taking the time and effort to read my work and give feedback, as well as Prof. ...... Figure 3 The causal pie model of three sufficient causes of a disease (Rothman, ...
QUANTIFICATION OF INDIVIDUAL SOURCES OF UNCERTAINTY IN THE DISEASE BURDEN ESTIMATES OF FINE PARTICLES IN FINLAND
Heli Päivi Hannele Lehtomäki MSc thesis Environmental Science February 2017
UNIVERSITY OF EASTERN FINLAND, Faculty of Science and Forestry, Environmental Science Heli Lehtomäki: Quantification of Individual Sources of Uncertainty in the Disease Burden Estimates of Fine Particles in Finland MSc thesis: 59 pages, 2 appendixes (3 pages) Supervisors: Otto Hänninen (National Institute for Health and Welfare, THL), Jarkko Tissari (University of Eastern Finland, UEF) February 2017
___________________________________________________________________________ Keywords: fine particles, environmental burden of disease, uncertainty, concentration response relationship
ABSTRACT Environmental disease burden has been widely studied for several factors. Disease burden studies are many times conducted to support decision makers. There are multiple sources of uncertainties in burden of disease estimates, and understanding the uncertainties, knowing their quantity and reporting them adequately is important to be able to take them into account in decision making and to be able to reduce them. Air pollution is among the biggest environmental risks globally. In this thesis the aim was to quantify selected uncertainties in the most recent national disease burden estimates for ambient fine particles (PM2.5) in Finland. The uncertainties were selected based on literature. They were quantified separately and then compared, to find the most significant uncertainties. The focus was set on parametric and model uncertainties. The selected uncertainties were related to the uncertainties in the relative risk and exposure estimates, selection of health endpoints, exposure characterisation and the shape of the concentration response function. This thesis has been conducted under the Health impacts of air pollution (ISTE) -project at the National institute for health and welfare (THL). The uncertainties were analysed for the ISTEproject’s best estimate for the disease burden of fine particles (PM2.5) which was 20,800 DALY in 2013. The differences caused by the quantified uncertainties were calculated using this estimate as a reference point. The quantified uncertainties resulted in 13,600 DALY/a and 42,700 DALY/a as a minimum and maximum estimate for the disease burden of fine particles, respectively. The biggest differences from the quantified uncertainties were caused by i) the selection of health endpoints; in this case the use of natural or cause-specific mortality, and ii) the shape of the concentration-response function.
ACKNOWLEDGEMENTS This thesis is a part of Health impacts of Air Pollution (Ilmansaasteiden terveysvaikutukset) ISTE -project which was conducted at the National Institute for Health and Welfare (THL) in 2015-2016. It was founded by the Ministry of the Environment and the Ministry of Social Affairs and Health. The aim of the project was to evaluate the health impacts of air pollution in Finland by calculating disease burden of 14 air pollutants. I am sincerely grateful to my main supervisor Dos. Otto Hänninen for his great guidance and the many discussions we had. I always felt welcome to come ask for advice which was important to me. I also want to thank my other supervisor Dr. Jarkko Tissari for all his guidance and encouraging comments. Furthermore, I am grateful for Dr. Marko Tainio for taking the time and effort to read my work and give feedback, as well as Prof. Matti Viluksela for reviewing my thesis. I would also like to thank other members of the ISTE-project’s working group: Arja Asikainen, Antti Korhonen and Isabell Rumrich. To my family: sydämelliset kiitokset kaikesta tuesta ja luottamuksesta! Special thanks also go to Alwin for his great support throughout my Master studies. I highly appreciate all your support.
ABBREVIATIONS AND DEFINITIONS BoD
Burden of disease
CI
Confidence interval
CRF
Concentration-response function
DALY
Disability adjusted life years
EBD
Environmental burden of disease
GBD
Global burden of disease
GHE
Global health estimates
IARC
International Agency for Cancer Research
IER
Integrated exposure response function
IHME
Institute of Health Metrics and Evaluation, University of Seattle
ISTE
Health impacts of air pollution -project in Finland
PAF
Population attributable fraction
RR
Relative risk
PM2.5
Fine particles; particles whose aerodynamic diameter is less than 2.5 µm
TMRED
Theoretical minimum risk exposure distribution
UR
Unit risk
YLD
Years lived with disability
YLL
Years of life lost due to premature death
WHO
World Health Organization
CONTENTS ABSTRACT ............................................................................................................................... 2 ACKNOWLEDGEMENTS ....................................................................................................... 3 ABBREVIATIONS AND DEFINITIONS ................................................................................ 4 CONTENTS ............................................................................................................................... 5 1
INTRODUCTION .............................................................................................................. 7
2
LITERATURE REVIEW ................................................................................................... 9 2.1
ENVIRONEMENTAL BURDEN OF DISEASE (EBD) STUDIES .......................... 9
2.2
UNCERTAINTIES IN ENVIRONMENTAL DISEASE BURDEN ANALYSIS ... 11
2.1
PARAMETRIC UNCERTAINTIES ......................................................................... 15
2.1.1
Descriptive and normative parameters ............................................................... 15
2.1.2
Random and systematic error ............................................................................. 15
2.2
MODEL UNCERTAINTIES .................................................................................... 17
2.2.1
Epidemiological studies and relative risk........................................................... 17
2.2.2
Shape of the concentration-response function ................................................... 21
2.2.3
Threshold ............................................................................................................ 24
2.2.4
Exposure modelling............................................................................................ 25
3
THE AIM OF THE WORK ............................................................................................. 27
4
MATERIALS AND METHODS ..................................................................................... 28
5
4.1
DISEASE BURDEN CALCULATION METHODS................................................ 28
4.2
ANALYSIS OF SELECTED UNCERTAINTIES .................................................... 30
RESULTS......................................................................................................................... 38 5.1
COMPARISON OF THE QUANTIFIED INDIVIDUAL UNCERTAINTIES........ 38
5.2
EXPOSURE AND RELATIVE RISK ESTIMATES ............................................... 40
5.3
CONCENTRATION RESPONSE RELATIONSHIP............................................... 42
5.4
SELECTION OF HEALTH ENDPOINTS ............................................................... 43
5.5
UNCERTAINTY RANGE FOR THE DISEASE BURDEN OF PM2.5 ................... 45
6
7
DISCUSSION .................................................................................................................. 46 6.1
RELATIVE RISK AND EXPOSURE ESTIMATES ............................................... 46
6.2
AVERAGE VS. EXPOSURE DISTRIBUTION ...................................................... 46
6.3
SHAPE OF THE CONCENTRATION RESPONSE RELATIONSHIP .................. 47
6.4
CUTOFF VALUE ..................................................................................................... 48
6.5
SELECTION OF HEALTH ENDPOINTS ............................................................... 48
6.6
OVERALL DISCUSSION ........................................................................................ 51
CONCLUSIONS .............................................................................................................. 53
REFERENCES ......................................................................................................................... 54 APPENDICES APPENDIX 1: DISEASE BURDEN CALCULATION METHODS ................................... 1 APPENDIX 2: THE IDENTIFIED CONCENTRATION RESPONSE FUNCTIONS ........ 3
7
1 INTRODUCTION Ambient air pollution is one of the main environmental risk factors worldwide (Brauer et al., 2015). The World Health Organization (WHO) has estimated that globally ambient air pollution kills around 3 million people every year (WHO, 2016a). The recent global burden of disease study ranked ambient particulate matter as the sixth leading risk globally in 2015 (Forouzanfar et al., 2016). Decades of toxicological, clinical, and epidemiological research have shown that there are significant associations between exposure to ambient air pollution and several adverse human health effects varying from short term respiratory symptoms to premature deaths (Schwarze et al., 2006; Pope & Dockery, 2006; Anenberg et al., 2016). Furthermore, ambient air pollution in general is classified as carcinogenic to humans (Group 1) (IARC, 2013). Humans are exposed to air pollution from the first weeks in the womb until the end of their lives. In some life stages and under different health conditions (e.g. babies, children, older people and people with long-term health conditions) people are more vulnerable to the adverse health effects of air pollution. This is important to consider since, for instance, adverse effects on babies and children can have an impact that lasts far into their future. (Royal College of Physicians, 2016.) WHO has provided air quality guidelines, and governments have set concentration limits for air pollutants. However, these do not define safe levels for the whole population. The health effects caused by ambient air pollution have high costs to society and businesses, health services and also to those individuals who suffer from illness and premature death. According to a recent report from the Royal College of Physicians the value of benefits from reducing emissions greatly exceeds the costs of reduction. (Royal College of Physicians, 2016.) Air pollution levels are comparatively low in Finland but air pollution is still the main environmental risk factor causing attributable diseases and exacerbating symptoms leading to higher morbidity among the exposed population and additionally causing premature deaths. Earlier national works on the disease burden of ambient air pollution have focused on fine particles (PM2.5) and ozone (O3) (e.g. EBoDE (Hänninen and Knol, 2011) and SETURI/TEKAISU (Asikainen et al., 2013). However, there is increasing evidence of additional air pollutants causing adverse health effects. In Finland 15 air pollutants are regulated by law. The Health Effects of Air Pollution (ISTE) – project (Hänninen et al., 2016, Lehtomäki et al., 2016) was carried out to evaluate the health effects of air pollutants in Finland. In the ISTE-project, 14 air pollutants were analysed for their disease burden. The disease burden estimates of fine particles and ozone have been analysed also previously and now their estimates were updated. For the remaining 12 ambient air pollutants the health impacts were analysed for the first time in Finland. The largest share of the disease burden was associated with fine particles which were estimated to cause 64% of the total disease burden of air pollution (Lehtomäki et al., 2016).
8
Environmental burden of disease assessments are often complex and the uncertainties of the results can be significant. There are several sources of uncertainty which adds difficulty to accurate uncertainty analysis (Knol et al., 2009). Uncertainties can for instance, relate to missing, incomplete and/or incorrect knowledge and considerations with ignorance and/or with a lack of awareness. (WHO, 2008.) Some parametric uncertainties are quite commonly addressed but there are several uncertainties which are often not measured but can, in addition, affect the results substantially. (Knol et al., 2009.) According to current knowledge, uncertainties mainly arise from concentration response functions (CRFs) and exposure estimates. (WHO, 2016b.) When uncertainties are not sufficiently evaluated they can result in poor decisions leading to wasted resources and even loss of lives (Evans, 2016). Therefore it is important to conduct an uncertainty analysis for disease burden studies and characterize the uncertainties transparently. This, in turn, ensures that decision-makers can take them adequately in to consideration. (WHO, 2008.) This thesis aims to evaluate the most significant uncertainties, focusing on model and parametric uncertainties, in the most recent burden of disease estimates for fine particles in Finland. A short overview of possible sources of uncertainty will be given and selected individual uncertainties will be quantified.
9
2 LITERATURE REVIEW 2.1 ENVIRONEMENTAL BURDEN OF DISEASE (EBD) STUDIES Humans are continuously exposed to different environmental risk factors throughout their lives. Risk factors can be carried by different media; for instance, by drinking water, food or indoor and outdoor air (Prüss-Üstün et al., 2003.) Depending on the amount of exposure and the characteristics of the factors, the risks can vary from minor to even life threatening risks. It is important to distinguish which environmental factors pose a real threat to human health under levels to which people are exposed. Environmental burden of disease (EBD) studies determine the attributable burden of disease caused by different environmental factors in global, national or regional burden of disease (Gibson et al., 2013). That makes it possible to determine which environmental risk factors are more severe than others. EBD estimates are based on burden of disease (BoD) methods but the focus lies on environmental risk factors. These methods are widely recognized and increasingly used in environmental health impact assessments. They are used to measure the health loss caused by a risk factor on a population level. Burden of disease methods combine morbidity (years lived with disability, YLD) and mortality (years of life lost, YLL) into one comparable unit called DALY (disability adjusted life years) (Figure 1). (Knol, 2010.)
Figure 1 Disability adjusted life years (DALY) combine morbidity (YLD) and mortality (YLL) components. (Based on de Hollander et al., 1999 translated from Hänninen et al., 2016).
