Mixture Effects of Environmental Contaminants

Digital Comprehensive Summaries of Uppsala Dissertations from the Faculty of Medicine 1059

ERIK LAMPA

ACTA UNIVERSITATIS UPSALIENSIS ISSN 1651-6206 ISBN 978-91-554-9122-2 UPPSALA urn:nbn:se:uu:diva-237690 2015 Dissertation presented at Uppsala University to be publicly examined in Frödingsalen, Ulleråkersvägen 40A, Uppsala, Friday, 6 February 2015 at 09:00 for the degree of Doctor of Philosophy (Faculty of Medicine). The examination will be conducted in Swedish. Faculty examiner: Professor Jonas Björk (Avdelningen för Arbets- och Miljömedicin, Lunds universitet).

Abstract Lampa, E. 2015. Mixture Effects of Environmental Contaminants. Digital Comprehensive Summaries of Uppsala Dissertations from the Faculty of Medicine 1059. 85 pp. Uppsala: Acta Universitatis Upsaliensis. ISBN 978-91-554-9122-2.

Chemical exposure in humans rarely consists of a single chemical. The everyday exposure is characterized by thousands of chemicals mainly present at low levels. Despite that fact, risk assessment of chemicals is carried out on a chemical-by-chemical basis although there is a consensus that this view is too simplistic. This thesis aims to validate a statistical method to study the impact of mixtures of contaminants and to use that method to investigate the associations between circulating levels of a large number of environmental contaminants and atherosclerosis and the metabolic syndrome in an elderly population. Contaminants measured in the circulation represented various classes, such as persistent organic pollutants, plastic-associated chemicals and metals. There was little co-variation among the contaminants and only two clusters of PCBs could be discerned. Gradient boosted CARTs were used to assess additive and multiplicative associations between atherosclerosis, as measured by the intima-media thickness (IMT) and the echogenicity of the intima-media complex (IM-GSM), and prevalent metabolic syndrome. Systolic blood pressure was the most important predictor of IMT while the influence of the contaminants was marginal. Three phthalate metabolites; MMP, MEHP and MIBP were strongly related to IM-GSM. A synergistic interaction was found for MMP and MIBP, and a small antagonistic interaction was found for MIBP and MEHP. Associations between the contaminants and prevalent metabolic syndrome were modest, but three pesticides; p,p’-DDE, hexachlorbenzene and trans-nonachlor along with PCBs 118 and 209 and mercury were the strongest predictors of prevalent metabolic syndrome. This thesis concludes that many contaminants need to be measured to get a clear picture of the exposure. Boosted CARTs are useful for uncovering interactions. Multiplicative and/or additive effects of certain contaminant mixtures were found for atherosclerosis or the metabolic syndrome.

Keywords: Mixtures, Environmental Contaminants, Atherosclerosis, Metabolic Syndrome, Boosting, Epidemiology

Erik Lampa, Department of Medical Sciences, Occupational and Environmental Medicine, Akademiska sjukhuset, Uppsala University, SE-75185 Uppsala, Sweden.

ISSN 1651-6206 ISBN 978-91-554-9122-2 urn:nbn:se:uu:diva-237690 (http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-237690) Till min familj

List of papers

This thesis is based on the following papers, which are referred to in the text by their Roman numerals.

I E Lampa, L Lind, A Bornefalk-Hermansson, S Salihovic, B van Bavel, and P M Lind. An investigation of the co-variation in circulating levels of a large number of environmental contaminants. Journal of Exposure Science and Environmental Epidemiology, 22(5):476–482, 2012

II E Lampa, L Lind, P M Lind, and A Bornefalk-Hermansson. The identiﬁcation of complex interactions in epidemiology and toxicology: a simulation study of boosted regression trees. Environ Health, 13(57), 2014

III E Lampa, A Bornefalk-Hermansson, P M Lind, and L Lind. Atherosclerosis in Humans and the Association to Environmental Contaminant Mixtures. Manuscript, 2014

IV E Lampa, A Bornefalk-Hermansson, L Lind, and P M Lind. Mixture Effects of Multiple Environmental Contaminants on the Metabolic Syndrome in a Human Population-based Sample. Manuscript, 2014

Reprints were made with permission from the publishers.

Contents

1 Introduction ...... 15 1.1 Environmental Contaminants ...... 16 1.1.1 Persistent Organic Pollutants ...... 16 1.1.2 Plastic-Associated Chemicals ...... 17 1.1.3 Metals ...... 18 1.2 Chemical mixtures ...... 18 1.2.1 Concentration Addition and Independent Action ...... 19 1.2.2 Experimental Studies on Mixtures ...... 22 1.2.3 Epidemiological Studies on Mixtures ...... 22 1.3 An Unsolved Issue ...... 23 1.4 Atherosclerosis ...... 24 1.4.1 Environmental Contaminants and Atherosclerosis ...... 24 1.5 The Metabolic Syndrome ...... 25 1.5.1 Environmental Contaminants and the Metabolic Syndrome ... 26

2 Aim ...... 28

3 Material and Methods ...... 29 3.1 The PIVUS Study ...... 29 3.1.1 Measured Contaminants ...... 29 3.1.2 Ultrasound Evaluation of the Carotid Artery ...... 30 3.2 Statistical Methods ...... 31 3.2.1 Identifying Clusters of Contaminants ...... 31 3.2.2 Identifying Suitable Markers ...... 32 3.2.3 Classiﬁcation and Regression Trees ...... 32 3.2.4 Stochastic Gradient Boosting ...... 34 3.2.5 Variable Importance and Interpretation ...... 35 3.2.6 Assessment of Interaction Effects ...... 35 3.2.7 Simulating Contaminant Data ...... 37

4 Results and Discussion ...... 39 4.1 Paper I ...... 39 4.1.1 Discussion, paper I ...... 39 4.2 Paper II ...... 41 4.2.1 Assessment of Interaction Effects in the Simulated Data ...... 43 4.2.2 Visualizing the Four-way Interaction ...... 46 4.2.3 Discussion, paper II ...... 48 4.3 Paper III ...... 50 4.3.1 IMT ...... 50 4.3.2 IM-GSM ...... 53 4.3.3 Discussion, paper III ...... 59 4.4 Paper IV ...... 59 4.4.1 Discussion, paper IV ...... 64

5 Summary ...... 65

6 General Discussion ...... 66

7 Svensk sammanfattning: Blandningseffekter av Miljögifter ...... 72 7.1 Studie I ...... 73 7.2 Studie II ...... 73 7.3 Studie III ...... 74 7.4 Studie IV ...... 74 7.5 Sammanfattning av Studierna ...... 74

8 Acknowledgement ...... 75

References ...... 77 List of Tables

Table 3.1: Mean values and standard deviations (SD) used in the simulation ...... 38

Table 4.1: Bootstrap validated parameters ...... 43 List of Figures

1.1 A mixture experiment ...... 20 1.2 Contaminants and cardiovascular disease ...... 25 3.1 Ultrasound image of the common carotid artery ...... 30 3.2 Example CART ...... 33 4.1 Dendrograms from PIVUS and NHANES ...... 40 4.2 Loading plot for the medium/high chlorinated PCB cluster . . 41 4.3 Loading plot for the low/medium chlorinated PCB cluster . . 42 4.4 Variable importance for SNRs 2, 1, 0.5 and 0.1 ...... 43 4.5 Interactions for SNR = 2 ...... 44 4.6 Interactions for SNR = 1 ...... 45 4.7 Interactions for SNR = 0.5 ...... 45 4.8 Interactions for SNR = 0.1 ...... 46 4.9 Two-way interactions with sex ...... 47 4.10 Partial dependence on BPA ...... 47 4.11 Partial dependence on OCDD ...... 48 4.12 Visualization of the four-way interaction ...... 49 4.13 Variable importance, IMT ...... 50 4.14 Partial dependencies, IMT ...... 51 4.15 Predicted effect, IMT ...... 52 4.16 Variable importance, IM-GSM ...... 53 4.17 Interaction strengths, IM-GSM ...... 54 4.18 MMP–MIBP interaction, IM-GSM ...... 55 4.19 MEHP–MIBP interaction, IM-GSM ...... 56 4.20 Partial dependencies, IM-GSM ...... 57 4.21 Predicted effects, IM-GSM ...... 58 4.22 Relative importance, MetS ...... 60 4.23 Interaction strengths, MetS ...... 61 4.24 Partial dependencies, MetS ...... 62 4.25 Estimated OR, MetS ...... 63 6.1 Parameter tuning, simulated mixture data ...... 68 6.2 Interaction assessment, simulated mixture data ...... 69 6.3 Joint effect, simulated mixture data ...... 70 Glossary

R2 Coefﬁcient of determination. 31 EC1 Effective concentration yielding 1% of the maximum effect. 19 EC50 Half maximal effective concentration. 19 NHANES National Health and Nutrition Examination Survey. 22 PIVUS Prospective Investigation of the Vasculature in Uppsala Seniors. 24

ADHD Attention deﬁcit hyperactive disorder. 22 AhR Aryl hydrocarbon receptor. 24 AIDS Acquired Immune Deﬁciency Syndrome. 18 Al Aluminum. 36 AMS Artery measurement software. 29 ATPIII Adult treatment panel III. 25 AUC Area under the receiver operating characteristic curve. 58

BDE47 2,2’,4,4’-tetra-bromodiphenyl ether. 28 BMI Body mass index. 49 BPA Bisphenol A. 17

CA Concentration addition. 19 CART Classiﬁcation and regression tree. 27 Cd Cadmium. 36 CI Conﬁdence interval. 61 Co Cobalt. 36 Cr Chromium. 36 Cu Copper. 36

DDT dichlorodiphenyltrichloroethane. 16

EWAS Environment-wide association study. 26

HCB Hexachlorobenzene. 28 HDL High-density lipoprotein. 25 Hg Mercury. 36 HOMA Homeostatic model assessment. 26 HRGC/HRMS High resolution gas chromatography / high resolution mass spectroscopy. 28

IA Independent action. 19 ICP-SFMS Inductively coupled plasma-sector ﬁeld mass spectrometry. 29 IM-GSM Intima-media grey scale median. 29 IMT Intima-media thickness. 29

LDL Low-density lipoprotein. 23

MBZP monobenzyl phthalate. 29 MCHP monocyclohexyl phthalate. 29 MEHHP mono-2-ethyl-5-hydroxyhexyl phthalate. 28 MEHP Mono-2-ethylhexyl phthalate. 25 MEOHP mono 2-ethyl-5-oxohexyl phthalate. 29 MEP Monoethyl phthalate. 29 MetS Metabolic syndrome. 25 MIBP Mono-isobutyl phthalate. 24 MINP monoisononyl phthalate. 29 MMP Mono-methyl phthalate. 24 Mn Manganese. 36 Mo Molybdenum. 36 MOP mono-n-octyl phthalate. 29

NCEP National Cholesterol Education Program. 25 Ni Nickel. 36 NOAEL No observed adverse effect limit. 21 NOEL No observed effect limit. 21

OC Organochlorine. 17 OCDD Octachlorodibenzo-p-dioxin. 28, 36 OR Odds ratio. 61 p,p’-DDE 1,1-Bis-(4-chlorophenyl)-2,2-dichloroethene. 26 PAC Plastic-associated Chemical. 15 Pb Lead. 36 PCB Polychlorinated Biphenyl. 16 PCB 118 2,3’,4,4’,5-Pentachlorobiphenyl. 36 PCB 126 3,3’,4,4’,5-Pentachlorobiphenyl. 36 PCB 153 2,2’,4,4’,5,5’-Hexachlorobiphenyl. 36 PCB 169 3,3’,4,4’,5,5’-Hexachlorobiphenyl. 36 PCB 170 2,2’,3,3’,4,4’,5-Heptachlorobiphenyl. 36 PCB 209 Decachlorobiphenyl. 36 PCDD Polychlorinated-p-dibenzodioxin. 16 PCDF Polychlorinated-p-dibenzofuran. 16 POP Persistent Organic Pollutant. 15 PPAR Peroxisome Proliferator-Activated Receptor. 17

RMSE Root-mean-squared error. 40 SD Standard deviation. 36 SE Standard error. 40 SNR Signal to noise ratio. 37

TEF Toxic equivalence factor. 21 TEQ Toxic equivalent. 21 TNC trans-nonachlor. 28

WHO World Health Organization. 25

Zn Zinc. 36

1. Introduction

Contamination of the environment has always followed civilization. Deter- mination of copper concentration in Greenland ice showed values exceeding natural levels beginning some 2,500 years ago (Hong et al., 1996). This was attributed to the crude, highly polluting, smelting technologies used for copper production in Roman and medieval Europe as well as in China. Smoke from sea coal ﬁres was a nuisance in London, England, by the late 13th century with king Edward I banning the use of sea coals in kilns by 1307 (Te Brake, 1975). Perhaps because of poor record keeping, but air pollution was not a serious nuisance in London until the mid 17th century with sea coal again being the cause. At this time, pollution was a local problem. Pollution from the use of coal became a serious health problem during the industrial revolution in the late 18th century. In October, 1948, weather con- ditions caused a smog to saturate the air of Donora, Pennsylvania, USA. In a town of 14,000 inhabitants, an estimated 5,000–7,000 individuals became ill and 20 died before rain could disperse the smog three days later (Helfand et al., 2001). A few years later in 1952, London, England, was struck by a smog that lingered for four days caused by the city’s millions of coal stoves and local factories. In the year that followed, some 12,000 excess deaths occured that has been attributed to the acute and persistent effects of the smog (Bell and Davis, 2001). In 1976, a chemical plant near Seveso, Italy, exploded releasing massive amounts of dioxin into the air. Acute effects were seen in vegetation, birds and domestic animals. Individuals who happened to be on the deposition path of the cloud developed nausea, headache and eye-irritation. When the exposed population were followed during 15 years after the accident, excess mortailty cases of cardiovascular and respiratory diseases as well as some cancers were observed (Bertazzi et al., 1998). The above are examples of high exposures in a relatively small area. The everyday chemical exposure for most of us is characterized by a contamination consisting of hundreds or thousands of chemicals, all at relatively low levels (Kortenkamp, 2014). Risk assessment of chemicals is currently conducted on a chemical by chemical basis by regulatory bodies although there is a consensus that effects of simultaneous exposure to multiple chemicals, so called mixture effects, exist.

15 1.1 Environmental Contaminants This thesis comprises environmental contaminants from three classes of chemicals; persistent organic pollutants (POPs), plastic-associated chemicals (PACs) and metals. Environmental contaminants are deﬁned herein as a chemical sub- stance, excluding drugs, normally absent in the environment which, in sufﬁ- cient concentration, can adversly affect living organisms.

1.1.1 Persistent Organic Pollutants POPs belong to a class of carbon-based contaminants which have long half lives, i.e. they persist in the environment. POPs are toxic to humans and wildlife. They bioaccumulate, i.e. they accumulate in tissue, and they bio- magnify, i.e. the concentrations tend to get higher further up in the food chain. POPs are semi-volatile; in general they have boiling points above water and may vaporize when exposed to temperatures above room temperature. This property facilitates transport through the atmosphere and they can thus be deteceted all over the world, even in places where they never have been used (Stockholm Convention, 2008b). As POPs pose a threat to human health, the Stockholm Convention restricted or banned the use of 12 of the most dan- gerous POPs (Stockholm Convention, 2008a). A major source of POPs, such as dioxins and polychlorinated biphenyls (PCBs), for the population living along the east coast of Sweden is fatty ﬁsh from the Baltic sea.

Dioxins Polychlorinated dibenzodioxins (PCDDs) and dibenzofurans (PCDFs), or dioxins for short, is a group of highly toxic polyhalogenated organic compounds. The major source of dioxins is combustion of organic material (United States Environmental Protection Agency, 2010) but dioxins also appear as by products of certain manufacturing processes, e.g. smelting, chlorine bleaching of paper and the production of some pesticides (World Health Organization, 2014). The majority of human exposure to dioxins is through the diet, mainly from dairy products, red meat and ﬁsh (World Health Organization, 2014).

Polychlorinated Biphenyls PCBs are man-made compounds consisting of two benzene rings with 1–10 clorine atoms attached yielding a total of 209 different conﬁgurations (congeners). They were used as coolants and insulating ﬂuids because of their stability and heat resistance but also as plastizicers in paints and stabilizing additives in PVC coatings of electric wires and electronic components (United States Environmental Protection Agency, 2013). PCB use has been restricted in Sweden since 1973 and completely banned since 1986 but detectable levels can still be found in goods manufactured prior to the ban and they are widespread throughout the environment.

16 Pesticides The Nobel Prize in Physiology or Medicine in 1948 was awarded to Paul Her- mann Müller “for his discovery of the high efficiency of DDT as a contact poisson against several arthopods” (Nobel Media AB, 2014). DDT was used successfully to eradicate malaria from large parts of the world but the use was questioned by Rachel Carson in the book Silent Spring which examines the effects of pesticides on wildlife (Carson, 1962). The book eventually led to a ban on agricultural use of DDT in the USA in 1972. DDT is still used to- day to fight malaria but the use is controversial where concerns about health and the environment are weighted against the efficency of DDT in killing mosquitoes (Bouwman et al., 2011). Pesticides can be classified according to the type of pest they control, i.e. herbicides, insecticides, fungicides or bac- tericides to name a few. Pesticides can also be classified according to their chemical composition, i.e. organochlorine pesticides, organophosphate pesticides, carbamate pesticides and so on. Organochlorine (OC) pesticides, one of which is DDT, are considered toxic to humans and tend to be persistent in the environment and bioaccumulate in the food chain. Some organophosphate pesticides are also highly toxic to humans but do not persist in the environment (United States Environmental Protection Agency, 2014).

