Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Sensitivity of exposure-response relationships to exposure assessment strategies in retrospective cohort… Friesen, Melissa Charmaine 2006

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2006-199763.pdf [ 5.86MB ]
Metadata
JSON: 831-1.0092944.json
JSON-LD: 831-1.0092944-ld.json
RDF/XML (Pretty): 831-1.0092944-rdf.xml
RDF/JSON: 831-1.0092944-rdf.json
Turtle: 831-1.0092944-turtle.txt
N-Triples: 831-1.0092944-rdf-ntriples.txt
Original Record: 831-1.0092944-source.json
Full Text
831-1.0092944-fulltext.txt
Citation
831-1.0092944.ris

Full Text

SENSITIVITY OF EXPOSURE-RESPONSE RELATIONSHIPS TO EXPOSURE ASSESSMENT STRATEGIES IN RETROSPECTIVE COHORT STUDIES by MELISSA CHARMAINE FRIESEN B.Sc, The University of British Columbia, 1997 M.Sc, The University of British Columbia, 2001 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE F A C U L T Y OF G R A D U A T E STUDIES (Occupational & Environmental Hygiene) THE UNIVERSITY OF BRITISH COLUMBIA September 2006 © Melissa Charmaine Friesen, 2006 Abstract Introduction: Exposure misclassification in epidemiologic studies results in attenuated exposure-response relationships. Using quantitative exposure estimates helps reduce exposure misclassification in retrospective studies, but it does not eliminate it. This study focused on two areas that impact the degree of exposure misclassification inherent in quantitative exposure estimates: 1) the choice of exposure indicator where the exposure o f interest is a component of a mixture; 2) the choice between expert judgment and measurement-based exposure assessment strategies. Methods: These themes were examined empirically using data from two Canadian retrospective occupational cohort studies: the B C Aluminum Smelter Cohort (n=6423) and the B C Sawmill Cohort (n=26,847). In the smelter cohort exposure to polycyclic aromatic hydrocarbons ( P A H ) , measured as benzene soluble materials ( B S M ) and benzo(a)pyrene (BaP), was examined in relation to mortality and cancer incidence. In the sawmill cohort three exposures classes were evaluated: 1) nonspecific dust and wood dust in relation to chronic obstructive pulmonary disease hospitalizations; 2) total chlorophenols, pentachlorophenols (PCP), and tetrachlorophenol (TCP) in relation to cancer incidence; and 3) noise estimated using expert-based and measurement-based approaches in relation to heart disease mortality. The shape, goodness of fit, and precision of the exposure-response relationships were evaluated using Poisson regression. Results: Using a more specific exposure measure that was more proximal to the causal agent(s) improved the precision of the exposure-response relationship by between 1-14% for BaP over B S M , by 10-30% for P C P over total chlorophenols, and by 218% for wood dust over nonspecific dust. Measurement-based noise estimates improved the precision by 12-108% over the use o f expert ratings. Accounting for hearing protection use in the measurement-based noise estimates improved the precision by 58% over the unadjusted estimates. The observed attenuation was not correlated with predicted attenuation from theoretical equations. Conclusions: Examining the precision of exposure-response relationships provided a quantifiable measure for evaluating different exposure assessment approaches. Refining the quantitative exposure estimates, through the use o f more proximal exposure measures and the use of exposure measurements, resulted in stronger, more precise exposure-response relationships. With the exception of the expert-based noise estimates, the more common, less specific exposure measures resulted in inconclusive exposure-disease associations. ii Table of Contents Abstract ii Table of Contents iii List of Tables v List of Figures vi List of symbols and abbreviations '. vii Acknowledgements viii Dedication ix Co-authorship statement x Chapter 1: Introduction 1 Research Objectives 5 References 8 Chapter 2: From expert-based to quantitative retrospective exposure assessment at a Soderberg aluminum smelter Introduction 12 Methods 13 Results 19 Discussion 20 Tables and Figures 24 References 31 Chapter 3 : Adequacy of benzo(a)pyrene and benzene soluble materials as indicators of exposure to polycyclic aromatic hydrocarbon mixtures Introduction 34 Methods 35 Results 37 Discussion 40 Tables and Figures 43 References 49 Chapter 4: Comparison of benzo(a)pyrene and benzene soluble materials as indices of exposure to polycyclic aromatic hydrocarbons for a retrospective cohort study of aluminum smelter workers Introduction 51 Methods 52 Results 55 Discussion 56 Tables and Figures 59 References 63 Continued. Chapter 5: Impact of the specificity of the exposure metric on exposure-response relationships in a large retrospective occupational cohort Introduction 66 Methods 67 Results 70 Discussion 72 Tables and Figures 75 References 80 Chapter 6: Comparison of expert and measurement-based exposure assessment of historical noise levels for a sawmill cohort Introduction 82 Methods 84 Results 87 Discussion 89 Tables and Figures 92 References 99 Chapter 7: General Discussion 102 Conclusions and Significance 109 Summary Table 111 References 112 iv List of Tables Table 2.1 Model coefficients and standard errors for CTPV and BaP potroom operations mixed effects models developed to predict exposures for 1977-2000 24 Table 2.2 Model coefficients and standard errors for the CTPV and BaP potroom maintenance mixed effects models developed to predict exposures for 1977-2000...25 Table 2.3 Updated CTPV and BaP Job Exposure Matrices by Job and Time Period 26 Table 3.1 Coal tar pitch in use at Soderberg aluminum smelter 43 Table 3.2 BaP/BSM ratio by work area and coal tar pitch mixture 44 Table 3.3 Correlation of benzo(a)pyrene with particulate PAHs in the potrooms, by pitch and job type 45 Table 3.4 Multiple linear regression of time and work area-specific factors that impact the BaP RARs 46 Table 3.5 Predicted BaP RARs and BaP Equivalency Factors by pitch type and potroom building for pot operators 47 Table 4.1 Exposure category cut points and mean cumulative exposure for BSM and BaP by health outcome 59 Table 4.2 Model parameters and precisions for the relationships between BaP and BSM exposure and bladder cancer, lung cancer, and acute myocardial infarction 60 Table 5.1 Model parameters and precisions for the relationships between dust and wood dust exposure and hospitalizations for COPD 75 Table 5.2 Model parameters and precisions for the relationships between chlorophenols exposure and N H L and kidney cancer 76 Table 6.1 Exposure category cut points and mean cumulative exposure for cumulative noise exposure 92 Table 6.2 Pearson correlation, Spearman correlation and weighted kappa between cumulative expert- and measurement-based noise estimates 93 Table 6.3 Model parameters and precisions for the relationships between cumulative noise exposure and acute myocardial infarction 94 Table 7.1 Summary of main findings on impact of exposure metric choice on exposure-response relationships 111 v List of Figures Figure 2.1 Multi-component procedure to develop quantitative C T P V and BaP exposure estimates 28 Figure 2.2 Predicted annual mean exposures for 1977-2000 from the C T P V model and BaP model for pot operators by potroom group 29 Figure 2.3 Predicted mean exposures for 1977-2000 from the C T P V model and BaP model ' for potroom repairman 30 Figure 3.1 Comparison o f mean BaP R A R for three potroom jobs and pitch type K o s c o / V F T (>75 / <25%) 48 Figure 4.1 Relationship between cumulative BaP and B S M exposure indices 61 Figure 4.2 Log-linear and log-log relationships between exposure and bladder cancer incidence, 20 year lag 61 Figure 4.3 Log-linear and log-log relationships between exposure and lung cancer incidence, 20 year lag 62 Figure 4.4 Log-linear and log-log relationships between exposure and acute myocardial infarction mortality, no lag 62 Figure 5.1 Log-linear and log-log relationships between dust and wood dust exposure and C O P D hospitalizations 77 Figure 5.2 Log-linear and log-log relationships between total chlorophenol and tetrachlorophenol exposure and N H L , 20 year lag 78 Figure 5.3 Log-linear and log-log relationships between total chlorophenol and tetrachlorophenol exposure and kidney cancer, 20 year lag 79 Figure 6.1 Comparison of expert-based and measurement-based cumulative noise estimates for subcohort 95 Figure 6.2 Agreement in the exposure category assigned each cohort subject between exposure assessment approaches 96 Figure 6.3 Log-linear and log-log relationships between acute myocardial infarction mortality and cumulative noise estimates for full cohort 97 Figure 6.4 Log-linear and log-log relationships between acute myocardial infarction mortality and cumulative noise estimates for subcohort 98 vi List of symbols and abbreviations A M I Acute myocardial infarction BaP Benzo(a)pyrene BC British Columbia BEF Benzo(a)pyrene equivalency factors B S M Benzene soluble materials CE Cumulative exposure CI Confidence interval COPD Chronic obstructive pulmonary disease CTPV Coal tar pitch volatiles dBA Decibel. Units of sound pressure level. dBA*year Unit of cumulative noise exposure. HPD Hearing protection device ICD International Classifications of Diseases J E M Job exposure matrix L e q A measured equivalent sound pressure over a period of time (usually an 8 or 12-hour shift in this study) N H L Non-Hodgkin's lymphoma NS dust Non-specific dust P A H Polycyclic aromatic hydrocarbons PCP Pentachlorophenol R A R Relative abundance ratio RR Relative risk SD Standard deviation TCP Tetrachlorophenol U B C University of British Columbia WCB Workers' Compensation Board of British Columbia (now called WorkSafe BC) vii Acknowledgements While one name is on the front of this dissertation, it would not have been possible without the support of many individuals. First, I thank my research supervisor, Paul Demers, and the members of my supervisory committee Hugh Davies, Nhu Le, and Kay Teschke. Their constructive advice throughout was essential and made my PhD experience an exceptional, positive experience. Their ability to recognize the scope of work that would be possible in a four year PhD was greatly appreciated. I also thank the joint Alcan/Canadian Auto Workers advisory committee for their comments and their cooperation in developing the exposure assessment for the BC Aluminum Smelter Cohort. I also thank Richard Lapointe, Barry Boudreault, Jim Thorne, Nela Walter, and Mary Lo for their assistance in the smelter study. The participation of the many workers in both the BC Sawmill Cohort and the BC Aluminum Smelter Cohort was greatly appreciated. Protecting the health of these workers is the reason why occupational health research is so important. A big thank you to all the faculty, staff and students in the School of Occupational and Environmental Hygiene. The collaborative nature of this department created an outstanding learning environment. Specific thanks to all the SOEH PhD students: I will miss our lunches. Special thanks also to Marilyn Borugian and Chris Bajdik who shared their personal PhD experiences with me and provided fantastic advice at the beginning of my degree. Finally, I would like to acknowledge the personal funding support provided through trainee awards given by the Michael Smith Foundation for Health Research, the BC Medical Services Foundation, and the Canadian Institutes for Health Research (CIHR). Their funding made it possible to complete this research in four years. I also appreciate the study-specific funding from CIHR, the BC Lung Association, US National Institute for Occupational Safety & Health, Alcan, and WorkSafeBC without which this research could not be completed. viii Dedication To my husband, Scott Gater, and to our family...without your support this thesis would not have been possible. "Call it a clan, call it a network, call it a tribe, call it a family. Whatever you call it, whoever you are, you need one. " Jane Howard ix Co-authorship statement This statement is to acknowledge my role and that of the co-authors in the research presented in this dissertation. The enumeration of the retrospective cohort studies and the determination o f the subjects' health outcomes had been conducted by the co-authors prior to this dissertation. The exposure assessment for chlorophenols and noise in the B C Sawmill Cohort were conducted by the co-authors prior to this dissertation. A s part o f work prior to my dissertation I was directly responsible for the development of historical dust and wood dust exposure estimates in the B C Sawmill Cohort, which included the development of predictive models and validating the estimates using exposure measurements and exposure determinants that had previously been collected by co-authors. M y role in the research presented in this dissertation was as follows. I was directly responsible for all aspects, including developing the methodology, gathering historical exposure measurements and determinants of exposure, and conducting the analyses, for the development of historical exposure estimates for polycyclic aromatic hydrocarbons ( P A H ) in the B C Aluminum Smelter Cohort (Chapter 2). I developed the approach and conducted all analyses for comparing the two measures of polycyclic aromatic hydrocarbons with other components of the P A H mixture (Chapter 3). 1 developed the framework for evaluating the impact of different exposure metrics on the exposure-response relationships and I conducted all epidemiologic analyses presented in this dissertation (Chapters 4-6). I prepared all aspects of the manuscript, including the text, tables and figures. M y supervisory committee and fellow co-authors provided feedback throughout my research and their comments have been incorporated into the final draft o f the manuscript. x Chapter 1: Introduction Quite likely the biggest challenge faced in retrospective studies o f occupational disease is to assess exposure. Chronic diseases, such as cancer, provide a particular challenge to exposure assessors due to long latency periods between exposure and disease onset. Epidemiological studies of chronic disease require workers' exposures to be estimated for their entire employment history. A paucity o f exposure measurements has limited many studies to the use of qualitative or surrogate exposure measures, such as ever/never exposed or duration o f exposure. These measures have been successful in identifying associations between exposure and disease when the risks are large and when exposure levels are high (Schneider, 2002). However, quantitative exposure assessment, which differentiates jobs and time periods by the intensity of exposure, is needed to reduce exposure misclassification when the associated risks and exposure levels are lower in order to detect associations that would otherwise be missed (Stewart and Herrick, 1991; Kauppinen, 1994). Quantitative exposure assessment has several additional benefits. It allows for the relationship between intensity o f exposure and risk o f disease (the exposure-response relationship) to be examined, which is an important criterion in establishing causality. It facilitates the comparison of risk between studies, countries, and industries. It also provides a solid base for policy initiatives, such as risk assessment, the establishing o f preventative strategies, and the development o f occupational exposure limits. Substantial effort is required to develop quantitative exposure estimates for retrospective studies, but this effort is necessary. A review of studies reporting both quantitative and surrogate exposure measures found that quantitative exposure estimates were more successful in identifying exposure-response relationships than using exposure duration (Blair and Stewart, 1992). The quantitative estimates were found to more often yield larger relative risk estimates and monotonically increasing exposure-response gradients in cancer studies compared to using exposure duration. Nevertheless, while developing quantitative exposure estimates helps reduce exposure misclassification in epidemiologic studies, it does not eliminate it. Simulation studies have demonstrated that exposure misclassification in quantitative exposure measures can both attenuate and distort exposure-response relationships (Dosemeci etal, 1990; Dosemeci and Stewart, 1996). The exposure assessment process is complex, with numerous aspects that can introduce or eliminate exposure misclassification. This dissertation focused on two specific areas o f quantitative exposure assessment strategies that can impact the degree of exposure misclassification inherent in the 1 exposure estimates: 1) the choice o f exposure indicator where the exposure of interest is a component o f a mixture; and 2) the choice between expert judgment and measurement-based exposure assessment strategies. Overarching both themes was the need for methods to evaluate the degree of exposure misclassification in the exposure estimates and to examine the impact of exposure misclassification on exposure-response relationships in epidemiologic studies. The remainder of the introduction discusses these two themes, methods for assessing the validity of exposure measures, and the specific research objectives of this dissertation. Theme 1: Exposure Indicators for Mixtures Exposure to mixtures has posed a particular challenge for exposure assessors. Exposure misclassification, and thus attenuation and distortion of exposure-response relationships, can result when the available exposure indicator is not highly correlated with the causal components (Ahrens and Stewart, 2003; Burdorf, 2003). Three situations where exposure misclassification could be introduced in an exposure indicator when exposure occurs as a mixture are discussed. One situation occurs when analytic methods are nonspecific, resulting in exposure misclassification i f the relationship between the chosen exposure indicator and the causal agents is not constant, such as across work areas, plants, or industries. For instance, the wood dust component of total dust has been found to vary across work areas and jobs, but the gravimetric sampling method used measures all particulate sources (Demers et ai, 2000). A second situation occurs when an analytic method measures only one component, although multiple components may contribute differentially to the mixture's toxicity. For instance, exposure to polycyclic aromatic hydrocarbons ( P A H ) is often measured using benzo(a)pyrene, a specific particulate-phase P A H , but other components of the P A H mixture also have toxic properties that could contribute to disease. A third situation occurs when analytic and sampling methods do not capture the relevant airborne particulate fraction. For instance, the respirable fraction may be more relevant than thoracic or inhalable fractions for specific health outcomes, but often only total (37mm) particulate levels are obtained. Exposure indicators for mixtures in epidemiologic studies can often be refined to better reflect a more proximal measure o f the causal agents. Pilot studies that examined the relationship between the available exposure indicators and more proximal measures have been used to convert from one measure to another, such as from dust to wood dust or between different size particle fractions (Friesen et al, 2005). However, these conversions are unable to consider all sources of variability in the relationship between one measure and another; they may introduce additional uncertainty and measurement error. Thus, choosing appropriate exposure indicators for exposure mixtures in 2 epidemiologic studies is nontrivial, with compromises made between imperfect measures of exposure that can be more accurately assessed compared to a more proximal exposure measure of the causal agents that has additional uncertainty. The impact of choosing one exposure indicator over another has been difficult to assess as rarely are more than one exposure indicator for an exposure mixture available in a single study. Theme 2: Expert versus Measurement-based Exposure Assessment Strategies Each epidemiologic study differs in the availability of exposure information, job task information, and other information needed to develop quantitative exposure estimates. Given this variability, it is not surprising that there is a lack of standardized methods for quantitative exposure assessment (Stewart, 1999). Quantitative exposure assessment has usually been conducted using expert raters, simple algorithms, or statistical models using a basic job exposure matrix (JEM) framework (Seixas and Checkoway, 1995). These methods have been used alone and together, depending on the availability of exposure measurements (Stewart et al, 1996). When exposure measurements are limited or lacking, expert-based exposure assessment is often the only option. Because of the frequent reliance on experts for exposure assessment in epidemiologic studies, their ability to accurately assess semi-quantitative and quantitative exposure levels for epidemiologic studies has received much attention. Committees of expert raters appear to be able to achieve reasonable agreement among themselves; however, their classifications have often been poorly correlated with directly measured exposure levels (Kromhout et ai, 1987; Hertzman et al, 1988; Hawkins and Evans, 1989; Teschke et al., 1989; Post et al, 1991;de Cocke/ al, 1996; Teschke et al, 2002; Friesen et al, 2003). Regardless of the experts' accuracy, many of these same studies have used those estimates in epidemiologic studies due to the lack of alternatives. On the other hand, the availability of exposure measurements for a measurement-based exposure assessment approach does not guarantee unbiased exposure estimates. The exposure measurements may have been collected under non-random or non-optimized sampling conditions resulting in an "alloyed" gold standard (Stewart et al, 1996; Mulhausen and Damiano, 1998). Or, as mentioned in the previous section, the sampling method may not adequately reflect the causal components. Both expert judgment and measurement-based exposure assessment approaches are subject to exposure misclassification, but there is limited empirical evidence of the comparative robustness of their exposure-response relationships due to the rarity of both approaches being used in the same 3 study. One such study found much higher relative risks (RR) for the measurement-based exposure estimates than for an expert-based JEM (RR 2.27 vs. 1.53) for the highest exposure group in a study of magnetic fields and brain cancer (Kromhout et al. 1999), but more empirical evidence is needed. Assessing the Validity of Exposure Estimates Assessing the validity of exposure estimates that are to be used in epidemiologic studies is a necessary, but often challenging step in the development of retrospective exposure estimates and thus is frequently neglected (Stewart, 1999; Hornung et al, 1994). It is unlikely that a true gold standard for exposure dose exists for most exposures; however, studies evaluating the validity of exposure assessment estimates provide a useful measure of their bias and precision and may serve to calibrate the exposure estimates. Direct methods of validating retrospective exposure estimates have been used in several studies. Methods include setting aside data or using data from other studies to compare the estimates obtained by a specific exposure assessment methodology (Hornung et al, 1994; Stewart et al., 1996; Glass et al, 2001; Burstyn et al, 2002; Friesen et ai, 2005) or designing studies to compare exposure estimates to current exposure measurements including biological measurements (Hertzman et al, 1988; Astrakianakis et al, 1998; Friesen et al, 2003). Such validation studies can be used to identify where the exposure assessment approach performs well. More importantly, identifying where it performs poorly can assist in improving the exposure assessment approach. By necessity, most validation studies have focused on the most recent time periods in the study even though the study may encompass many decades. As a result, this direct validation of exposure estimates provides little guidance on the potential misclassification for the earlier time periods in a retrospective epidemiologic study. Alternatively, indirect methods of validation can be used to evaluate the accuracy of exposure estimates by examining the sensitivity of exposure-response relationships to the choice in exposure metric. This is particularly useful when there is a previously known exposure-response relationship, such as the risk of silicosis or rate of benzene poisoning (Dosemeci et al, 1994; Dosemeci et al, 1997; t Mannetje et al, 2002). However, in the absence of an established relationship, the shape, magnitude, and fit of exposure-response relationships can be examined (Loomis et al, 1999). Simulation studies have been used to determine the direction and magnitude of specific patterns of exposure misclassification on exposure-response relationships (Dosemeci et al, 1990; Flegal et al, 1991; Veierod and Laake, 2001; Richardson and Loomis, 2004), but empirical evidence has been limited. Examining the robustness of exposure-response relationships can provide a measure of the overall accuracy of the retrospective exposure estimates. As the relationship is based on cumulative 4 exposure which aggregates each worker's exposure across all jobs and time periods, it does not directly inform researchers on which aspects o f a particular exposure assessment strategy (i.e. specific time periods or jobs) performed poorly. However, indirect validation attempts can include sensitivity analyses where some exposure assessment decisions are evaluated, such as the relative weights applied to historical exposure estimates when exposure measurements were lacking (Loomis et al, 1998; Kromhout et al, 1999). In contrast to the conventional confidence intervals which represent only the random e'rror in the risk estimate, determining the robustness of exposure-response relationships to exposure assessment decisions can provide more information about the uncertainty in the exposure-response relationship (Kromhout et al, 1999). The use o f indirect validation to examine the impact on the exposure-response relationship of exposure assessment decisions is becoming more common. The impact o f grouping schemes for exposure assessment has received the greatest attention thus far (Kromhout and Heederik, 1995; Seixas and Sheppard, 1996; Kromhout et al, 1997; van Tongeren et al, 1999; Heederik and Attfield, 2000; Werner and Attfield, 2000). The choice in exposure metric (average vs. cumulative), exposure category cut points, and historical weighting of exposures has also received some attention (Checkoway and Rice, 1992; Armstrong et al, 1994; Loomis et al, 1998; Kromhout et al, 1999; Burstyn et al, 2003; Richardson and Loomis, 2004; Loomis et al, 2005). As yet little attention has been given to either o f the two themes of this dissertation, exposure indicators for mixtures or expert-versus measurement-based strategies. Research Objectives The overall objective o f this doctoral thesis was to examine the sensitivity of the exposure-response relationships to exposure assessment decisions as a measure o f their validity in the following two thematic areas: 1) choice o f exposure indicator for mixtures; and 2) expert judgment versus measurement-based approaches to exposure assessment. These themes were examined empirically using two large retrospective cohort studies based in British Columbia (BC) , Canada. The B C Aluminum Smelter Cohort consists o f approximately 6,400 workers employed at least 3 years between 1954 and 2000 at one Soderberg aluminum smelter (Chapters 2-4). The B C Sawmil l Cohort consists of approximately 28,000 sawmill workers employed at one of 14 sawmills for at least one year between 1950 and 1995 (Chapters 5-6). Examining the impact of exposure assessment decisions w i l l help determine i f more sophisticated quantitative exposure assessment approaches provide more accurate and precise exposure-response relationships with less exposure 5 misclassification. The specific research questions for the five studies included in this dissertation are described below. Theme 1: Exposure Indicators for Mixtures In this theme, more commonly available, but less specific exposure measures are compared to refined exposure estimates hypothesized to be more proximal measures of the causal agents. While both the less specific and more proximal measures are subject to exposure misclassification, the more proximal exposure measures are expected a priori to result in stronger, more precise exposure-response relationships. In Chapter 2 quantitative exposure estimates for two different exposure measures [benzo(a)pyrene (BaP) and benzene soluble materials (BSM)] of exposure to a mixture polycyclic aromatic hydrocarbons (PAH) were developed for the BC Aluminum Smelter Study. These two exposure indicators were developed independently from each other, using a measurement-based exposure assessment approach wherever possible. The exposure estimates developed here were used in the epidemiologic analyses reported later in Chapter 4. In Chapter 3 the relationship between two commonly evaluated measures of PAH exposure (BaP, BSM) and other components of the PAH mixture within the aluminum smelter environment were assessed as a direct measure of their validity. Multiple components of the PAH mixture are believed to contribute to the mixture's toxicity. The factors that influence the relationship between the two commonly evaluated measures of PAH exposure in the aluminum smelter and other PAH components were evaluated as a measure of the stability of the relationship and provided an indication of the potential exposure misclassification that could occur if these factors were not taken into account. In Chapter 4 the validity of the BaP and BSM exposure measures developed in Chapter 2 were indirectly assessed by examining the robustness of the exposure-response relationships in the BC Aluminum Smelter Study. Three health outcomes were examined: bladder cancer incidence, lung cancer incidence, and acute myocardial infarction mortality. In Chapter 5 the validity of specific versus nonspecific exposure indicators of wood dust and chlorophenol exposure were indirectly assessed by examining the robustness of the exposure-response relationships in the BC Sawmill Cohort. The exposure metrics for wood dust and chlorophenol exposure were previously developed and reported elsewhere (Friesen et al, 2005; 6 Demers et al, 2006). The following exposure-disease associations were evaluated: wood dust and hospitalization for chronic obstructive pulmonary disease; and chlorophenols and cancer incidence (non-Hodgkin lymphoma, kidney cancer). Thematic Area 2: Expert Judgment versus Measurement-based Exposure Assessment Strategies In this theme, expert-based exposure estimates are compared to measurement-based exposure estimates. The measurement-based estimates are hypothesized to result in stronger, more precise exposure-response relationships. In Chapter 6 the validity of expert ranked and statistical model based occupational noise estimates were indirectly assessed by examining the robustness of the exposure-response relationship with acute myocardial infarction in the BC Sawmill Cohort. The exposure metrics for occupational noise exposure were previously developed and reported elsewhere (Davies, 2002; Ostry et al, 1998). Although not part of this thesis, a study testing the validity of expert-based PAH exposure estimates for the BC Aluminum Smelter Cohort found that there was only moderate agreement (weighted kappa = 0.39) between the expert estimates and exposure measurements (Friesen et al, 2003). Hence, a measurement-based exposure assessment, described in Chapter 2, was used to develop quantitative exposure indicators of PAH exposure for this cohort. Overall Discussion The five papers in this dissertation are connected through two common threads: 1) Does more sophisticated exposure assessment approaches lead to measurable improvements in exposure-response relationships; and 2) how does one measure whether an exposure assessment approach reduces exposure misclassification? Substantial effort is being spent on improving exposure assessment for epidemiologic studies with the a priori expectation that true associations between exposure and disease will be detected because exposure misclassification is minimized. Simulation studies support this expectation, but there is limited empirical evidence of the degree of improvement resulting from more sophisticated exposure assessment strategies. The main findings of this dissertation, its strengths and limitations, and these common threads are discussed in Chapter 7. 