Hänninen and Knol (2011) describe four methods for calculating disease burden named 1 A, 1 B, 2 A and 2 B (see appendix Figure A1 and Figure A2). Methods 1A and 1B are for relative
10
risk functions and 2A and 2B for unit risk functions. Methods 1A and 2A can be used when there is background disease burden data available for the health outcome. Methods 1B and 2B are used in the cases when background disease burden data is not available and disability weights (DW) have to be applied. 1A, 1B and 2A are based on population attributable fraction (PAF) which indicates the percentage of the population which have attributable risk due to exposure. PAF is combined with background disease burden estimates to get the excess disease burden estimates. Environmental burden of disease estimates can be used to, for instance, raise awareness and help policy makers identify high risk populations and prioritize actions in health, and reduce the impacts of environmental health risks by basing policy actions or interventions on estimated health gains. (Prüss-Üstün et al., 2003.) The use of a common unit allows comparative evaluations and therefore, also helps with setting priorities (Knol, 2010). EBD assessments can also assist policy makers with weighting the pros and cons of alternative situations since the health gains can be estimated directly from the exposure. (Prüss-Üstün et al., 2003.) A few worldwide evaluations have estimated different risk factors’ share in the global burden of disease (GBD). The first one was commissioned by the World Bank in the 1990s. That study was done in collaboration with WHO and the Harvard School of Public Health. (Gibson et al., 2013.) It brought to light otherwise hidden or neglected health challenges like mental illness and road injuries, thus having a significant impact on health policy. This study was updated in 2002 and 2004. (IHME, 2013.) After that, the next comprehensive global burden of disease study was conducted by the Institute for Health Metrics and Evaluation (IHME) in 2012. The study was funded by the Bill & Melinda Gates Foundation and had some methodical changes compared to the original GBD study. Age weighting and discounting of the healthy life years had led to heavy debates. As a result, these were dropped and thus all healthy life years were considered equally valuable. (IHME, 2013.) The latest update of the GBD study was calculated for 2015 (Forouzanfar et al., 2016). In 2013, WHO published the new Global Health Estimates (GHE) study which has also been updated for the year 2015 (WHO, 2016a). They adopted methodical changes from the GBD 2010 study, and therefore these studies are not comparable with WHO’s earlier global burden of disease estimates. (WHO: Global Health Estimates (GHE).
11
2.2 UNCERTAINTIES IN ENVIRONMENTAL DISEASE BURDEN ANALYSIS Since the current evidence on the relationship between exposure and disease is solid enough for many environmental risk factors, it is possible to make quantitative estimates of the disease burden. However, many other risk factors have not been documented adequately. For instance, risk factors with long latency periods or nonspecific health effects may not be registered yet. (Prüss-Üstün et al., 2003.) Furthermore, knowledge and data needed in EBD estimates are often incomplete. In addition, the assessment can be very complex, leading to significant uncertainties of the results. (Knol et al., 2009.) So like all estimates, also environmental disease burden methods have some degree of uncertainty in them. Thus EBD assessments have to be interpreted with caution and should be viewed as the best current estimates of the magnitude of health problems caused by environmental factors. (Prüss-Üstün et al., 2003.) The need for uncertainty analysis in risk assessments have been recognized internationally. European Food Safety Authority (EFSA) has published draft guidance for uncertainty analysis of EFSA’s scientific assessments (EFSA Scientific Committee, 2016). It provides a flexible framework for uncertainty analysis. A first step of the uncertainty analysis is to develop assessment strategy. In assessment strategy the problem is formulated by clarifying the scope of the assessment and developing the conceptual framework for the assessment. The question formulation is an essential part of the assessment process. In the planning of the uncertainty analysis it is important to take into account possible time and resource limitations. (EFSA scientific committee, 2016). When the strategy is made, the possible sources of uncertainties can be identified. There are usually multiple sources of uncertainty and it is not necessary to evaluate all the uncertainties individually but to select some of the uncertainties for individual assessment. Individual assessment is needed for uncertainties which are not addressed or covered by the standard procedure. These case-specific sources of uncertainty can be evaluated quantitatively or qualitatively if qualitative assessment is not possible. Uncertainties which have not been assessed individually can be evaluated collectively in the combined evaluation. Individually qualitatively assessed uncertainties and uncertainties which are not individually assessed by any method are recommended to quantify by expert judgement. Those can be then combined to individually quantified uncertainties. Uncertainties which are not possible to include to the quantitative assessment should be described and added to the overall characterization of identified uncertainties with the quantitatively expressed combined uncertainty. Figure 2 illustrates this procedure. (EFSA Scientific Committee, 2016)
12
Figure 2 The illustration of the main elements of combined uncertainty analysis. (EFSA Scientific Committee, 2016)
In this thesis the focus lays on quantitative assessment of selected individual uncertainties. Uncertainties in environmental burden of disease estimates can arise from different sources. They can be divided into groups according to their nature. This classification can be done in several ways. The widely accepted classification compiled by the US Environmental Protection Agency (EPA) is used in this thesis (US EPA, 1992). Uncertainties are divided into three components: 1. Model uncertainty 2. Parameter uncertainty 3. Scenario uncertainty Model uncertainties may arise because of the gaps in the scientific theory which can lead to false predictions of causal inferences, errors in understanding relationships, and oversimplified models of reality. (Bailar & Bailer, 1999.) Here the focus lays on model uncertainties related to the link between exposure and health outcomes, the shape of the concentration response functions, the setup of a cutoff point and the exposure characterization. Parameter uncertainties are “traditional” uncertainties for which quantitative evaluations can be easier to conduct than for model uncertainties. They may arise from measurement errors (e.g. random errors in analytic devices and systematic bias), the type of data used (e.g. generic or surrogate), misclassification of subjects, random sampling errors, and other kinds of nonrepresentativeness in the analysis (Bailar & Bailer, 1999). Here parameter uncertainty is
13
called parametric uncertainty. The parametric uncertainties chapter will focus on uncertainties related to the exposure and relative risk estimates. Uncertainties related to input data like background disease burden and diagnostic problems will be shortly discussed as well. Scenario uncertainties are related to missing or incomplete information. They include errors in information, aggregation errors, errors in professional judgement and errors related to incomplete analysis. (US EPA, 1992.) The classification presented above is a rather simple way of classifying uncertainties. This classification is not strict since uncertainties may overlap and therefore even experts’ opinions may differ on categorizing some of the uncertainties (WHO, 2008). The US EPA classification of uncertainty has been used by, for instance, the International Programme on Chemical Safety (IPCS) (WHO, 2008). The above-mentioned classification will be used in this thesis, with the focus on model and parametric uncertainties. In addition, there have been several other attempts to classify uncertainties. Here detailed classification of uncertainties by Knol et al. (2009) (Table 1) will be presented shortly. They adopted some existing typologies and applied them with some further developments to uncertainties in environmental burden of disease assessments. Their classification has been used with some modification in, for instance, by Bouwknegt and Havelaar (2015).
Table 1 Typology of uncertainty (Knol et al., 2009).
14
Location of uncertainty indicates where the uncertainty arises. Knol et al., 2009 divided this category in three locations: model structure (referred to as model in this thesis), parameters and input data. Model location includes uncertainties from structure and from the relationship between the variables that describe the system. Parameters are the constants in functions (i.e. relative risk (RR) or disability weight (DW)) which define the relationships between variables. Uncertainties in input data can arise from instance concentrations and background disease burden or incidence data. (Knol et al., 2009.) The nature of the uncertainty describes the underlying cause of the uncertainty. It can be divided into two types: ontic and epistemic uncertainty. Epistemic uncertainty arises from the lack of knowledge. With more information, this uncertainty would decrease. (Bouwknegt and Havelaar, 2015.) Ontic uncertainty is related to the intrinsic properties of the system and it is many times referred to as variability (Knol et al., 2009). The range of uncertainty relates to the way it can be expressed, which can be either statistic or scenario. A statistical uncertainty can be expressed in statistical terms like confidence intervals. Scenario uncertainties are often referred to as what-if statements. They can only be specified in terms of a range of possible events and not in probabilities or in terms of a change. (Knol et al., 2009.) Recognized ignorance is uncertainty which can be recognized but for which useful estimates cannot be established due to incomplete understanding of the process (Knol et al., 2009). Thus the sources of uncertainty can be identified but the uncertainty cannot be taken into account in the modelling and is therefore ignored (Bouwknegt and Havelaar, 2015). Methodological unreliability is caused by weaknesses in methodological quality. They are often impossible to quantify but qualitative judgements can be used to express in which ways the scientific knowledge is limited (Knol et al., 2009). Value diversity among analysts refers to different personal values and normative judgements among analysts in scientific practice (Knol et al., 2009). Values can differ, for instance, between choices of model or dataset use. Value diversity is lower when the majority of analysts agree with each other on what procedure to follow. Uncertainties may also arise from how to deal with a number of non-representative datasets or when selecting study area or cohort. (Bouwknegt and Havelaar, 2015.) The next chapters will elaborate the various uncertainties arising from EBD assessments. It is good to keep in mind that the division into different groups of uncertainty is man-made and therefore there can be different views on which uncertainties belong to which group. In practice uncertainties can arise in overlapping areas which makes it difficult to classify uncertainties (WHO, 2008).
15
2.1 PARAMETRIC UNCERTAINTIES 2.1.1 Descriptive and normative parameters There are two kinds of parameters: descriptive and normative. Descriptive parameters are, for example, relative risks, background disease burden estimates, attributable fractions and exposure estimates. Descriptive parameters describe something as it is, whereas normative parameters value something. Normative parameters are related to, for instance, disability weights of health conditions, age weights and discount factors. In other words, those parameters are related to how we value things and they can be given very different values depending on the person. Therefore, normative parameters are subjective interpretations for which no true values exist. Contrary to descriptive parameters which do not depend on opinions and should be the same when the same data is used. (Knol et al., 2009.) Normative parameters create uncertainty because of their nature. When no true value exists different studies may use different values and all of them can get wrong results because it is simply not known what the true value is of the parameter they are using (Knol et al., 2009). Disability weights are a good example of a normative parameter. They are used to present the severity of a health condition in a comparable quantity. Disability weights are measured on a scale from 0 to 1, with 0 presenting a state in which a person is completely healthy and 1 a state which is equivalent to death. Disability weights are defined by surveys in which paired comparison questions are used. Respondents are asked to select the person which they consider the healthiest individual between two hypothetical individuals with different health states. With these surveys 183 health states were covered. (Salomon et al., 2015.) There are several uncertainties related to defining disability weights. For instance, the question of “who is healthier” might be ambiguous and difficult to answer. There are problems comparing disabilities and diseases with each other’s; one might for example consider a person with an amputation to be healthy even though they are partly disabled. Even though one might not consider that person to be sick, they probably still think that the condition is not desirable and that being disabled might limit that person’s life to some extent. But that might not be seen in a survey were the only question is “who is healthier”. (Nord, 2015.)
2.1.2 Random and systematic error Measurements suffer from two types of errors: random error which leads to unprecise results and systematic error which affects the accuracy of the estimate. Random errors in direct measurements are the most studied and the best understood uncertainty. This uncertainty depends on the range of the variation between observations and the number of observations taken. Random error can be decreased by taking more measurements i.e. increasing n. There are several statistical techniques for quantifying this uncertainty like, for instance, confidence intervals and standard deviation. (Morgan and Henrion, 1992.) When random errors get smaller, the results become more precise.