1.1.2 Plastic-Associated Chemicals Plastic-associated chemicals (PACs) are chemicals that are used in the production of plastics. They are not persistent and not acutely toxic as the POPs but humans are constantly exposed to them.

Phthalates Phthalates are substances added mainly to plastics to add stability, ﬂexibility, transparancy and/or durability. They are also used in e.g. vinyl ﬂooring, rain clothes, personal care products and as coatings on prescription drugs (Centers for Disease Control and Prevention, 2013; Kelley et al., 2012). Some phthalates are known agonists of the peroxisome proliferator-acivated receptors (PPARs) that play a role in the lipid and carbohydrate metabolism (Hurst and Waxman, 2003). There are thus concerns that phthalates can disrupt the deli- cate metobolic balances in the human body. Phthalates are metabolized in the body and are usually excreted within hours of exposure (Centers for Disease Control and Prevention, 2013).

Bisphenol A Bisphenol A (BPA) is a high production volume chemical which is used mainly to produce plastics and is a key monomer in the production of epoxy resins. BPA is also used in cash register thermal paper where it is present in its non- polymerized form which is likely to be more available for exposure than the

17 polymerized BPA (Swedish Chemicals Agency, 2012). Health effects of BPA has been much debated with several animal studies showing adverse effects. If human exposure to BPA is high enough to cause adverse health effects is widely debated (Mirmira and Evans-Molina, 2014) although detectable levels can be found in the majority of all human beings. It is generally believed that the main exposure route for humans is through the diet, although that view has been challenged due to the limited amount of data available (Vandenberg et al., 2013). BPA is used as coatings in tinned cans to keep the food or bev- erage from coming into contact with the can itself. Upon entering the body through the gastrointestinal tract, it is conjugated in the liver and excreted from the blood via the kidneys. The unconjugated form of BPA is thus not present in the body for longer periods although BPA entering the body via different routes may linger longer, for hours or days (Vandenberg et al., 2013). Although soluble in lipid, BPA does not seem to bioaccumulate.

1.1.3 Metals Metals are elements with good electrical and thermal conductivity that have the ability to lose electrons to other compounds. Humans are constantly sur- rounded by metals in various forms. Many metals are essential. Iron is an essential part of haemoglobin, while chromium, copper, manganese and zinc are required as enzyme cofactors. Most other metals have no known biological functions (Florea et al., 2012). Exposure to cadmium, lead and mercury has been associated to various health effects including kidney damage, neu- rological damage and osteoporosis (Järup, 2003). Although many metals are essential and naturally occuring, exposure to high levels of them have been associated with adverse health effects (Florea et al., 2012).

1.2 Chemical mixtures A recent EU report by Kortenkamp et al. (2009) deﬁnes chemical mixtures as:

• Substances that are mixtures themselves, e.g. the Aroclor series of PCB mixtures1

• Products that contain more than one chemical, e.g. cosmetics

• Chemicals jointly emitted from production sites, during transport processes, consumption and/or recycling processes

1http://www.epa.gov/osw/hazard/tsd/pcbs/pubs/aroclor.htm

18 • Several chemicals that might occur together in environmental media (water, soil, air), food items, biota and human tissues, as a result of emission from various sources, via multiple pathways.

Treating patients with several drugs to achieve an enhanced effect is common. Some diseases, like cancer and AIDS, require a combination therapy which works better than if each drug would have been administered alone (Ro- drigues, 2008). An example of serious biological consequences of drug–drug interactions was described by Psaty et al. (2004). Statins and fibrates are two classes of lipid lowering drugs. When the statin Lipobay was introduced to the US market, several other lipid lowering drugs were already in use. Lipobay contained cerivastatin as the active ingredient and interactions occured between cervivastatin and gemfibrozil, the active ingredient in other widely used drugs. Gemfibrozil inhibited the metabolic conversion of cerivastatin leading to internal accumulation of cerivastatin with levels far exceeding those of patients who took cerivastatin on its own, which led to sometimes severe toxic effects and an excess number of deaths (Psaty et al., 2004). This led to Lipobay being withdrawn from the market. Much is known regarding drug–drug interactions but less so about interactions between other types of chemicals. Figure 1.1 is taken from Faust et al. (2001) where algae were exposed to different active substances in herbicides (s-triazines), both in isolation and in two mixtures. The concentrations of the two mixtures were varied but the concentration ratio of the s-triazines remained constant. Concentration-response curves were fitted to each s-triazine and EC50 and EC1 were estimated from the fitted functions. The two mixtures were prepared so that all s-triazines were present in concentraions corresponding to their EC50 and EC1 values. When applied in isolation, the effects of the s-triazines were modest but when combined the observed effect was larger than the sum of the individual effects.

1.2.1 Concentration Addition and Independent Action Although many concepts to describe mixture effects have been proposed, two concepts have emerged; concentration addition (CA) and independent action (IA) (Greco et al., 1995). Concentration addition, or Loewe additivity (Loewe and Muischneck, 1926), looks at mixture effects from the standpoint that one chemical in a mixture can be replaced, in part or totally, by an equi-effective concentration of another chemical without changing the total effect. CA im- plies that every chemical contributes to the response of the mixture, regardless of concentration or individual effect. If additivity holds for a combination of

19 Figure 1.1. Figure 4 in Faust et al. (2001). Comparison of the total effects of s-triazine mixtures with the single effects of all 18 mixture components. (A) Concentrations of individual mixture components equal 1/18 of individual EC50 values. (B) Concentra- tions of individual mixture components equal individual EC1 values. CA, IA, predictions of the total effect according to concentration addition and independent action, respectively. Reprinted with permission from the publisher.

20 N exposures, the CA model can be written as N c ∑ i = 1 (1.1) i=1 Ei

In the above equation, Ei represents the concentration of chemical i associated th with a certain response and ci represents the concentration of the i chemical in combination with the other N − 1 chemicals yielding the same response. If the left hand side of equation 1.1 is less than one, there is a synergistic effect and if the left hand side is greater than one, the effect is antagonistic. It can be shown algebraically that this synergy corresponds to an interaction term with a regression coefficient larger than zero in a regression model whereas the an- tagonism corresponds to an interaction term with a coefficient less than zero. The additive case, i.e. the coefficient for the interaction term being equal to zero, corresponds to equation 1.1 (Gennings et al., 2005). Isoboles, i.e. curves of constant effect, can be used as a graphical assessment of the nature of the joint effect in the case of binary mixtures. If the mixture effect is additive, the isoboles are straight negatively sloped lines (see Howard and Webster, 2009, figure 1). CA was extended by Howard and Webster (2009) to include effects of partial agonists which renders straight line isoboles of varying slopes. A widespread application of CA is the concept of the toxic equivalence factor (TEF) (van den Berg et al., 1998). The TEF is defined for dioxin-like chemicals where they are expressed in terms of the dose of 2,3,7,8-TCDD needed to induce the same effect. An assessment of the combined effect is then a matter of adding up the equivalent doses. The toxic equivalent, TEQ, for a mixture is calculated as the sum of the constiuent chemicals concentrations multiplied by their TEFs. Independent action, or Bliss independence (Bliss, 1939), on the other hand assumes that the effect of a mixture of chemicals at different concentrations is equal to the product of each chemical’s effect applied individually at the spe- cific concentrations. A mixture of chemicals where all chemicals are present in concentrations below their individual no observed (adverse) effect limits (NOEL/NOAEL), i.e. the lowest concentration in which (adverse) effects are observed, would thus generate no observable (adverse) effects. Although not specifically stated by Loewe and Muischneck (1926) and Bliss (1939), CA is regarded to be applicable for mixtures in which the constituent chemicals act through a common mode of action whereas IA is applicable for mixtures in which the constituent chemicals act through different modes of action.

A Note on Interactions Toxicologists and epidemiologists define interaction as departure from additivity, but the concept of additivity differs between the two fields (Howard and Webster, 2013). Epidemiologists have argued that interactions should be defined as departures from risk or rate additivity whereas the statistical approach

21 is to include a product term in a regression model. Interaction in the statistical sense is thus scale dependent since an interaction on the risk ratio scale does not necessarily imply an interaction on the risk difference scale and vice versa. This has led to the concept of biologic interaction which is supposed to provide more mechanistic insight than statistical interaction (VanderWeele, 2009) although that view has been challanged by statisticians who argue that what is called biologic interaction is a special case of statistical interaction using additive odds or rate models (Andersen and Skrondal, 2010). Toxicologists have formulated a variety of null models all originating from mechanistic consider- ations. In this thesis, interaction is deﬁned in the spirit of the unifying concept presented in Gennings et al. (2005); if the slope of the concentration-response curve of one chemical changes with changing levels of another chemical, there is an interaction between the two. This is assessed on the scale of the linear predictor, i.e. the log-odds scale for a binary response and the scale on which the response is measured for a numerical response.

1.2.2 Experimental Studies on Mixtures Faust et al. (2003) exposed freshwater algae to mixtures of 16 chemicals with strictly different modes of action. The resulting mixture effect was predicted well by IA whereas CA slightly overestimated the joint effect. Both IA and CA predicted the joint toxicity of 16 similarly acting chemicals on Vibrio ﬁs- cheri (Altenburger et al., 2000) although IA underestimated EC50 by a factor of three. Junghans et al. (2006) studied the toxicological effects of environ- mentally realistic mixtures of pesticides on the reproduction of freshwater algae. The mixture contained chemicals with both common and diverse modes of action and the joint toxicity was well predicted by CA, with IA underesti- mating the total effect, although EC50 was well predicted by both IA and CA. Backhaus et al. (2004) studied the joint effects of 12 herbicides on green algae reproduction inhibition. The herbicides have a common mode of action and CA predicted the mixture effect well, as expected, although IA predicted the joint effect equally well. There are also situations where the nature of the mixture effect is different depending on whether CA or IA is used to describe the mixture effect (Altenburger et al., 2005).

1.2.3 Epidemiological Studies on Mixtures Few epidemiological studies have examined the joint effect of two contaminants. Froelich et al. (2009) studied the effect of co-exposure to lead and prenatal tobacco on the prevalence of Attention deﬁcit hyperactive disorder (ADHD) in almost 2,600 children using data from the National Health and Nutrition Examination Survey (NHANES). They found that having been exposed to tobacco and having high levels of blood lead increased the odds of

22 prevalent ADHD eight times while having not been exposed to tobacco dou- bled the odds of ADHD for the same blood lead levels. Claus Henn et al. (2012) studied the effect of co-exposure to lead and manganese on neurodevelopmental deﬁciencies. A cohort of 455 children were followed from birth until 36 months of age. Lead and manganese were measured at 12 and 24 months and neurodevelopment was measured every six months between 12 and 36 months of age. Results show that the negative effect of lead was more pronounced with simultaneous high levels of manganese. A number of epidemiological studies have examined mixture effects without explicitly modeling interactions. Gennings et al. (2010) studied the effect of four antiestrogenic PCBs and 12 estrogenic PCBs on the risk of endiometriosis. Instead of modeling each PCB in single contaminant models, a weighted linear combination of scaled PCBs was used. PCB 114 accounted for almost 100% of the relationship between the antiestrogenic PCBs and endiometriosis. The relationship between the estrogenic PCBs and endiometriosis was mainly driven by PCBs 99 and 188. A similar result was obtained by Herring (2010) who studied the associations between seven PCBs and their pairwise interactions on the risk of endiometriosis using a Bayesian hierarchical model. The model allowed different degrees of shrinkage for main effects and interactions with interactions more penalized than main effects. PCB 114 had a posterior inclusion probability of 99% while the other PCBs had inclusion probabilities of around 20%. An interaction between PCBs 142 and 206, with a posterior inclusion probability of 91%, was found.

1.3 An Unsolved Issue Controlled experiments are considered the gold standard regarding scientiﬁc evidence (Hammer et al., 2009). There are, however, questions that cannot be answered using controlled experiments, such as human health effects arising from exposure to multi-contaminant mixtures. The issues when assessing multi-contaminant mixtures in epidemiological studies were summarized by Billionnet et al. (2012); correlations between contaminants can be high making it difﬁcult to discern the effect of the individual contaminants, interactions between the contaminants may occur, the concentration-response curves may be non-linear and the interactions themselves may also be non-linear. There is a great need for statistical methods that can handle these types of issues.

23 1.4 Atherosclerosis Atherosclerosis is a disease of the innermost layer of the artery walls, called the intima. The disease is initiated when LDL-cholesterol, penetrating the artery wall, become trapped and oxidized. This oxidation triggers an inflam- matory response in which monocytes and T-lymphocytes penetrates the artery wall and binds to the oxidized LDL forming foam cells consisting of fat-laden macrophages. The accumulation of foam cells creates a necrotic core. Smooth muscle cells moves from the second layer of the artery wall, the media, to create a fibrous cap, consisting of collagen and elastin, around the necrotic core to prevent it coming into contact with the blood. This plaque buildup narrows the artery and obstructs the blood flow. Atherosclerosis is the major underlying cause of some major cardiovascular diseases like coronary heart disease, stroke, and peripheral artery disease. The most likely underlying mechanism is the rupture of a lipid-rich plaque which can create an occlusion of the artery leading to ischemia. Since coronary heart disease and stroke are two of the major causes of death worldwide, the understanding of atherosclerosis formation is crucial for mortality reduction.

1.4.1 Environmental Contaminants and Atherosclerosis Several animal studies have linked exposure to a wide variety of environmental contaminants to the development of atherosclerosis or atherosclerosis-related risk factors (Kirkley and Sargis, 2014). Figure 1.2 displays how environmental contaminants (pollutants) contribute to the development of vascular disease. The exact mechanisms by which environmental contaminants act to induce atherosclerosis is not known, but PCBs, dioxin and particulate matter have been shown to up-regulate the expression of MCP-1, a protein that attracts circulating monocytes to sites of inflammation (Vogel et al., 2007; Majkova et al., 2009). Endothelial dysfunction and inflammation have been shown in animal models investigating exposure to cadmium and PCBs (Hennig et al., 2002; Majkova et al., 2009) and activation of the aryl hydrocarbon receptor (AhR) by dioxin exposure has been shown to induce vascular inflammation (Wu et al., 2011). Few epidemiological studies have investigated the relationships between environmental contaminants, measured in the circulation, and human atherosclerosis. Three cross-sectional studies have originated from the Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS) cohort. Plaque occurence in the carotid arteries, as judged by ultrasound measurements, was more prevalent with increasing levels of several PCBs (Lind et al., 2012b). Elevated levels of the highly chlorinated PCBs and PCB 126 were also related to a more echolucent vascular wall, suggesting an effect on the lipid infiltration of the vascular wall. The phthalate metabolite mono-methyl phthalate (MMP) was related to the echogenecity of the artery wall in an inverted U-

24 Figure 1.2. Figure 1 in Kirkley and Sargis (2014). Contributions of environmental pollutants to cardiovascular disease pathology. This ﬁgure presents compounds and mechanisms related to the development of vascular disease observed in either in vivo exposure studies or in vitro cellular models. Reprinted with permission from the publisher. shaped manner (Lind and Lind, 2011). Elevated levels of BPA, mono-isobutyl phthalate (MIBP) and MMP were associated with a more echogenic vascular wall, while elevated levels of mono-2-ethylhexyl phthalate (MEHP) were associated with a more echolucent vascular wall. Nickel levels were found to be related to prevalent atherosclerotic plaques, while elevated aluminium and chromium levels both were related to a more echolucent vascular wall (Lind et al., 2012a). A study in the Korean Elderly Environment Panel showed associations between DEHP, the parent compound of MEHP, and oxidative stress as well as insulin resistance (Kim et al., 2013).

1.5 The Metabolic Syndrome Although the clustering of several risk factors was described in the 1920’s, it was not until 1988 when Reaven and other groups described the existence of a metabolic syndrome (Reaven, 1988; Lind et al., 1988). WHO proposed a deﬁnition of the syndrome in 1998 by which insulin resistance or an impaired glucose control was the key component of the syndrome (Alberti and Zim- met, 1998). In 2001, the National Cholesterol Education Program (NCEP) expert panel (National Cholesterol Education Program, 2002) suggested an alternative deﬁnition of the metabolic syndrome (MetS), not including insulin

25 resistance, as the co-occurence of at least three of the following

• Waist circumference > 102 cm for men and > 88 cm for women

• Serum triglyceride levels > 1.7 mmol/L

• HDL cholesterol < 1.03 mmol/L for men and < 1.29 mmol/L for women

• Blood pressure > 130/85 mmHg or use of antihypertensive medication

• Fasting blood glucose > 5.6 mmol/L

Although other definitions have been proposed, the so-called NCEP/ATPIII definition was widely accepted following some minor modifications and has been used extensively in clinical research. The usefulness of the MetS has been questioned, since the magnitude of risk of future cardiovascular events associated with the MetS is not greater than the sum of the risks associated with the individual components of the syndrome (Sundström et al., 2006a,b). Although the MetS does not include any unique information in terms of risk prediction, it is useful as a descriptive term for patients with multiple risk factors. As the prevalence of the MetS is increasing in both high and low- income countries (World Health Organization, 2011), it is anticipated that the incidence rates of cardiovascular diseases will start to increase again due to the higher burden of risk factors that will accompany the obesity epidemic seen worldwide.

