2}$, children and the elderly accumulate more of their daily
exposure indoors (92.01% and 87.97%) than young adults and adults
(82.04% and 81.2%) reflecting the differences in time they spend in that
environment and the pollutant concentrations they are exposed too during
their day. This pattern is similar for PM$_{2.5}$ exposure.

With regard to which transport modes contribute most to exposure, for
PM$_{2.5}$ the ranking is driving, underground & DLR, walking and then
the bus, with all other transport modes less than 1% of daily
contribution to exposure. For NO$_{2}$ the ranking is slightly
different, being driving, then walking, then the bus, then the
underground & DLR – reflecting the balance of pollutant types in
different environments. Noticeably when comparing age groups, young
adults get ten times more of their daily PM$_{2.5}$ exposure (5.22%)
from the underground than children do (0.44%), and four times more than
the elderly (1.19%). Comparisons between active and passive travel are
also interesting, for example when looking at the subjects overall,
passive travel constitutes 6.84% of someones daily PM$_{2.5}$ exposure,
compared to 2.85% of their time, but for active travel these figures are
2.13% and 1.19% respectively, meaning that on a minute-by-minute basis,
active travel results in lower exposure. This pattern is similar for
NO$_{2}$, where 10.52% of exposure comes from passive travel in 2.85% of
their time, compared to 4.32% of exposure from 1.19% of time.

### Comparing methods of exposure estimation {#subsec:comparing_exposure_methods}

As discussed extensively so far in previous chapters, the primary
function of the development of the LHEM is to consider the variation and
potential exposure miss-classification occurring by using different
exposure metrics. Table \[tab:comparing\_methods\] below summarises
(mean, median and interquartile range) the exposures of the 45,079
people in the LTDS dataset using five different exposure metrics to
provide side-by-side comparison. Notably all but the LHEM methods work
on a static-outdoor basis, only the LHEM attempts to model movements and
the mitigating effect of the indoor environment.

<span> | l | l | l | l | </span> & Mean & Median & Interquartile Range\
\
Background monitoring site & 16.33 & 12 & 8-19\
Roadside monitoring site & 24.45 & 22 & 14-32\
Postcode & 13.49 & 13.53 & 13.14-13.84\
Residential address & 13.54 & 13.62 & 12.99-14.16\
LHEM Model & 8.48 & 8.23 & 7.80 – 8.66\
\
Background monitoring site & 35.96 & 30.08 & 19.1-49.18\
Roadside monitoring site & 97.05 & 90.25 & 61.12-126.5\
Postcode & 34.56 & 34.59 & 31.20-37.59\
Residential address & 34.34 & 34.45 & 30.65-38.29\
LHEM Model & 13 & 12.34 & 10.82-14.64\

For both NO$_{2}$ and PM$_{2.5}$ the LHEM calculates lower exposure than
any of the other exposure metrics. The roadside monitoring site gives
the highest general exposures, followed by the background monitoring
site, followed by postcode and residential address which are almost
identical (although with residential address giving a larger
inter-quartile range), and then the LHEM. Figures
\[fig:address\_point\_v\_lhem\_no2\_hist\] and
\[fig:address\_point\_v\_lhem\_pm25\_hist\] below show histogram plots
of exposure at the residential address, compared to exposure using the
LHEM, for NO$_{2}$ and PM$_{2.5}$ respectively, to aid understanding of
the distribution of exposures (residential address exposure method is
chosen for comparison to the LHEM, rather than say postcode or
monitoring site, as this is currently seen as the most accurate method
due to it’s fine spatial detail).

![Daily mean exposure to NO$_{2}$ comparing residential address exposure
with the LHEM<span
data-label="fig:address_point_v_lhem_no2_hist"></span>](address_point_v_lhem_no2_hist)

![Daily mean exposure to PM$_{2.5}$ comparing residential address
exposure with the LHEM<span
data-label="fig:address_point_v_lhem_pm25_hist"></span>](address_point_v_lhem_pm25_hist)

When looking at the 90th to 10th percentile range of exposures from
these two methods as oppose to the inter-quartile ranges, there is
little difference in the relative size of the range for PM2.5 (2.08
($\mu \text{g m}^{-3}$) for LHEM and 2.15 ($\mu \text{g m}^{-3}$) for
residential address). However for NO2 the range is twice as large in the
residential method than it is in the LHEM ( 14.36
($\mu \text{g m}^{-3}$) compared to 7.64 ($\mu \text{g m}^{-3}$))

If we now plot each individuals NO$_{2}$ and PM$_{2.5}$ exposure using
the residential method, against the LHEM method, and colour-code the
plots by whether the person left their house or not during the day of
the survey, we can see that the change in exposure seems to be due to
their movement around the urban environment away from their house (See
Figure \[fig:lhem\_address\_whether\_leave\_house\] below).

![Comparing LHEM v. residential address exposure results, colour-coded
by whether the subject left their house or not<span
data-label="fig:lhem_address_whether_leave_house"></span>](lhem_address_whether_leave_house)

### Highly exposed people {#subsec:highly_exposed_people}

Table \[tab:whopmandno2levels\] is reproduced from Section
\[subsec:urbanenvironments\] below for convenience and shows the
acceptable mean daily and annual PM and NO$_{2}$ concentrations as
prescribed by the World Health Organisation. They are now used as a
reference number with which to identify the numbers of people in the
LTDS who have high average exposures when using the residential address
method, and then when using the LHEM method. This comes with the caveat
that the WHO limits are designed to be referenced against outdoor air
quality, and so are suitable for the residential address method, but are
not for the LHEM which introduces indoor and in-transport
microenvironments to modelling exposure. However there are not limits by
the WHO or EU or indeed any regulatory body currently that take account
of this type of exposure modelling, and so are used as indicative values
for comparison only. Specifically, the annual average values are used as
the LTDS data is designed to be typical of a days movements and
activities of each person.

               Annual mean ($\mu \text{g m}^{-3}$)   24 hour mean ($\mu \text{g m}^{-3}$)
  ------------ ------------------------------------- --------------------------------------
  PM$_{2.5}$   10                                    25
  PM$_{10}$    20                                    50
  NO$_{2}$     40                                    200

  : Table of ’acceptable’ WHO PM$_{2.5}$ and NO$_{2}$ levels<span
  data-label="tab:whopmandno2levels"></span>

Exposure estimates undertaken using a residential address exposure
method find that 14% of the subjects (6,469 out of 45,079) have a daily
NO$_{2}$ exposure higher than the WHO value of 40 $\mu \text{g m}^{-3}$.
However using the LHEM model, less than 1% (18 people) have an average
exposure over this value. For PM$_{2.5}$ the residential exposure method
finds that 100% of the subjects in London have an average exposure of
higher than 10 $\mu \text{g m}^{-3}$. Whereas the LHEM finds only 8%
(3,491 out of 45,079) of people above this limit.

### Exposure peaks {#subsec:exposure_peaks_by_age_group}

In the background section \[subsec:longtermvshortterm\] which considered
the relative importance of short-term exposure compared to longer-term
exposure to an individual, it was noted that hyper-short-term exposure
had not been explored in epidemiological style health studies so far due
to the difficulties in collecting this data and subsequent linkages to
health records and outcomes. Whilst the LHEM does not have the
capability to fully answer this question, it can be used to explore the
variation in concentrations between micro-environments, and the time
that each of the 45,079 subjects spend in environments of elevated
concentrations. The table below therefore classifies the percentage of
time that, on averages across the people in that age group, is spent in
environments where the concentration is higher than the WHO annual mean
levels for ’acceptable’ (with the same caveat as noted in the previous
section about it’s use in the LHEM). This is 10 $\mu \text{g m}^{-3}$
for PM$_{2.5}$, and 40 $\mu \text{g m}^{-3}$ for NO$_{2}$.

  Age category          Percent of time in high PM$_{2.5}$   Percent of time in high NO$_{2}$
  --------------------- ------------------------------------ ----------------------------------
  Child (5-17)          13.1% (8.8%-16.7%)                   1.3% (0%-1.7%)
  Young adult (18-29)   16.0% (12.5%-16.1%)                  3.8% (0%-6.1%)
  Adult (30-59)         15.8% (11.9%-20.3%)                  3.7% (0.1%-5.8%)
  Elderly (&gt;=60)     13.7% (8.9%-17.3%)                   2.0% (0%-2.6%)

  : Table of time of day in environments above ’acceptable’ WHO
  PM$_{2.5}$ and NO$_{2}$ levels (I/Q range in brackets)<span
  data-label="tab:people_time_above"></span>

In Section \[subsec:comparing\_exposure\_methods\] the LHEM calculated
that residential address-based exposure estimates appeared to be
over-estimating exposure, showing LHEM histogram plots with much lower
distributions of exposure to both PM$_{2.5}$ and NO$_{2}$. However
despite these lower daily averages, we can see that the LHEM also finds
that, depending on age category, between 13.1% and 16% of people’s time
is spent in high concentrations of PM$_{2.5}$, and between 1.3% and 3.8%
for NO$_{2}$.

### Geographical missclassification {#subsec:geographical_missclassification}

As the LTDS dataset contains the residential address of the subjects,
the percentage difference between the residential and LHEM models can be
calculated and then mapped to investigate whether there are any
geographical patterns in the data that are not apparent from the results
presented so far e.g. do people that live in North London have greater
missclassification than those that live in inner London? A cumulative
distribution function plot of the missclassification between the two
models, expressed as a percentage, is shown first in Figure
\[fig:cumulative\_missclass\_dist\] (below), in order to examine the
range of values.

![Percentage missclassification between LHEM and residential address
exposure methods, plotted as a cumulative distribution plot<span
data-label="fig:cumulative_missclass_dist"></span>](cumulative_missclass_dist)

As is evident from these plots, for the majority of subjects in the
dataset, for both PM$_{2.5}$ and NO$_{2}$, the LHEM calculates their
exposure as being around 30% - 50% lower than the residential address
method. There are many ways in which this aspect of the LHEM results
could be interrogated, but as an example, a map of the addresses of the
subjects where the LHEM calculates an **increase** in exposure compared
to residential address was now created to look for obvious spatial
patterns (Figure \[fig:address\_lhem\_increases\]).

![The residential address of subjects whose exposure increased by using
the LHEM method compared to residential address<span
data-label="fig:address_lhem_increases"></span>](address_lhem_increases)

Looking at the NO$_{2}$ map first (shown right), there appears to be
fewer people with an increased NO$_{2}$ LHEM exposure in the South-East
of London. Whether this is a interesting result and perhaps to do with
travel behaviour, or whether it is a result of the number of LTDS
subjects being less in that region, is examined by plotting out all the
respondants below in Figure \[fig:addresses\].

![The residential address of subjects whose exposure increased by using
the LHEM method compared to residential address<span
data-label="fig:addresses"></span>](addresses)

As can be seen, there appears to be slightly less people in the dataset
in the South-East of London, which may explain this pattern (and is
reflective of the general population of london
(<http://londondatastore-upload.s3.amazonaws.com/instant-atlas/borough-profiles/atlas.html>)).
There could also be additional factors as work, however this would
require much more detailed analysis of this aspect of the LHEM, and some
sort of regression analysis, so is not attempted in this introduction to
the model.

For the PM$_{2.5}$ map, it appears that the people living in Central
London are less likely to have increased PM$_{2.5}$ using the LHEM
exposure method compared to the residential method, than those living
further outside of central London. It seems likely that as high levels
of PM$_{2.5}$ are found on the London Underground, and that people
living in central London would not need to use the London Underground as
frequently or for as long time periods compared to those living in outer
London, that this is the reason for this geographical pattern. Although
as with the NO$_{2}$ discussion above, this is not investigated further
here and is merely suggested as an area for future exploration and as a
demonstration of the LHEMs capabilities.

### Pollutant correlation {#subsec:pollutant_correlation}

A further area that the LHEM may be of use in health studies is in the
separation of health effects from different pollutants. @Brunekreef2007
for example notes that many cohort studies have looked at the negative
health effects of NO$_{2}$, but he questions whether NO$_{2}$ is a
surrogate for other pollutants such as PM$_{2.5}$, which may be actually
causing the health effects. The studies he reviewed found it difficult
to look at the relative effects of each pollutant, as they are so
strongly correlated. Indeed, using the residential address method, our
exposure estimates of the LTDS subjects find that, as the other studies
do, NO$_{2}$ and PM$_{2.5}$ are well correlated with a Pearson’s R of
0.90 (95% CI 0.90 to 0.90) (shown left in Figure
\[fig:correlation\_comparisons\_no2\_pm25\] below). In contrast, using
the LHEM (Figure \[fig:correlation\_comparisons\_no2\_pm25\], right)
shows a much more complicated picture of the relationship/correlation
between NO$_{2}$ and PM$_{2.5}$, and produces a Pearson’s R of 0.66 (95%
CI 0.66 to 0.67). Though as the relationship is not of a linear fashion
using the LHEM, Pearson’s R is not a valid comparison any longer, and a
more complex statistical examination would be required - probably along
the lines of a regression model as briefly discussed in Section
\[subsec:geographical\_missclassification\] - but again that is not
attempted here.

![Daily mean exposure to PM$_{2.5}$ v. NO$_{2}$ using residential
address exposure method (left) and the LHEM (right)<span
data-label="fig:correlation_comparisons_no2_pm25"></span>](correlation_comparisons_no2_pm25)

This difference between using the LHEM and residential address estimates
is an important finding since it has the potential to separate the
health effects of NO2 and PM2.5, as part of any ongoing public health
research.

### Susceptible groups and exposure {#subsec:susceptible_groups}

Studies have shown that the elderly and children are more susceptible to
adverse health effects as a result of poor air quality than other age
groups (@Wang2015, @WorldHealthOrganization2013a). Given this increased
risk, the two tables below show summary statistics of exposure to air
quality using the LHEM (Table \[tab:lhem\_comparing\_age\_categories\])
by age-category, for PM$_{2.5}$ and NO$_{2}$, compared to using the
residential address method (Table
\[tab:address\_comparing\_age\_categories\]).

