Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Determinants of dust exposure in sawmills : a comparison of fixed-effects and mixed-effects predictive… Friesen, Melissa Charmaine 2001

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

Item Metadata

Download

Media
831-ubc_2001-0379.pdf [ 3.79MB ]
Metadata
JSON: 831-1.0090072.json
JSON-LD: 831-1.0090072-ld.json
RDF/XML (Pretty): 831-1.0090072-rdf.xml
RDF/JSON: 831-1.0090072-rdf.json
Turtle: 831-1.0090072-turtle.txt
N-Triples: 831-1.0090072-rdf-ntriples.txt
Original Record: 831-1.0090072-source.json
Full Text
831-1.0090072-fulltext.txt
Citation
831-1.0090072.ris

Full Text

D E T E R M I N A N T S O F D U S T E X P O S U R E IN S A W M I L L S : A C O M P A R I S O N O F F I X E D - E F F E C T S A N D M I X E D - E F F E C T S P R E D I C T I V E S T A T I S T I C A L M O D E L S by M E L I S S A C H A R M A I N E F R I E S E N B . S c , The University of British Columbia 9 1997 A THESIS S U B M I T T E D IN P A R T I A L F U L F I L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F M A S T E R O F S C I E N C E in T H E F A C U L T Y O F G R A D U A T E STUDIES School of Occupational and Environmental Hygiene We accept this thesis as conforming to the required standard T H E U N I V E R S I T Y O F BRITISH C O L U M B I A August 2001 © Melissa Charmaine Friesen, 2001 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University _ of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. -Department of QcCu^aMonili L-£nyiAOrXon^n4^ I^^U^JZJ The University of British Columbia Vancouver, Canada Date DE-6 (2/88) Abstract Personal dust measurements (n=1516) collected over the period 1981-1997 for both research and compliance purposes were used to construct two statistical models to predict historical dust exposures for a cohort of 14 B . C . sawmills and 28,000 workers. Two multiple linear regression models were built: (1) a fixed-effects model, with potential exposure determinants designated as fixed effects; and (2) a mixed-effects model, with job title designated as a random effect and all other variables as fixed effects. The two predictive models were validated against personal dust exposures (n=213) from a large interior mill that was not part of the sawmill cohort. The two models explained 36% of the dataset's variability. The predicted values were strongly correlated with observed values for both models (fixed-effects model: Pearson r=0.617; mixed-effects model: Pearson r = 0.619). The fixed-effects model predicted 56 of 58 jobs within ± 0 . 2 mg/m 3 of the job geometric means. The mixed-effects model predicted 54 jobs within ± 0 . 2 mg/m . Multiple linear regressions revealed that the most important determinants of wood dust exposure were process group, coastal mill location(-), number of employees(+), and annual production levels per sawmill size(-). The two models underestimated the validation mill's geometric mean exposure level by 0.5 mg/m . On average, outdoor jobs were underestimated by 0.8 mg/m and indoor jobs by 0.3 mg/m . Precisions within the datasets were poor, with GSD's of bias of 2.36 for both models in the modeling dataset and 2.50 and 2.45 for the fixed-effects and mixed-effects models, respectively, in the validation dataset. The predicted values from the two models were nearly perfectly correlated for both the model building dataset (Pearson r=0.975) and the validation dataset (Pearson r=0.985). The mixed-effects model provided no improvement in predictive ability over the fixed-effects model. Several jobs in the validation mill were predicted within the range of normal day-to-day variability, but a few jobs were significantly underestimated, suggesting that the models are only generalizable to mills of similar size, level of technology, and building/yard conditions. ii Table of Contents Abstract ii Table of Contents iii List of Tables v List of Figures vi 1. Introduction 1 2. Background on Cohort Mills and Sawmill Processes 2 3. Literature Review 5 3.1 Quantitative Retrospective Exposure Assessment 5 3.2 Statistical Modeling for Retrospective Exposure Assessment 6 3.3 Modeling Considerations - Grouping Strategies 8 3.4 Modeling Considerations - Choice of Exposure Metric 8 3.5 Modeling Considerations - Model Validation 9 3.6 Reported Wood Dust Exposure Determinants 10 4. Methods .- 12 4.1 Exposure Data — 12 4.2 Potential Exposure Determinants 15 4.3 Data Pre-treatment 19 4.4 Statistical Analysis 20 4.4.1 Preliminary Analyses 20 4.4.2 Model Building: Fixed-Effects Model 21 4.4.3 Model Building: Mixed-Effects Model 21 4.4.4 Model Validation 22 5. Results 24 5.1 Description of the Datasets : 24 5.1.1 Model Building Dataset 24 5.1.2 Validation Dataset 29 5.2 Predictive Models 31 5.2.1 Preliminary Analyses 31 5.2.2 Fixed-Effects Model 34 5.2.3 Mixed-Effects Model 41 5.3 Comparison of Fixed-Effects and Mixed-Effects Models 45 5.4 External Validation of Models 49 iii 6. Discussion 53 6.1 Interpretation of the Model Parameters in the Fixed-Effects and Mixed-Effects Models 53 6.2 Predictive Ability of the Fixed-Effects and Mixed-Effects Models Compared to Model Building Dataset 55 6.3 Predictive Ability of the Fixed-Effects and Mixed-Effects Models Compared to the Validation Dataset 57 6.4 Limitations of the Predictive Models 59 6.5 Conclusions 62 References 64 Appendix A - Form of Fixed-Effects Model Appendix B - Form of Mixed-Effects Model and SAS P R O C M I X E D Procedure iv List of Tables Table 1 Comparison of mill location and 1995 mill parameters, including size, 4 production levels and species, by mill for the 14 cohort sawmills and the validation sawmill. Table 2 Exposure Determinants Reported by Four Previously Published 11 Determinants of Exposure Studies on the Wood Industry. Table 3 Comparison of Number of Measurements, Sampling Method Used, 14 Source, Purpose and Sampling Strategy by Dataset. Table 4 Number, Arithmetic and Geometric Means, Concentration Range, and 26 Source of Dust Measurements by Mi l l . Table 5 Mean Exposure Levels (mg/m ) in Model Building Dataset by Process 27 Group from Highest to Lowest. Table 6 Mean Exposure Levels (mg/m ) in Model Building Dataset by Job 28 (n>10) from Highest to Lowest. Table 7 Mean Exposure Levels (mg/m3) in Validation Dataset by Job (n>5) 30 from Highest to Lowest. Table 8 Descriptive data, coefficients, and standard errors for independent 37-38 variables in the fixed-effects and mixed-effects multiple regression models of wood dust (mg/m3) concentrations (log-transformed, base e). Table 9 Coefficients and standard errors for job entered as a random effect in 43-44 the mixed-effects multiple regression model of wood dust (mg/m3) concentrations (log-transformed, base e). Table 10 Mean, geometric standard deviation, range, model residual and 46 correlation of the fixed-effects and mixed-effects models compared to the model building dataset. Table 11 Predicted dataset geometric mean, bias, precision and correlation of the 50 fixed-effects model and the mixed-effects model compared to the validation dataset. Table 12 Bias and precision of the fixed-effects model and the mixed-effects 51 models compared against the validation dataset by job (n > 5). v List of Figures Figure 1 Map of British Columbia indicating the locations of the cohort mills 3 and the validation mill. Figure 2 Distribution of Dust Measurements in the Model Building Dataset by 25 Year and Data Source. Figure 3 Distribution of Dust Measurements in Model Building Dataset by 27 Concentration (mg/m3). Figure 4 Distribution of Dust Measurements in Validation Dataset by 30 Concentration (mg/m3). Figure 5 Cook's distances compared to the predicted wood dust concentrations 36 (mg/m 3, on log scale) from the fixed-effects model. Figure 6 Residuals compared to the predicted wood dust concentration (mg/m , 39 log-transformed, base e) by data source from the fixed-effects model. Figure 7 Residuals compared to the predicted wood dust concentration (mg/m 3, 40 log-transformed, base e) by categorized observed concentration (log-transformed, base e) from the fixed-effects model. Figure 8 Residuals compared to the predicted values of wood dust 42 concentration (mg/m 3, log-transformed, base e) from the mixed-effects model. Figure 9 Predicted wood dust concentrations (mg/m ) of the fixed-effects 46 model compared to the predicted wood dust concentrations (mg/m3) of the mixed-effects model. Figure 10 Job geometric means and their 95% confidence intervals from the 47-48 model building dataset, fixed-effects model and mixed-effects model. Figure 11 Job geometric means and their 95% confidence intervals from the 52 validation dataset, fixed-effects model and mixed-effects model. vi 1. Introduction A team of researchers from the University of British Columbia (UBC) has been studying the health of a cohort of sawmill workers. This cohort includes approximately 28,000 workers who were employed in 14 sawmills (hereafter referred to as the Sawmill Cohort) in British Columbia (B.C.) for at least one year between 1950 and 1995. The original study was initiated to examine the link between fungicide exposures and cancer incidence and mortality. The research program has since been expanded to include additional exposures, including wood dust, noise, and job strain, and additional health outcomes, such as fertility, birth outcomes, and heart disease mortality (Teschke et al, 1998). In 1995, the International Agency for Research on Cancer classified wood dust as a carcinogen, but the strongest evidence was from workers thought to be exposed to hardwoods (broadleaf trees) at high dust levels and very little exposure-response data was available (IARC, 1995). As B . C . sawmills mill only softwoods (conifers), the Sawmill Cohort provides an ideal opportunity to investigate cancer and mortality outcomes associated with softwoods. With the aim to reconstruct the exposure histories (1950-1995) of the cohort subjects, each mill in the Sawmill Cohort was visited to obtain a thorough technological history pertinent to wood dust exposure. Four of the mills were involved in an extensive inhalable and thoracic wood dust monitoring program conducted in 1996/97 (Davies et al, 1999; Teschke et al, 1998). Additional wood dust exposure measurements for B . C . sawmills were available, including measurements collected for two previously published studies (Teschke et al, 1999a; Teschke et al, 1994), measurements collected by a private consulting group on behalf of industry to evaluate compliance with regulations, and measurements collected by hygienists from the Workers' Compensation Board of B . C . (WCB) for compliance purposes. The aim of this project was to develop, validate, and compare two predictive statistical models (a fixed-effects model and a mixed-effects model) using the previously collected wood dust measurements and information gathered on potential exposure determinants. The model validation procedure will assist the U B C research team in choosing which model to 1 use to develop an exposure matrix to assess wood dust exposures in a mill-, job- and time period-specific manner for all 14 cohort mills for the time period 1950 to 1995. The completed exposure assessment will be used to examine the dose-response relationship of softwood dust exposure on health. 