7 References Ahrens W, Stewart P (2003) Retrospective exposure assessment. Exposure assessment in occupational and environmental epidemiology. Ed. M . J. Nieuwenhuijsen. Oxford, Oxford University Press: 104. Armstrong B, Tremblay C, Baris D, Theriault G (1994) Lung cancer mortality and polynuclear aromatic hydrocarbons: A case-cohort study of aluminum production workers in Arvida, Quebec, Canada. Am J Epidemiol; 139 250-62. Astrakianakis G, Band PR, Le N, Bert J, Janssen B, Svirchev L M , Tang C, Anderson J, Keefe A (1998) Validation of a mill-specific job-exposure matrix in the British Columbia pulp and paper industry. Appl Occup Environ Hyg; 13 1-7. Blair A, Stewart PA (1992) Do quantitative exposure assessments improve risk estimates in occupational studies of cancer? Am J Ind Med; 21 53-63. Burdorf A (2003) Analysis and modelling of personal exposure. Exposures assessment in occupational and environmental epidemiology. Ed. M . J. Nieuwenhuijsen. Oxford, Oxford University Press. Burstyn I, Boffetta P, Burr GA, Cenni A, Knecht U, Sciarra G, Kromhout H (2002) Validity of empirical models of exposure in asphalt paving. Occup Environ Med; 59 620-4. Burstyn I, Boffetta P, Kauppinen T, Heikkila P, Svane O, Partanen T, Stucker I, Frentzel-Beyme R, Ahrens W, Merzenich H, Heederik D, Hooiveld M , Brunekreef B, Langard S, Randem BG, Jarvholm B, Bergdahl IA, Shaham J, Ferro G, Kromhout H (2003) Performance of different exposure assessment approaches in a study of bitumen fume exposure and lung cancer mortality. Am J Ind Med; 43 40-8. Checkoway H, Rice C H (1992) Time-weighted averages, peaks, and other indices of exposure in occupational epidemiology. Am J Ind Med; 21 25-33. Davies HW (2002) Exposure to occupational noise and risk of cardiovascular disease: A retrospective cohort study. Interdisciplinary Studies. Vancouver, University of British Columbia. PhD. de Cock J, Kromhout H, Heederik D, Burema J (1996) Experts' subjective assessment of pesticide exposure in fruit growing. Scand J Work Environ Health; 22 425-32. Demers PA, Davies HW, Friesen M C , Hertzman C, Ostry A, Hershler R, Teschke K (2006) Cancer and occupational exposure to pentachlorophenol and tetrachlorophenol. Cancer Causes Control; In Press. Demers PA, Teschke K, Davies HW, Kennedy SM, Leung V (2000) Exposure to dust, resin acids, 8 and monoterpenes in softwood lumber mills. Am Ind Hyg Assoc J; 61 521-8. Dosemeci M , McLaughlin JK, Chen JQ, Hearl F, McCawley M , Wu Z, Chen RG, Peng K L , Chen A L , Rexing SH, et al. (1994) Indirect validation of a retrospective method of exposure assessment used in a nested case-control study of lung cancer and silica exposure. Occup Environ Med; 51 136-8. Dosemeci M , Rothman N, Yin SN, Li G L , Linet M , Wacholder S, Chow WH, Hayes RB (1997) Validation of benzene exposure assessment. Ann N Y Acad Sci; 837 114-21. Dosemeci M , Stewart P (1996) Recommendations for reducing the effects of exposure misclassification on relative risk estimates. OccupHyg; 3 169-76. Dosemeci M , Wacholder S, Lubin JH (1990) Does nondifferential misclassification of exposure always bias a true effect toward the null value? Am J Epidemiol; 132 746-8. Flegal K M , Keyl PM, Nieto FJ (1991) Differential misclassification arising from nondifferential errors in exposure measurement. Am J Epidemiol; 134 1233-44. Friesen M C , Davies HW, Teschke K, Marion S, Demers PA (2005) Predicting historical dust and wood dust exposure in sawmills: Model development and validation. J Occup Environ Hyg; 2 650-8. Friesen M C , Demers PA, Spinelli JJ, Le ND (2003) Validation of a semi-quantitative job exposure matrix at a Soderberg aluminum smelter. Ann Occup Hyg; 47 477-84. Glass DC, Gray CN, Adams G G , Manuell RW, Bisby JA (2001) Validation of exposure estimation for benzene in the Australian petroleum industry. Toxicol Ind Health; 17 113-27. Hawkins NC, Evans JS (1989) Subjective estimation of toluene exposures: A calibration study of industrial hygienists. Appl Indust Hyg; 4 61-68. Heederik D, Attfield M (2000) Characterization of dust exposure for the study of chronic occupational lung disease: A comparison of different exposure assessment strategies. Am J Epidemiol; 151 982-90. Hertzman C, Teschke K, Dimich-Ward H, Ostry A (1988) Validity and reliability of a method for retrospective evaluation of chlorophenate exposure in the lumber industry. Am J Ind Med; 14 703-13. Hornung RW, Greife A L , Stayner LT, Steenland NK, Herrick RF, Elliott LJ, Ringenburg V L , Morawetz J (1994) Statistical model for prediction of retrospective exposure to ethylene oxide in an occupational mortality study. Am J Ind Med; 25 825-36. Kauppinen TP. (1994) Assessment of exposure in occupational epidemiology. Scand J Work Environ Health; 20 19-29. Kromhout H, Heederik D (1995) Occupational epidemiology in the rubber industry: Implications of 9 exposure variability. Am J Ind Med; 27 171-85. Kromhout H, Loomis DP, Kleckner RC (1999) Uncertainty in the relation between exposure to magnetic fields and brain cancer due to assessment and assignment of exposure and analytical methods in dose-response modeling. Ann N Y Acad Sciences; 895 141-55. Kromhout H, Loomis DP, Kleckner RC, Savitz DA (1997) Sensitivity of the relation between cumulative magnetic field exposure and brain cancer mortality to choice of monitoring data grouping scheme. Epidemiol; 8 442-5. Kromhout H, Oostendorp Y, Heederik D, Boleij JS (1987) Agreement between qualitative exposure estimates and quantitative exposure measurements. Am J Ind Med; 12 551-62. Loomis A, Kromhout H, Kleckner RC, Savitz DA (1998) Effects of the analytical treatment of exposure data on associations of cancer and occupational magnetic field exposure. Am J Ind Med; 34 49-56. Loomis D, Salvan A, Kromhout H, Kriebel D (1999) Selecting indices of occupational exposure for epidemiologic studies. Occup Hyg; 5 73-91. Loomis D, Richardson DB, Elliott L (2005) Poisson regression analysis of ungrouped data. Occup Environ Med; 62 325-9. Mulhausen JR, Damiano J (1998) A strategy for assessing and managing occupational exposures. Fairfax, Virginia: AIHA Press. Ostry A (1998) Psychosocial job strain and coronary heart disease in a cohort of blue collar workers. Department of Health Care and Epidemiology. Vancouver, University of British Columbia. PhD. Post W, Kromhout H, Heederik D, Noy D, Duijzentkunst RS (1991) Semiquantitative estimates of exposure to methylene chloride and styrene: The influence of quantitative exposure data. Appl Occup Environ Hyg; 6 197-204. Richardson DB, Loomis D (2004) The impact of exposure categorisation for grouped analyses of cohort data. Occup Environ Med; 61 930-5. Schneider T (2002) Improving exposure assessment requires measurements and modeling. Scand J Work Environ Health; 28 367-9. SeixasNS, Checkoway H (1995) Exposure assessment in industry specific retrospective occupational epidemiology studies. Occup Environ Med; 52 625-33. Seixas NS, Sheppard L (1996) Maximizing accuracy and precision using individual and grouped exposure assessments. Scand J Work Environ Health; 22 94-101. Stewart P (1999) Challenges to retrospective exposure assessment. Scand J Work Environ Health; 25 505-10. 10 Stewart P A , Herrick RF (1991) Issues in performing retrospective exposure assessment. Appl Occup Environ Hyg; 6 421-27. Stewart PA, Lees PS, Francis M (1996) Quantification of historical exposures in occupational cohort studies. Scand J Work Environ Health; 22 405-14. t Mannetje A , Steenland K , Checkoway H, Koskela RS, Koponen M , Attfield M , Chen J , Hnizdo E, DeKlerk N, Dosemeci M (2002) Development of quantitative exposure data for a pooled exposure- response analysis of 10 silica cohorts. Am J Ind Med; 42 73-86. Teschke K , Hertzman C, Dimich-Ward H, Ostry A , Blair J , Hershler R (1989) A comparison of exposure estimates by worker raters and industrial hygienists. Scand J Work Environ Health; 15 424-9. Teschke K , Olshan AF, Daniels J L , De Roos A J , Parks C G , Schulz M , Vaughan T L (2002) Occupational exposure assessment in case-control studies: Opportunities for improvement. Occup Environ Med; 59 575-93. van Tongeren M J , Kromhout H, Gardiner K , Calvert I A , Harrington J M (1999) Assessment of the sensitivity of the relation between current exposure to carbon black and lung function parameters when using different grouping schemes. Am J Ind Med; 36 548-56. Veierod MB, Laake P (2001) Exposure misclassification: Bias in category specific poisson regression coefficients. Stat Med; 20 771-84. Werner M A , Attfield MD (2000) Effect of different grouping strategies in developing estimates of personal exposures: Specificity versus precision. Appl Occup Environ Hyg; 15 21-5. 11 Chapter 2: From expert-based to quantitative retrospective exposure assessment at a Soderberg aluminum smelter 1 Introduction In the last fifteen years significant advancements have been made in quantitative assessment methods for retrospective occupational studies. Many methods to estimate quantitative exposure levels have been used alone and in conjunction with other techniques, with the options dictated by the availability of exposure measurements (Stewart et al, 1996). Calculation of means and the use of simple algorithms are limited to time~periods with available exposure measurements. The use of statistical models, both deterministic and empirical, is increasing because of their ability to borrow information in a rigorous and repeatable manner to predict exposure levels for circumstances where no measurements were made (Dement et al., 1983; Eisen et ai, 1984; Yu et ai, 1990; Hornung et al., 1994; Kromhout et al., 1994; Burstyn et al., 2000). Experts have been frequently asked to estimate exposures for studies and time periods where measurements are limited or lacking (Spinelli et ai, 1991; Astrakianakis et ai, 1998; Kauppinen et al., 2002). However, numerous studies have found that the experts' exposure estimates are often only poorly to moderately correlated with directly measured exposure levels (Kromhout et al, 1987; Hertzman et al, 1988; Hawkins and Evans, 1989; Teschke et al, 1989; Post et al, 1991; de Cock et al, 1996; Teschke et al, 2002; Friesen et al, 2003). Their findings support the rigorous use of exposure measurements over expert opinion whenever possible. A 15-year study update at a vertical stud Soderberg aluminum smelter together with 25 years of personal coal tar pitch volatiles (CTPV) exposure measurements collected at the smelter provided an opportunity to update the original expert-based semi-quantitative job exposure matrix (JEM) (Spinelli et al, 1991). A comparison of the original JEM with available exposure measurements found that the expert committee's exposure estimates were only moderately correlated with a measurement-based exposure assessment strategy (Spearman's rho = 0.42) (Friesen et al, 2003). Additionally, several areas where improvements were possible were identified, including better characterization of 1 A version of this chapter has been published. Friesen MC, Demers PA, Spinelli JJ, Le ND. (2006) From expert-based to quantitative retrospective exposure assessment at a Soderberg aluminum smelter. Ann Occup Hyg; 50 359-70. doi:10.1093/annhyg/mel003 12 transitions between exposure categories, accounting for exposure differences between pot lines, and providing finer discrimination between exposure categories. The study update was also an opportunity to quantitatively assess benzo(a)pyrene exposure levels, one component of coal tar pitch volatiles. Benzo(a)pyrene (BaP) has been suggested to be a more specific marker of the carcinogenic potential of the potroom fumes (Theriault et al, 1984; Armstrong et al, 1986; Armstrong et al, 1994; Farant and Gariepy, 1998). The use of non-specific exposure metrics such as CTPV may introduce exposure misclassification, in particular when the causal component may be differentially related to the broader exposure mixture, i.e. by work area. To move from an expert-based to a measurement-based exposure assessment for this study update required combining several approaches to maximize the information from exposure measurements, which we described in this paper. The method used depended on the type of information and availability of exposure measurements for each job. Statistical models to predict annual exposure levels for each job within the potrooms were the foundation of the quantitative exposure assessment. Measurement-based approaches were used wherever possible; however, the use of expert opinion after calibration of their estimates with measurement-based approaches was necessary for estimating exposures for the earliest time periods. With a measurement-based approach, finer discrimination between exposures was possible compared to the original study. The move to a quantitative CTPV and BaP JEM is expected to reduce exposure misclassification so that we may better assess the exposure-response relationships between potroom exposures and cancer incidence and mortality in the 15 year aluminum smelter study update. Methods Description of Facility The updated study cohort consists of 6395 men and 597 women who worked for 3 or more years at a vertical Soderberg aluminum reduction plant in British Columbia (BC), Canada between 1954-2000. Aluminum is produced in carbon-lined steel shells ("pots") by electrolytic reduction at 1000°C of alumina (A1203) to aluminum in molten cryolite (Na 3AlF 6). In a Soderberg smelter the pots are supplied with an anode paste consisting of petroleum coke and coal tar pitch. In the potrooms the paste is baked in situ resulting in the continuous generation of CTPV, a complex mixture that includes known carcinogens such as polycyclic aromatic hydrocarbons. There are approximately 900 1 3 operating pots at this smelter in 7 pot lines. Each pot line represents a group of pots connected in series on the same electrical circuit. The pot lines were grouped (Lines 1-2, 3-5, 7-8) based on construction dates to reflect similarities in technology and building ventilation and to reflect working units used administratively by the smelter. Exposure to coal tar pitch volatiles occurs primarily to operators and maintenance workers in the potrooms, in potlining and potshell repair areas, and the carbon plant where anode paste is prepared. There is negligible exposure to coal tar pitch volatiles in the casting, "wharf/transportation, power operations, and administration departments. A modernization program was implemented at this smelter in 1975 to reduce the generation of potroom fumes and lower worker exposure. Over the next ten years improvements were made in several areas: building ventilation was improved, potroom vehicle and crane cabs were enclosed and cab air filters were improved, double burners and improved gas collection ducts were installed on the pots to capture the emissions, dry anode technology was implemented to reduce emissions at the top of the anode, automated studpulling was implemented to minimize the amount of time workers spent above the pots, crane maintenance bays were installed and improved to separate crane maintenance workers from the potrooms, and a new carbon plant was constructed. Many of these technological improvements took several years to fully implement and each pot line group was updated at different points in time. For instance, dry anode technology was implemented in lines 1-2 in 1985-88, in lines 3-5 in 1979-81 and 1984-87, and in lines 7-8 in 1995-98. Additional details of these technological changes are included in Friesen et al. ( 2003). Exposure Measurements Between 1975 and 2001, 2624 personal CTPV (measured as benzene soluble materials) and 1275 personal BaP exposure measurements were collected by the company and by a regulatory agency. Samples collected by the regulatory agency for compliance purposes accounted for 13% of CTPV measurements and 69% of BaP measurements. Personal exposure measurements were predominantly collected in the potrooms, accounting for 92% of the CTPV measurements and 86% of the BaP measurements. Approximately 180 and 100 CTPV area measurements were collected within the potrooms pre-1975 and 1976-79, respectively. Samples were collected using 37-mm sampling cassettes with fibreglass filters, desorbed with benzene. Company CTPV exposure measurements collected pre-1982 were analyzed by using a moving wire detector (Pye LCM-2) to carry the benzene extract to a flame ionization detector (Alcan Arvida Research Laboratory Method, unknown year). For company measurements after 1982 and all 14 compliance measurements, the benzene was evaporated and the residue was weighed (Alcan Kitimat Laboratories Standard Method 2020, 1983; Workers' Compensation Board of BC Method 3350). Benzo(a)pyrene was analyzed by evaporating an aliquot of the benzene extract to dryness and redissolving the extract with acetonitrile. The acetonitrile aliquot was analyzed by liquid chromatography with a fluorescence detector (Workers' Compensation Board of BC Method 2102). Exposure Assessment Process The exposure assessment strategy consisted of several stages to maximize the use of the exposure measurements (Figure 2.1). While a large number of personal CTPV and BaP measurements had been collected by the company and a regulatory agency from the mid-1970s onwards, exposure measurements were not available for all exposed jobs. For 1977-2000 exposure estimates, these stages included modelling exposures for potroom operations and maintenance jobs, calculating mean exposures for non-potroom locations, and extrapolating exposures for jobs without measurements. Estimating pre-1977 exposure levels involved backwards extrapolation of 1977 exposure levels. These stages are described in detail below. Develop Statistical Models Personal full-shift measurements (minimum duration 6 hours) were used to build statistical models to predict CTPV (1977-2001) and BaP (1975-2001) exposures. Samples less than the detection limit were assigned an exposure 1/V2 times the detection limit (CTPV: n=51; BaP: n=10) (Hornung and Reed, 1990). Separate mixed effects models were developed for potroom operations jobs (CTPV model: n=1868; BaP model: n=773) and for potroom maintenance jobs (CTPV model: n=505; BaP model: n=230). The potroom operations model for CTPV contained 6 jobs: anode assistant, anode operator, controlman, equipment operator, studblast operator, foreman, and cell operator (ref: cell operator & foreman). For the BaP potroom operations model, no exposure estimates for anode assistant could be determined as no measurements were collected on that job. The potroom maintenance model contained 8 jobs: crane maintenance (unspecified trade), gas collection (also known as exhaust maintenance), cell startup operator, stud repair, potroom repairman, electrician, welder, and millwright (reference). A key aspect of the statistical model was modelling the time trend as a linear spline. Using a linear spline allowed for different linear rates of change in each specified time interval while constraining the functions for each interval to meet at the interval boundaries (Harrell, 2001). A linear spline function can be modelled as fixed effects in any linear regression procedure by transforming the 15 independent time variable. For instance, if exposure 7 was to be explained by a linear spline function of a time variable Xwith four time periods with boundaries (knots) of a, b, and c, then three additional variables (X-a), (X-b), and (X-c) would need to be created and offered in the model as fixed effects in addition to variable X(Equation 2.1). Except for the main time effect X, the additional spline terms must be constrained to be positive values or else set to 0. Overall linearity in Xcan be tested by testing H0: fi2 = fi3 = P4 = 0. t Y = Bo + P,X+ B2(X-a, if >0) + %{X-b, if >0) + %{X-c, if >0) [2.1] The time intervals used in the models were based on start and end dates of technological improvements made in the smelter that were previously identified (Friesen et al., 2003). These dates were used to determine time periods of similar technology and periods of transition, with each time period covering two or more years. The implementation dates of technological improvements varied by potroom group, so a separate time trend was modelled for each potroom group, with with 5-6 time periods per potroom group. The time period cut points used in the potroom operations models were: 1981, 1983, 1985, 1987, and 1989 for lines 1-2; 1979, 1983, 1986, and 1988 for lines 3-5; and 1980, 1983, 1995, and 1998 for lines 7-8. In the maintenance models no potroom line specific time trend could be determined as maintenance services was centralized for many time period. Instead, four time periods (cut points: 1979, 1981, 1983, 1985) were specified dividing the major period of technological change into two year periods. Note that the variable 'Year-1977' is the main time effect in all models: we have subtracted 1977 to set the baseline to equal 0, but unlike the additional terms to allow a spline time trend, this main time effect is not constrained to be a positive value. Other variables offered as fixed effects in the models were job, potroom group (Lines 1-2, Lines 3-5, Lines 7-8), season (winter/summer), measurement source (company or regulatory agency), and a job-time period interaction term. A random worker term was included to account for correlation within repeated measurements on the same worker. The models were constructed using the PROC MIXED restricted likelihood method using the natural log of the exposure measurements as the dependent variable (SAS version 8.0, SAS Institute Inc., Cary North Carolina, USA). The model structure is given in equation 2.2, where Yy was the CTPV or BaP exposure level for the fth individual on the/ t h day; Xtj was the log-transformed (base e) exposure level; /?„ was the intercept; Yfid x (determinant of exposure d) was the summation of all model coefficients for the fixed effects, including main effects, interaction terms and linear spline terms, multiplied by the value of the fixed effect (0/1, or continuous variable); b, is the random effect to account for between-worker differences not accounted 16 for by the fixed effects; and £,y is the within-worker differences not accounted for by the fixed effects. Both b, and e,y are assumed to be statistically independent and approximately normally distributed with a mean value of 0 and with variances o 2 «and a2w, respectively. An example of how to apply each model is provided in the footnotes of the tables listing model parameters. ln(Yij) = Xy = P„ + x (determinant of exposure d) + bj + Sy [2.2] r The models were constructed using a manual backwards regression procedure. For the CTPV models, in each iteration the variable with the highest p-value was eliminated, until all variables had a p-value < 0.10. For the BaP models, jobs that had been significant in the CTPV model were kept in the model regardless of p-value because measurements for many jobs were sparse and the aim was to predict exposures rather than determine statistical differences. Residual plots were examined for patterns in unexplained variance. Cook's D was used to identify influential values in the models. The models were used to predict an annual arithmetic mean Y„,k for each job (m) in each year (k, 1977-2000) using equation 2.3, where X mk is the predicted exposure (natural log-transformed) from A 2 A 2 the regression equation, GB is the estimated between-worker variance component, and o~w is the estimated within-worker variance component. Within each time period, the annual arithmetic means for each job were averaged to obtain a time-period specific exposure estimate. A 2 A 2 Ymk =exp[X„,k + 0.5((TB+CTw)] [2.3] Calculation of Means and Extrapolation of Exposures The models did not predict exposures for jobs with no exposure measurements or jobs in work areas outside the potrooms. However, the models were used in conjunction with estimates of the proportion of time spent in exposed areas for each job to assign jobs in the models to one of seven exposure categories for CTPV and BaP for each time period. For CTPV, the four categories used in the original study (None, 0.01-0.2, 0.2-1, >1 mg/m3) were divided for finer discrimination of exposure levels. The seven CTPV categories were none, 0.01-0.09, 0.10-0.19, 0.20-0.39, 0.40-0.99, 1.0-1.9, and >2.0 mg/m3 (measured as benzene soluble materials). For BaP, the categories were determined from observing the range of predicted exposures over time from the BaP exposure models. The seven BaP categories were none, 0.05-0.49, 0.5-0.9, 1.0-2.9, 3.0-6.9, 7.0-13.9, and >14.0ug/m3. 17 For some jobs in non-potroom plant areas (cathode lining, the anode paste plant, and potshell repair) a limited number of exposure measurements were available. The arithmetic mean was calculated from the available personal full-shift measurements and used to determine the appropriate exposure category. For jobs/time periods that had no exposure measurements, each job was compared to the time-specific mean for jobs with measurements with similar tasks but that may have differed in the amount of time spent in exposed areas. For instance, one maintenance job (Job 1) with no measurements that spent 25% of their work day in the potrooms was compared to another maintenance job (Job 2) that spent 75% of their work day in the potrooms with a predicted CTPV exposure of 0.7 mg/m3 for that time period from the CTPV maintenance model. Job 1 was then assigned to the CTPV exposure category "0.2-0.4 mg/m3" (0.7 x 0.25/0.75 = 0.23 mg/m3). The type of tasks and the proportion of time spent in exposed areas for each job and time period were determined from interviews with senior employees. Time period differences occurred primarily in maintenance and service departments as departments were reorganized, centralized, and decentralized at various times, which were represented in the work histories as changes in department names. As in this example, the exposure category assigned to the job was often in a lower exposure category than its comparison job as these jobs spent less time in exposed areas than the jobs with measurements. Jobs with no exposure measurements were primarily maintenance, technician, and engineering jobs that spent only a portion of their time in exposed areas and the remainder of their time in trade shops, offices, or other non-exposed plant areas (i.e. casting, wharf, power operations). Backwards Extrapolation of 1977 Exposure Estimates The time period 1954-76 was a relatively stable technological period at the smelter with few major changes. The original exposure assessment identified two major improvements that would be expected to decrease exposure levels: 1) improvements to the pot gas collection system (gas skirts and gas burners) in 1960; and 2) an expansion of the carbon plant in 1966, which included ventilation and exhaust improvements. There were no measurements available for this time period aside from a limited number of area measurements in the early 1970s. The updated measurement-based 1977 exposure levels were used to calibrate the original expert-based exposure estimates. We then applied the relative changes in exposure levels from the original assessment to assign exposure categories for each job and time period pre-1977. 18 Results The fixed effects from the potroom operations and potroom maintenance models explained approximately 45% and 40% of the variability in the data for CTPV and 27% and 19% for BaP, respectively (Tables 2.1-2.2). The source of measurement (company or regulatory agency) was not significant in the models. Season was only significant for the BaP potroom operations model. For CTPV four to five distinct time periods with different rates of change in exposure were identified for each potroom group for potroom operations jobs whereas only two distinct time periods were identified for BaP exposures. A short-term peak in BaP exposures in 1981 is accounted for by the use of a specific pitch type in the smelter that was only in use for one month; this pitch had a substantially higher level of BaP than other pitch types. This peak was not observed in the maintenance model as no measurements of maintenance workers were collected in during the one month this pitch was used. No statistical differences between electricians, welders, and millwrights could be determined due to the large variability in exposures. Similarly, maintenance job differences pre-1981 were not observed, although job differences in later time periods were observed with the inclusion of job-time period interaction terms. Other pitch differences could not be detected in the BaP models due to the use of year in the models and the lack of overlap between pitch types. For both CTPV and BaP the total variance was smaller for the operations models than the maintenance models. For CTPV, the between- and within-worker variance components were greater for the maintenance model (0.328 and 0.489, respectively) than for the operations model (0.087 and 0.371). For BaP, the between-worker variance was not estimatable (~0) in the potroom operations model due to very identifiable few repeated measures on the same worker, but the within-worker variance was large (1.27). The BaP maintenance model had much larger between-worker (1.67) and within-worker (0.624) variance components than the other models. Total variance was much greater in the BaP models than the CTPV models. The between-worker variance component contributed more to the total variance in the maintenance models than in the operations models. The models' predicted CTPV and BaP exposure levels for pot operator and potroom maintenance workers are displayed in Figures 2.2 and 2.3, respectively. Since the time trend was modelled on a log-transformed scale, the linear trend for each time period in the model translates to an exponential change once retransformed to the original exposure units. To account for the seasonal differences in BaP exposure levels for potroom operations jobs, the annual exposure level was weighted by the proportion of months in each season (5/12 * summer exposure levels + 7/12 * winter exposure levels). 