1.5.1 Environmental Contaminants and the Metabolic Syndrome Several environmental contaminants have been associated to components of the MetS or to the MetS as such but in experimental and epidemiological studies. Ruzzin et al. (2010) exposed rats to a high fat diet consisting of either crude or refined fish oil from farmed Atlantic salmon. Rats who were fed crude fish oil developed insluin resistance and abdominal obesity. Exposure to POPs, and especially OC pesticides, led to inhibition of insulin action in adipocytes. Lee et al. (2007) studied the associations between POPs and the prevalence of metabolic syndrome in the NHANES 1999–2002 examination cycles. They used data from a cross-sectional study on 19 different POPs in 721 subjects who were free from diabetes. OC Pesticides were strongly related to the MetS with a five-fold increase in odds when comparing the highest versus lowest quartile. Associations between the PCBs and MetS were weaker, with a two- fold increase in odds when the highest and lowest quartiles were compared. The authors also investigated how the five different components of the syn-

26 drome were related to the POP levels. An increased waist circumference was mainly related to non-dioxin like PCBs and OC pesticides. Increased serum triglyceride levels were mainly related to dioxin-like PCBs and OC pesticides, while only OC pesticide levels were related to low HDL-cholesterol. Elevated blood pressure was only related to PCDFs, while increased fasting glucose was mainly related to non-dioxin like PCBs and OC pesticides. Stahlhut et al. (2007) studied the associations between urinary concentrations of six phthalate metabolites and waist circumference and insluin resistance assessed by the homeostatic model assessment (HOMA). Increased levels of most metabolites were associated with an increased waist circumference and with an increased HOMA even after adjusting for potential confounders. Sergeev and Carpenter (2010) studied the association between recidency in areas with waste sites containing POPs and MetS-related hospitalization rates in the state of New York, USA. Recidency were deﬁned as zip codes, classi- ﬁed as POP zip codes or non-POP zip codes. During three years of follow-up, a 39% increase in MetS-related hospitalization incidence was observed comparing POP zip codes to non-POP zip codes. In a paper from the PIVUS cohort, Lind et al. (2013) studied associations between 76 environmental and lifestyle factors and prevalent MetS using an EWAS approach (Patel et al., 2010). p,p’-DDE, trans-nonachlor (TNC) and PCBs 105, 118, 194 and 209 showed the strongest associations of the studied contaminants.

27 2. Aim

The overall aim of this thesis was to investigate to relationships between circulating levels of environmental contaminants and atherosclerosis and the metabolic syndrome in an elderly population. The more speciﬁc aims of the individual papers were

I To investigate the co-variation between the contaminants and to assess whether some contaminants could be used as markers for a larger number of contaminants.

II To assess, using simulated data, the applicability of boosted CARTs to predict mixture effects.

III To investigate additive and multiplicative mixture effects of a large number of environmental contaminants on two measures of atherosclerosis using boosted CARTs.

IV To investigate additive and multiplicative mixture effects of a large number of environmental contaminants on the metabolic syndrome using boosted CARTs.

28 3. Material and Methods

3.1 The PIVUS Study Data were obtained from the Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS) study. Eligible to the study were all subjects aged 70 living in the community of Uppsala, Sweden. The subjects were randomly chosen from the register of community living. A total of 1,016 subjects participated in the baseline investigation, giving a participation rate of 50.1%. The subjects went through an extensive physical examination and were subject to blood withdrawal. Blood samples were drawn in the morning after an overnight fast. No medication or smoking was allowed after midnight. The participants were asked to answer a questionnaire about their medical history, smoking habits and regular medication. Blood pressure was measured by a calibrated mercury sphygmomanometer in the non-cannulated arm to the nearest mmHg after at least 30 minutes of rest and the average of three recordings was used. Lipid variables and fasting blood glucose were measured by standard laboratory techniques. The MetS was deﬁned according to the NCEP/ATPIII criteria (National Cholesterol Education Program, 2002).

3.1.1 Measured Contaminants Persistent Organic Pollutants POP levels were analysed in stored plasma samples (Salihovic et al., 2011). A total of 23 POPs were measured: 16 PCBs; five OC pesticides: p,p’-DDE, hexachlorobenzene (HCB), cis-chlordane, trans-chlordane and trans-nonachlor (TNC); octachlorinated dibenzo-p-dioxin (OCDD); and one polybromi- nated diphenyl ether (BDE47). Briefly, samples were prepared using solid- phase extraction (SPE) and analysed using high resolution gas chromatography coupled to high resolution mass spectrometry (HRGC/HRMS). Among the 23 POPs measured, the detection rate was above 95.5% for all POPs except OCDD (80.6%) and BDE47 (72.2%). Two OC pesticides, cis-chlordane and trans-chlordane, with detection rates of less than 10% in the study population were not included in the final statistical analyses. Plasma concentrations were lipid adjusted in paper I but not in papers III and IV.

Plastic-Associated Chemicals Human serum was analyzed for levels of BPA and ten phthalate metabolites; mono-2-ethyl-5-hydroxyhexyl phthalate (MEHHP), mono 2-ethyl-5-oxohexyl

29 phthalate (MEOHP), MEHP, monobenzyl phthalate (MBZP), monocyclohexyl phthalate (MCHP), monoethyl phthalate (MEP), MIBP, monoisononyl phthalate (MINP), MMP, and mono-n-octyl phthalate (MOP) at ALS Canada following the general procedures presented by the Centers for Disease Control and Prevention (Calafat et al., 2005; Silva et al., 2003). Detectable levels were found in almost all subjects for four phthalate metabolites, MEHP, MEP, MIBP and MMP. Further details can be found in Olsén et al. (2011).

Metals All 11 metal elements in this study were determined in whole blood. The analysis was performed using inductively coupled plasma-sector ﬁeld mass spectrometry, ICP-SFMS, after microwave-assisted digestion with nitric acid (Ro- dushkin and Axelsson, 2000) according to a method accredited for 10 of the 11 metal elements tested, with Al being unaccredited. Further details can be found in Olsén et al. (2012b),

3.1.2 Ultrasound Evaluation of the Carotid Artery The carotid arteries were investigated by external B-mode ultrasound imaging (Acuson XP128 with a 10 MHz linear transducer, Acuson Mountain View, California, USA). The images were digitised and imported into the AMS (Artery Measurement Software) automated software for dedicated analysis of the intima-media thickness (IMT) and the echogenicity of the intima-media complex (intima-media grey scale median, IM-GSM). A 10 mm segment with good image quality was chosen for IMT analysis from the common carotid artery close to the bifurcation as shown in ﬁgure 3.1. The reported IMT and IM-GSM values are mean values from both sides.

Figure 3.1. Ultrasound image of the common carotid artery. The bifurcation is visible slightly to the left of the center and the internal as well as the external carotid arteries are visible on the left side. The intima-media complex of the far wall, marked by the arrow, was used for determination of IMT and IM-GSM.

30 A region of interest was placed manually around the intima-media segment that was evaluated for IMT and the software calculates the echogenicity of the intima-media complex from an analysis of the individual pixels within the region of interest on a scale from 0 (black) to 256 (white). Thus, a low IM- GSM value possibly indicates a lipid-rich wall, while a high value represents a vascular wall rich in collagen or calcium. The measurements were repeated in 30 random subjects, giving a coefﬁcient of variation of carotid artery IMT of 7.2% and 7.5% for IM-GSM.

3.2 Statistical Methods 3.2.1 Identifying Clusters of Contaminants Hierarchical cluster analysis was used to assess possible clustering of contaminants. For this analysis, the similarity was deﬁned using Hoeffding’s D statistic (Hollander and Wolfe, 1999) which allows for non-linear and non- monotonic dependencies between the variables. The value of 30 · D ranges from -0.5 to 1 with -0.5 indicating complete independence and 1 indicating complete dependence. In the presence of a large number of ties however, the value could be smaller.

Complete-linkage clustering was used in which the distance (1-similarity) between clusters at each stage is determined by the distance between the two points, one from each cluster, that are farthest away from each other. Cluster stability was assessed using multiscale multistage bootstrapping (Shimodaira, 2002). The procedure assesses the frequency of which the observed clustering is similar in repeated resampling, with sampling fractions ranging from 0.5 to 1.4, from the original data. These observed frequencies correspond to bootstrap probability values (BP) and z-scores for the BP-values for each sampling fraction were calculated according to

z = Φ−1(BP) where Φ−1(·) is the inverse of the standard normal distribution function. For each cluster, a line is then ﬁtted to the z-values by √ √ z(λ)=α λ + β λ −1 where λ denotes the sampling fraction. Asymptotically unbiased probability values (AU) were then calculated for each cluster by

AU = Φ(−α + β).

A stable cluster was deﬁned as a cluster having an AU > 0.95. 10,000 bootstrap replicates for each sampling fraction were performed to evaluate the cluster

31 stability. Additional analyses were done by examining the clustering pattern by body mass index (BMI) divided into three groups, < 25 (n = 342), 25–30 (n = 449) and > 30 (n = 225).

3.2.2 Identifying Suitable Markers Principal component analysis (PCA) was used to identify possible marker contaminants within each stable cluster. To render the PCA more effective, the contaminants were transformed using a nonlinear, additive transformation (Harrell Jr., 2001). Briefly, each contaminant was expanded as restricted cubic splines with five knots and the transformation that had the highest correlation with the best linear combination of the remaining contaminants, i.e. the first canonical variate, was used. Each cluster/sub-group was represented by the first principal component score resulting from a PCA of the variables in each cluster/sub-group. Multiple linear regression was then used to model the cluster/sub-group score. Each contaminant was modeled using restricted cubic splines with three knots placed at the 10th,50th and 90th percentiles of the contaminant distribution. The decrease in the coefficient of determination (R2 resulting from dropping each contaminant at a time until no contaminants remained in the model was assessed. The contaminants left in a model that had an R2 of at least 0.95 against the cluster/sub-group summary score were then defined as markers for the entire cluster/sub-group.

3.2.3 Classification and Regression Trees Classification and Regression Trees (CARTs) (Breiman et al., 1984) are very simple yet powerful. They partition the data into a set of disjoint regions and approximate the outcome with a constant value within these regions. This is accomplished via a series of binary splits in the input variables. The CART is grown in a top-down fashion by first finding the variable and split point that optimizes a statistical criterion, e.g. the residual sum of squares. Within each formed subset the optimal split is determined using the subset of observations passing through the previous split. This is repeated until the number of observations left is too low to be split, typically <10. A CART consisting of a single split is said to have depth one (d = 1), a CART with two splits is said to have depth two (d = 2) and so on. CARTs are thus able to fit complex interactions as each split after the first is conditional on the former split. This means that if a higher order interaction is present, its lower order components are also present. A CART of depth d can allow interactions of at most order d but usually contains combinations of interactions and nonlinear effects, with the latter being handled via successive splits on the same variable. The fitted CART can then be visually assessed for any interactions and/or nonlinear effects. The ability to automatically handle interactions and nonlinear effects

32 makes CARTs attractive in the study of mixture effects. Further details on CARTs can be found elsewhere (Hastie et al., 2009; Harrell Jr., 2001; Ripley and Venables, 2002). Figure 3.2 shows a CART. To form predictions from this CART one starts at the top node and follow the splits to the right or to the left depending on the values of x1 and x2 respectively. Nodes containing the predictions are called terminal nodes.

yes x2 < 0.59 no

x1 < 0.35 x1 < 0.46

0.029 x2 < 0.35 x1 < 0.21 x1 < 0.8

0.11 x1 < 0.77 0.056 0.28 0.5 x2 < 0.84

0.23 0.41 0.66 0.91

Figure 3.2. Example CART. Values within boxes are predicted values conditional on the splitting rules in each node above.

CARTs are easily interpretable but have several drawbacks. One drawback is the selection bias towards variables with many possible split points (Breiman et al., 1984). Another issue is that CARTs are highly variable: a small change in the outcome data can lead to a different CART. Purely additive relationships are poorly approximated by CARTs and much information is lost due to the binary splits of the input variables. Predictions from CARTs are usually somewhat crude and they also tend to overﬁt the data because of the amount of searching done. The price paid is that stable CARTs that cross-validate well

33 usually consist of no more than a few terminal nodes and are thus not very discriminating (Harrell Jr., 2001).

3.2.4 Stochastic Gradient Boosting In the language of statistical learning, single CARTs are called weak learners because of their poor predictive performance. Stochastic gradient boosting (Friedman, 2002) (hereafter called boosting) is a numerical technique created around the idea that many weak learners can be combined into a strong learner, which is called en ensemble, with superior predictive performance. The goal is to accurately map a set of explanatory variables x to an outcome variable y via a function F(x), which is usually called the target function, estimated by an additive expansion M Fˆ(x)= ∑ βmb(x;γm) (3.1) m=1 where M is the number of weak learners; βm are the expansion coefficients and b(x;γm)are individual weak lerners characterized by the parameters γm (Hastie et al., 2009). Accuracy is defined by a loss function L(y,F) which represent the loss in predicting y with F(x). A detailed description of gradient boosting is beyond the scope of this paper but with CARTs as the weak learners the algorithm briefly works as follows

1. Initialize Fˆ0(x) to a constant α.

2. Randomly sample a fraction η from the data without replacement.

3. Using η, compute the negative gradient of the loss function, zm = −∇L, and ﬁt a depth d CART, g(x), predicting z.

4. Update Fˆm(x) ← Fˆm−1(x)+λρg(x).

5. Iterate steps 2 through 4 M times.

In step 4, ρ is the step size along the gradient and λ is a shrinkage parameter which slows down the learning to reduce overfitting. The parameters M, d and λ can be tuned using the bootstrap or cross-validation, although a value of d 5 is often a reasonable starting point Hastie et al. (2009). For ( , )= 1 ( − )2 squared error loss L y F 2 y F the negative gradient is the ordinary residual, so each iteration in the above algorithm fits a CART predicting the residuals from the CART fitted in the previous step. For absolute error loss L(y,F)=|y − F| the negative gradient is the sign of the residual making it more robust to skewed outcomes than the squared error loss function. Loss functions for binary and multinomial data as well as Poisson and time to event

34 (survival) data are also available (Ridgeway, 2013). The subsampling in step 2 not only reduces computing time but also usually improves predictive performance (Friedman, 2002). A typical value of η is 0.5 meaning that in each step a random sample of half the data is used to grow the CART but η can be smaller or larger depending on the sample size. More comprehensive descrip- tions of boosting are given elsewhere (Hastie et al., 2009; Friedman, 2002; Friedman et al., 2000; Friedman, 2001; Elith et al., 2008). An additional step was made in step 3 in papers III and IV. Before each split, a random subset pS of the predictors are chosen as candidates for the split.

3.2.5 Variable Importance and Interpretation A single CART is easily interpretable, but this feature is lost in the gradient boosted model, which usually contains hundreds or thousands of CARTs. The gradient boosted model also does not provide regression coefficients, confidence intervals or p-values for the independent variables, so the difficulty of understanding and evaluating the model is increased. Variable importance and partial dependence plots are two tools that aid interpretation. The measure of variable importance in boosted CARTs is based on the number of times a variable is involved in a split, weighted by the squared improvement of the model as a result of the split. The measure thus incorporates both additive as well as interaction effects. Graphical visualization of the fitted function as a function of one or more of the explanatory variables provides a comprehensive summary of its dependence on the variables, especially if the function is dominated by additive terms and/or lower-order interactions. The partial dependence of a subset S of the explanatory variables can be estimated by

1 N FˆS(xS)= ∑ F(xS,x−S(i)) (3.2) N i=1 where x−S(i) denotes the data values of the variables not in S. FˆS(xS) is the effect of a subset S of variables on the outcome after accounting for the average effect of the other variables not in S. For boosted CARTs, FˆS(xS) can be calculated from the individual CARTs without reference to the data which would otherwise be computationally very expensive (Friedman, 2001).

3.2.6 Assessment of Interaction Effects The H statistic was deﬁned by Friedman and Popescu (2008) as a measure of interaction strength. The idea behind it is that if two variables x j and xk do not interact with each other, the function Fjk(x j,xk) can be written as the sum of

35 two functions; one that does not depend on xk and one that does not depend on x j, i.e. Fjk(x j,xk)=Fj(x j)+Fk(xk). The statistic Hjk is related to the fraction of variance of Fjk(x j,xk) not captured by Fj(x j)+Fk(xk) and ranges from 0 to 1, with larger values indicating stronger interaction effects. For two-way 2 interactions Hjk is defined as N 2 ∑ = Fˆ (xij,x ) − Fˆj(xij) − Fˆ (x ) H2 = i 1 jk ik k ik (3.3) jk ∑N ˆ 2 ( , ) i=1 Fjk xij xik where i = 1,2,...,N is the number of observations in the data. The interaction = 2 strength Hjk is then calculated as Hjk Hjk. The H statistic is not restricted to two-way interactions and generalizes to interaction effects of any order. H can be used to assess whether a particular variable interacts with any other variable by noting that F(x)=Fj(x j)+F− j(x− j) if variable x j does not interact with any other variable and by inserting the relevant partial dependencies in 3.3. H is not comparable to the traditional way of assessing interactions via regression coefficients as it is a relative measure. Even if an interaction is absent from F(x) the sample based estimate of H will not necessarily be zero as sampling fluctuations may introduce spurious interactions in Fˆ(x). A parametric bootstrap procedure can be used to generate a null distribution for H in which artificial outcome data containing only additive effects is generated according to y˜i = FA(xi)+ yp(i) − FA(xp(i)) (3.4)

or in the case of binary outcome,y ˜i = bi, where bi is a Bernoulli random variable generated according to

1 Pr(bi = 1)= (3.5) 1 + exp(−FA(xi))

In equation 3.4, p(i) is a random permutation of the integers 1,...,N. FA(x) is the closest fit to the target containing no interaction effects. This could be accomplished by restricting the depth of the CARTs to d = 1. Nonlinear effects are still captured by the sequential nature of the boosting algorithm even if the individual CARTs are restricted to contain a single split. Other methods could also be used to fit the additive model, e.g. using Generalized Additive { , }N Models (Wood, 2006). The full model is then fitted to the data y˜i xi 1 , where x are the original data. H is then calculated and corresponds to what could be expected if no interactions are present in the target function. The process is repeated many times, and a null distribution for H is obtained, which is hereafter denoted by H0. By comparing the observed value of H to H0, an idea is obtained of which variables participate in interactions and the order of these interaction effects (Friedman and Popescu, 2008).