  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --
     **<span>Age category</span> & **<span>Mean</span> & **<span>Median</span> & **<span>I/Q Range</span> & **<span>5^th^ to 95^th^</span>\            
     & Child (5-17) & 13.53 & 13.63 & 13.02 to 14.14 & 12.03-14.62\                                                                                    
     & Young Adult (18-29) & 13.63 & 13.74 & 13.10 to 14.24 & 12.14-14.77\                                                                             
     & Adult (30-59) & 13.54 & 13.63 & 12.99 to 14.16 & 12.06-14.66\                                                                                   
     & Elderly (&gt;60) & 13.44 & 13.53 & 12.90 to 14.06 & 11.99-14.59\                                                                                
     & Child (5-17) & 34.21 & 34.44 & 30.73 to 37.96 & 24.69-42.14\                                                                                    
     & Young Adult (18-29) & 35.26 & 35.49 & 31.35 to 39.17 & 25.55-43.59\                                                                             
     & Adult (30-59) & 34.40 & 34.53 & 30.67 to 38.35 & 24.72-42.52\                                                                                   
     & Elderly (&gt;60) & 33.53 & 33.58 & 29.76 to 37.52 & 24.13-42.01\                                                                                
     **********                                                                                                                                        
  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --

  : PM$_{2.5}$ and NO$_{2}$ residential address exposure results by age
  category (n=45,079, concentrations in $\mu \text{g m}^{-3}$)<span
  data-label="tab:address_comparing_age_categories"></span>

  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --
     **<span>Age category</span> & **<span>Mean</span> & **<span>Median</span> & **<span>I/Q Range</span> & **<span>5^th^ to 95^th^</span>\            
     & Child (5-17) & 8.11 & 8.11 & 7.75 to 8.42 & 7.09-8.96\                                                                                          
     & Young Adult (18-29) & 8.92 & 8.39 & 7.91 to 9.03 & 7.2-13.12\                                                                                   
     & Adult (30-59) & 8.66 & 8.33 & 7.86 to 8.82 & 7.16-12.32\                                                                                        
     & Elderly (&gt;60) & 8.20 & 8.110 & 7.72 to 8.46 & 7.1-9.26\                                                                                      
     & Child (5-17) & 11.82 & 11.68 & 10.39 to 12.97 & 8.47-15.84\                                                                                     
     & Young Adult (18-29) & 13.74 & 13.21 & 11.31 to 15.71 & 9.06-19.66\                                                                              
     & Adult (30-59) & 13.53 & 12.99 & 11.10 to 15.43 & 8.83-19.64\                                                                                    
     & Elderly (&gt;60) & 12.17 & 11.77 & 10.36 to 13.43 & 8.34-17.25\                                                                                 
     **********                                                                                                                                        
  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --

  : PM$_{2.5}$ and NO$_{2}$ LHEM exposure results by age category
  (n=45,079, concentrations in $\mu \text{g m}^{-3}$)<span
  data-label="tab:lhem_comparing_age_categories"></span>

This next table (Table
\[tab:percent\_change\_comparing\_age\_categories\]) now shows
percentage change between the two exposure methods i.e. percentage
change between Tables \[tab:lhem\_comparing\_age\_categories\] and
\[tab:address\_comparing\_age\_categories\].

  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --
     **<span>Age category</span> & **<span>Mean</span> & **<span>Median</span> & **<span>I/Q Range</span> & **<span>5^th^ to 95^th^</span>\            
     & Child (5-17) & -40.06% & -40.50% & -40.48% to -40.45% & 41.06%-38.71%\                                                                          
     & Young Adult (18-29) & -34.56% & -38.94% & -39.62% to -36.68% & 40.69%-11.17%\                                                                   
     & Adult (30-59) & -36.04% & -38.88% & -39.49% to -37.71% & 40.63%-15.96%\                                                                         
     & Elderly (&gt;60) & -38.99% & -40.06% & -40.16% to -39.83% & 40.78%-36.53%\                                                                      
     & Child (5-17) & -65.45% & -66.09% & -66.19% to -65.84% & 65.69%-62.41%\                                                                          
     & Young Adult (18-29) & -61.03% & -62.78% & -63.92% to -59.89% & 64.54%-54.90%\                                                                   
     & Adult (30-59) & -60.67% & -62.38% & -63.81% to -59.77% & 64.28%-53.81%\                                                                         
     & Elderly (&gt;60) & -63.70% & -64.95% & -65.19% to -64.21% & 65.44%-58.94%\                                                                      
     **********                                                                                                                                        
  -- ---------------------------------------------------------------------------------------------------------------------------------------- -- -- -- --

  : PM$_{2.5}$ and NO$_{2}$ residential address exposure results by age
  category (n=45,079, concentrations in $\mu \text{g m}^{-3}$)<span
  data-label="tab:percent_change_comparing_age_categories"></span>

Firstly, these results show that there is more variation in exposure
between age groups when using the LHEM compared to using the residential
address method. Secondly, they allow us to consider the different
exposure of each age group from each method, and lastly they find
different patterns e.g. when using the LHEM exposure method, children
have the lowest exposure of all age groups to both PM$_{2.5}$ and
NO$_{2}$ (and young adults the highest) – however when using the
residential address method, the lowest exposure is found in the elderly
instead of the children.

Discussion {#sec:2Discussion}
----------

The LHEM exposure model developed in this chapter, using inputs from the
LTDS-X in the previous Chapter, calculates detailed spatial and
temporally defined exposure estimates on a level not seen in similar
studies. The detail in the inputs separates it in particular, such as
the number of participants, the time-activity and demographic detail for
them, hourly and spatially resolved CMAQ-Urban data, mass-balance
modelling of some of the transport modes, and a very detailed model of
indoor exposure. The closest study at this time is that of @Dhondt2012,
who modelled 8000 people, at 15 minute intervals, with only four
microenvironments, and less detailed spatial air pollutant inputs. The
results in Dhondt were quite different to the results presented here
however. For NO$_{2}$ they found that the static address-method was
underestimating by an average of 1.2%, unlike the LHEM which finds an
overestimation (13 $\mu \text{g m}^{-3}$ for residential address, and
8.48 $\mu \text{g m}^{-3}$ with the LHEM, meaning the address method is
overestimating). Interestingly @DeNazelle2013, discussed in Section
\[sec:dynamic\_hybrid\_models\] and in Figure
\[fig:time\_no2\_activity\_spaces\], found fairly similar NO$_{2}$
results to the LHEM in Barcelona in their measurements-based study. They
found that people spend 94% of their time indoors, and 6% in transport,
where they accumulated 83% and 7% of their exposure. Their study
population all worked at the University, and as such are likely to fall
in the 18-59 age-group categories, and therefore the LHEM results fit
very well with these modelled results.

As was demonstrated in the results section \[sec:2results\], the LHEM
exposure model can allow detailed investigation and calculation of
exposure and exposure missclassification. It was used to take a ’first
glance’ at the relative importance of microenvironments on exposure, the
calculation of exposure for different age groups, exposure
missclassification, highly exposed people, exposure peaks, whether there
were any geographical patterns to exposure missclassification,
correlating pollutants, and susceptible groups. Each of these areas was
only briefly explored to demonstrate the potential of the model, and
further analysis is needed. The model has other uses however, for
example each of the LTDS subjects has many more demographic attributes
that have not been explored such as ethnicity, income, gender, how many
cars the household owns, distance from tube stations etc. that all may
play an important role in better understanding exposure. The model can
also be used for policy applications which are not explored here, for
example what effect would changing 5% of the subjects journeys from car
to walking have on their overall exposure, and how would this vary by
age group?

There are a number of ways in which the LHEM could be improved and some
elements of the model that might be considered for change in the next
iteration. Firstly, the number of people could be increased, as extra
LTDS data is now available from TfL. Adding large numbers of people is
likely to increase the strength of the conclusions arrived at, however
as many of the exposure results are only marginally different including
new subjects may actually change this. It is difficult to say at this
stage. Another area that could be reconsidered is the use of the
building I/O ratios from @Taylor2014. This was an excellent dataset for
this research, and to my knowledge unparallelled at this time, however
it calculates average ratios on a postcode level, for residential
buildings, and therefore may not be suitable for modelling other types
of buildings such as office blocks. A new iteration of this dataset
might be an option, or indeed developing of an entirely new method. With
regard to the transport modelling, the use of static numbers for the
London Underground also seem far from ideal. The figures used for
PM$_{2.5}$ were taken from a small number of measurements on one stretch
of the Underground by researchers at KCL, when clearly the
concentrations will vary on different lines, in different sections of
those lines, and perhaps by time of day and season. For NO$_{2}$ the
concentrations were taken from a study in Paris, and face similar
problems about their applicability to London, and across the network.
Refining this aspect of the model will be the focus of the next Chapter.

Perhaps most importantly for the long-term future of this type of model,
and perceptions of the results that it produces, is how to validate the
estimates. The most likely route to do this would seem to be using
personal mobile monitoring, and then comparing this dataset to what the
LHEM outputs for similar journeys and days of the subjects.

Conclusions {#sec:2conclusions}
-----------

Results Section \[subsec:time\_exposure\_microenvironments\] showed that
the LTDS subjects spend most of their time indoors, and thus
understanding indoor exposure and being able to model exposure to indoor
pollutants (aswell as ingress from outdoor pollutants) is important in
understanding exposure in general. With the caveat that the CMAQ-Urban
input to the LHEM only models exposure to outdoor pollution, it finds
that people are exposed to around 85% of their daily NO$_{2}$ and 90% of
their daily PM$_{2.5}$ exposure while indoors - although this varied
slightly by age group, with children and the elderly accumulating more
of their daily exposure from the indoors environment that adults and
young adults, due to the slightly increased time that they spend
indoors, and naturally a reflection of this, the increased time that
adults and young adults spend in transport in environments of higher
concentrations that indoors. Comparing exposure in transport modes is
difficult, as except for the underground, they are a function of the
outdoor concentrations. However on a time/exposure basis, active travel
can be seen to result in lower exposure than passive travel.

The different exposure values found in results Section
\[subsec:comparing\_exposure\_methods\] should help epidemiologists
understand that the means and ranges of exposures that are currently
being used in health studies are perhaps not appropriate, and that they
may be over-estimating exposure across the population (and using
incorrect ranges). The LHEM finds that estimates based on monitoring
sites, postcodes or residential address are all overestimating exposure
for both PM$_{2.5}$ and NO$_{2}$, by varying degrees depending on the
method in comparison. Interestingly, when comparing postcode and
residential address estimates across the 45,079 people, they are found
to be almost identical. Given this, perhaps the recent drive for
individual residential addresses for completing exposure analysis is
unrequired, and postcode estimates are sufficient to reflect the
variation in exposure using static analysis methods (although studies
discussed in the Introduction typically took annual averages, rather
than the hourly variation that CMAQ-Urban models). When the LHEM
exposure estimates are plotted against residential address estimates,
with the points coloured by whether people left their house during the
day, this factor appears to be the main reason for the differences
between the LHEM and residential address method exposure estimates
(instead of perhaps the indoor-outdoor ratios, which would only result
in lower estimates but with the same overall profile). Introducing the
LHEM to exposure estimates creates peaks and ranges in a persons daily
exposure that are not seen using other methods.

In results section \[subsec:highly\_exposed\_people\], it was
demonstrated how the LHEM can investigate the numbers of people who are
accumulating daily mean exposures above the WHO limits for PM$_{2.5}$
and NO$_{2}$. As the static models of exposure do not take account of
the subjects movements, or normally the effects of exposure being
different when people are indoors, if a subjects house happens to be in
an area of very high concentrations then that persons mean exposure is
taken to be high. When actually the LHEM (compared to residential
exposure method) demonstrates that when the subjects daily activities
are taken into account, their daily mean exposure is actually
(generally) lower. The result of this with the LHEM, is that much fewer
subjects are found to be living in areas above the WHO limits.

As the LHEM calculates exposure on a minute-by-minute basis, Section
\[subsec:exposure\_peaks\_by\_age\_group\] showed how it can also be
used to consider much shorter periods of high exposure in a subjects day
that other models are generally unable to do. Using this aspect of the
LHEM, we find that the LTDS subjects are exposed to levels of
’unacceptable’ (as defined by the WHO) PM$_{2.5}$ and NO$_{2}$ levels
for between 13.1% and 16% and 1.3% and 3.8% respectively, depending on
age group.

By calculating the percentage difference between the residential address
method, and the LHEM method, we were able to see how for most people the
difference in exposure is between (approximately) -50% and -20%, but how
there were was a long-tail of distributions and some people were found
to have higher mean exposure estimates when using the LHEM. The people
with higher LHEM than residential exposure estimates were mapped, to see
if there was any specific geographical distribution to these results,
but none was apparent.

The LHEM was now used to see whether PM$_{2.5}$ and NO$_{2}$ exposures
were found to be correlated, as they are in many other health effect
studies (discussed in Section \[subsec:pollutant\_correlation\]. First
the residential PM$_{2.5}$ and NO$_{2}$ were checked for correlation to
see if the data and subjects in this research were similar to others,
and they were found to have a R^2^ of 0.90, confirming that they were.
The same analysis using the LHEM however resulted in an lower
correlation of 0.66 i.e. poorly correlated. However there actually
appear to be two different correlations within the scatter-plot, and
thus this requires further investigation. This difference between
pollutants using the LHEM may allow future health studies to better be
able to estimate the differences in health effects from individual
pollutants.

Finally the LHEM was used to investigate how exposure varies by age
groups in Section \[subsec:susceptible\_groups\]. It demonstrated how
each age group has fairly similar ranges and means of exposure, and
similar missclassification between the LHEM and residential exposure
estimates.

Exposure to on the London Underground {#chap:monitoring_on_underground}
=====================================

Aim {#sec:3aim}
---

Create an exposure model for PM$_{2.5}$ on the London Underground

Objectives {#sec:3objectives}
----------

-   Measure PM$_{2.5}$ across the London Underground network

-   Transcribe a time-location diary and link to pollutant data

-   Import other relevant data

-   Analyse data to characterise PM$_{2.5}$ on the London Underground

Background {#sec:3background}
----------

The London Underground (otherwise known as ’The Tube’) has around 402
kilometres of track covering the Greater London Area, around which 52%
is overground and 48% is underground across 11 lines
(@TransportforLondon2014a). The network currently has annual passenger
numbers of 1.305 billion, and is the main source of transport for the
population of London.

![A map of the London Underground<span
data-label="fig:tube_map"></span>](tube_map)

In the previous chapter (Section \[sec:the\_london\_underground\])
concentrations of 94 $\mu \text{g m}^{-3}$ and 51 $\mu \text{g m}^{-3}$
for PM$_{2.5}$ and NO$_{2}$ respectively were used as simple exposure
estimates to represent exposure while the LTDS-X subjects were
travelling on the London Underground network. These concentrations,
particularly for PM$_{2.5}$, represented some of the highest exposures
that the subjects encountered, way in excess of the concentrations found
at their residential address. However the concentrations used, taken
from the studies cited at the time, are the means of a wide range of
measurements. For example the PM$_{2.5}$ data that was collected by Dr.
Barratt (personal communication) upon which the mean of 94
$\mu \text{g m}^{-3}$ was based, has variation below and above this
mean. Taking a small sub-sample of a journey between Waterloo and Bond
Street on the Jubilee Line the PM$_{2.5}$ varied between 22
$\mu \text{g m}^{-3}$ and 140 $\mu \text{g m}^{-3}$ (1 minute averaging
time).

The evidence of variation in concentrations along the lines, supporting
the assertion that a more disaggregated method of estimating exposure on
the London Underground is needed, is further strengthened by the studies
discussed in Section \[sec:transport\] (Transport exposure) where very
different concentrations were found between studies and within studies.
Figure \[fig:pm\_tube\_summary\] summarised the concentrations across
various studies of underground train exposure (with results for
PM$_{2.5}$ on the London Underground being 202 $\mu \text{g m}^{-3}$ in
@Adams2001 and 246 $\mu \text{g m}^{-3}$ in @Pfeifer1999a). Specifically
within the Adams paper, a mean of 238.7 $\mu \text{g m}^{-3}$ for
PM$_{2.5}$ was found in lines below ground and a mean of 29.3
$\mu \text{g m}^{-3}$ on above ground lines suggesting that whether the
line is over or under ground contributes to levels of PM$_{2.5}$. Adams
also noted that there was no statistical difference between
concentrations at different times of day, or between the two seasons
when testing occurred (summer, June 1999 and winter, February 2000).

However these studies have tended to have (i) relatively short time
periods of measured concentrations (ii) study only small areas or fixed
sites of the network (iii) summarise by giving an overall mean and
confidence intervals, and (iv) are limited in their scope and attempts
to explore the variation of concentrations e.g. by mapping their data.
Additionally, the more comprehensive of these studies (in particular
@Hurley2003) have framed their findings in terms of exposure on the
underground network as an occupational hazard, for example comparing to
the exposure of welders, when as we know the actual people who are being
exposed to this air are from all ages, backgrounds and of varying
degrees of health.