19 The updated CTPV and BaP job exposure matrices for all jobs are presented in Table 2.3. For the epidemiological analyses, cumulative exposure was calculated. For jobs not in the models the midpoints of the exposure categories were used. The highest exposure categories were assigned a value of 2.5 mg/m3 for CTPV exposure and 18 ug/m3 for BaP exposure, respectively. For the time period 1977-2000, the models were used to estimate exposure for approximately 57% of the exposed personryears and accounted for nearly all of the high and moderately exposed jobs at the smelter. Direct calculation of exposure was used for 6% of the exposed-person-years. The remaining 37% of the exposed person-years required adjusting the model-based exposure estimates based on proportion of time spent in the potrooms, and represented jobs with low exposures. Substantial changes occurred in the 1977-2000 CTPV estimates compared to the original estimates. These differences occurred because we better characterized when exposure changes occurred through the use of the linear spline time trends and because we used more exposure categories. Discrepancies between a measurement-based and the original CTPV exposure estimates for this time period were previously reported in a validation study of the original exposure estimates (Friesen et al, 2003). For the most part, the updated CTPV exposure estimates for 1954-76 did not differ substantially from the original exposure estimates. Small differences occurred due to the finer discrimination in exposure categories and the resulting change in category midpoints (i.e. 0.6 mg/m3 became 0.7 mg/m3; 0.1 became 0.15 or 0.05 mg/m3 in updated study). Whereas the maximum exposure assigned in the original study was 1.5 mg/m3, 2.5 mg/m3 was now assigned to two jobs: anode assistant, 1954-76; gas collection, 1954-60. A few jobs originally assigned to the medium exposure category (0.2-1.0 mg/m3) were now assigned to the 0.2-0.4 mg/m3 CTPV category, reducing their assigned exposure from 0.6 to 0.3 mg/m3; these jobs included mtce/tech-moderate, cathode lining, and carbon plant maintenance. CTPV exposure estimates for 1977 for anode operators (0.3 mg/m3) and studblast operators (0.7 mg/m3) suggested that their original exposure estimates (0.1 and 0.6 mg/m3, respectively) for 1954-1960 were too low, thus their updated exposure estimates were increased to 0.3 and 1.0 mg/m3 CTPV, respectively. Discussion The advantage of a statistical model over direct calculation of means is its ability to borrow information across jobs and time periods to predict exposures where direct measurements were not 20 available (Stewart et al, 1996). However, due to limitations in the type of determinants offered in models, exposure levels for some jobs or time periods may not be directly obtained from predictive models. In this study, exposure estimates from statistical models served as the foundation of the exposure assessment methodology and were used to directly estimate exposures for the moderate and highly exposed potroom jobs. The models were also used indirectly to extrapolate exposure levels for the lower exposed jobs that spend only a fraction of their time in exposed areas and to calibrate expert exposure ratings for the earliest time periods. ' A key aspect of the models developed here for predicting exposures for a retrospective health study was the use of a linear spline time trend. Calendar year is often used as a surrogate for technological and administrative changes not captured by available exposure determinants. Most commonly year is offered as a continuous variable which assumes a linear change in the log-transformed exposure with time (Symanski et al, 1998; Symanski etal, 1998; Burstyn etal, 2000). Technological changes entered as dichotomous variables (yes/no) are often treated as having an "instantaneous" impact on exposure levels. More realistically, technological improvements are incorporated over time and there may be periods of stability between periods of technological change. Production and economic factors may also cause nonlinear exposure changes with time (Symanski et al, 1998). Nonlinear time trends have previously been accounted for by the use of quadratic terms (Hornung et al, 1994) or by offering categorical time periods in the model that provides step-wise changes in exposure but do not allow information to be borrowed from adjoining time periods (Dement et al, 1983). One previous study has used a linear spline term to describe the time trend in exposures for retrospective exposure assessment (Raaschou-Nielsen et al., 2002). In the case of technological improvements in this aluminum smelter, technological improvements occurred over a period of years. To account for varying rates of exposure changes over time, we instead used information available on technological changes to define time periods where technology was similar or under transition and then modelled the exposure time trend for each time period while constraining the exposures to be equal at the interval boundaries. This has the added advantage of borrowing information from adjoining time periods, and while more flexible than incorporating a linear, quadratic, or categorical time trend, it also required several assumptions. Our approach assumed that all jobs in the model have the same proportional decrease in exposure within each time period (and by pot line group for the potroom operations model). In the potroom operations model the time trend was driven by the exposure levels of pot operators, a job that accounts for one-third of all exposed person-years. Pot operators were sampled annually and accounted for approximately half 21 of all exposure measurements. We were unable to estimate a job-specific time trend and thus we may have missed some job-specific impacts of technological changes, but because of the open and large environment of the potrooms this assumption is reasonable and was necessary given insufficient samples and significant gaps between sampling campaigns for most jobs. The models presented here assumed constant variance over the time periods, pot lines, and jobs; however, ignoring changes in the variance'components would impact the arithmetic means. We tested this assumption by allowing heterogeneity in the variance components by time period in the models and also by visually inspecting plots of the geometric standard deviations for jobs over time, but we observed no substantive differences between pot lines within the same job or over time (not shown). However, we did observe that maintenance jobs had much greater variability than operations jobs as has previously been reported by Kromhout et al. (1993), which is not surprising due to the non-routine nature and varying locations of their tasks. As such, we chose to develop the maintenance jobs and operations jobs models separately. The time trends and variance components were substantially different for the maintenance jobs models compared to the operations jobs models, confirming that it is important to explore possible heterogeneity of variance in models developed for retrospective exposure assessment. The proportion of variability in the data explained by the models described here was typical of models developed for retrospective exposure assessment (Burstyn and Teschke, 1999). The BaP measurements were much more variable than the CTPV (dataset GSD 5.1 vs. 2.8, respectively). The resulting BaP models explained a much smaller proportion of the variability in the data than the CTPV models. Fewer distinctions between rates of exposure change by time periods were observed and the predicted changes in BaP exposure were generally less steep than for CTPV. The cause of the increased variability in BaP exposure levels is not certain; however, some potential but untested explanations include differences in analytical techniques between company and compliance samples (though source of measurement was not significant in the models) and BaP is known to less stable than some other polycyclic aromatic hydrocarbons (Aubin and Farant, 2000). The drawback of using jobs as predictors in the models is that they cannot directly predict exposure levels for jobs where no measurements were present. An alternative option would have been to identify process conditions, control measures, and specific job tasks that influence exposure, such as was done by Romundstad et al (Romundstad et al, 1999). Using more specific determinants of exposure could increase the proportion of variability explained by the fixed effects in the model. We 22 were limited in our ability to do so as operating condition and job task information were not collected alongside the exposure measurements, though some parameters could be extracted from plant records. Additionally, to be useful potential exposure determinants need to be available for the entire length of the retrospective study and exposure measurements must exist for every combination of process parameters. Rather, our goal was to predict average annual personal exposures for each job. Thus, using dates of technologically similar time periods and surrogates such as job was sufficient to borrow information across jobs and time periods to estimate average annual exposures for the health study, but would be insufficient for making decisions about control measures. The models were limited to predicting the time trend for years where measurements were available. Back extrapolation to earlier periods requires some knowledge of the shape of the time trend. Limiting the use of predictive models to specific time ranges is not uncommon. For example, Hornung et al (Hornung et al., 1994) limited the predictive model for ethylene oxide exposures to 1978 forward and for earlier time periods assumed a steady exposure level equivalent to 1978 levels. The exposure measurements covered the majority of the technological improvements at this smelter. The smelter modernization program began in 1975 so some continuation of the exposure decrease was expected between 1975 and 1977 and is supported by area measurements collected in the early 1970s. Prior to 1975 few major changes took place. The impact of these changes on exposure is impossible to calculate given the lack of exposure measurements; thus, decisions made by the original study's exposure assessment committee were used after calibration of their exposure estimates to the measurement-based 1977 exposure levels. With this calibration, and the finer discrimination provided with seven compared to the original four categories, we expect these exposure estimates to be improved over the original estimates. Even when exposure measurements are available many assumptions are inherent in quantitative exposure assessment for retrospective studies. Yet quantitative exposure estimates have been found to more often yield larger risk estimates and sharpened exposure-response gradients suggesting that exposure misclassification is being reduced (Blair and Stewart, 1992). Changes in the exposure-response relationship when moving from an expert-based semi-quantitative exposure assessment to quantitative exposure estimates using statistical models as its foundation will be examined in the aluminum smelter study update. In addition, the use of benzo(a)pyrene as a more specific marker of exposure will be assessed. 23 Table 2.1 Model coefficients (P) and standard errors (SE) for CTPV and BaP potroom operations mixed effects models developed to predict exposures for 1977-2000 CTPV Model BaP Model Model Effect a b e 3 SE 3 SE Intercept -0.626 0.069 0.750 0.090 (year: 1977, pot operators, foremen) Potlines 1-2 x Year - 1977 -0.274 0.025 -0.200 0.037 x Year - 1981, if >0C 0.255 0.045 0.171 0.045 x Year - 1985, if >0 -0.184 0.073 ~ ~ x Year - 1987, if >0 0.466 0.099 ~ — x Year - 1989, if >0 -0.356 0.070 — — Potlines 3-5 x Year - 1977 -0.534 0.056 -0.177 0.025 x Year - 1979, if >0C 0.484 0.080 — — x Year - 1983, if>0 -0.136 0.051 0.137 0.036 x Year - 1986, if >0 0.160 0.037 — — x Year - 1988, if >0 — — — — Potlines 7-8 x Year - 1977 — — -0.325 0.033 x Year - 1980, if >0C -0.630 0.034 ~ — x Year- 1983, if >0 0.620 0.040 0.308 0.047 x Year - 1995, if >0 -0.101 0.050 — — x Year - 1998, if >0 — — — — Anode Assistant 1.342 0.122 d d Anode Operator 0.215 0.075 0.077 0.126 x 'Year>1985' -0.592 0.121 -0.964 0.265 Controlman -0.353 0.114 -0.207 0.220 x 'Year >1985' -1.144 0.219 -1.535 0.462 Equipment Operator -0.212 0.103 -0.781 0.221 Studblast Operator 0.388 0.109 -0.425 0.279 x 'Year > 1985' -0.640 0.189 — — Winter ~ — 0.385 0.091 Pitch Type: 100% Koppers - - 2.208 0.183 "All variables except year were entered as dichotomous variables (0=No/l=Yes). b Variables starting with an "x" are interaction terms that apply only when the value of the preceeding main effect is "1". c In linear spline time trends the coefficients for the second term through remaining terms are constrained to apply only when the variable (year - boundary) is a positive value. d No BaP measurements were available for this job, so job was not entered in the model. e Example calculation: Arithmetic mean CTPV exposure for Anode Operator in 1987 in lines 1-2: log-transformed exposure = intercept + p(Lines 1-2, Year-1977)+P(Lines 1-2, Year-1981, if>0)+p(Lines 1-2, Year-1985, if>0)+P(Lines 1-2, Year-1987, if>0)+ P(Lines 1-2, Year-1989, it>0)+ p(AnodeOperator)+ P(Anode Operator x Year> 1985) = -0.626-0.274( 1987-1977, ifi>0)+0.255( 1987-1981, if>0)-0.184(1987-1985, if>0)+0.466(1987-1987, if>0)-0.356(l987-1989, if >0)+0.215-0.592 = -2.58; Arithmetic Mean=exp[log-transformed exposure + 0.5(o2B+o2w)]= exp[-2.58+0.5(0.087+0.371)]= 0.095mg/m3 24 Table 2.2 Model coefficients (P) and standard errors (SE) for the CTPV and BaP potroom maintenance mixed effects models developed to predict exposures for 1977-2000 CTPV Model BaP Model Variable3"" P SE p SE Intercept 0.549 0.339 0.067 0.141 (year: 1977, Millwright) Year-1977 ~ — -0.553 0.101 Year-1979, if> 0C -1.478 0.176 Year-1981, if > 0 1.434 0.177 Year-1983, if > 0 ~ ~ 0.548 0.109 Year-1985, if> 0 — — Crane Maintenance (Trade not specified) -1.444 0.381 -Gas Collection ~ ~ x Year within 1981-85 0.295 0.157 -x Year> 1986 1.022 0.154 1.179 0.260 Cell Startup Operator - - 1.098 0.528 x Year within 1981-85 1.013 0.216 x 'Year> 1986' 0.610 0.285 Stud Repair - - --x Year within 1981-85 1.047 0.294 Potroom Repairman — — Potroom Repairman x 'Year> 1986' 0.307 0.144 Electrician - -Welder — — Crane Lube & Serviceman - ~ ~ a All variables except year were entered as dichotomous variables (0=No/l=Yes). b Variables starting with an "x" are interaction terms that apply only when the value of the preceeding main effect is "1". 0 In linear spline time trends the coefficients for the second term through remaining terms are constrained to apply only when the variable (year - boundary) is a positive value. d Example calculation: Arithmetic mean CTPV exposure for Cell Startup Operator in 1987: log-transformed exposure = intercept +|3(Year-1979, if>0)+P(Year-1981, if>0)+ P(Cell Startup Operator x Year within 1981-85) + P(CelI Startup Operator x Year > 1986) = 0.549-1.478(1987-1979, if >0)+1.434(1987-1981, if >0)+1.013(0)+0.610(l)=-2.061; Arithmetic Mean=exp[log-transformed exposure + 0.5(o2B+o2w)]= exp[-2.061+0.5(0.328+0.489)]= 0.19 mg/m3 25 Table 2.3 Updated CTPV (mg/m3) and BaP (ug/m3) Job Exposure Matrices by Job and Time Period a Work Area/Job CTPV BaP Potrooms 1-2 54-59 60-76 77-80 81-85 86-00 54-59 60-76 77-80 81- 86-85 00 Anode Assistant (M) 2.5 2.5 1.8 0.8 - 18 18 11 11 Controlman (M) 0.7 0.7 0.3 0.15 0.05 5 5 2.7 1.8 0.3 Studblast Operator (M) 1 1 0.7 0,3 0.12 5 5 2.2 1.4 0.9 Anode Operator (M) 0.3 0.3 0.3 0.3 0.1 5 5 3.6 2.3 0.6 Equipment Operator (M) 0.7 0.7 0.4 0.2 0.12 3 3 1.5 1 0.6 Pot Operator & Foreman (M) 1.5 1.5 0.5 0.2 0.15 7 7 3.3 2.2 1.4 Potrooms 3-5 54-59 60-76 77-78 79-85 86-00 54-59 60-76 77-78 79- 86-85 00 Anode Assistant (M) 2.5 2.5 2 0.7 ~ 18 18 11 11 — Controlman (M) 0.7 0.7 0.4 0.13 0.05 5 5 3.3 1.8 0.2 Studblast Operator (M) 1 1 0.8 0.3 0.07 5 5 2.6 1.5 0.7 Anode Operator (M) 0.3 0.3 0.3 0.2 0.06 5 5 4.4 2.4 0.4 Equipment Operator (M) 0.7 0.7 0.4 0.16 0.07 3 3 1.8 1 0.5 Pot Operator & Foreman (M) 1.5 1.5 0.5 0.2 0.09 7 7 4 2.2 1 Potrooms 7-8 54-59 60-76 77-80 81-82 83-00 54-59 60-76 77-80 81- 83-82 00 Anode Assistant (M) 2.5 2.5 2 1.1 - 18 18 11 11 __ Controlman (M) 0.7 0.7 0.05 0.2 0.05 5 5 2.3 1.2 0.2 Studblast Operator (M) 1 1 1 0.4 0.08 5 5 1.9 1 0.4 Anode Operator (M) 0.3 0.3 0.3 0.3 0.07 5 5 3.1 1.6 0.3 Equipment Operator (M) 0.7 0.7 0.5 0.2 0.07 3 3 1.3 0.69 0.3 Pot Operator & Foreman (M) 1.5 1.5 0.7 0.3 0.09 7 7 2.9 1.5 0.5 Other Potroom Jobs 54-59 60-76 77-80 81-85 86-00 54-59 60-76 77-80 81- 86-85 00 Potroom Training 0.7 0.7 0.3 0.15 0.05 5 5 2 2 0.7 Planner/Manager/Toolroom 0.15 0.15 0.15 0.05 0.05 0.7 0.7 0.7 0.2 0.2 Cell Startup Operator (M) 1.5 0.7 0.6 0.3 0.14 11 5 1.5 2.7 2.5 Potroom Labourer 0.7 0.7 0.7 0.3 0.15 5 5 2 2 0.7 Crustbreaker, Ore Truck 1.5 1.5 - - ~ 7 7 Potroom & Crane Mtce (M) 0.7 0.7 0.6 0.12 0.1 5 5 1.5 0.9 0.9 Gas Collection (M) 2.5 1.5 0.6 0.15 0.2 11 5 1.5 0.9 2.8 Potliner/Equipment Op (M) 0.7 0.7 0.7 0.12 0.08 5 5 2 2 0.7 Stud Repair (M) 0.7 0.7 0.6 0.3 0.08 2 2 0.7 0.7 0.2 Duclaux Installman ~ 0.7 0.7 0.7 0.7 - 11 11 11 11 Mtce/Tech-Moderateb 0.3 0.3 0.3 0.05 0.05 5 5 2 0.7 0.7 Mtce/Tech-Infrequentb 0.15 0.15 0.15 0.05 0.05 2 2 0.7 0.2 0.2 Mtce/Tech-Minimal b, 0.05 0.05 0.05 0.05 0.05 0.7 0.7 0.2 0.2 0.2 Drivers/Janitors/Warehouse Non-Potroom Areas Potshell Repair (C) 0.15 0.15 0.1 0.1 0.1 2 2 0.7 0.7 0.7 26 Craneman (non potroom) 0.15 0.15 0.05 0.05 0.05 0.7 0.7 0.2 0.2 0.2 Coke Calcining Op (C) 0.05 0.05 0.05 0.05 0.05 0.2 0.2 0.2 0.2 0.2 Cathode Lining (C) 0.3 0.3 0.15 0.15 0.05 2 2 0.7 0.7 0.2 Carbon Plant 54-66 67-76 77-82 83-00 54-66 67-76 77-82 8 3 -00 Maintenance (C) 0.3 0.15 0.15 0.15 5 5 5 2 Hot Paste Operator 1.5 - - - 18 -- -- --Dryer/Grinder/Furnace Op 0.15 0.15 0.15 - 2 2 2 ~ Plant Operators (C) 0.7 0.3 0.3 0.05 11 5 5 0.2 Potlining Mix Manufacturing 0.15 0.15 0.05 0.05 2 2 2 2 a Point estimates are listed for jobs in models. Category midpoints are listed for jobs not in models. CTPV categories: 0, 0.01-0.09, 0.10-0.19, 0.20-0.39, 0.40-0.99, 1.0-1.9, and >2.0 mg/m3. BaP categories: 0, 0.05-0.49, 0.5-0.9, 1.0-2.9, 3.0-6.9, 7.0-13.9, and >14.0 pg/m3. The highest exposure categories were assigned a value of 2.5 mg/m3 for CTPV and 18 ug/m3 for BaP. b Refers to amount of time spent in potroom. Moderate 20-50%; Infrequent 5-20%; Minimal < 5%. Includes trades, supervisors, engineers and technicians. (M) Exposure estimates from 1977-2000 were derived from statistical models. Annual predicted exposures were averaged over the specified time periods for these jobs to obtain the time-period specific estimate. (C) Some personal exposure measurements were available for this job/work area for the later time periods. 27 Figure 2.1 Multi-component procedure to develop quantitative CTPV and BaP exposure estimates C T P V & B a P p e r s o n a l e x p o s u r e m e a s u r e m e n t s , 1 9 7 7 - 2 0 0 0 ^~ Or ig ina l exper t -b a s e d C T P V e x p o s u r e matr ix , 1 9 5 4 - 1 9 8 8 v Va l i da t i on of exper t - b a s e d C T P V es t ima tes , 1 9 7 7 - 1 9 8 8 ' D e v e l o p stat is t ica l m o d e l s 1 9 7 7 - 2 0 0 0 [Point Es t ima tes ] T C a l c u l a t i o n of m e a n s & ex t rapo la t ion 1 9 7 7 - 2 0 0 0 [Ca tego r i ca l Es t ima tes ] T B a c k w a r d s ex t rapo la t ion f rom 1977 es t ima tes 1 9 5 4 - 2 0 0 0 [Ca tego r i ca l Es t ima tes ] T C T P V & B a P e x p o s u r e e s t i m a t e s ' / ' 1 9 5 4 - 2 0 0 0 ; * Results reported in Friesen et al (2003). 28 Figure 2.2 Predicted annual mean exposures for 1977-2000 from the CTPV model (x) and BaP model (•) for pot operators by potroom group 0.8 0.7 -0.6 -0.5 0.4 0.3 0.2 0.1 Potlines 1-2 • X x * • • • • ******xx| 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 - 0.0 m Potlines 3-5 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 O L TO C O 0.8 -, 0.7 0.6 -I Potlines 7-8 Xxxx , xxxxxxxxxxxxx x - 8.0 - 7.0 - 6.0 - 5.0 - 4.0 - 3.0 |- 2.0 1.0 - 0.0 o. ro m Y e a r 29 Figure 2.3 Predicted mean exposures for 1977-2000 from the CTPV model (x) and BaP model (•) for potroom repairman 30 References Armstrong B, Tremblay C, Baris D, Theriault G (1994) Lung cancer mortality and polynuclear aromatic hydrocarbons: A case-cohort study of aluminum production workers in Arvida, Quebec, Canada. Am J Epidemiol; 139 250-62. Armstrong BG, Tremblay C G , Cyr D, Theriault GP (1986) Estimating the relationship between exposure to tar volatiles and the incidence of bladder cancer in aluminum smelter workers. Scand J Work Environ Health; 12 486-93. Astrakianakis G, Anderson J, Band PR, Keefe A, Bert J, Le N, Fang R (1998) Job-exposure matrices and retrospective exposure assessment in the pulp and paper industry. Appl Occup Environ Hyg; 13 663-70. Aubin S, Farant JP (2000) Benzo[b]fluoranthene, a potential alternative to benzo[a]pyrene as an indicator of exposure to airborne pahs in the vicinity of Soderberg aluminum smelters. J Air Waste Manag Assoc; 50 2093-101. Blair A, Stewart PA (1992) Do quantitative exposure assessments improve risk estimates in occupational studies of cancer? Am J Ind Med; 21 53-63. Burstyn I, Kromhout H, Kauppinen T, Heikkila P, Boffetta P (2000) Statistical modelling of the determinants of historical exposure to bitumen and polycyclic aromatic hydrocarbons among paving workers. Ann Occup Hyg; 44 43-56. Burstyn I, Teschke K (1999) Studying the determinants of exposure: A review of methods. Am Ind Hyg Assoc J; 60 57-72. de Cock J, Kromhout H, Heederik D, Burema J (1996) Experts' subjective assessment of pesticide exposure in fruit growing. Scand J Work Environ Health; 22 425-32. Dement JM, Harris RL, Jr., Symons MJ, Shy C M (1983) Exposures and mortality among chrysotile asbestos workers. Part I: Exposure estimates. Am J Ind Med; 4 399-419. Eisen EA, Smith TJ, Wegman DH, Louis T A , Froines J (1984) Estimation of long term dust exposures in the Vermont granite sheds. Am Ind Hyg Assoc J; 45 89-94. Farant JP, Gariepy M (1998) Relationship between benzo(a)pyrene and individual polycyclic aromatic hydrocarbons in a Soderberg primary aluminum smelter. Am Ind Hyg Assoc J; 59 758-65. Friesen M C , Demers PA, Spinelli JJ, Le ND (2003) Validation of a semi-quantitative job exposure matrix at a Soderberg aluminum smelter. Ann Occup Hyg; 47 477-84. Harrell FEJ (2001) Regression modeling strategies: With applications to linear models, logistic regression, and survival analysis: Springer-Verlag New York, Inc. 31 Hawkins NC, Evans JS (1989) Subjective estimation of toluene exposures: A calibration study of industrial hygienists. Appl Ind Hyg; 4 61-68. Hertzman C, Teschke K, Dimich-Ward H, Ostry A (1988) Validity and reliability of a method for retrospective evaluation of chlorophenate exposure in the lumber industry. Am J Ind Med; 14 703-13. Hornung RW, Greife A L , Stayner LT, Steenland NK, Herrick RF, Elliott LJ, Ringenburg V L , Morawetz J (1994) Statistical model for prediction of retrospective exposure to ethylene oxide in an occupational mortality study. Am J Ind Med; 25 825-36. Hornung RW, Reed L D (1990) Estimation of average concentration in the presence of non-detectable values. Appl Occup Environ Hyg; 5 46-51. Kauppinen T, Teschke K, Astrakianakis G, Boffetta P, Colin D, Keefe A, Korhonen K, Liukkonen T, Nicol A M , Pannett B, Westberg H (2002) Assessment of exposure in an international study on cancer risks among pulp, paper, and paper product workers. Am Ind Hyg Assoc J; 63 254-61. Kromhout H, Oostendorp Y, Heederik D, Boleij JS (1987) Agreement between qualitative exposure estimates and quantitative exposure measurements. Am J Ind Med; 12 551-62. Kromhout H, Swuste P, Boleij JS (1994) Empirical modelling of chemical exposure in the rubber-manufacturing industry. Ann Occup Hyg; 38 3-22. Kromhout H, Symanski E, Rappaport SM (1993) A comprehensive evaluation of within- and between-worker components of occupational exposure to chemical agents. Ann Occup Hyg; 37 253-70. Post W, Kromhout H, Heederik D, Noy D, Duijzentkunst RS (1991) Semiquantitative estimates of exposure to methylene chloride and styrene: The influence of quantitative exposure data. Appl Occup Environ Hyg; 6 197-204. Raaschou-Nielsen O, Hansen J, Thomsen BL, Johansen I, Lipworth L, McLaughlin JK, Olsen JH (2002) Exposure of Danish workers to trichloroethylene, 1947-1989. Appl Occup Environ Hyg; 17 693-703. Romundstad P, Haldorsen T, Ronneberg A (1999) Exposure to PAH and fluoride in aluminum reduction plants in Norway: Historical estimation of exposure using process parameters and industrial hygiene measurements. Am J Ind Med; 35 164-74. Spinelli JJ, Band PR, Svirchev L M , Gallagher RP (1991) Mortality and cancer incidence in aluminum reduction plant workers. J Occup Med; 33 1 150-5. Stewart PA, Lees PS, Francis M (1996) Quantification of historical exposures in occupational cohort studies. Scand J Work Environ Health; 22 405-14. 32 Symanski E, Kupper L L , Hertz-Picciotto I, Rappaport SM (1998) Comprehensive evaluation of long-term trends in occupational exposure: Part 2. Predictive models for declining exposures. Occup Environ Med; 55 310-6. Symanski E, Kupper LL, Rappaport SM (1998) Comprehensive evaluation of long-term trends in occupational exposure: Part 1. Description of the database. Occup Environ Med; 55 300-9. Teschke K, Hertzman C, Dimich-Ward H, Ostry A, Blair J, Hershler R (1989) A comparison of exposure estimates by worker raters and industrial hygienists. Scand J Work Environ Health; 15 424-9. Teschke K, Olshan AF, Daniels JL, De Roos AJ, Parks C G , Schulz M , Vaughan T L (2002) Occupational exposure assessment in case-control studies: Opportunities for improvement. Occup Environ Med; 59 575-93. Theriault G, Tremblay C, Cordier S, Gingras S (1984) Bladder cancer in the aluminium industry. Lancet; 1 947-50. Yu RC, Tan W-Y, Mathew RM (1990) A deterministic mathematical model for quantitative estimation of historical exposure. Am Ind Hyg Assoc J; 51 194-201. 