2. Background on Cohort Mills and Sawmill Processes The fourteen cohort mills formed the study base of a cohort enumerated for an earlier study of chlorophenates and cancer. These mills were selected to be representative of a cross-section of current mill technology and regions in B . C . The validation sawmill was a large interior sawmill that took part in a study of respiratory health of sawmill workers undertaken by U B C researchers (Teschke et al, 1999a). A l l mills are large commercial mills producing between 30 and 356 million board feet per year and employing between 150 and 500 employees (Table 1). Although specific equipment, log storage methods, and tree species may vary between mills, they share an overall common process. Prior to use logs are stored and sorted either in the water or on land (Boom, Log Yard1). The logs are cut to a specified length and debarked (Log Processing). In the sawmill, the logs are sawn to desired dimensions determined by the quality of the wood and the customers' requirements (Sawing, Sawmill Non-sawing). Further smoothing of the lumber surface may be required and is processed in the planermill (Planer). Waste wood is processed to create two sawmill by-products: wood chips and sawdust (hog fuel) (Chip and Hog). The lumber produced is graded, sorted, and then packaged, ready for shipment (Sorting and Packaging, Lumber Yard). Specialty trades maintain the saws (Sawfiling). Traditional trades may work primarily inside the lumber mill complex or in their trade-shops (Mill Maintenance, Non-mill Maintenance, Powerhouse). Cleaning, general labour, office, and supervisory staff round out the mills' work force (Cleaning/Labouring, Office, Foremen). 1 For modeling purposes, sawmill jobs were grouped into process groups based on related job tasks. These process groups are listed in parentheses next to their associated functions in the above description of sawmill processes. 2 FIGURE 1 Map of British Columbia1 indicating the locations of the cohort mills and the validation mill. CM 8 CM 4 & CM 6 CM 9 CM 7 CM 2 IM = interior sawmill C M = coastal sawmill CM 1, CM 3, & CM 5 Map adapted from the B . C . Forest Regions Map from the B . C . Ministry of Forests website: http://www.for.gov.be.ca/pab/publctns/getintch/provmap.htm 3 a a «« "3 <u a a a 9 -O o i_ a .9 a -3 a "3 a <u cu a s 2 a a ca a .2 _ 88 a *• vs g 5 a a £ ° ** ° 2 'C ° a-* a ^ ® -a _C UJ ^ 5 at es o I/) Oft o S S u SS Qj> CO 5 o o = at CO c. 111 O & 7 O. Z E < CS 3 i) u > ts s o u 2? o CU -+-» 00 cu o o ON u o E rt1 o o ro CN O ON oo a o 2 cu •3 •4-» "3 0 0 o o o T f .a bO 3 O Q o O S CU T f O <L> d o 2 "3 GO cu o o o 5 u T3 4> O O o cu to +H "3 CU o o o .a 60 3 O Q o 1 CU o o o Vi O ON T f T f m so o oo ON oo ON 00 ON c x m CN V co to -4-» "3 GO <L> o o o PH 3 O Q o o 1 X o o CN oo O T f 0 0 o o 5 o oo _ o CO T f oo CD W5 CD I* PH CD O o 00 .a PH (30 3 O Q o o S CD ac o o CN OS CN OS 0) CD CD CD d d d d o o o o 2 2 2 2 CD (Z) o o o_ CN .a PH 0 0 3 O Q o o CD <C J N o SO CN CN Os ON CD "3 0 0 CD o o o PH O0 3 O Q o o S IU T f IU "3 o .a PH 0 0 3 O Q «a =a =a «a u o S CD o ro cu "3 0 0 CD o B CU ac <« .a PH oo 3 O Q o o o o T f V~l O ro ro CN Os t CN 00 v© VO Os ON OV O - H ON rt rt ON (0 in ( N o o CN ON O O <U |"H CB J3 —1 co CU PH O o o o o o o o o^  ro rt WO CN OS NO ro O ro ro vi so ON O O o o so ON O o T3 TD o o rt ^ .a .a .a .a cu CD (U CU d .s .3 .3 CU CU CU CU o o o s s e s Cl, OH rt t/5 on t/5 O V~l Vi ro ro CN o o 00 Vi ro CN o so ON B •2 B B B B B B B B B CO CO CO CO co CO CO CO CO CO C3 C3 CS « csj CC C3 CO C3 CO O O O o O o O O O O O o U U U O O u U U U U .9 .2-3 4 3. Literature Review 3.1 Quantitative Retrospective Exposure Assessment Establishing a dose-response relationship is a major criterion for establishing causality between an exposure and a health effect. Past methods of retrospective exposure assessment have typically utilized exposure surrogates, such as occupation and the length of employment, or employed semi-quantitative exposure assessment methods (i.e. assigning jobs to low, medium, and high exposure categories). These methods have often been found to attenuate the dose-response relationship due to their potential for significant exposure misclassification (Stewart & Herrick, 1991). In today's work environment with lower exposure levels and confounding due to multiple exposures, the use of quantitative exposure assessment is becoming more important to limit exposure misclassification in order to detect health risks (Kauppinen, 1994; Stewart & Herrick, 1991). Several methods have been used to estimate quantitative levels for exposure assessment in cohort studies. These methods use available hygiene measurements and include subjective estimation by people with knowledge of the specific processes and time periods being assessed, arithmetic calculations (calculation of means, simple algorithms and marginal means), the use of multipliers to account for process changes (including data-derived factors, subjective factors and deterministic factors), and statistical modeling (Stewart etal, 1996a; Seixas & Checkoway, 1995; Stewart & Dosemeci, 1994; Goldberg et al, 1993). Unfortunately, quantitative estimation of historical occupational exposures has several difficulties to overcome: quantitative exposure data are usually available only for more recent time periods, available data are usually limited to "high" exposure jobs, available measurements may be short-term or area samples, and industrial operations rarely remain constant across time because of changes in job definitions, raw materials, process parameters and equipment, and other factors (Stewart et al, 1996b; Smith et al, 1991; Seixas & Checkoway, 1995). While several studies have observed an association between exposure to wood dust or a history of employment in wood-related occupations and cancer, only one study has reported 5 the risk of cancer in relation to quantitative wood dust exposure categories (Kauppinen et al, 1993). At best, other cancer epidemiological studies of wood dust exposure in wood related occupations ordinally classified workers according to level of wood dust exposure (for example: somewhat dusty, dusty, and very dusty) (Demers and Boffetta, 1998). Kauppinen et al (1993 & 1988) developed plant and period specific job exposure matrices to assign mean level and cumulative wood dust exposures to a small case-control study (4 Hodgkin's disease; 8 non-Hodgkin's lymphoma, 12 leukemia; and 152 controls) nested within a cohort of 7307 Finnish woodworkers from 35 plants. Historical and current hygienic data were utilized, using extrapolation when necessary to fill the gaps in the matrices, to assign job 3 3 groupings to a four categories: not exposed (<0.1 mg/m ); low (0.1-1 mg/m ); moderate (1-5 mg/m 3) and high (over 5 mg/m 3). No indications of increased risk or an exposure-response relationship with respect to respiratory cancers and wood dust were found in this study, which had an estimated mean level of exposure of 1 mg/m to mainly softwood species. 3.2 Statistical Modeling for Retrospective Exposure Assessment Statistical modeling can be a powerful tool for quantitative exposure assessments as it has the ability to consider the dataset as a whole and use the interrelation between the data segments to provide unbiased estimates of any one part of the matrix in a reproducible way (Stewart et ai, 1996b; Seixas & Checkoway, 1995). Predictive models are descriptive in nature and are developed by fitting the model using available exploratory variables to the observed data (Hornung et al, 1994; Hornung, 1991; Y u et al, 1990). Examples of the use of statistical modeling in retrospective exposure assessment include the road construction industry (Burstyn et al, 2000), the sterilization industry (Hornung et al, 1994), rubber manufacturing industry (Kromhout et al, 1994), the granite industry (Eisen et al, 1984), and the asbestos textile industry (Dement et al, 1983). For retrospective exposure assessment, a good predictive model should be able to reconstruct past exposure in a quantitative manner for the entire span of interest. Explanatory variables should be chosen rationally, preferably choosing variables known a priori to influence exposure (Yu et al, 1990). Independent variables that are not necessarily causal in nature, i.e. 6 calendar year, may be added to the model as a surrogate for unmeasured variables, such as improvement in work practices due to increased awareness of potential health effects, to accommodate trends in available sampling data (Burstyn & Teschke, 1999; Hornung et al, 1994; Y u et al, 1990). However, a predictive model is useful only i f the variables included in the model are available for all or nearly all of the jobs and plants and the entire time period of interest (Hornung et al, 1994). Other quantitative exposure assessment methods, such as extrapolation, may have to be used to fill any gaps that the data and determinants do not cover. While most predictive models have designated all independent variables as fixed effects, setting job category or worker as random effects is becoming more frequently used to explain additional variability in the dependent (exposure) variable (Burnstyn et al, 2000; Teschke et al, 1999a;Nylander-French^a/, 1999; Rappaport etal, 1999; Hall, 1999). The term "mixed-effects model" has been coined to describe models incorporating both fixed and random effects. By setting a variable as a random effect, the model takes into account the variability within the chosen factor, with an adjustment coefficient calculated for every value of that factor. The use of mixed-effects models in modeling personal exposures has allowed either the within-worker and between-worker or the within-job and between-job variance components, dependent on what variable was chosen as the random effect, to be calculated. These variance components have the potential of improving the model's predictive ability by accounting for additional variability in the measurements (i.e. within-worker or within-job variability) beyond that explained by the fixed effects. This application is important, as it has become clear that occupational exposures vary greatly both within workers over time and between workers in the same job (Rappaport et al, 1999). In a comparison of a fixed-effects model (where job was offered as a fixed effect) and mixed-effects model (where job was offered as random effect), Hall (1999) found that the variability explained by the model was slightly greater for the fixed-effects model, but that the mixed-effects model was 5% more precise in predictive ability. 