The aim of this chapter is to provide a more detailed understanding of
pollutant concentrations on the London Underground. The focus will be on
PM$_{2.5}$ given that their are no obvious sources of NO$_{2}$ on the
London Underground and that concentrations in the literature are found
to be similar to ambient concentrations. Additionally, we did not
currently have any portable measurement equipment for NO$_{2}$
available. Specifically, measurements will be taken within the tube
train during journeys across the network, and then combined with a
manually completed diary which will note the section of track or station
that the concentrations were recorded at. As noted by Adams above,
whether a train is underground or overground may be important in
understanding concentration levels, so depth data will also be sourced
and joined to the existing data, and then further this data will be
joined to a geographic representation of the tube network. The result of
this research will be a geographically defined dataset of PM$_{2.5}$ on
the London Underground that can be used in modelling Londoners exposure
whilst travelling on the tube.

Methods {#sec:3methods}
-------

### Measurements

In attempting to better understand PM$_{2.5}$ levels on the tube, a
TSI-Sidepak was carried while sitting in a passenger cabin and
journeying around the network. One line was sampled per day (over a
number of months), with the aim being to cover every section and station
of the line at least twice. For the simpler lines, with no spurs, this
was a case of starting the equipment at one end, journeying to the
other, changing trains, and then making a return journey. However for
the more complex lines such as the DLR or the Central line where there
are many different spurs and sections of lines, a pragmatic approach was
taken whereby various sections were repeated to get as complete coverage
as possible, resulting in some sections being repeated more than twice.
Journey times are summarised in below.

  ----------------------------------------------- -- --
  **<span>Line</span> & **<span>Minutes</span>\      
  Victoria & 97\                                     
  Circle & 160\                                      
  Northern & 246\                                    
  Bakerloo & 116\                                    
  Jubilee & 253\                                     
  District & 201\                                    
  Piccadilly & 205\                                  
  Docklands Light Railway & 258\                     
  Metropolitan & 255\                                
  Central & 222\                                     
  ****                                               
  ----------------------------------------------- -- --

  : Time spent collecting air quality measurements on the London
  Underground, by line<span
  data-label="tab:time_on_the_underground"></span>

Consideration was taken of the possible causes of variation in the
concentrations, and how these might effect the results. When monitoring
concentrations of PM$_{2.5}$ near roads, and to a lesser degree away
from roads (background), there is normally a diurnal variation and
seasonal variation that is caused by emissions from increased traffic
and weather conditions respectively. This issue was taken to be of
negligible importance in our sampling of the London Underground, as the
concentrations seen in the pilot data and in previous studies have found
no evidence of diurnal variation. Supporting this approach to the
sampling is the work of Adams, also discussed in the introduction to
this chapter, where they found little difference between concentrations
in different seasons and times of day. Further, the effect of passenger
numbers and movement on concentration levels was taken to be negligible,
as particle concentration levels from this are insignificant in scale,
in the same manner (@Ferro2004a). Given this, re-suspension of particles
from the movements of the trains seem to be the main cause of elevations
in particles.

#### Equipment - TSI-Sidepak {#sec:equipment_tsi_sidepak}

A TSI-Sidepak (@TSI2015) was used to measure PM$_{2.5}$ on the London
Underground. The device uses a light-scattering technique, and is shown
below in Figure \[fig:sidepak\]. It weighs approximately 16 ounces, and
is 10.7 x 9.4 x 7.1 cm in size. Whilst in use the pump makes a low level
of noise, due to the pump sucking in the air. The device was placed in a
backpack, and an inlet tube connected and fed out the top of the bag.
For the sampling period, this backpack was then placed on the seat of a
carriage (or occasionally on the researchers knees when the carriage was
busy). This device was chosen for this purpose due to it’s small size,
ease of use, low level of noise, and use in other published personal
exposure studies (@Huang2015, @Han2015, @Yu2016). The time-resolution
for collecting data was set to one minute intervals, and at the end of
each sampling period the data was extracted using the TSI software, and
then loaded into a SQL database.

![A TSI-Sidepak for measuring PM$_{2.5}$<span
data-label="fig:sidepak"></span>](sidepak)

According to the TSI website @TSI2015, the sidepak is calibrated “*to
the respirable fraction of standard ISO 12103-1, A1 Test Dust (formerly
Arizona Test Dust) \[which\] allows comparisons between measurements
where the source or type of dust is predominately the same*”. Therefore
when this device is used it needs calibration factors calculating and
applying to the data, to accurately reflect the concentrations it is
recording. Doing so is relatively simple to do when placed alongside a
gravimetric measurement device, and is commmon in robust studies such as
@Torrey2015 where they found the Sidepak was overestimating by a factor
of about 1.3 and @Jiang2011 where the Sidepak was overestimating
concentrations by a factor of 3. Within London, Dr Barratt has
calculated a correction factor of 0.6 for use of the sidepak in outdoor
environments (personal communication, 2016). However to our knowledge no
correction factor exists for use in the London Underground, and as such
this needed calculating. Briefly, as part of a separate research project
at King’s, a TSI-Sidepak was placed in a small cabin on the platform of
Hampstead station (a Northern line station of the tube network). Also in
this cabin a ThermoFisher Partisol (@ThermoFisherScientific2016) was
installed and both instruments had inlets pushed through holes in the
ceiling of the cabin to sample the air over a three week period. The
process and results are described in full in (REF TO BEN/DAVE PAPER TO
COME), but in summary the result is that a correction factor of times
two should be applied when the device is sampling PM$_{2.5}$ in the
tube, to give a correct reading. Or rather, this correction should be
applied to the proportion of the PM that is attributable to the tube,
rather than external air (which should have the london correction factor
of 0.6 applied). This process is illustrated with some simulated data in
Figure \[fig:tube\_correction\_example\] below.

![Simulated tube data showing the proportion of the data that would be
scaled by 2.0 (green box) and the proportion of the data that would be
scaled by 0.6 (blue box) if the London background concentration at that
time was 7 $\mu \text{g m}^{-3}$ (red line)<span
data-label="fig:tube_correction_example"></span>](tube_correction_example)

To apply this scaling factor, the daily average London background
concentration of PM$_{2.5}$ was taken from the North Kensington
monitoring station (part of the LAQN) for each of the days that the
sampling occurred, and this scaling method was therefore applied to all
of the London Underground air quality data presented below.

### Tube diary {#tube_diary}

In order to map PM$_{2.5}$ concentrations across the tube network,
location data was needed to combine with the PM$_{2.5}$ data collected
by the TSI-Sidepak. Due to around 50% of the network being underground,
where GPS use is not possible, a diary was kept and then transcribed
into SQL code, before being loaded into the same database as the Sidepak
data. An example of the SQL code is shown below.

    INSERT INTO tube_diary VALUES('Bakerloo', 'Elephant & Castle', '2015-02-04 08:07:00', '2015-02-04 08:09:00', 'platform', 'floor', 1);
    -- Started tube journey North
    INSERT INTO tube_diary VALUES('Bakerloo', 'Elephant & Castle', '2015-02-04 08:07:00', '2015-02-04 08:07:00', 'tube', 'shelf', 2);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Lambeth North', '2015-02-04 08:13:00', '2015-02-04 08:13:00', 'tube', 'shelf', 3);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Waterloo', '2015-02-04 08:15:00', '2015-02-04 08:15:00', 'tube', 'shelf', 4);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Embankment', '2015-02-04 08:16:00', '2015-02-04 08:16:00', 'tube', 'shelf', 5);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Charing Cross', '2015-02-04 08:17:00', '2015-02-04 08:17:00', 'tube', 'shelf', 6);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Piccadilly Circus', '2015-02-04 08:19:00', '2015-02-04 08:19:00', 'tube', 'shelf', 7);
    INSERT INTO tube_diary VALUES('Bakerloo', 'Oxford Circus', '2015-02-04 08:21:00', '2015-02-04 08:21:00', 'tube', 'shelf', 8);

This location data was then linked to the Sidepak data by time and date.

### Station and line locations

To be able to investigate any spatial patterns in the data collected, we
needed to add spatial attributes to the data. Whilst the tube diary and
Sidepak data describe in text and numbers that, for example, the
PM$_{2.5}$ levels are 37 $\mu \text{g m}^{-3}$ on the stretch of line
between Aldgate East and Liverpool Street at 9:32am, we do not have the
geographical location of that stretch of track (or indeed the stations
at either end). A GIS network of the London Underground was therefore
manually created in a PostGIS database. The process is summarised below:

-   Download station locations (latitude/longitude/name) from TfL as a
    CSV file

-   Compare this list against a tube map to ensure 100% of stations
    were present.

-   Identify missing stations (14), use Google Maps to note their
    lat/long, and add these to the TfL CSV

-   Loaded CSV into database

-   Using a printed tube map, make a manual note of each section of
    track, including the stations that it joins, and the line of the
    track

-   Digitise this information to a CSV (start station coordinates, end
    station coordinates, line ID) and load into PostGIS

-   Use the PostGIS makeline() function to create a GIS linestring
    between each station for each line as appropriate

-   Visualise and error check

The key SQL code to create this is shown in Appendix . Loaded into QGIS
for visualisation the network is shown below in Figure
\[fig:postgis\_tube\_map\].

![PostGIS Tube Map<span
data-label="fig:postgis_tube_map"></span>](postgis_tube_map)

The map looks different to the traditional London Underground map due to
the lines between stations being straight, rather than adjusted for
cartographic purposes. The created lines are also not exactly
geographically correct, however it was decided that creating a GIS
dataset to represent the real-life complexity of the exact routes of the
lines was not needed for this work.

### London Underground station characteristics {#subsec:station_characteristics}

As briefly mentioned above, one of my hypotheses is that line and
station depths influence PM$_{2.5}$ concentrations on the network.
However the data to add a ’z’ attribute (depth) of each station was not
freely available via the London Datastore or similar data sources. I was
able to find a Freedom of Information (FOI) request made by Hamechan
Madhoo in January 2013
(<https://www.whatdotheyknow.com/request/depth_of_tube_stations_and_tube#incoming-366374>),
in which TfL provided this data. These data were downloaded, quality
checked, errors corrected, and then loaded as a CSV file into the
PostGIS database, and then matched to the existing data by station name
and line. Stations on the DLR are all above ground except for Bank, and
were missing from this dataset, so for simplicity they were all assigned
a depth z value of -10 i.e. 10 metres above ground.

### Trains {#subsec:trains}

Data regarding the types of trains that are operated on each line was
also sourced from from the TfL website (@TransportforLondon2016),
thinking it might be useful, and is summarised in Table
\[tab:train\_type\_on\_the\_underground\] below.

  ----------------------------------------------- -- --
  **<span>Line</span> & **<span>Train</span>\        
  Victoria & 2009 stock\                             
  Circle & ’S’ stock (2010)\                         
  Northern & 1995 stock\                             
  Bakerloo & 1972 stock\                             
  Jubilee & 1996 stock\                              
  District & ’D’ stock (1980)\                       
  Piccadilly & 1973 stock\                           
  Docklands Light Railway & ’B07’ stock (2005)\      
  Metropolitan & ’S’ stock (2010)\                   
  Central & 1992 stock\                              
  Hammersmith & City & ’S’ stock (2010)\             
  ****                                               
  ----------------------------------------------- -- --

  : Train type running on each tube line<span
  data-label="tab:train_type_on_the_underground"></span>

Results {#sec:3results}
-------

### Timeline concentrations {#subsec:timeline_concentrations}

To gain an initial understanding of the concentrations and variation
between each line, a time-line graph of all the data was created (Figure
\[fig:all\_lines\_pm25\]). The x axis was taken as minutes on the line,
hence the lines that had more monitoring completed on them (sometimes
due to the length of the line, sometimes due to availability of the
researcher) continue further to the right of the graph than others,
noting that there were repeats of each line section.

![Timelines of PM$_{2.5}$ on the tube<span
data-label="fig:all_lines_pm25"></span>](all_lines_pm25)

Although plotting all the lines on top of each other in this way makes
the data hard to interpret, it does immediately show the large variation
in some lines, and the lack of variation in others. There are clear
peaks and troughs, most apparent in the Northern line data , compared to
the DLR or Metropolitan lines which although still have variation, the
scale is much smaller. Figure \[fig:all\_lines\_individual\_pm25\_p1\]
below plots the lines individually to enable better understanding of the
patterns.

![Timelines of PM$_{2.5}$ on the tube (Note differing axis
scales)](all_lines_individual_pm25_p1 "fig:")
\[fig:all\_lines\_individual\_pm25\_p1\]

![image](all_lines_individual_pm25_p2)
\[fig:all\_lines\_individual\_pm25\_p2\]

We can see that the concentrations vary by line and within lines, for
example the Hammersmith & City line has a maximum of around 170
$\mu \text{g m}^{-3}$, compared to lines such as the Central line with
concentrations upto 500 $\mu \text{g m}^{-3}$ and the Victoria line of
upto 900 $\mu \text{g m}^{-3}$. Additionally, for some of the lines,
there are clear patterns in the data. The Jubilee line has four clear
peaks, and similarly the District line two clear peaks. Referencing
these against the diary information that was collected, the peaks
coincide with the train being underground, and then ’flat’ low levels of
PM$_{2.5}$, when the train was above ground and exposed to ambient air.

### Line averages {#subsec:line_averages}

Box and whisker plots using the ggplot2 package of R (@ggplot2) were
created in Figure \[fig:boxplot\_pm25\_lines\] compare concentrations
between lines.

![PM${2.5}$ $\mu \text{g m}^{-3}$ on the tube, summaried by line<span
data-label="fig:boxplot_pm25_lines"></span>](pm25_on_underground)

This visualisation gives a much clearer picture of the concentrations
and variations found between lines, and although the boxplot has
identified a number of concentrations on each line as being ’outliers’,
it is worth noting that this is likely not representative of bad data or
device error, it is actually that there are a number of places on some
of the lines with substantially higher concentrations that the median.
This is most apparent with the Victoria Line where there are
concentrations recorded of over 900 $\mu \text{g m}^{-3}$ compared to
the median of around 280 $\mu \text{g m}^{-3}$. It is interesting to
note how the Circle, District, Metropolitan, Docklands Light Railway,
and Hammersmith & City lines all have noticeably lower medians than the
other lines and a relatively small inter-quartile range. Given that the
environment of the Tube varies between fully exposed to the outside air
(in the manner of a normal train), to semi-covered, to fully
underground, it seems a reasonable premise that these different
environments are effecting the rises and falls in concentrations as per
the timelines in Figure \[fig:all\_lines\_individual\_pm25\_p1\]. This
was tested in Section \[subsec:concentrations\_v\_depth\]

### Concentrations v. Depth {#subsec:concentrations_v_depth}

To compare concentrations by line, the mean station depth of all
stations on a line was calculated and plotted against the mean
PM$_{2.5}$ concentration for that line (Figure
\[fig:concentrations\_depth\_summary\])

![Mean concentrations v. mean station depths by tube line<span
data-label="fig:concentrations_depth_summary"></span>](depth_pm25_summary)

Plotting the average line depths against average line concentrations
appeared to back-up that depth is an important determinant of levels of
PM$_{2.5}$. The shallower lines have the lowest average concentrations,
and the deepest lines the highest. To explore this further, figure
\[fig:depth\_plot\_per\_line\_pm25\] plots the concentrations for each
station of each line in the same manner. Each point is the mean of the
PM$_{2.5}$ concentrations recorded at that station.