33 Chapter 3: Adequacy of benzo(a)pyrene and benzene soluble materials as indicators of exposure to polycyclic aromatic hydrocarbon mixtures2 Introduction Occupational and environmental exposure to polycyclic aromatic hydrocarbons (PAH) occurs as complex mixtures as a result of incomplete combustion. The specific PAHs present are dependent on raw materials and process conditions. Several PAHs have been classified as carcinogens based on animal studies (NTP, 2005; Straif et al, 2005). An increased cancer risk has been well established in workers exposure to coal derived PAH mixtures, such as in aluminum smelters and coke ovens, (Straif et al., 2005), but much uncertainty remains in the exposure-response relationship and the specific roles of individual PAHs are not known (Boffetta et al, 1997; Armstrong et al, 2004; Straif et al, 2005). Exposure measurements of benzene soluble materials (BSM, a measure of the PAH mixture) and benzo(a)pyrene (BaP, a specific particulate-phase PAH) account for the bulk of the measurements evaluating PAH exposure in aluminum smelters (Benke et al, 1998). As a result, BaP and BSM have been the primary exposures evaluated in epidemiologic studies. There have been insufficient measurements collected for other components of the PAH mixture to evaluate those components in retrospective studies thus far. Using BaP or BSM in epidemiologic studies assumes that their relationship with the toxic potential of the mixture remains constant across the scope of the study. However, there is evidence that the relative distribution of the specific PAHs in the PAH mixtures varies due to smelter technology, anode paste composition, work shifts, and jobs (Farant and Gariepy, 1998; Sanderson et al, 2005; Armstrong etal, 1986; Armstrong et al, 1994). Knowing how these commonly assessed measures, BaP and BSM, are related to other components may allow for retroactively assessing exposure to these other components (Farant and Gariepy, 1998). It may also facilitate comparisons of exposure-response relationships across studies and industries. In this study we examined time-specific factors (anode paste composition, season) and work area-specific factors (potroom building, job) that influence the relative distribution of the P A H 2 A version of this chapter has been submitted for publication. Friesen MC, Demers PA, Spinelli JJ, Le ND. Adequacy of benzo(a)pyrene and benzene soluble materials as indicators of exposure to polycyclic aromatic hydrocarbon mixtures. (Submitted) 34 components within an aluminum smelter using personal measurements collected over 22 years. To evaluate differences in the relative contribution of components, BaP relative abundance ratios (BaP RARs), a ratio of the concentration of a specific PAH divided by the concentration of BaP, were used in regression analyses to facilitate comparisons between locations and time periods that may differ greatly in concentration (Farant and Gariepy, 1998; Sanderson and Farant, 2000; Sanderson et al., 2005). We also calculated a measure of the carcinogenicity of the mixture by applying factors to each PAH that represent their relative toxicity compared to benzo(a)pyrene; the relative contributions of each PAH were then summed to obtain a BaP equivalency factor. Methods Description of Facility This vertical-stud Soderberg aluminum smelter consists of over 900 operating pots located in seven pot lines, with each line representing a group of pots connected in series on the same electrical circuit. The seven pot lines are grouped for both statistical analysis and within the smelter for work organization purposes; these groupings represent real differences across the smelter in technology, process conditions, and building properties (i.e. ventilation, gas collection). A modernization program was implemented in 1975 to reduce exposure. Over the next ten years improvements that reduced exposure levels by over 80% (Chapter 2) were made in several areas, including ventilation and gas collection technology, coal tar pitch sources, implementation of dry anode technology, enclosure of potroom vehicles and cranes, and automated studpulling to minimize workers' time spent above the pots. An overview of the technological changes implemented at the smelter has been previously reported (Friesen et al., 2003). The coal tar pitch used to form the anode paste varied over the smelter's history (Table 3.1). The pitch mixtures were grouped into 6 different pitch types by their composition; no measurements were available for the Sumitomo coal tar pitch. There is conflicting information on what pitch was used prior to 1980, with both V F T and Kosco mentioned. From 1980-82, different pitch combinations were tried. From 1983 onwards a mixture of Kosco and V F T pitch was used. We assumed that the coal tar pitch type remained constant during the gaps in the records (1989-92, 1994-98). Exposure Measurements Between 1978 and 2001, 576 measurements analyzed for both BSM and BaP were collected by the company and by a regulatory agency (duration: 241-708 minutes, median 454 min). In the same time 35 period, 479 measurements were analyzed for BaP and a selection of other particulate PAHs (duration: 85-841 minutes, median 457 min.). The specific PAHs that were analyzed changed overtime and depended on whether collected by the company or the regulatory agency, with 9 particulate-phase PAHs measured regularly. Samples collected by the regulatory agency for compliance purposes accounted for 32% of the BaP/BSM measurements and 89% of the PAH particulate measurements. Measurements were predominantly collected in the potrooms, accounting for 90%> of both the BaP/BSM mea'surements and the PAH particulate measurements. Samples were collected using 37-mm sampling cassettes with fibreglass filters, desorbed with benzene. All measurements were personal samples, worn by an individual worker on his/her shoulder or lapel area. Company BSM exposure measurements collected pre-1982 were analyzed using a moving wire detector (Pye LCM-2) to carry the benzene extract to a flame ionization detector (Alcan Arvida Research Laboratory Method, unknown year). For company measurements after 1982 and all compliance measurements, the benzene was evaporated and the residue was weighed (Alcan Kitimat Laboratories Standard Method 2020, 1983; Workers' Compensation Board of BC Method 3350). Benzo(a)pyrene and other particulate PAHs were analyzed by evaporating an aliquot of the benzene extract to dryness and redissolving the extract with acetonitrile. The acetonitrile aliquot was analyzed by liquid chromatography with a fluorescence detector (Workers' Compensation Board of B C Method 2102). Statistical Analyses The correlation between BSM and BaP, and BaP and other particulate PAH exposures were examined. The relationships between BaP and other particulate PAHs were examined using the BaP relative abundance ratios (BaP RAR), where the exposure level of one PAH was divided by the BaP exposure level (unitless). The relationship between BaP and BSM was examined using the BaP/BSM ratio, which was calculated with BSM exposure levels in the denominator (BaP/BSM ratio, p.g BaP:mg BSM), to be consistent with previously reported BaP/BSM ratios (Armstrong et al, 1986; Armstrong et al, 1994). Since we were examining ratios, only measurements with exposure levels greater than the limit of detection were included (BaP/BSM: 47 excluded; PAH/BaP varied by PAH) (Farant and Gariepy, 1998). Group differences in the ratios were examined using Kruskal-Wallis non-parametric tests. The dependent variables, BaP/BSM ratio and BaP RARs, were log-transformed (base e) for all linear regression analyses as they were positively skewed. Linear regression was used to examine the 36 factors potentially influencing the BaP/BSM ratio and the BaP RARs. Factors examined were season (Summer: May-October; Winter: November-April), coal tar pitch type (Table 3.1), pot line group (Lines 1-2, 3-5, 7-8), job, and work area. Interactions between coal tar pitch type and pot line group and between job group and pot line group were also examined. A potential explanatory variable was only offered into the model if it was associated with the dependent variable in simple linear regression (p-value < 0.20). The models were developed using a manual backwards regression procedure: at each iteration the factor with the highest p-value was removed, until all factors had p-values < 0.10. BaP Equivalency Factor (BEF) We calculated the BaP equivalency factors (BEF) in this study by multiplying the BaP RAR by the relative potency factors that were calculated for aluminum smelter emissions (Willes et al, 1992; Sanderson et al., 2005). The BEF for each PAH were summed to obtain the total BEF to compare differences between pitch types, jobs, and potroom buildings. Resul ts Relationship Between BaP and B S M The correlation between BaP and BSM varied by type of anode paste, with moderate to high correlation for Kosco/VFT pitch (r=0.81) and Koppers pitch (r=0.65), but low correlation for pre-1980 pitch (r=-0.07) and VFT/Koppers pitch (r=0.23). Significant limitations had been found with the analytical method used pre-1980 by the regulatory agency, which may account for the lack of a relationship between BaP and BSM for this pitch type. Based on a sampling campaign in April 1995, there was no significant difference in the BaP/BSM ratio due to source of measurement (company and regulatory agency samples) (Chapter 2). As anode paste type was highly correlated with the source of measurement, comparison for other types or time periods was not possible. A stronger BaP/BSM relationship was observed for anode paste plant operators and maintenance workers (r=0.99) than for potroom workers (r=0.80) for Kosco/VFT pitch, which accounted for the majority of measurements. No significant difference in the BaP/BSM ratio for pot operators was observed for the Kosco/VFT anode paste when the proportion of Kosco and V F T in the mixture was taken into account using linear regression (p=0.4, R2<0.01), supporting this grouping. Eight job groups with BaP and BSM measurements were identified. The BaP/BSM relationship could only be determined for the most recent anode paste type for six job groups (Table 3.2). Potroom 37 operations jobs were grouped the jobs into two categories: near field and far field. Near field jobs spent the majority of their time in the potroom right at the pot side or on top of pot (e.g. controlmen, exhaust maintenance, and pot repairman). Far field jobs spent some time at the pot side or on top of pot, but also spent considerable time in potroom vehicles or passageways (e.g. anode operator, pot operator, and equipment operator). Near field jobs had a significantly higher BaP/BSM ratio than far field jobs (p<0.001). The correlation between BaP and BSM was higher for near field jobs (r=0.83) than for far field jobs (r=0.75) for Kosco/VFT pitch. For the anode paste plant, no job differences in the BaP/BSM ratio were observed (p=0.3). There were insufficient measurements to examine job-specific ratios within other work areas. Five of seven work groups had significantly different BaP/BSM ratios than potroom operations far field jobs. The BaP/BSM ratios for cathode rodding and potlining mix plant were based on very few measurements. In a multiple linear regression model, 27% of the variability in the log-transformed BaP/BSM ratio was explained by work area and anode paste type; season and potroom building differences were not significant. Since the model provided no additional information, only the descriptive statistics of the BaP/BSM ratio by work area and anode paste type are presented (Table 3.2). Relationship Between BaP and Other PAHs Only the PAH profile in the potroom environment could be assessed as no other work area had more than four personal or area particulate PAH measurements. BaP was very strongly correlated with individual particulate PAHs, with most correlations above 0.9 (Table 3.3). The correlations decreased if all anode paste types were considered together suggesting that while the relationship between BaP and other PAHs is strongly linear, the slope differs by anode paste composition. This was confirmed by examining the factors that influence the magnitude of the BaP/PAH relative abundance ratios in linear regression analyses (Table 3.4). Anode paste composition, seasonal, potroom building, and job differences were observed in the log-transformed BaP RARs and explained between 23 and 89% of the variability in the dataset, dependent on the PAH (Table 3.4). Seasonal differences were significant for 3 of 8 PAHs. Differences in the BaP RARs by potroom building (Lines 1-2, Lines 3-5, Lines 7-8) were observed for all PAHs except benzo(k)fluoranthene. 38 Pot operators accounted for 70 % of the measurements. Anode assistants, who spent the majority of their exposed time at the top of the pots had significantly different BaP RARs for 5 of 6 PAHs. The BaP RARs for exhaust maintenance workers, who repair and maintain the pots' gas collection systems, were also significantly different than pot operators for 5 of 6 PAHs. For anode operators, who are primarily within the crane cabs and were the second most measured job, the BaP RARs were significantly different than pot operators for 3 of 8 PAHs. For equipment operators, the BaP RARs were only significantly different than pot operators for one PAH. ' For controlman, pot startup operator, and pot maintenance, the BaP RARs were significantly different that pot operators for approximately half of the PAHs. The job differences in the PAH profile for anode operators, pot startup operators, and pot operators for Kosco/VFT (>75/<25) anode paste is illustrated in Figure 3.1. Unlike the BaP/BSM ratio, the BaP RARs were impacted by the proportion of Kosco (50% vs. >75%) in the anode paste for three of the five PAHs. The BaP RARs for Koppers anode paste was consistently different than Kosco/VFT (>75/<25) for all the PAHs. For Koppers/VFT, the BaP RARs were significantly different than Kosco/VFT (>75/<25) for 5 of 6 PAHs assessed. Predicted BaP RARs and BaP Equivalency Factors With the exception of Koppers anode paste which was used for only one month, the total BEFs were relatively consistent across anode paste types (range: 2.3-2.8) although the relative contribution of each PAH varied across anode paste types (Table 3.5). Excluding Koppers pitch, the contribution of benzo(ghi)pyrene (34-46%), benzo(a)pyrene (35-43%) and chrysene (7-22%) to the total BEF was larger due to their larger BaP potency factors than benzo(e)pyrene (2-4%) and benzo(k)fluoranthene (0.1-0.3%). The more potent dibenzo(a,h)anthracene accounted for only 1-2% of the mixture's carcinogenicity for Kosco/VFT. The total BEFs were consistent across the potroom buildings (range: 2.3-2.5). Job differences in the total BEF ranged from 2.1 to 3.5 for Kosco/VFT (>75/<25) pitch type and pot lines 1-2. Five jobs had a total BEF of 2.5: anode assistant, anode operator, controlman, equipment operator, and pot operator. The latter four of these jobs had nearly identical BaP RARs across the five PAHs, whereas the predicted BaP RARs for anode assistant differed from these jobs for all PAHs. Exhaust maintenance jobs had the highest total BEF (3.5). Due to differences in the predicted BaP RARs for most PAHs, the total BEFs for pot startup operator (2.1), pot maintenance (2.7), and exhaust maintenance (3.5). 39 Discussion In this study we found that BaP was strongly correlated with other particulate PAHs supporting the use of BaP as a marker of the PAH mixture. B S M was only moderately correlated with BaP and hence other PAHs, and thus was a less precise indicator of the PAH mixture than BaP. We identified several factors that influenced the relationship between BaP and the other PAHs, namely the source of the coal tar pitch (anode paste composition), work area, and job. Previous studies have suggested that a single smelter-specific P A H profile for a given time period could be used to relate BaP exposure to other PAHs (Farant and Gariepy, 1998; Sanderson et al, 2005). However, the differences observed here suggest that a single smelter-specific P A H profile would not account for important within-smelter differences and thus may result in misclassification error from using a single marker for the carcinogenic potential of the P A H mixture. Coal tar pitch is the source of PAHs in the aluminum smelter, thus the differences in the relationships between BaP, B S M , and other particulate PAHs related to the type of coal tar pitch used for the anode paste as observed here and by Sanderson and Farant (Sanderson et al, 2005) are not surprising. The type of Soderberg aluminum smelter (horizontal-stud vs. vertical-stud) has been previously shown to impact the P A H profile (Sanderson et al, 2005); however, we observed differences in the PAH profile due to much smaller differences in technology and process conditions within the same type of smelter. The work shift differences in the P A H profile recently observed by Sanderson et al (Sanderson et al, 2005) provides additional support for within-smelter variability in the PAH profile due to small differences in conditions. Work area and job differences have been previously observed between BaP and B S M (Armstrong et al, 1986, 1994), but not between BaP and other PAHs (Farant and Gariepy, 1998; Sanderson et al, 2005). The reasons for the job differences in the P A H profile are not clear and could not be investigated in this study. Jobs that entailed a larger proportion of exposed time in potroom vehicles and cranes were more similar to each other compared to other jobs that entailed a greater proportion of exposed time beside an open pot or above the pot near the anode. This may suggest that patterns of P A H emission differ within the potroom or may suggest that different PAHs may preferentially bind to different size particles which potroom vehicle air filters may differentially remove. BaP degradation is unlikely to account for the job differences observed here, as we found similar pitch, job, potroom building, and seasonal differences using benzo(b)fluoranthene relative abundance ratios, which has been suggested as a more stable indicator of P A H for environmental exposure (Aubin and Farant, 2000). 40 This study's strength is the availability of personal exposure measurements analyzed for different components of the mixture over a 22 year time period. However, some of the differences observed may be due to artifacts in the dataset and other unmeasured factors may also influence the relationships between BaP, BSM, and other PAHs. For instance, sampling and analytical methods used by the company changed in 1982, but the impact of this change was not possible to examine as it coincided with changes in coal tar pitch source. A study examining the determinants of exposure that includes information collected on process conditions, work tasks, and proximity to open pots and top of anode to confirm and expand on our findings is needed. In addition, our study was limited to examining factors that impact the relationship between BaP and selected particulate PAHs. Dibenzo(a,I)pyrene is believed to be more than ten times more potent than benzo(a)pyrene (Straif et al, 2005); however, it was not routinely analyzed at the smelter so its contribution to the mixture's toxicity could not be assessed. Factors that impact the relationship between BaP and volatile and semi-volatile PAHs may be different and should also be examined. An alternative metric for epidemiologic studies of PAHs would be a measure of the mixture's carcinogenicity that accounts for both the exposure levels and relative toxicity of each component of the mixture to facilitate comparisons between studies and industries exposed to PAHs (Krewski et al, 1989). In this study, the total BaP equivalency factor showed more variation within the same time period due to job differences, than due to potroom building, or coal tar pitch source. In comparison, larger differences in the total BEF have been observed between the two types of Soderberg aluminum smelters (vertical-stud and horizontal-stud) (Sanderson et al, 2005). Using this aggregated metric requires several assumptions, including assuming that the toxicity of the mixture's components are additive and that an animal model is relevant. However, the biggest challenge is the dearth of measurements analyzed for the components of the mixture for all locations, jobs, and process conditions of interest (Armstrong et al, 2004). For instance, at this smelter there are no measurements analyzed for multiple PAHs for work areas other than the potrooms. As more information on the relative abundance of specific PAHs are being evaluated this approach may become feasible in future studies (Farant and Gariepy, 1998; Sanderson and Farant, 2000; Sanderson et al, 2005). While our study is specific to the primary aluminum smelter industry, it provides important insight into factors influencing the relationship between two common exposure metrics used in epidemiologic studies of PAH exposure and other components of the PAH mixture. The identification of within and between-industry differences in the PAH mixtures is important for 41 understanding the role that individual PAHs may contribute to cancer risk. Understanding the components of the PAH mixture helps identify appropriate indicator metrics and facilitates comparisons between studies and industries. Until these relationships are better understood, benzo(a)pyrene seems to be a reasonable choice, albeit an imperfect indicator, due to the strong correlation between BaP and other PAHs for a given set of conditions and due to the relative abundance of BaP exposure measurements. 42 Table 3.1 Coal tar pitch in use at Soderberg aluminum smelter Year Coal Tar Pitch Coal Tar Pitch Group Pre-1980 Unknown (VFT or Kosco?) Pre-1980 1980 50% VFT & 50% Koppers VFT/Koppers January, 1981 100% Koppers Koppers 1981 50% VFT & 50% Koppers VFT/Koppers 1981 (Potline3) Sumitomo Sumitomo A January & February, 1982 100%o Kosco Kosco/VFT (>75/<25) March - December, 1982 50% VFT & 50% Kosco Kosco/VFT (50/50) 1983 75% Kosco & 25% VFT Kosco/VFT (>75/<25) 1984-85 83% Kosco & 17%> VFT Kosco/VFT (>75/<25) 1986-87 85% Kosco & 15%> VFT Kosco/VFT (>75/<25) 1988 87% Kosco & 13%> VFT Kosco/VFT (>75/<25) 1993 89% Kosco & 11% VFT Kosco/VFT (>75/<25) 1999 100%> Kosco Kosco/VFT (>75/<25) A Not assessed as no measurements collected. 43 Table 3.2 BaP/BSM ratio (ug:mg) by work area and coal tar pitch mixture. GM GSD N PITCH: Kosco/VFT (>50/50), 1982-2000 Potroom Operations (1) Far: Pot Operator, Equipment Op, Studblast <• 7.2 2.18 328 Op, Potlining, Startup Op; Anode Op (2) Near: Anode Assistant, Foreman, 14.6* 3.52 44 Controlman Potroom Maintenance 7.4* 3.92 56 Gas Collection 11.2* 2.08 65 Cathode Rodding 8.9 2.73 5 Anode Paste Plant 11.9* 2.28 37 Coke Calcining 0.3* 10.9 7 Potlining Mix Plant 4.7 5.36 3 PITCH: Koppers, 1981 Potroom Operations (1) Far: Pot Operator, Equipment Op, Studblast 36.1 2.42 37 Op, Potlining, Startup Op; Anode Op (2) Near: Anode Assistant, Foreman, 39.9 3.83 6 Controlman PITCH: VFT/Koppers, 1980-1981 Potroom Operations (1) Far: Pot Operator, Equipment Op, Studblast 3.0 2.80 41 Op, Potlining, Startup Op; Anode Op (2) Near: Anode Assistant, Foreman, 5.2 2.22 6 Controlman PITCH: Unknown, pre-1980 Potroom Operations (1) Far: Pot Operator, Equipment Op, Studblast 9.1* 4.96 17 Op, Potlining, Startup Op; Anode Op # p<0.10 in Mann-Whitney two sample comparison compared to Potroom Operations "Far" jobs and Kosco/VFT (50/50) coal tar pitch mixture. 44 Table 3.3 Correlation (Pearson's r) of benzo(a)pyrene with particulate PAHs in the potrooms, by pitch and job type. Number of measurements in parentheses (). Pitch 1975-78 Pitch 0.95 0.98 0.99 0.99 0.99 - - - 0.95 (50) (49) (51) (51) (51) (51) 50% VFT & 50% Koppers 0.96 0.98 0.95 0.62 0.69 - - 0.96 0.85 (63) (66) (71) (64) (66) (18) (71) 50% Kosco, 50% VFT -- 0.86 0.99 0.97 0.99 -- - 0.99 0.95 (29) (10) (22) (34) (29) (14) >75% Kosco, <50% VFT 0.97 0.94 0.98 0.96 0.97 0.98 0.95 0.95 0.93 (96) (162) (137) (124) (162) (96) (65) (65) (62) All pitch types 0.94 0.88 0.82 0.85 0.85 0.97 0.98 0.94 0.79 (209) (306) (269) (261) (313) (95) (96) (112) (198) 45 Table 3.4 Multiple linear regression of time and work area-specific factors that impact the BaP RARs I n B A F I n B B F InBEP I n B G H I P I n B K F I n C H R Y I n D A H A InIND P ( S E ) P (SE) P ( S E ) P ( S E ) P ( S E ) P ( S E ) P (SE) p ( S E ) Intercept 0.83 0.71 0.75 -0.04 -0.41 0.52 -2.04 -0.59 (0.06) (0.03) (0.09) (0.05) (0.03) (0.08) (0.04) (0.07) Season: Winter — -0.09 — -0.11 — — — 0.37 Pitch typeH (0.04) (0.06) (0.08) Unknown Pitch (78-79) A -0 .02 B -0.24 0.24 - 0.17 B A A (0.04) (0.11) (0.08) 0 . 0 1 B (0.12) (0.07) Koppers (100%) -2.22 A -2.31 -0.32 -0.32 -1.28 A A (0.08) (0.35) (0.08) (0.07) (0.42) K o p p e r s / V F T (50/50) 0 .18 B -0.14 -0.17 0.17 0.38 -0.72 A A (0.11) (0.04) (0.10) (0.07) (0.06) (0.11) K o s c o / V F T (50/50) -0.23 A -0.94 0 .15 B -1.50 -0 .26 8 A A (0.10) (0.18) (0.10) (0.15) (0.14) K o s c o / V F T (>75/<25) Reference Potline Lines 3-5 , -0.13 — -0.31 -0.14 — -0.23 -0.17 -0.13 (0.08) (0.10) (0.05) (0.10) (0.07) (0.08) L ines 7-8 -0.45 -0.18 -0.36 - - -0.51 -0.19 -0.24 (0.10) (0.04) (0.12) (0.12) (0.11) (0.13) L ines 1-2 Reference Jobs Anode Assistant -0.87 — -0.34 -0.28 0.29 0.43 A A (0.14) (0.15) (0.10) (0.10) (0.17) Anode Operator -0.24 -0.13 - - ~ - - 0.38 (0.09) (0.05) (0.13) Controlman A -0.54 A - -0.69 - -0.68 -0.52 (0.12) (0.21) (0.15) (0.18) Equipment Operator - -0.40 - - - - - -(0.10) Exhaust Maintenance 1.05 0.22 0.72 0.67 0.48 - A A (0.23) (0.09) (0.19) (0.14) (0.14) Pot Startup Operator - -0.24 - -0.66 - - -1.13 -0.81 (0.10) (0.16) (0.13) (0.15) Pot Maintenance 0.59 - 0.93 - 0.40 - A A (0.28) (0.57) (0.19) Pot Operator Reference R2 0.89 0.34 0.36 0.23 0.44 0.32 0.50 0.47 variance 0.14 0.06 0.32 0.20 0.17 0.49 0.09 0.11 # measurements 160 233 213 377 340 345 101 104 A no measurements to assess effect B pitch type not significant (p>0.10) -- location or job effect not significant (p>0.10) BAF = Benzo(a)fluoranthene; BBF = Benzo(b)fluoranthene; BEP = Benzo(e)pyrene; BGHIP = Benzo(ghi)pyrene; CHR = Chrysene; DAHA = Dibenzo(ah)anthrance; IND=Indeno(l,2,3-cd)pyrene 46 Table 3.5 Predicted BaP RARs and BaP Equivalency Factors (BEF) by pitch type and potroom building for pot operators BaP Predicted BaP Relative Abundance Ratios Potency Factors Unknown Koppers Koppers/ Kosco/ Kosco/ Kosco/ Kosco/ (1978-79) V F T VFT V F T VFT V'FT (50/50) (>75/<25) (>75/<25) • (>75/<25) Lines 1-2 Lines 1-2 Lines 1-2 Lines 1-2 Lines 1-2 Lines 3-5 Lines 7-8 Predicted BaP RAR Benzo(a)fluoranthene n/a - 0.25 2.75 1.82 2.29 2.01 1.46 Benzo(b)fluoranthene 0.1 1.99 ~ 1.77 - 2.03 2.03 1.70 Benzo(e)pyrene* 0.05 1.67 0.21 1.79 0.83 2.12 1.55 1.48 Benzo(ghi)pyrene* 1.0 1.22 0.70 1.14 1.12 0.96 0.83 0.96 Benzo(k)fluoranthene* 0.01 0.65 0.48 0.97 0.15 0.66 0.66 0.66 Chrysene* 0.26 1.99 0.47 0.82 1.40 1.68 2.12 1.01 Dibenzo(a,h)anthracene 1.4 - - - - 0.13 0.11 0.11 lndeno( 1,2,3-cd)pyrene 0.1 - - - - 0.55 . 0.49 0.44 Benzo(a)pyrene* 1.0 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Total BEF of selected (*) PAHs'' 2.83 1.84 2.45 2.51 2.47 •2.30 2.51 *Several PAHs are not included in the calculation of total BEF as these PAHs were not analyzed for all pitch types. B Willes etal (1992) c Sum of BEF x BaP RAR for selected PAHs 47 Figure 3.1 Comparison of mean BaP RAR (±2 standard errors) for three potroom jobs and pitch type Kosco/VFT (>75 / <25%) 250 2DQ ] 150 1 •S. m wo 050 ODD II Anode Op • Potroom Op • Pot Startup Op te 5 ffi i i B$.)F B<3h\f B lKf d$y\ Chry D^hyv hdsno d$J B(a)F = Benzo(a)fluoranthene; B(b)F = Benzo(b)fluoranthene; B(e)P = Benzo(e)pyrene; B(ghi)P = Benzo(ghi)pyrene; Chry = Chrysene; D(ah)A = Dibenzo(ah)anthrance; Indeno=Indeno( 1,2,3-cd)pyrene 48 References Armstrong B, Hutchinson E, Unwin J, Fletcher T (2004) Lung cancer risk after exposure to polycyclic aromatic hydrocarbons: A review and meta-analysis. Environ Health Perspect; 112 970-8. Armstrong B, Tremblay C, Baris D, Theriault G (1994) Lung cancer mortality and polynuclear aromatic hydrocarbons: A case- cohort study of aluminum production workers in Arvida, Quebec, Canada. Am J Epidemiol; 139 250-62. Armstrong BG, Tremblay C G , Cyr D, Theriault GP (1986) Estimating the relationship between exposure to tar volatiles and the incidence of bladder cancer in aluminum smelter workers. Scand J Work Environ Health; 12 486-93. Aubin S, Farant JP (2000) Benzo[b]fluoranthene, a potential alternative to benzo[a]pyrene as an indicator of exposure to airborne PAHs in the vicinity of Soderberg aluminum smelters. J Air Waste Manag Assoc; 50 2093-101. Benke G, Abramson M , Sim M (1998) Exposures in the alumina and primary aluminium industry: An historical review. Ann Occup Hyg; 42 173-89. Boffetta P, JourenkovaN, Gustavsson P (1997) Cancer risk from occupational and environmental exposure to polycyclic aromatic hydrocarbons. Cancer Causes Control; 8 444-72. Farant JP, Gariepy M (1998) Relationship between benzo(a)pyrene and individual polycyclic aromatic hydrocarbons in a Soderberg primary aluminum smelter. Am Ind Hyg Assoc J; 59 758-65. Friesen M C , Demers PA, Spinelli JJ, Le ND (2003) Validation of a semi-quantitative job exposure matrix at a Soderberg aluminum smelter. Ann Occup Hyg; 47 477-84. Krewski D, Thorslund T, Withey J (1989) Carcinogenic risk assessment of complex mixtures. Toxicol Ind Health; 5 851-67. NTP (2005) Polycyclic aromatic hydrocarbons. Report on carcinogens. Eleventh edition, U.S. Department of Health and Human Services, Public Health Service. Sanderson E, Kelly P, Farant JP (2005) Effect of Soderberg smelting technology, anode paste composition, and work shift on the relationship between benzo[a]pyrene and individual polycyclic aromatic hydrocarbons. J Occup Environ Hyg; 2 65-72. Sanderson E G , Farant JP (2000) Use of benzo[a]pyrene relative abundance ratios to assess exposure to polycyclic aromatic hydrocarbons in the ambient atmosphere in the vicinity of a Soderberg aluminum smelter. J Air Waste Manag Assoc; 50 2085-92. 49 Straif K, Baan R, Grosse Y, Secretan B, El Ghissassi F, Cogliano V. Carcinogenicity of polycyclic aromatic hydrocarbons. Lancet; 6 931-932. Tremblay C, Armstrong B, Theriault G, Brodeur J (1995) Estimation of risk of developing bladder cancer among workers exposed to coal tar pitch volatiles in the primary aluminum industry. Am J Ind Med; 27 335-48. Willes RF, Friar S, Orr J, Lynch B (1992) Application of risk assessment to point sources of polycyclic aromatic hydrocarbons (PAHs). J Air Waste Manag Assoc; 42 75-100. 