7 3.3 Modeling Considerations - Grouping Strategies In exposure assessment for epidemiology, there has recently been a recognition of the benefit of grouped, instead of individual, estimates of mean exposure to reduce the amount of random variability associated with the estimate (Heederik & Attfield, 2000; Werner and Attfield, 2000; Tielemans et al, 1998; Seixas and Sheppard, 1996). Gains in precision of the estimate by grouping mean a tradeoff with losses in specificity; however, increased precision of exposure estimates usually results in substantially less attenuated dose-response relationships. Grouping strategies are most beneficial when the ratio of within-worker variance to between-worker variance is large, the contrast between exposure categories is large, i f sufficient measurements per category have been taken so that the estimate for that category is relatively precise, or if measuring workers repeatedly is not feasible (Heederik & Attfield, 2000; Tielemans et al, 1998). Variability in exposure concentrations from day-to-day can range over orders of magnitude (geometric standard deviations (GSDs) as high as 4 to 5 have been commonly noted). Even within groups of subjects chosen to have similar exposure distributions, the G S D can be high (Rappaport et al, 1993; Seixas & Checkoway, 1995). Characteristics routinely used to establish similarly exposed groups (job title and location) have been demonstrated to be only marginally related to the between-worker component of variation because a given job title may encompass different tasks and work environments (Rappaport et al, 1993; Goldberg et al, 1993). Modeling allows grouping to occur, while accounting for additional exposure determinants, such as work practices and characteristics within a job group that may account for a significant component of the variability in exposure levels. 3.4 Modeling Considerations - Choice of Exposure Metric The exposure metric used in exposure assessments depends on the health effect of interest and the availability of exposure information (Kauppinen, 1991). For chronic health hazards, the arithmetic mean has been found to best represent the body burden of the substance 8 (Rappaport, 1991). In occupational wood dust exposures, most of the airborne dust mass is contributed by particles larger than 10 um, which deposit in the nasal cavity (Hinds, 1988). As the increased risk of nasal cancer is believed to be the result of direct deposition of wood dust in the nasal cavity, inhalable particulate mass sampling is required to capture the exposure of interest in relation to nasal cancer (Hinds, 1988; A C G I H , 2001). The wood dust measurements collected in B . C . have utilized various sampling methods, including both open- and closed-face 37 mm sampling cassettes, inhalable samplers such as the GSP and the 7-hole sampler, a thoracic fraction sampler (PM10), and a respirable fraction sampler (37 mm sampler with cyclone separator). To increase the number of wood dust measurements that can be used in a predictive model, conversion factors can be used so that all the measurements approximate the inhalable dust measurements (Burstyn et al, 2000; Heederik & Attfield, 2000; Gao et al, 2000; Liden et al, 2000; Martin & Zalk, 1988; Hinds, 1988; Beaulieu et al, 1980). However, conversion factors between sampling methods are dependent on the underlying particulate size distribution and the homogeneity of the particulate shape, so the use and choice of conversion factors must be carefully evaluated. 3.5 Modeling Considerations - Model Validation Validating the statistical model is not routine in most studies, often because of the lack of historical measurements or the small number of measurements available, but it is an important step of model development (Seixas & Checkoway, 1995; Stewart & Dosemeci, 1994; Hornung, 1991). Cross-validation, or data splitting, is a common method used to assess the model's predictive capability, referred to as the external validity, against a reserved dataset (Hornung, 1991). Choices for the reserved dataset include withholding a randomly selected subset, withholding a non-randomly selected subset (i.e. data from a certain number of plants/facilities from a multi-facility study) or using the sampling results from a future sampling campaign or another researcher's study; the later two choices providing a more severe test of validity that more clearly examines the generalizability of the model (Burstyn & Teschke, 1999). Other methods for validating statistical models include sensitivity analysis (which identifies exposure factors that drive the model), field testing, comparing models against deterministic models based on physical principles, and by soliciting expert opinion or examining published data (Burstyn & Teschke, 1999; Hornung, 1991). 3.6 Reported Wood Dust Exposure Determinants Several studies have been conducted to determine exposure levels within sawmills and to determine factors that influence exposure (Table 2) (Hall et al, 2001; Douwes et al, 2000; Duchaine et al, 2000; Hall, 1999; Teschke et al, 1999a; Teschke et al, 1999b; Teschke et al, 1994). Exposure levels were found to depend on the mill department, with the planing department having the highest mean exposure. Higher exposure levels were associated with interior mills, indoor jobs, close proximity to specific machinery (such as chippers, planers and multiple saws), cleaning up sawdust, planing kiln-dried lumber, and milling of spruce and pine. Lower exposure levels were associated with coastal mills, outdoor jobs, personnel enclosures, and alpine fir or mixed tree species. Exposures were found to gradually decrease over time. Certain variables that are more likely to vary between mills rather than within mills, such as production level and log storage methods, have not been adequately assessed since most studies have been limited to one or two mills (Teschke et al, 1999a; Teschke et al, 1994). In one study which examined multiple worksites, including sawmills and other industries where wood dust exposures occur, variables such as year, state, and industry were found to be related to dust exposure level, but are likely surrogates for other factors that directly influence exposure, such as the use of engineering control measures that were not available (Teschke et al, 1999b). 10 TABLE 2 Exposure Determinants Reported by Four Previously Published Determinants of Exposure Studies on the Wood Industry. Hall et al, 2001 Teschke et al, 1999a Teschke et al, 1999b Teschke et al, 1994 Department V V V Mill Location V V Job Location (Indoor/Outdoor) V V x Production Levels V X Year V V Company V Job V V Season X x Species X V Industry V Purpose of Sampling X x Number of Employees X Booth Enclosure V V Proximity to Saws V V Job Tasks V Wood Condition V Ventilation X x Weather V x Enclosure of Saws X V Variable was found to be significant in statistical analyses. X Variable was not significant exposure determinant because it was not associated with exposure in simple regression analyses (potentially due to lack of variability in the factor), it was strongly correlated with other variables, or it was not significant in statistical models. 11 4. Methods 4.1 Exposure Data A database of dust exposures in B . C . sawmills (referred to as the sawmill database) was created from personal gravimetric samples collected from five separate sources (Table 3). "Field Blank" samples, area (stationary) samples, records without job titles, and records that failed other quality control criteria specified by the original researcher, such as large differences in pre- and post-sample flow rates, were removed from the individual datasets prior to being merged into the sawmill database. Only dust measurements taken with inhalable or total-dust samplers (open-face 37 mm cassette, closed-face 37 mm cassette, seven hole sampler, or GSP) at the 14 cohort mills and one additional interior mill were included (n=1791). A l l other mills were excluded as no exposure determinants were collected for these mills. Thoracic and respirable sampling methods were excluded from the dataset as these particulate size specific samplers are designed to have 50% capture efficiencies at approximately 10 pm and 4 um, respectively ( A C G I H , 2001), and as such are not related to the inhalable fraction where the majority of the mass is contributed by particulate greater than 10 um (Hinds, 1988). The sawmill database included information on date of sample, season, sampling method used, mill name, flow rate, sampling duration, dust concentration (in mg/m 3), original job title, standardized job title, mill department, tree species, wood condition, and mill geographic location. Not all datasets included information on each variable. Tree species refers to the type of tree milled at the time of study. Wood condition provides information on whether the wood the worker was exposed to was green (not treated) or kiln-dried. M i l l geographic location refers to whether the mill was a coastal mill or located in the interior of B . C . A member of the Sawmill Cohort research team standardized the original job titles and departments. The standardization was based on setting all analogous job titles to one standard identifier by dropping modifiers or standardizing spelling (i.e. '#l-edger', 'left-edger', and 'edgerman' became 'Edger Operator') and by standardizing terminology when 12 different mills used different terms to refer to the same tasks, processes and locations (i.e. 