![Concentrations recorded at stations v. station depth<span
data-label="fig:depth_plot_per_line_pm25"></span>](depth_plot_per_line_pm25)

The relationship between depth and concentrations that was apparent in
Figure \[fig:concentrations\_depth\_summary\] is not so clear in Figure
\[fig:depth\_plot\_per\_line\_pm25\], i.e., once the depth and
concentrations are plotted for each individual station as opposed to a
line average. Although there does still seem to be a relationship
between depth and PM$_{2.5}$, it does not seem to be such a
straight-forward one as increasing depth equals increasing
concentrations. Taking the District line as an example (Figure
\[fig:depth\_plot\_district\_line\]), the stations mostly either have a
depth of 5 metres above ground, or 5 metres below ground. The stations
above ground ($<0$ metres) tend to have low concentrations, but the
stations below ground ($>0$ metres) have both high and low
concentrations, highlighted with black and blue boxes below (Figure
\[fig:depth\_plot\_district\_line\]).

![Concentrations recorded at District line stations v. station
depth<span
data-label="fig:depth_plot_district_line"></span>](depth_plot_district_line)

Similarly taking the Central line (Figure
\[fig:depth\_plot\_central\_line\]), there is an increase in PM$_{2.5}$
as depth increases, except for a few outlier stations (highlighted with
a blue box). These two stations (Gants Hill and Wanstead) have a depth
of 18-20 metres, but concentrations of only 100 $\mu \text{g m}^{-3}$,
unlike other stations with similar depths where the PM$_{2.5}$ is in the
range 300-500 $\mu \text{g m}^{-3}$ (marked with a black box).

![Concentrations recorded at Central line stations v. station depth<span
data-label="fig:depth_plot_central_line"></span>](depth_plot_central_line)

Manually inspecting the depths of Wanstead and Gants Hill, and the
stations surrounding them, it is clear that they are in an area of the
Central line network where stations are shallow, and that they are an
exception. This could mean that shallow stations tend to be more well
ventilated due to natural air circulation, and that this cleaner air is
being moved down the tunnels to Gants Hill and Wanstead by the movement
of the train, or indeed that the train cabin is being ’flushed’ with
cleaner air at those station platforms and that it does not build back
up to higher concentrations by only going one stop. A plot of the
geography of the line, along with concentrations and depth was made in
Figure \[fig:fake\_central\_line\_pm25\] to better understand this
point.

![Central line locations, depths and PM$_{2.5}$<span
data-label="fig:fake_central_line_pm25"></span>](fake_central_line_pm25)

From the plot we can see that Wanstead and Gants Hill (towards the right
of the graph) are both medium depth stations (shown as red circles), and
as such we might expect them to have pollutant concentrations similar to
other medium depth stations such as Shepherd’s Busy, Holland Park,
Bethnal Green and Lancaster Gate. But the PM$_{2.5}$ levels are actually
more like that of above ground and shallow stations, such as those on
the section of the line where Wanstead and Gants Hill are located,
suggesting that depth of station and line is not directly related to
PM$_{2.5}$ concentrations, and that distance from outdoors or shallow
stations is important.

Taking a further example of stations and concentrations that do not seem
to fit the pattern of deep equals high, and shallow equals low, we can
see that there is a station on the Northern line which has a depth of
less than 0, i.e. above ground, but concentrations of over 200
$\mu \text{g m}^{-3}$. From manual inspection of the data, this station
is Golders Green, on the North-West spur of the Northern end of the
Northern line. The mean value of 200 $\mu \text{g m}^{-3}$ shown, is
actually an average of four four recorded values at that location, 356
$\mu \text{g m}^{-3}$, 14 $\mu \text{g m}^{-3}$, 374
$\mu \text{g m}^{-3}$ and 74 $\mu \text{g m}^{-3}$, which upon further
inspection relate to the train arriving at Golders Green having just
emerged from the tunnel and deep station of Hampstead (the higher
values), and the train having arrived from East Finchley (an outside
station further North on the track), which relate to the lower values.
The difference in these values is quite extreme and suggests that for
stations that are outside or shallow, but close to deeper stations, the
direction of travel influences the PM$_{2.5}$ concentrations for that
location.

As the tube arrives at Hampstead from within the deep tunnels further
South, the carriage must still be full of air that has accumulated over
the journey, which will get partly flushed out by the doors opening at
Golders Green, but not before the device (with one minute resolution)
has recorded high concentrations at Golders Green, which then drop by
the next station at East Finchley. Conversely, when the tube arrives at
Golders Green from East Finchley, the air inside the carriage is much
cleaner as it has not been deep underground previously. The
concentrations inside the carriage only start to rise to levels of 200 -
400 $\mu \text{g m}^{-3}$ once tube has gone into the tunnel at
Hampstead and started to be exposed to the higher concentrations.

To further explore this theory, the data for Oxford Circus on the
Victoria line was examined. There were four measurements taken while on
the train at that station, which were 140 $\mu \text{g m}^{-3}$, 322
$\mu \text{g m}^{-3}$, 152 $\mu \text{g m}^{-3}$ and 362
$\mu \text{g m}^{-3}$. The lower two measurements being just as the
train had arrived heading Northbound and Southbound respectively.
Direction of travel seeming to have have very little effect on the
concentrations. The higher two values, the second and fourth, being the
minute after the train had arrived , and presumably after the doors had
been opened and the air had exchanged with the platform air that had
been stirred up the train arriving.

In summary, from this small sample, it seems that PM$_{2.5}$
concentrations in the carriage at stations which are underground, and
are not close to a platform or section of line that is outdoor, are not
effecting by the direction of travel. Conversely, those that are near
outdoor sections of line and platforms, are. The effect of this
variation is partly minimised due to our study design, i.e. we measured
each station arriving and departing from different directions.

### Spatial distribution of tube air quality {#subsec:spatial_distribution_tube_air}

Figure \[fig:tube\_map\_pm25\] shows all the tube stations in London and
their location, with the size of the point used to indicate the mean
levels of PM$_{2.5}$ recorded i.e. larger circles on the map, show
higher concentrations.

![Locations of tube stations and PM$_{2.5}$ levels<span
data-label="fig:tube_map_pm25"></span>](tube_map_pm25)

As expected from the results of the previous sections, there is a
relationship between depth and PM$_{2.5}$ concentrations, with stations
in more central London having higher concentrations, due to the lines
being deeper and more frequently underground than in outer London.
However as noted in at Golders Green, taking means at stations can hide
important variation, so whilst this figure is useful for giving a
general impression of the spatial variation of concentrations, a more
complicated approach might be more useful for modelling exposure.

### PM build-up and dissipation {#subsec:tube_air_build_up}

In Figure \[fig:all\_lines\_individual\_pm25\_p1\], where timelines of
PM$_{2.5}$ are shown on each line, it is possible to see how on certain
lines and at certain places concentrations fell to levels similar to
background concentrations. To investigate how quickly the air in the
carriage falls to these levels for passengers the Jubilee line was taken
as an example (due to it’s clear differences between high and low
concentrations) and examined in more detail. Figure
\[fig:all\_lines\_individual\_pm25\_p1\] has been re-created and
annotated as Figure \[fig:decay\_time\_jubilee\] below. In addition to
the PM$_{2.5}$ concentrations during journeys, the sections where
concentrations fall down to around background PM$_{2.5}$ levels (red
line, taken from the ’Kensington and Chelsea - North Ken’ background
monitoring site) have been highlighted with a black rectangle. To aid
interpretation the station names have also been added for those
sections.

![Locations of tube stations and PM$_{2.5}$ levels<span
data-label="fig:decay_time_jubilee"></span>](decay_time_jubilee)

Taking the first and third highlighted sections above, concentrations
build-up inside the cabin while the train is underground, but then start
to fall after Canada Water, reaching background levels by the time the
train is at West Ham. The depth of these stations are, in order, Canada
Water (18m), Canary Wharf (18m), North Greenwich (15m), Canning Town
(2), and West Ham (0m). These stations do not have the step-by-step
change in depth in the way that the concentrations do. If anything, the
first three stations in this subset might be grouped as deep, and then
the final two as surface. However the concentrations do not immediately
change from high to background, they gradually decline. This suggests
that in addition to depth being an indicator of concentrations inside
the tube trains, distance from cleaner outside air, and it’s exchange
with air inside the cabin when the doors open, also effects
concentrations. To elaborate, the concentrations between Canada Water
and North Greenwich fall by about 40%, despite there only being a small
change in depth. I suggest that the change in concentrations is due to
the doors opening at North Greenwich, and an air exchange happening with
the air on the platform, which is cleaner ’platform air’ than is the
case at Canada Water, due to North Greenwich being closer to a surface
station (Canning Town). Now taking the second and fourth sections
highlighted in Figure \[fig:decay\_time\_jubilee\] the stations under
consideration are Swiss Cottage (17m), Finchley Road (3m), West
Hampstead (5m), Kilburn (7m) and Willesden Green (5m). Here we see a
similar pattern, in that the concentrations do not immediately drop to
background levels in the way that might be expected if depth was the
sole determinant. It takes a couple of stations (and subsequent doors
opening and air exchange) for the air in the train cabin to be
sufficiently replaced with cleaner outside air.

### Train stock {#subsec:tube_train_stock}

Mean concentrations per line were now calculated (in the same manner as
section \[subsec:line\_averages\]), but rather than grouping by line,
the train stock data was linked, and boxplots created with that as the
categorical variable (Figure \[fig:pm25\_by\_stock\])

![PM$_{2.5}$ $\mu \text{g m}^{-3}$ on the tube, summaried by stock
type<span data-label="fig:pm25_by_stock"></span>](pm25_by_stock)

There does not seem to be any clear pattern to PM$_{2.5}$ concentrations
when examining the data using train stock as a variable. Excluding the
B07 2005 stock (as they are soley on the DLR), the oldest trains (D 1980
stock) have the lowest concentrations, and the 2009 stock the highest.
But they are only used on the District line and the Victoria lines
respectively, and as we have seen there are large variations within the
length of those lines which suggest that there are other factors (namely
depth and distance from exposed station platforms) which are influential
and not linked to train stock.

### Revising LHEM exposure estimates {#subset:revising_lhem_exposure}

As discussed at the beginning of this chapter, whilst travelling on the
London Underground network the LTDS-X subjects were assigned PM$_{2.5}$
exposure concentrations of 95 $\mu \text{g m}^{-3}$, per minute. We can
now see that this is a simplistic representation of the variability
within the network. A final aim for this area of research is to create a
detailed spatial GIS model layer which can be used for exposure
assessments of the population of London while on the tube, however as an
intermediary step, two LTDS subjects who used the London Underground
during their day were chosen at random, and their exposure recalculated,
but with mean line concentrations taken from this new data, to give a
small example of the possible effects. Their LHEM exposure before this
new method is shown in Figure \[fig:lhem\_exposure\_timeline\], with the
exposure of 95 $\mu \text{g m}^{-3}$ while the subjects are on the tube,
shown in blue.

![PM$_{2.5}$ $\mu \text{g m}^{-3}$ exposure for LHEM subjects
60601534101 and 70737511101<span
data-label="fig:lhem_exposure_timeline"></span>](recalculating_lhem_pre)

The mean daily exposure for these two subjects was 10.06
$\mu \text{g m}^{-3}$ and 12.69 $\mu \text{g m}^{-3}$ respectively. We
now manually inspect the routes that these two subjects took their
journeys on, and substitute mean line concentrations from this new
dataset. Subject 60601534101 begun their journey at Finsbury Park, and
ended it at Arnos Grove, taking the Piccadilly line. The mean PM$_{2.5}$
recorded on the Piccadilly line was 63 $\mu \text{g m}^{-3}$, and this
is therefore substituted instead of the 95 $\mu \text{g m}^{-3}$ used.
For subject 70737511101, they began their journey at Rayners Lane, and
ended it Canning Town, taking the Metropolitan line for approximately
the first third of their journey, and the Jubilee line for the remaining
two thirds. So the first third of their journey the mean Metropolitan
line concentration of 67 $\mu \text{g m}^{-3}$ is used, and for the
second two thirds of their journey the mean Jubilee line concentration
of 138 $\mu \text{g m}^{-3}$ is used. The new timelines are shown below
in .

![Revised PM$_{2.5}$ $\mu \text{g m}^{-3}$ exposure for LHEM subjects
60601534101 and 70737511101<span
data-label="fig:lhem_exposure_post_timeline"></span>](recalculating_lhem_post)

The new mean daily exposure for these two subjects was 11.73
$\mu \text{g m}^{-3}$ and 12.99 $\mu \text{g m}^{-3}$ respectively, an
increase of 17% and 2.4%. The former being higher, as although the
journey was shorter the concentrations were much higher for that period
than previously modelled. Alongside this increase in mean exposure, both
subjects for a short period are also exposed to higher peaks in
concentrations than in the old method.

Discussion {#sec:3Discussion}
----------

This dataset is, to the best of my knowledge, the largest, most
systematic, and detailed collection of PM$_{2.5}$ concentrations on the
London Underground. There are few studies which have studied PM$_{2.5}$
on the tube, however in a review of exposure on metro systems in 2007
@Nieuwenhuijsen2007 collated various studies, and specifically within
London found ranges of between 130–200 $\mu \text{g m}^{-3}$, 157–247
$\mu \text{g m}^{-3}$, and 12–264 $\mu \text{g m}^{-3}$, mostly from the
@Adams2001a studies which completed around sixty journeys on the tube in
2001. The data collected for this chapter is superior as the Adams work
only collected data for short sections of repeat journeys on the
Piccadilly, Bakerloo, Northern and District lines for comparison with
other transport modes, and no spatial analysis was undertaken (or indeed
possible as the data was collected on filters and therefore lacked
spatial and temporal granularity).

In addition to peer-reviewed academic work, the other main source of
data on tube PM$_{2.5}$ measurements is from occupational health work
commissioned by TfL, the most well know being the “Assessment of health
effects of long-term occupational exposure to tunnel dust in the London
Underground” led by Hurley and published in 2003 (@Hurley2003). This was
commissioned with the purpose of looking at occupational exposure to
drivers and station staff, with only a small section concerning
passengers, and therefore most of the results focus on the personal
exposure of drivers and staff. There is no consideration of the
variability by line, or within line, or an attempt at understanding the
viability. In summary, the work completed and data collected is
appropriate for for the purpose it was commissioned for, but not there
are no spatial or temporal attributes linked to the data to enable
further investigation or to use the results in other ways. Nonetheless
the PM$_{2.5}$ levels were found to be in the range 270–480
$\mu \text{g m}^{-3}$ on station platforms, and passengers average
expose was taken as 200 $\mu \text{g m}^{-3}$, although a number of
broad assumptions were made to arrive at this figure. As with
@Nieuwenhuijsen2007, these concentrations are not dissimilar to those we
measured.

By collecting PM$_{2.5}$ concentrations and linking them to
time-resolved location data across the whole of the network, and further
calibrating the PM$_{2.5}$ measurements using newly calculated scaling
factors, we have been able to offer the most complete understanding of
the variation and levels of pollutant in this environment that is used
by millions of people everyday. There are however some issues and
considerations with the methods and data collection that need
discussion, and which might effect our findings.

Firstly, all of our measurements are taken from within the carriage of
the tube. So whilst there is often discussion of stations within this
chapter, this is actually the tube train (normally with doors open)
pausing at the platform of a station for a minute or two before moving
away again, and the reader must be careful not to misinterpret the
findings as such.