50 Chapter 4: Comparison of benzo(a)pyrene and benzene soluble materials as indices of exposure to polycyclic aromatic hydrocarbons for a retrospective cohort study of aluminum smelter workers 3 Introduction Several individual polycyclic aromatic hydrocarbons (PAHs) have been classified as carcinogens or probable carcinogens based on animal studies (NTP, 2005; Straif etal, 2005); however, workplace exposures are to PAH mixtures, not individual PAHs. Exposure to coal tar derived mixtures that include PAHs have been found to be carcinogenic in several industries, but the specific role of PAHs is not known (Straif et al, 2005). In primary aluminum smelting exposure to coal tar pitch volatiles has been associated with increased risk of bladder and lung cancer, and has raised questions about increased risks of other cancers and cardiovascular disease (Armstrong et al, 1986; Theriault et al, 1988; Spinelli etal, 1991; Armstrong et al, 1994; Ronneberg and Andersen, 1995; Ronneberg et al, 1999; Romundstad, Andersen et al, 2000; Romundstad, Haldorsen et al, 2000; Spinelli et al, 2006). Coal tar pitch volatiles are a complex mixture that includes over 100 polycyclic aromatic hydrocarbons, several of which are known carcinogens (Skogland, 1991; NTP, 2005). Coal tar pitch volatiles may also contain very low levels of amino and nitro polycyclic aromatic hydrocarbons, which are also known bladder carcinogens (Roussel et al, 1991; Farant and Ogilvie, 2002). While PAHs are suspected, the specific causal agents of the increased cancer risk in aluminum smelters are not known (Tremblay et al, 1995; Farant and Gariepy, 1998; Straif et al, 2005). The occurrence of PAH exposure as a mixture creates significant challenges for exposure assessment in epidemiologic studies. One challenge is choosing an appropriate indicator of the carcinogenic or toxic potential of the mixture. Epidemiologic studies have used various indicators, including measurements of individual hydrocarbons or the mixture, to describe exposure to PAHs and to examine dose-response relationships. These measures have included benzene soluble materials (BSM), total particulate PAHs, and benzo(a)pyrene (BaP) due to routine monitoring of these components at most smelters (Benke et al, 1998). Of these metrics, BaP has been proposed to be a 3 A version of this chapter has been submitted for publication. Friesen M C , Demers PA, Spinelli JJ, Lorenzi MF, Le ND. Comparison of benzo(a)pyrene and benzene soluble materials as indices of exposure to polycyclic aromatic hydrocarbons for a retrospective cohort study of aluminum smelter workers (Submitted) 51 more specific indicator of exposure to PAHs as a class of carcinogens and thus a better indicator of the carcinogenic potential of coal tar pitch volatiles than BSM (Tremblay et al, 1995; Farant and Gariepy, 1998). Initial comparisons between BaP and BSM as exposure indices in aluminum smelter studies have found that BaP provided a slight improvement in the dose-response for bladder cancer (Armstrong et al, 1986; Tremblay et al, 1995), whereas BSM provided a better fitting dose-response for lung cancer (Armstrong et al, 1994). A recent follow-up health study at a vertical stud Soderberg aluminum smelter provided an opportunity to compare quantitative BSM and BaP exposure indices in analyses of mortality and cancer incidence. In contrast to other aluminum smelter studies that used the work-area specific relationship between BaP and BSM to derive BaP exposure estimates (Armstrong et al, 1986; Armstrong et al, 1994), the BaP exposure index used here was derived independently of BSM exposure levels (Chapter 2). In this paper, our objective was to examine the dose-response relationships to determine which of the exposure indices, BSM or BaP, provided a better marker for the causal component(s) of coal tar pitch volatiles. The shape of the dose-response relationship was examined using both a log-linear model [ln(relative risk)=R*cumulative exposure] and a log-log model [ln(relative risk)= (3 *ln(cumulative exposure+1)] structure. The focus of the dose-response relationships presented in this paper were bladder cancer incidence, lung cancer incidence, and acute myocardial infarction mortality because a substantial number of cases were available for examining the slope of dose-response relationships and monotonically increasing risks have been observed in categorical analyses (Spinelli et al, 2006). Methods Study Population The study population consisted of 6423 males with 3 or more years of work at the aluminum smelter or its power generating station since operation began in 1954 through 1997. The cohort members' work histories up to December 31, 1997 (job title, department, starting date, and stopping date for each job assignment) were obtained from company records. Cohort members were linked using probabilistic linkage techniques to the national mortality database (1954-1999) to ascertain their cause of death and to the National Cancer Registry (1969-1999) to identify cancer diagnoses. Active follow-up from the original study, including pension records, advertisements in company newsletter, union lists, and contact of last employers and family members, supplemented vital status information (Spinelli et al, 1991). Vital status was also ascertained through linkage to the B C Client Registry 52 administered by the BC Ministry of Health (records of health care recipients on the provincial medical system since 1984) to determine their last known dates residing in British Columbia (BC). Workers were censored at their last contact dates if they did not link to the BC Client Registry and their location at last contact was the company or not in Canada or if they had been censored in the original study prior to 1985 and did not link to the BC Client Registry. Workers who were not censored and did not die during the follow-up period (1954-1999) were assumed to be alive at the end of the study period of December 31, 1999. Workers with contact information (current workers and pensioners) were sent a self-administered questionnaire to request information on their smoking habits. The results were supplemented with smoking information obtained from a similar mailed questionnaire to current workers in the original study. Exposure Assessment Exposure were evaluated using two different measures: benzene soluble material (BSM) and Benzo(a)pyrene (BaP). BSM provides a measure of the mixture and includes all components of the particulate exposure that are extractable by benzene, including polycyclic aromatic hydrocarbons (PAH). BaP is a single particulate-phase PAH and has been found to be a good indicator of both individual and total particulate PAHs, but not necessarily of volatile PAHs (Farant and Gariepy, 1998; Sanderson et al, 2005). The relationship between BaP, BSM and other PAHs has been found to vary between work areas, jobs, and technological processes (Armstrong et al, 1986; Ronneberg, 1995; Farant and Gariepy, 1998; Sanderson etal, 2005). Job exposure matrices (JEM) for BSM and BaP were developed independently of each other using several exposure assessment approaches to maximize the use of personal exposure measurements (Chapter 3). Briefly, statistical models were developed to derive annual arithmetic means for each potroom (reduction plant) operations and potroom maintenance job from 1977 to 2000. For non-potroom locations, mean exposures were directly calculated. Exposure estimates for jobs without exposure measurements were extrapolated from exposure estimates from the statistical models after adjusting for the amount of time in exposed areas. Estimating pre-1977 exposure levels involved backwards extrapolation of 1977 exposure levels. Data Analysis To examine the slopes and precisions of the dose-response relationships with cumulative BSM and BaP exposures, we selected diseases with greater than 50 cases that had significant dose-response trends based on categorical analyses in the full mortality and cancer incidence study (Spinelli et al, 53 2006). Three diagnoses met this criteria: bladder cancer (ICD-9: 188; 90 cases, including in-situ cases); lung cancer (ICD-9: 162; 147 cases), and acute myocardial infarction (ICD-9: 410; 184 deaths). The cumulative exposure indices were lagged according to the best fitting latency periods observed in the categorical analyses. For bladder and lung cancer incidence exposure was lagged 20 years; for acute myocardial infarction exposure was not lagged. Iriternal comparisons, using workers in the lowest exposure category as the reference group, were used to examine dose-response relationships. The Life Table Analysis System (PCLTAS, NIOSH) software was used to generate files for analysis. Relative risks were calculated using maximum likelihood methods after adjustment for age, calendar year, and smoking status using Poisson regression (SAS version 8.0, SAS Institute Inc., Cary North Carolina, USA). Six exposure categories were used: an unexposed group (BSM: <0.05 mg/m3-yr; BaP: <0.5 ug/m3-yr), and five exposed groups with cumulative exposure cut points defined by distributing the remaining cases into five equal size groups. As such, different exposure cut points were identified for each outcome and exposure metric. Because of the different units of the BaP and BSM exposure metrics, a case-based definition of exposure cut points provided a comparison that was less sensitive to a priori cut points and equalized the standard errors for each exposure category. This approach has previously been found to minimize bias in grouped analyses of cohort data (Richardson and Loomis, 2004). Smoking status was defined as ever (57% of workers, reference group), never (19%), or unknown (23%). Ten-year calendar period and five-year age categories were used. Categories were combined when necessary to ensure all categories had a minimum of ten cases to improve precision. Ninety-five percent confidence intervals were based on the standard error of the coefficients derived from the model. Exposure was treated as a continuous variable in Poisson regression by assigning the mean cumulative exposure to all subjects in each category. The mean cumulative exposure was weighted by the person-years each subject contributed (Richardson and Loomis, 2004). Two dose-response model structures were assessed: the log-linear model, where the relative risk, RR, increased logarithmically with a linear increase in cumulative exposure, C E (equation 4.1); the log-log model, where the relative risk increased logarithmically with log-transformed cumulative exposure (equation 4.2) (Steenland and Deddens, 2004). Log-linear model: Ln(RR) = p*CE (4.1) Log-log model: Ln(RR) = p*ln(CE+l) (4.2) 54 To determine which of the exposure metrics and which dose-response relationship provided the best fit for each health outcome, we compared two measures: 1) the change in the -21oglikelihood model fit statistic with and without exposure in Poisson regression, while adjusting for age, calendar year, and smoking as fixed effects; 2) the precision of the dose-response relationship based on the Wald statistic (slope/standard error) (Kromhout et ai, 1997). Resul ts The cumulative exposure indices for BSM and BaP were very highly correlated (r=0.94), but substantial deviations from a linear relationship are observed (Figure 4.1). Adjusting the correlation for duration of employment did not weaken the relationship (partial correlation=0.93) since duration of employment was only moderately correlated with cumulative exposure of either metric (r=0.4). Approximately half of the workers worked in non-exposed areas of the plant. Categorization of cumulative exposure based on the distribution of cases slightly decreased the relationship between the exposure metrics, with weighted kappa values between 0.85-0.87 and percent agreement of category assignment between 74-77%, with the slight variation dependent on the health outcome. The median BSM cumulative exposures of the exposed workers were 2.6 and 2.4 mg/m3-yr for no lag and 20 year lag, respectively, and the maximum cumulative exposure was 46 mg/m3-yr. The median BaP cumulative exposures were 20 and 18 u.g/m3-yr for no lag and 20 year lag, respectively, and the maximum cumulative exposure was 300 pg/m3-yr. The proportions of workers in the unexposed category with no lag were 19 and 20% for BSM and BaP, respectively. With a 20 year lag the proportions unexposed were 36 and 37% for BSM and BaP, respectively. The category cut points and their mean cumulative exposures were disease-specific (Table 4.1). For bladder cancer incidence, the BSM and BaP cumulative exposure metrics had a significant dose-response relationship (Figure 4.2). The BaP metric had the highest relative risk (3.0) with the highest cumulative exposure, the most consistent monotonic increase, the largest improvement in model fit, and had the most precise slope for the dose-response relationship (Table 4.2). For both BSM and BaP, the log-log dose-response model structure had the greatest improvement in model fit (as measured by -2loglikelihood) and in the precision of the slope (B/SE) compared to the log-linear model. Visually, the dose-response relationship for BaP was best described by a linear relationship 55 between increasing risk and cumulative exposure, but this model structure could not be assessed in Poisson regression using the SAS statistical program. For lung cancer incidence, both the BSM and BaP cumulative exposure metrics had a significant dose-response relationship (Figure 4.3). The highest cumulative exposure categories had significant elevations in lung cancer risk for both metrics, with relative risks of 2 for BSM and 1.8 for BaP. The shapes of the dose-response curves were very similar. For both metrics, the log-log mode/ structure provided a greater improvement in model fit compared to the log-linear model (Table 4.2). The slope of the dose-response was most precise for BaP with the log-log model. For acute myocardial infarction, we observed a monotonically increasing risk with increasing cumulative exposure for both metrics in the categorical analyses, but the slopes of the dose-response relationships were not significant for either metrics and no individual exposure category had significantly elevated risks (Figure 4.4, Table 4.2). The highest relative risks were 1.4 for both BSM and BaP. The BaP and BSM cumulative exposure metrics with the log-log model structure provided nearly identical improvements in model fit and their slopes were equally precise. Discussion Despite a high correlation between the BaP and BSM cumulative exposure metrics we observed differences in model fit and in the precision of the dose-response relationships. For all three health outcomes, BaP cumulative exposure provided the greatest improvement in model fit and had the more precise slope. The strength of the dose-response relationships with BaP provides support that BaP is more closely related to the etiological agents than BSM. However, the improvements in the dose-response relationship of BaP over BSM were only slight. Thus, similar conclusions were made regarding associations between coal tar pitch volatiles exposure and the health outcome of interest regardless of exposure metric. Previous aluminum smelter studies have assessed BaP exposure using the ratio of BaP to BSM, since much less sampling data is available for BaP than for BSM (Armstrong et al, 1986; Armstrong et al, 1994). These studies have accounted for work area and job differences in the BaP/BSM ratio, but have assumed through necessity that these ratios were stable throughout the study period. A ratio-derived BaP JEM for this cohort that accounted for the anode paste composition and work area differences was less precise than the independently assessed BaP exposure (not shown). In this study, 56 we assessed BaP exposure independently from BSM so we could account for time-varying factors that influenced BaP exposure, but we were unable to account for differences in the relationship between BaP and other PAHs which has been found to vary across time, job, and work area (Chapter 3). Exposure misclassification may be introduced by the use of an individual compound if its relationship with the causal components varies. A quantitative comparison of two different exposure metrics is complicated when the metrics have different units of exposure or scales of exposure. To facilitate comparisons between BSM and BaP, we defined exposure categories based on the cumulative exposure of the cases to minimize bias due to cutpoint selection and treated exposure as a continuous variable in Poisson regression (Richardson and Loomis, 2004). Both the improvement in model fit and the precision of the dose-response relationship indicated that cumulative BaP exposure was the better index for all three health outcomes. However, both measures are sensitive to the units and scale of the exposure metric due to the logarithmic relationship with relative risk and exposure. For the log-linear model, changing the scale of the exposure metric (i.e. dividing BaP by ten to be on the same scale as BSM) did not change the magnitude of the improvement in model fit, but did slightly change the magnitude of the precision. The log-log model was particularly sensitive to the scale of these measures, with both the improvement in model fit and the magnitude of the precision changing with the scale of the metric. Regardless, BaP provided a better fit in all cases. A leveling off of risk at the highest exposure categories (log-log model) has been observed in many retrospective studies. Nondifferential exposure misclassification has been shown in simulation studies to have this effect (Dosemeci et al, 1990). The highest exposure categories were strongly influenced by pre-1977 exposure levels where no exposure measurements were available. While we extrapolated backwards from 1977 exposure estimates, assumptions regarding the shape of the time trend was required (Chapter 2). The linear relationship between bladder cancer and BaP exposure seen here provides indirect support for the validity of the BaP exposure metric. Sensitivity analyses could be conducted in future analyses to examine the impact of different assumptions regarding the backwards extrapolation (Loomis et al, 1998; Kromhout et al, 1999). Other possible explanations for the leveling off of risk at high exposure categories include confounding by other risk factors, healthy worker survivor effect; depletion of the number of susceptible people at high exposure levels; a natural limit on relative risks for diseases with high background rates; and the saturation of key disease pathways (Stayner et al, 2003). Several of these reasons may be present in this cohort; 57 however, these explanations would be expected to have a similar impact for both the BaP and BSM exposure indices. Other aluminum smelter exposures have been proposed as indicators of PAH, namely benzo(b)fluoranthene and naphthalene. Aubin and Farant (2000) suggested benzo(b)fluoranthene as a more stable indicator of PAH environmental exposure as BaP levels are more influenced by environmental factors; however, it was very highly correlated with BaP (r=0.97) within this smelter. Rappaport et al. (2004) proposed naphthalene as a potential surrogate for occupational PAH exposure as it is the most abundant PAH and because it is present almost entirely in the gas phase and therefore is easily measured. The relationship between BaP and naphthalene at this smelter is not known as naphthalene has not been regularly monitored. Naphthalene has been found to be only moderately correlated (r=0.54) with BaP at another Soderberg aluminum smelter (Farant and Gariepy, 1998), suggesting that it would be useful to examine since it measures alternative exposures. The correlation between BaP and amino- and nitro-PAHs has not been reported, but amino- and nitro-PAH levels were very low in one aluminum smelter and were found to be well controlled by current ventilation processes (Farant and Ogilvie, 2002). The consistently strong and more precise dose-response relationships observed with the BaP metric provides support for the use of BaP as a measure of exposure to the PAH mixture in examinations of cancer and heart disease in aluminum smelters. The strength of the comparison between dose-response relationships presented here was the substantial number of cases for each outcome, the examination of the continuous dose-response relationships, and the measurement-based exposure assessment strategy that independently assessed BaP and BSM exposure levels. The leveling off of risk at high exposure categories may result from exposure misclassification, but may also result from time-varying differences between BaP and other specific causal exposure agents, or suggest that different causal components may be associated with the different disease outcomes. Morbidity outcomes, such as hospitalization for respiratory or cardiovascular disease, may provide a better comparison of the usefulness of BaP and BSM as indices of coal tar pitch volatiles exposure as there is more confidence in the exposure estimates for more recent time periods. The relationship between BaP and other coal tar pitch volatiles constituents and factors that influence that relationship should continue to be examined to shed light on the specific causal agents. 58 Table 4.1 Exposure category cut points and mean cumulative exposure for BSM and BaP by health outcome Cumulative BaP (ug/m3-yr) Cumulative B S M (mg/m3-yr) Bladder Cancer Incidence Lung Cancer Incidence Acute Myocardial Infarction Bladder Cancer Incidence Lung Cancer Incidence Acute Myocardial Infarction Category Cut Points" Unexposed 0.50 0.50 0.50 0.05 0.05 0.05 20 t h %ile 7.63 11.32 6.21 0.70 0.89 0.65 40 t h %ile 19.99 27.25 20.99 2.14 3.61 2.43 60 t h %ile 38.90 52.05 42.45 5.30 7.45 5.77 80 , h %ile 69.61 81.70 80.61 11.80 12.41 12.42 Mean Cumulative Exposureh 1 (Unexposed) 0 0 0 0 0 0 2 3.5 4.9 3.1 0.33 0.40 0.33 3 13.3 18.4 13.1 1.33 1.97 1.44 4 28.4 37.8 30.3 3.57 5.18 3.89 5 51.6 64.7 57.7 7.94 9.49 8.31 6 108.3 119.6 122.0 20.25 20.87 21.31 a Category cut points based on cumulative exposure of cases. b Mean cumulative exposure of all subjects, weighted by person-years contributed by subject. 59 Table 4.2 Model parameters and precisions for the relationships between BaP and BSM exposure and bladder cancer, lung cancer, and acute myocardial infarction Cumulat ive B S M (mg/m3-yr) Cumulat ive B a P (ug/m'-yr) Change in -21oglikelihood° B S E B / S E Change in -21oglikeIihood° B S E B / S E Bladder Cancer Incidence, 20 yr lag Log-L inear M o d e l * 7.92 L o g - L o g M o d e l ' 9.95 0.0446 0.3323 0.0150 0.1060 2.97 3.13 10.77 8.90 0.0095 0.2082 0.0028 0.0713 3.39 2.92 Lung Cancer Incidence, 20 yr lag Log-L inear Mode l 3.90 L o g - L o g Mode l 4.67 0.0248 0.1744 0.0122 0.0804 2.03 2.17 3.22 5.31 0.0039 0.1162 0.0021 0.0511 1.86 2.27 Acute Myocardial Infarction, no lag Log-L inear Mode l 1.52 L o g - L o g Mode l 1.84 0.0124 0.0943 0.0098 0.0691 1.27 1.36 1.32 1.88 0.0020 0.0611 0.0017 0.0447 1.18 1.37 RR=Relative Risk, B=slope parameter from model, CE=cumulative exposure, SE=standard error, B/SE=precision a Change in - 21oglikelihood compared to model with age, calendar year and smoking status, but no exposure b Log-linear: LogRR = B * CE; c Log-log: LogRR = B * ln(CE + 1) 60 Figure 4.1 Relationship between cumulative BaP and BSM exposure indices 50 -i i r 45 -X 40 -300 400 Cumulative BaP (ug/m3-yr) 61 Figure 4.3 Log-linear (—) and log-log (—) relationships between exposure and lung cancer incidence, 20 year lag Figure 4.4 Log-linear (—) and log-log (—) relationships between exposure and acute myocardial infarction mortality, no lag 62 References Armstrong B, Tremblay C, Baris D, Theriault G (1994) Lung cancer mortality and polynuclear aromatic hydrocarbons: A case- cohort study of aluminum production workers in Arvida, Quebec, Canada. Am J Epidemiol; 139 250-62: Armstrong BG, Tremblay C G , Cyr D, Theriault GP (1986) Estimating the relationship between exposure to tar volatiles and the incidence of bladder cancer in aluminum smelter workers. Scand J Work Environ Health; 12 486-93. Aubin S, Farant JP (2000) Benzo[b]fluoranthene, a potential alternative to benzo[a]pyrene as an indicator of exposure to airborne pahs in the vicinity of Soderberg aluminum smelters. J Air Waste Manag Assoc; 50 2093-101. Benke G, Abramson M , Sim M (1998) Exposures in the alumina and primary aluminium industry: An historical review. Ann Occup Hyg; 42 173-89. Dosemeci M , Wacholder S, Lubin JH (1990) Does nondifferential misclassification of exposure always bias a true effect toward the null value? Am J Epidemiol; 132 746-8. Farant JP, Gariepy M (1998) Relationship between benzo(a)pyrene and individual polycyclic aromatic hydrocarbons in a Soderberg primary aluminum smelter. Am Ind Hyg Assoc J; 59 758-65. Farant JP, Ogilvie D (2002) Investigation of the presence of amino and nitro polycyclic aromatic hydrocarbons in a Soderberg primary aluminum smelter. Am Ind Hyg Assoc J; 63 721-5. Kromhout H, Loomis DP, Kleckner RC (1999) Uncertainty in the relation between exposure to magnetic fields and brain cancer due to assessment and assignment of exposure and analytical methods in dose-response modeling. Ann N Y Acad Sciences; 895 141-55. Kromhout H, Loomis DP, Kleckner RC, Savitz DA (1997) Sensitivity of the relation between cumulative magnetic field exposure and brain cancer mortality to choice of monitoring data grouping scheme. Epidemiol; 8 442-5. Loomis A, Kromhout H, Kleckner RC, Savitz DA (1998) Effects of the analytical treatment of exposure data on associations of cancer and occupational magnetic field exposure. Am J Ind Med; 34 49-56. NTP (2005) Polycyclic aromatic hydrocarbons. Report on carcinogens, eleventh edition. U.S. Department of Health and Human Services, Public Health Service. Rappaport SM, Waidyanatha S, Serdar B (2004) Naphthalene and its biomarkers as measures of occupational exposure to polycyclic aromatic hydrocarbons. J Environ Monit; 6 413-6. 63 Richardson DB, Loomis D (2004) The impact of exposure categorisation for grouped analyses of cohort data. Occup Environ Med; 61 930-5. Romundstad P, Andersen A, Haldorsen T (2000) Cancer incidence among workers in six Norwegian aluminum plants. Scand J Work Environ Health; 26 461-9. Romundstad P, Haldorsen T, Andersen A (2000) Cancer incidence and cause specific mortality among workers in two Norwegian aluminum reduction plants. Am J Ind Med; 37 175-83. Ronneberg A (1995) Mortality and cancer morbidity in workers from an aluminium smelter with prebaked carbon anodes-part I: Exposure assessment. Occup Environ Med; 52 242-9. Ronneberg A, Andersen A (1995) Mortality and cancer morbidity in workers from an aluminium smelter with prebaked carbon anodes—part II: Cancer morbidity. Occup Environ Med; 52 250-4. Ronneberg A, Haldorsen T, Romundstad P, Andersen A (1999) Occupational exposure and cancer incidence among workers from an aluminum smelter in western Norway. Scand J Work Environ Health; 25 207-14. Roussel R, Gaboury A, Lariviere C (1991) Aromatic amines in the workplace atmosphere of an aluminum smelter. Light Metals; 503-7. Sanderson E, Kelly P, Farant JP (2005) Effect of Soderberg smelting technology, anode paste composition, and work shift on the relationship between benzo[a]pyrene and individual polycyclic aromatic hydrocarbons. J Occup Environ Hyg; 2 65-72. Skogland M (1991) A survey of the PAH problem in the aluminum industry. Light Metals; 497-502. Spinelli JJ, Band PR, Svirchev L M , Gallagher RP (1991) Mortality and cancer incidence in aluminum reduction plant workers. J Occup Med; 33 1 150-5. Spinelli JJ, Demers PA, Le ND, Friesen M C , Lorenzi MF, Fang R, Gallagher RP (2006) Cancer incidence and mortality in aluminum reduction plant workers. Submitted. Stayner L, Steenland K, Dosemeci M , Hertz-Picciotto I (2003) Attenuation of exposure-response curves in occupational cohort studies at high exposure levels. Scand J Work Environ Health 29 317-24. Steenland K, Deddens JA (2004) A practical guide to dose-response analyses and risk assessment in occupational epidemiology. Epidemiology; 15 63-70. Straif K, Baan R, Grosse Y, Secretan B, El Ghissassi F, Cogliano V (2005) Carcinogenicity of polycyclic aromatic hydrocarbons. Lancet; 6 931-932. Theriault GP, Tremblay CG, Armstrong BG (1988) Risk of ischemic heart disease among primary aluminum production workers. Am J Ind Med; 13 659-66. 64 Tremblay C, Armstrong B, Theriault G, Brodeur J (1995) Estimation of risk of developing bladder cancer among workers exposed to coal tar pitch volatiles in the primary aluminum industry. Am J Ind Med; 27 335-48. 65 Chapter 5: Impact of the specif ic i ty of the exposure metric on exposure-response relat ionships in a large retrospective occupat ional c o h o r t ^ Introduction Exposure-response relationships are sensitive to the choice of exposure indicator, as a result studies may result in spurious or inconclusive findings if the exposure indicator chosen does not adequately reflect the etiological agent (Armstrong, White et al, 1994; Loomis et ai, 1999). Theoretically, the exposure indicators should be based on toxicological mechanisms of the exposure and disease being studied (Ahrens and Stewart, 2003). In reality, the exposure indicators are often chosen based on the types of exposure measurements available, but this does not guarantee that the measurements capture the relevant components (Ahrens and Stewart, 2003). Exposure to mixtures is a particular challenge, where analytic methods may be nonspecific and provide a measure of the mixture without regard for the varying toxicities of the components. A common example of this is assessing exposure to airborne particles using simple gravimetric methods; for instance, measuring all dust when only wood dust is of interest. Pilot studies examining the relationship between the available indicator and the more proximal measure may provide valuable information on the potential extent of exposure misclassification. In addition, the pilot studies can be used to refine the available exposure estimates to reflect the more proximal measure. Examples of such studies include those used to calculate the wood dust fraction of particulate exposure in sawmills (Teschke et al, 1999), the benzo(a)pyrene fraction of benzene soluble materials (Armstrong et al, 1986; Armstrong, Tremblay et al, 1994), and to convert from one particulate size fraction to another (Daviess a/., 1999). The attenuation and distortion of exposure-response relationships due to exposure misclassification has been examined using simulation studies (Dosemeci et al, 1990; Flegal et al, 1991; Veierod and Laake, 2001; Richardson and Loomis, 2004), but there is limited empirical examples due to the rarity of multiple exposure assessment approaches being available within the same study. In this study we examined how moving from two measures of exposure mixtures (total particulate, total A version of this chapter has been accepted for publication. Friesen MC, Demers PA, Davies HW, Teschke K (2007) Impact of the Specificity of the Exposure Metric on Exposure-Response RelationshipsEpidemiol; 18(l):88-94. 66 chlorophenols) to more specific exposure indicators (wood dust, pentachlorophenol and tetrachlorophenol) impacted exposure-response relationships in a large retrospective cohort of sawmill workers. For both classes of exposure, moving to more specific exposure agents required substantial additional effort in the retrospective exposure assessment. We used the refined exposure estimates to examine the relationships of particulate and wood exposures to hospitalizations for chronic obstructive pulmonary disease (COPD), and to examine the relationships of total chloropherfol and specific chlorophenols to the incidence of non-Hodgkin lymphoma (NHL) and kidney cancer. Methods Study Population The study population consisted of 26,847 male workers employed for a minimum of one year between 1950 and 1995 at one of 14 sawmills in British Columbia, Canada, and has been described elsewhere (Demers et al, 2006). The cohort members' personal identification and work histories were abstracted from company records. Cohort members were linked using probabilistic linkage techniques to the Canadian Cancer Incidence Data Base (1969-1995). A subcohort of 11,273 sawmill workers with a minimum of one year of employment and at least one day worked between 1985 and 1995 were linked through the BC Client Registry administered by the BC Ministry of Health (records of health care recipients on the provincial medical system since 1984). Through the BC Client Registry, all hospitalizations incurred by the subcohort were identified and their diagnosis codes (ICD-9) were ascertained (1985-1998). Vital status for the full cohort was ascertained using linkages to the national mortality database, sawmill pension records, BC motor vehicle license records, and union records. Those with social insurance numbers were also linked to Revenue Canada's income tax file. The focus of this paper was non-Hodgkin lymphoma (NHL) incidence (92 cases) and kidney cancer incidence (79 cases) for the full cohort as significant exposure-response relationships were observed in categorical analyses and there were a substantial number of cases available to examine continuous exposure-response relationships with chlorophenol exposure.17 In the subcohort we examined the relationship between hospitalization for chronic obstruction pulmonary disease (COPD, 132 cases) and wood dust. Exposure Assessment The exposure assessment for non-specific particulate and wood dust is summarized in detail elsewhere (Friesen et al, 2005). Briefly, Job exposure matrices (JEM) for non-specific particulate 67 and wood dust were constructed from predictive models that used 1399 inhalable particulate measurements collected from 1981-1997. For the wood dust J E M , the model was constructed after applying adjustment factors to measurements to subtract the estimated mass of non-wood particulates. A l l sawmills in this study processed only softwoods. The wood dust adjustment factors were based on a study that used the abietic acid content, a wood extractive common to softwood trees, to estimate the proportion of non-specific particulate that was wood dust (Demers et al, 2000). Industrial hygienists assigned each job at the 14 sawmills to one of 4 categories o f proportions of wood dust (10, 30, 70, and 100%), based on proximity to dust sources and relative abietic acid content o f the particulate in those jobs. The predictive models, together with historical data gathered on the exposure determinants included in the models, were used to predict particulate and wood dust exposures for all cohort subjects and all time periods. The exposure assessment for chlorophenols has been described elsewhere (Demers et al, 2006), and is summarized here. Sodium salts of pentachlorophenol (PCP) and tetrachlorophenol (TCP) were used extensively in coastal sawmills from 1950 to 1990 to control the growth of sap-staining fungi (Teschke et al, 1994). The relative proportions o f PCP and T C P in the fungicide formulations varied over time and the specific formulations and the dates of their use were identified from mi l l records. To assess hours of dermal exposure per year, structured interviews were conducted with senior workers at each mil l with at least 5 years experience in each time period during which processes and formulations remained constant (9-20 workers per period). The senior workers' exposure ratings were reliable (group intraclass correlation coefficients=0.69 to 0.94) (Teschke etal, 1996) and correlated with urinary chlorophenol levels (0.47 to 0.76) (Hertzman et al, 1988; Teschke et al, 1989). To create specific exposure indices for P C P and T C P , the proportions of P C P and T C P in the fungicide formulations were used as weights to adjust the senior workers' estimates of hours of dermal exposure per year. In the analyses, cumulative exposure to chlorophenols was quantified as full-time equivalent (FTE) years where 1 F T E equaled 2000 hours of dermal exposure. Data Analysis The correlations between the metrics were examined using Pearson correlation (r) and partial correlation (r p) on the untransformed and log-transformed continuous cumulative exposure estimates. The percent agreement between the exposure category assignments (6 categories, described below) was calculated. The level of agreement between the categorical assignments of the different metrics was calculated using weighted kappa (K w) using quadratic weights where closer agreement in categories was penalized less than larger discrepancies (Armstrong, White et al, 1994). 