'gate operator' and 'dropsorter' are different terms both referring to the worker who makes preliminary grading decisions and directs lumber within the sawmill). In a few cases of rare jobs they were pooled (e.g. 'scowman' refers to all jobs loading chips into rail cars, trucks, as well as scows). The 87 Research-1,90 Research-2, 37 Research-3, 96 Compliance-1, and 39 Compliance-2 original job titles were standardized into 70 jobs. The original departments were similar to each other, requiring only spelling standardization or simple grouping or splitting of departments based on jobs to be standardized into 11 departments. Detection limits varied with time and sampling method. For the statistical analyses, samples below the limit of detection were assigned a value of XA the detection limit. The following quality criteria were applied to the datasets: > 42 (2.3 %) samples were excluded because the individual spent less than 75% of the day at one job. Only the Research-1, Research-3, and Compliance-1 datasets provided this level of detail. For the other sources, it was assumed that the job title given represented greater than 75%) of that individual's work that day. Samples that were rejected by this criterion accounted for 4.5% of the Research-1, 4.4% of Research-3, and 2.2 % of the Compliance-1 samples. > 12 (0.7%) samples were excluded because they had a sample duration of less than 2 hours for production jobs, or less than 4 hours for non-production jobs. These cut points were set because shorter sampling durations may not accurately characterize mean shift exposure. The cut point for non-production jobs was more stringent as longer sampling durations are required to capture the true mean shift exposure due to the task-dependent variability in exposure. > 3 (0.2%) samples were excluded as likely being erroneous, having concentrations greater than 100 mg/m . The remaining measurements had a maximum concentration of 43 mg/m . The final data set consisted of 1734 records (96.8 % of initial sawmill database) that met the quality control criteria. 13 a E es Ji OS OT M l e S es OT o a. S PH —S IS1 o o 0X1 _c "H< & es OT U <y C3 Q O N O N O N Q u o PI D oo CN NO co CD O N O N 1 ) CD >>CN O cu .2 S £ cn CD J 3 cx — W « S f cn — o cx cx g - I £2 -o § § If CD S c o CN CN CN I JS Pi © ° o O O N O O N CN ~ cn JD <5 2 S o n» cn O t—1 1 ca U CQ T l -CN oo CN CN J = o 1> cn ai J 2 O 3 T3 CX I OT "E. S o O o U CN CD O o O —? oo ° O N O O N S CN O N — DC .3 &! * CD -a u JD CX "o OT c n O N » 2 .1 J 3 00 3 cx oo e S ™ ^ O S o « ^ CS cn co CD cn cn J D CX cn 3 < OT CD cn e ' O N c o N O CD O c cd O O 14 4.2 Potential Exposure Determinants Potential exposure determinants to include in the predictive models were identified from variables collected alongside the dust measurements, the mills' technological histories (information obtained from mill production reports, schematics, walk-through surveys, and interviews with key staff during site visits), and from determinants of exposure studies on the wood industry. Variables that could not be assessed with a high degree of confidence for all jobs, mills, and the entire time period (1950-1995) were excluded from further consideration. The variables that were available for the scope of the study were added to the dataset on a record-by-record basis. These variables can be divided into study specific, mill specific, and job or process group specific variables and were defined as follows: Study Specific Variables: "Sampling Method" was available for all data sources and was coded categorically into one of four sampling methods: open-face 37-mm cassette, closed-face 37-mm cassette, 7-hole sampler, or GSP. "Season" was defined from the sample date included in all data sources and coded as an indicator variable. Samples collected May through October were classified as summer samples ('0') and samples collected November through April were classified as winter samples (' 1'). "Year" was available from the sample date and was added as a continuous variable after zeroing the year, with 1980 representing the zero value (in order to reduce the size of the intercept term in the model). "Sampling Strategy" was defined as the purpose of the study and was classified as to research (Research-1, Research-2, Research-3), compliance (Compliance-2), and mixed (Compliance-1 and subset of Compliance-2 that was collected for research purposes) and treated as a categorical variable. 15 Mill Specific Variables: "Log Storage Method' was obtained from the technological histories and site visits and was defined as the location where the logs were stored prior to milling. Three categories were defined: salt water stored logs, land stored logs, and a combination of fresh water and land storage. "Mill Location" was obtained from the geographic location of the mill and coded as an indicator variable into coastal mills (' 1') and interior mills ('0'). "Tree Species" was available as observational data from some sources and available either in the original database or from written reports. Where species was unknown, industry publications were consulted (Madison's; Random Lengths; B . C . Ministry of Forests; Directory of Forest Products; Big Book - The Buyer's & Seller's Guide). Seven categories of softwood tree species based on the predominant species milled were defined as a categorical variable for the purposes of modeling into the following groups: spruce/pine/fir (SPF); SPF + other species; hemlock; hemlock + other species; western red cedar (WRC); W R C + other species; and douglas fir. "Annual Production Levels" in million board feet (Mmbf) were obtained from industry publications (Madison's; Random Lengths; B . C . Ministry of Forests; Directory of Forest Products; Big Book - The Buyer's& Seller's Guide) and from data obtained directly from the mills. Production levels were categorized into four levels: < 100 Mmbf, 100-150 Mmbf, 150-275 Mmbf, and > 275 Mmbf. Where gaps existed, the production level category was interpolated from the available data. "Size of sawmill" was obtained from mill schematics and defined as the area of the mill's sawmill in square meters (m 2). Mills were being constantly upgraded; however, blueprints were at best only available when a mill was rebuilt or underwent a major upgrade. This variable was categorized to capture significant mill changes, yet reduce misclassification due to the smaller changes. The mills were categorized into one of four categories: less than 6000 m 2 , 6000-7500 m 2 , 7500-13000 m 2 , and greater than 13000 m 2 . 16 "Number of Employees" was obtained from the work history files collected from the cohort mills. A s the number of employees changed over a year, for the purposes of analyses the number of employees employed on July 1 s t of the year for each mill was calculated, and then categorized as follows: less than 200 employees, 200-350 employees, and greater than 350 employees. "Production per m2" was calculated by dividing the annual production by the area of the sawmill, and then categorized to provide a measure of density. The categories were as follows: less than 0.0125 Mmbf/m 2 , 0.0125-0.025 Mmbf/m 2 , 0.025-0.050 Mmbf/m 2 , and greater than 0.050 Mmbf/m 2 . "Production per employee" was calculated by dividing the annual production by the number of employees, and then categorized to provide a measure of level of automation. The four categories were less than 0.2 Mmbf/employee, 0.2-0.4 Mmbf/employee, 0.4-0.6 Mmbf/employee, and greater than 0.6 Mmbf/employee. "Planermill on Site" was defined as to whether the mill had a planermill (' 1') or no planermill ('0'). The planer mill process typically results in a finer dust (Hinds, 1988) and is typically one of the most highly exposed areas (Douwes et al, 2000; Duchaine et al, 2000; Teschke et al, 1999;Teschke et al, 1994). The dust produced in the planer mill settles on the lumber and may be disturbed in later handling, so the existence of a planer mill on site can affect the exposures in process groups that handle the milled lumber. Only two mills (both coastal mills) did not have a planer mill. Job or Process Group Specific Variables: "Job Title" was obtained from the standardized job titles available in the sawmill database. There were seventy job title categories. 17 "Process Group" was determined from the standardized job title. Each job was assigned to one of 16 process groups defined by related job tasks and processes (see Section 2 for groups). "Department" was obtained from the original data sources. Usually, department was defined as the administrative unit to which each individual worker was assigned. For jobs that were found in more than one area of the mill, such as Cleanup, Labour, or Foremen, department provided additional information on what area of the mill the specific worker was assigned. "Wood Condition1'' was obtained from the original data sources and, where missing, obtained from the mills' technological histories. Wood condition was defined as to whether lumber milled was green (untreated) (' 1') or was a mix of green and kiln-dried lumber ('0'). A separate kiln-dried only variable was not possible because most mills only kiln-dry a portion of the lumber, so when this information was missing from the original data source determining whether the exposure was mixed or kiln-dried was impossible. "Job Location", coded as inside ( T ) or outside ('0') jobs were obtained from the original data sources and, where missing, obtained from the mills' technological histories. Jobs were defined as inside jobs if the job was located in the sawmill, planer mill, powerhouse, maintenance buildings, or chip and hog buildings. Jobs were defined as outside i f they were in the yard, in vehicles, or in small outbuildings where no milling or maintenance operations took place. "In Booth" was obtained from the original data sources and, where missing, obtained from the mills' technological histories. Booth enclosure was coded as an indicator variable, with ' 1' representing working in booth and '0' representing no booth. Jobs were defined as being in booths if greater than 10% of their job tasks took place in a booth or enclosure separating the worker from dust sources. Vehicle enclosures were not defined as booths and were considered separately (see next variable). 