Another area that might warrant further refinement is around the effect
that passenger numbers have on concentrations. The theory of the
’personal cloud’, that being that a persons movements and activity can
stir-up particles into the air that otherwise may have settled on
surfaces. With so many people moving inside trains and on the platforms
of the stations, this might contribute to increased concentrations
independently of other factors i.e. the movement of the trains. This
said, studies (@Ferro2004a) have found that this personal cloud effect
to only increase concentrations by a few $\mu \text{g m}^{-3}$, which
when set alongside the 100s of $\mu \text{g m}^{-3}$ being recorded are
quite insignificant, and thus I feel that the effect of passenger
numbers can be largely ignored in efforts to understand PM$_{2.5}$
levels in this environment.

Similarly, @Adams2001a discusses that wind direction and outdoors
concentrations may be linked to increased concentrations in the tube
system, with particles being blown into the tunnels and recirculating.
But with background concentrations of PM$_{2.5}$ in London normally
around the 10–15 $\mu \text{g m}^{-3}$ mark, it seems unlikely that
pollutants from outside the system are contributing in a meaningful way
to the high concentrations found (when the tube is in the underground
environment anyway).

Regarding the findings related to depth, these must also be considered
alongside the fact that the depth data obtained and used in this
research related to station platform depths, and was not detailed enough
to enable understanding of depth between stations. This would have been
useful particular for Section \[subsec:tube\_air\_build\_up\] where the
build-up and dissipation of PM$_{2.5}$ between stations was considered.
Alongside monitoring equipment with a higher time resolution, say 3-4
seconds rather than 1 minute, it would be easier to understand how the
concentrations vary between stations and along stretches of track.

A further complication to the findings we have so far, is that of
ventilation settings and train stock. Some of the lines only have one
stock-type running on them, but some lines have a variety. Also within
these varieties different ventilation settings are available, and vary
as to whether they are in use/functioning or not. It would be useful to
repeat a sample of this data collected with a specific focus on
ventilation to understand this variable more.

The data collected in this study is an important step forward in better
estimating the exposure of people who live and commute in London (as was
demonstrated with two random LHEM subjects). With further work, a
dataset can be created which will allow vastly improved estimation of
the exposure of millions of people travelling on the tube, which in most
exposure studies involving London is not currently considered at all.
Epidemiological studies are still geocoding peoples address or perhaps
postcodes, and then taking annual concentration maps, normally with
small ranges between the maximum and minimums, when there are millions
of people every day who are spending prolonged periods of time in
environments that have PM$_{2.5}$ concentrations of over 500
$\mu \text{g m}^{-3}$. It might be that the there are strong links
between people who use the London Underground for more than an hour a
day, and those who develop COPD issues. At the moment there has been no
way to investigate this on a population level, but this research makes
developing a research question about this subject now possible.

Future work in this area should focus on creation of a geographically
defined data layer or database which can be used alongside passenger
modelling (such as done by a Dr Reades, a colleague in the Geography
department of King’s: <https://www.youtube.com/watch?v=F6qsh1KBW-E>).
This dataset should be sufficiently geographically defined, to have
different exposure estimates related to each section of each tube line,
rather than the intermediate step used in Section
\[subset:revising\_lhem\_exposure\] whereby line means were calculated.
Preferably it should also take into account the direction of travel, as
this was seen to alter PM$_{2.5}$ exposure dramatically. One option
would be to create a Tube network within a pgRouting database, and
assign the concentrations as ’cost parameters’ to each stretch of the
network. Once built this would allow input of an origin and destination,
and then the algorithms would calculate a route between the two stations
that minimised exposure, with an output of exposure along the route.
Although the LHEM currently makes use of the TfL routing API for this
step of the model, and therefore a layer that can be easily incorporated
to this step would be a more appropriate approach. Each minute that an
individual is on the Tube would be a location and timestamp, and also a
field containing the name of the Tube line the journey is on, so these
coordinates could be ’snapped’ to the nearest Tube segment, for the
appropriate line of travel, and then the relevant concentrations
extracted from the line data and assigned to that minute of travel.

Conclusions {#sec:3conclusions}
-----------

PM$_{2.5}$ data collected on the London Underground varies between 0
$\mu \text{g m}^{-3}$ (or rather, below the limit of detection for the
instrument) and 990 $\mu \text{g m}^{-3}$, with a mean of 129
$\mu \text{g m}^{-3}$ and a median of 63 $\mu \text{g m}^{-3}$.

There is a large variation when comparing lines against each other, and
between sections of the same line. The Victoria Line for example has the
highest PM$_{2.5}$ concentrations recorded, of 990
$\mu \text{g m}^{-3}$, but a mean of 436 $\mu \text{g m}^{-3}$ and a
minimum of 66 $\mu \text{g m}^{-3}$.

When ranked in decreasing order of mean PM$_{2.5}$ concentrations, the
results are; Victoria Line (436 $\mu \text{g m}^{-3}$), Northern (219
$\mu \text{g m}^{-3}$), Piccadilly (199 $\mu \text{g m}^{-3}$), Bakerloo
(164 $\mu \text{g m}^{-3}$), Central (119 $\mu \text{g m}^{-3}$),
Jubilee (115 $\mu \text{g m}^{-3}$), Metropolitan (67
$\mu \text{g m}^{-3}$), Circle (41 $\mu \text{g m}^{-3}$), District (40
$\mu \text{g m}^{-3}$), Hammersmith & City (36 $\mu \text{g m}^{-3}$),
Docklands Light Railway (18 $\mu \text{g m}^{-3}$) . Stations and
sections of lines also have large variation. The Southern section of the
Victoria line had some of the highest concentrations, particularly on
the stretch of track between Brixton, Stockwell and Victoria with
concentrations of 400-500 $\mu \text{g m}^{-3}$, compared to many
stretches of track on more exposed lines such as the Circle, Hammersmith
& City and the Metropolitan lines were concentrations were less than 10
$\mu \text{g m}^{-3}$

Increasing depth was considered to be an indicator of increasing
PM$_{2.5}$ concentrations, illustrated by the linear relationship seen
in Figure \[fig:concentrations\_depth\_summary\] which looked at average
concentrations by line and plotted them against average depth. However
when this plot was disaggregated and considered by individual line,
station, and direction of travel, this relationship became more complex.
There were high concentrations recorded while on the tube at stations
that are not particularly deep, and conversely low concentrations
recorded while on the tube at stations that are very deep. The
explanation for this variation is due to the environments immediately
previously experienced by the train, and distance from those
environments. Trains that have just exited areas of high concentrations
bring polluted air with them which takes time to dissipate, and
conversely tube trains that have been in cleaner air take a little while
for concentrations in the carriages to increase. This increase and
decrease seems likely to mainly happen at platforms when doors are
opened and closed, but also to a lesser degree during movement of the
train (which needs further investigation).

The spatial representation of the stations in London, ignorant of
direction of travel and depth data, showed that higher higher
concentrations tend to be found in central London areas.

Finally, by calculating simplistic line averages in lieu of a more
complex dataset that will be created in the future, two randomly chosen
LHEM daily exposure estimates were recalculated. Both subjects daily
exposure was increased, one by 17% and one by 2%. Repeating this new
method across the entire tube-using population of London is likely to
lead to a general increase in exposure from the LHEM, but with some
people experiencing lower overall exposure (due to only travelling on
section of the tube that are cleaner than previously presumed).

Evaluating dynamic exposure models {#chap:evaluating_dynamic_exposure_models}
==================================

Aim {#sec:4aim}
---

Develop an understanding of methods to evaluate predictions of exposure
from hybrid models

Objectives {#sec:4objectives}
----------

-   Establish how to use mobile monitoring equipment to replicate static
    monitoring site datasets (which are used for air quality
    model evaluation)

-   Collect mobile monitoring data representative of a journey from the
    LHEM

-   Model exposure of this journey using LHEM methodology

-   Compare the monitored and modelled exposures

Background {#sec:4background}
----------

The focus of this PhD research so far has been on developing a dynamic
exposure model to better understand exposure to urban air pollution in
the population of London. Having reconstructed the time-activity of the
population, their exposure to PM$_{2.5}$ and NO$_{2}$ was modelled, and
then refined, with further investigation of the London Underground
micro-environment. This next chapter will consider how to evaluate the
NO$_{2}$ results that are calculated using an exposure model of this
style.

I defined the features that a hybrid exposure model should have in
Section \[sec:dynamic\_hybrid\_models\] as *“It should have highly
temporal and spatially resolved air quality inputs which consider both
indoor and outdoor sources (including regional and local source for the
latter), it should be able to model infiltration rates for different
modes of transport and building types, it should reflect the multiple
micro-environments that people spend their time in (and take account of
the temporal resolution of these) and finally it should (for linkage
through to epidemiological end-points) be able to consider different
breathing rates to quantify exposure and dose for multiple pollutants”*.
Section \[sec:dynamic\_hybrid\_models\] examined models that were within
this wider field (of varying levels of complexity), but it was
noticeable that there was little evaluation of the exposure predictions
that they made. There are to my knowledge no established protocols for
evaluating exposure predictions from a hybrid/dynamic exposure model, as
this type of method and field of research is relatively new. In
addition, as the field grows the exact approaches are being refined and
vary between studies, meaning one evaluation method would unlikely be
fit for the next study. Possible sources of error in this type of model
are classified in Table \[tab:exposure\_error\_table\], with a brief
description, and whether they are unique to exposure models of this
kind.

  **Type of error**                                            **Description**                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   **Hybrid specific**
  ------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
  Air quality annual average monitoring site predictions       Evaluation exercises of air quality models using high quality monitoring site data demonstrate that they (in our case CMAQ-UK) can make predictions of annual averages of most pollutants with reasonable accuracy.                                                                                                                                                                                                                                                                                                                                                                                                                                                                               No, CMAQ-UK and other similar models are commonly used in static exposure studies.
  Air quality annual average non-monitoring site predictions   How air quality models predict concentrations in locations that are not readily available for evaluation via monitoring sites is not well understood.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             No, CMAQ-UK and other similar models are commonly used in static exposure studies.
  Air quality temporal resolution                              Annual averages are often used for exposure studies, which are relatively easy to quantify against monitoring sites (see above). However the complexity of the hybrid model we are now considering uses hourly diurnal profiles, which vary in their accuracy of prediction over time, and therefore the accuracy of their input to exposure varies by time of the day, and day of the week.                                                                                                                                                                                                                                                                                                      Partly. Not many static studies of large groups of people consider exposure at high temporal resolution (because they normally use annual averages), and even less quantify and include the errors that this produces.
  Micro-environmental modelling                                Mass-balance models and I/O ratios to estimate the concentrations of pollutants within microenvironments, in relation to outdoor concentrations, are inherently subject to variation due to the inexact inputs. Literature reviews to establish ‘best-guess’ I/O ratios are common but their transfer-ability to other counties/cities is often unknown. Monte-Carlo simulation of the inputs to create a range of predictions can be undertaken to understand their impact.                                                                                                                                                                                                                      Yes. Static exposure studies normally use outdoor concentrations for exposure assessment. Exposure assessments of the health effects of environments i.e. indoor, may use micro-environmental modelling and I/O ratios, but not in conjunction with people’s movements, multiple environments, and high spatial-temporal air quality.
  Temporal representative errors of exposure                   Exposure predictions for a person over a time period can be made, but the degree to which these predictions are representative of that person’s exposure for that period of time are unknown. In the LHEM the exposures were calculated based on the person’s previous days’ movements, and the respondents were asked whether this was representative of their typical day; but quantifying the difference between the day of the data collected and their ‘typical’ day in terms of exposure is not explored.                                                                                                                                                                                   Yes. Hybrid exposure studies that consider the exposure of individuals through space-time, especially those that seek to frame exposure results in terms of longer-term health effects, need to develop methods to estimate the variability of representativeness error.
  Representative errors of groups and populations              Extrapolating exposure predictions from a small group of people (e.g. a classroom of 30 ten year olds in South London), to larger groups of people (e.g. all school children in London) can be controlled in part by statistical sampling techniques and appropriate power calculations. But this can only be done with prior data/knowledge of the population, which while fairly simple to do for basic demographics using Census data and similar, is much more difficult to do for exposure prediction models. To ensure representativeness exposure sample calculations need to ensure that important drivers of exposure i.e. tube usage, are included alongside more standard variables.   No, other exposure methods have this issue, but it is often not addressed fully. Studies with larger numbers of participants would be expected to have a better range of exposures and be more representative.

  : Errors and uncertainty in exposure models<span
  data-label="tab:exposure_error_table"></span>

The next sections of the background to this chapter are discussed around
the types of uncertainty from Table \[tab:exposure\_error\_table\]
above.

### Air quality annual average monitoring site predictions {#air_quality_annual_average_predictions}

Air quality models used in hybrid exposure studies are often evaluated
against annual averages from monitoring sites in the cities of the
study. For example @DeNazelle2013 study in Barcelona used a dispersion
model (@Lao2011) for the NO$_{2}$ air quality inputs to their exposure
model, and showed good performance/agreement (Table
\[tab:fig:de\_nazelle\_model\_performance\]) with errors in the range of
-12% to +10%.

![Performance statistics of an air quality model used in
@DeNazelle2013<span
data-label="tab:fig:de_nazelle_model_performance"></span>](de_nazelle_model_performance)

By understanding the scale of these errors in the air quality model, if
it is presumed that they are uniform in time and space across the study
area (discussed more below), then they can be incorporated into a static
exposure study of outdoor air at household addresses or similar.

### Air quality annual average non-monitoring site predictions {#air_quality_annual_average_non_site_predictions}

The predictions of air quality models, away from monitoring sites, is
less well understood. Taking our 2016 CMAQ-UK model the NO$_{2}$ map of
London shown in Figure \[fig:cmaq\_and\_monitoring\_sites\] (below)
illustrates the location of monitoring sites i.e. where the predictions
have been evaluated and understood. Clearly there are many areas which
are not evaluated; there are 16 million cells in the raster, and only
242 sites (many of which are either not operational, do not have a high
enough data capture rate for comparison, or do not measure the
pollutants we are interested in, meaning the actual number is far less).
This means that only  0.0015% of the locations shown in Figure
\[fig:cmaq\_and\_monitoring\_sites\] have been evaluated and the
possible error (between model prediction and measurement) understood.

![CMAQ-UK for London + Air Quality monitoring sites<span
data-label="fig:cmaq_and_monitoring_sites"></span>](images/cmaq_and_monitoring_sites.png)

For these non-monitoring site locations, we presume that the model
functions in the same manner, with the same error levels. In reality it
seems likely that there are varying degrees of accuracy. This may be
less of a concern within an exposure model looking at health effects
between areas of high/low concentrations, providing it represents the
spatial pattern adequately i.e. predicting lower concentrations and
higher concentrations in the right places, even if the concentrations
themselves are not exactly right. As part of this chapter I am going to
aim to better understand how the model predicts away from monitoring
sites. It is impractical to place a monitoring site at every location,
so I will develop a method whereby I use mobile monitoring devices
in-lieu of monitoring sites. As background to this, I searched for
studies which have already attempted this. There is a small amount of
literature which has tried to evaluate air quality away from monitoring
sites using mobile monitors. @Buteau2017 compared a number of methods
for predicting concentrations of O3 and NO$_{2}$ in Montreal, Canada at
the level three-character postal code area ( 6km$^{2}$ square).
Alongside traditional spatial interpolation using inverse-distance
weighting, and nearest monitoring site methods, they also based a
land-use regression model upon a dense monitoring survey. However, these
were semi-permanent monitors at fixed sites for 9 months of a year, and
so whilst their results were encouraging the method is not really
transferable to portable monitors that I intend to use. @Shi2016 also
created a land-use regression model, with mobile measurements as an
input, which achieved good results, but their study was based on
repeated measurements using vehicles sampling along a set route for 14
consecutive days, and the choice of 14 rather than say 20 seemed to be a
function of their available sampling time rather than calculated using
statistical analysis.