36 3.2.7 Simulating Contaminant Data Artiﬁcial data were simulated based on the contaminants measured in the PIVUS study. All contaminants were assumed to follow log-normal distributions and the PCBs were assumed to correlate to a certain degree. Let X be the matrix of contaminants. Then

log(X) ∼ MVN(μ,Σ) with μ representing the vector of log-scale mean values and Σ is the covariance matrix deﬁned in the simulation as Σ Σ = PCB 0 0 ΣRest Empirical log-scale mean values were used as μ. The emprical covariances for the six PCBs were used in ΣPCB. All other contaminants were uncorrelated, i.e. ΣRest is a diagonal matrix with variances. Simulation parameters are shown in table 3.1. ΣPCB was deﬁned as

⎛ 153 118 170 209 126 169 ⎞ 153 0.19 ⎜ ⎟ 118 ⎜ 0.17 0.32 ⎟ ⎜ ⎟ 170 ⎜ 0.15 0.12 0.17 ⎟ ΣPCB = ⎜ ⎟ 209 ⎜ 0.12 0.07 0.15 0.27 ⎟ 126 ⎝ 0.18 0.15 0.16 0.14 0.85 ⎠ 169 0.13 0.10 0.13 0.12 0.16 0.22

The target function F(xS) used in the simulations was generated similar to the target function in (Friedman and Popescu, 2008), as

−3(1−s[PCB 170]2) −3(1−s[p,p’-DDE]2) F(xS)=11 · e · e · − ( − [ ]2) − ( − [ ]2) e 2 1 s MMP · e 2 1 s Cd − 1.6sin2 (π · s[OCDD]) + s[BPA](0.6 + 1.8 · I[Sex = Male]) (3.6) The function s[x] in equation 3.6 transforms x to range somewhat uniformly between 0 and 1 for numerical convenience and I[Ω] equals 1 if the logical condition Ω is true and 0 otherwise. The variables selected in 3.6 were chosen so that one of the correlated PCBs as well as one contaminant from each class (metals, phthalates) would be part of the target. The target function, F(xS), thus includes a four-way interaction between PCB 170, p,p’-DDE, MMP and Cd and a non-linear dependency on OCDD which is U-shaped on the log- scale. Also included is a BPA by sex interaction.

37 Table 3.1. Mean values and standard deviations (SD) used in the simulation Contaminant Mean value (SD) PCB 118 5.27 (0.56) PCB 126 3.68 (0.89) PCB 153 7.24 (0.41) PCB 169 5.11 (0.45) PCB 170 6.19 (0.38) PCB 209 3.23 (0.52) MIBP 2.99 (1.09) MMP 0.51 (1.13) OCDD 0.99 (0.59) HCB 5.54 (0.46) TNC 4.91 (0.61) p,p’-DDE 7.52 (0.93) BDE47 2.73 (0.69) BPA 1.18 (0.88) MEHP 1.82 (1.43) MEP 2.49 (0.71) Al -0.43 (0.39) Cd 0.96 (0.63) Co 0.44 (0.63) Cr 2.58 (0.48) Cu 2.55 (0.15) Hg 2.18 (0.67) Mn 4.93 (0.30) Mo 2.31 (0.39) Ni 4.54 (0.97) Pb -2.49 (0.49) Zn 4.55 (0.13)

The response y was then generated as yi = F(xiS)+εi where εi ∼ N(0,σ) with σ chosen to obtain signal to noise ratios (SNRs) of 2, 1, 0.5 and 0.1 re- 2 σ ( ) = F xS spectively. The signal to noise ratio is defined as SNR σ 2 , i.e. the ratio of the target function’s variance to the noise variance. A large value of the SNR indicates more signal than noise and a stronger relationship between the outcome and the predictors. The SNRs were chosen to represent a strong relationship (SNR = 2), a moderate relationship (SNR = 1), a weaker relationship (SNR = 0.5) and a very weak relationship (SNR = 0.1). The coefficients were chosen so that each variable in equation 3.6 would have approximately the same relative influence when SNR = 2. The SNRs present in the simulated data were within 10% of the target SNRs.

38 4. Results and Discussion

4.1 Paper I Figure 4.1 shows how the contaminants clustered in the PIVUS and the NHANES studies. Two stable clusters could be discerned in the PIVUS data. One cluster corresponded to low/medium chlorinated PCBs, i.e. PCBs with ≤ six chlorine atoms and one cluster corresponding to medium/high chlorinated PCBs, i.e. PCBs with ≥ six chlorine atoms.

Two principal components were enough to explain > 90% of the medium/high chlorinated cluster variability while three principal components were needed to explain >90% of the low/medium chlorinated cluster variability. All PCBs in the medium/high cluster loaded negative on the first principal component, whereas PCBs 194, 206 and 209 loaded negative on the second component as well. PCBs 156, 157, 170, 180, and 189 loaded positive on the second component, suggesting two subgroups within the medium/high chlorinated cluster (figure 4.2). All contaminants in the low/medium chlorinated cluster, except HCB, loaded negative on the first principal component. For the contaminants that loaded positive on the third component (TNC, PCBs 99, 138, and 153) only TNC loaded negative on the second component. PCBs 74, 105, and 118 all loaded negative on the third component, suggesting two subgroups within the low/medium chlorinated cluster as well (figure 4.3). PCBs 118 and 153 explained 91% and 95% of their respecive sub groups’ variances and PCBs 170 and 209 both explained 81% of their sub groups’ variances.

4.1.1 Discussion, paper I Contaminants measured in the circulation co-varied little as two clusters of PCBs could be found. The number of chlorine atoms in the PCBs is the major determinant of lipid solubility and thereby of the toxicokinetic properties of the contaminants. PCBs with a high degree of chlorination have a con- siderably longer half-life than those with a lower degree of chlorination. It is therefore reasonable to assume that the two PCB clusters are derived from the same sources of exposures and are retained in the body in a similar way because of their similar elimination half-life. However, similar elimination half-life is not sufﬁcient for the POPs to cluster in paper I. p,p’-DDE, a break- down product of DDT, has a similar elimination rate to the highly chlorinated PCBs, but is not found in the medium/high-chlorinated PCB cluster. This

39 lsesaeecoe ndse oe tponly). (top boxes dashed in enclosed are clusters 40 4.1. Figure and NHANES irrhclcutraayi ftecnaiat rmthe from contaminants the of analysis cluster Hierarchical bto)suisuigHoeffding’s using studies (bottom)

Dissimilarity (1 − 30 * Hoeffding D) Dissimilarity (1 − 30 * Hoeffding D)

0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0

PCB189 Mo Ni PCB206 MEP BPA PCB209 MMP Zn PCB156 MIBP PCB126 DDE PCB157 BDE47 Hg PCB194 Pb PCB169 PCB170 PCB194 NHANES PCB206

PCB180 PCB209 PIVUS PCB157 Cd PCB189 PCB156 Pb PCB180

D PCB170 Hg OCDD HCB stesmlrt esr.Stable measure. similarity the as OCDD TNC PCB99 PCB99 PCB153 PCB138 PCB74 PCB138 PCB118 PCB105 PCB153 Cr MEHP PCB74 Al Cu PCB105 Mn Cd PCB118 Co PIVUS (top) ● PCB156

● PCB157

PCB170 ● ● PCB189 PCB180 ● 0.2 0.0 0.2 0.4 0.6 − ● PCB194 Second principal component

0.4 ● PCB206 − 0.6 ● PCB209 −

−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6

First principal component

Figure 4.2. Component loadings on the two ﬁrst principal components resulting from a principal component analysis on the contaminants in the medium/high chlorinated cluster. Broken lines indicate a zero loading. might indicate that the route of exposure for p,p’-DDE is not similar to that seen for the high-chlorinated PCBs. Another explanation might be the timing of the peak of exposure. In Sweden, DDT was banned several years before PCBs were banned. Thus, even if the toxicokinetics of p,p’-DDE and the medium/high chlorinated PCBs share similarities in terms of elimination half- life, the timing of exposure might inﬂuence the degree of co-variation between the contaminants. p,p’-DDE has often been used as a marker for POP exposure. The results in paper I suggest a poor correlation between p,p’-DDE and the other POPs. An important implication of this result is that it is necessary to measure many contaminants to get the greater picture needed for studies of mixture effects.

4.2 Paper II Table 4.1 shows the bootstrap validated root mean squared error (RMSE), R2, the optimal M and d as well as the M and d chosen by the one standard error (SE) rule. Figure 4.4 shows the ten most influential variables from the different scenarios. The seven variables present in the target function were correctly identified among the ten most important variables in the first three scenarios. For SNR = 0.1, the correct variables were not identified among the top ten as sex came at 14th place in the variable importance ranking. A few unimportant variables

41 Third principal component ≥ 0

PCB138

● ● PCB99

●

PCB153 0.2 0.0 0.2 0.4 0.6 − Second principal component ● TNC 0.4 − 0.6 −

−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6

First principal component

Third principal component < 0

● HCB

PCB105 ●

● PCB118

● PCB74 0.2 0.0 0.2 0.4 0.6 − Second principal component 0.4 − 0.6 −

−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6

First principal component

Figure 4.3. Component loadings on the two ﬁrst principal components resulting from a principal component analysis on the contaminants in the low/medium chlorinated cluster. Broken lines indicate a zero loading.

42 (Mn, Pb and two PCBs) came before PCB 170 in the importance ranking as well.

20 40 60 80 100 SNR = 2 SNR = 1 Mn ● Mn ● PCB 169 ● PCB 153 ● PCB 153 ● PCB 169 ● Sex ● Sex ● OCDD ● OCDD ● MMP ● MMP ● Cd ● Cd ● BPA ● PCB 170 ● PCB 170 ● BPA ● p,p'−DDE ● p,p'−DDE ●

SNR = 0.5 SNR = 0.1 PCB 126 ● PCB 170 ● PCB 169 ● PCB 169 ● Mn ● MMP ● Sex ● PCB 126 ● OCDD ● OCDD ● PCB 170 ● Pb ● MMP ● Mn ● Cd ● Cd ● BPA ● BPA ● p,p'−DDE ● p,p'−DDE ●

20 40 60 80 100 Relative influence

Figure 4.4. Variable importance for the ten most important variables for each SNR. The importance measure has been scaled so that the most important variable has a value of 100.

4.2.1 Assessment of Interaction Effects in the Simulated Data The top left panel of figure 4.5 shows the strengths of the total interaction effects involving each of the ten most influential variables for SNR = 2. Dots are observed values of H and boxes represent the derived null distributions of H for each variable. The top left panel of figure 4.5 shows that p,p’-DDE, PCB 170, BPA, Cd, MMP and Sex all seem to be involved in interactions, as the observed values of H are well outside the null distribution, whereas OCDD, though it is an important variable, does not seem to interact with any other of the top ten variables.

Table 4.1. Optimal parameters and parameters chosen according to the one SE rule with R2 as the metric. RMSE and R2 values are bootstrap validated using 250 resam- ples. Optimal One SE rule dMRMSE R2 dMRMSE R2 SNR = 2 8 3,900 1.11 0.57 6 3,500 1.11 0.57 SNR = 1 8 3,000 1.50 0.40 6 2,700 1.52 0.39 SNR = 0.5 10 2,100 2.04 0.23 6 2,400 2.04 0.23 SNR = 0.1 10 1,300 4.29 0.02 5 1,400 4.29 0.02

43 The top right panel of figure 4.5 shows interaction strengths for two-way interactions with p,p’-DDE for SNR = 2. PCB 170 is clearly involved in interactions with p,p’-DDE, as are Cd and MMP. Sex and BPA were seen to be involved in interactions but do not interact with p,p’-DDE, as their observed values of H are well inside their respective null distributions. The bottom left and right panels of figure 4.5 shows interaction strengths for three- and four-way interactions with p,p’-DDE, PCB 170 (left panel) and Cd (right panel) for SNR = 2. The four interacting variables p,p’-DDE, PCB 170, Cd and MMP have been correctly identified as important variables and as variables participating in interactions. The null distributions for H in the bottom panels of figure 4.5 are very narrow, however, so even small observed values of H could become significant.

0.0 0.1 0.2 0.3 Total interaction Two−way interactions with p,p'−DDE

Mn ● Mn ●

● PCB 169 PCB 169 ● PCB 153 ● PCB 153 ● Sex ● Sex ● OCDD ● OCDD ● MMP ● MMP ● Cd ● Cd ● BPA ●

● PCB 170 ● BPA

p,p'−DDE ● PCB 170 ●

Three−way interactions with p,p'−DDE and PCB 170 Four−way interactions with p,p'−DDE, PCB 170 and Cd

Mn ● Mn ●

PCB 169 ● PCB 169 ●

PCB 153 ● PCB 153 ● Sex ● Sex ● OCDD ● OCDD ● MMP ●

MMP ● Cd ●

BPA ● BPA ●

0.0 0.1 0.2 0.3 Interaction strength

Figure 4.5. Interactions for SNR = 2. Black dots represent observed values of H, and boxes represent the null distributions H0. Small tick marks represent values of the null distribution below or above the 5th and 95th percentiles respectively.

Figures 4.6 and 4.7 show the same for SNR = 1 and SNR = 0.5 as figure 4.5 does for SNR = 2. The top left panels of figures 4.6 and 4.7 show the total interaction strengths. The effect of the narrow null distributions is apparent in the lower left panel of figure 4.7. A spurious three-way interactions involving p,p’-DDE, Cd and PCB 169 could be seen, although the observed value of H is small. The correct four-way interactions were identified, however (figures 4.6 and 4.7, lower right panels). The top left panel of figure 4.8 shows the strengths of the total interaction effects when SNR = 0.1. Only p,p’-DDE and BPA seem to be involved in interactions and neither the correct two-way interactions (top right panel) nor the correct three-way (bottom panels) interactions were identified. Figure 4.9 shows interaction strengths for the two-way interactions with sex for SNR = 2, 1 and 0.5. BPA is clearly interacting with sex in each of the three

44 0.0 0.1 0.2 0.3 Total interaction Two−way interactions with p,p'−DDE

Mn ● Mn ●

● PCB 153 PCB 153 ● PCB 169 ● PCB 169 ● Sex ● Sex ● OCDD ● OCDD ● MMP ● MMP ● Cd ● Cd ● PCB 170 ●

● PCB170 ● PCB 170

p,p'−DDE ● PCB170 ●

Three−way interactions with p,p'−DDE and PCB 170 Four−way interactions with p,p'−DDE, PCB 170 and Cd

Mn ● Mn ●

PCB 153 ● PCB 153 ●

PCB 169 ● PCB 169 ● Sex ● Sex ● OCDD ● OCDD ● MMP ●

MMP ● Cd ●

BPA ● BPA ●

0.0 0.1 0.2 0.3 Interaction strength

Figure 4.6. Interactions for SNR = 1. Black dots represent observed values of H, and boxes represent the derived null distributions H0. Small tick marks represent values of the null distribution below or above the 5th and 95th percentiles respectively.

0.0 0.1 0.2 0.3 Total interaction Two−way interactions with p,p'−DDE

PCB 126 ● PCB 126 ●

● PCB 169 PCB 169 ● Mn ● Mn ● Sex ● Sex ● OCDD ● OCDD ● PCB 170 ● PCB 170 ● MMP ● MMP ● Cd ●

● BPA ● Cd

p,p'−DDE ● BPA ●

Three−way interactions with p,p'−DDE and Cd Four−way interactions with p,p'−DDE, Cd and MMP

PCB 126 ● PCB 126 ●

PCB 169 ● PCB 169 ●

Mn ● Mn ● Sex ● Sex ● OCDD ● OCDD ● PCB 170 ●

PCB 170 ● MMP ●

BPA ● BPA ●

0.0 0.1 0.2 0.3 Interaction strength

Figure 4.7. Interactions for SNR = 0.5. Black dots represent observed values of H, and boxes represent the null distributions H0. Small tick marks represent values of the null distribution below or above the 5th and 95th percentiles respectively.

45 0.0 0.1 0.2 0.3 Total interaction Two−way interactions with p,p'−DDE

PCB 170 ● PCB170 ●

● PCB 169 PCB169 ● MMP ● MMP ● PCB 126 ● PCB126 ● OCDD ● OCDD ● Pb ● Pb ● Mn ● Mn ● Cd ●

● BPA ● Cd

p,p'−DDE ● BPA ●

Three−way interactions with p,p'−DDE and Pb Three−way interactions with p,p'−DDE and PCB 126

PCB170 ● PCB170 ●

PCB169 ● PCB169 ●

MMP ● MMP ●

PCB126 ● OCDD ●

OCDD ● Pb ●

Mn ● Mn ●

Cd ● Cd ●

BPA ● BPA ●

0.0 0.1 0.2 0.3 Interaction strength

Figure 4.8. Interactions for SNR = 0.1. Black dots represent observed values of H and boxes represent the null distributions. Small tick marks represent values of the null distribution below or above the 5th and 95th percentiles respectively. scenarios. SNR = 0.1 is not included in figure 4.9 as sex was not found among the ten most important variables. Partial dependences on BPA conditioned on sex are seen in figure 4.10 with SNR = 2 (top left panel), SNR = 1, (top right panel) and SNR = 0.5 (bottom left panel). The non-linear dependence on OCDD is captured well as is shown in figure 4.11 although the U-shape is not as clear for SNR = 0.1 as it is for the other SNRs.