18 "In Vehicle" was defined from the job title information. A l l jobs were defined as to whether the job was operating a vehicle (' 1') or not operating a vehicle ('0'), and coded as an indicator variable. 4.3 Data Pre-treatment Measurement data that used 37mm closed-face cassettes and the seven-hole sampler were converted to inhalable dust concentrations as defined by the GSP sampler using inter-sampler ratios calculated by Davies et al (1999) for the cohort mills. Measurements taken using the 37mm open-face cassette were converted to 37mm closed-face equivalent concentrations using Beaulieu et aVs (1980) conversion factor prior to using Davies et a/'s conversion factors. The Research-3 dataset of a non-cohort interior B . C . sawmill that met the inclusion criteria (n=217) were removed from the data set prior to analyses to be used for model validation ("validation dataset"). The remaining data (n=1517) was used for modeling ("model building dataset"). Prior to beginning modeling, the job classifications in the model building dataset were reviewed. Any job with less than four observations was combined into another job category based on similarity of tasks. A l l jobs in the process group Planing (Tilthoist, Planer Feeder, Planermari) were combined as their geometric means were not statistically different (p=0.989, A N O V A ) . Jobs in the process group Boom {Boat Operator, Boomman, Slipman) were combined due to low number of observations per job and because their job geometric means were not statistically different (p=0.749, A N O V A ) . Jobs in the process group Office {Watchman, Clerk, First Aid, Manager/Superintendent) were combined due to very low number of observations (n=6) and similar exposure levels (p=0.308, A N O V A ) . The process groups Boom and Office were further combined into one process group as their process group geometric means were not statistically different (p=0.904, A N O V A ) . 19 One job, Bucker, with only one observation in the model building dataset was removed from the dataset due to the inability to satisfactorily reclassify the job. The Bucker'% job tasks are unique, being the only sawmill job with continuous use of a chain saw. With the removal of Bucker from the datasets, the model building dataset had 1516 observations and the validation dataset had 213 observations. 4.4 Statistical Analysis Mixed-effects modeling was conducted using S A S (PROC M I X E D procedure), all other statistical analyses were conducted using SPSS. 4.4.1 Preliminary Analyses A probability plot of the data revealed that the exposure data was positively skewed and resembled a lognormal distribution. A probability plot revealed the log-transformed data (base e) was approximately normal. As such, dust concentrations were log-transformed (base e) prior to analyses. The potential exposure determinants were initially evaluated using simple linear regressions analysis. Only independent variables associated with wood dust concentration with p < 0.20 were considered for inclusion in the predictive models. Potential nested variables were identified, such as variables only applicable in one department, process group, or job category, or where the effect of the variable is expected to change dependent on the mill area. These terms were evaluated in initial univariate analyses in their nested form. The correlations between ordinal and indicator variables were examined using Pearson r statistic. Among pairs with r > 0.60, only one determinant was selected for entering into the model. The determinant was selected based on which factor more logically explained the associated exposure and/or which factor was more consistently and reliably available over the scope of the study mills, jobs, and time spans. Categorical variables with no intrinsic 20 order (nominal variables) were evaluated for correlations using symmetric measures (Cramer's V and contingency coefficient) (Fisher et al, 1993). 4.4.2 Model Building: Fixed-Effects Model A manual backwards regression procedure using generalized linear model ( G L M ) was used to develop the fixed-effects model. A l l variables meeting the simple linear regression and correlation criteria were offered into the model as fixed effects. Variables with the highest p-value (>0.10) were eliminated one at a time, until all included variables had a p-value < 0.10. Categorical variables, such as process group, were entered and removed as a group based on the statistical significance of the entire group. Residual plots were examined for patterns in unexplained variance and Cook's D were used to identify influential values. The general model used to develop the fixed-effects model is described by the following expression: Ln(concentration) = Y + PiX, + p 2 X 2 + . . . + pnXn + e where Y (the intercept) is the background concentration not explained when all independent variables are equal to zero; Pi . . . P„ are the regression coefficients of each independent variable; Xi . . . Xn are the values of the independent variables, entered as a continuous variable or indicator variable; and e is the residual error in model. When variables were entered as nested variables, their contribution to the model equation is in the following form: PiX; = PJXAJXBJ where ' i ' is the variable of interest, X A was an indicator variable (0/1) indicating whether variable ' i ' applies and X B is the value of the nested variable ' i ' . 4.4.3 Model Building: Mixed-Effects Model A l l variables meeting the simple linear regression and correlation criteria were offered into the mixed model as fixed effects, with the exception of job category. Job category was entered in the mixed model as a random effect. A manual backwards regression procedure using SAS's P R O C M I X E D procedure was used to develop the model. The between-job and 21 within-job variance was estimated to determine if the random effects explain any additional variation in the model. Fixed effect variables with the highest p-value (>0.10) were eliminated one at a time, until all included variables had a p-value < 0.10. Categorical variables, such as process group, were entered and removed as a group based on the statistical significance of the entire group. Residual plots were examined for patterns in unexplained variance and Cook's D were used to identify influential values. The general model used to develop the mixed-effects model is described by the following expression: Ln(concentration) = Y + P1X1 + p 2 X 2 + . . . + P„Xn + Xi + e where Y (the intercept) is the background concentration not explained when all independent variables are equal to zero; Pi . . . P„ are the regression coefficients of each independent variable; X i . . . X n are the values of the independent variables, entered as a continuous variable or indicator variable; $ is the random effect of the i t h job, with each job having a unique %; and 8 is the random within-job variation. Nested variables entered in the mixed-effects model are entered in the same form described above for the fixed-effects model. The model's algorithm assumes that # and 6 are normally distributed with zero means and 2 2 variances BJO" (between-job variance) and WJO" (within-job variance), respectively, and are mutually independent. 4.4.4 Model Validation The validation dataset was used to compare the predictive abilities of the two (fixed-effects and mixed-effects) models, as measured by the models' bias and precision using a cross-validation technique described by Hornung et al (1991) and by evaluating correlations between the predicted and measured concentrations using Pearson r. Bias is the average difference between the estimates and the true exposure level and can be calculated as the mean of the residuals, which is the difference between observed and predicted values (Equation 1). Precision is a measure of the degree of variability of the estimate of a true exposure level and can be calculated as the standard deviation of the residuals (Equation 2). 22 Bias = ^ (y>- yd 1=1 Precision = [0"- yi)-bias]2 no-I [Equation 1] [Equation 2] where yi is the predicted exposure level for the ith set of exposure factors in the validation dataset; y; is the measured exposure level for the rth set of exposure factors, and no is the number of observations in the validation dataset. Bias and precision were calculated on the log-transformed values. The bias was then converted into mg/m (Equation 3) and precision converted into the geometric standard deviation of bias (Equation 4). Bias in mg/m 3 = exp[Ln(geometric mean) + Bias] - Geometric Mean [Equation 3] Geometric Standard Deviation of Bias = exp(Precision) [Equation 4] 95% confidence intervals (95% CI) around the job geometric means were calculated as test-based confidence intervals on the log-transformed measurements and then converted to mg/m to obtain the confidence interval. 23 5. Results 5.1 Description of the Datasets 5.1.1 Model Building Dataset Dust measurements were available for 13 of the 14 cohort mills, with an average of 117 observations per mill. Two mills had less than 10 measurements and 7 mills had more than 100 measurements (Table 4). The measurements were made between 1981 and 1997, with an average of 89 measurements per year, with some years having no measurements to a maximum of 413 measurements in 1997 (Figure 2). The later eight years, between 1989 and 1997, accounted for 93% of the measurements. Interior mills accounted for 25% (n=386) of the measurements. Inside jobs accounted for 81% (n=1231) of the observations. O f the measurements from inside jobs, 14% (n=177) of the measurements involved the worker spending greater than 10% of their time in a booth. The dust measurements were converted to inhalable dust equivalents with the use of inter-sampler ratios (Davies et al, 1998). Sixty-five per cent of the measurements were taken with open-face 37 mm cassettes, 5% with closed-face 37 mm cassettes, 3% with seven-hole samplers, and 27% with the GSP. The data was positively skewed and resembled a lognormal distribution (Figure 3). The arithmetic mean (AM) of the dataset was 1.4 mg/m , with a range of 0.03 to 42 mg/m . The geometric mean (GM) was 0.73 mg/m and the geometric standard deviation (GSD) was 2.98. O f the observations, 3.8% (n=57) were greater than 5 mg/m . The process group geometric means ranged from 0.19 to 1.58 mg/m (Table 5). The highest exposed process groups were Cleaning/Labouring ( G M = 1.58 mg/m ) and Planing ( G M = 1.13 mg/m ). The Boom/Office process group had the lowest means ( G M = 0.20 mg/m") exposures. O f the 58 job categories, 39 jobs had more than 10 measurements. The job geometric means ranged from 0.19 to 1.63 mg/m 3 (Table 6). Nine job categories had geometric means greater 24 than 1 mg/m 3 and nine job categories had geometric means less than 0.5 mg/m 3. The most highly exposed jobs were Cleanup ( G M = 1.63 mg/m3) and Labour ( G M = 1.27 mg/m3) and the lowest exposed jobs were Barker Operator ( G M = 0.23 mg/m 3) and Boom ( G M = 0.19 mg/m3). FIGURE 2 Distribution of Dust Measurements in the Model Building Dataset by Year and Data Source. 