16
Statistical confidence intervals indicate the probability that the true value exists within the range of the intervals. For instance, a 95% confidence interval (95% CI) indicates that the true value has a 95% probability of being in the given range. In other words, if the calculation is repeated 100 times, in theory 95 times the true value should be in the confidence interval range and five times it falls outside of this range. (Morgan and Henrion, 1992.) Even if the measurements would be very precise, they may be inaccurate due to systematic error. Systematic error can arise from biases in measurement instruments like imprecise calibration and mistakes in the experimental procedure like faulty reading of the measurement results. Systematic error is the difference between the true value and the mean value towards the measurements converge when measurements are repeated. (Morgan and Henrion, 1992.) Systematic error is also sometimes referred to as bias. Bias is commonly classified in selection bias, information bias and confounding. Selection bias refers to the selection of a study population from the source population. Information bias refers to the misclassification of a study population with respect to disease or exposure status. Information bias thus concerns selected study population. Confounding is caused when effects of the studied exposure are mixed with effects of some other factor. It can occur when exposed and nonexposed study groups have different background disease risks. When diseased and healthy or exposed and unexposed people have an equal chance to be misclassified a false negative results is tend to occur. This is called non-differential misclassification and it leads to a relative risk to go towards null value of 1.0. It is important to adjust study design for possible confounding factors like age and smoking status. Confounders can cause bias on the exposure-response association, and in an extreme situation it can even change the direction of an effect. (Pearce et al., 2007.) Systematic error can be decreased by carefully designing the experiment and calibrating the measurement instruments while also carefully analysing the assumptions that are used in the calculations. Even if a measurement would be adjusted for all the known sources of systematic error, the unknown sources would still remain. Because those sources are unknown, their possible magnitude is hard to estimate. Since systematic error cannot be reduced by additional observations, it becomes a dominant error in large studies. (Morgan and Henrion, 1992.) It is important to estimate the likely direction and magnitude of the unknown biases which cannot be avoided or controlled. Systematic error can only be reduced by changing the study design. (Pearce et al., 2007.)
17
2.2 MODEL UNCERTAINTIES 2.2.1 Epidemiological studies and relative risk A lot of research has been done on the adverse health effects of air pollution and many epidemiological studies have shown that there is a wide range of adverse health effects associated with air pollution. However, it is difficult to relate health effects to specific health endpoints to find out the attributable fraction caused by single air pollutant and explain which properties of the air pollutant are causing the effects. Air pollution always consists of a mixture of different air pollutants. Since people are exposed to this mixture it makes it difficult to assess whether a certain health effect is caused by one particular air pollutant or if that pollutant partly or completely correlates with other pollutants in the mixture leading to a dependency error. A dependency error can occur through the incorrect conclusion that there is a correlation between an adverse health effect and particular air pollutant. (WHO, 2016b) One known example of correlation between air pollutants is between fine particles (PM2.5) and nitrogen dioxides (NO2). In WHO’s recommendation it was estimated that NO2 can correlate with PM2.5 up to 33% (Héroux et al., 2015). Currently, air quality standards and abatement strategies are based on the total mass of suspended particles. Thus all particulates are treated as equally toxic. This assumption does not fit well with basic toxicological principles. (Schwarze et al., 2006.) Particulate matter is a complex mixture of solid and liquid particles mainly consisting sulphates, nitrates, ammonia, sodium chloride, black carbon, mineral dust and water (WHO, 2016a). The chemical and physical properties of particulate matter can vary substantially between different emission sources which can modify the relative toxicity of PM2.5 (Tainio et al., 2010). Yet it is not known which emission sources or properties of PM2.5 are responsible for the adverse health effects (Burnett et al., 2014). However, several studies have been done on this topic and, for instance, Siponen et al., 2014 studied the short term effects of PM2.5 from five source categories; regional and long-range transport (LRT), traffic, biomass combustion, sea salt and pulp industry. Association between source specific PM2.5 and markers of systematic inflammation were analysed. The best evidence was found for the relation of levels of air pollution and inflammation in the cases of LRT, traffic and biomass combustion. (Siponen et al., 2014.) Causality is an essential point when exposure to a certain air pollutant is being linked with a specific health outcome. Causation can never be proven so we cannot know for sure that some exposure X causes a health outcome Y. Though there are some criteria which can make the association stronger (Lucas & McMichael, 2005). Perhaps the most popular criteria for evaluating causality are Bradford Hill’s criteria of causality, which were published for the first time over 50 years ago (Hill, 1965). Bradford Hill’s criteria have nine points which are briefly described below (Lucas & McMichael, 2005):
18
1. Strength
A strong association is more likely to be causal than a weak association.
2. Consistency
A causal explanation for an association is more likely to be true if the same answer has been received in different situations, proand retrospectively and in different study populations.
3. Specificity
When exposure is associated with a specific group of health outcomes, the causal association is strengthened.
4. Temporality
Exposure must come before a health effect.
5. Biological gradient The likelihood of a causal association is increased if a biological gradient or dose-response curve can be demonstrated. But there are cases where the relationship is difficult to demonstrate and furthermore, thresholds and non-linearity of the association can make it more difficult to determine. 6. Plausibility
A causal association is biologically plausible. Though Bradford Hill notes that what we see as biologically plausible depends on the biological knowledge at that time. Thus, he says that this criterion cannot be demanded.
7. Coherence
The cause-and-effect interpretation of an association should follow the known facts of natural history and biology of the disease.
8. Experiment
Manipulation of exposure changes the frequency of the outcome. Bradford Hill saw this criterion as the most important support of a causal relationship.
9. Analogy
Clear-cut analogies may make otherwise weak associations stronger.
The above list can help in considering whether a reported relation could be causal, but it is not a check-list with which someone could distinguish causal and non-causal relations. Defining causal relations is difficult because multiple factors can be related to a certain outcome. On an individual level all factors which has led to the appearance of an outcome, are as strong. On a population level though it is possible to find out how some of the factors are stronger related to the outcome. (Rothman, 2012.) Ken Rothman has proposed a conceptual model of causation called The SufficientComponent Cause Model also referred as “The causal pie model” and “Rothman’s causal pies”. There are three pies on the model and each pie presents a theoretical causal mechanism for a given outcome (Figure 3). One pie consists of different factors which are single component causes. These causes together form a causal mechanism which can also be called a sufficient cause. All three pies are sufficient causes of the specific outcome even though they consist of different components. If any of the single components are lacking from the sufficient cause, the outcome will not occur. That leads to a conclusion that we do not have to know all the single causes of sufficient cause for some outcome, since if reducing one we can
19
already prevent the adverse health outcome. (Rothman, 2012.) Like in the example, in all three pies there is a single cause A (Figure 3). If we remove that, the outcome will not occur due to any of the presented causal mechanisms.
Figure 3 The causal pie model of three sufficient causes of a disease (Rothman, 2012).
On an individual level, all the single components in the causal pie are as important since they are all needed for the health outcome to occur. On a population level though, epidemiologist can define causes to be stronger or weaker. That is because, on a population level, it is possible to see whether some cause is involved in a large proportion of cases (a strong cause) or only in some cases (a weak cause). Therefore, in a causal pie model, factors can be divided into causes and factors which are not causes. It should be kept in mind though that the strength of a factor can vary from population to population. (Rothman, 2012.) The population attributable fraction is between 0─100% but in cases where a factor has a protective effect the PAF value is negative. The single component causes can attribute to the health outcome up to 100%. There can be several single causes in a causal mechanism. They sum up together and cause the outcome. Their sum can add up to more than 100%. That is due to interactions between causes. One example of interacting causes could be alcohol consumption, tobacco smoking and cancer. (Rothman, 2012.) Air pollutants can cause an additional risk of occurrence for adverse health effects. The risk relation between an adverse health effect and an air pollutant is often described with a concentration response function (CRF). (WHO, 2016b.) A dose-response or concentration response relationship exists when change in the exposure causes a quantifiable change in the effect. Relationship can also exist in cases where observed effect is not quantifiable but it is present or not present when the percentage of the population responding with effect depends on the exposure. (Moffett et al., 2014) Concentration response functions indicates how the excess risk changes when exposure changes. CRFs are usually derived from epidemiological studies which are facing the problem that people are exposed to air pollution mixture which adds uncertainty to the CRFs. In
20
addition in epidemiological studies there are always some assumptions made during the study which can cause some degree of uncertainty in the results. (WHO, 2016b.) Epidemiological studies provide the most relevant information for hazard identification because studies are done on humans. Though there are various difficulties associated with obtaining and interpreting epidemiological information. One of the problems with epidemiological studies is that it is difficult to get exact levels and duration of exposure because, unlike toxicological studies, epidemiological studies are not controlled experiments. The main advantage of toxicological studies is that they are well controlled but the biggest problem is that they are not tested on humans, expect in some few cases, and it can be difficult to extrapolate results from an animal or cell experiments to humans. (Committee on Risk Assessment of Hazardous Air Pollutants, 1994.) In Table 2 the pros and cons of epidemiological, toxicological and clinical studies are listed. Table 2 Pros and cons of epidemiological, toxicological and clinical studies (Based on Klaassen, 2013)
Epidemiological studies + Relevant species
- Poor quantitative exposure assessment
+ Natural exposure
- Low power to find associations
+ Long-term and low-level effects
- Usually short term - Many covariates
Toxicological studies + Experimental conditions well defined
- Interspecies extrapolation
+ Minimal amount of confounders
- High doses
+ Measurements easier
- Short exposure times
+ Show biological plausibility Clinical studies + Controlled exposure
- Artificial exposure
+ Few covariates
- Only acute effects - Health hazards - Need of volunteers
Concentration response functions are many times studied in other places than where they are later applied. Extrapolating CRFs found in one population to others where people can have different lifestyles, age structures and medical care adds uncertainties to the evaluation (Anenberg et al., 2016). Most of the epidemiological studies are conducted in North America
21
and Europe where concentrations are usually much lower than in developing countries (Anenberg et al., 2016). That leads to a problem that CRFs have many times not been studied in their whole range of exposure where they are later on applied (WHO, 2016b). The HRAPIE working group states in the report that mortality data for natural mortality tends to be more reliable than for cause-specific mortality. However, air pollution has not been linked reliable to all natural causes of death. That is why cause-specific assessments are defensible. (WHO, 2013.) HRAPIE recommendations suggest using natural all-cause mortality in Europe because in Europe risk estimates and precise background national data for all-cause mortality are available. The GBD project used cause-specific mortality because patterns of causes of death vary greatly globally. (Héroux et al., 2015.) Whether one chooses to use natural or cause-specific mortality can have a significant impact on the results.
2.2.2 Shape of the concentration-response function The disease burden estimates depends greatly on the shape of the concentration-response function (CRF) in addition to the magnitude of the coefficient. Associations between fine particles and increased natural mortality and cause-specific diseases have been reported for annual ambient concentrations from approximately 5 to 30µg/m3. Those associations have not been well evaluated in areas where annual PM2.5 exposures can exceed 100µg/m3. There is a concern that using relative risk estimates derived from studies at lower PM2.5 concentrations would produce unrealistically large estimates in those areas. (Burnett et al., 2014.) To tackle this issue and cover the entire range of exposure the IHME institute has developed the integrated exposure-response (IER) model (Lim et al., 2012). Traditionally concentration response functions for PM2.5 and adverse health effects are thought to be linear or log-linear. However, recent studies suggested that the concentration response function is likely to be supra linear (concave) for a wide range of exposure, including high levels. Supra linearity has also been suggested for low concentrations. Linear and supra-linear functions differ slightly at low concentrations and greatly when the concentrations are high (over 40µg/m3) (Figure 4) (Pope et al., 2015).
22
Figure 4 Illustration of supra linear and linear shape of concentration response function (X-axis PM2.5 (µg/m3)) (Pope et al., 2015).
Integrated exposure response functions cover the entire range of exposure. This has been made possible by combining evidence from different studies concerning ambient air pollution (AAP), second-hand smoking (SHS), household air pollution (HAP) and active smoking (AS). (WHO, 2016b.) The IER functions are drawn according to the type specific relative risks from different sources of exposure. The IER model includes causes of mortality in adults for: ischemic heart disease (IHD), cerebrovascular disease (stroke), chronic obstructive pulmonary disease (COPD) and lung cancer (LC) (Figure 5). Furthermore, they have developed function for the incidence of acute lower respiratory infections (ALRI) in children under 5 years old (Figure 6). (Burnett et al., 2014.)
23
Figure 5 Integrated exposure response functions for A: ischemic heart disease (IHD), B: stroke, C: chronic obstractive pulmonary disease (COPD) and D: lung cancer (LC) (Burnett et al., 2014).