4.2.2 Visualizing the Four-way Interaction The four-way interaction between p,p-’DDE, PCB 170, Cd and MMP for SNR = 0.5 is seen in figure 4.12. The x- and y-axes of each panel represent p,p’- DDE and PCB 170 levels respectively. Cd and MMP are represented as shin- gles Sarkar (2008) which are overlapping intervals used to represent continuous variables in a high-dimensional setting. Panels going left to right represent increasing levels of Cd while panels going bottom to top represent increasing levels of MMP. The bar to the right of the figure provides the color codes for the predicted outcome. The bottom left panel of figure 4.12 shows the joint effect of p,p’-DDE and PCB 170 while CD and MMP are both at low levels. The synergistic effect is hardly discernable. Following the panels right or up from the bottom left panel shows the joint effect when Cd or MMP increases. The synergistic effect becomes clearer, although it is still small. Following the diagonal from the bottom left panel shows the joint effect of p,p’-DDE and PCB 170 as Cd and MMP both increase, and the synergistic effect is obvious in the top right panel.

46 0.0 0.1 0.2 0.3 0.4 SNR = 2 SNR = 1 SNR = 0.5

Mn ● Mn ● PCB 126 ●

PCB 169 ● PCB 153 ● PCB 169 ●

PCB 153 ● PCB 169 ● Mn ●

Sex OCDD ● ● OCDD ●

MMP ● MMP ● PCB 170 ●

Cd ● Cd ● MMP ●

BPA ● BPA ● Cd ●

PCB 170 ● PCB 170 ● BPA ●

p,p'−DDE ● p,p'−DDE ● p,p'−DDE ●

0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 Interaction strength

Figure 4.9. Two-way interactions with sex. Black dots represent observed values of H and boxes represent the null distributions H0. Small tick marks represent values of the null distribution below or above the 5th and 95th percentiles respectively.

010203040 SNR = 2 SNR = 1 SNR = 0.5

1.5

1.0 F M Partial dependence on BPA

0.5

0 10203040 0 10203040 BPA

Figure 4.10. Partial dependence on BPA. Partial dependence on BPA conditioned on sex for SNR = 2 (left panel), SNR = 1 (middle panel), SNR = 0.5 (right panel). Solid lines are the partial dependencies for females, and the dashed lines are the partial dependencies for males.

47 0 5 10 15 20 SNR = 2 SNR = 1

1.2

1.0

0.8

0.6

0.4

SNR = 0.5 SNR = 0.1

1.2

1.0 Partial dependence on OCDD

0.8

0.6

0.4

0 5 10 15 20 OCDD

Figure 4.11. Partial dependence on OCDD. Partial dependence on OCDD for SNR = 2 (top left panel), SNR = 1 (top right panel), SNR = 0.5 (bottom left panel) and SNR = 0.1 (bottom right panel).

4.2.3 Discussion, paper II Paper II showed that boosted CARTs may be a useful tool to generate hypotheses regarding interactions. A four-way interaction was uncovered with an SNR as low as 0.5. Some signs of spurious interactions were found making the issue of validation very important. A repeated split-sample validation approach was suggested in the paper although further simulations have since revealed that the results in the paper may be fortuitous. Splitting the data is a random process and the impact of different values of the random seed needs to be investigated furher. Another approach to internally validate the ﬁndings, that does not involve data-splitting, may be to estimate the optimism in H by using the bootstrap. A bootstrap sample is drawn from the data and the ensemble is built on that. H is then estimated in both the original data and the bootstrap sample. The difference in estimated H between the bootstrap sample and the original data is a measure of the optimism in estimating H in the original data. By repeating the process many times and subtracting the average optimism, an optimism corrected value of H can be obtained and compared with the null distribution. This technique has been used to estimate various indices of model performance (Efron and Tibshirani, 1994; Harrell Jr., 2001).

48 Figure 4.12. Visualization of the four-way interaction. The x- and y-axes of each panel represent p,p’-DDE and PCB 170 respectively. Levels of Cd increase with panels going left to right, and levels of MMP increase with panels going bottom to top. The plotted ranges are from the 10th to the 90th percentiles of each variable’s distribution to ease interpretation.

49 4.3 Paper III 4.3.1 IMT Associations between the studied contaminants and IMT were modest. An ensemble consisting of 6,300 depth 2 CARTs acheived an R2 of 0.065. Using the one SE rule, an R2 of 0.058 was acheived with an ensemble of 4,800 depth 1 CARTs, indication that no interactions are present. Figure 4.13 shows the relative inﬂuence of the studied variables on IMT. Systolic blood pressure was by far the most important predictor followed by BMI and LDL cholesterol. Two contaminants, PCB 126 and Cd, followed but their inﬂuences on IMT were marginal.

Antihypertensive medication ● Diabetes medication ● Cu ● Pb ● Education ● Mo ● MEP ● Mn ● BDE47 ● Zn ● Diabetes ● Ni ● Cr ● PCB 118 ● OCDD ● Hg ● HCB ● Sex ● Al ● PCB 169 ● Statin use ● Co ● p,p'−DDE ● PCB 153 ● BPA ● TNC ● MIBP ● PCB 170 ● MEHP ● Smoking status ● Physical activity ● Diastolic blood pressure ● PCB 209 ● MMP ● Triglycerides ● HDL cholesterol ● Cd ● PCB 126 ● LDL cholesterol ● BMI ● Systolic blood pressure ●

0 20406080100

Figure 4.13. Variable importance for the studied variables when predicting IMT.

Partial dependencies on the six most important variables are shown in ﬁg- ure 4.14. IMT increases with increasing systolic blood pressure, BMI, LDL cholesterol levels, PCB 126 and Cd levels while it decreases with increasing HDL cholesterol levels. Figure 4.15 shows the predicted change in IMT in μm with corresponding bootstrapped 95% conﬁdence intervals for an IQR increase in the studied contaminants.

50 tapdpitie9%cndneitras h ml ikmrsa h otmof bottom the at variables. marks the tick of deciles small are The panel each intervals. conﬁdence 95% pointwise strapped 4.14. Figure Predicted IMT (mm) 0.80 0.85 0.90 0.95 0.80 0.85 0.90 0.95 0102030 1234567 0 5 200 150 100 Systolic blood pressure ata eedniso h ih otipratvralswt boot- with variables important most eight the on dependencies Partial LDL cholesterol Cd 0 0 0 400 300 200 100 0 02 03 045 40 35 30 25 20 123 HDL cholesterol PCB 126 BMI 0.80 0.85 0.90 0.95 51 PCB 126 ● 7.91 [ 0.42 , 34.98 ] Cd ● 3.23 [ −0.98 , 24.65 ] MMP ● −0.14 [ −8.58 , 6.05 ] PCB 209 ● 7.11 [ −0.06 , 30.31 ] MEHP ● 4.84 [ −6.90 , 25.92 ] PCB 170 ● 4.05 [ 0.74 , 21.08 ] MIBP ● −1.03 [ −16.04 , 5.66 ] TNC ● 0.40 [ −9.46 , 8.54 ] BPA ● −0.05 [ −6.12 , 8.62 ] PCB 153 ● 1.38 [ −1.01 , 11.02 ] p,p'−DDE ● −1.35 [ −22.34 , 3.02 ] Co ● 1.86 [ −3.35 , 24.19 ] PCB 169 ● 0.31 [ −8.21 , 9.11 ] Al ● 0.79 [ −1.75 , 20.91 ] HCB ● 0.18 [ −4.95 , 11.18 ] Hg ● 0.49 [ −4.39 , 13.94 ] OCDD ● −0.16 [ −8.82 , 4.85 ] PCB 118 ● −0.26 [ −11.86 , 3.95 ] Cr ● −0.29 [ −13.32 , 4.49 ] Ni ● 0.24 [ −9.13 , 8.21 ] Zn ● 1.10 [ −2.89 , 17.57 ] BDE47 ● −0.05 [ −15.67 , 3.36 ] Mn ● 0.04 [ −5.10 , 12.31 ] MEP ● 0.23 [ −5.22 , 9.10 ] Mo ● −0.16 [ −9.25 , 8.65 ] Pb ● 0.19 [ −5.40 , 13.74 ] Cu ● 0.43 [ −2.88 , 11.33 ]

−40.00 −20.00 0.00 20.00 40.00

Figure 4.15. Predicted effects on IMT for an IQR increase in the studied contaminants, with bootstrapped 95% conﬁdence intervals.

52 4.3.2 IM-GSM An ensemble consisting of 4,000 depth 10 CARTs achieved an R2 of 0.289. Using the one SE rule, an R2 of 0.287 was achieved using 3,800 depth 7 CARTs. The maximum achieved R2 using 11,400 depth 1 CARTs was 0.274 suggesting that interaction effects may be present. Figure 4.16 shows the relative inﬂuence of the studied variables on IM-GSM. The three phtalate metabolites MMP, MEHP and MIBP stand out as the most important predictors followed by four contaminants; PCB 126, p,p’-DDE, BPA and Ni, and two traditional risk factors for atherosclerosis; BMI and serum triglycerides, although the latter six variables have limited inﬂuence in predicting IM-GSM compared to the three phtalate metabolites.

Diabetes medication ● Sex ● Diabetes ● Antihypertensive medication ● Statin use ● Education ● Exercise ● LDL cholesterol ● Smoking status ● BDE47 ● Mn ● Zn ● PCB 118 ● Diastolic blood pressure ● Cu ● Co ● HDL cholesterol ● TNC ● Systolic blood pressure ● PCB 153 ● Hg ● Pb ● Cd ● PCB 170 ● Al ● Mo ● HCB ● MEP ● OCDD ● PCB 209 ● Cr ● PCB 169 ● Ni ● Triglycerides ● BPA ● p,p'−DDE ● BMI ● PCB 126 ● MIBP ● MEHP ● MMP ●

0 20406080100 Relative influence

Figure 4.16. Variable importance for the studied variables when predicting IM-GSM.

The top left panel of ﬁgure 4.17 shows the total interaction strength evaluated for the nine most important variables. The three phtalate metabolites MMP, MEHP and MIBP all seem to be involved in interactions as the observed values of H are outside their respective null distributions. The top right, bottom left and bottom right panels of ﬁgure 4.17 show two-way interactions with MMP, MEHP and MIBP respectively. MMP seems to interact with MIBP whereas MIBP seems to interact with both MMP and MEHP. There are also

53 indications that p,p’-DDE may be involved in interactions with both MMP and MIBP, although the importance of p,p’-DDE is small compared to the phthalates’.

0.00 0.05 0.10 0.15 Total interaction Two−way interaction with MMP

Ni ● Ni ●

● Triglycerides Triglycerides ● BPA ● BPA ● p,p'−DDE ● p,p'−DDE ● BMI ● BMI ● PCB 126 ● PCB 126 ● MIBP ●

● MEHP ● MIBP

MMP ● MEHP ●

Two−way interaction with MEHP Two−way interaction with MIBP

Ni ● Ni ●

Triglycerides ● Triglycerides ●

BPA ● BPA ●

p,p'−DDE ● p,p'−DDE ●

BMI ● BMI ●

PCB 126 ● PCB 126 ●

MIBP ● MEHP ●

MMP ● MMP ●

0.00 0.05 0.10 0.15

Figure 4.17. Interaction strengths for total interaction (lower left panel), two-way interactions with MMP (bottom right panel), two-way interactions with MEHP (top left panel) and two-way interactions with MIBP (top right panel). Black dots represent observed values of H while boxes represent the null distribution H0. Small tick marks indicate values outside the 5th and 95th percentiles.

Figure 4.18 shows the two-way interaction between MMP and MIBP. Keep- ing MIBP at low levels, IM-GSM increases with increasing MMP. Likewise, IM-GSM increases with increasing MIBP levels while keeping MMP constant at low levels. The increase is however more pronounced with a simultaneous increase in both MMP and MIBP suggesting a synergistic effect. The two-way interaction with MEHP and MIBP is shown in figure 4.19. IM-GSM decreases with increasing MEHP while keeping MIBP at low levels. The decrease is less pronounced at higher levels of MIBP suggesting a small antagonistic effect. Associations between IM-GSM and the six most important contaminants, not participating in interactions, are shown in figure 4.20. Predicted changes in IM-GSM for an IQR increase in the contaminants that did not participate in interactions are shown in figure 4.21.

54 2^10 95

2^8

2^6 MIBP 80

75 2^4

2^−22^−1 2^0 2^1 2^2 2^3 2^4 MMP

Figure 4.18. Predicted IM-GSM as a function of MMP (horizontal axis) and MIBP (vertical axis). Regions of the joint density of MMP and MIBP where there are insuf- ﬁcient data are white.

55 90 2^8

2^6 80 MIBP

2^4

2^0 2^2 2^4 2^6 2^8 MEHP

Figure 4.19. Predicted IM-GSM as a function of MEHP (horizontal axis) and MIBP (vertical axis). Regions of the joint density of MEHP and MIBP where there are insufﬁcient data are white.

56 ikmrsa h otmo ahpnlaedclso h variables. the of deciles are panel small each The of bottom intervals. the conﬁdence at 95% marks pointwise tick bootstrapped with interactions, in ing 4.20. Figure Predicted IM−GSM 70 75 80 70 75 80 012345 0 0 300 200 100 0 0010010020000 15000 10000 5000 0 ata eedniso h i otipratvrals o participat- not variables, important most six the on dependencies Partial Triglycerides p,p' PCB 126 − DDE 0 0 0 800 600 400 200 0 01 025 20 15 10 5 0 03 050 40 30 20 BPA BMI Ni 70 75 80 57 PCB 126 ● −2.42 [ −4.99 , −0.77 ] p,p'−DDE ● −1.94 [ −4.88 , −0.74 ] BPA ● 1.89 [ −0.73 , 3.49 ] Ni ● 1.14 [ −0.47 , 3.00 ] PCB 169 ● 0.74 [ 0.09 , 3.27 ] Cr ● −0.19 [ −1.68 , 1.42 ] PCB 209 ● −0.27 [ −2.45 , 0.21 ] OCDD ● 0.42 [ −0.71 , 2.21 ] MEP ● 0.29 [ −1.13 , 1.76 ] HCB ● 0.26 [ −1.16 , 1.78 ] Mo ● 0.65 [ −0.31 , 2.83 ] Al ● 0.51 [ −0.30 , 2.39 ] PCB 170 ● 0.42 [ −0.35 , 1.89 ] Cd ● 0.14 [ −1.23 , 1.23 ] Pb ● 0.63 [ 0.33 , 3.96 ] Hg ● −0.88 [ −3.97 , −0.77 ] PCB 153 ● 0.25 [ −0.38 , 1.92 ] TNC ● 0.01 [ −1.43 , 1.39 ] Co ● 0.49 [ −0.15 , 2.53 ] Cu ● 0.84 [ 0.46 , 3.92 ] PCB 118 ● −0.33 [ −2.58 , 0.34 ] Zn ● −0.42 [ −2.53 , 0.30 ] Mn ● −0.03 [ −1.46 , 1.14 ] BDE47 ● 0.57 [ −0.23 , 2.84 ]

−6.00 −2.00 0.00 2.00 4.00

Δ IM−GSM

Figure 4.21. Predicted effects on IM-GSM for an IQR increase in the variables that did not participate in interactions when predicting IM-GSM, with bootstrapped 95% conﬁdence intervals.

58 4.3.3 Discussion, paper III Why the combination of MIBP and MMP associate to IM-GSM in a multiplicative and not in an additive fashion is not known. Phthalates are known ligands to the nuclear PPAR receptors (Lovekamp-Swan et al., 2003; Hurst and Waxman, 2003), and intuitively one would expect the concentrations of these two PPAR ligands to add to each other regarding any vascular response. However, several PPAR receptor types are present and the relative activation of these different receptors has been shown to differ between different phthalates (Venkata et al., 2006). The relationship between MMP and MiBP and IM-GSM is in line with less lipid infiltration in the vascular wall. This fits well with the finding that agonists of the PPARγ receptor could have a bene- ficial effect on atherosclerosis when administered in humans (Sugawara et al., 2012). The antagonistic effects of MEHP and MIBP is however less evident. The parent compound of MEHP, DEHP, is also known as a PPAR receptor agonist (Lovekamp-Swan et al., 2003; Hurst and Waxman, 2003). However, during the recent years it has become evident that some phthalates also could interfere with other receptors, like the androgen receptors, estrogen receptors, the Ah-receptor and also to disrupt the thyroid hormonal system (Ernst et al., 2014; Shi et al., 2013; Wang et al., 2013; Wakui et al., 2013; Zhai et al., 2014). Thus, antagonistic effects of some of the phthalates might be explained by different bindings to receptors with divergent effects. Amongst the environmental contaminants with additive effects regarding IM-GSM, PCB 126 was one of the most powerful and acted in an inverse way vs IM-GSM. PCB 126 is the PCB measured in paper III with the highest dioxin-like activity and the highest TEQ (van den Berg et al., 1998). This is of particular interest since dioxin exposure in the atherosclerosis-prone ApoE knock-out mice resulted in enhanced atherosclerosis development (Dalton et al., 2001). Activation of the Ah-receptor by dioxin or dioxin-like PCBs induces formation of proinflammatory mediators (Dalton et al., 2001; Kraemer et al., 1996; Puga et al., 1997, 2004) as well as reactive oxygen species (Kopf et al., 2008; Park et al., 1996; Shertzer et al., 1998), two important parts in atherosclerosis formation. Also p,p’-DDE was related to IM-GSM in an additive and inverse fashion. DDT and its active metabolite p,p’-DDE are known androgen receptor agonists (Kelce et al., 1995) and it is well known that the androgen receptor plays a role in atherosclerosis formation (Bourghardt et al., 2010). There are however no experimental treatment studies using DDT so the causal role of this OC pesticide is still unknown.

4.4 Paper IV The predictive ability of the studied variables on the MetS was modest. A boostrap validated AUC of 0.684 was acheived using an ensemble consist-

59 ing of 3,000 depth 6 CARTs. Using the one SE rule, an AUC of 0.683 was acheived using 2,600 depth 4 CARTs. Using 8,000 depth 1 CARTs, an AUC of 0.677 was achieved indicating thatany interactions, if present, are of limited importance. Figure 4.22 displays the relative importance of the studied variables. p,p’-DDE was the most important predictor of prevalent MetS followed by PCBs 118 and 209, HCB, TNC and Hg.