450 400 350 300 250 200 150 100 50 0 • Compliance-2 • Research-1 M Compliance-1 • Research-2 1981 1982 1 983 1 984 1985 1986 1987 1988 1989 1990 1 991 1992 1993 1 994 1995 1996 1997 25 TABLE 4 Number, Arithmetic and Geometric Means, Concentration Range, and Source of Dust Measurements by Mill. Arithmetic Standard Geometric Range of Mill N* Mean Deviation Mean Concentrations Source & Years Available (mg/m3) (mg/m3) (mg/m3) (mg/m3) Coastal Mill 1 178 0.65 1.30 0.36 0.1 - 14.6 Research-2, n= 156, 1989 Compliance-2, n=22, 1996 Coastal Mill 2 79 0.48 0.91 0.26 0.1 -7.2 Research-2, n= 179, 1989 Compliance-1, n=55, 1995 Coastal Mill 3 386 1.18 1.96 0.82 0.06 - 28.0 Compliance-2, n=331, 1982-83, 1987-88, 1993, 1996 Coastal Mill 4 54 1.70 1.48 1.26 0.13-7.9 Compliance-1, n=54, 1995 Coastal Mill 5 8 3.40 5.60 1.78 0.73 - 17.2 Compliance-2, n=8, 1997 Coastal Mill 6 72 1.35 1.81 0.90 0.13 - 13.3 Compliance-1, n=72, 1995 Coastal Mill 7 144 1.05 2.26 0.54 0.03 - 17.7 Research-l,n=1418, 1997 Compliance-2, n=3, 1990 Coastal Mill 8 83 1.60 3.73 0.84 0.13-29.2 Compliance-1, n=37, 1995 Compliance-2, n=46, 1983-85, 1988 Coastal Mill 9 2 0.39 0.02 0.39 0.38 - 0.40 Compliance-2, n=2, 1986 Coastal Mill 10 124 1.29 3.93 0.40 0.04 - 38.4 Research-l,n= 124, 1996-97 Coastal Mill 11 0 - - - - -Interior Mill 1 135 1.91 3.56 0.88 0.05 - 42.2 Research-1, n=135, 1996-97 Interior Mill 2 136 2.65 4.09 1.63 0.16-40.9 Research- l,n= 128, 1997 Compliance-2, n=8, 1981, 1985 Interior Mill 3 115 2.27 3.93 1.64 0.04 - 38.4 Compliance-1, n=96, 1995 Compliance-2, n=19, 1983, 1985-86 * N = Number of Measurements 26 FIGURE 3 Distribution of Dust Measurements in Model Building Dataset by Concentration (mg/m3). 600 1 1 <Q 500 0.0-.5 1.0- 1.5 2.0-2.5 3.0-3.5 4.0-4.5 .5-1.0 1.5-2.0 2.5-3.0 3.5-4.0 4.5-5.0 Dust Concentration (mg/m3) TABLE 5 Mean Exposure Levels (mg/m3) in Model Building Dataset by Process Group from Highest to Lowest. Model Building Dataset Validation Dataset Process Group N GM GSD AM N GM GSD AM Cleaning/Labouring 108 1.58 2.63 2.71 18 4.38 3.1 8.29 Planing 95 1.13 2.8 1.75 20 3.71 1.14 4.44 Sawfiling 80 1.00 2.44 1.62 Powerhouse 15 0.86 3.77 1.88 Mill Maintenance 99 0.83 3.61 2.13 1 2.33 . . . 2.33 Non-mill Maintenance 34 0.80 2.61 1.24 Non Sawing (Sawmill) 183 0.79 2.61 1.24 7 2.05 1.43 2.15 Sawing 202 0.79 3.03 1.75 27 0.87 1.86 1.04 Chip and Hog 61 0.73 2.71 1.30 9 1.94 5.23 7.63 Log Yard 7 0.68 3.39 1.33 12 2.97 3.46 5.24 Sorting and Packaging 400 0.66 2.62 1.13 70 1.37 1.98 1.79 Lumber Yard 85 0.52 2.84 0.93 23 2.28 2.05 2.97 Foreman 46 0.42 2.52 0.62 Log Processing 73 0.30 3.19 0.59 26 0.74 2.23 0.99 Office/Boom 28 0.20 2.27 0.28 All 1516 0.73 2.98 1.42 213 1.65 2.67 2.98 N=number of measurements; GM=geometric mean; AM=arithmetic mean; GSD=geometric standard deviation; 27 T A B L E 6 Mean Exposure Levels (mg/m3) in Model Building Dataset by Job (n>10) from Highest to Lowest. Job Category Number of Geometric Mean Geometric Standard Arithmetic Mean Observations (mg/m3) Deviation (mg/m3) GM > 1.0 mg/m3 Cleanup 93 1.63 2.75 2.91 Labour 11 1.27 1.58 1.42 Benchman 21 1.20 3.13 2.38 Tailsawyer 52 1.19 2.64 1.79 Strip Handler 13 1.15 2.83 2.09 Planer 95 1.13 2.80 1.75 Resawyer 64 1.11 2.77 2.15 Sawfitter 29 1.06 1.77 1.30 Mechanic 22 1.01 1.99 1.25 GM 0.75 -1.0 mg/m3 Spotter 30 0.96 2.56 2.24 Utility 27 0.95 2.36 1.50 Grinderman 17 0.92 2.64 1.68 Chipper Feeder 38 0.89 2.46 1.53 Millwright 41 0.85 2.29 1.18 Trimmer Operator 132 0.78 2.48 1.29 Gangsawyer 17 0.75 2.03 0.95 GM 0.50 - 0.75 mg/m3 Sawfiler 13 0.73 2.51 1.05 Oiler 15 0.70 5.10 3.03 Dropsorter 73 0.70 2.44 0.99 Forklift 32 0.67 3.13 1.38 Offbearer 75 0.64 2.51 1.19 Grader 67 0.64 2.16 0.82 Log Deck Operator 17 0.63 3.22 1.08 Stacker Operator 53 0.63 3.32 1.17 Edger Operator 60 0.61 2.77 1.00 J-Bar Sorter Patrolman 17 0.59 2.69 0.86 Carrier Driver 11 0.57 2.48 0.77 Package Press Operator 18 0.56 2.69 0.89 Timber Deckman 15 0.55 2.36 0.77 Electrician 20 0.54 3.90 1.92 GM < 0.5 mg/m3 Blockman 11 0.46 3.25 0.79 Department Foremen 46 0.42 2.53 0.62 Head Sawyer 17 0.39 3.94 0.82 Lumber Straightner 12 0.33 2.32 0.47 Tallyman 20 0.31 2.59 0.52 Crane Operator 17 0.26 1.95 0.31 CutOff Sawyer 31 0.26 3.39 0.54 Barker Operator 25 0.23 2.41 0.32 Boom 22 0.19 2.44 0.29 28 5.1.2 Validation Dataset The validation dataset consisted of 213 dust measurements made using a seven-hole sampler from one interior B . C . sawmill milling non-allergenic softwood in 1996. The seven-hole sampler dust measurements were converted to equivalent inhalable dust levels based on the GSP with the use of inter-sampler ratios (Davies et al, 1998). The data was positively "3 skewed and resembled a lognormal distribution, with a range of 0.06 to 43 mg/m , an •3 -3 arithmetic mean of 2.98 mg/m , a geometric mean of 1.65 mg/m , and a geometric standard deviation of 2.67 (Figure 4). 14% (n=29) of the measurements were greater than 5 mg/m . Inside jobs accounted for 66% (n=140) of the observations. O f the measurements from inside jobs, 14%> (n=20) of the measurements involved the worker spending time in a booth. Only 24 of the 58 job categories and 10 of the 16 process groups were present in the -3 validation dataset. The process group geometric means ranged from 0.74 to 4.38 mg/m (Table 5). The highest exposed process groups were Cleaning/Labouring ( G M = 4.38 •3 "3 mg/m ) and Planing ( G M = 3.71 mg/m ). This mill did not have a Boom process group, as all logs were land stored. Jobs in the Foremen, Non-mill Maintenance, Office, Powerhouse, and Sawfiling process groups were not sampled in this study. O f the 24 job categories, 18 jobs had 5 or more measurements. The job geometric means -3 ranged from 0.65 to 6.70 mg/m (Table 7). Eight job categories had geometric means greater than 2 mg/m and two job categories had geometric means less than 1 mg/m . The most highly exposed jobs were Janitor (6.70 mg/m 3) and Strip Handler (4.60 mg/m 3) and the -3 lowest exposed jobs were Cut-off Sawyer (0.65 mg/m ) and Chipnsaw Operator (0.97 mg/m 3). 29 FIGURE 4 Distribution of Dust Measurements in Validation Dataset by Concentration (mg/m3). 60 T 0.0- 5 1.0- 1.5 2.0-2.5 3.0-3.5 4.0-4.5 .5-1.0 1.5-2.0 2.5-3.0 3.5-4.0 4.5-5.0 Dust Concentration (mg/m3) TABLE 7 Mean Exposure Levels (mg/m3) in Validation Dataset by Job (n>5) from Highest to Lowest. Number of Geometric Mean Geometric Standard Arithmetic Mean Job Category Observations (mg/m3) Deviation (mg/m3) GM > 2 mg/m3 Janitor 8 6.70 3.95 13.51 Strip Handler 8 4.60 1.95 5.46 Planer 20 3.71 1.78 4.44 Cleanup 10 3.12 2.26 4.11 Log Loader Operator 12 2.97 3.46 5.24 Kiln Operator 7 2.82 1.97 3.36 Cat Operator 5 2.49 9.98 12.58 Forklift 14 2.06 2.20 2.87 GM 1-2 mg/m3 Dropsorter 5 1.90 1.44 0.35 Stacker Operator 14 1.36 1.68 1.53 Barker Operator 6 1.31 2.12 1.74 Trimmer Operator 8 1.26 1.28 1.29 J-Bar Sorter Patrolman 10 1.20 1.29 1.24 Package Press Operator 16 1.09 1.70 1.24 Edger Operator 15 1.04 1.70 1.17 Grader 14 1.02 2.00 1.25 GM < 1 mg/m3 Chipnsaw Operator 8 0.97 1.71 1.14 CutOff Sawyer 16 0.65 2.11 0.77 30 5.2 Predictive Models 5.2.1 Preliminary Analyses Initial simple linear regressions of all variables as main effects removed season (p=0.225) and 26 of 58 jobs (p>0.2) from further consideration. Year was also removed in simple linear regressions. While year was significant in univariate analysis (p=0.016), the direction of the effect was contrary to previously published determinants of exposure studies (Hall et al, 2001; Teschke et al, 1999b) as it indicated an increase in exposure with time. Reviewing the distribution of measurements by year and data source (Figure 1) demonstrated that only the Compliance-2 data source sampled over more than two years; suggesting that any time trend seen may be an artifact due to study and mill differences, rather than a true time trend. M i l l location (coastal or interior) was highly correlated with tree species and log storage methods. Log storage method was dependent on mill location: coastal mills stored their logs in salt water, interior mills stored their logs on land or a combination of land and freshwater. A l l interior sawmills milled the same mixture of tree species: spruce, pine and fir (SPF). The coastal mills had more variety in tree species, but did not use SPF. A n examination of the coefficients for tree species (entered as a categorical variable) when entered in the full model revealed that only SPF was statistically different (p<0.10) and that the coastal tree species were not statistically different from the reference group (Douglas Fir). A s such, mill location was chosen to offer in the multiple regression models, and would be representative of both log storage and tree species. Production per m 2 and production per employee were strongly correlated as both were based on the annual production. Dust levels in a building might be expected to be dependent on the dust produced divided by the volume in the building and adjusted by the ventilation rate. While mill volume was not available, the mill area is closely related to volume, so production per m 2 was believed to provide the best potential explanatory measure and was chosen over production per employee. 