### Air quality temporal predictions {#air_quality_temporal_predictions}

Air quality models that use a higher temporal resolution have more
sources of error to quantify and include in an exposure model than one
which uses annual averages. As an annual average value, a difference of
+/- 5 $\mu \text{g m}^{-3}$ between the model and the measurements can
be factored into the exposure prediction relatively easily. But if the
model is predicting a concentration every hour then this error will vary
for each hour i.e. 8760 hours resulting in 8760 model/measurement
differences.

### Micro-environmental modelling {#microenvironmental_modelling}

Within dynamic exposure models, exposure in microenvironments is often
calculated in relation to outdoor concentrations. Therefore, in addition
to the uncertainty of the air quality predictions in a place and time,
further uncertainty is added when the outdoor concentration is
‘converted’ to a microenvironment concentration (whether this is done
using an I/O ratio for indoor concentrations, or a mass-balance model as
used in Section \[sec:in\_vehicle\_modelling\]). Exposures in the LHEM,
for all environments except the London Underground, used the CMAQ-UK air
quality model as an input. @Dhondt2012 described the possible sources of
error within the in-vehicle section of their dynamic model, and noted
that the air quality inputs may not be of a high enough spatial
resolution to capture concentration gradients near roads adequately. But
they did not seek to evaluate the individual-level daily/weekly/hourly
concentrations, and indeed due to the type of the study (an aggregated
population) they would have found it difficult to do so (due to the
large numbers of individuals). Similarly and whilst this was evaluated
at monitoring sites (Figure
\[tab:fig:de\_nazelle\_model\_performance\]), no monitoring attempt at
individual level exposure evaluation was undertaken.

### Representative errors {#representative_errors}

Exposure predictions for an individual or group, over a time period, can
be made using dynamic models. But the degree to which these predictions
are representative of that exposure for that period of time is a source
of uncertainty. Similarly, the degree to which predicting exposure for a
group or groups of people, and extrapolating to the wider population, is
a further source of confusion and error. In the LHEM, the exposures for
each individual were calculated based on the person’s previous days’
movements, and they were asked whether this was representative of their
typical day. Analysing the responses to the latter question showed us
that most of the subjects had a fairly set pattern of movement (and
therefore likely exposure), but the degree of the variance between their
typical day and a non-typical day is unknown, and indeed how many
‘non-typical days’ do they have? One way to evaluate the exposure
predictions from our model would have been to distribute personal
monitors to all the LTDS participants, for an entire year, collect and
process the data, and then compare it to the LHEM predictions to see
whether our ‘snap shot’ day exposure was close to their annual average
day exposure. However, there were  45,000 subjects, meaning a huge
number of personal monitors and logistical support would have been
needed, and it was therefore unfeasible. A method to make this more
manageable might be to develop a statistical model whereby an
appropriate sample-size is calculated. That is, just sampling a
percentage of the 45,000, but in the knowledge that they will have a
similar distribution of exposures to the total population of the study;
then giving this smaller subset of participants monitors for a prolonged
period of time and understanding the daily changes in their exposure.
This is explored more in Methods (Section \[sec:4methods\])

Methods {#sec:4methods}
-------

We can now see that there are a number of difficult and challenging
sources of error to consider in evaluating a hybrid exposure model. Air
quality models have quantifiable errors in their predictions at the
monitoring sites they are evaluated against, errors at locations away
from monitoring sites which are not well understood, further variation
and errors when using a high time resolution air quality model, errors
in the micro-environmental modelling, and more errors in terms of the
representativeness of the exposure predictions arrived at. The following
methods section outlines how I attempt to mitigate some of these sources
of error, remove some of the others, and produce an evaluation of one
short journey; comparing the LHEM prediction of this journey to measured
exposure using personal monitoring. The process is summarised below in
Figure \[fig:comparison\_method\_summary\].

![Method summary<span
data-label="fig:comparison_method_summary"></span>](images/comparison_method_summary.jpg)

### Modelled Air Quality {#subsec:4modelledairquality}

The exposures in the LHEM contained Londoners movement data covering
twelve months of the year between 2005 and 2011. The CMAQ-UK air quality
model used was an annual average hourly weekday/Saturday/Sunday model of
2011 air quality, and linked to the movement of the individual based on
what day of the week their time-activity data was recorded. The air
quality model for this chapter is a 2016 CMAQ-UK output, with
appropriate emissions and meteorology, cropped to London, but otherwise
unchanged in methods from the 2011 model. When evaluated on an annual
average basis i.e. one value for each 20m x 20m grid square of London,
this new 2016 model performs well (Table
\[tab:cmaq\_performance\_stats\_2016\]).

  **Pollutant**   **Number of data**   **Mean obs**   **Mean Mod**   **FAC2**   **MB**   **MGE**   **NMB**   **NMGE**   **RMSE**   **r**   **COE**
  --------------- -------------------- -------------- -------------- ---------- -------- --------- --------- ---------- ---------- ------- ---------
  CO              2                    369.79         318.75         1          -51.04   72.08     -0.14     0.19       88.32      1       0.50
  NO$_{2}$        73                   51.80          48.49          1          -3.30    9.01      -0.06     0.17       13.79      0.84    0.52
  NO$_{x}$        72                   134.74         108.52         0.97       -26.23   39.52     -0.19     0.29       61.50      0.78    0.43
  O$_{3}$         16                   29.74          35.40          1          5.66     6.59      0.19      0.22       7.81       0.83    0.07
  PM$_{10}$       56                   22.76          22.74          1          -0.03    2.93      0.00      0.13       4.14       0.57    0.22
  PM$_{2.5}$      23                   13.08          12.34          1          -0.74    1.85      -0.06     0.14       2.30       0.72    0.20

  : Annual average performance statistics for CMAQ-UK 2016<span
  data-label="tab:cmaq_performance_stats_2016"></span>

We can see that the RMSE values for the pollutants range from 88.32 PPB
for CO down to 2.3 $\mu \text{g m}^{-3}$ for PM$_{10}$. For NO$_{2}$,
which is the pollutant being focused on in this chapter, the RMSE is
13.79 $\mu \text{g m}^{-3}$. An exposure model using annual average
NO$_{2}$ predictions from this 2016 CMAQ-UK air quality model could
incorporate these errors; a prediction of 40 $\mu \text{g m}^{-3}$,
presuming a normal distribution of errors, would therefore actually be
anywhere within the range of 26.21 $\mu \text{g m}^{-3}$ to 53.79
$\mu \text{g m}^{-3}$. The exposure-health relationships investigated by
epidemiologists could then take this into account. Moving to a higher
temporal scale, the daily-hourly version of this same model (Table
\[tab:cmaq\_performance\_stats\_2016\_hourly\]) performs less well for
NO$_{2}$ than the annual average shown above (Table
\[tab:cmaq\_performance\_stats\_2016\]).

  **Pollutant**   **Number of data**   **Mean obs**   **Mean Mod**   **FAC2**   **MB**   **MGE**   **NMB**   **NMGE**   **RMSE**   **r**   **COE**
  --------------- -------------------- -------------- -------------- ---------- -------- --------- --------- ---------- ---------- ------- ---------
  CO              24520                408.92         338.26         0.78       -70.65   175.22    -0.17     0.43       259.18     0.53    0.23
  NO$_{2}$        656419               52.10          49.01          0.84       -3.09    18.63     -0.06     0.36       28.32      0.69    0.35
  NO$_{x}$        649220               136.27         110.28         0.68       -25.99   70.72     -0.19     0.52       129.38     0.63    0.36
  O$_{3}$         148043               28.84          34.31          0.66       5.46     12.88     0.19      0.45       17.20      0.74    0.30
  PM$_{10}$       49461                22.96          22.87          0.89       -0.09    7.07      0.00      0.31       11.89      0.70    0.35
  PM$_{2.5}$      188237               13.15          12.44          0.85       -0.71    3.72      -0.05     0.28       5.76       0.86    0.51

  : Daily-hourly performance statistics for CMAQ-UK 2016<span
  data-label="tab:cmaq_performance_stats_2016_hourly"></span>

The NO$_{2}$ predictions have an R value down from 0.84 to 0.68 and a
RMSE up from 13 $\mu \text{g m}^{-3}$ to 28.32. Using a higher temporal
scale model has increased the uncertainty of the air quality
concentration predictions to the model.

To evaluate the LHEM using this air quality as an input, we needed to
decide what time period of exposure to try and replicate with
measurements. Evaluating at a higher time resolution than annual average
was preferable, but a very high time resolution would be difficult as
this would mean needing to take a huge number of measurements. For
example supposing we wanted to evaluate exposure on Saturday’s between
11am and 2pm and how the LHEM modelled this (within the context of a
journey); measurements would need to be taken continuously on the
journey in question every Saturday throughout the year between 11am and
2pm to compare the measurements with the modelled air quality output
(putting the spatial and micro-environmental factors aside for a
moment). To simplify the evaluation, and make the experiment a
reasonable undertaking, the timeframe that it would be performed over
was restricted, so rather than cover a whole year we considered August
and September only, on weekdays, and between the hours of 9am of 10am.
To create an appropriate air quality model for this time period, 365
days of CMAQ-UK layers were used (each with 24 hours). A loop was
written in R to download the raw CMAQ-UK NetCDF model files, convert
them to standard raster R files, combine the individual files for each
hour of the two months in question, discard data that was not 9am to
10am, and then take the mean for each grid square. The concentrations
were outputted in PPB, so were converted to $\mu \text{g m}^{-3}$ using
the EU standard of 1.9125. Figure 3 shows the result of this process
i.e. the models predicted air quality for a typical weekday in August or
September between 9am and 10am.

![CMAQ-UK air quality model for 9am to 10am on weekdays in
August/September<span
data-label="fig:cmaq_aug_sept"></span>](cmaq_aug_sept.png)

### Micro-environmental adjustments {#subsec:4microenvironmentaladjustments}

As discussed in the introduction to this chapter, and within Section
\[subsec:microenvironments\], calculating exposure within
micro-environments in relation to outdoor concentrations is not a well
understood area of science and seems to be susceptible to a great deal
of variation depending on conditions of the micro-environment (and other
factors). To simplify, I decided to effectively remove this uncertainty
from my evaluation by considering the exposure of a journey which did
not need micro-environmental modelling, looking at journeys of the LHEM
that take concentrations directly from CMAQ-UK i.e. walking and cycling.
This is not to say that this is an unimportant area for investigation,
due to the large amount of time that people spend indoors it clearly is,
but I need to compartmentalise this chapter in order to make the
research achievable, and this is an area that can be reasonably removed.

### Sample size {#subsec:samplesize}

Having established that I will look to evaluate exposure on a cycling
journey between 9am and 10am on weekdays in August and September, I now
needed to establish how many cycling journeys (and measurements) would
be needed to represent modelled exposure of this same journey. Repeat
measurements will be needed as although CMAQ-UK has well resolved
temporal and spatial inputs such as vehicle numbers and emissions by
time of day and day of the week, with similar meteorological inputs, it
cannot take account of inter-day variability of an unpredictable nature
e.g. a car breaking down leading to a traffic jam, and increased
emissions for an hour of a Saturday morning. By taking repeat
measurements, the effects of these type of events on the evaluation
should be reduced. But exactly how many we are not sure. We can consider
this question in a theoretical framework to start-with, based around the
LHEM. If we presume that the mean daily PM2.5 exposures of the 45,000
LTDS subjects was 15 ug/m3, with a standard deviation of 2.5 ug/m3, and
we use a random number generator with these parameters to assign each
subject an exposure, we end up with a distribution like Figure
\[fig:theoretical\_lhem\_pm25\].

![Theoretical LHEM exposures of 45,000 subjects based on a pre-defined
mean of 15 $\mu \text{g m}^{-3}$ and a standard deviation of 2.5
$\mu \text{g m}^{-3}$ (shown in blue)<span
data-label="fig:theoretical_lhem_pm25"></span>](theoretical_lhem_pm25.png)

This distribution of these exposures has a mean of 15
$\mu \text{g m}^{-3}$, a standard deviation of 2.5
$\mu \text{g m}^{-3}$, a max of 25.19 $\mu \text{g m}^{-3}$ and a min of
5.61 $\mu \text{g m}^{-3}$. If these were the real exposures of the LHEM
subjects, and we wanted to undertake measurements on a sub-sample to
evaluate our model, and we decided we would be happy with our sub-sample
having a mean within 10% of the population mean (15
$\mu \text{g m}^{-3}$), with 95% confidence, we can then solve a sample
size calculation (REF) and arrive at the answer of 43. That is, we only
need to sample 43 people to evaluate the model within these parameters.
However this is just going to give us a sample distribution, that we can
be relatively certain has a mean that is within the area highlighted in
red in Figure \[fig:theoretical\_lhem\_pm25\_samples\], but may or may
not have the same shape of distribution, maximum, minimum, standard
deviation etc. of the population. We are also likely to miss evaluating
the exposures of the subjects who have exposures in the highest and
lowest percentiles of the data, which as discussed in Section
\[subsec:longtermvshortterm\] may be important for understanding health
effects.

![Theoretical LHEM exposures of 45,000 subjects based on a pre-defined
mean of 15 $\mu \text{g m}^{-3}$ and a standard deviation of 2.5
$\mu \text{g m}^{-3}$ (shown in blue), with the mean of a sub-sample of
43 subjects (within the red area)<span
data-label="fig:theoretical_lhem_pm25_samples"></span>](theoretical_lhem_pm25_samples.png)

To properly capture the distribution of the population in a sample we
need to do stratified random sampling, ensuring proportional
representation of each strata (exposure concentrations in our case).
However, to do this we would already need to have a priori understanding
of the exposures of the larger group, and the drivers of the exposures,
in order to be able to calculate the numbers of samples required from
each strata. Given the difficultly in trying to evaluate a typical day
of an individual in the LHEM and how these fit against an annual
average, I will again simplify. By just evaluating the modelled and
measured exposure of a single cycling journey from the LHEM most of
these issues are put aside. I just need to calculate how many
measurements of the journey on a weekday between 9am and 10am in August
and September are needed to represent (within predefined boundaries) the
typical exposure on the journey to compare to the model. I calculated
this using one-minute data from a monitoring site (as a known,
continuous, and reliable dataset), and then running sample size
calculations on it, as a proxy for a CMAQ-UK grid square in the model. I
took one year of one minute NO2 monitoring data from the site
‘Wandsworth - Putney High Street’, removed the data outside of our time
period of interest (weekdays, 9am to 10am in August and September), and
then calculated summary statistics on the remaining data (Shown in Table
\[tab:wa7\_summary\_stats\]). I chose this monitoring site as the site
collects one-minute resolution NO$_{2}$ data, which aligns with the
temporal resolution of the equipment I intend to use for monitoring.