Smoking status ● Education ● OCDD ● MIBP ● Al ● PCB 126 ● Female sex ● MEHP ● Ni ● Zn ● Mo ● MEP ● Co ● BDE47 ● Cd ● MMP ● PCB 170 ● PCB 153 ● Physical activity ● Mn ● PCB 169 ● Pb ● Cr ● Cu ● BPA ● Hg ● TNC ● HCB ● PCB 209 ● PCB 118 ● p,p'−DDE ●

0 20406080100

Figure 4.22. Relative importance for the studied variables when predicting prevalent MetS.

No evidence of interactions could be found (figure 4.23) as all observed values of H are within their respective null distributions. Figure 4.24 displays the partial dependencies with bootstraped 95% pointwise confidence intervals on each of the six most important variables. The prevalence of MetS increased with increased concentrations of p,p’-DDE, PCB 118, HCB and TNC while it decreased with increasing concentrations of PCB 209 and Hg. Figure 4.25 shows estimated Odds ratios (OR) and 95% bootstrapped confidence intervals (CI) for an IQR increase the studied contaminants in order of ther relative importance in predicting prevalent MetS.

60 Hg ●

TNC ●

HCB ●

PCB 209 ●

PCB 118 ●

p,p'−DDE ●

0.00 0.05 0.10

Figure 4.23. Total interaction strength for the six most important variables. Black dots represent observed values of H, and boxes represent the null distributions H0.

61 62 variable. each for intervals. deciles percentile are bootstrap panel 95% each are of intervals bottom conﬁdence the Pointwise at marks Tick MetS. prevalent 4.24. Figure Probability of prevalent MetS 0.1 0.2 0.3 0.4 0020 004000 3000 2000 1000 0 0000 20000 500010000 0 ata eedniso h i otipratvralswe predicting when variables important most six the on dependencies Partial p,p' HCB − DDE 0 0 0 800 600 400 200 0 0 001500 1000 500 0 PCB 118 TNC 1020304050600 010150 100 50 0 PCB 209 Hg 0.1 0.2 0.3 0.4 p,p'−DDE ● 1.59 [ 1.22 , 2.12 ] PCB 118 ● 1.31 [ 1.08 , 1.74 ] PCB 209 ● 0.69 [ 0.55 , 0.86 ] HCB ● 1.14 [ 1.01 , 1.43 ] TNC ● 1.14 [ 1.00 , 1.39 ] Hg ● 0.79 [ 0.63 , 0.94 ] BPA ● 1.16 [ 0.99 , 1.51 ] Cu ● 1.16 [ 1.01 , 1.44 ] Cr ● 1.03 [ 0.90 , 1.17 ] Pb ● 0.86 [ 0.71 , 1.00 ] PCB 169 ● 0.92 [ 0.76 , 1.05 ] Mn ● 0.98 [ 0.86 , 1.11 ] PCB 153 ● 1.03 [ 0.97 , 1.13 ] PCB 170 ● 0.88 [ 0.74 , 0.97 ] MMP ● 0.98 [ 0.87 , 1.08 ] Cd ● 1.08 [ 0.95 , 1.31 ] BDE47 ● 1.03 [ 0.93 , 1.18 ] Co ● 0.92 [ 0.77 , 1.06 ] MEP ● 0.98 [ 0.85 , 1.08 ] Mo ● 0.98 [ 0.87 , 1.05 ] Zn ● 1.09 [ 0.95 , 1.27 ] Ni ● 1.05 [ 0.95 , 1.20 ] MEHP ● 1.01 [ 0.92 , 1.12 ] PCB 126 ● 1.01 [ 0.90 , 1.11 ] Al ● 1.09 [ 0.97 , 1.29 ] MIBP ● 0.97 [ 0.87 , 1.05 ] OCDD ● 1.04 [ 0.94 , 1.19 ]

0.50 1.00 1.50 2.00 2.50

Figure 4.25. Estimated OR and 95% bootstrapped CI for an IQR increase in the studied contaminants.

63 4.4.1 Discussion, paper IV The relationships in paper IV between p,p’-DDE, HCB, TNC and prevalent MetS are in line with previous findings (Lee et al., 2007; Son et al., 2010; Al-Othman et al., 2014; Lee et al., 2014; Gauthier et al., 2014) where elevated levels have been associated to components of the MetS. Using data from the NHANES study, Lee et al. (2007) showed that PCBs were linked to the MetS and its components. In paper IV, the directions of the association vs the MetS were different for PCBs 118 and 209. This is in line with previous observations that higly chlorinated and less chlorinated PCBs are related to obesity and fat mass in divergent ways (Rönn et al., 2011; Dirinck et al., 2011). This is probably due to kinetic properties of the different PCBs, since the half-life of PCB 118 and PCB209 differ substantially with generally a longer half-life for PCBs with a high number of chlorine atoms. Since PCBs mainly are stored in blood, a negative association versus fat mass will be observed during the first years of exposure, but with time this negative association is converted to a positive association due to the elimination of the compounds from the body. However, the timing in shift in the direction of association is largely dependent on the half-life and thereby a positive relation could well be detected in obese people, the hallmark of the MetS, for PCBs with a relative short half-life while on the same time a negative association could be found to PCBs with a longer half-life, as described by Wolff et al. (2007). Uemura et al. (2009) showed that several of the PCB congeners, including PCB 118, were related to the MetS, as in paper IV. However, PCB 209 was not measured in that study or in the NHANES study, so no direct comparisons regarding that particular contaminant could be performed. Increased levels of Hg, Cu and Pb have previously been related to some of the cardiovascular risk factors included in the MetS (Olsén et al., 2012a; Rhee et al., 2013; Chung et al., 2014). The inverse relationship between Hg and prevalent MetS in paper IV differs in direction to what was reported by (Chung et al., 2014) and (Rhee et al., 2013). Hg levels were far lower in the PIVUS study and it is possible that increased values of Hg in the range found here is indicative of a diet rich in fish which has been shown to be protective with regards to the MetS (Panagiotakos et al., 2007; Zaribaf et al., 2014).

64 5. Summary

• There was little co-variation among the studied contaminants. Two clusters of PCBs could be discerned with the number of chlorine atoms as the deﬁning characteristic.

• Gradient boosted CARTs can be a useful tool to search for interactions in epidemiological data. The power to detect interactions that are not small compared to the main effects was adequate for a sample size of 1,000.

• The studied contaminants showed little association with IMT but interactions among three phthalate metabolites were found regarding IM- GSM.

• Additive effects for some contaminants could be seen regarding the MetS.

65 6. General Discussion

The general conclusion in this thesis is that both additive and multiplicative associations between circulating levels of environmental contaminants and various health endpoints exist. A natural question is to ask whether these associations are causal, i.e. that the contaminants actually cause the endpoints, or if they are merely associated. There are four principal reasons for associations in epidemiological studies; bias, confounding, chance and cause (Jepsen et al., 2004). Bias denotes deviation from the truth and all observational studies suffer from it (Grimes and Schulz, 2002). Selection bias refers to the bias arising when the sample in which the analyses are made is not representative of the population for which statements are made. The individuals invited to participate in the PIVUS study were randomly selected from a well-defined population and of the 2,025 invited, 1,016 individuals participated. Lind et al. (2006) studied the cardiovascular disorders and medications in 100 non-participants. Cardiovascular medication use did not differ between participants and non- participants but diabetes, angina pectoris, congestive heart failure and stroke were more prevalent in the non-respondents than in the respondents. The sample used in papers I, III and IV could thus be slightly healthier, in the cardiovascular sense, than the population which the sample was drawn from. This is unlikely to have any major impact on the generalizability of the results. That being said, the results are only generalizable to the elderly population in Uppsala county and need to be confirmed in other populations. Information bias and measurement error are two other sources of bias which relate to the recording of individual factors (Hammer et al., 2009). In PIVUS, physical activity and smoking status were self-reported. Garriguet and Col- ley (2014) compared self-reported physical activity to measurements made by accelerometers in a Canadian population of various ages. Self-reported and measured physical activity were weakly correlated with elderly tending to overestimate their physical activity. If this is true for Swedish elderly as well is not known. The impact of misclassified physical activity is likely to be low as it had modest influence on the studied endpoints. Smokers are generally less likely to respond to questionnaires about smoking resulting in missing values or misclassifying current smokers as non-smokers. The fraction of missing values for smoking status in PIVUS was low and smoking was of minor importance for the results. Potential misclassification of smoking status is likely of no concern for these results. No adjustment for measurement errors was made in the analyses. Measure- ment error can be due e.g. to poorly callibrated instruments, the measurement

66 procedure or human error. If the error is random, results will typically be biased towards the null. A systematic measurement error can bias the result either way depending on the direction of the error. It is not reasonable to assume random measurement errors for the contaminants measured in PIVUS. Contaminant concentrations near the limit of detection are generally very low and the actual measurements are more prone to influences from the laboratory environment or instrument condition than measurements at higher contaminant concentrations. All chemical analyses, except the analysis for Al, were carried out by accredited laboratories using state of the art instruments so it is not unreasonable to assume that the effect of measurement errors on the associations seen in papers III and IV is small. Confounding is the blurring of effects where a variable is associated with the exposure and the outcome but is not in the causal pathway between the exposure and the outcome. Common methods to deal with confounding in epidemiological studies include stratification, in which data are analyzed sep- arately within defined groups of the confounding variable(s) and the results are weighted together, matching, in which controls and cases sharing some characteristics are matched together, and regression adjustment, where confounding variables are included in a regression model. If environmental contaminants cause altered cholesterol and/or triglyceride levels which in turn cause atherosclerosis, then cholesterol and triglyceride levels are not considered confounders and should not be adjusted for in a regression model if causal inferences are to be made about contaminant exposure. Boosted CARTs are made for predicting the outcome rather than generating inferences about it, and including known predictors of the outcome can help gauge the contribu- tion of the contaminants relative to the known predictors. Predictions from a regression model is often used to estimate effects from non-linear relationships (Harrell Jr., 2001) and while no formal inferences can be made from a boosted CART model, it can be used to generate hypotheses that can be formally tested in other studies and/or experiments. Confounding from so- cioeconomic status, occupational history, diet and other unmeasured lifestyle factors cannot be ruled out as they were not included in the analysis due to lack of data. The ability of boosted CARTs to uncover complex interactions in simulated data was demonstrated in paper II. Although applied on a variety of different prediction problems (Hastie et al., 2009), boosted CARTs have rarely been used to evaluate experimental data. Seilkop et al. (2012) used Multivari- ate Adaptive Regression Trees (MART), which is another name for gradient boosted CARTs, to study the pro-atherosclerotic response in mice from exposure to contaminants found in combustion emissions. They found no substan- tial departures from additivity upon examining pairwise partial dependencies but acknowledged that this may be due to the limited sample size. Greco et al. (1995, pg. 350) provides simulated experimental data with two variables called D1 and D2 and a small amount of synergism between the two. Full de-

67 tails of the data generation are given in Greco et al. (1995). Figure 6.1 shows the bootstrap validated RMSE as a function of the number of CARTs (M) and CART depth (d) when predicting the square root of the outcome. The subsampling fraction η was set to 0.8 due to the small number of observations. An ensemble consisting of 20,000 depth d = 2 CARTs predicted the response better than any ensemble consisting of depth d = 1 CARTs suggesting a potential interaction between D1 and D2.

Max Tree Depth 12● ●

2.2 ● ●

2.0

1.8 ● ●

RMSE (Bootstrap) 1.6 ●

●

● ● 1.4 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

5000 10000 15000 20000 # Boosting Iterations

Figure 6.1. Bootstrap validated RMSE versus number of CARTs and CART depth

Figure 6.2 shows the observed interaction strengths along with their null distributions. The observed values of H are well outside their respective null distributions meaning that both D1 and D2 were involved in interactions.

Figure 6.3 displays the joint effect of D1 and D2 on the simulated response. While this does not prove the applicability of boosted CARTs to study mixture effects, the fact that the method uncovered the interaction effect simulated using the CA concept is encouraging.

68 D2 ●

D1 ●

0.00 0.02 0.04 0.06 0.08

Figure 6.2. Interaction strengths for D1 and D2.

69 5 80

70 4

3 50 2 D 40

2 30

20 1

0 0

0 1020304050

Figure 6.3. Predicted joint effect of D1 and D2 on the simulated response.

70 The exposure must precede the outcome if a causal statement is to be made about the observed associations. Exposure to environmental contaminants is likely to be continuous and co-occur with the development of atherosclerosis and the different components of the MetS. Papers III and IV used cross- sectional data in which the exposure and outcome are measured at a single time-point. How representative a single measurement is for long-term exposure is an important question. Townsend et al. (2013) studied the within- person reproducibility of single BPA and eight phthalate metabolites measured in urine. The conclusion was that a single measurement of BPA was not stable over a longer time period. Seven phthalate metabolites showed reasonable stability over time. These results were based on two cohorts of nurses in the USA and whether the same applies to an elderly population in Sweden and contaminants measured in the circulation is unknown. POPs are persistent and widespread in the environment and a single measurement should better reﬂect long-term exposure than the plastic associated chemicals. Despite the short half-lives of the plastic-associated chemicals, detectable levels were found in almost all individuals in PIVUS. The results in this thesis should thus be considered exploratory and hypotheses generating until conﬁrmed in other studies and/or experiments.

Based on the results in this thesis, future studies need to measure many different contaminants to get a clearer picture of the exposure. There is a need for repeated measurements to assess longitudinal effects. Boosted CARTs are promising but need to be evaluated further, especially methods for internal validation of discovered interactions. Experiments should complement epidemiological studies where applicable.

71 7. Svensk sammanfattning: Blandningseffekter av Miljögifter

Människor exponeras varje dag för ett stort antal kemikalier. Den kunskap om kemikaliers toxiska effekt som finns idag, och som riskbedömningar baseras på, grundar sig i de allra flesta fall på studier av en kemikalie i taget. Bland- ningseffekter är välkända när det gäller läkemedel men kunskapen om eventuella blandningseffekter av kemikaler, som betraktas som miljögifter, i vår vardag är liten. Experimentellt har man visat hur kemikalier som var för sig inte ger någon stor effekt kan samverka när de blandas samman och ge en effekt som är större än om deras enskilda effekter skulle summeras. På grund av ett mycket stort antal olika kombinationer av kemikalier i vår omgivning så har existerar mycket få epidemiologiska studier hos människa. Att studera blandningseffekter i större omfattning med traditionella statistiska metoder skulle kräva enorma mängder data. Moderna metoder för datautvinning har nyligen fått uppmärk- samhet i litteraturen som möjliga metoder att tillämpa på detta problem. Ett regressionsträd är en samling logiska regler för prediktion av ett ut- fall. Samverkan mellan olika variabler hanteras naturligt vilket gör metoden lämplig för att studera blandningseffekter. Enkla regressionsträd har dock dålig prediktiv förmåga. Stokastisk gradientboosting är en metod som kom- binerar många enkla regressionsträd till en ensamble med överlägsen prediktiv förmåga jämfört med enkla träd. Nackdelen jämfört med traditionella statistiska metoder är att metoden inte lämpar sig för hypotesprövning eller infer- ens, utan är ett prediktionsverktyg. Åderförkalkning är en progressiv sjukdom i artärväggarna där LDL-kolesterol oxideras och tas upp av makrofager som bildar skumceller. Denna process ger upphov till en kronisk inflammation i vilken ett plack byggs upp. Placktillväx- ten leder till minskad blodgenomströmning och kan, i de fall ett plack spricker, leda till total ocklusion av artären och efterföljande ischemi i vävnaden, till ex- empel hjärtinfarkt. Det metabola syndromet är en ansamling riskfaktorer för kardiovaskulär sjukdom hos en och samma individ. Syndromet definieras som förekomst av minst tre av följande komponenter: högt blodtryck, höga nivåer av triglyc- erider, låga halter av HDL-kolesterol, bukfetma och förhöjt fasteblodsocker. Syndromet har inte mer prediktivt värde än sina ingående beståndsdelar, men är en användbar beskrivande term för individer med multipla riskfaktorer. Denna avhandling syftar till att undersöka blandningseffekter av 37 miljögifter uppmätta i blod hos 1016 70-åringar i Uppsala län, ingående i PIVUS-kohorten,

72 i relation till åderförkalkning samt det metabola syndromet. Miljögifterna representerade tre klasser; persistenta organiska miljögifter, plast-associerade kemikalier och metaller.

7.1 Studie I Studie I syftade till att undersöka hur de 37 uppmätta miljögifterna samvarierade. En liten grad av samvariation innebär att många olika miljögifter måste mätas för att få en klar bild av exponeringen. Hierarkisk klusteranalys användes för att hitta grupper av samvarierande miljögifter och stabiliteten hos de funna klustren undersöktes med resampling- metoder, där ﬂertalet stickprov dras från data och grupperingen i dessa stickprov jämförs med den ursprungliga grupperingen. Inom varje stabilt kluster utfördes principalkomponentanalys för att undersöka om någon eller några miljögifter kunde användas som markörer för klustret. Generellt var samvariationen liten, men två kluster av PCBer kunde urskiljas. Den gemensamma nämnaren för klustren var antalet kloratomer på PCBerna. PCBer med färre än sex kloratomer bildade ett kluster medan PCBer med ﬂer än sex kloratomer bildade ett kluster. PCBer med sex kloratomer åter- fanns i båda klustren, där plana PCBer samvarierade med de PCBer med lä- gre kloreringsgrad och ej plana PCBer samvarierade med PCBer med högre kloreringsgrad. Vid sidan av PCB 126 och 169 kunde PCB 118 och 153 användas som markörer för de PCBer med lägre kloreringsgrad medan PCB 170 och 209 kunde användas som markörer för PCBer med högre kloreringsgrad.