Figure 6 Predicted values of IER model (solid line) and 95% confidence intervals (dashed line) and typespecific RRs (points) and 95% confidence intervals (error bars) for ALRI in infants (Burnett et al., 2014).
24
2.2.3 Threshold The integrated exposure response model uses cutoff value which above the excess risk is modelled. The IHME GBD2010 study invented a theoretical-minimum-risk exposure distribution (TMRED) term for the cutoff they are using. They set the TMRED value for PM2.5 at 5.8-8.8µg/m3 (Lim et al., 2012). The TMRED was defined based on two criteria (Héroux et al., 2015): i)
the availability of evidence from epidemiological studies that support a continuous reduction of risk from the current levels to the chosen TMRED
ii)
a counterfactual population distribution of exposure that is theoretically possible at the population level
Lim et al. (2012) suggested using a positive counterfactual concentration which is bound by the minimum concentrations observed in the studies used for estimation of risk and some low percentile of the PM2.5 distribution. The largest cohort study on exposure distribution of PM2.5, the CPS II cohort (Krewski et al., 2009) was used. They used the fifth percentile for the upper bound (8.8µg/m3) and minimum percentile for the lower bound (5.8µg/m3). With this approach the IER model does not calculate excess risk when long-term exposure to PM2.5 is less than 5.8µg/m3. Burnett et al. (2014) claims that even though excess risk may extend below 5th percentile of the distribution, the risk estimates are statistically unstable and highly uncertain. Though, they are not suggesting that there is convincing evidence of no adverse health effects below any specific concentration based on biological considerations i.e. biological threshold. The IER model can be adapted for different counterfactuals in the case where there is new evidence supporting an increase of risk at lower concentrations. There is, for instance, already one study in Canada which shows a positive association between PM2.5 and mortality at concentration as low as 2µg/m3 (Crouse et al. 2012). (Burnett et al., 2014.) The WHO working group’s recommendations for concentration response functions are relative risk functions without a threshold, but it is assumed that there are no great differences between the results when these recommended CRFs are used instead of IER functions used in the GBD2010 study. That is because medium to long term concentrations of PM2.5 in most areas of Europe are not below 5.8µg/m3. (Héroux et al., 2015.) The CAFÉ (Clean Air for Europe) project adopted a “no threshold” assumption for their analysis like some other earlier major health impact assessments (ExternE, Künzli and colleagues, US EPA etc.). They looked at thresholds at the population level, meaning that a threshold is a concentration below which there is no increase in risk of adverse health effects in any of the exposed populations at risk. This interpretation was the most consistent with the evidence and understanding which was available at that time. Though, in CAFÉ-project the
25
health effects were calculated only for anthropogenic contributions to air pollution. (Hurley et al., 2005.) WHO Air Quality Guidelines recommend aiming for the lowest concentration of PM possible. They set air quality guidelines in a global update in 2005 (WHO, 2006) for annual mean PM2.5 at 10µg/m3. WHO claimed that this level is the lowest level at which total, cardiopulmonary and lung cancer mortality have been shown to increase with more than 95% confidence in response to long-term exposure to PM2.5. (WHO, 2006) This guideline value is already ten years old and since then more research has been done on the health effects of fine particles in lower exposure levels (e.g. Crouse et al., 2012). WHO is currently updating the global air quality guidelines (WHO, 2017).
2.2.4 Exposure modelling Estimating population exposure levels is challenging because pollution levels vary over time and space. People can also move several times throughout their lifetimes making the estimation even more difficult. (Crouse et al., 2015.) Ground monitors of air pollutants do not have a full geographical coverage. Usually disease burden studies have to rely on modelled exposure. Modelled exposure levels have to also be used in cases where disease burden is calculated for future exposure as a result of some policy implementations or technological improvements. (WHO, 2016b.) Model uncertainties in exposure estimates can arise from modelling errors (i.e. not all parameters are considered) or from relation or dependency errors (i.e. incorrect conclusions from correlations) (WHO, 2008). A model is always a simplification of the reality. Uncertainties related to modelled exposure estimates are based upon modelling errors and relation i.e. dependency errors. Modelling errors are related to model boundaries, assumptions, level of detail, extrapolation and implementation, and technical aspects of the model. Dependency errors arise from a lack of consideration or incorrect inference of dependencies between parameters. (WHO, 2008.) To be precise, even though the term of exposure is used here, it does not refer to exposure but to concentration. Disease burden studies usually utilize concentration data instead of exposure data because exposure data would be much harder to get on a population level than concentration data. When exposure estimates are not available, outdoor concentrations can be used as proxies (Knol et al., 2009). Therefore, in the field of air pollution there is much more known about the relationship between air pollution concentrations and health effects than about exposure and health effects. (Settimo, 2015.) Exposure to air pollutants mainly occurs indoors but measurement stations for air pollutants are usually outside monitoring stations. Exposure to ambient air pollutants does also occur inside but the diffusion level at which the outdoor air pollutants infiltrates indoors depends on several factors like ventilation and the indoor-outdoor temperature gradient. Furthermore, personal factors affect the exposure rate of individuals. For instance physical activity,
26
respiratory rate and metabolism are factors which can influence an individual’s exposure rate. Therefore, individuals can have different exposure rates, even though the concentration of air pollutants they are exposed to is the same. (Settimo, 2015.)
27
3 THE AIM OF THE WORK The aims of this thesis are to (i) select uncertainties for the uncertainty analysis of the disease burden estimates of fine particles and (ii) to estimate their magnitude and relative significance on the corresponding estimates. The selected uncertainties will be quantified with relation to the most recent Finnish national estimate for the disease burden of fine particles to identify the most significant uncertainties. The range of the disease burden will be evaluated based on the quantified uncertainties. The specific objectives are: Quantify the key parametric uncertainties affecting health impact estimates a. Accuracy of the relative risk estimates resulting from epidemiological studies b. Accuracy of the exposure: uncertainties in the estimated population exposures Quantify the uncertainties related to model choices: c. Exposure characterization using average exposure and full exposure distribution d. Effect of the concentration response functions’ shape on the disease burden estimates using three shapes: log-linear, linear and supra-linear e. Impact of a cutoff point on the integrated exposure response function f. Selection of health endpoints: comparison of natural mortality, causespecific mortality calculated using relative risk for natural mortality and cause-specific mortality calculated using cause-specific relative risks
28
4 MATERIALS AND METHODS 4.1 DISEASE BURDEN CALCULATION METHODS The disease burden calculations were based on comparative risk assessment methods. These methods have been introduced by Prüss-Üstün et al. (2003) and they have been used in several national and multi-national health impact assessments such as the Environmental Burden of Disease in European countries (EBoDE) -project (Hänninen & Knol, 2011). The uncertainty analysis was carried out based on the ISTE-project’s disease burden calculations for PM2.5. The detailed representation of the methods of the ISTE-project can be found on a methodological report by Lehtomäki et al., 2015. Disease burden (BoD) is calculated by combining mortality (YLL) and morbidity (YLD) compounds into one unit called disability adjusted life years (DALY) (equation 1). The YLL component is calculated by multiplying the number of deaths at each age by the standard life expectancy at the age of death. The YLD component is calculated by multiplying the number of incident or prevalent (Lim et al., 2012) cases in the population by disability weight and average duration of disability. (Prüss-Üstün et al., 2003; Gibson et al., 2013.) 𝐵𝑜𝐷 = 𝑌𝐿𝐿 + 𝑌𝐿𝐷
(1)
Method 1A was used in the ISTE-project for PM2.5 disease burden calculations (see attachment Figure A1). WHO’s GHE 2012 background disease burden data were applied (WHO: Global health estimates for 2000-2012). As in the recent global burden of disease studies, also here age weighting and discounting were not applied and calculations were based on prevalence of the disease instead of incidence. The methods used for the exposure assessment in the ISTE-project are explained in detail in a report by Korhonen et al., 2015. The exposure assessment was mainly based on the Finnish measurement stations’ network and their data for 2013 (Table 3). Table 3 Finnish PM2.5 measurement stations concentrations in 2013 by station types (translated from Korhonen et al., 2015).
Station type
Number of stations (n)
Mean concentration (µg/m3)
Standard deviation (SD) (µg/m3)
All stations
31
6.6
1.6
Traffic
10
7.0
1.2
Industrial
8
7.0
1.7
Suburban background
3
7.6
1.1
Urban background
5
6.0
0.9
29
The list of all concentration response functions for PM2.5 which were identified in the ISTEproject can be found in the appendix 2 (Table A1). Three concentration response functions were selected to calculate disease burden of PM2.5 (Table 4). Selection of concentration response functions was done based on the latest recommendations (Héroux et al., 2015). The recommendations’ pollutant-outcome pairs which were classified as having enough data available to enable quantification of effect and the effects being additional (group A*) were chosen for the analysis. An exception was made for natural mortality (Table 4): the concentration response function was used to calculate cause-specific mortality, instead of mortality from all natural causes. Mortality was calculated for cardiovascular diseases, lung cancer and respiratory diseases and infections. Further information will be provided in chapter 4.2 where methods for the use of specific and non-specific mortality are explained. Table 4 The selected concentration response functions for PM2.5 (Héroux et al., 2015).
a
Health endpoint
Relative risk (95% CI)
Mortality (natural causes)a
1.062 (1.04-1.083)
Cardiovascular diseases, hospital admissions
1.0091 (1.0017-1.0166)
Respiratory diseases, hospital admissions
1.019 (0.9982-1.0402)
The function was used only for cause-specific diseases
The uncertainty analysis was mainly conducted by using a model in Microsoft Excel which was created for disease burden calculations during the ISTE-project, here referred to as the ISTE-model. The model includes all the concentration response functions found through literature review, concentrations and confidence intervals of exposure estimates for 14 air pollutants and WHO’s GHE2012 background disease burden data. The model includes several macros which enables disease burden calculations for all included air pollutants and for different combinations of pollutant groups as well as health endpoints and health endpoint groups. It is possible to add other risk factors and new concentration response functions to the model. For this thesis the calculations were only made for fine particles.
30
4.2 ANALYSIS OF SELECTED UNCERTAINTIES Conceptual overview of the individual sources of uncertainties in the environmental burden of disease is based on WHO 2016b and complemented by Knol et al. 2009 and author’s judgement (Figure 7). The uncertainties for the quantification were selected from the identified uncertainties. The selected uncertainties were related to health endpoints selection, shape of the concentration response functions, uncertainties in risk estimates and in exposure. Selection criteria were that uncertainties are considered to be significant and/or they were possible to quantify with the data from the ISTE-project.
Figure 7 Conceptual overview of the model and parametric uncertainties related to different phases of disease burden assessments.
31
Uncertainties in risk and exposure estimates Parametric uncertainties related to the risk estimates were quantified by using risk estimates’ 95 % confidence intervals (CI) (Table 4). In the ISTE-project the uncertainties for the exposure were estimated and 95% confidence intervals were defined (Korhonen et al., 2015). The population weighted annual mean was 6.8µg/m3 (95% CI: 6.1; 7.5µg/m3). Uncertainties in the relative risk estimates were calculated keeping the exposure estimate central and calculating the minimum and maximum estimates by taking the lower or higher confidence interval values for risk estimate. Calculations were similar for the uncertainties in exposure estimates but in the calculations the lower and upper exposure estimates were used with the central estimate for relative risk.
Average exposure vs. exposure distribution The population weighted average exposure was used for the calculations in the ISTE-project. However, that is a simplification and does not take into account the distribution of exposure. The exposure distribution and the population weighted average exposure for PM2.5 in Finland are defined in Korhonen et al., 2015 (Figure 8).
Figure 8 Cumulative log-normal PM2.5 exposure distribution and population weighted annual average.