68 Dose-response relationships were examined using internal comparisons within the cohort, where the workers in the lowest exposure category served as the reference group. The cumulative exposure indices were lagged according to the best fitting latency periods observed in categorical analyses reported elsewhere (Demers et al, in preparation; Demers et al, 2006). For COPD hospitalizations no lag was used; for NHL and kidney cancer incidence exposure was lagged 20 years. The Life Table Analysis System (PCLTAS, NIOSH) software was used to generate files for analysis. Six exposure' categories were used: an unexposed group and five exposed groups with cumulative exposure cut points defined by distributing the remaining cases into five equal size groups. Relative risks were calculated using maximum likelihood methods after adjustment for age, calendar year, and race using Poisson regression (SAS version 8.0, SAS Institute Inc., Cary North Carolina, USA). Race was based on surname; for the full cohort 5.9% were classified as South Asians and 1.6% were classified as East Asian. Ten-year calendar period and age categories were used; to improve precision, categories were combined when necessary to ensure all categories had a minimum often cases. Exposure was treated as a continuous variable in Poisson regression by assigning the mean cumulative exposure to all subjects in each category. Two dose-response model structures were assessed: the log-linear model, where the relative risk, RR, increased logarithmically with a linear increase in cumulative exposure, CE, after subtracting the mean exposure level, x,„ of the reference group [Ln(RR) = (3*(CE-x„)]; the log-log model, where the relative risk increased logarithmically with log-transformed cumulative exposure [Ln(RR) = P*ln(CE-;c„ +1)] (Steenland and Deddens, 2004). Ninety-five percent confidence intervals were based on the standard error of the coefficients derived from the model. To examine the fit of the exposure-response relationships we compared two measures: 1) the change in the -2 log likelihood model fit statistic with and without exposure in Poisson regression (1 degree of freedom), while adjusting for age, calendar year, and race;(Loomis et al, 1999) and 2) the precision of the dose-response relationship based on the Wald statistic (slope/standard error) (Kromhout et al., 1997). A Wald Statistic equal or greater than 1.96 corresponds to a p-value of 0.05 or lower based on one degree of freedom. We also compared the observed attenuation of the slope of the exposure-response relationship with the expected attenuation assuming that the more specific marker was the "gold standard". The expected slope for the less precise exposure indicator, /52, is a function of the slope of the more 69 precise exposure indicator, pi, and the correlation between the two exposure indicators, pi2\ Pi= Pn Pi (Armstrong, White et ai, 1994). The expected attenuation was calculated using the correlation between the untransformed continuous cumulative exposure metrics for the log-linear models and with the log-transformed continuous cumulative exposure metrics for the log-log models. The expected and observed slopes were then used to calculate the percentage attenuation in relative risk for the 80 l h percentile of exposure. r Results Non-specific Particulate vs. Wood Dust For the subcohort the median and maximum cumulative non-specific particulate exposure was 9.8 and 220 mg/m3-yr, respectively. For wood dust the median and maximum cumulative exposure was 6.8 and 89 mg/m3-yr, respectively. There was moderately-high correlation between the continuous cumulative non-specific particulate and wood dust (untransformed: r=0.68, rp=0.46; log-transformed: r=0.82, rp=0.63). Categorization of cumulative exposure based on the distribution of cases resulted in moderate agreement between exposure categories for particulate and wood dust (Kw=0.67; percent agreement=57%). In the sawmill environment, particulate exposure is ubiquitous and no workers were unexposed; the reference exposure group was defined as less than 5.0 mg/m3-years to ensure a minimum of ten cases. In exposure-response analyses the mean exposure of the reference group (2.7 and 2.3 mg/m3-years for non-specific particulate and wood dust, respectively) was subtracted from the mean cumulative exposure of the other exposure categories. The proportion of workers assigned to the reference group was 30% for non-specific particulate and 40% for wood dust. The highest exposure category was defined based on the 80"' percentile of cumulative exposure of the exposed cases, with cut points of >32 mg/m3-yr for non-specific particulate and >27 mg/m3-yr for wood dust. There was limited evidence of a relationship between non-specific particulate and COPD hospitalizations with minimal change in model fit, very poor model precision, and no consistent increase in risk across exposure categories (Figure 5.1A, Table 5.1). However, wood dust demonstrated a roughly monotonic exposure-response relationship with increasing exposure (Figure 5.1B). The log-log model for wood dust provided the best model fit based on change in -2 log likelihood and the precision of the exposure-response slope (Table 5.1). 70 Chlorophenols For the full cohort, maximum cumulative exposure was 29, 21, and 11 FTE for total chlorophenol, PCP, and TCP, respectively. There was a moderately-high correlation between the continuous cumulative exposure estimates of total chlorophenol and either PCP (untransformed: r=0.88, rp=0.82; log-transformed: r=0.83, rp=0.75) or TCP (untransformed: r=0.78, rp=0.71; log-transformed: r=0.96, rp=0.77)). Exposure to the two specific chlorophenols was moderately associated (untransformed: r=0.39, rp=0.26; log-transformed: r=0.46, rp=0.63). Categorization of cumulative exposure based on the distribution of cases resulted in moderate agreement between exposure categories for total chlorophenols and PCP (Kw=0.67; percent agreement=66%) and for total chlorophenols and TCP (Kw=0.60; percent agreement: 60% for NHL, 56%o for kidney cancer). The reference group was defined as less than 0.25 FTE. Because the mean exposure for the reference group was negligible, it was not subtracted from the cumulative exposure in exposure-response analyses. The proportions of workers assigned to the reference group were 40, 58, and 60% for total chlorophenol, PCP, and TCP, respectively. The highest exposure category was based on the 80th percentile of cumulative exposure for the exposed cases, with cut points of >6.2 FTE for total chlorophenol, >5.9 FTE for PCP, and >3.5 FTE for TCP. For kidney cancer, the highest exposure category cut points were >5.6 FTE for total chlorophenol, >5.0 FTE for PCP, and >2.5 FTE for TCP. The risk of N H L incidence increased roughly monotonically with exposure to total chlorophenols and PCPs (Figure 5.2A, 5.2B), but not with TCP (not shown). The best fitting and most precise exposure-response relationship for N H L was with PCP, with improvements of 10%> in model precision over total chlorophenol (Table 5.2). For PCP, the log-linear and log-log models provided nearly identical measures of goodness of fit and precision. TCP was not associated with NHL, with negligible improvement in model fit based on 1 degree of freedom and a very imprecise exposure slope. Similarly, the risk of kidney cancer incidence increased roughly monotonically with total chlorophenol and PCP (Figure 5.3A, 5.3B), but not with TCP (not shown). The log-log model for PCP was the best fitting and most precise exposure-response relationship for kidney cancer, with improvements of 30% in model precision over total chlorophenols (Table 5.2). TCP exposure was not associated with increased risk of kidney cancer, with negligible improvement in model fit based on 1 degree of freedom and very imprecise exposure slopes. 71 Observed vs. Expected Attenuation The correlations between exposure metrics were used to calculate the expected attenuations in the relative risk; these were compared to the observed attenuations for the 80th percentile exposure for exposed cases. For the dust-COPD models, the observed attenuations in relative risk were nearly two times greater than expected (log-linear model: 25% observed vs. 14% expected; log-log model: 34% observed vs. 18% expected). For the chlorophenol-cancer relationships, the log-linear models also showed observed attenuations twice the expected (NHL: 16% observed vs. 8% expected; kidney cancer: 9% observed vs. 13% expected), but the log-log models had much smaller than expected observed attenuation in relative risk (NHL: 7% observed vs. 17% expected; kidney cancer: 12% observed vs. 19% expected). Discussion This study demonstrates the importance of developing exposure metrics as specific to the disease-causing agent as possible, particularly when the composition of mixed exposures varies by work areas, jobs, and time periods. For both of the sawmill exposures examined here, the more specific indicator was more strongly associated with the health outcomes of interest than the less specific measure. The stronger the correlation between the nonspecific and specific exposure indicators, the less attenuation we observed in the exposure-response relationship with the use of the non-specific exposure indicator. We did not observe a consistent association between the predicted and expected attenuation of the exposure-response relationships in this study, as we observed attenuation both less than and greater than expected depending on the exposure. The continuous exposure-response relationships presented here were the result of using grouped analyses to apportion person-time prior to modeling the relationship. This is a common and often necessary approach in cohort studies using Poisson regression (Checkoway et al, 2004), but this violation from a true continuous analysis may account for the lack of agreement between observed and expected attenuation. Basing exposure groups on the exposure distribution among cases, as was used here, has been found to minimize bias introduced by group-based analyses in Poisson regression, but not necessarily completely eliminate it (Richardson and Loomis, 2004). Nevertheless, we observed attenuation up to two times greater than expected based on the correlation in exposure metrics, which provides a clear caution to not rely on the 72 predicted attenuation from pilot studies and instead use the most proximal indicator of exposure wherever possible. Choosing the best indicator has been considered an "exceedingly ambitious goal" and is dependent on both the current state of knowledge about the association and empirical evidence, in addition to the specific goal of the study (Loomis et al, 1999). Several performance measures have been suggested to identify the "best" indicators of exposure, including magnitude of risk, model goodness-of-fit, and precision of the slope (Blair and Stewart, 1992; Kromhout et al, 1997; Loomis et al, 1999, Blair, 1992 #28), although some concern has been raised about the validity of using the magnitude of the risk (Salvan et al, 1995). In this study, all these measures consistently indicated that, for the health outcomes examined, wood dust and pentachlorophenol were better exposure indicators than the corresponding measures of the mixtures. This study does not assess the impact of exposure misclassification when the more specific, or proximal, indicator is measured with error. For all exposures and health outcomes, the log-log model provided the better fitting and more precise exposure-response relationship. This leveling off of risk at the highest exposure categories has been observed in many retrospective studies and is a likely result of nondifferential exposure misclassification (Dosemeci etal, 1990); although other possible explanations, such as confounding and the healthy worker survivor effect, may contribute to this exposure-response shape (Stayner et al, 2003). We were unable to account for smoking status, which may be an important confounder, but the results from a smoking survey in this population found that smoking rates were not correlated with exposure (Demers et al, 2006). Nondifferential misclassification beyond what was already inherent in the less specific exposure indicators was potentially introduced in the group-based weights used to convert nonspecific particulate to wood dust and from total chlorophenols to PCP and TCP. While strong exposure-response relationships were observed for both the specific exposure indicators, wood dust and PCP, it is likely that neither metric is a perfect indicator of the proximal causes of the diseases in question. The association between wood dust and COPD hospitalizations may be due to extractives in the wood, such as resin acids and monoterpenes (Demers et al, 1997). Some of these extractives have been measured in a sawmill environment and are known to vary in exposure by tree species and work area (Teschke et al, 1999; Demers et al, 2000), but at this time there is insufficient information to use these markers of exposure for a retrospective epidemiologic study. Assessment of chlorophenol exposure was limited to estimates of contact hours, but the true dose depends on the 73 availability of the chlorophenols to be absorbed by the skin and would vary by job tasks, proportion of exposed skin, choices of work clothes, hygiene behaviors, and other factors. In addition, while increased risk of disease was observed using the more specific exposure metrics, the diseases studied are likely multi-causal illnesses. Nevertheless, the significant associations observed here improve our knowledge regarding the potential etiologic agents. Exposure assessment for epidemiologic studies continues to be challenged by {imitations in analytic methods and a paucity of measurements of exposures to specific etiological agents, especially for exposure mixtures. The use of more proximal measures of the hypothesized etiological agents resulted in larger risk estimates and better fitting and more precise exposure-response relationships in this study, whereas the more commonly used nonspecific measures resulted in much weaker or inconclusive associations. This study demonstrated the benefit from using exposure measurements and expert exposure estimates together with supplementary exposure information, such as a pilot study examining the work area differences in the wood dust component of particulate exposures or historical mill records that record chemical formulations. Our knowledge of the etiologic agents in epidemiologic studies will benefit from examining exposure-response relationships with multiple exposure indicators that provide independent information. 74 Table 5.1 Model parameters and precisions for the relationships between dust and wood dust exposure and hospitalizations for COPD Change in -2 log likelihood" B S E Model Precision B / S E R R @ 80 t h percentile exposure o f cases Total Particulate Log-Linear Mode l 0.1 0.0019 0.0064 0.30 1.06 L o g - L o g Model 0.7 0.0730 0.0884 0.83 1.28 Wood Dust Log-Linear Mode l 3.2 0.0141 0.0077 1.83 1.41 L o g - L o g Model 7.4 0.2032 0.0771 2.64 1.93 RR=Relative Risk, B=slope parameter from model with units (mg/nr'-year)" 1, CE=cumulative exposure with units mg/nr'-year, SE=standard error, B/SE=precision a Change in - 2loglikelihood compared to model with age, calendar year and race, but no exposure b Log-linear: L n R R = B * ( C E - x 0 ) ; x 0 = average exposure in reference group c Log-log: L n R R = B * ln (CE- x 0 + 1) 75 Table 5.2 Model parameters and precisions for the relationships between chlorophenol exposure and NHL and kidney cancer Non-Hogkin lymphoma, 20 year lag Kidney Cancer, 20 yr lag Chang ;e in B SE Model RR @ 80th Change in B SE Model RR @ 80* -2 log Precision percentile -2 log Precision percentile likelihood" B/SE exposure likelihood" B/SE exposure of cases of cases Total chlorophenol Log-Linear Model* 1.6 0.0438 0.0335 1.31 1.31 0.5 0.0272 0.0360 0.76 1.16 Log-Log Modef 2.7 0.2414 0.1444 1.67 1.61 1.6 0.1918 0.1497 1.28 1.43 Pentachlorophenol Log-Linear Model 3.2 0.0750 0.0408 1.84 1.56 1.6 0.0587 0.0446 1.32 1.34 Log-Log Model 3.3 0.2831 0.1551 1.83 1.73 2.7 0.2700 0.1619 1.67 1.62 Tetrachlorophenol Log-Linear Model 1.0 0.1085 0.1044 1.04 1.46 0.4 0.0682 0.1074 0.64 1.19 Log-Log Model 0.7 0.2204 0.2637 0.84 1.39 0.7 0.2126 0.2551 0.83 1.30 RR=Relative Risk, B=slope parameter from model with units FTE"1, CE=cumulative exposure with units FTE, SE=standard error, B/SE=precision a Change in - 21oglikelihood compared to model with age, calendar year and race, but no exposure b Log-linear: LogRR = B * CE; c Log-log: LogRR = B * ln(CE + 1) Figure 5.1 Log-linear (—) and log-log (—) relationships between dust (A) and wood dust (B) exposure and COPD hospitalizations 10.0 1.0 4 13 C u m i i l t i t i v e D u s t (IIKJ/III 77 Figure 5.2 Log-linear (—) and log-log (—) relationships between total chlorophenol (A) and pentachlorophenol (B) exposure and NHL, 20 year lag B 0 2 4 6 8 10 12 Cumulative Pentachlorophenol Exposure (FTE) 78 Figure 5.3 Log-linear (—) and log-log (—) relationships between total chlorophenol (A) and pentachlorophenol (B) exposure and kidney cancer, 20 year lag 10 Cd 0) CD A 01 .1 . . . —. 1 0 2 4 6 8 10 Cumulative Total Chlorophenol Exposure (FTE) 0.1 -I , , , , 1 0 2 4 6 8 10 Cumulative Pentachlorophenol Exposure (FTE) 79 References Ahrens W, Stewart P (2003) Retrospective exposure assessment. Exposure assessment in occupational and environmental epidemiology. M . J . Nieuwenhuijsen. Oxford, Oxford University Press: 104. Armstrong B, Tremblay C, Baris D, Theriault G (1994) Lung cancer mortality and polynuclear ' aromatic hydrocarbons: A case- cohort study of aluminum production workers in Arvida, Quebec, Canada. Am J Epidemiol; 139 250-62. Armstrong BG, Tremblay C G , Cyr D, Theriault GP (1986) Estimating the relationship between exposure to tar volatiles and the incidence of bladder cancer in aluminum smelter workers. Scand J Work Environ Health; 12 486-93. Armstrong BK, White E, Saracci R (1994) Principles of exposure assessment in epidemiology. Oxford: Oxford University Press. Blair A, Stewart PA (1992) Do quantitative exposure assessments improve risk estimates in occupational studies of cancer? Am J Ind Med; 21 53-63. Checkoway H, Pearce N, Kriebel D (2004) Research methods in occupational epidemiology. Oxford: Oxford University Press. Davies HW, Teschke K, Demers PA (1999) A field comparison of inhalable and thoracic size selective sampling techniques. Ann Occup Hyg; 43 381-92. Demers PA, Davies HW, Friesen M C , Hertzman C, Ostry A, Hershler R, Teschke K (2006) Cancer and occupational exposure to pentachlorophenol and tetrachlorophenol. Cancer Causes Control; In Press. Demers PA, Teschke K, Davies HW, Kennedy SM, Leung V (2000) Exposure to dust, resin acids, and monoterpenes in softwood lumber mills. Am Ind Hyg Assoc J; 61 521-8. Demers PA, Teschke K, Kennedy SM (1997) What to do about softwood? A review of respiratory effects and recommendations regarding exposure limits. Am J Ind Med; 31 385-98. Dosemeci M , Wacholder S, Lubin JH (1990) Does nondifferential misclassification of exposure always bias a true effect toward the null value? Am J Epidemiol; 132 746-8. Flegal K M , Keyl PM, Nieto FJ (1991) Differential misclassification arising from nondifferential errors in exposure measurement. Am J Epidemiol; 134 1233-44. Friesen M C , Davies HW, Teschke K, Marion S, Demers PA (2005) Predicting historical dust and wood dust exposure in sawmills: Model development and validation. J Occup Environ Hyg; 2 650-8. 80 Hertzman C, Teschke K, Dimich-Ward H, Ostry A (1988) Validity and reliability of a method for retrospective evaluation of chlorophenate exposure in the lumber industry. Am J Ind Med; 14 703-13. Kromhout H, Loomis DP, Kleckner RC, Savitz DA (1997) Sensitivity of the relation between cumulative magnetic field exposure and brain cancer mortality to choice of monitoring data grouping scheme. Epidemiol; 8 442-5. Loomis D, Salvan A, Kromhout H, Kriebel D (1999) Selecting indices of occupational exposure for epidemiologic studies. Occup Hyg; 5 73-91. Richardson DB, Loomis D (2004) The impact of exposure categorisation for grouped analyses of cohort data. Occup Environ Med; 61 930-5. Salvan A, Stayner L, Steenland K, Smith R (1995) Selecting an exposure lag period. Epidemiol; 6 387-90. Stayner L, Steenland K, Dosemeci M , Hertz-Picciotto I (2003) Attenuation of exposure-response curves in occupational cohort studies at high exposure levels. Scand J Work Environ Health; 29 317-24. Steenland K, Deddens JA (2004) A practical guide to dose-response analyses and risk assessment in occupational epidemiology. Epidemiology; 15 63-70. Teschke K, Demers PA, Davies HW, Kennedy SM, Marion SA, Leung V (1999) Determinants of exposure to inhalable particulate, wood dust, resin acids, and monoterpenes in a lumber mill environment. Ann Occup Hyg; 43 247-55. Teschke K, Hertzman C, Dimich-Ward H, Ostry A, Blair J, Hershler R (1989) A comparison of exposure estimates by worker raters and industrial hygienists. Scand J Work Environ Health; 15 424-9. Teschke K, Marion SA, Jin A, Fenske RA, van Netten C (1994) Strategies for determining occupational exposures in risk assessments: A review and a proposal for assessing fungicide exposures in the lumber industry. Am Ind Hyg Assoc J; 55 443-9. Teschke K, Marion SA, Ostry A, Hertzman C, Hershler R, Dimich-Ward H, Kelly S (1996) Reliability of retrospective chlorophenol exposure estimates over five decades. Am J Ind Med; 30 616-22. Veierod MB, Laake P (2001) Exposure misclassification: Bias in category specific poisson regression coefficients. Stat Med; 20 771-84. 81 Chapter 6: Comparison of expert and measurement-based exposure assessment of historical noise levels for a sawmill cohort1 Introduction Minimizing exposure misclassification that can result in attenuated and distorted exposure-response relationships is an important aspect of epidemiologic studies (Dosemeci et al, 1990). Developing quantitative exposure estimates helps reduce exposure misclassification by differentiating jobs and time periods by the intensity of exposure, often resulting in the ability to detect associations that would otherwise be missed (Stewart and Herrick, 1991). While measurement-based exposure assessment approaches are preferred, experts such as occupational hygienists, engineers, and senior workers have been frequently asked to assess semi-quantitative and quantitative exposure levels for epidemiologic studies when exposure measurements are limited or lacking. Because expert-based exposure assessment may be the only option in many epidemiologic studies, the ability of experts to accurately estimate exposure levels has received much attention. The results from studies testing experts' exposure estimates against exposure measurements have been mixed. While experts' estimates are often only poorly correlated with exposure measurements, they are often able to successfully rank the order of exposed jobs (Kromhout et al, 1987; Hertzman et al, 1988; de Cock et al, 1996). Experts' exposure estimates were often more accurate when exposure measurements were available to anchor the exposure estimates or when the experts were able to perceive the exposure of interest (Teschke el al, 2002). Many epidemiologic studies are now assessing the validity of the expert exposure ratings prior to their use (Hertzman et al, 1988; Astrakianakis et al, 1998), but regardless of the experts' accuracy many studies have used those estimates in epidemiologic analyses due to the lack of alternatives. To estimate the potential attenuation of exposure-response relationships due to exposure misclassification, theoretical equations based on the correlation between the exposure metric and a "gold standard" can be used (Armstrong et al, 1994). This approach has been used to select between different exposure assessment strategies, such as different grouping schemes (Tielemans et al, 1998; ' A version of this chapter has been submitted for publication. Friesen MC, Davies HW, Ostry A, Teschke K, Hertzman C, Demers PA (Submitted August 2006). 82 van Tongeren etal, 1999; Mwaiselage etal, 2005). However, in a recent study comparing less specific and more specific measures of exposure mixtures (i.e. total dust vs. wood dust) the attenuation observed in the epidemiologic analyses was greater than predicted from theoretical equations (Chapter 5). The validity studies used to estimate the potential attenuation are based on the assumption that the exposure measurements used are the gold standard, but the exposure measurements may themselves be "alloyed" due to non-optimized sampling conditions that may underestimate the true exposure variability (Stewart et al, 1996; Maulhausen and Damiano, 1998). Also, validity studies typically evaluated only the most recent time periods so the validity of the experts' estimates is not known across all time periods (Hertzman et al, 1988; Post et al, 1991; de Cock et al, 1996; Astrakianakis et al, 1998). Rarely are both expert- and measurement-based exposure assessment approaches available in the same study, thus there is limited empirical evidence of the robustness of exposure-response relationships to expert- versus measurement-based exposure assessment strategies. One such study found much higher relative risks (RR) for the measurement-based job exposure matrix than for an expert-based job exposure matrix (RR 2.27 vs. 1.53) for the highest exposure group in a study of magnetic fields and brain cancer (Kromhout et al, 1999). Even when such comparisons are made, it is difficult to directly compare exposure assessment approaches when cumulative exposure is treated as a categorical variable in epidemiologic analyses. The magnitude of the relative risks are sensitive to the category cut points and often only a qualitative evaluation of the consistency of the exposure-response trend across exposure categories can be made (Loomis et al, 1998; Richardson and Loomis, 2004). There is also evidence that the "best" exposure metric should not be chosen based on the magnitude of the relative risk due to the potential for nondifferential misclassification to create bias away from the null (Salvan et al, 1995). Instead, offering exposure as a continuous variable in analyses provides an opportunity to examine improvements in model fit and in the precision of the exposure variable's slope in the model (Loomis et al, 1999; van Tongeren etal, 1999). In this study we examined the sensitivity of the exposure-response relationship between cumulative noise exposure and acute myocardial infarction mortality using both an expert-based and measurement-based exposure assessment approach in a retrospective cohort of male sawmill workers. Occupational noise exposure has been associated with heart disease in this sawmill cohort using both expert-based (Ostry, 1998) and measurement-based exposure assessment (Davies et al, 2005). While both studies used the same study cohort, they differed in the specific heart disease outcome and in the analyses method used; thus the results are not directly comparable. 