7.2 Studie II I studie II studerates tillämpbarheten hos boostade regressionsträd i simulerade data. Data simulerades utifrån data som användes i studie I. Korrelationer simulerades mellan PCBerna och utfallet innehöll en samverkan mellan fyra miljögifter, ett icke-linjärt beroende samt en samverkan mellan ett miljögift och en binär variabel. Fyra olika styrkor på sambanden simulerades. De variabler som var involverade i samverkanseffekter identifierades kor- rekt i de tre fall där sambanden var starkast. Endast då bruset var 10 gånger starkare än den simulerade signalen lyckades inte de boostade regressionsträ- den identifiera de korrekta variablerna. Några falska samverkanseffekter identifierades också. Förmågan att identifiera samverkande variabler för en stickprovsstorlek på 1000 individer bedömdes som förhållandevis god givet att samverkanseffek- terna inte är för små.

73 7.3 Studie III Studie III var en tvärsnittsstudie där sambanden mellan ett ﬂertal miljögifter och åderförkalkning undersöktes. Karotisartärerna undersöktes med ultraljud för att bedöma graden av åderförkalkning. Två mått användes för åderförkalkn- ing; tjockleken på intima-media-komplexet (IMT) och intima-media-komplexets ekogenomsläpplighet (IM-GSM). En hög ekogenomsläpplighet (låg IM-GSM) har tidigare visats vara en prediktor för framtida kardiovaskulär död. Resul- tatet från studie I utnyttjades och av de 37 miljögifterna ingick 27 stycken i analysen tillsammans med övriga traditionella riskfaktorer; kolesterol, blod- fetter, blodtryck, rökning, kön, utbildning, fysisk aktivitet och BMI. Systoliskt blodtryck var den variabel som var starkast associerad till IMT. Sambanden mellan IMT och de studerade miljögifterna var svaga. Alla in- gående variabler förklarade tillsammans ca. 6% av variationen i IMT. Tre ftalatmetaboliter, MMP, MIBP och MEHP var starkt associerade till IM-GSM. En synergistisk samverkan mellan MMP och MIBP samt en svag antagonistisk samverkan mellan MIBP och MEHP kunde urskiljas. Tillsam- mans förklarade alla ingående variabler 29% av variationen i IM-GSM.

7.4 Studie IV I studie IV undersöktes tvärsnittssambanden mellan miljögifterna och förekom- sten av det metabola syndromet. I analysen inkluderades de 27 miljögifterna samt utbildning, kön och fysisk aktivitet. Prevalensen av det metabola syndromet var 23%. Sambanden var överlag blygsamma, men tre pesticider; p,p’- DDE, hexaklorbensen och trans-nonachlor visade på positiva samband med prevalent metabolt syndrom. PCB 118 visade även den ett positivt samband medan PCB 209 och kvicksilver visade inversa samband, dvs ökade halter av de båda var associerade med en minskad prevalens av det metabola syndromet.

7.5 Sammanfattning av Studierna Miljögifter mätta i blod samvarierade inte i någon hög grad. Detta innebär att man behöver mäta ett stort antal miljögifter för att få en klar bild av exponeringen. Boostade regressionsträd är ett verktyg för att studera blandningseffekter i epidemiologiska studier och kan användas för att generera hypoteser som sedan kan testas i experiment eller andra studier. Ftalatmetaboliterna MMP och MIBP samverkade synergistiskt och MIBP och MEHP antago- nistiskt med avseende på IM-GSM. Pesticiderna p,p’-DDE, hexaklorbensen och trans-nonachlor samt PCB 118 visade positiva samband med prevalent metabolt syndrom medan PCB 209 och kvicksilver visade inversa samband. Resultaten skall ses som explorativa och hypotesgenererande tills dess de är bekräftade i experiment eller i oberoende data.

74 8. Acknowledgement

Många har bidragit till att arbetet med denna avhandling gått, kanske inte alltid på en rak räls, men framåt och i mål. Speciellt vill jag tacka

Mina tre handledare; Monica Lind, Lars Lind och Anna Bornefalk-Her- mansson. Ingen kan undvika att ryckas med av Monicas passion för detta ämne, så även jag till slut. Tack Lars för din förmåga att se vägar framåt när planerna för arbetet hastigt behöver ändras. Tack Anna för att ha hållit mig på mattan när jag svävat iväg med alla galna idéer.

Lenita Öqvist, Jennie Lindström, Britt-Marie Löfgren och Nina Lifven- dahl för att det är ni som får verksamheten att fungera och har koll på allt.

Nuvarande och gamla kollegor på Arbets- och miljömedicin i Uppsala, ingen nämnd, ingen glömd.

Samira Salihovic och Bert van Bavel. Utan er hade det inte funnits data att räkna på. Tack Samira för alla kemilektioner.

Doktorandkollegorna Monika Rönn, Magnus Helgesson och Erika Ax. PIVUS post-docs Johanna Penell och Margareta Halin Lejonklou.

Gänget på UCR, framför allt Lars Berglund och Bodil Svennblad för att jag fått låna ett skrivbord under denna tid. Johan Lindbäck som stod ut med att ha mig sittande i rummet och att gång på gång ha bevisat att R kan göra det bättre.

Birgitta Sembrant och Anette McLoughlin för att ha guidat mig rätt i den administrativa djungeln.

Professor Chisato Mori and Dr. Emiko Todaka, Chiba University, Japan for the nice collaboration and for treating my wife and I to a wonderful evening in Geneva.

Familjen Larsson, den äldre; Birgit och Lars Göran för att ni är de bästa svärföräldrar man kan ha.

75 Familjen Larsson, den yngre; Magnus, Caroline, Hampus, Saga och Al- ice. Hur jag och Magnus kunde spela ihop Londonresan är fortfarande ett mysterium.

Mina föräldrar Kerstin och Svante. Tack för att ni stöttat och trott på mig oavsett vad jag hittat på. Lillasyster Anna och svåger Hans.

De viktigaste personerna i mitt liv; Axel och Anton för att ni är de ﬁnaste som ﬁnns och Annica, min bästa vän. Tack för den energi och glädje du ger mig och för att jag får dela livet med dig. Jag älskar er!

76 References

A A Al-Othman, S H Abd-Alrahman, and N M Al-Daghri. DDT and its metabolites are linked to increased risk of type 2 diabetes among Saudi adults: a cross-sectional study. Environ Sci Pollut Res, 2014. K G Alberti and P Z Zimmet. Definition, diagnosis and classification of diabetes mellitus and its complications. part 1: diagnosis and classification of diabetes mellitus provisional report of a WHO consultation. Diabet Med, 15:539–53, 1998. R Altenburger, T Backhaus, W Boedeker, M Faust, M Scholze, and L H Grimme. Predictability of the toxicity of multiple chemical mixtures to Vibrio fischeri: mixtures composed of similarly acting chemicals. Environ Toxicol Chem, 19: 2341–2347, 2000. R Altenburger, H Walter, and M Grote. What contributes to the combined effect of a complex mixture? Environ Sci Technol, 38:6353–62, 2005. P K Andersen and A Skrondal. Biological Interaction from a statistical point of view. Technical Report 10/1, Department of Biostatistics, University of Copenhagen, 2010. T Backhaus, M Faust, M Scholze, P Gramatica, M Vighi, and L H Grimme. Joint algal toxicity of phenylurea herbicides is equally predictable by concentration addition and independent action. Environ Toxicol Chem, 23:258–64, 2004. M L Bell andDLDavis. Reassessment of the lethal London fog of 1952: novel indicators of acute and chronic consequences of acute exposure to air pollution. Environ Health Perspect, 109(Suppl 3):389–394, 2001. P A Bertazzi, I Bernucci, B Brambilla, D Consonni, and A C Pesatori. The Seveso studies on early and long-term effects of dixin exposure: a review. Environ Health Perspect, 106(Suppl 2):625–633, 1998. C Billionnet, D Sherrill, and I Annesi-Maesano. Estimating the Health Effects of Exposure to Multi-Pollutant Mixture. Ann Epidemiol, 22:126 – 141, 2012. C I Bliss. The toxicity of poissons applied jointly. Ann J Appl Biol, 26:585–615, 1939. J Bourghardt,ASWilhelmson, C Alexanderson, K De Gendt, G Verhoeven, A Krettek, C Ohlsson, and Å Tivesten. Androgen receptor-dependent and independent atheroprotection by testosterone in male mice. Endocrinology, 151: 5428–37, 2010. H Bouwman, H van den Berg, and H Kylin. DDT and malaria prevention: addressing the paradox. Environ Helath Perspect, 119:744–7, 2011. L Breiman, J Friedman, R Olshen, and C Stone. Classification and Regression Trees. Wadsworth, 1984. A M Calafat, Z Kuklenyik, J A Reidy, S P Caudill, J. Ekong, and L L Needham. Urinary concentrations of Bisphenol A and 4-nonylphenol in a human reference population. Environmental Health Perspectives, 113(4):391–395, 2005.

77 D O Carpenter, K Arcaro, and D C Spink. Understanding the Human Health Effects of Chemical Mixtures. Environmental Health Perspectives, 110(Supplement 1): 25–42, February 2002. R Carson. Silent Spring. Houghton Mifﬂin Co, 1962. Centers for Disease Control and Prevention. National Health and Nutrition Examination Surveys (NHANES), Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey Data. Hyattsville, MD: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention. http://www.cdc.gov/nchs/nhanes/nhanes2003-2004/nhanes03_04.htm, 2003. Centers for Disease Control and Prevention. CDC - NDP - Factsheet - Phthalates, 2013. URL http://www.cdc.gov/biomonitoring/phthalates_factsheet.html. Accessed November 24, 2014. J Y Chung, M S Seo, J Y Shim, and Y J Lee. Sex differences in the relationship between blood mercury concentration and metabolic syndrome risk. J Endocrinol Invest., pages 1–7, 2014. B Claus Henn, L Schnaas, A S Ettinger, J Schwartz, H Lamadrid-Figueroa, M Hernández-Avila, C Amarasiriwardena, H Hu, D C Bellinger, and R O Wright. Associations of early childhood manganese and lead coexposure with neurodevelopment. Environ Health Persp, 120:126 – 136, 2012. T P Dalton,JKKerzee, B Wang, M Miller, M Z Dieter, J N Lorenz, H G Shertzer, D W Nerbert, and Puga A. Dioxin exposure is an environmental risk factor for ischemic heart disease. Cardiovasc Toxicol, 1:285–298, 2001. E Dirinck, P G Jorens, A Covaci, T Geens, L Roosens, H Neels, I Mertens, and L van Gaal. Obesity and persistent organic pollutants: possible obesogenic effect of organochlorine pesticides and polychlorinated biphenyls. Obesity (Silver Spring), 19:709–14, 2011. B Efron andRJTibshirani. An Introducion to the Bootstrap. Monographs on Statistics and Probability. Chapman & Hall/CRC, 1994. J Elith, J R Leathwick, and T Hastie. A working guide to boosted regression trees. J Anim Ecol, 77:802 – 813, 2008. J Ernst, J C Jann, R Biemann,HMKoch, and B Fischer. Effects of the environmental contaminants DEHP and TCDD on estradiol synthesis and aryl hydrocarbon receptor and peroxisome proliferator-activated receptor signalling in the human granulosa cell line KGN. Mol Hum Reprod, 20:919–28, 2014. M Faust, R Altenburger, T Backhaus, H Blanck, W Boedeker, P Gramatica, V Hamer, P Gramatica, M Scholze, M Vighi, and Grimme L H. Predicting the joint algal toxicity of multi-component s-triazine mixtures at low-effect concentrations of individual toxicants. Aquatic Toxicology, 56, 2001. M Faust, R Altenburger, T Backhaus, H Blanck, W Boedeker, P Gramatica, V Hamer, M Scholze, M Vighi, and L H Grimme. Joint algal toxicity of 16 dissimilarly acting chemicals is predictable by the concept of independent action. Aquat Toxicol, 63:43–63, 2003. A-M Florea, D Büsselberg, and D Carpenter. Metals and Disease. J Toxicol, 2012: 1–2, 2012.

78 J Friedman, T Hastie, and R Tibshirani. Additive Logistic Regression: A Statistical View of Boosting. Ann Stat, 28:337 – 407, 2000. J H Friedman. Greedy Function Approximation: A Gradient Boosting Machine. Ann Stat, 29:1189 – 1232, 2001. J H Friedman. Stochastic gradient boosting. Comput Stat Data An, 38:367 – 378, 2002. J H Friedman and B E Popescu. Predictictive learning via rule ensembles. The Annals of Applied Statistics, 2(3):916–954, 2008. T E Froelich, B P Lanphear, P Auinger, R Hornung, J R Epstein, J Braun, and R S Kahn. Association of tobacco and lead exposures with attention-deficit/hyperactivity disorder. Pediatrics, 124:e1054 – e1063, 2009. D Garriguet and R C Colley. A comparison of self-reported leisure-time physical activity and measured moderate-to-vigorous physical activity in adolescents and adults. Health Rep, 16:3–11, 2014. M S Gauthier, R Rabasa-Lhoret, D Prud’homme, A D Karelis, D Geng, B van Bavel, and Ruzzin J. The Metabolically Healthy But Obese Phenotype Is Associated With Lower Plasma Levels of Persistent Organic Pollutants as Compared to the Metabolically Abnormal Obese Phenotype. J Clin Endocrinol Metab, 99: E1061–E1066, 2014. C Gennings, W H Carter, Jr, R A Carchmann,LKTeuschler, J E Simmons, and E W Carney. A unifying concept for assessing toxicological interactions: changes in slope. Toxicol Sci, 88:287–97, 2005. C Gennings, R Sabo, and E Carney. Identifying subsets of complex mixtures most associated with complex diseases: polychlorinated biphenyls and endometriosis as a case study. Epidemiology, 21 Suppl 4:S77–84, 2010. W R Greco, G Bravo, andJCParsons. The Search for Synergy: a Critical Review from a Response Surface Perspective. Pharmacol Rev, 47:331 – 385, 1995. D A Grimes and K F Schulz. Bias and causal associations in observational research. The Lancet, 359:248–52, 2002. G P Hammer,JBduPrel, and Blettner M. Avoiding bias in observational studies: part 8 in a series of articles on evaluation of scientific publications. Dtsch Arztebl Int, 106:664–8, 2009. F E Harrell Jr. Regression Modeling Strategies. With Applications to Linear Models, Logistic Regression and Survival Analysis. Springer, second edition, 2001. T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Data Mining, Inference and Prediction. Springer, second edition, 2009. W H Helfand, J Lazarus, and P Theerman. Donora, Pennsylvania: an environmental disaster of the 20th century. Am J Public Health, 91:331, 2001. B Hennig, P Meerarani, R Slim, M Toborek, A Daugherty, A E Silverstone, and L W Robertson. Proinflammatory properties of coplanar PCBs: in vitro and in vivo evidence. Toxicol Appl Pharmacol, 15, 2002. A H Herring. Nonparametric bayes shrinkage for assessing exposures to mixtures subject to limits of detection. Epidemiology, 21 Suppl 4:S71–6, 2010. M Hollander andDAWolfe. Nonparametric statistical methods. John Wiley & Sons, New York, 1999.

79 S Hong, J-P Candelone,CCPatterson, and C F Boutron. History of Ancient Copper Smelting Pollution During Roman and Medieval Times Recorded in Greenland Ice. Science, 272, 1996. GJHowardandTFWebster. Generalized concentration addition: A method for examining mixtures containing partial agonists. J Theor Biol, 259:469–477, 2009. GJHowardandTFWebster. Contrasting Theories of Interaction in Epidemiology and Toxicology. Environ Health Perspect, 121:1–6, 2013. C H Hurst andDJWaxman. Activation of PPARα and PPARγ by environmental phthalate monoesters. Toxicol Sci, 74:297–308, 2003. P Jepsen, S P Johnsen, M W Gillman, and H T Sørensen. Interpretation of observational studies. Heart, 90:956–960, 2004. M Junghans, T Backhaus, M Faust, M Scholze, and L H Grimme. Application and validation of approaches for the predictive hazard assessment of realistic pesticide mixtures. Aquat Toxicol, 76:93–110, 2006. L Järup. Hazards of heavy metal contamination. Br Med Bull, 68:167–82, 2003. W R Kelce, C R Stone,SCLaws,LEGray,JAKemppainen, andEMWilson. Persistent DDT metabolite p,p’-DDE is a potent androgen receptor antagonist. Nature, 375:581–5, 1995. K E Kelley, S Hernández-Díaz, E L Chaplin, R Hauser, and A A Mitchell. Identiﬁcation of phthalates in medications and dietary supplement formulations in the United States and Canada. Environ Health Perspect, 120:379–84, 2012. JHKim,HYPark, S Bae, Y H Lim, and Y C Hong. Diethylhexyl phthalates is associated with insulin resistance via oxidative stress in the elderly: a panel study. PLoS One, 8:e71392, 2013. A G Kirkley and R M Sargis. Environmental endocrine disruption of energy metabolism and cardiovascular risk. Curr Diab Rep, 14:494, 2014. P G Kopf, J K Huwe, and Walker M K. Hypertension, cardiac hypertrophy, and impaired vascular relaxation induced by 2,3,7,8-tetrachlorodibenzo-p-dioxin are associated with increased superoxide. Cardiovasc Toxicol, 8:181–93, 2008. A Kortenkamp. Low dose mixture effects of endocrine disrupters and their implications for regulatory thresholds in chemical risk assessment. Curr Opin Pharmacol, 19:105–111, 2014. A Kortenkamp, T Backhaus, and M Faust. State of the art report on Mixture Toxicity. http://ec.europa.eu/environment/chemicals/pdf/report_ Mixture%20toxicity.pdf, 2009. S A Kraemer, K A Arthur, M S Denison, W L Smith, and DeWitt D L. Regulation of prostaglandin endoperoxide H synthase-2 expression by 2,3,7,8,-tetrachlorodibenzo-p-dioxin. Arch Biochem Biophys, 330:319–28, 1996. E Lampa, L Lind, A Bornefalk-Hermansson, S Salihovic, B van Bavel, and P M Lind. An investigation of the co-variation in circulating levels of a large number of environmental contaminants. Journal of Exposure Science and Environmental Epidemiology, 22(5):476–482, 2012. E Lampa, A Bornefalk-Hermansson, L Lind, and P M Lind. Mixture Effects of Multiple Environmental Contaminants on the Metabolic Syndrome in a Human Population-based Sample. Manuscript, 2014.