  --------------------------------------------------------------------------- --
  **<span>Statistic</span> & **<span>NO$_{2}$ $\mu \text{g m}^{-3}$</span>\   
  Minutes & 2400\                                                             
  Mean & 90.92\                                                               
  Median & 77.70\                                                             
  Max & 1074.4\                                                               
  Min & 11.67\                                                                
  Standard deviation & 64.23\                                                 
  ****                                                                        
  --------------------------------------------------------------------------- --

  : Site WA7 summary statistics for weekday, 9am-10am during<span
  data-label="tab:wa7_summary_stats"></span>

Now that I had summary statistics for the period of time of interest, I
calculated how many samples were required within this time period to
create a similar distribution and mean, as per the theoretical example
from Figures \[fig:theoretical\_lhem\_pm25\] and
\[fig:theoretical\_lhem\_pm25\_samples\] using Equation
\[eq:sample\_size\_equation\_pt1\]
(@PennStateEberlyCollegeofScience2017). By taking this reliable and
continuous dataset, in a known location, we extrapolate the results to
locations where there is not a fixed site data collection method
available.

$$\chi = \textit{Z}^2 \times \frac{\sigma^2}{moe^2} 
  \label{eq:sample_size_equation_pt1}$$

Where $\chi$ is the calculated sample size, *Z* is the z-score of the
desired confidence level, $\sigma$ is the standard deviation of the
population (concentrations), and *moe* is the allowed margin of error
(in $\mu \text{g m}^{-3}$). By writing a loop in the R programming
language (@RFoundationforStatisticalComputing2014), and testing a
variety of input variables to this equation, I was able to see how many
samples would be required in different scenarios of confidence levels
and allowable margins of error. The results are shown in Table
\[tab:no2\_samples\_needed\]:

  ------------------ ---------------------------------------- -----------------
  Confidence level   Allowed margin of error                  Sample required
  hline              10% ($\pm$ 9 $\mu \text{g m}^{-3}$ )     1325
                     15% ($\pm$ 14 $\mu \text{g m}^{-3}$ )    589
                     20% ($\pm$ 18 $\mu \text{g m}^{-3}$ )    331
                     25% ($\pm$ 23 $\mu \text{g m}^{-3}$ )    212
                     30% ($\pm$ 27 $\mu \text{g m}^{-3}$ )    147
                     35% ($\pm$ 32 $\mu \text{g m}^{-3}$ )    108
                     10% ($\pm$ 9 $\mu \text{g m}^{-3}$ )     767
                     15% ($\pm$ 14 $\mu \text{g m}^{-3}$ )    341
                     20% ($\pm$ 18 $\mu \text{g m}^{-3}$ )    192
                     25% ($\pm$ 23 $\mu \text{g m}^{-3}$ )    123
                     30% ($\pm$ 27 $\mu \text{g m}^{-3}$ )    85
                     35% ($\pm$ 32 $\mu \text{g m}^{-3}$ )    63
                     10% ($\pm$ 9 $\mu \text{g m}^{-3}$ )     540
                     15% ($\pm$ 14 $\mu \text{g m}^{-3}$ )    240
                     20% ($\pm$ 18 $\mu \text{g m}^{-3}$ )    135
                     25% ($\pm$ 23 $\mu \text{g m}^{-3}$ )    86
                     30% ($\pm$ 27 $\mu \text{g m}^{-3}$ )    60
                     35% ($\pm$ 32 $\mu \text{g m}^{-3}$ )    44
                     10% ($\pm$ 9 $\mu \text{g m}^{-3}$ )     414
                     15% ($\pm$ 14 $\mu \text{g m}^{-3}$ )    184
                     20 % ($\pm$ 18 $\mu \text{g m}^{-3}$ )   103
                     25% ($\pm$ 23 $\mu \text{g m}^{-3}$ )    66
                     30% ($\pm$ 27 $\mu \text{g m}^{-3}$ )    46
                     35% ($\pm$ 32 $\mu \text{g m}^{-3}$ )    34
                     10% ($\pm$ 9 $\mu \text{g m}^{-3}$ )     327
                     15% ($\pm$ 14 $\mu \text{g m}^{-3}$ )    145
                     20% ($\pm$ 18 $\mu \text{g m}^{-3}$ )    82
                     25% ($\pm$ 23 $\mu \text{g m}^{-3}$ )    52
                     30% ($\pm$ 27 $\mu \text{g m}^{-3}$ )    36
                     35% ($\pm$ 32 $\mu \text{g m}^{-3}$ )    27
  ------------------ ---------------------------------------- -----------------

  : Numbers of samples required for each confidence level and allowable
  margin of error based on measurements from WA7<span
  data-label="tab:no2_samples_needed"></span>

From these calculations we can see that to get a result which is within
10 $\mu \text{g m}^{-3}$ of the mean, with 99% confidence, 1325 samples
would need to be taken, at random in the time window, in a grid square.
Given my aim is to take measurements on a cycling journey, and as there
are only 44 weekdays in August and September, attaining 1325 samples is
not achievable. I decided to take at least 27 samples, which would give
a 80% confidence interval (CI) and 32 $\mu \text{g m}^{-3}$ margin of
error (MOE) in my results. Clearly it would be better to have may more
samples and stronger results with a higher CI and lower MOE, but this
was not possible in the time-frame. This issue is explored further in
the Discussion, (Section \[sec:4Discussion\]).

### Measuring journey exposure {#subsec:measuringjourneyexposure}

Now that the time period of comparison was set (9am to 10am, weekday
mornings, August to September) and the minimum number of samples
required had been established (27) the data collection was planned.
Collecting NO2 data on 27 journeys at one-minute resolution would be
difficult, as King’s do not own a reliable and portable NO$_{2}$ sensor
that can monitor at this frequency (indeed there are few commercially
available at all). At the suggestion of a colleague, I decided to
collect black carbon (BC) data instead using a microaeth (@Hansen1984,
@Aethlabs2016), and then convert this to NO$_{2}$, the microaeth being a
well understood and reliable device used in many previous exposure
studies (/cite<span>Cheng2013</span>, /cite<span>Viana2015</span>). To
calculate the conversion factor between BC and NO$_{2}$, NO$_{2}$ and BC
data was downloaded from the Marylebone Road monitoring site (WA7) which
collects both pollutants at 1-minute resolution, and then a linear
regression model was created (Figure
\[fig:black\_carbon\_no2\_conversion\]).

![Linear regression between black carbon and NO$_{2}$<span
data-label="fig:black_carbon_no2_conversion"></span>](images/black_carbon_no2_conversion.png)

This gave a conversion of NO$_{2}$ = 41 + 9.1 x BC, with a R^2^ of
0.474. A higher R^2^ would have been preferable to give more confidence
to the BC-NO2 conversion process. This is again explored in the
Discussion ( Section \[sec:4Discussion\] ).

With regards to further sources of error in the evaluation of the LHEM
using monitored data, it was presumed that the Microaeth AE51 was giving
perfectly accurate readings of black carbon. Accepting that this is a
limitation and in reality the device has been shown to have (relatively
modest) measurement errors ( @Cheng2013, @Viana2015 ).

A monitoring campaign was then completed over August and September,
taking measurements on 28 cycling journeys between Kennington Park
(51.484535, -0.109529) and Waterloo (51.505627, -0.111812) between 8am
and 9am using the Microaeth AE51. The journey between Kennington Park
and Waterloo was chosen due to convenience for the researcher and is
purely arbitrary; the method could have been applied anywhere within the
model domain.

### Modelling journey exposure {#subsec:modellingjourneyexposure}

Using the CMAQ-UK air quality layer as an input I now modelled the
exposure of the journey using the same methods as used in the LHEM from
Chapter \[chap:the\_lhem\]. The route and concentrations along it are
shown in Figure \[fig:dual\_exposure\_cycling\] below.

![Cycling journey between Kennington Park and Waterloo, overlain on
modelled CMAQ-UK concentrations for 9am to 10am on weekdays in August to
September<span
data-label="fig:dual_exposure_cycling"></span>](images/dual_exposure_cycling.png)

### Data processing {#subsec:dataprocessing}

Before the measured exposure datasets could be compared to the modelled
exposure of the same journey the following steps were completed:

-   Download raw CSV data from the Microaeth. Remove unnecessary columns
    and header meta-data.

-   Import to R

-   Divide black carbon concentrations by 1000 to convert into
    micrograms per metre cubed (for comparison with the NO$_{2}$)

-   Manually fill in missing GPS data (device does not record position
    for first few minutes of operation, which was unknown at the time of
    data collection)

-   Correct poor quality GPS data / drift by snapping GPS points to the
    nearest road segment (Shown in Figure 9)

-   Create line segments instead of point positions.

![Locations of sampled BC concentrations along the cycle journey
(purple), which have been snapped to the road (pink) to correct for GPS
drift<span
data-label="fig:snapping_cycle_to_road"></span>](images/snapping_cycle_to_road.png)

The final step of creating line segments from points was required as the
microaeth concentration that is stored each minute is at the end of one
minute of sampling, meaning that the GPS position is the mean
concentration from the previous minute of movement. To enable linking of
this concentration with the CMAQ-UK grid squares during that period of
movement, I created a SpatialLinesDataFrame line between each GPS point,
and the previous GPS point in that journey, and assigned the
concentration to that whole line. The result of this processing of the
monitoring data was 27 lines, stored as a SpatialLinesDataFrame (SLDF),
split into segments of varying lengths, along the length of the journey
from Kennington Park to the Waterloo. The start of one of the journeys
is shown in Figure \[fig:cmaq\_cycle\_route\_zoomed\] below. Although
not shown in the figure, the line has concentration attributes linked to
it.

![Start of a cycling journey from Kennington Park to Waterloo, cycle
route shown by a black line, CMAQ model concentrations shown
behind.<span
data-label="fig:cmaq_cycle_route_zoomed"></span>](images/cmaq_cycle_route_zoomed.png)

Using the SLDF of measured concentrations, and the CMAQ-UK raster of
modelled concentrations, two result datasets were now created:

-   **Concentration comparison**: The mean concentration from each
    journey, compared to the mean concentration of the grid cells that
    the journey intersected.

-   **Spatial comparison**: For each grid cell the route intersected, I
    took the mean concentration from the 27 monitored lines that went
    through it. I then calculated the difference, by grid square,
    between the monitoring concentrations and the modelled
    concentrations and outputted this as a new raster file.

Results {#sec:4results}
-------

### Concentration comparisons {#subsec:concentrationcomparisons}

Figure \[fig:grouped\_journey\_boxplots\] shows a boxplot summary graph
of the 27 monitored journeys (right, black) compared to the modelled
journey (left, red).

![Box-plot of modelled cycling journey exposure compared to monitored
cycling exposure<span
data-label="fig:grouped_journey_boxplots"></span>](images/grouped_journey_boxplots.png)

From this summary plot we can see that in general the monitoring
campaigns found higher concentrations than modelling the journey. When
visualised individually (Figure \[fig:journey\_boxplots\]) we can see
the variation more clearly. Noting that earlier we calculated that 27
repeat measurements were needed for a 80% confidence interval and 30
$\mu \text{g m}^{-3}$ margin of error comparison, and should therefore
really only be considering this data in aggregate form..

![Box-plots of monitored cycling journeys (black) compared to the
modelled cycling journey (red). Note that due to the device numbering
system the sessions were numbered between 15 and 43, but there are not
43 sessions in the graph.<span
data-label="fig:journey_boxplots"></span>](images/journey_boxplots.png)

Note that due to the device numbering system the sessions were numbered
between 15 and 43, but there are not 43 sessions in the graph.

The summary statistics for these two datasets are shown in Table
\[tab:modelled\_monitored\_cycling\_summary\]. Taking the raw monitored
data against the modelled data, the lower boundaries of the datasets are
similar, but the monitoring found mean and median concentrations
approximately double that of the modelling, and the maximum monitored
value is way in excess of the model. If we consider the monitored data,
and include the allowable margin of error defined in the sample size
calculation of 30 $\mu \text{g m}^{-3}$, then the agreement of the means
at the lower boundary of the margin of error are much closer (64.19
$\mu \text{g m}^{-3}$ for monitored compared to 44.82
$\mu \text{g m}^{-3}$ for modelled).

  -- ------------------------------------------------------------------------------------------------------------------------- --
     **<span>Modelled NO$_{2}$ $\mu \text{g m}^{-3}$</span> & **<span>Monitored NO$_{2}$ $\mu \text{g m}^{-3}$ (MOE)</span>\   
     Minimum & 37.58 & 40.35\                                                                                                  
     1st Quartile & 40.48 & 68.08\                                                                                             
     Median & 44.57 & 80.81\                                                                                                   
     Mean & 44.82 & 94.19 (64.19-124.20)\                                                                                      
     3rd Quartile & 46.19 & 102.70\                                                                                            
     Maximum & 80.33 & 376\                                                                                                    
     ****                                                                                                                      
  -- ------------------------------------------------------------------------------------------------------------------------- --

  : Comparison of measured and modelled exposure on the cycling
  journey<span
  data-label="tab:modelled_monitored_cycling_summary"></span>

### Spatial Comparisons {#subsec:spatialcomparisons}

For reference, figure \[fig:cycling\_journey\] shows the journey from
Kennington Park to Waterloo that we have been comparing.

![Map of cycling route between Kennington Park and Waterloo<span
data-label="fig:cycling_journey"></span>](images/cycling_journey.png)

Figure \[fig:monitored\_modelled\_summary\] below shows the monitoring
data, the modelled data, and a comparison of the two on this route.

![Map of monitored, modelled and difference concentrations along the
journey (NO$_{2}$ derived from black carbon)<span
data-label="fig:monitored_modelled_summary"></span>](images/monitored_modelled_summary.png)

By spatially comparing the results in this manner we can see that the
monitoring campaign found the highest concentrations (&gt; 100
$\mu \text{g m}^{-3}$) around the junction of Kennington Road and
Kennington Lane (towards the South of the journey), near Lambeth North
Station (a busy junction), and again near The Old Vic theatre (another
busy junction). These locations of high concentrations are also found in
the modelled data, but the difference against other stretches of the
journey is not as large, and compared to the monitored data the
concentrations are not as high. Considering the difference map (shown
right in Figure \[fig:monitored\_modelled\_summary\]), except for the
first 5-10% of the journey, most of the journey is underestimated by the
model, between 0 and 25 $\mu \text{g m}^{-3}$ in each grid square. The
largest differences are on Baylis Road, between Lambeth North Station
and the Old Vic theatre, where the monitoring found concentrations to be
up to 50 $\mu \text{g m}^{-3}$ higher.

Discussion {#sec:4Discussion}
----------

As discussed in the Background (Section \[sec:4background\]), to my
knowledge, there have not been any studies which have attempted to
evaluate a dynamic approach to exposure in this manner. Studies have
tended to measure certain micro-environments for a time-span often
determined by practical constraints such as battery life, staff time and
convenience; and then use these as empirical comparisons to their model
outputs. This piece of research was novel in that exposure was modelled,
and then an attempt to evaluate the predictions (of an example journey)
was undertaken by calculating how many personal monitoring samples would
be needed, and then going out to collect that data. Despite various
sources of possible error, this research demonstrated this as a possible
process, and highlighted the difficulties of it. Reasonable ‘ball-park’
results were found, and clear spatial patterns were discernible.

There were many issues and challenges to overcome whilst undertaking
this evaluation that will have impacted the accuracy of the results. The
main one being (mostly unavoidable) sources of error at each stage of
the process. On the modelling side of the comparison CMAQ-UK has been
shown to perform well when evaluated against monitoring sites, but it is
far from perfect – the input the modelling uses for this chapter has
NO$_{2}$ R values of 0.75. Against this, it was calculated that 27
samples would be needed to get a monitored concentration estimate that
had a 30 $\mu \text{g m}^{-3}$ MOE (and 80% CI). Black carbon samples
were collected to do this, using a MA300 Microaeth, which has not yet
been evaluated by academic literature (although previous models of
similar equipment has R^2^ values of 0.8), and these data were converted
to NO$_{2}$ using a linear regression with an R^2^ of 0.47. These
multiple sources of error are bound to have confounded the results, but
to what degree is uncertain. Post-hoc, it is not possible to go back and
improve the CMAQ-UK model, or to collect more samples, or to improve the
accuracy of the portable monitor, but it is interesting to reconsider
the conversion process of black carbon to NO$_{2}$ from Figure
\[fig:black\_carbon\_no2\_conversion\]. Visually the intercept of 41
looks high, and if this was theoretically changed to 0, then the
measured concentrations would all be reduced, and the comparison between
modelled and monitored would be much closer. Figures
\[fig:grouped\_journey\_boxplots\_new\_intercept\] and
\[fig:monitored\_minus\_cmaq\_route\_cells\_concs\_zero\_intercept\]
below show revised boxplot comparisons and spatial comparisons.