32
The population attributable fractions were calculated for both: exposure distribution and average exposure and then compared. The PM2.5 exposure distribution was assumed to be lognormally distributed with 1.6µg/m3 standard deviation (Korhonen et al., 2015). The mean exposure estimate was 6.8µg/m3. The PAF estimates for the average exposure and for the distribution and the corresponding disease burden estimates were calculated for natural mortality (Table 4). The relative risks (RR) were reported per 10 µg/m3 and they were transformed per 1µg/m3 (equation 2). Relative risks at prevailing exposure were calculated using equation 3. For exposure distribution the average exposures of the percentiles were used. 𝑅𝑅1 = 𝑅𝑅1/10
(2)
𝑅𝑅𝐸 = 𝑅𝑅1 E
(3)
Where RR is the risk increment per 10µg/m3, RR1 is the risk increment per 1 µg/m3, RRE is the relative risk per prevailing exposure and E is the PM2.5 exposure. The population attributable fractions for each percentile were calculated using the following equation: 𝑃𝐴𝐹 =
𝑓×(𝑅𝑅 𝐸 −1) 𝑓×(𝑅𝑅 𝐸 −1)+1
(4)
The PAF values for exposure distribution were calculated for 2, 3, 4, 5, 6, 10, 100, 1k, 10k, 100k and 1M quantiles. They were summed up to get the overall PAF from exposure distribution which was then compared with the PAF from average exposure. The difference between mean exposure and exposure distribution was calculated using the average for mean exposure as a reference point.
Shape of the concentration response function and cutoff point Two different shapes, linear and supra-linear, were compared with the log-linear function used in the ISTE-project. This was done to analyse the impact of the concentration response function’s shape on the disease burden estimates. The function for log-linear and linear shapes was the relative risk function for natural mortality (1.06 per 10µg/m3) (Table 4). Relative risk per unit of exposure (RR’) was calculated using the function 2 and 5. Linear unit risk function was derived from the relative risk function using the equation (5). 𝐸
𝑈𝑅 = 10 × (𝑅𝑅 − 1) + 1
(5)
33
To compare the differences in the shape of the functions the population attributable fractions were calculated using formula 4. For unit risk function the RRe was replaced with UR. The supra-linear integrated exposure response function was compared with the log-linear natural mortality function. Ischemic heart disease was chosen as a supra linear function from the health endpoints which were included in the integrated exposure response model. Ischemic heart disease has the highest background disease burden from all the health endpoints included in the IER-model (Figure 9).
Figure 9 Background disease burden estimates (WHO: Global health estimates for 2000-2012) in Finland for the health endpoints which were included in the IER model (Burnett et al., 2014).
Relative risk data for the supra-linear IER function was downloaded (GHDx: Global burden of disease study 2010 (GBD 2010) - Ambient air pollution risk model 1990-2010)). The relative risk data was reported per 1 µg/m3 increases of PM2.5 exposure and RR values were reported using 3 significant numbers. The data was used as such. The comparison was made throughout the whole exposure range, 0-300 µg/m3 (Figure 10).
34
Figure 10 Supra-linear integrated exposure response (IER) function with 95% confidence intervals for ischemic heart disease and relative risk (RR) function for natural mortality. The risk increase is for annual mean concentrations of PM2.5.
Log-linear relative risk function starts to increase from 0µg/m3 onwards but the IER function starts to increase after the cutoff value which is the so called TMRED value. Figure 11 shows RR and IER functions at low concentrations where the effect of TMRED can be seen clearly.
Figure 11 Supra-linear integrated exposure response (IER) function and its 95% confidence intervals for ischemic heart disease and relative risk (RR) function for natural mortality at lower PM2.5 exposures.
35
The effect of the cutoff value was tested by shifting the IER function to start from 0, 2 and 6µg/m3 (Figure 12). The form of the function was kept the same. According to the author’s knowledge, there are no published predictions of the IER model below TMRED value. Therefore this method was used to get a rough estimate of how different cutoff values would affect the burden of disease estimates. Exposure distribution was used for these calculations because the shape of the IER function varies along the concentration range and at low concentrations it is very sensitive to point estimates due to the cutoff point.
36
A
B
C
Figure 12 The supra-linear integrated exposure response function shifted into three different cutoff points: A) 6 µg/m3, B) 2 µg/m3 and C) 0 µg/m3 while the shape is kept same. Figure shows how much of the PM 2.5 exposure distribution will be covered by the supra-linear function at each cutoff point.
37
Selection of health endpoints The HRAPIE workgroup recommended use of natural mortality for quantification of PM2.5 health impacts (Héroux et al., 2015). In the ISTE-project mortality was calculated using the recommended function but only including background disease burden from cardiovascular diseases, respiratory diseases and infections and lung cancer (Table 5). All-cause natural mortality was calculated to evaluate the difference between using natural mortality which includes all causes of death, excluding injuries, as opposed to specific causes of death which were included in the ISTE calculations. All-cause natural mortality and cause-specific morality were calculated using method 1A (Appendix 1) and the background disease burden for natural mortality which was calculated by subtracting disease burden caused by injuries (GHE code 151) from all causes (GHE code 0). Morbidity (YLD) was not included in the calculations for all-cause mortality.
Table 5 Background disease burden estimates for selected health endpoints in Finland in 2012 (WHO: Global health estimates for 2000-2012). Disease burden in disability adjusted life years (DALY), mortality (YLL), morbidity (YLD) and deaths.
Disease burden Health endpoint DALY
YLL
YLD
Deaths (n)
YLL/death
1,420,000
835,000
584,000
47,700
18
Cardiovascular diseases
343,000
305,000
37,200
20,300
15
Respiratory diseases
60,000
28,000
32,200
1,620
17
Respiratory infections
12,700
7,160
5,500
381
19
Lung cancer
47,600
47,100
524
2,150
22
All natural causes
Cause-specific disease burden was calculated with two approaches: using concentration response function for natural mortality and deriving cause-specific diseases from that as in the ISTE-project, and using separate cause-specific concentration response functions. For causespecific calculations both, mortality and morbidity were taken into account. The causespecific concentration response functions were from Hänninen et al., 2011 and they were based on Pope et al. (2002) and WHO (2006). The coefficients per 1µg/m3 increase in PM2.5 were: 1.0077 (95% CI: 1.0020, 1.0132) and 1.012 (95% CI: 1.004, 1.020), respectively for cardiopulmonary diseases and lung cancer. Both functions were for adults over 30 years old.
38
5 RESULTS 5.1 COMPARISON OF THE QUANTIFIED INDIVIDUAL UNCERTAINTIES The magnitudes of individual uncertainties are presented as differences in percentages in comparison to the model choices made in the ISTE-project (Figure 13). The original model choices are used as reference points with regard to the results from alternative methods are compared. The parametric uncertainties are calculated for the central estimate of the disease burden of fine particles. A comparison between exposure distribution and average exposure was done by using the average exposure as a reference point. In a comparison of the different shapes of the concentration response functions the reference point was natural mortality calculated using a log-linear function. Health endpoint comparison was done using a causespecific mortality derived from natural mortality as a reference point.
Figure 13 Differences caused by the individual uncertainties on the disease burden of PM 2.5. The results from the calculation methods applied in the ISTE-project are used as reference points (diamonds) and the percentages indicate the difference between the reference point and the alternative method. (RR = relative risk, CI = confidence interval, CPD = cardiopulmonary disease).
39
Parametric uncertainties are rather symmetrically spread around the best estimate of PM2.5 disease burden. From parametric uncertainties the uncertainty rising from the 95% CI of concentration-response functions are larger than from the 95% CI of exposure estimates. Model uncertainties are only causing changes to one way, either increasing or decreasing the estimate. All-cause mortality would change the estimate most dramatically. It would lead to 79% increase in the estimate in comparison to cause-specific mortality which was used in the ISTE-project. While using cardiopulmonary and lung cancer health endpoints would have led to 36% increase. Exposure distribution and linear unit risk leads only to small increases of the disease burden. When supra-linear IER function with cutoff at 6µg/m3 is used it would lead to 23 % decrease of the disease burden. All the single uncertainties, expect selection of health endpoints, were smaller than 1/3 of the best estimate of PM2.5 disease burden. None of the individual uncertainties would change the order of magnitude of the result.
40
5.2 EXPOSURE AND RELATIVE RISK ESTIMATES The disease burden of PM2.5 estimated in the ISTE-project was 20,800 DALY in 2013. Uncertainties from the epidemiological studies on the relative risk caused uncertainty of 10,000 DALY which is 48% of the disease burden. Uncertainties from the exposure estimates were together 4,100 DALY which is 20% of the disease burden. Uncertainties in the exposure were smaller than uncertainties related to concentration response functions (Figure 14).
Figure 14 Disease burden estimates of fine particles uncertainties related to 95 % confidence intervals for concentration response functions (CRF) and for exposure.
The uncertainties from the concentration response functions confidence intervals were from 15,700 to 25,600 DALY and from exposure estimation 18,800 to 22,900 DALY (Table 6).
Table 6 The disease burden estimates and differences in comparison to the central estimate for confidence intervals (CI) of concentration response functions (CRFs) and exposure.
CRFs 95%CI Disease burden (DALY) Difference
Exposure 95%CI
Upper
Lower
Upper
Lower
25,600
15,700
22,900
18,800
23 %
-25 %
9.8 %
-9.9 %
41
Average exposure vs. exposure distribution In the ISTE-project a population weighted average exposure was used instead of the whole exposure distribution. This led to an error which is -4.13 % when compared to estimate calculated using exposure distribution with 1 million quantiles and 6.81µg/m3 mean exposure. Population attributable fraction (PAF) for natural mortality was 4.01% and 4.18% using population weighted average exposure and exposure distribution, respectively. When exposure distribution is divided to 100 or more quantiles the relative error settles around 4.1% (Figure 15).
Figure 15 Population attributable fraction (PAF) values for using exposure distribution and population weighted average exposure. The relative error is on a secondary axis.
42
5.3 CONCENTRATION RESPONSE RELATIONSHIP The log-linear, linear and supra-linear shapes of the concentration response functions lead to differences in the population attributable fractions thus affecting the disease burden estimates (Figure 16). The log-linear and the linear shaped functions have very small differences on relative risk estimate on small concentrations. Also the effects on population attributable fraction (PAF) are very small especially on low concentrations. At PM2.5 concentration of 6.8µg/m3 the difference in PAF is 0.0035 %. The differences are bigger when log-linear and supra linear shapes are compared. Supra-linear function gives higher relative risk estimates on concentrations of 13 to 116µg/m3 with cutoff value 6µg/m3. The cutoff value in the IER function does affect the risk estimates greatly especially at low concentrations (Figure 11).
Figure 16 Log-linear, linear and integrated exposure response (IER) function and their effect on the population attributable fraction (PAF).
When IER function is forced to start from 0µg/m3 (TMRED 0) it gives higher risk estimates since it will then cover the whole PM2.5 exposure distribution. The change can be seen the best at the low concentrations. Different TMRED values affect the disease burden estimates greatly (Figure 12). Use of the IER function with 6µg/m3, 2µg/m3 and 0µg/m3 cutoff points would change the ISTE-estimate -23%, -16% and 31%, respectively.
43
5.4 SELECTION OF HEALTH ENDPOINTS Use of coefficient for all-cause (natural) mortality recommended by HRAPIE working group leads to 32,100 DALY/a when background disease burden of all natural causes of death are taken into account. When same coefficient is used for cause-specific diseases (cardiovascular diseases, respiratory diseases and infections and lung cancer) the disease burden is 17,900 DALY/a, which is remarkably smaller estimate (Figure 17). Selected mortality health endpoints can only explain 56 % of the all-cause mortality’s disease burden. Cardiovascular diseases are the main cause of disease burden from the selected causes of deaths.
Figure 17 Disease burden of PM2.5 related to natural mortality (A) and cause-specific mortality (B). Both A and B are calculated using a concentration response function for natural mortality.
Comparison was also made between cause-specific mortality derived from the natural mortality and disease burden of cardiopulmonary diseases and lung cancer calculated using cause-specific concentration response functions. Natural mortality only includes mortality compound and not morbidity which have likely been causing a loss of healthy life years before the premature death. When cause-specific health endpoints are used it is possible to also take into account morbidity compound. Though, morbidity compound only has a minor share of cause-specific mortality’s disease burden estimates (Table 7).