80 E Lampa, A Bornefalk-Hermansson, P M Lind, and L Lind. Atherosclerosis in Humans and the Association to Environmental Contaminant Mixtures. Manuscript, 2014. E Lampa, L Lind, P M Lind, and A Bornefalk-Hermansson. The identiﬁcation of complex interactions in epidemiology and toxicology: a simulation study of boosted regression trees. Environ Health, 13(57), 2014. D H Lee, I K Lee, M Porta, M Steffes, and D R Jacobs, Jr. Relationship between serum concentrations of persistent organic pollutants and the prevalence of metabolic syndrome among non-diabetic adults: results from the National Health and Nutrition Examination Survey 1999-2002. Diabetologia, 50:1841–1851, 2007. Y M Lee, K S Kim, S O Kim, N S Hong, S J Lee, and D H Lee. Prospective associations between persistent organic pollutants and metabolic syndrome: A nested case-control study. Sci Total Environ, 496:219–225, 2014. L Lind, S Jakobsson, H Lithell, B Wengle, and S Ljunghall. Relation of serum calcium concentration to metabolic risk factors for cardiovascular disease. BMJ, 297:960–3, 1988. L Lind, N Fors, J Hall, K Marttala, and A Stenborg. A comparison of three different methods to determine arterial compliance in the elderly: the Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS) study. J Hypertens, 24:1075–82, 2006. P M Lind and L Lind. Circulating levels of bisphenol A and phthalates are related to carotid atherosclerosis in the elderly. Atherosclerosis, 218:207–13, 2011. P M Lind, L Olsén, and L Lind. Circulating levels of metals are related to carotid atherosclerosis in elderly. Sci Total Environ, 416:80–8, 2012. P M Lind, B van Bavel, S Salihovic, and L Lind. Circulating levels of persistent organic pollutants (POPs) and carotid atherosclerosis in the elderly. Environ Health Perspect, 120:38–43, 2012. P M Lind, U Risérus, S Salihovic, B van Bavel, and L Lind. An environmental wide association study (EWAS) approach to the metabolic syndrome. Environment Int, 55:1–8, 2013. S Loewe and H Muischneck. Effect of combinations: mathematical basis of problem. Arch Exp Pathol Pharmakol, 114:313–326, 1926. T Lovekamp-Swan, A M Jetten, and Davis B J. Dual activation of PPARα and PPARγ by mono-(2-ethylhexyl) phthalate in rat ovarian granulosa cells. Mol Cell Endocrinol, 201:133–41, 2003. Z Majkova, E Smart, M Toborek, and B Hennig. Up-regulation of endothelial monocyte chemoattractant protein-1 by coplanar PCB77 is caveolin-1-dependent. Toxicol Appl Pharmacol, 15:1–7, 2009. P Mirmira and C Evans-Molina. Bisphenol A, obesity, and type 2 diabetes mellitus: genuine concern or unnecessary preoccupation? Translational research, 164:13 – 21, 2014. National Cholesterol Education Program. Third report of the National Cholesterol Education Program (NCEP) expert panel on detection, evaluation, and treatment of high blood cholesterol in adults (Adult Treatment Panel III). Final report. Circulation, 106:3143–3421, 2002.

81 Nobel Media AB. The Nobel Prize in Physiology or Medicine 1948, 2014. URL http: //www.nobelprize.org/nobel_prizes/medicine/laureates/1948/. Accessed November 27, 2014. L Olsén, E Lampa, D A Birkholz, L Lind, and P M Lind. Circulating levels of bisphenol A (BPA) and phthalates in an elderly population in Sweden, based on the Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS). Ecotoxicol Environ Saf, 75:242–248, 2011. L Olsén, P M Lind, and L Lind. Gender differences for associations between circulating levels of metals and coronary risk in the elderly. International journal of hygiene and environmental health, 215:411–417, 2012. L Olsén, P M Lind, and L Lind. Gender differences for associations between circulating levels of metals and coronary risk in the elderly. Int J Hyg Environ Health, 215:411–417, 2012. D B Panagiotakos, C Pitsavos, Y Skoumas, and C Stefanadis. The association between food patterns and the metabolic syndrome using principal components analysis: The ATTICA study. J Am Diet Assoc., 107:979–87, 2007. J Y Park, M K Shigenaga, and B N Ames. Induction of cytochrome P4501A1 by 2,3,7,8-tetrachlorodibenzo-p-dioxin or indolo(3,2-b)carbazole is associated with oxidative DNA damage. Proc Natl Acad Sci USA, 93:2322–7, 1996. S K Park, H K Son, S K Lee, J H Kang, Y S Chang, D R Jacobs, Jr, and D H Lee. Relationship between serum concentrations of organochlorine pesticides and metabolic syndrome among non-diabetic adults. J Prev Med Public Health., 43: 1–8, 2010. C J Patel, J Bhattacharya, and A J Butte. An Environment-Wide Association Study (EWAS) on Type 2 Diabetes Mellitus. PLoS ONE, 5:e10746, 2010. B M Psaty, C D Furberg,WARay,andNSWeiss. Potential for conflict of interest in the evaluation of suspected adverse drug reactions: use of cerivastatin and risk of rhabdomyolysis. JAMA, 292:2622–31, 2004. A Puga, A Hoffer, S Zhou, J M Bohm, G D Leikauf, and H G Shertzer. Sustained increase in intracellular free calcium and activation of cyclooxygenase-2 expression in mouse hepatoma cells treated with dioxin. Biochem Pharmacol, 54: 1287–1296, 1997. A Puga, M A Sartor, M Y Huang,JKKerzee,YDWei,CRTomlinson, C S Baxter, and M Medvedovic. Gene expression profiles of mouse aorta and cultured vascular smooth muscle cells differ widely, yet show common responses to dioxin exposure. Cardiovasc Toxicol, 4:385–404, 2004. G M Reaven. Banting lecture 1988. Role of insulin resistance in human disease. Diabetes, 37:1595–607, 1988. S Y Rhee,YCHwang,JTWoo,DHSinn, S O Chin, S Chon, and Y S Kim. Blood lead is significantly associated with metabolic syndrome in Korean adults: an analysis based on the Korea National Health and Nutrition Examination Survey (KNHANES), 2008. Cardiovasc Diabetol., 12, 2013. G Ridgeway. gbm: Generalized Boosted Regression Models, 2013. URL http://CRAN.R-project.org/package=gbm. R package version 2.1. B D Ripley andWNVenables. Modern Applied Statistics with S. Springer, New York, NY, 4th edition, 2002.

82 D A Rodrigues, editor. Drug–Drug Interactions. CRC Press, second edition, 2008. I. Rodushkin and M D Axelsson. Application of double focusing sector field ICP-MS for multielemental characterization of human hair and nails. part I. Analytical methodology. Science of the Total Environment, 250(1–3):83–100, 2000. J Ruzzin, R Petersen, E Meugnier, L Madsen, E J Lock, H Lillefosse, T Ma, S Pesenti, S B Sonne, T T Marstrand, M K Malde, Z Y Du, C Chavey, L Fajas, A K Lundebye, C L Brand, H Vidal, K Kristiansen, and L Frøyland. Persistent organic pollutant exposure leads to insulin resistance syndrome. Environmental health perspectives, 118:465–471, 2010. M Rönn, L Lind, B van Bavel, S Salihovic, K Michaëlsson, and P M Lind. Circulating levels of persistent organic pollutants associate in divergent ways to fat mass measured by DXA in humans. Chemosphere, 85:335–343, 2011. S. Salihovic, L. Mattioli, G. Lindström, L. Lind, P M Lind, and B. van Bavel. A rapid method for screening of the stockholm convention POPs in small amounts of human plasma using SPE and HRGC/HRMS. Chemosphere, 86(7):747–753, 2011. D Sarkar. Lattice: Multivariate Data Visualization with R. Springer, New York, 2008. URL http://lmdvr.r-forge.r-project.org. ISBN 978-0-387-75968-5. S K Seilkop, M J Campen, A K Lund, J D McDonald, and J L Mauderly. Identification of chemical components of combustion emissions that affect pro-atherosclerotic vascular responses in mice. Inhal Toxicol, 24:270–87, 2012. A V Sergeev and D O Carpenter. Increase in metabolic syndrome-related hospitalizations in relation to environmental sources of persistent organic pollutants. Int J Environ Res Public Health, 8:762–76, 2010. H G Shertzer, D W Nebert, A Puga, M Ary, D Sonntag, K Dixon, L J Robinson, E Cianciolo, and T P Dalton. Dioxin causes a sustained oxidative stress response in the mouse. Biochem Biophys Res Commun, 253:44–8, 1998. W Shi, S Wei, X X Hu, G J Hu, C L Chen,XRWang, J P Giesy, andHXYu. Identification of thyroid receptor antagonists in water sources using mass balance analysis and monte carlo simulation. PLoS One, 8, 2013. H Shimodaira. An Approximately Unbiased Test of Phylogenetic Tree Selection. Systematic Biology, 51(3):492–508, 2002. M J Silva, N A Malek, C C Hodge, J A Reidy, K Kato, D B Barr, L L Needham, and J W Brock. Improved quantitative detection of 11 urinary phthalate metabolites in humans using liquid chromatography-atmospheric pressure chemical ionization tandem mass spectrometry. Journal of Chromatography B Analytical Technologies in the Biomedical and Life Sciences, 789(2):393–404, 2003. H K Son, S A Kim, J H Kang, Y S Chang,SKPark, S K Lee, D R Jacobs, Jr, and D H Lee. Strong associations between low-dose organochlorine pesticides and type 2 diabetes in Korea. Environ Int, 36:410–4, 2010. R W Stahlhut, E van Wijngaarden, T D Dye, S Cook, andSHSwan.Concentrations of urinary phthalate metabolites are associated with increased waist circumference and insulin resistance in adult U.S. males. Environmental health perspectives, 115: 876–882, 2007. Stockholm Convention. The 12 initial POPs, 2008. URL http://chm.pops.int/ TheConvention/ThePOPs/The12InitialPOPs/tabid/296/Default.aspx. Accessed November 27, 2014.

83 Stockholm Convention. The pops, 2008. URL http: //chm.pops.int/TheConvention/ThePOPs/tabid/673/Default.aspx. Accessed November 24, 2014. A Sugawara, A Uruno, K Matsuda, T Saito-Ito, T Funato, A Saito-Hakoda, M Kudo, and S Ito. Effects of PPARγ agonists against vascular and renal dysfunction. Curr Mol Pharmacol, 5:248–54, 2012. J Sundström, U Risérus, L Byberg, B Zethelius, H Lithell, and L Lind. Clinical value of the metabolic syndrome for long term prediction of total and cardiovascular mortality: prospective, population based cohort study. BMJ, 332:878–82, 2006. J Sundström, E Vallhagen, U Risérus, L Byberg, B Zethelius, C Berne, L Lind, and E Ingelsson. Risk Associated With the Metabolic Syndrome Versus the Sum of Its Individual Components. Diabetes Care, 29:1673–4, 2006. Swedish Chemicals Agency. Bisphenol A (BPA), 2012. URL https://www.kemi.se/en/Content/In-focus/Bisphenol-A/. Accessed November 25, 2014. W H Te Brake. Air Pollution and Fuel Crises in Preindustrial London, 1250–1650. Technology and Culture, 16:357–339, 1975. M K Townsend, A A Franke, X Li, F B Hu, and A H Eliassen. Within-person reproducibility of urinary bisphenol A and phthalate metabolites overa1to3year period among women in the Nurses’ Health Studies: a prospective cohort study. Environ Health, 12:1–9, 2013. H Uemura, K Arisawa, M Hiyoshi, A Kitayama, H Takami, F Sawachika, S Dakeshita, K Nii, H Satoh, Y Sumiyoshi, K Morinaga, K Kodama, T Suzuki, M Nagai, and T Suzuki. Prevalence of metabolic syndrome associated with body burden levels of dioxin and related compounds among Japan’s general population. Environ Health Perspect, 117:568–73, 2009. United States Environmental Protection Agency. Dioxin, 2010. URL http://cfpub.epa.gov/ncea/CFM/nceaQFind.cfm?keyword=Dioxin. Accessed November 24, 2014. United States Environmental Protection Agency. Basic Information | Polychlorinated Biphenyls (PCBs), 2013. URL http://www.epa.gov/epawaste/hazard/tsd/pcbs/about.htm. Accessed November 24, 2014. United States Environmental Protection Agency. Types of pesticides, 2014. URL http://www.epa.gov/pesticides/about/types.htm. Accessed November 27, 2014. M van den Berg, L Birnbaum, A T Bosveld, B Brunström, P Cook, M Feeley, J P Giesy, A Hanberg, R Hasegawa,SWKennedy, T Kubiak, J C Larsen,FXvan Leeuwen, A K Liem, C Nolt, R E Peterson, L Poellinger, S Safe, D Schrenk, D Tillitt, M Tysklind, M Younes, F Waern, and T Zacharewski. Toxic equivalency factors (TEFs) for PCBs, PCDDs, PCDFs for humans and wildlife. Environ Health Perspect, 106:775–792, 1998. L N Vandenberg, P A Hunt, J P Myers, andFSVomSaal. Human exposures to bisphenol A: mismatches between data and assumptions. Rev Environ Health, 28: 37–58, 2013. T J VanderWeele. Sufﬁcient Cause Interactions and Statistical Interactions. Epidemiology, 20:6–13, 2009.

84 N G Venkata, J A Robinson, P J Cabot, B Davis, G R Monteith, and S J Roberts-Thomson. Mono(2-ethylhexyl)phthalate and mono-n-butyl phthalate activation of peroxisome proliferator activated-receptors α and γ in breast. Toxicol Lett, 163:224–34, 2006. C F Vogel, N Nishimura, E Sciullo, P Wong, W Li, and F Matsumura. Modulation of the chemokines KC and MCP-1 by 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) in mice. Arch Biochem Biophys, 15:169–75, 2007. S Wakui, M Shirai, M Motohashi, T Mutou, N Oyama,MFWempe, H Takahashi, T Inomata, M Ikegami, H Endou, and M Asari. Effects of in Utero Exposure to Di(n-butyl) Phthalate for Estrogen Receptors α, β, and Androgen Receptor of Leydig Cell on Rats. Toxicol Pathol, 42:877–887, 2013. SWWang,SSWang,DCWu,YCLin,CCKu,CCWu,CYChai,JNLee,EM Tsai,CLLin,RCYang,YCKo,HSYu,CHuo, C P Chuu, Y Murayama, Y Nakamura, S Hashimoto, K Matsushima, C Jin, R Eckner, C S Lin, S Saito, and K K Yokoyama. Androgen receptor-mediated apoptosis in bovine testicular induced pluripotent stem cells in response to phthalate esters. Cell Death Dis,4, 2013. M S Wolff, H A Anderson, J A Britton, and N Rothman. Pharmacokinetic variability and modern epidemiology–the example of dichlorodiphenyltrichloroethane, body mass index, and birth cohort. Cancer Epidemiol Biomarkers Prev, 16:1925–30, 2007. S N Wood. Generalized Additive Models. An Introduction with R. Chapman & Hall/CRC, Boca Raton, FL, 2006. ISBN 978-1-58488-474-3. World Health Organization. Global status report on noncommunicable diseases 2010, 2011. World Health Organization. Dioxins and their effects on human health, 2014. URL http://www.who.int/mediacentre/factsheets/fs225/en/. Accessed on December 1, 2014. D Wu, N Nishimura, V Kuo, O Fiehn, S Shahbaz, L van Winkle, F Matsumura, and C F Vogel. Activation of aryl hydrocarbon receptor induces vascular inﬂammation and promotes atherosclerosis in apolipoprotein E-/- mice. Arterioscler Thromb Vasc Biol, 31:1260–7, 2011. F Zaribaf, E Falahi, F Barak, M Heidari,AHKeshteli, A Yazdannik, and A Esmaillzadeh. Fish consumption is inversely associated with the metabolic syndrome. Eur J Clin Nutr., 68:474–80, 2014. W Zhai, Z Huang, L Chen, C Feng, B Li, and T Li. Thyroid endocrine disruption in zebraﬁsh larvae after exposure to mono-(2-ethylhexyl) phthalate (MEHP). PLoS One, 9, 2014.

85 Acta Universitatis Upsaliensis Digital Comprehensive Summaries of Uppsala Dissertations from the Faculty of Medicine 1059 Editor: The Dean of the Faculty of Medicine

A doctoral dissertation from the Faculty of Medicine, Uppsala University, is usually a summary of a number of papers. A few copies of the complete dissertation are kept at major Swedish research libraries, while the summary alone is distributed internationally through the series Digital Comprehensive Summaries of Uppsala Dissertations from the Faculty of Medicine. (Prior to January, 2005, the series was published under the title “Comprehensive Summaries of Uppsala Dissertations from the Faculty of Medicine”.)

ACTA UNIVERSITATIS UPSALIENSIS Distribution: publications.uu.se UPPSALA urn:nbn:se:uu:diva-237690 2015