![Boxplot comparison of modelled and monitored concentrations with an
intercept of 0 between BC and NO$_{2}$<span
data-label="fig:grouped_journey_boxplots_new_intercept"></span>](images/grouped_journey_boxplots_new_intercept.png)

![Spatial comparison of modelled and monitored concentrations with an
intercept of 0 between BC and NO$_{2}$<span
data-label="fig:monitored_minus_cmaq_route_cells_concs_zero_intercept"></span>](images/monitored_minus_cmaq_route_cells_concs_zero_intercept.png)

Putting measurement errors and modelling errors aside, there are
practical improvements that could be implemented in repeating this work
which might lead to improved results. The most notable being the sample
size. If the journey had been shorter, or more researchers were
available, it would have been possible to collect a much larger number
of samples, which according to the sample size calculations would have
increased the reliability of the results. A more practical long-term
approach to evaluating air quality model inputs to a hybrid exposure
model could be to fix reliable portal monitors to cars or public
transport and increase sample size in this manner. For example, mobile
monitors might be attached to the outside of a fleet of No.59 buses
which run along the route I cycled, meaning each grid square would be
sampled 10-15 times between 9am and 10am, resulting in 400-500
measurements over the same time period. Indeed, our group are already
trialling a similar method, but with the monitors inside a bus to
measure the changes in passenger exposure from the electrification of a
bus route in London.

Focusing on the equipment, an unforeseen issue in collecting the
monitoring data was a mismatch between cycling speed, the resolution of
the CMAQ-UK grid squares, and the temporal resolution of the Microaeth
MA300. I wanted to measure concentrations in each grid square along a
route, but as the Microaeth was set to 1-minute resolution, the grid
squares are 20m by 20m, and typical cycling speed is 15 km/h (or 250
metres per minute), this meant that each minute of Microaeth data
covered 250/20 = 12.5 grid squares. Repeating this experiment, a
researcher could calculate the sampling rate of the device based upon
the likely speed of the movement, and the air quality model resolution
it will be compared to, as per equation \[eq:sampling\_rate\_equation\]

$$sampling rate_{(minutes)} = \frac{AQ_{(metres)}}{speed_{(metres/minute)}}
  \label{eq:sampling_rate_equation}$$

Though this would of course be constrained by the settings available in
the sampling device.

Within the GIS and data processing section of this work, the main issue
was the lack of accuracy of the GPS points that were output from the
device, how to snap these appropriately to roads (presuming an available
roads dataset), and how to link this to CMAQ-UK concentration grid
squares for analysis. This was time-consuming and required manual
editing at times. Automating this for a large scale campaign would be
challenging but necessary. Junctions were a particular difficulty, for
instance snapping GPS points at a left-turn always means that the point
is snapped to part of the road either side of the junction and never at
the actual junction itself (Figure \[fig:gps\_snapping\_error\]); which
might be where the highest concentrations occur and the researcher is
most interested.

![Accurate GPS points shown in green, GPS error shown in red. Correction
should move the point to location shown by green arrow, but simple
‘snapping’ will move it to direction of red arrow.<span
data-label="fig:gps_snapping_error"></span>](images/gps_snapping_error.png)

Conclusions {#sec:4conclusions}
-----------

The main conclusions from this chapter are method-orientated, rather
than definitive answers about the reliability of a hybrid exposure
method. The amount of data collection and data processing were both
extremely onerous in order to even evaluate one journey, requiring
skills in data processing, GIS, statistics and personal monitoring. To
repeat this method on a larger scale initial consideration should be
given to the monitors that are going to be used, and in which possible
transport modes, to ensure that they provide a high enough temporal and
spatial resolution for alignment with the modelled air quality being
used by the exposure modelling. Similarly, testing the monitors
extensively to ensure that the parameters required are output
immediately rather than taking a short period of time to internally
calibrate themselves when turned on (as I unfortunately found with the
GPS sensor of the Microaeth which led to hours of manual correction).
The sample size calculations here demonstrated a method (given a fixed
reliable ‘base’ dataset) to estimate the number of samples that are
required to arrive at monitored concentrations that have a low margin of
error and high confidence. For this piece of research, it was calculated
that over 1,000 measurements would have been needed to give a low margin
of error with high confidence, but it is worth noting that this figure
depends on the pollutant of interest, the base data, and the
geographical area. It could be much higher or lower. Regarding results
rather than method, I found that the CMAQ-UK NO$_{2}$ model used as an
input to the London Hybrid Exposure Model is underestimating
concentrations for the journey I chose. Whether this can be extrapolated
to draw wider conclusions about the model are unclear. The findings have
large caveats due to multiple, mostly unavoidable, sources of error in
the method e.g. the black carbon to NO$_{2}$ conversion and sample size,
that require fuller exploration of their impact and propagation.

Discussion & future work
========================

The original hypothesis of this research was that “*Individual and
population level air pollutant exposure can be estimated using
time-activity surveys, GIS and routing tools, and then coupled with high
resolution spatio-temporal air quality models to facilitate a greater
understanding of the health impacts of air pollution and how public
health risks can be reduced*”. Overall, this research has proved this
hypothesis through the creation of a tool that allows scientists to
examine individual-level and group exposure to ranges of pollutants in a
range of microenvironments. Chapters \[chap:the\_ltdsx\] and
\[chap:the\_lhem\], which were the main chapters centred around the
creation of the tool, have been published as @Smith2016 (the model
development and use), and then subsequently used by @Tonne2018 to
examine socioeconomic and ethnic inequalities in exposure to poor air
quality. Chapter \[chap:monitoring\_on\_underground\], refining the
exposure estimates for users of the London Underground after this
microenvironment was found to be so important to Londoners exposure, is
in the process of being written-up as an academic paper and will provide
an open-source dataset (’TubeAir’) for calculating exposure of London
Underground users within other studies, or as a stand-alone policy tool.
Chapter \[chap:evaluating\_dynamic\_exposure\_models\], on how to
evaluate exposure models of this kind, diverted away from the main
hypothesis of the research, but during chapters \[chap:the\_ltdsx\],
\[chap:the\_lhem\] and \[chap:monitoring\_on\_underground\] the
reliability and representativeness of the results became something of
interest (and importance) to consider, and so this seemed necessary.\
In more detail, Chapter \[chap:the\_ltdsx\] demonstrated how a dataset
that was originally purposed for assessing transport demand could be
adapted, processed and re-purposed to create the LTDS-X, a high
resolution spatial and temporal time-activity dataset, including
demographics, that is representative of the daily movements of the
population of London. The main limitations with using this dataset as an
input to the exposure modelling, were that the data were London-centric,
and ’porting’ this method to another city or country would require a
similar or replacement dataset. Although whilst undertaking this
research other datasets have been investigated and considered as
replacements, for example in the 2011 UK Census, a new question of
’workplace zones’ was asked of the population, and research by Dr Reis
(Centre for Ecology and Hydrology, UK) is using the responses as a way
to model population-level movement for exposure modelling. Using this as
a basis, travel exposure from workplace zone to workplace zone by each
mode of transport could be simulated and the exposure calculated, and
then an indoor exposure module ’plugged-in’ to create a UK-wide exposure
model. The spatial detail would not be as high as the LTDS-X and LHEM,
but the coverage would be much larger. Given Census’ are common in many
countries around the world, this could be replicated in other places
given time.

It is worth noting that whilst this dataset was specifically re-purposed
for use as an input to air quality exposure modelling, it could
similarly be used for exposure to other things such as perhaps
ultra-violet light and risks of skin-cancer, or to estimate where the
population are drinking water and therefore links between poor quality
water and health.

Away from exposure the dataset has been useful in other areas of work
that the Environmental Research Group is involved in. In 2012 it was
used to look at the number of cross-Borough car trips being undertaken
as part of the London Atmospheric Emissions Inventory (LAEI), and in
2018 was used to identify whether active travel has increased or
decreased in the London Borough of Waltham Forest as part of a contract
piece of work.\
In Chapter \[chap:the\_lhem\], having built the LTDS-X, this was
combined with CMAQ-UK and microenvironmental modelling methods to
quantify at unprecedented detail the exposure of the London population
to poor air quality. The size, detail and possibilities for future
research of this model were demonstrated by focusing on exposure
missclassification, and this work was published in @Smith2016 (which has
13 citations as of May 2018). The applications for this model are
already being taken forward in other exposure studies, including the
CLUE II led by Imperial College London looking at the air quality and
noise exposure of children in London, and the COPE (Characterisation of
COPD Exacerbations using Environmental Exposure Modelling) study led by
King’s College London. As was highlighted in the Chapter, it is our hope
that this type of tool can increasingly be used for policy applications
by the likes of TfL and London Boroughs to better understand the effects
of policy interventions on exposure. Indeed, Waltham Forest are using it
to better understand the effects on cyclist exposure of introducing
segregated cycle lanes, and we are collecting data on behalf of TfL to
quantify changes in bus passenger exposure following retrofitting of
cleaner engines.

The areas of this model that were identified as most in need of
improvement were the modelling within microenvironments, which include
time in enclosed transport (bus, car, train, tube), and time indoors. As
little was known about the exposure of passengers on the London
Underground, static values were used as proxy based on a small sampling
campaign by Dr Barratt. Over the course of this section of research, it
transpired that this was an important determinant of high daily
exposures in the population, and therefore the objectives and research
plan for Chapter \[chap:monitoring\_on\_underground\] were developed.\
Chapter \[chap:monitoring\_on\_underground\] involved an extensive
mobile monitoring campaign on the London Underground, and creation of a
dataset for estimating PM$_{2.5}$ passenger exposure during journeys on
the network. This research has directly led to a sub-group of the
Committee on the Medical Effects of Air Pollution (COMEAP) being formed,
and a report (currently in final draft) on “available evidence on the
health risks associated with particulate matter exposure in the London
Underground” being written. Within the Environmental Research Group at
King’s, Dr Green has also been contracted by TfL to undertake further
monitoring on station platforms. The data and research should be
published later in 2018 which is expected to lead to further
opportunities and incorporation into other studies.

DISCUSS EVAL HERE CHAPTER \[chap:evaluating\_dynamic\_exposure\_models\]

As discussed in the Background (Section \[sec:4background\]), to my
knowledge, there have not been any studies which have attempted to
evaluate a dynamic approach to exposure in this manner. Studies have
tended to measure certain micro-environments for a time-span often
determined by practical constraints such as battery life, staff time and
convenience; and then use these as empirical comparisons to their model
outputs. This piece of research was novel in that exposure was modelled,
and then an attempt to evaluate the predictions (of an example journey)
was undertaken by calculating how many personal monitoring samples would
be needed, and then going out to collect that data. Despite various
sources of possible error, this research demonstrated this as a possible
process, and highlighted the difficulties of it. Reasonable ‘ball-park’
results were found, and clear spatial patterns were discernible.

There were many issues and challenges to overcome whilst undertaking
this evaluation that will have impacted the accuracy of the results. The
main one being (mostly unavoidable) sources of error at each stage of
the process. On the modelling side of the comparison CMAQ-UK has been
shown to perform well when evaluated against monitoring sites, but it is
far from perfect – the input the modelling uses for this chapter has
NO$_{2}$ R values of 0.75. Against this, it was calculated that 27
samples would be needed to get a monitored concentration estimate that
had a 30 $\mu \text{g m}^{-3}$ MOE (and 80% CI). Black carbon samples
were collected to do this, using a MA300 Microaeth, which has not yet
been evaluated by academic literature (although previous models of
similar equipment has R^2^ values of 0.8), and these data were converted
to NO$_{2}$ using a linear regression with an R^2^ of 0.47. These
multiple sources of error are bound to have confounded the results, but
to what degree is uncertain. Post-hoc, it is not possible to go back and
improve the CMAQ-UK model, or to collect more samples, or to improve the
accuracy of the portable monitor, but it is interesting to reconsider
the conversion process of black carbon to NO$_{2}$ from Figure
\[fig:black\_carbon\_no2\_conversion\]. Visually the intercept of 41
looks high, and if this was theoretically changed to 0, then the
measured concentrations would all be reduced, and the comparison between
modelled and monitored would be much closer. Figures
\[fig:grouped\_journey\_boxplots\_new\_intercept\] and
\[fig:monitored\_minus\_cmaq\_route\_cells\_concs\_zero\_intercept\]
below show revised boxplot comparisons and spatial comparisons.

Putting measurement errors and modelling errors aside, there are
practical improvements that could be implemented in repeating this work
which might lead to improved results. The most notable being the sample
size. If the journey had been shorter, or more researchers were
available, it would have been possible to collect a much larger number
of samples, which according to the sample size calculations would have
increased the reliability of the results. A more practical long-term
approach to evaluating air quality model inputs to a hybrid exposure
model could be to fix reliable portal monitors to cars or public
transport and increase sample size in this manner. For example, mobile
monitors might be attached to the outside of a fleet of No.59 buses
which run along the route I cycled, meaning each grid square would be
sampled 10-15 times between 9am and 10am, resulting in 400-500
measurements over the same time period. Indeed, our group are already
trialling a similar method, but with the monitors inside a bus to
measure the changes in passenger exposure from the electrification of a
bus route in London.

Focusing on the equipment, an unforeseen issue in collecting the
monitoring data was a mismatch between cycling speed, the resolution of
the CMAQ-UK grid squares, and the temporal resolution of the Microaeth
MA300. I wanted to measure concentrations in each grid square along a
route, but as the Microaeth was set to 1-minute resolution, the grid
squares are 20m by 20m, and typical cycling speed is 15 km/h (or 250
metres per minute), this meant that each minute of Microaeth data
covered 250/20 = 12.5 grid squares. Repeating this experiment, a
researcher could calculate the sampling rate of the device based upon
the likely speed of the movement, and the air quality model resolution
it will be compared to, as per equation \[eq:sampling\_rate\_equation\]

Though this would of course be constrained by the settings available in
the sampling device.

Within the GIS and data processing section of this work, the main issue
was the lack of accuracy of the GPS points that were output from the
device, how to snap these appropriately to roads (presuming an available
roads dataset), and how to link this to CMAQ-UK concentration grid
squares for analysis. This was time-consuming and required manual
editing at times. Automating this for a large scale campaign would be
challenging but necessary. Junctions were a particular difficulty, for
instance snapping GPS points at a left-turn always means that the point
is snapped to part of the road either side of the junction and never at
the actual junction itself (Figure \[fig:gps\_snapping\_error\]); which
might be where the highest concentrations occur and the researcher is
most interested.

Conclusions {#conclusions}
===========

Code Listings
=============

The following code contains pertinent examples of key R and SQL code
used in this research

An example of in-vehicle modelling using a mass-balance approach {#code:in_vehicle_example}
----------------------------------------------------------------

Creating the geographical missclassification graphs and maps {#code:create_geographical_results}
------------------------------------------------------------

Creating a London Underground GIS file {#code:make_tube_network_from_scratch}
--------------------------------------

Creating example distribution plots for the LHEM {#code:example_distribution_plot}
------------------------------------------------