44
Table 7 Fine particles (PM2.5) disease burden in Finland in 2013 calculated as natural mortality, three cause-specific diseases derived from natural mortality and two cause-specific diseases using cause-specific concentration response functions (CRFs). The disease burden (DALY) is divided into mortality (YLL) and morbidity (YLD). The numbers of premature deaths associated with PM2.5 exposure are also shown.
Health endpoint
Burden of disease DALY
Reference
YLL
YLD
Deaths
32,100
0
1,900
Héroux et al., 2015
A: Mortality from all natural causes Natural mortality
32,100
B: Cause-specific mortality derived from the natural mortality Cardiovascular diseasesa
13,600
12,100
1,460
812
Héroux et al., 2015
Respiratory diseases and infectionsa
2,430
1,340
1,080
80
Héroux et al., 2015
Lung cancera
1,910
1,890
21
86
Héroux et al., 2015
C: Cause-specific mortality using cause-specific concentration response functions
a
Cardiopulmonary
19,900
16,800
3,050
1,110
Pope et al., 2002
Lung cancer
3,710
3,670
41
168
Pope et al., 2002
Calculated using the natural mortality concentration response function
Cause-specific mortality calculated using coefficient for natural mortality in comparison to all-cause natural mortality leads to 79% difference. When cause-specific coefficients are used for cardiopulmonary and lung cancer the disease burden is 36% higher than cause-specific mortality derived from the natural mortality.
45
5.5 UNCERTAINTY RANGE FOR THE DISEASE BURDEN OF PM 2.5 When relevant uncertainties are taken into account for the disease burden estimate of PM2.5 in Finland the minimum estimate is 13,600 DALY/a and the maximum estimate is 42,700 DALY/a (Figure 18). The reference point is at 20,800 DALY/a. In minimum and maximum calculations it is assumed that all the factors which bring the result lower or the factors which make it higher are present at the same time. The differences in comparison to the reference point are -35% smaller and 105% higher for minimum and maximum estimates, respectively.
Figure 18 Minimum and maximum estimates for the disease burden of PM 2.5. Each part indicates how much they would add or reduce the central estimate (reference point presented as orange diamond). The minimum estimate is calculated using the lower confidence intervals (CI) for exposure and concentration response functions’ coefficients while for the maximum estimate the upper values are used. In addition, all natural causes of death and the exposure distribution are taken into account in the maximum estimate.
46
6 DISCUSSION 6.1 RELATIVE RISK AND EXPOSURE ESTIMATES Epidemiological studies commonly report 95% confidence intervals for concentration response functions. Thus, it is easy to calculate the uncertainty due to the relative risk estimate, which is quite often done. The results showed that the uncertainty from concentration response functions can be one of the greatest uncertainties in disease burden analysis. Therefore, it is very important to do an uncertainty analysis for this uncertainty and report the results adequately. Even though the PM2.5 measurement stations network in Finland has not been planned out for exposure estimation, the uncertainties in the exposure estimation still turned out to be smaller than for the relative risk estimates. The confidence intervals for exposure estimates can give valuable information of the exposure and precision of the exposure measurements. Unfortunately, they are not commonly reported. With different air pollutants this uncertainty can vary greatly. It would be important to also calculate the results of minimum and maximum exposure estimates instead of only for the central one.
6.2 AVERAGE VS. EXPOSURE DISTRIBUTION Modelled exposure estimates are always simplifications of reality. Use of population weighted average exposure makes the calculations simpler than when the whole exposure distribution is taken into account. This simplification leads to a small error which was quantified to be around 4%. The error settled at 300 quantiles in the sensitivity analysis. When exposure distribution is divided into more quantiles it can be assumed that the exposure distribution illustrates the real exposure better in the population. In the ISTE-project, the population average mean exposure was used for the disease burden calculations, causing a small underestimation of the health effects. If the average exposure estimate presents the exposure distribution on the population well, then this can be used instead of the population exposure distribution for linear or log-linear concentration response functions. This is especially the case at low concentrations. It should still be kept in mind that when average exposure is used it might lead to a small underestimation of the disease burden. Exposure distribution should always be used in cases where concentration-response function differs non-linearly along the concentration range like the supra-linear integrated exposure response function.
47
6.3 SHAPE OF THE CONCENTRATION RESPONSE RELATIONSHIP The difference between the log-linear and linear functions on PAF estimates is very small on the low exposure range (0.036%). In the light of these results the log-linear and linear concentration response functions do not differ so much in regard of the shape that it would have a notable impact on the results at low concentrations. However, the use of relative risk or unit risk functions can lead to bigger differences in the results due to different calculation methods. Relative risk functions are used for so called topdown calculations where an exciting estimate of the disease burden or number of cases is used for evaluating the proportion of the disease caused by a chosen factor. This is why relative risk estimates cannot give higher estimates than the actual disease burden or number of cases. Unit risk approach can also be used as a bottom-up method where the calculations rely on disability weights and unit risk factors without using any real estimate for the already exciting burden. Therefore, using a unit risk approach can lead to an even higher disease burden for a certain factor than even exciting cases. Shapes of the log-linear and supra linear integrated exposure response (IER) functions differ varying along the exposure range (Figure 10). It depends on the concentration which one gives higher estimates. The IER functions have different shapes for every health endpoint from which the function for ischemic heart disease had the clearest supra-linear shape. For ischemic heart disease the IER function gives higher risk estimates on a range of 6 to 55µg/m3. The IER curve has a steep curve on that range and after that the slope starts to get flatter. WHO published their estimates for the disease burden of ambient air pollution in 2012 using PM2.5 as an indicator pollutant (WHO, 2016a). They used IHME’s integrated exposure response functions for the evaluation. Their estimate for the disease burden in Finland was 5,774 DALY for 2012 which is about 70% smaller than the ISTE-result. The IER functions are used in global burden of disease studies in order to get better estimates of the disease burden, also in areas with a high exposure. This can however lead to problems in areas with low concentrations. Use of the IER functions can underestimate the disease burden due to the use of a cutoff value in the IER model. At the moment it can be difficult to decide which shape to use in quantification of health impacts of PM2.5. Fortunately it is expected that the Air Quality Guidelines which are currently under update will provide recommendations for the shape of the concentrationresponse functions in addition to numerical concentration limits (WHO, 2017).
48
6.4 CUTOFF VALUE Disease burden estimates are sensitive to cutoff points but they are not made uncertain by it since cutoff values are pre-decided, known values (WHO, 2016b). However, uncertainties are related to threshold values as there is no level of PM2.5 exposure which has yet been proven safe for humans. The cutoff value is very crucial at small concentrations. Since PM2.5 concentrations are relatively low in Finland the use of a cutoff value can lead to a dramatic reduction in the burden of disease estimates. In this thesis, the results were calculated using three cutoff values 6; 2 and 0µg/m3 that had -23%, -16% and 31% impact on the disease burden results; respectively. In this thesis, it was assumed that the shape would stay the same even if the TMRED value would be different, making it possible to estimate some of the differences caused by the TMRED value. The main problem is that the IER functions have not been formed yet for concentrations lower than the TMRED value. Most probably, the shape of the function under the TMRED value will differ, so these results are inaccurate. IHME has however published Rcode which could possibly help to more accurately analyse the effect of the TMRED value. The TMRED value is defined so that it is possible to revise with the current research results and exposure estimates. Since there is new research available on the health effects of PM2.5 at low concentrations, it would be time to update the IER curve to a lower TMRED value. From the definition of the TMRED value for Finland, a better value could be 2 since we have evidence that there are areas with PM2.5 concentrations of 2µg/m3. Moreover, there are epidemiological studies which prove that risk reduces till 2µg/m3 (Crouse et al., 2014). Whether there is a threshold for the health effects of PM2.5 is still unknown and hard to prove. As Levy (2016) stated “Lack of proof does not equate to proof of lack” which is contradicted to use of cutoff to assume zero effect on low concentrations where there is little knowledge on the health effects. In addition, use of cutoff in the IER curve leads to a function which has zero effect up to TMRED and after that suddenly increases steeply and there are no empirical evidence supporting that. (Levy 2016.)
6.5 SELECTION OF HEALTH ENDPOINTS When all-cause natural mortality is related to PM2.5 exposure it is assumed that there is a primary cause leading to a disease, while exposure to PM2.5 exacerbates the symptoms causing a premature death. In that case, exposure to PM2.5 does not cause morbidity but only mortality since the disease is not a result of the exposure to PM2.5. When cause-specific mortality is calculated, it is assumed that exposure to PM2.5 causes a disease which leads to a premature death. These are two different approaches and they have a significant impact on the
49
disease burden. Here, the use of natural mortality would have resulted in 79% higher disease burden estimates than cause-specific mortality. All-cause mortality always leads to bigger estimates than using cause-specific health endpoints. For instance, Götchi et al. (2015) estimated that using all-cause mortality can lead to 2-3 times higher estimates than when using cause-specific health endpoints. However, cause-specific health endpoints have higher excess risk due to a stronger connection to the risk factor. Therefore, summing up cause-specific diseases calculated with their separate risk estimates leads to higher disease burden than summing up cause-specific diseases which have been derived from the natural mortality. The use of a natural mortality coefficient for causespecific diseases can lead to an error since the risks for cause-specific diseases might not increase in the same relation as the risk for natural mortality does. In this thesis, it was estimated that using cause-specific concentration response functions for cardiopulmonary diseases and lung cancer would have led to a 36% higher estimate than cause-specific diseases calculated using the all-cause natural mortality concentration response function. In the ISTE-project, the disease burden of PM2.5 was evaluated to be 20,800 DALY/a. As said before, the natural mortality in the ISTE-project was calculated using background disease burden only for cardiovascular diseases, respiratory diseases and infections, and lung cancer. In addition, the disease burden caused by cardiovascular and respiratory diseases was calculated using the concentration response function related to hospital admissions. This might lead to double counting to some degree since cardiovascular and respiratory diseases were also included into calculations for cause-specific mortality. In a case where all disease burden from cardiovascular and respiratory diseases hospital admission would be overlapping with cause-specific mortality, the disease burden would be overestimated by 2,870 DALY. The comparison for the health endpoint selection was done using the ISTE cause-specific morality (17,900 DALY/a) as a reference point. If the all-cause natural mortality and disease burden caused by cardiopulmonary diseases and lung cancer would have been compared to the whole disease burden of PM2.5 (20,800 DALY/a), the differences would have been 54% and 13%, respectively. In the earlier national study, PM2.5 disease burden has been estimated to be 14,000 DALY/a (Asikainen et al., 2013). The methodological changes e.g. dropping age weighting and discounting, as well as population ageing have increased the recent estimate. In comparison to the previous estimate, it seems proper to use cause-specific mortality instead of all-cause natural mortality. This is a rather big question to consider in environmental burden of disease studies since the choice of health endpoints and concentration-response functions can lead to highly different results. One uncertainty which was not analysed here is related to how effects from short- and longterm exposure should be calculated. Can short-term exposure indicate long-term mortality or can it only be associated with short-term acute mortality? Hospital admissions are also used to
50
derive concentration response functions but can they be used in health impact assessments? Clear guidelines and recommendations on which concentration response functions can be used, are needed.
51
6.6 OVERALL DISCUSSION The true values of the environmental disease burden are not known. Thus in disease burden studies it is not possible to quantify the model errors by comparing the estimated values to the true values. Therefore, uncertainty analysis is needed to get a picture of the uncertainties’ magnitudes. (Hänninen et al., 2005.) Many times, only parametric uncertainty in concentration-response functions is taken into account when disease burden estimates are analysed for their uncertainties. Parametric uncertainty in concentration-response function is a significant source of uncertainty and easy to quantify. However, in cases where only parametric uncertainty is evaluated, other significant sources of uncertainty will not be covered. Quantification of parametric uncertainty in concentration-response functions could, at least, work as a minimum level of uncertainty analysis, as long as it is kept in mind that it will not cover all the uncertainties present in the estimates. This thesis includes quantitative estimates for selected uncertainties present in the disease burden estimates of fine particles. Figure 19 shows the impact of the quantified uncertainties on the disease burden estimates of PM2.5. In order to make the uncertainty analysis complete, the uncertainties which were not possible to quantify should be quantitatively analysed or given a quantitative estimate according to expert judgement. However, it is not necessary or even possible to analyse all individual uncertainty sources separately. For an overall picture of the uncertainties in the assessment the individual uncertainties should be combined, including as many uncertainties as appropriate and possible within the time available. Uncertainty analyses should be in line with the resources, scope and time schedule of the analysis. (EFSA Scientific Committee, 2016.)