83 In this paper we examine the shape of the exposure-response relationship with cumulative exposure entered as a continuous variable. Improvements in model fit and precision of the exposure variable (slope parameter/standard error) are examined as measures of the performance of the expert- and measurement-based exposure assessment approaches. In addition we compare the observed attenuation with the theoretical attenuation. Methods Study Population The study population was originally enumerated for a study of fungicide exposure and cancer and later expanded to include other health outcomes and other exposures, including noise, stress and wood dust (Hertzman et al, 1997; Teschke et al, 1998). The study cohort consisted of 27,499 male sawmill workers employed one or more years between 1950 and 1995 in 14 sawmills in British Columbia (BC), Canada, with probablistic linkage to the Canadian mortality database (1950-1995). Vital status was ascertained using linkages to sawmill pension records, BC motor vehicle license records, union records, and tax records. In this substudy, we evaluated noise exposure and risk of acute myocardial infarction (AMI) mortality (ICD-9: 410; 910 deaths). The study protocol was approved by the Behavioural Research Ethics Board of the University of British Columbia. Expert Noise Assessment The expert exposure assessment team consisted of four job evaluators with over 20 years experience in BC sawmills. Since the late 1960s ajoint union/management job evaluation system has been in place in BC sawmills to evaluate jobs with respect to psychological and physical demand, control over skill use, control over decision making, and physical and other hazards. The four job evaluators were interviewed in 1997 and asked to rate a generic sawmill using a shortened version of a demand/control instrument to obtain scores for psychosocial variables for 54 job titles (Ostry, 1998; Ostry et al, 2001). They were asked to assess noise exposure by ranking the statement "The job was noisy" on a four point scale, with l=strongly disagree and 4=strongly agree. Rankings between integers were allowed. The noise estimates for each job were pooled and used in the expert-based noise job exposure matrix. A reliability study for a subset of 28 jobs found that the inter-rater reliability was moderately-high (group intraclass correlation coefficient: 0.79) (Ostry et al, 2001). The same reliability study found that the pooled expert noise estimates had a high to moderate correlation with pooled self-reports (r=0.71) and pooled mill-level estimates from senior workers (r=0.78). Expert-based estimates of cumulative noise exposure for each worker were calculated as a 84 product of the unit-less rating and duration in years, with units of years, weighted by rating (rating weighted-years). Measurement-based Noise Assessment Full-shift, personal noise dosimetry measurements (n=1900) from the participating sawmills were used to build empirical regression models that were used to predict noise exposure estimates for over 3800 unique mill/job/time period combinations (Davies, 2002; Davies et al, 2005). Data on the potential explanatory variables were gathered through interviews, site visits, and historical records. Cumulative exposure for each worker was calculated as a product of predicted exposure intensity (from the empirical models) and duration, with units of "dBA-year" on a logarithmic scale [dBA-year = 10*log E(10 d B A / 1 0 *time)]. Adjusting for Hearing Protection Device Use Noise dosimetry does not capture the hearing protection reduction from the use of hearing protection devices (HPD) that was common in this industry after 1978 and thus may systematically overestimate exposure. The modeled noise estimates were arithmetically adjusted for the use of hearing protection devices (HPD) using real-world hearing-protection efficacy data and the prevalence of device use among BC sawmill workers obtained from annual hearing test data (Davies, 2002; Davies et al, 2005). HPD use was not accounted for in the expert noise assessment. A subcohort of those whose sawmill employment terminated before June 30, 1970, believed to have had minimal HPD use, was defined to examine the performance of the modeled and expert noise exposure estimates prior to the regular and mandatory HPD use in the sawmills (n=8700, 520 AMI deaths). Statistical Analyses The expert- and measurement-based cumulative exposure indices were compared for the full and subcohort using Pearson and Spearman correlation. Measures of agreement (percent agreement and weighted kappa) for the categorical exposure assignments were examined. Exposure-response relationships were examined using internal comparisons within the cohort, where the reference group was the workers in the lowest exposure category. The reference group for the measurement-based noise estimates was defined in previous analyses as <100 dBA-years, which corresponds to 30 years at 85 dBA and was equivalent to the 25th percentile of the exposure 85 distribution of the cases (Davies et ai, 2005). For consistency and roughly equal size reference groups between exposure assessment methods, the reference group for the expert-based noise estimates was defined by the 25th percentile of the exposure distribution of the cases and was equivalent to <14.5 rating-weighted-years. The remaining cases were distributed into into five equal size groups for each metric. Each category was assigned the mean cumulative exposure weighted by person-years contributed by each study subject in that category (Table 6.1) (Richardson and Loomis, 2004). The Life Table Analysis System (PCLTAS, NIOSH) software was used to apportion person-years for analysis. Relative risks were calculated using maximum likelihood methods in Poisson Regression after adjustment for age, calendar year, and race (SAS version 9.1, SAS Institute Inc., Cary North Carolina, USA). Ten-year calendar and age periods were used; they were combined where necessary to ensure a minimum of ten cases per category. Exposure was offered both as a categorical exposure variable and as a continuous variable in separate analyses. To treat exposure as a continous variable, the mean cumulative exposure for each category was assigned. Two common shapes of the exposure-response model structures were examined (Equations 6.1-6.2), where RR is the relative risk, C E is the mean cumulative exposure for the exposure category, and x„ is the mean exposure level of the reference group. Log-linear model: Ln(RR) = p * ( C E - * „ ) [6.1] Log-log model: Ln(RR) = p * In(CE — x„+ 1) [6.2] The model fit and model precision of the exposure-response relationships using exposure as a continuous variable were used to compare the expert- and measurement-based noise estimates. Improvements in model fit were compared by examining the change in the model deviance with and without the exposure variable, while including all other variables. The model precision was calculated based on the Wald statistic: slope of the exposure variable divided by its standard error. A Wald statistic equal or greater than 1.96 corresponds to a p-value of 0.05 or lower based on one degree of freedom. The analyses treating exposure as a categorical variable were used to confirm that the simple parametric models were reasonable approximations of the shape of the exposure-response relationships. The expected attenuation from an imperfect exposure metric, RRx, can be calculated as a function of the true relative risk, RRj, and the correlation between the imperfect and true exposure measure, p2ix, 86 according to equation 6.3 (Armstrong et ah, 1994). The expected and observed attenuation assuming that the exposure measure with the strongest exposure-response relationships is the gold standard were compared. 2 RRX =RRrTX [ 6 3 ] Resul ts Comparison of Cumulative Exposure Measures The expert-based cumulative noise estimates had a moderately high correlation with the unadjusted measurement-based cumulative noise estimates for both the full cohort and subcohort, but were only moderately correlated with the HPD adjusted noise estimates for the full cohort (Table 6.2). The unadjusted and HPD adjusted measurement-based cumulative noise estimates also had a moderately-high correlation. The relationships between the cumulative measures were stronger when no linear relationship was assumed, which was not unexpected due to the logarithmic nature of the noise dosimetry units, dBA. However, regression equations describing the relationship between the two cumulative measures found that a linear relationship resulted in almost identical R 2 to the logarithmic relationship (data not shown). The strength of the correlation between exposure measures is driven by the duration of exposure aspect of the cumulative exposure: adjusting for exposure duration resulted in much lower partial correlations between the expert-based noise estimates and the unadjusted and HPD adjusted measurement-based noise estimates (0.32 and 0.12, respectively, in the full cohort). The partial correlation between unadjusted and HPD adjusted measurement-based noise estimates remained moderate (0.51) when exposure duration was taken into account. A plot of the expert-based versus measurement-based cumulative noise estimates illustrates the non-linearity between the measures (Figure 6.1). While only the cumulative exposure for the subcohort subjects is shown, the relationship had the same shape for the full cohort data. As the cumulative exposure was categorized for Poisson regression analyses, measures of agreement between the categorical assignments were also examined. There was stronger agreement, based on weighted kappa values, between the unadjusted and HPD adjusted measurement-based noise exposure categories (Table 6.2). In the full cohort the agreement between the expert-based exposure categories was stronger for the unadjusted measurement-based noise exposure categories than for the HPD adjusted noise exposure categories. In the subcohort only moderate agreement between expert-and measurement-based approaches was observed. 87 The proportion of agreement between the exposure categories when the different exposure assessment approaches are compared is shown in Figure 6.2. The six exposure categories were assigned values of 1 through 6 to determine the discrepancy in the categorical assignment between exposure assessment methods. For example, if one method assigned an individual to exposure category 3, a discrepancy of ± 2 would correspond to the comparison method having assigned an exposure category of 1 or 5. The proportion of individuals assigned to the same category was 51-72% depending on the exposure assessment methods (Figure 6.2). The greatest agreement between exposure categories was between the unadjusted and HPD adjusted measurement-based cumulative noise estimates, with 87% of the study subjects assigned to the same or within one exposure category. For the expert vs. measurement-based approaches, 78-80% of the study subjects were assigned to the same category or were within one exposure category. Comparison of Exposure-Response Relationships For the full cohort, only the HPD adjusted measurement-based cumulative noise estimates approached statistical significance (p<0.1), with the log-log model provided a slightly better fit than the log-linear model (Table 6.3, Figure 6.3). However, the categorical analyses for the HPD adjusted measurement-based noise estimates did not show a clear monotonic trend with increasing exposure. The expert-based and unadjusted measurement-based exposure-response relationships were very imprecise and provided minimal improvements in model fit. Stronger exposure-response relationships were observed in the subcohort with minimal HPD use. Both the measurement- and expert-based cumulative noise estimates had a statistically significant exposure-response relationship, with the log-linear model providing a better fit (Table 6.3). The categorical analyses for both exposure assessment approaches demonstrated a fairly consistent monotonic increase in disease risk with increasing exposure (Figure 6.4). The two highest exposure categories for both metrics were statistically significant. The measurement-based cumulative noise estimates provided a better model fit and more precise exposure-response relationship. The relative risks for the highest exposure group was higher in the subgroup than in the full cohort (RR=1.4 vs. 1.2). Comparison of Observed and Expected Attenuation The relative risks associated with the 80th percentile of the cases' exposure distribution (from Table 6.1) were compared. For the full cohort, the highest predicted relative risk was 1.2 for the HPD 88 adjusted measurement-based cumulative noise estimates. The relative risks from the unadjusted measurement-based and expert-based cumulative noise estimates were attenuated by 5.1 and 7.0 % for the log-log model, respectively, compared to the predicted attenuation of 7.8 and 9.6% compared to the HPD adjusted measurement-based approach. For the subcohort with minimal HPD use, the predicted relative risks were similar at 1.4 and 1.3 for the log-log and log-linear models, respectively for the 80th percentile of the cases' exposure distribution. The relative risks for the expert-based noise estimates were attenuated by 3.7% compared to the 15% predicted for the log-linear model, assuming the measurement-based approach was the gold standard. Discussion Occupational noise exposure was associated with acute myocardial infarction mortality with both the measurement- and expert-based noise exposure assessment approaches. In the full cohort, adjusting the measurement-based cumulative noise exposure estimates to account for HPD use provided the best fitting exposure-response relationship, with no exposure-response trends observed for the unadjusted measurement-based or expert-based noise levels. The strength of the relationship in the full cohort was much weaker than observed in a subgroup with minimal HPD use. In the subcohort with minimal HPD use the expert-based noise estimates resulted in very similar exposure-response relationships to the measurement-based noise estimates; however, the measurement-based approach was more precise. The relatively strong relationship between the two exposure assessment approaches suggests that noise is an exposure that experts may able to estimate fairly accurately. This finding is supported by two studies examining the validity of self-reported and expert-based noise assignments. Ising et al (Ising et al, 1997) found that individuals were well able to self-report noise exposure, with a strong correlation (Spearman r=0.84) between the self-reported exposure categories and noise dosimetry. Similarly, a separate study in the BC Sawmill cohort research group found that mill-specific expert-based noise rankings were strongly correlated (Pearson r=0.80) with noise dosimetry (H.W. Davies, unpublished data). The pooled job evaluators' noise exposure rankings that were used here were moderately correlated (r=0.61) with measurements. 89 The expert-based exposure estimates were more strongly correlated with the measurement-based noise exposure estimates prior to accounting for exposure attenuation due to HPD use. The expert-based exposure estimates accounted for job differences in noise exposure, but did not account for mill or time period differences. In contrast, the measurement-based noise estimates were based on statistical models that included explanatory variables to account for mill- and time period-differences in exposure level (Davies, 2002). In calculating the expert-based cumulative noise levels, the expert-based noise ratings were treated as if there were equal differences in noise levels between each unit. Because the expert-based estimates were ranks, there was no clear way to account for the magnitude of noise protection provided by HPD use in the expert assessment. However, studies testing the validity of the experts against exposure measurements could also be used to calibrate the expert ratings against units of noise exposure. The relationship between expert- and measurement-based exposure assessment methods was substantially driven by the exposure duration component. In this analysis only the HPD adjusted measurement-based cumulative noise estimates were associated with heart disease, with no associations with the unadjusted measurement-based cumulative noise levels or with the expert-based cumulative noise levels. This suggests that the exposure intensity adds a very important contribution to the cumulative noise exposure metric used here. Other exposure metrics have been recently suggested for examining noise-induced hearing loss, including measures of variability and "peakiness" (Seixas et al, 2005). These alternative metrics may also be important in studying noise-attributable heart disease which is hypothesized to be stress-mediated (Bjorntorp, 1997), but are unlikely to be easily estimated by experts. The observed attenuation was similar to the expected attenuation in the full cohort, but was less than one-quarter of that expected in the subcohort. While the theoretical equations provide guidance on the potential attenuation, the analyses presented here violate several of the assumptions of the equations. The theoretical equations assume that one measure is the "gold standard"; however, the measurement-based noise estimates used in this study are also subject to exposure misclassification and are imperfect measures of the true noise dose (Stewart et al, 1996). The theoretical equations also assume classical measurement error, which attenuates exposure-response relationships (Armstrong et al, 1994). Classical measurement error results from using an exposure measure that is not perfectly correlated with the proximal cause of the disease, from sampling strategy errors that do not account for all sources of variability in exposure levels, and from measurement device errors. However, the group-based exposure assessment used here includes both classical and Berkson-type 90 measurement error. Berkson error results when group means are assigned to individuals and typically results in an unbiased exposure-response relationship with reduced precision (Seixas and Sheppard, 1996; Steenland et al, 2000; Heid et al, 2004). Berkson type error would occur at two stages of the exposure assessment used here: firstly, when job means were assigned for each job a subject held; and secondly, when the cumulative exposure metrics were categorized to apportion person-years, with subsequent assignment of the group mean for the analysis of the "continuous" exposure-response relationship. Clearly caution should be used in interpreting the attenuation predicted by theoretical equations. The results are specific to this study, but suggest that occupational noise may be one exposure that experts with specific experience within the industry may be able to rank well. Quantitative exposure estimates are useful not only to reduce exposure misclassification, but also to assist in developing occupational exposure limits and for comparisons of exposure-response relationships across studies. To be useful for these purposes, the expert-based estimates would need to be calibrated against units of noise exposure. The exposure protection provided by hearing protective devices appears to make a substantial contribution to noise exposure misclassification in a retrospective exposure assessment of sawmills workers. Future work is underway in this cohort to develop better adjustment factors to account for HPD use by determining the factors that influence self-reported HPD use. While the relative risks observed in this study are not high, the high prevalence of occupational noise exposure across industries combined with a high background rate of heart disease suggests that the burden of heart disease related to occupational noise exposures may be high. 91 Table 6.1 Exposure category cut points based on the cumulative exposure of the cases and the mean cumulative exposure weighted by person-years for cumulative noise exposure Full Cohort Sub Cohort (<1 970 employment) Cumulative Cumulative Cumulative Cumulative Cumulative Noise Exposure Noise Exposure, Expert Noise Noise Exposure Expert Noise (dBA-years) HPD adjusted Rankings (rating (dBA-years) Rankings (rating (dBA-years) weighted-years) weighted-years) Category Cut Points Unexposed 100.0 100.0 14.5 100.0 14.5 20 t h %ile 103.2 103.0 29.9 102.9 24.6 40 , h %ile 106.0 105.6 48.9 105.6 37.7 60 t h %ile 108.8 108.4 65.6 108.4 54.2 80 t h %ile 112.0 111.8 85.7 111.5 75.2 Mean Cumulative Exposure 1 (Unexposed) 95.5 86.3 8.60 95.8 8.60 2 101.7 101.6 22.0 101.5 19.4 3 104.6 104.3 39.7 104.2 31.1 4 107.4 107.0 57.0 106.9 46.2 5 110.4 110.1 75.3 110.0 64.6 6 115.5 114.9 104.4 114.8 96.8 92 Table 6.2 Pearson correlation (r), Spearman correlation (rho) and weighted kappa (i between cumulative expert- and measurement-based noise estimates Continuous Exposure Categorical r rho K w Full Cohort Measurement-based vs. r Expert Noise HPD Adjusted Measurement-based vs. Expert Noise Measurement-based, unadjusted vs. HPD Adjusted Measurement-based noise Subcohort Measurement-based vs. 0.63 0.75 0.51 Expert Noise 0.71 0.51 0.66 0.81 0.57 0.74 0.61 0.49 0.70 93 Table 6.3 Model parameters and precisions (p/SE) for the relationships between cumulative noise exposure and acute myocardial infarction Full Cohort Change in deviance" SE Sub-cohort (employed <1970) Change in P/SE deviance P SE P/SE Cumulative exposure (dBA-years) Log-linear Model * Log-log M o d e l ' 0.84 0.0045 0.0050 0.90 7.27 0.0195 0.0072 2.71 1.21 0.0337 0.0298 1.13 6.07 0.0969 0.0396 2.45 Cumulative exposure, HPD adjusted (dBA-years) Log-linear Model Log-log Model 2.79 3.24 0.0056 0.0033 0.0446 0.0249 1.70 1.79 n/a Cumulative expert noise estimates (rating weighted-years) Log-linear Model Log-log Model 0.77 0.0009 0.0010 0.90 5.56 0.0041 0.0017 2.41 0.74 0.0175 0.0204 0.86 4.40 0.0616 0.0296 2.08 RR=relative risk, p=slope parameter from model, CE=cumulative exposure, SE=standard error, (3 /SE=precision, n/a=not applicable; a Change in model deviance compared to model with age, calendar year and race, but no exposure b Log-linear model: Log RR = |3 * (CE-JC„), .r„=mean cumulative exposure of the reference group c Log-log model: Log RR= |3*ln(CE-.x„+l) 94 Figure 6.1 Comparison of expert-based (log-scale) and measurement-based cumulative noise estimates for subcohort 140 -> 95 Figure 6.2 Agreement in the exposure category assigned each cohort subject between exposure assessment approaches • Same S ± 1 0 ± 2 • ± 3 B ± 4-5 Categories Expert / Expert / M s m t / H P D Expert / Msmt HPD Adj Adj Msmt Msmt Msmt Full Cohort Subcohort Exposure Assessment Approach 96 Figure 6.3 Log-linear (—) and log-log (—) relationships between acute myocardial infarction mortality and cumulative noise estimates for full cohort a: / 1 v _ ,1 , -i i 85 90 95 100 105 110 115 120 C u m u l a t i v e M e a s u r e m e n t - b a s e d N o i s e ( d B A - y e a r ) 85 90 95 -100 105 110 115 120 0 20 40 60 80 100 120 C u m u l a t i v e M e a s u r e m e n t - b a s e d Cumulative Expert -based N o i s e , H P D a d j u s t e d ( d B A - y e a r ) Noise (rating weighted-year) Figure 6.4 Log-linear (—) and log-log (—) relationships between acute myocardial infarction mortality and cumulative noise estimates for subcohort 98 References Armstrong BK, White E, Saracci R (1994) Principles of exposure assessment in epidemiology. Oxford: Oxford University Press. Astrakianakis G, Band PR, Le N , Bert J, Janssen B, Svirchev L M , Tang C, Anderson J, Keefe A (1998) Validation of a mill-specific job-exposure matrix in the British Columbia pulp and paper industry. Appl Occup Environ Hyg; 13 1-7. Bjorntorp P (1997) Stress and cardiovascular disease. Acta Physiol Scand Suppl; 640 144-8. Davies HW (2002) Exposure to occupational noise and risk of cardiovascular disease: A retrospective cohort study. Interdisciplinary Studies. Vancouver, University of British Columbia. PhD. Davies HW, Teschke K, Kennedy SM, Hodgson MR, Hertzman C, Demers PA (2005) Occupational exposure to noise and mortality from acute myocardial infarction. Epidemiol; 16 25-32. de Cock J, Kromhout H, Heederik D, Burema J (1996) Experts' subjective assessment of pesticide exposure in fruit growing. Scand J Work Environ Health; 22 425-32. Dosemeci M , Wacholder S, Lubin JH (1990) Does nondifferential misclassification of exposure always bias a true effect toward the null value? Am J Epidemiol; 132 746-8. Heid IM, Kuchenhoff H, Miles J, Kreienbrock L, Wichmann HE (2004) Two dimensions of measurement error: Classical and Berkson error in residential radon exposure assessment. J Exposure Analysis Environ Epidemiol; 14 365-77. Hertzman C, Teschke K, Dimich-Ward H, Ostry A (1988) Validity and reliability of a method for retrospective evaluation of chlorophenate exposure in the lumber industry. Am J Ind Med; 14 703-13. Hertzman C, Teschke K, Ostry A, Hershler R, Dimich-Ward H, Kelly S, Spinelli JJ, Gallagher RP, McBride M , Marion SA (1997) Mortality and cancer incidence among sawmill workers exposed to chlorophenate wood preservatives. Am J Public Health; 87 71-9. Ising H, Babisch W, Kruppa B, Lindthammer A, Wiens D. (1997) Subjective work noise: A major risk factor in myocardial infarction. Soz Praventivmed; 42 216-22. Kromhout H, Loomis DP, Kleckner RC (1999) Uncertainty in the relation between exposure to magnetic fields and brain cancer due to assessment and assignment of exposure and analytical methods in dose-response modeling. Ann N Y Acad Sciences; 895 141-55. Kromhout H, Oostendorp Y, Heederik D, Boleij JS (1987) Agreement between qualitative exposure estimates and quantitative exposure measurements. Am J Ind Med; 12 551-62. 99 Loomis A, Kromhout H, Kleckner RC, Savitz DA (1998) Effects of the analytical treatment of exposure data on associations of cancer and occupational magnetic field exposure. Am J Ind Med; 34 49-56. Loomis D, Salvan A, Kromhout H, Kriebel D (1999) Selecting indices of occupational exposure for epidemiologic studies. Occup Hyg; 5 73-91. Mwaiselage J, Bratveit M , Moen B, Yost M (2005) Variability in dust exposure in a cement factory in Tanzania. Ann Occup Hyg; 49 511-9. Ostry A (1998). Psychosocial job strain and coronary heart disease in a cohort of blue collar workers. Department of Health Care and Epidemiology. Vancouver, University of British Columbia. PhD. Ostry AS, Marion SA, Demers PA, Hershler R, Kelly S, Teschke K, Mustard C, Hertzman C (2001) Comparison of expert-rater methods for assessing psychosocial job strain. Scand J Work Environ Health; 27 70-5. Post W, Kromhout H, Heederik D, Noy D, Duijzentkunst RS (1991) Semiquantitative estimates of exposure to methylene chloride and styrene: The influence of quantitative exposure data. Appl Occup Environ Hyg; 6 197-204. Richardson DB, Loomis D (2004) The impact of exposure categorisation for grouped analyses of cohort data. Occup Environ Med; 61 930-5. Salvan A, Stayner L, Steenland K, Smith R (1995) Selecting an exposure lag period. Epidemiol; 6 387-90. Seixas N, Neitzel R, Sheppard L, Goldman B (2005) Alternative metrics for noise exposure among construction workers. Ann Occup Hyg; 49 493-502. Seixas NS, Sheppard L (1996) Maximizing accuracy and precision using individual and grouped exposure assessments. Scand J Work Environ Health; 22 94-101. Steenland K, Deddens JA, Zhao S (2000) Biases in estimating the effect of cumulative exposure in log-linear models when estimated exposure levels are assigned. Scand J Work Environ Health; 26 37-43. Stewart PA, Herrick RF (1991) Issues in performing retrospective exposure assessment. Appl Occup Environ Hyg; 6 421-27. Stewart PA, Lees PS, Francis M (1996) Quantification of historical exposures in occupational cohort studies. Scand J Work Environ Health; 22 405-14. Teschke K, Olshan AF, Daniels JL, De Roos AJ, Parks C G , Schulz M , Vaughan T L (2002) Occupational exposure assessment in case-control studies: Opportunities for improvement. Occup Environ Med; 59 575-93. 100 Teschke K, Ostry A, Hertzman C, Demers PA, Barroetavena M C , Davies HW, Dimich-Ward H, Heacock H, Marion SA (1998) Opportunities for a broader understanding of work and health: Multiple uses of an occupational cohort database. Can J Public Health; 89 132-6. Tielemans E, Kupper L L , Kromhout H, Heederik D, Houba R (1998) Individual-based and group-based occupational exposure assessment: Some equations to evaluate different strategies. Ann Occup Hyg; 42 115-9. van Tongeren MJ, Kromhout H, Gardiner K, Calvert I A, Harrington J M (1999) Assessment of the sensitivity of the relation between current exposure to carbon black and lung function parameters when using different grouping schemes. Am J Ind Med; 36 548-56. 101 Chapter 7: General D iscuss ion It is well understood and mentioned in the limitations of most epidemiologic studies that nondifferential exposure misclassification attenuates relative risk estimates. There is substantial evidence from simulation studies demonstrating this phenomenon (Flegal et al., 1986; Dosemeci et al., 1990; Flegal etal., 1991; Birkett, 1992; Dosemeci and Stewart, 1996; Veierod and Laake, 2001; Richardson and Loomis, 2004). Distortion of the exposure-response relationship may also result from exposure misclassification when exposure is categorized (Dosemeci etal., 1990). Few studies have empirically examined the impact of exposure misclassification of quantitative exposure estimates for retrospective cohort studies. Those studies that have examined the impact of exposure assessment decisions have been primarily focused on grouping strategies, category cut points, and historical weighting of exposures (Kromhout and Heederik, 1995; Seixas and Sheppard, 1996; Kromhout et al., 1997; van Tongeren et al., 1999; Heederik and Attfield, 2000; Werner and Attfield, 2000; Checkoway and Rice, 1992; Loomis et al., 1998; Kromhout et al., 1999; Burstyn et ai, 2003; Richardson and Loomis, 2004; Loomis et ai, 2005). These studies primarily compared different methods of using the same measurement data, but did not compare exposure metrics that were developed using different approaches to refine exposure estimates. This dissertation empirically evaluated the impact of exposure assessment strategies on the sensitivity of exposure-response relationships for two areas of quantitative exposure assessment that have received little attention: 1) exposure indicators when the causal agent(s) are a component of a mixture, and 2) expert judgment versus measurement-based exposure assessment. The research presented in this dissertation was unique in its ability to compare multiple exposure metrics developed to represent the same causal agents, as it is rare to have multiple measures of the same exposure within a single study. The comparisons between multiple exposure metrics were possible because of the availability of two large cohorts with a considerable number of cases. The substantial power provided by these cohorts allowed for simple parametric models of the exposure-response relationships to be evaluated. Using such models rather than solely categorical analyses facilitated these comparisons by providing a way to quantitatively measure the exposure metric's performance in the exposure-response relationship, namely, the improvements in the model fit and the precision of the exposure-response relationship. This dissertation involved quantitative exposure estimates of four broad classes of occupational exposure: polycyclic aromatic hydrocarbons (PAH, Chapters 2-4), dust/wood dust (Chapter 5), 102 chlorophenols (Chapter 5), and noise (Chapter 6). Substantial effort was involved in refining the exposure estimates to obtain exposure estimates hypothesized to be more proximal measures of the causal agents (Chapter2, Demers etal., 2006; Friesen etal., 2005; Davies, 2002; Ostry, 1998). This dissertation demonstrated that this additional effort was a necessary and worthwhile step in the development of the retrospective exposure estimates for the epidemiologic studies. Substantial attenuation of relative risks occurred when using exposure metrics developed using quantitative approaches that would generally be considered the best practice in exposure assessment. Yet with these "best" measures, and in particular non-specific dust, total chlorophenols, and noise estimates without adjustment for use of hearing protective devices, the exposure-response relationships were weak and inconclusive. In this chapter, the main findings of this research are recapped under within the two principle themes. In addition, specific issues in common throughout the individual chapters of the dissertation are discussed, including expected versus observed attenuation, issues in comparing exposure metrics, the shape of exposure-response relationships, and accounting for protective equipment use; and future directions for research are presented. Theme 1: Exposure Indicators for Mixtures Exposure to mixtures for retrospective studies is typically evaluated using either an individual component of the mixture or a measure of the mixture. Both options assume that the chosen indicator has a constant relationship with toxic potential of the mixture, whether contributed by an individual or multiple components. If the relationship is not constant across all time periods, jobs, and sites of a retrospective study, exposure misclassification may result. This was the case for PAHs, chlorophenols, and wood dust in this dissertation. More proximal measures of the causal agents for these exposures had been developed as part of this dissertation (BaP/BSM, Chapter 2) or previously (all others). These refined estimates were developed to minimize the exposure misclassification that results from sources of variability in the relationship between the more commonly avalable, but less specific exposure measure and the causal agents. The impact of moving to more proximal measures compared to the less specific exposure measures was evaluated (Chapters 4-5). For all three exposures examined, the more proximal measures consistently resulted in improved exposure-response relationships in the epidemiologic analysis (Chapters 4-5). A summary of the main findings is provided (Table 7.1). The use of the more specific exposure measure improved the precision of the exposure-response relationship by between 1-14% for BaP over BSM, by 10-30% for 103 PCP over total chlorophenols, and by 218% for wood dust over nonspecific dust. The use of the less specific exposure measure resulted in attenuation of the relative risk by 2-13% for BSM, by 3-7% by total chlorophenols, and 34% by nonspecific dust. Not surprisingly, as the correlation between the measures became weaker the amount of observed attenuation with the use of the less specific exposure measure became larger. The strong correlation between BaP and BSM in the BC Aluminum Smelter Cohort resulted in similar conclusions regarding exposure-disease associations regardless of the exposure metric. In contrast, the much weaker correlation between dust and wood dust resulted in no observable association for COPD hospitalizations and dust, but a clear and statistically significant association was observed with wood dust. With a correlation between that of dust/wood dust and BaP/BSM, total chlorophenols and PCP demonstrated similar shapes of the exposure-response relationship with both N H L and kidney cancer in categorical analyses, but only PCP approached statistical significance in simple parametric models. These substantial improvements in the exposure-response relationships were observed despite several limitations: 1) the inability to account for all sources of variability in the relationships between the less and more specific measures; 2) the increased uncertainty introduced in the more specific measures through the use of group-based weights; and 3) despite being developed with fewer exposure measurements. Work area differences in the wood dust component of total dust were observed and were accounted for in the wood dust exposure estimates through the use of group-based weights, but time period or mill differences have not yet been studied and thus could not be taken into account (Demers et ai, 2000). BaP was strongly correlated with other components of the PAH mixture, but the process-related differences in the relationship between BaP and other components of the PAH mixture observed in Chapter 3 could not be taken into account because BaP's relationship with other components is not known for all process conditions in the study. Also, the BaP exposure estimates were based on models that used approximately half the exposure measurements that were available for BSM and explained only half the variability in the data compared to the corresponding BSM models (Chapter 2). Studies that examine the sources of variability between two or more measures, such as the comparison in Chapter 3, are a useful method for developing more refined or alternative exposure metrics. For example, principle components analyses could be used to identify important co-exposures in the mixture to combine into one metric (Burstyn et al, 2000; Burstyn, 2004; Vermeulen et al., 2004). Alternatively, the relationship between a commonly measured exposure and another component of the mixture could be used to retroactively assess exposure to the other component that may be more strongly associated with the disease of interest. 