52
Figure 19 Impact of the quantified uncertainties on the disease burden estimates of PM 2.5 from highest to lowest.
There are rather few studies and a relatively small amount of knowledge on the quantities of the uncertainties in the burden of disease studies. In Finland though, a previous study has been conducted related to this topic. Tainio et al. (2010) researched the uncertainties in the health risks caused by PM2.5 from different emission source types in Finland. The uncertainties investigated in their study were somewhat different since their focus laid on different emission types, which were not addressed in this thesis. Consistent with these two studies, the uncertainties related to relative risk estimates were considered very high. Tainio et al. (2010) found that uncertainties in the relative risk estimates were much higher than other uncertainties included in their study, while uncertainties in the relative risk estimates were quantified as the second highest in this study as well. Medina et al. (2013) present three important points regarding uncertainties in health impact assessments. They state that it is important to acknowledge uncertainties and provide a broad picture of the possible impacts. In addition, it is important to clarify to stakeholders which uncertainties can possibly be reduced and how. Here, for instance, we could conclude that more precise risk estimates for epidemiological studies would decrease the uncertainty of the estimates. Last, they point out that regardless of the possible uncertainties, health impact assessments are important, as without quantitative estimates there can be an impression of no risk.
53
7 CONCLUSIONS Estimation of environmental burden of disease is a commonly used approach to guide policy makers. Disease burden studies contain uncertainties, which can be difficult to quantify, and which are not always adequately evaluated. Uncertainty analysis does not remove the uncertainties, but recognizing uncertainties is essential in order to take them into account and to be able to reduce them for more accurate and precise estimates. This work aimed to quantify the selected uncertainty components and their relevance for the overall disease burden. The quantified uncertainties were related to the recent national disease burden estimates calculated for PM2.5 exposure which is, according to the previous work, the most important environmental risk factor in Finland. Parametric uncertainties were distributed quite symmetrically around the central estimate. The 95% confidence intervals of exposure and relative risk estimates lead to roughly ± 10% and 25 to 23% differences, respectively. The uncertainties related to the epidemiological studies seem to be bigger than the uncertainties related to the exposure estimates. Model uncertainties were estimated using different approaches and comparing the results to the reference point i.e. national estimate of PM2.5 disease burden. Model uncertainties were more diverse and not as equally spread as parametric uncertainties. They would have all increased the reference point except for the supra-linear integrated exposure response function, which would have decreased the estimate. Major differences arise from the use of a supra-linear shaped concentration response function (-23%), and the use of cause-specific mortality instead of natural mortality (79%). Minor differences are related to the use of exposure distribution instead of average exposure (4%), and linear exposure response function instead of log-linear (0.04%). The ranges of parametric and model uncertainties were -25 to 23% and -23 to 79%, respectively. The biggest individual uncertainties in the disease burden estimate of PM2.5 are related to the selection of health endpoints, parametric uncertainties in concentration-response functions and the shape of the concentration-response function. The individual uncertainties did not change the order of magnitude of the inspected estimate. When quantified uncertainties were taken into account at the same time the uncertainty was slightly more than one order of magnitude. All the uncertainties listed here might not be relevant to calculate for every uncertainty analysis of disease burden studies. However, the choices behind health endpoint selection, the use of a cutoff point and the types of the concentration response functions are relevant to document. This documentation can help repeating the calculations for some other areas or time frames as well as explain the differences between separate disease burden studies. Regardless of their uncertainties, disease burden studies are important for characterizing the risks of different factors and assisting policy makers in evaluating and prioritizing policy actions. They give a picture of the possible health risks related to a risk factor and help identify the factors which pose a real threat to human health.
54
REFERENCES Adami, H. O., Berry, C. L., Breckenridge, C. B., Smith, L. L., Swenberg, J. A., Trichopoulos, D., Weiss, N. S., & Pastoor, T. P. (2011). Toxicology and epidemiology: improving the science with a framework for combining toxicological and epidemiological evidence to establish causal inference. Toxicological Sciences, 122(2), 223-234. Anenberg, S. C., Belova, A., Brandt, J., Fann, N., Greco, S., Guttikunda, S., Héroux, M-E., Hurley, F., Krzyzanowaki, M., Medina, S., Miller, B., Pandey, K., Roos, J., & Van Dingenen, R. (2015). Survey of Ambient Air Pollution Health Risk Assessment Tools. Risk Analysis. DOI: 10.1111/risa.12540 Asikainen A., Hänninen O. & Pekkanen J. (2013). Ympäristöaltisteisiin liittyvä tautitaakka Suomessa. Ympäristö ja Terveys 5 (44), 68–74. Bailar, J. C., & Bailer, A. (1999). Risk Assessment‐the Mother of All Uncertainties: Disciplinary Perspectives on Uncertainty in Risk Assessment. Annals of the New York Academy of Sciences, 895(1), 273-285. Bouwknegt, M., & Havelaar, AH. (2015). Uncertainty assessment using the NUSAP approach: a case study on the EFoNAO tool. EFSA supporting publication 2015:EN-663, 20 pp. https://www.efsa.europa.eu/en/supporting/pub/663e (Accessed 29 December 2016) Brauer, M., Freedman, G., Frostad, J., van Donkelaar, A., Martin, RV., Dentener, F., Van Dingenen, R., Estep, K., Amini, H., Apte, JS., Balakrishnan, K., Barregard, L., Broday, DM., Feigin, V., Ghosh, S., Hopke, PK., Knibbs, LD., Kokubo, Y., Liu, Y., Ma, S., Morawska, L., Sangrador, JLT., Shaddick, G., Anderson, HR., Vos. T., Forouzanfar, MH., Burnett, RT., & Cohen, A. (2015). Ambient air pollution exposure estimation for the Global Burden of Disease 2013. Environmental Science & Technology. doi: 10.1021/acs.est.5b03709. Burnett, R. T., Pope III, C. A., Ezzati, M., Olives, C., Lim, S. S., Mehta, S., ... & Anderson, H. R. (2014). An integrated risk function for estimating the global burden of disease attributable to ambient fine particulate matter exposure. Environmental Health Perspectives, 122(4), 397. Committee on Risk Assessment of Hazardous Air Pollutants. Science and Judgement in Risk Assessment. National Academy Press, Waschington, D.C. 1994. http://www.nap.edu/catalog/2125/science-and-judgment-in-risk-assessment (Accessed 2nd May 2016). Crouse, D. L., Peters, P. A., van Donkelaar, A., Goldberg, M. S., Villeneuve, P. J., Brion, O., Khan, S., Atari, D., Jerrett, M., Pope, A. III, Brauer, M., Brook, J., Martin, R., Stieb, D., & Burnett, R. (2012). Risk of nonaccidental and cardiovascular mortality in relation to longterm exposure to low concentrations of fine particulate matter: a Canadian national-level cohort study. Environmental health perspectives, 120(5), 708.
55
Crouse DL, Peters PA, Hystad P, Brook JR, van Donkelaar A, Martin RV, Villeneuve PJ, Jerrett M, Goldberg MS, Pope CA III, Brauer M, Brook RD, Robichaud A, Menard R, Burnett RT. (2015). Ambient PM2.5, O3, and NO2 exposures and associations with mortality over 16 years of follow-up in the Canadian Census Health and Environment Cohort (CanCHEC). Environ Health Perspect 123:1180–1186; http://dx.doi.org/10.1289/ehp.1409276 EFSA (European Food Safety Authority) Scientific Comittee, 2016. Guidance on Uncertainty in EFSA Scientific Assessment - Draft version for internal testing. https://www.efsa.europa.eu/en/topics/topic/uncertainty (Assessed 21st of October 2016). Evans, J. S. (2016). Characterizing Uncertainty in Estimates of Mortality Risk from Exposure to Ambient Fine Particulate Matter. Risk Analysis, 36(9), 1748-1750. Forouzanfar, M.H., Afshin, A., Alexander, L.T., Anderson, H.R., Bhutta, Z.A… & Murray, C.J.I. (2016). Global, regional, and national comparative risk assessment of 79 behavioral, environmental and occupational, and metabolic risks or clusters of risks, 1990–2015: a systematic analysis for the Global Burden of Disease Study 2015. Lancet. GHDx: Global burden of disease study 2010 (GBD 2010) - Ambient air pollution risk model 1990-2010. http://ghdx.healthdata.org/record/global-burden-disease-study-2010-gbd-2010ambient-air-pollution-risk-model-1990-2010 (Accessed 9 December 2016). Gibson, J. M., Brammer, A., Davidson, C., Folley, T., Launay, F., & Thomsen, J. (2013). Environmental burden of disease assessment. Springer Science & Business Media. Götschi, T., Tainio, M., Maizlish, N., Schwanen, T., Goodman, A., & Woodcock, J. (2015). Contrasts in active transport behaviour across four countries: How do they translate into public health benefits?. Preventive medicine, 74, 42-48. Hänninen, O.O., Lebret, E., Tuomisto, J.T., & Jantunen, M.J. (2005). Characterization of Model Error in a Simulation of PM2.5 Exposure Distributions of the Working Age Population in Helsinki, Finland. JAWMA. 55: 446-457. Hänninen, O., & Knol, A. (2011). European Perspectives on Environmental Burden of Disease Estimates for Nine Stressors in Six European Countries. National Institute for Health and Welfare (THL), Finland, Helsinki, Finland, No.1/2011. Hänninen, O., Korhonen, A., Lehtomäki, H., Asikainen, A., & Rumrich, I. (2016). Ilmansaasteiden terveysvaikutukset. Ympäristöministeriön raportteja 16/2016. http://urn.fi/URN:ISBN:978-952-11-4604-6 (Accessed 1 August 2016) Héroux, M. E., Anderson, H. R., Atkinson, R.., Brunekreef, B., Cohen, A., Forastiere, F., Hurley, F., Katsouyanni, K.., Krewski, D., Krzyzanowski, M., Künzli, N., Mills, I., Querol, X., Ostro, B., & Walton, H. (2015). Quantifying the health impacts of ambient air pollutants: recommendations of a WHO/Europe project. International Journal of Public Health, 60(5), 619-627. doi:10.1007/s00038-015-0690-y
56
Hill, A. B. (1965). The environment and disease: association or causation?. Proceedings of the Royal Society of Medicine, 58(5), 295. Hurley F, Hunt A, Cowie H, Holland M, Miller B, Pye S, Watkiss P, 2005. Methodology Paper (Volume 2) for Service Contract for carrying out cost-benefit analysis of air quality related issues, in particular in the clean air for Europe (CAFE) programme. AEA Technology Environment, 133 pp. IARC (2013). Outdoor air pollution a leading environmental cause of cancer deaths. International Agency for Research on Cancer (October 17 2013). https://www.iarc.fr/en/media-centre/iarcnews/pdf/pr221_E.pdf (Accessed 29 December 2016). IHME (2013). The Global Burden of Disease: Generating Evidence, Guiding Policy. Institute for Health Metrics and Evaluation, University of Washington. http://www.healthdata.org/sites/default/files/files/policy_report/2013/GBD_GeneratingEviden ce/IHME_GBD_GeneratingEvidence_FullReport.pdf (Assessed 12 December 2016). Klaassen, C. D., & Amdur, M. O. (Eds.). (1996). Casarett and Doull's toxicology: the basic science of poisons (Vol. 5). New York: McGraw-Hill. Korhonen, A., Asikainen, A., Rumrich, I., & Hänninen, O. (2015). Ilmansaasteille altistuminen Suomessa vuonna 2013. Ilmanlaadun mittausverkostoon pohjautuen. ISTEraportti. National Institute for Health and Welfare (THL). http://urn.fi/URN:NBN:fife201601122445 (Accessed 20 May 2016) Knol, A. B., Petersen, A. C., Van der Sluijs, J. P., & Lebret, E. (2009). Dealing with uncertainties in environmental burden of disease assessment. Environ Health, 8(21), 10-1186. Knol, A. B. (2010). Health and the Environment: assessing the impacts, addressing the uncertainties. Utrecht University, the Netherlands. Krewski, D., Jerrett, M., Burnett, R., Ma, R., Hughes, E., Shi, Y., Turner, M., Pope, C.A 3 rd, Thurston, G., Eugenia, E., Thun, M., Beckerman, B., DeLuca, P., Finkelstein, N., Ito, K., Moore, D K., Newbold, K., Ramsay, T., Ross, Z., Shin, H., & Tempalski, B. (2009). Extended follow-up and spatial analysis of the American Cancer Society study linking particulate air pollution and mortality. Research report Health Effects Institute, 140, 5-114. Ma, R., Hughes, E., Shi, Y., Turner, M. C., Pope III, C. A., Thurston, G., ... & Thun, M. J. (2009). Extended follow-up and spatial analysis of the American Cancer Society study linking particulate air pollution and mortality (No. 140). Boston, MA: Health Effects Institute. MacMahon, B., & Pugh, T. F. (1970). Epidemiology: principles and methods. Epidemiology: principles and methods.