104 Theme 2: Expert Judgment vs. Measurement-based Exposure Assessment Occupational noise exposure was strongly associated with acute myocardial infarction mortality with both the expert-based and measurement-based noise estimates in sawmill workers with minimal HPD use (Chapter 6). The strong performance of expert noise ratings suggests that noise may be an exposure that experts may be able to estimate fairly accurately. Previous studies have found that experts have been better able to accurately assess exposure levels for exposures that are easy to perceive (Teschke et al, 2002). Noise exposure is ubiquitous, familiar, and has a strong impact on one's senses: individuals have been well able to differentiate between noise levels (Ising et al, 1997). However, the model-based noise estimates resulted in a 12% improvement in the precision of the exposure-response relationship in this subcohort over the use of the expert ratings. In the full BC Sawmill Cohort, however, only the HPD-adjusted model-based noise estimates were associated with AMI mortality (Chapter 6). This was not surprising as HPD use had become widespread in the 1970s. Both the unadjusted model-based and expert-based noise estimates would result in exposure misclassification as neither accounted for the exposure attenuation provided by HPD use. Adjusting the experts' noise ratings to account for HPD use was not possible in this study as the experts' ratings were on a unitless scale. Further refinement of the expert-based noise estimates, including accounting for HPD use, would be possible by calibrating their exposure rankings with noise dosimetry. The expert ratings would also need to be translated into noise units (i.e. dBA) prior to being used for risk assessment or policy initiatives such as disease compensation and exposure limits. Expected vs. Observed Attenuation The use of less specific exposure metrics resulted in attenuation of the relative risks in the range of 2-34% for the four exposure classes and multiple health outcomes examined in this dissertation (Table 7.1). There was no consistent relationship between observed and expected attenuation: the observed attenuation was found to be greater than expected based on theoretical models for some exposures/health outcomes and less than expected for others. Various reasons may account for this inconsistency. One explanation, discussed in Chapter 6, may be due to the Berkson error structure of grouped data, which results in more accurate, but less precise exposure-response relationships (Steenland et al, 2000). An alternative explanation, discussed in Chapter 5, may be that the more specific exposure measure used for comparison purposes is not a true gold standard. The correlation between the less specific measure and the true dose may not be accurately reflected by the observed correlation between the less specific and the hypothesized more proximal measure that was used to 105 calculate the expected attenuation. The analyses used in this dissertation have violated some of the assumptions of the theoretical equations predicting the expected attenuation, discussed in Chapter 5, which were not developed specifically for Poisson regression using grouped exposure. Clearly, these theoretical equations provided useful guidance in estimating the potential attenuation of exposure-response relationships and therefore could be useful in a priori decisions of which exposure metrics to examine in epidemiologic analyses. However, these equations should not be used for a posteriori adjusting the exposui'e-response relationships. Issues in Comparing Exposure Metrics The primary measures used to evaluate improvements in exposure-response relationship in this dissertation were improvements in model fit and the precision of the exposure-response relationship using simple parametric models (log-log model, log-linear model). These measures provided a quantifiable evaluation tool to compare different exposure metrics for the same exposure-disease association. For each exposure-disease association (Chapters 4-6), both the model fit and model precision consistently identified the same exposure metric as the "better measure". The change in model fit with and without the variable of interest has been used for comparison purposes in several studies (Armstrong et al, 1986; Seixase/a/., 1993; Sal van et al, 1995; Youk et al, 2001), and has been suggested by Loomis et al. (1999) as the preferred method. However, the magnitude of the improvement in the model fit is not easily compared across studies or health outcomes as the baseline model fit is dependent on the other variables included in the model, such as age and calendar year. In contrast, the model precision takes into account both the magnitude of the exposure-response slope and its standard error and thus provides a unit-less measure that can be compared across studies and health outcomes. It has been used as performance indicator in only a few occupational studies thus far (Kromhout et al, 1997; van Tongeren et al, 1999). An added challenge occurs when the exposure metrics being compared have different units or scales. This was the case for measures of PAH (Chapter 4) and occupational noise (Chapter 6). With different units or scales the magnitude of the slope cannot be compared, but the magnitude of relative risk is comparable for a defined exposure difference, such as interquartile range or a specified percentile. The performance measures used in this dissertation were relatively insensitive to differences in units and scale. Some variability in the precision of the exposure-response relationship was attributable to the logarithmic aspect of the log-linear and log-log exposure-response models. However, decisions regarding the best fitting exposure metric based on model precision were robust to ten-fold differences in the scale for comparisons of measures of PAH exposure (Chapter 4). 106 Simple parametric models require strong assumptions on the shape of the exposure-response relationship. Smoothing and spline functions could be used and they do not require as strong assumptions on the shape of the exposure-response relationship; however, comparisons between exposure-response relationships using different exposure metrics are complicated by the absence of interpretable parameters. To confirm that the simple parametric models are reasonable approximations, the results from categorical analyses were presented alongside the simple parametric models to confirm the shape of the exposure-response relationship. Evaluating the performance of exposure metrics from categorical analyses is typically limited to comparisons of the p-value from trend tests (using either ordinal or mean value for the categories), the magnitude of relative risk in the highest exposure categories, and qualitative comparisons of whether exposure trend across categories increases in a monotonic manner. However, these comparisons are sensitive to the exposure category cut points and assume that the highest relative risk is the least biased even though there is evidence that nondifferential misclassification may create bias away from the null (Dosemeci et al, 1990; Salvan et al, 1995; Loomis et al, 1999; Richardson and Loomis, 2004). Shape of the Exposure-Response Relationship The log-log exposure-response model provided the best fitting and most precise exposure-response relationships for most exposure metrics and health outcomes examined in this dissertation (Table 7.1). The attenuation of risk with increasing exposure described by the log-log models may be a function of exposure misclassification, but may also result from the healthy worker effect, saturation of key disease pathways, confounding by other risk factors, depletion of the number of susceptible people at high exposure levels, and a natural limit for diseases with high background rates and is consistent with earlier findings among retrospective cohort studies (Stayner et al, 2003). These other explanations for the shape of the exposure-response relationship are expected to remain constant within a study/health outcome and as a result be relatively robust to the exposure metrics used. Thus, the differences in the model fit and precision observed in this dissertation were expected to be related to exposure misclassification in the exposure metrics evaluated. The log-linear model provided the better fitting and more precise exposure-response relationship for only two exposure-disease associations: BaP and bladder cancer incidence in the aluminum smelter cohort (Chapter 4); and occupational noise and AMI mortality in the sawmill subcohort with minimal HPD use (Chapter 6). The linear relationships provides indirect support for the accuracy of these exposure metrics as less distortion of the exposure-response relationship was observed 107 compared to the other exposure metrics. The log-log relationships for the other exposure-disease associations may result from the many factors listed above, but also suggests that exposure misclassification in these more refined, and more proximal, exposure metrics still remains, resulting in a distortion of the exposure-response relationship in the higher exposure categories. Accounting for Personal Protective Equipment Examining the impact of exposure measures that account for the exposure attenuation provided by personal protective equipment, such as respirators and hearing protection, was not a primary theme of this dissertation, however, this issue was relevant for both noise exposure in sawmills and PAH exposure in aluminum smelters. The noise estimates adjusted for the use of hearing protective devices resulted in more precise, stronger exposure-response relationships with acute myocardial infarction in the full sawmill cohort than the unadjusted exposure estimates which resulted in weak, inconclusive exposure-disease associations (Chapter 6). Failing to account for respirator use in PAH exposure estimates overestimated exposure in the low categories by 30% and the high categories by 12%; however, there was minimal impact on exposure-response relationships due to the long lag time (20 years) of the cancer diseases of interest (Friesen et al, 2004). As seen here, the use of protective devices may contribute substantially to exposure misclassification if their use is not taken into account. The magnitude of the exposure misclassification due to protective equipment is dependent on the date of widespread implementation, the lag time of the disease, the prevalence and frequency of use, and the true attenuation provided by their use. Adjusting for protective equipment use, such as hearing protection or respirators, has rarely been applied in retrospective occupational studies (Smith et ai, 1980; Swaen et al, 1998; Rando et al, 2001; Davies, 2002; Dosemeci et al, 2002), but is an important aspect to consider as the expanding use of protective devices. Recommendations for Future Work There are numerous aspects of the exposure assessment process that can introduce or eliminate exposure misclassification that have not been addressed in this dissertation. The evaluation framework used here is simple and can easily be extended to empirically examining other exposure assessment decisions. Three specific areas of exposure assessment decisions that arose from aspects of this research, but were not explored, would benefit from a consistent, quantitative evaluation of 108 their impact in epidemiologic studies. These three areas were the following: 1) evaluating differences between modeling approaches used to predict historical exposure; 2) evaluating the impact of assumptions regarding historical exposure levels for periods prior to the availability of exposure measurements; and 3) evaluating the impact of heterogeneity of the variability in exposures across jobs, work areas, and time periods. While used here for Poisson regression, the framework used here for comparing exposure metrics is directly extendable to other statistical models such as Cox proportional hazards regression and logistic regression where a slope and'standard error are reported. This dissertation has only begun to explore the differences between the observed attenuation and the attenuation predicted by theoretical equations. This inconsistency was observed in Poisson Regression analyses, but the relationship between observed and predicted attenuation may be more consistent in logistic regression and Cox proportional hazards models and should be examined in these other models. The relationship between observed and predicted attenuation may be better explained by theoretical equations that account for the variance components, such as within- and between-worker or group variance, for the exposure grouping used in Poisson regression. This area of research would benefit from both empirical and simulation studies. A theoretical framework for predicting attenuation that more accurately reflects the expected attenuation is useful for making decisions early in the exposure assessment process as developing full job exposure matrices for all potential exposure indicators requires substantial work. Conclusions and Significance This dissertation demonstrated empirically that substantial attenuation occurs with exposure measures based on the more commonly available exposure measurements. Only the more refined exposure measures, hypothesized to be more proximal measures of the causal agents, resulted in statistically significant exposure-response relationships. This dissertation also demonstrated that while expert-based quantitative exposure estimates can result in strong exposure-disease associations, associations were stronger when a measurement-based approach was used. Substantial improvements in the exposure-response relationship were made with the more proximal exposures. However, the frequent observation of attenuation of risk with increasing exposures, best described by the log-log model, suggests that some exposure misclassification in the more proximal measures remains. The comparisons between exposure-response relationships using different exposure metrics included in this dissertation assumed that the better metric would result in stronger, more precise exposure-109 response relationships. Some of these improvements may be due to chance; however, the findings were consistent across multiple exposure-disease associations in two large cohorts. This dissertation provides a framework based on quantifiable measures to evaluate exposure metrics used in retrospective cohort studies. A consistent, quantifiable approach is necessary because the theoretically predicted attenuation correlated poorly with the observed attenuation throughout this research. The key aspects of this evaluation framework included assigning exposure cut points based on the exposure distribution of the cases, using simple parametric models to describe the exposure-response relationship, and using the Wald Statistic (slope/standard error) to evaluate the precision of the exposure-response relationship. This framework was used across all exposure-disease associations in this dissertation and can be used in any retrospective cohort study that includes quantitative assessment of exposure. For this framework to be used, however, quantitative estimates of each exposure metric of interest for all jobs and time periods in the study are needed. The theoretically predicted attenuation based on the correlation between exposure metrics obtained from validation studies may serve to reduce the potential number of exposure metrics to be considered; however, due to the lack of agreement observed here these equations should not be used to a posteriori adjust the exposure-response relationships. Continuous exposure-response relationships have not been previously reported for wood dust and COPD hospitalizations, for occupational noise and AMI mortality, or for chlorophenols and N H L and kidney cancer incidence. For these exposure-disease associations, the simple parametric models evaluated here will facilitate future risk assessment and will assist regulators in setting exposure limits. The use of pilot studies, validation studies, and other supplementary information to refine exposure estimates is increasing, with the assumption that these refined estimates will reduce exposure misclassification. This assumption has rarely been tested and has not been rigorously or quantifiably assessed for the two themes evaluated in this dissertation. The findings of this research support the a priori hypothesis that the more refined exposure estimates have less exposure misclassification. Still, every effort should be made to include validation steps in the development of retrospective exposure estimates to assess key exposure assessment decisions. Improving our knowledge of exposure-response relationships will assist occupational hygienists and epidemiologists in our most important objective, improving and protecting workers' health. 110 Table 7.1 Summary of main findings on impact of exposure metric choice on exposure-response relationships Exposure Metr ic Best Fit t ing Metr ic and Mode l by Health Outcome Theme 1: Exposure Indicators for Mixtures Benzene soluble materials Bladder Cancer: BaP , log-linear ( B S M ) vs. benzo(a)pyrene (BaP) L u n g Cancer: BaP , log-log (Ch . 4) Dust vs. W o o d Dust (Ch. 5) Total Chlorophenols vs. Pentachlorophenol ( P C P ) (Ch . 5) Total Chlorophenols vs. Tetrachlorophenol (TCP) (Ch. 5) A M I Mortal i ty: BaP , log-log C O P D Hospitalizations: W o o d Dust, log-log Non-Hodgk in lymphoma: P C P , log-log Kidney cancer: P C P , log-log Non-HodglJ in lymphoma: Total Chlorophenols, log-log Kidney cancer: Total Chlorophenols, log-log Theme 2: Expert Judgment vs. Measurement-based Exposure Assessment Measurement-based Full Cohort (unadjusted & H P D adjusted) A M I Mortal i ty: H P D adjusted measurement-based estimates, log-log vs. Expert No i se (Ch. 6) Subcohort A M I Mortal i ty: Measurement-based estimate, log-linear Correlation Improvement in precision (%)' 0.94 0.68 0.78 Unadj vs. H P D adj: 0.66 Expert vs. Unadj: 0.71 Expert vs. H P D adj: 0.51 Expert vs. Unadj: 0.63 Observed vs. Expected Attenuation (%) 2 14 5 10 30 12.6 vs. 7.4 5.9 vs. 5.8 2.4 vs. 3.1 34 vs. 18 7 vs. 15 3 vs. 14 No relationship with T C P 58 31 108 12 Expert: 7.0 vs. 7.8 Unadj: 5.1 vs. 9.6 —— : — j. . „ m u n a n , l u f t - i n i c c u e . u ai u.bJ  3 7 vs 15 o f T p e V e m e n t m P r e C ' S i 0 n [ B / S E ] ^ ° f b 6 S t f ' t t i n g m e t r i C ( S U b S C r l p t 1 } ^ l £ S S S p e C i f ' C m e t r l c ( 5 u b s C r ' P t 2>: P / S E V P /SE] 2 ; B=S,lope estimate; SE=standard error 2 compared to best fitting metric (and model) from 2 n d column C O P D = chronic obstructive pulmonary disease; H P D = hearing protection device; A M I = acute myocardial infarction References Armstrong BG, Tremblay CG, Cyr D, Theriault GP (1986) Estimating the relationship between exposure to tar volatiles and the incidence of bladder cancer in aluminum smelter workers. Scand J Work Environ Health; 12 486-93. BirkettNJ (1992) Effect of nondifferential misclassification on estimates of odds ratios with multiple levels of exposure. Am J Epidemiol; 136 356-62. Burstyn I (2004) Principal component analysis is a powerful instrument in occupational hygiene inquiries. Ann Occup Hyg; 48 655-61. Burstyn I, Boffetta P, Kauppinen T, Heikkila P, Svane O, Partanen T, Stucker I, Frentzel-Beyme R, Ahrens W, Merzenich H, Heederik D, Hooiveld M , Brunekreef B, Langard S, Randem BG, Jarvholm B, Bergdahl IA, Shaham J, Ferro G, Kromhout H (2003) Performance of different exposure assessment approaches in a study of bitumen fume exposure and lung cancer mortality. Am J Ind Med; 43 40-8. Burstyn I, Kromhout H, Kauppinen T, Heikkila P, Boffetta P (2000) Statistical modelling of the determinants of historical exposure to bitumen and polycyclic aromatic hydrocarbons among paving workers. Ann Occup Hyg; 44 43-56. Checkoway H, Rice C H (1992) Time-weighted averages, peaks, and other indices of exposure in occupational epidemiology. Am J Ind Med; 21 25-33. Davies HW (2002) Exposure to occupational noise and risk of cardiovascular disease: A retrospective cohort study. Interdisciplinary Studies. Vancouver, University of British Columbia. PhD. Demers PA, Davies HW, Friesen M C , Hertzman C, Ostry A, Hershler R, Teschke K (2006) Cancer and occupational exposure to pentachlorophenol and tetrachlorophenol. Cancer Causes Control 15(9):939-45; In Press. Demers PA, Teschke K, Davies HW, Kennedy SM, Leung V (2000) Exposure to dust, resin acids, and monoterpenes in softwood lumber mills. Am Ind Hyg Assoc J; 61 521-8. Dosemeci M , Alavanja M C , Rowland AS, Mage D, Zahm SH, Rothman N, Lubin JH, Hoppin JA, Sandler DP, Blair A (2002) A quantitative approach for estimating exposure to pesticides in the agricultural health study. Ann Occup Hyg; 46 245-60. Dosemeci M , Stewart P (1996) Recommendations for reducing the effects of exposure misclassification on relative risk estimates. Occup Hyg; 3 169-76. Dosemeci M , Wacholder S, Lubin JH (1990) Does nondifferential misclassification of exposure • always bias a true effect toward the null value? Am J Epidemiol; 132 746-8. Flegal K M , Brownie C, Haas JD (1986) The effects of exposure misclassification on estimates of 112 relative risk. Am J Epidemiol; 123 736-51. Flegal K M , Keyl PM, Nieto FJ. (1991) Differential misclassification arising from nondifferential errors in exposure measurement. Am J Epidemiol; 134 1233-44. Friesen M C , Davies HW, Teschke K, Marion S, Demers PA (2005) Predicting historical dust and wood dust exposure in sawmills: Model development and validation. J Occup Environ Hyg; 2 650-8. Friesen M C , Demers PA, Spinelli JJ, Le ND (2004) Impact of respirator use on retrospec'tive exposure estimates of potroom fumes at an aluminum smelter. Exposure Assessment. Utrecht, The Netherlands. Heederik D, Attfield M (2000) Characterization of dust exposure for the study of chronic occupational lung disease: A comparison of different exposure assessment strategies. Am J Epidemiol; 151 982-90. Ising H, Babisch W, Kruppa B, Lindthammer A, Wiens D (1997) Subjective work noise: A major risk factor in myocardial infarction. Soz Praventivmed; 42 216-22. Kromhout H, Heederik D (1995) Occupational epidemiology in the rubber industry: Implications of exposure variability. Am J Ind Med; 27 171-85. Kromhout H, Loomis DP, Kleckner RC (1999) Uncertainty in the relation between exposure to magnetic fields and brain cancer due to assessment and assignment of exposure and analytical methods in dose-response modeling. Ann N Y Acad Sciences; 895 141-55. Kromhout H, Loomis DP, Kleckner RC, Savitz DA (1997) Sensitivity of the relation between cumulative magnetic field exposure and brain cancer mortality to choice of monitoring data grouping scheme. Epidemiology; 8 442-5. Loomis A, Kromhout H, Kleckner RC, Savitz DA (1998) Effects of the analytical treatment of exposure data on associations of cancer and occupational magnetic field exposure. Am J Ind Med; 34 49-56. Loomis D, Richardson DB, Elliott L (2005) Poisson regression analysis of ungrouped data. Occup Environ Med; 62 325-9. Loomis D, Salvan A, Kromhout H, Kriebel D (1999) Selecting indices of occupational exposure for epidemiologic studies. Occup Hyg; 5 73-91. Ostry A (1998) Psychosocial job strain and coronary heart disease in a cohort of blue collar workers. Department of Health Care and Epidemiology. Vancouver, University of British Columbia. PhD. Rando RJ, Shi R, Hughes JM, Weill H, McDonald AD, McDonald JC (2001) Cohort mortality study of North American industrial sand workers. Iii. Estimation of past and present exposures to 113 respirable crystalline silica. Ann Occup Hyg; 45 209-16. Richardson DB, Loomis D (2004) The impact of exposure categorisation for grouped analyses of cohort data. Occup Environ Med; 61 930-5. Salvan A, Stayner L, Steenland K, Smith R (1995) Selecting an exposure lag period. Epidemiol; 6 387-90. Seixas NS, Robins T G , Becker M (1993) A novel approach to the characterization of cumulative exposure for the study of chronic occupational disease. Am J Epidemiol; 137 463-71. Seixas NS, Sheppard L (1996) Maximizing accuracy and precision using individual and grouped exposure assessments. Scand J Work Environ Health; 22 94-101. Smith TJ, Ferrell WC, Varner MO, Putnam RD (1980) Inhalation exposure of cadmium workers: Effects of respirator usage. Am Ind Hyg Assoc J; 41 624-9. Stayner L, Steenland K, Dosemeci M , Hertz-Picciotto I (2003) Attenuation of exposure-response curves in occupational cohort studies at high exposure levels. Scand J Work Environ Health 26(6):470-5;29 317-24. Steenland K, Deddens JA. (2004) A practical guide to dose-response analyses and risk assessment in occupational epidemiology. Epidemiol; 15 63-70. Steenland K, Deddens JA, Zhao S (2000) Biases in estimating the effect of cumulative exposure in log-linear models when estimated exposure levels are assigned. Scand J Work Environ Health; 26 37-43. Swaen G M , Bloemen LJ, Twisk J, Scheffers T, Slangen JJ, Collins JJ, ten Berge WF, Sturmans F (1998) Mortality update of workers exposed to acrylonitrile in The Netherlands. Scand J Work Environ Health; 24 10-6. Teschke K, Olshan AF, Daniels JL, De Roos AJ, Parks C G , Schulz M , Vaughan T L (2002) Occupational exposure assessment in case-control studies: Opportunities for improvement. Occup Environ Med; 59 575-93. van Tongeren MJ, Kromhout H, Gardiner K, Calvert I A, Harrington JM (1999) Assessment of the sensitivity of the relation between current exposure to carbon black and lung function parameters when using different grouping schemes. Am J Ind Med; 36 548-56. Veierod MB, Laake P (2001) Exposure misclassification: Bias in category specific poisson regression coefficients. Stat Med; 20 771-84. Vermeulen R, Li G, Lan Q, Dosemeci M , Rappaport SM, Bohong X, Smith MT, Zhang L, Hayes RB, Linet M , Mu R, Wang L, Xu J, Yin S, Rothman N (2004) Detailed exposure assessment for a molecular epidemiology study of benzene in two shoe factories in China. Ann Occup Hyg; 48 105-16. 114 Werner M A , Attfield M D (2000) Effect of different grouping strategies in developing estimates of personal exposures: Specificity versus precision. Appl Occup Environ Hyg; 15 21-5. Youk A O , Marsh G M , Stone RA, Buchanich JM, Smith TJ (2001) Historical cohort study of us man-made vitreous fiber production workers: III. Analysis of exposure-weighted measures of respirable fibers and formaldehyde in the nested case-control study of respiratory system cancer. J Occup Environ Med; 43 767-78. 115 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0092944/manifest

Comment

Related Items