57
Medina, S., Ballester, F., Chanel, O., Declercq, C., & Pascal, M. (2013). Quantifying the health impacts of outdoor air pollution: useful estimations for public health action. Journal of epidemiology and community health, jech-2011. Moffett, D., Mumtaz, M., Sullivan, JR. D. & Fowler, B. General Considerations of DoseEffect and Dose-Response Relationships. In a book Nordberg, G. F., Fowler, B. A., & Nordberg, M. (Eds.). (2014). Handbook on the Toxicology of Metals. Academic Press. Nord, E. (2015). Uncertainties about disability weights for the Global Burden of Disease study. The Lancet Global Health, 3(11), e661-e662. Lehtomäki, H., Asikainen, A., Rumrich, I. & Hänninen, O. (2015). Ilmansaasteiden tautitaakka Suomessa. ISTE-raportti. National Institute for Health and Welfare (THL). (In Finnish) http://urn.fi/URN:NBN:fi-fe201601122444 Lehtomäki, H., Korhonen, A., Asikainen, A. & Hänninen, O. (2016). Ulkoilman saasteiden aiheuttamat terveyshaitat Suomessa. Ympäristö ja terveys -lehti 8 (47), 22–27. Levy, J. I. (2016). Fine Particulate Matter, Risk Assessment, and Risk Management. Risk Analysis, 36(9), 1745-1747. Li, F., Huang, J. H., Zeng, G. M., Yuan, X. Z., Liang, J., & Wang, X. Y. (2012). Multimedia health risk assessment: A case study of scenario-uncertainty. Journal of Central South University, 19, 2901-2909. Lim, S. S., Vos, T., Flaxman, A. D., Danaei, G., Shibuya, K., Adair-Rohani, H., ... & Aryee, M. (2012). A comparative risk assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study 2010. The Lancet, 380(9859), 2224-2260. Lucas, R. M., & McMichael, A. J. (2005). Association or causation: evaluating links between" environment and disease". Bulletin of the World Health Organization, 83(10), 792795. Parascandola, M., & Weed, D. L. (2001). Causation in epidemiology. Journal of Epidemiology and Community Health, 55(12), 905-912. Pearce, N., Checkoway, H., & Kriebel, D. (2007). Bias in occupational epidemiology studies. Occupational and environmental medicine, 64(8), 562-568. doi: 10.1136/oem.2006.026690 Pope III, C. A., Burnett, R. T., Thun, M. J., Calle, E. E., Krewski, D., Ito, K., & Thurston, G. D. (2002). Lung cancer, cardiopulmonary mortality, and long-term exposure to fine particulate air pollution. Jama, 287(9), 1132-1141. Pope III, C. A., & Dockery, D. W. (2006). Health effects of fine particulate air pollution: lines that connect. Journal of the air & waste management association, 56(6), 709-742.
58
Pope, C. A., Burnett, R. T., Krewski, D., Jerrett, M., Shi, Y., Calle, E. E., & Thun, M. J. (2009). Cardiovascular mortality and exposure to airborne fine particulate matter and cigarette smoke shape of the exposure-response relationship. Circulation, 120(11), 941-948. Pope III, C. A., Cropper, M., Coggins, J., & Cohen, A. (2015). Health benefits of air pollution abatement policy: Role of the shape of the concentration–response function. Journal of the Air & Waste Management Association, 65(5), 516-522. Prüss-Üstün, A., Mathers C., Corvalán, C. & Woodwards, A. Introduction and methods: assessing the environmental burden of disease at national and local levels. Geneva, World Health Organization, 2003. (WHO Environmental Burden of Disease Series, No. 1). Rothman, K. J. (2012). Epidemiology: an introduction. Oxford University Press. 2nd edition. Royal College of Physicians. Every breath we take: the lifelong impact of air pollution. Report of a working party. London: RCP, 2016. https://www.rcplondon.ac.uk/projects/outputs/every-breath-we-take-lifelong-impact-airpollution (accessed 17 November 2016). Salomon, J. A., Haagsma, J. A., Davis, A., de Noordhout, C. M., Polinder, S., Havelaar, A. H., Cassini, A., Devleesschauwer, B., Kretzchmar, M., Speybroeck, N., Murray, C. J. & Vos, T. (2015). Disability weights for the Global Burden of Disease 2013 study. The Lancet Global Health, 3(11), e712-e723. Settimo, G. (2015). Support Additional Air Pollutants in Calculations for Quantitative Health Risk Assessment. Contract No. 2015/588109. World Health Organization. Siponen, T., Yli-Tuomi, T., Aurela, M., Dufva, H., Hillamo, R., Hirvonen, M. R., Huttunen, K., Pekkanen, J., Pennanen, A., Salonen, I., Tiittanen, P., Salonen R. O. & Lanki, T. (2015). Source-specific fine particulate air pollution and systematic inflammation in ischaemic heart disease patients. Occupational and environmental medicine, 72(4), 277-283. Schwarze, P. E., Øvrevik, J., Låg, M., Refsnes, M., Nafstad, P., Hetland, R. B., & Dybing, E. (2006). Particulate matter properties and health effects: consistency of epidemiological and toxicological studies. Human & experimental toxicology, 25(10), 559-579. Tainio, M., Tuomisto, J. T., Pekkanen, J., Karvosenoja, N., Kupiainen, K., Porvari, P., ... & Kukkonen, J. (2010). Uncertainty in health risks due to anthropogenic primary fine particulate matter from different source types in Finland. Atmospheric Environment, 44(17), 2125-2132. US EPA. Guidelines for exposure assessment [R]. Washington D C: Office of Research and Development, Office of Health and Environmental Assessment, 1992. Vock, D. M., Atchison, E. A., Legler, J. M., McClure, D. R., Carlyle, J. C., Jeavons, E. N., & Burton, A. H. (2011). Accounting for model uncertainty in estimating global burden of disease. Bulletin of the World Health Organization, 89(2), 112-120.
59
Wang, Y., Kloog, I., Coull, B. A., Kosheleva, A., Zanobetti, A., & Schwartz, J. D. (2016). Estimating Causal Effects of Long-Term PM 2.5 Exposure on Mortality in New Jersey. Environ Health Perspect. DOI: 10.1289/ehp.1409671 WHO (2006). Air quality guidelines: global update 2005: particulate matter, ozone, nitrogen dioxide, and sulfur dioxide. World Health Organization Regional Office for Europe, Copenhagen. WHO (2008). Uncertainty and data quality in exposure assessment. World Health Organization. WHO (2013). Review of evidence on health aspects of air pollution – REVIHAAP Project citation. World Health Organization. WHO 2016a. Ambient air pollution: A global assessment of exposure and burden of disease. World Health Organization. http://apps.who.int/iris/bitstream/10665/250141/1/9789241511353-eng.pdf?ua=1 (accessed 9 November 2016). WHO 2016b. Health risk assessment of air pollution - general principles. World Health Organization Regional Office for Europe, Copenhagen. x + 30 pages. ISBN 978 92 890 51316. http://www.euro.who.int/en/health-topics/environment-and-health/airquality/publications/2016/health-risk-assessment-of-air-pollution.-general-principles-2016 (accessed 4 February 2016). WHO: Global health estimates for 2000-2012. http://www.who.int/healthinfo/global_burden_disease/estimates/en/ (Assessed 20 December 2016). WHO: Global Burden of Disease (GBD) http://www.who.int/healthinfo/global_burden_disease/gbd/en/ (assessed 12 February 2016). WHO 2017. Evaluation of WHO air quality guidelines: past, present and future. World Health Organization Regional Office for Europe, Copenhagen. vi + 32 pages. ISBN 9789289052306 http://www.euro.who.int/__data/assets/pdf_file/0019/331660/Evolution-air-quality.pdf?ua=1 (accessed 2 March 2017).
APPENDICES
APPENDIX 1: DISEASE BURDEN CALCULATION METHODS Disease burden calculation methods can be divided into methods which can be applied when relative risk functions are available (Figure A1) and to methods which are utilized when unit risk functions are used (Figure A2). When background disease burden data are available methods (A) are used. Methods (B) can be applied when incidence or prevalence data are used.
Figure A1 Disease burden calculation methods 1A and 1B for relative risk functions (based on Hänninen & Knol, 2011)
1 (3)
APPENDICES
Figure A2 Disease burden calculation methods 2A and 2B for unit risk functions (based on Hänninen & Knol, 2011)
2 (3)
APPENDICES
3 (3)
APPENDIX 2: THE IDENTIFIED CONCENTRATION RESPONSE FUNCTIONS Table A1 The concentration response functions identified for fine particles (PM2.5) (based on Lehtomäki et al., 2015). (CVD = cardiovascular diseases, COPD = chronic obstructive pulmonary disease, IHD = ischemic heart disease, HA = hospital admissions) Unit
CR functions
Health endpoint
µg m^-3
Central
Lower 95% CI
Upper 95% CI
Mortality (natural)
10
1.062
1.04
1.083
Héroux et al.. 2015
CVDs (includes stroke). HA
10
1.0091
1.0017
1.0166
Héroux et al.. 2015
Respiratory diseases (HA)
10
1.019
0.9982
1.0402
Héroux et al.. 2015
RADs (Restricted Activity Days)
10
1.047
1.042
1.053
Héroux et al.. 2015
Work days lost
10
1.046
1.039
1.053
Héroux et al.. 2015
Mortality (all)
10
1.0123
1.0045
1.0201
Héroux et al.. 2015
Mortality (all)
10
1.06
1.04
1.08
Hoek et al.. 2013
CVDs and diabetes
10
1.12
1.08
1.15
Pope et al.. 2004
Cardiopulmonary (mortality)
10
1.09
1.03
1.16
Pope et al.. 2002
Cardiopulmonary (mortality)
1
1.0077
1.002
1.0132
Pope et al.. 2002
Chronic bronchitis (incidence)
1
5.3E-05
1.7E-06
0.00011
Hurley et al.. 2005
Lung cancer
10
1.0723
1.0148
1.1331
Yang et al.. 2015
Lung cancer
1
1.012
1.004
1.02
Pope et al.. 2002
17.5
1.03
1.01
1.05
Chang et al.. 2015
Mortality (natural)
5
1.07
1.02
1.13
Beelen et al.. 2014
Lung cancer
10
1.09
1.04
1.14
Hamra et al.. 2014
Stroke
10
1.060
1.000
1.130
Shin et al.. 2014
Headache
Reference