31 Number of employees was correlated with mill size. Sampling strategy was also highly correlated with the number of employees and mill size, suggesting that compliance samples were most likely taken at the larger mills. As a measure of mill size was accounted for in production per m 2 , and because compliance samples were taken mainly at the large mills, the number of employees was chosen to be entered into the model. Process group was correlated with job location (inside/outside). Since process group was defined based on job tasks and the location of the tasks within the mill, most process groups contained either completely inside jobs (Sawing, Planing, Mill Maintenance, Non-mill Maintenance, Sawfiling, and Powerhouse) or completely outside jobs (Boom, Lumber Yard, Log Yard). To account for this correlation, the job location variable was only nested within the process groups for which there were variability in job location (Log Processing, Sawmill Non-sawing, Foremen, Chip and Hog, Cleaning/Labouring, Sorting & Packaging, and Office). After variables were removed for co-linearity, nested variables were identified, and then reassessed with regression analyses in their nested forms. Nesting was considered when a variable was specific to a mill area (i.e. wood condition) or process group (i.e. inside/outside), or where the effect of the variable is expected to change dependent on the mill area. Both booth and vehicle were nested within job location. Booth has been found to reduce exposures in previous studies (Teschke et al, 1999a); however, the magnitude of the reduction might be expected to be different i f the job was inside or outside the mill. To account for this difference, booth was nested within job location. Four combinations represent this term in the model: (1) inside = yes & booth = yes; (2) inside = yes & booth = no; (3) outside = yes & booth = yes; and (4) outside = yes & booth = no. Vehicle jobs were de facto nested within the outside jobs (in the form outside=yes x vehicle job=yes or no) as no vehicle jobs were located inside. 32 Due to strong correlations and because some variables were only relevant to specific process groups, several variables were nested within process group: inside/outside, department, planermill, and wood condition. These nested variables were all coded as 0/1. Because process group was entered as a main effect, the nested terms provided an adjustment factor on the process group estimate, which applied only when the nested variable equaled '1'. For most process groups, the administrative department was constant. For process groups where more than one administrative department existed, department was nested within process group in the form 'specific process group' x 'department' where both are indicator variables (0/1). If the nested department term stayed in the model, it provided an adjustment term on the process group coefficient specific to that department. After assessing the nested terms in simple regression analyses, department was entered in the models nested within the following process groups: Chip and Hog, Sorting & Packaging, Lumber Yard, Cleaning/Labouring, Foremen, and Mill Maintenance. The nested department term was not significant (p > 0.2) for the other process groups. Job location (inside/outside) was nested within process group due to strong correlations with process group. After assessing the nested job location term in simple regression analyses to remove insignificant nested terms, job location (1 = inside; 0 = outside) was entered in the models nested within the following process groups: Chip and Hog, Log Processing, Sorting & Packaging, Foremen, and Sawmill Non-sawing. The nested job location variable stayed in the model only when the inside dust exposures were different than the outside dust exposures within a process group. The variable "Planermill on site" was nested in the process groups where a planermill was expected to increase exposure, specifically the process groups that receive the milled lumber. The type of lumber output is different depending on the existence of a planer mill on site. The planer mill process typically results in a finer dust and is typically one of the most highly exposed areas (Hinds, 1988). The dust produced in the planer mill settles on the lumber and is disturbed in later handling. After assessing the nested "Planermill on site" term in simple regression analyses to remove insignificant nested terms, "Planermill on 33 site" was entered in the models in nested form in the following process groups: Chip and Hog, Mill Maintenance, Lumber Yard, and Sorting & Packaging. Since wood is typically kiln-dried just before or just after the planing process, wood condition was nested in the Planing process group and in process groups that follow the planing process. After assessing the nested job location term in simple regression analyses to remove insignificant nested terms, wood condition was entered in the models in nested form within the following process groups: Sorting and Packaging and Lumber Yard. 5.2.2 Fixed-Effects Model The fixed-effects model procedure yielded a fixed-effects model with an R of 0.381 and an adjusted R 2 of 0.363 (Table 8). The predicted values of the fixed-effects model ranged from 0.11 to 5.37 mg/m , with a geometric mean of 0.73 mg/m and geometric standard deviation of 1.96. The model's residual variance was 0.759. A l l Cooks' distances were all less than 0.05 indicating no individual values were influential (Figure 5). •y The magnitude of the partial R for each variable can be used to determine which variables contribute the most to explaining the dataset. Four variables explain the majority of the 9 9 model: coastal versus interior (partial R = 0.128), process group (partial R = 0.107), number of employees (partial R 2 = 0.035), and production per m 2 (partial R 2 = 0.031). The •y model intercept also is a significant contributor (partial R = 0.028). Working in an indoor booth had a partial R 2 of 0.010. Thirteen of the 58 jobs stayed in the model as fixed effects, with partial R 2 ' s between 0.002 and 0.016. The variables nested in process group (job •y location, planermill, department, and wood condition) had partial R 's of between 0.002 and 9 9 0.013. A direct comparison of the partial R 's to the model R is not possible as the sum of 9 9 the partial R 's add up to 0.461, which is slightly higher than the model R , illustrating that the individual effect of a variable is reduced when combined with the other variables. Residual versus predicted value plots showed a non-random pattern: a downward sloping linear pattern in the lower left quadrant (Figure 6). Comparison of the residuals to model 34 variables identified that the pattern was a result of a group of measurements below the limit of detection in the Research-2 Dataset. In this dataset, 81 of 235 (34.4%) of the measurements were below the limit of detection. To explore this further, the observed concentration was categorized into 8 equally distributed categories. The residuals versus predicted value plot was re-plotted with the observed concentration category as a label (Figure 7). The same downward trend was seen for all concentration categories and is an expected pattern. A s the pattern seen was concentration dependent and resulted from a group of samples with the same value, the residual plot was deemed acceptable. Coastal mills were found to have a geometric mean wood dust concentration an average of 1.3 mg/m 3 lower than interior mills. Working in a booth decreased exposures for inside jobs by an average of 0.5 mg/m 3 , but was not significant for outdoor jobs. Exposures were lower in smaller mills, as defined by number of employees: mills with 200-350 employees had average exposures of 0.2 mg/m 3 higher than mills with less than 200 employees, and mills with greater than 350 employees had average exposures of 0.6 mg/m 3 greater than mills with less than 200 employees. As production per m 2 increased, exposures decreased: mills with production per m 2 of less than 0.0125 Mmbf/m 2 had average exposures of 0.8 mg/m 3 higher than mills with greater than 0.050 Mmbf/m 2 ; mills with 0.0125 to 0.025 Mmbf/m 2 had average exposures of 0.6 mg/m 3 higher than the highest group, and mills with 0.025 to 0.050 Mmbf/m had average exposures of 0.08 mg/m higher than the highest group. The coefficients for the 13 jobs in the model were adjustment factors on the job's process group mean, with the magnitude of the coefficient based on the difference between the job's geometric mean and the job's corresponding process group's geometric mean. 35 FIGURE 5 Cook's distances compared to the predicted wood dust concentrations (mg/m , on log scale) from the fixed-effects model. O z O o CD o c (0 .04 .03 .02 .01 co 0.00 H i * 1 ^ Q CO Iv: o o O .01 o o °ifl ^ ° u <g° .6 .8 1 8 10 Predicted Values from Fixed-Effects Model (mg/m3) 36 T A B L E 8 Descriptive data, coefficients, and standard errors for independent variables in the fixed-effects and mixed-effects multiple regression models of wood dust (mg/m3) concentrations (log-transformed, base e). Fixed-Effects ModelA Mixed-Effects Model3 Variable0 ,„ . . Standard Partial Coefficent ^ r 2 „ ,_ . , Standard Coefficient _, Error Intercept -0.633 0.196 0.028 * D -0.622 0.303# Mill Geographic Location: Coastal1 (n=l 130) -1.305 0.089 0.128$D -1.294 0.091$ PG E : Log Processing (n=73) -0.145 0.245 F -0.015 0.381 x Job Location: Indoors2 (n=48) 0.714 0.237 0.006 J 0.670 0.242$ PG: Foremen (n=46) -0.268 0.394 -0.271 0.550 x Job Location: Indoors (n=40) 1.009 0.383 0.005$ 1.011 0.384$ PG: Cleaning/Labouring (n=108) 2.175 0.274 $ 2.037 0.436$ x Wood Condition: Green4 (n=89) -0.887 0.223 0.011} -0.859 0.226$ x Department: Sawmill5 (n=45) 0.346 0.177 0.003* 0.350 0.178# x Department: Maintenance (n=7) 0.600 0.354 0.002* 0.786 0.427* PG: Mill Maintenance (n=99) 0.360 0.251 0.614 0.364* x Planermill on site: Yes 6 (n=75) 0.940 0.210 0.013 J 0.919 0.212$ PG: Sorting and Packaging (n=400) 0.632 0.245 # 0.551 0.335$ x Planer mill on site: Yes (n=357) 0.424 0.150 0.005 J 0.491 0.152# x Wood Condition: Green (n=329) -0.221 0.125 0.002* -PG: Lumber Yard (n=85) 0.384 0.324 0.291 0.395 x Planer mill on site: Yes (n=70) 0.557 0.261 0.003 # 0.625 0.261# PG: Sawmill Sawing (n=202) 1.118 0.189 t 1.209 0.321$ PG: Sawmill Non-sawing (n=183) 1.023 0.183 $ 1.116 0.323# PG: Chip and Hog (n=61) 1.308 0.210 t 0.939 0.349# PG: Planing (n=95) 1.351 0.190 + + 1.352 0.427 PG: Non-mill Maintenance (n=34) 1.405 0.232 $ 1.150 0.374$ PG: Sawfiling (n=80) 1.297 0.193 t 1.276 0.338 PG: Powerhouse (n=15) 1.587 0.281 t 1.604 0.424 PG: Log Yard (n=7) 0.384 0.324 0.576 0.533 PG: Office and Boom (n=28) reference - - reference — Job 8 : Welder (n= 10) 1.439 0.294 0.016$ -- --Job: Carpenter (n=9) 0.736 0.309 0.004 #D . - -Job: Tailsawyer (n=52) 0.523 0.144 0.009$ - -Job: Strip Handler (n=13) 0.480 0.255 0.002* - -Job: Log Deck Operator3 (n=17) 0.469 0.245 0.002* - -Job: Resawyer (n=64) 0.430 0.139 0.006 J - -Job: Trimmer Operator (n=132) 0.426 0.105 0.01 IJ - -Job: Forklift (n=32) 0.382 0.215 0.002 * - -37 Fixed-Effects ModelA Mixed-Effects Model8 Variable0 Coefficient Standard Error Partial R2 Coefficient Standard Error Job: Scowman (n=8) -0.633 0.336 0.002* - -Job: Crane Operator (n=17) -0.707 0.266 0.005 J - -Job: Chemicals (n=4) -0.835 0.449 0.002* - -Job: Pipefitter & Sheet Metal Workers (n=6) -1.047 0.424 0.004 # - -Job: Cat Operator (n=7) -1.102 0.354 0.007 J -- — Production / m 2 (Mmbf/m2) 0.031H < 0.0125 (n=522) 0.710 0.140 i 0.683 0.142$ 0.0125-0.025 (n=474) 0.626 0.141 + + 0.585 0.142$ 0.025-0.050 (n=385) 0.100 0.410 0.085 0.124 > 0.050 (n=135) reference - reference — Number of Employees 0.035H < 200 (n=355) -0.529 0.076 X -0.543 0.079$ 200 - 350 (n=486) -0.310 0.068 % -0.286 0.069$ > 350 (n=675) reference — reference w — Indoor Job in Booth (n=177) -0.380 0.097 0.010$ -0.343 0.105$ A See Appendix A for the complete form of the fixed-effects model form. 8 In the mixed-effects model, job was entered as a random effect. The mixed-effects model incorporates both the coefficients for fixed effect and random effect variables. See Table 9 for the mixed-effects job coefficients. See Appendix B for complete mixed-effects model form. c Bulleted (x) variables were nested within the specified process group. The coefficient for the nested variable applies when the observation meets both the process group criteria and the nested variable's listed criteria. D Statistical significance of predictor variable in model: * p<0.10; # p<0.05; $ p<0.01. E PG = Process Group F Process Group was entered as a categorical variable: the partial R 2 for all process groups = 0.107. G - implies variable not included in model. H Production/m2 and Number of Employees were entered as categorical variables: the given partial R2 refers to the combined contribution of all categories in the variable. The following variables are indicator variables (0/1) with the following coding: 1 Mill geographic location: 1 = coastal, 0 = interior 2 Job Location: 1 = indoors, 0 = interior 3 Job: 1 = specified job, 0 = other job 4 Wood Condition: 1 = green lumber; 0 = mix of green and kiln-dried lumber 5 Department: 1 = specified department, 0 = other departments 6 Planermill on site: 1 = planermill on site, 0 = no planermill 38 FIGURE 6 Residuals compared to the predicted wood dust concentration (mg/m , log-transformed, base e) by data source from the fixed-effects model. 39 FIGURE 7 Residuals compared to the predicted wood dust concentration (mg/m , log-transformed, base e) by categorized observed concentration (log-transformed, base e) from the fixed-effects model. 4 0 5.2.3 Mixed-Effects Model The mixed-effects model was almost identical in predictive variables to the fixed-effects model (Table 8, above). Only one model variable included in the fixed-effects model (wood condition nested within the Sorting & Packaging process group) was not significant in the mixed-effects model. Other than job category as a random effect (Table 9), the mixed-effects model did not add any potential determinants of exposure that were not already identified in the fixed-effects model. Only 5 of the 58 jobs were significant (p<0.1) as random effects; these 5 jobs were also significant in the fixed-effects model. The magnitudes of the effects of the exposure determinants were nearly identical to that of the fixed-effects model, with the exception of process group. Since each job was assigned to one process group, adding job category as a random effect influenced the process group estimate. The greatest change in process group estimate occurred in process groups where a job category was a significant predictor in the fixed-effects model. The predicted values of the mixed-effects model ranged from 0.12 to 5.42 mg/m , with a geometric mean of 0.73 mg/m and geometric standard deviation of 1.92 mg/m . The residual patterns were very similar to the fixed-effects model (Figure 8). The resulting mixed-effects model had a between-job variance of 0.0923 (p<0.005), a within-job variance of 0.763, and total residual of 0.856, indicating that the addition of job category as random effect explained 10.8% of the total residual in the model. When job category was entered as a random effect without entering any fixed effects, the total residual was 1.2247, with job category explaining 19.5% (variance = 0.239) of the residual. To determine the between- and within-process group variances, the mixed-effects model was rerun with the same variables, except that process group was entered as a random effect instead of a fixed effect and that job was not entered into the model. The within-process group variance (0.81) was twice that of the between-process group variance (0.40). 41 FIGURE 8 Residuals compared to the predicted values of wood dust concentration (mg/m3, log-transformed, base e) from the mixed-effects model. TABLE 9 Coefficients and standard errors for job entered as a random effect in the mixed-effects multiple regression model of wood dust (mg/m3) concentrations (log-transformed, base e). JobA Number of Observations Job CoefficientB Standard Error Significant in Fixed Effects Modelc Log Processing Barker Operator 25 -0.138 0.217 Cut-off Sawyer 31 -0.089 0.214 Log Deck Operator 17 0.227 0.223 Yes Sawmill Sawing Chipnsaw Operator 4 -0.030 0.255 Edger Operator 60 -0.179 0.163 Gangsawyer 17 -0.104 0.201 Head Sawyer 17 -0.175 0.201 Quadsaw Operator 10 0.064 0.222 Resawyer 64 0.303 * 0.165 Yes Spotter 30 0.121 0.183 Sawmill Non-sawing Dropsorter 73 -0.051 0.166 Lumber Straightener 12 -0.169 0.215 Tailsawyer 52 0.379 ** 0.173 Yes Timber Deckman 15 -0.176 0.206 Utility 27 -0.014 0.188 Wood Picker 4 0.029 0.255 Planing Tilthoist/Planerman/Planer Feeder 95 0.000 0.304 Chip and Hog Cat Operator 7 -0.340 0.242 Yes Chip Screen Tender / Feeder 38 0.304 0.210 Hog Operator 8 0.166 0.239 Scowman 8 -0.131 0.239 Yes Foremen Department Foremen 46 0.000 0.304 Cleaning / Labour Cleanup 93 0.129 0.245 Janitor 4 -0.069 0.269 Labour 11 -0.060 0.240 Mill Maintenance Carpenter 9 0.248 0.227 Yes Chemical 4 -0.349 0.256 Yes Electrician 20 -0.261 0.200 Millwright 41 -0.142 0.184 Oiler 15 -0.151 0.209 Welder 10 0.653 ** 0.223 Yes 43 Job* Number of Job Standard Observations Coefficient8 Error Significant in Fixed-Effects Model0 Powerhouse Fireman 6 0.177 0.268 Powerhouse Maintenance 9 -0.177 0.268 Non-mill Maintenance Machinist 6 0.098 0.253 Mechanic 22 0.218 0.238 Pipefitter / Sheet Metal Worker 6 -0.316 0.253 Yes Sawfiling Benchman 21 0.127 0.208 Grinderman 17 -0.120 0.213 Sawfiler 13 -0.071 0.220 Sawfitter 29 0.064 0.203 Sorting & Packaging Grader 67 -0.089 0.146 J-bar Sorter Operator/Stacker Operator 53 -0.106 0.152 J-bar Sorter Patrolman 17 -0.177 0.192 Offbearer 75 -0.194 0.146 Package Press Operator 18 -0.002 0.189 Stenciller 5 0.395 0.246 Strip Handler' 13 0.240 0.204 Yes Tallyman 20 -0.283 0.186 Trimmer Operator 132 0.212 0.139 Yes Lumber Yard Car Loader / Blockman 11 0.065 0.221 Carrier Driver 11 -0.146 0.221 Crane Operator 17 -0.439 ** 0.208 Yes Forklift 32 0.323 * 0.189 Yes Kiln Operator 7 0.055 0.237 Truck Driver 7 0.138 0.237 Log Yard Log Loader Operator Office/Boom Office/Watchman/Clerk Boat Operator/Boomman/Slipman 28 0.000 0.001 0.304 0.261 A In the mixed-effects model, job was entered as a random effect. The mixed-effects model incorporates both the coefficients for fixed effect and random effect variables. See table 8 for the coefficients of the fixed effects. See Appendix B for complete mixed-effects model form. B Statistical significance of job as a random effect: * p < 0.10; ** p < 0.05. Unmarked jobs had p-value > 0.1. c Yes = job was a significant (p<0.1) variable in fixed-effects model (Table 8). 44 5.3 Comparison of Fixed-Effects and Mixed-Effects Models The variability in exposures was markedly reduced in the predictive models compared to the model dataset, with geometric standard deviations much narrower (1.92 and 1.96) than the model dataset (2.98) (Table 10). Both empirical models predicted a single exposure (0.06%) above 5 mg/m (maximum exposure of 5.4 mg/m ), compared to the model building dataset where 3.8% of the model dataset were greater than 5 mg/m3, with a maximum exposure of 41 mg/m3. Comparisons of predicted versus observed values yielded strong correlations (Pearson r) of 0.617 for fixed-effects model and 0.619 for mixed-effects model. The unexplained residual in the models were very similar, with residuals of 0.759 and 0.763, for the fixed-effects model and mixed-effects model respectively. The predicted values (transformed to mg/m ) of the fixed-effects model and the mixed-effects model were nearly perfectly correlated (Pearson r = 0.975) (Figure 9). However, one job stood out as the fixed-effects model predicted substantially higher exposures than the mixed-effects model for Welder. The mixed-effects model underestimated the model building dataset's job mean for Welder by approximately 1 mg/m (Figure 10b), while the fixed-effects model predicted an identical job mean to the model building dataset. The job geometric means calculated from the model dataset and from the two predictive models visually appear very similar for most jobs (Figures 10a and 10b). All job means from both predictive models fall within the 95% confidence interval around the model building dataset's job geometric means. The magnitude of the 95% confidence intervals around the predicted job means are of similar magnitude by job for both predictive models, and both predictive models have narrower 95% confidence intervals around the job mean than that of the model building dataset. The fixed-effects model predicted 56 of 58 jobs within ±0.2 mg/m3 of the job geometric means, with 47 jobs within ±0.1 mg/m3. The mixed-effects model predicted 54 jobs within ±0.2 mg/m , with 48 jobs within ±0.1 mg/m . Both models underestimated Foreman (fixed-effects model, bias= -0.76 mg/m3; mixed-effects model, bias = -0.45 mg/m ) and Stenciller (fixed-effects model, bias= -0.56 mg/m ; mixed-effects model, 45 bias = -0.40 mg/m3). The mixed-effects model also underestimated Welder (bias = -0.99 mg/m ) and Carpenter (bias = -0.30 mg/m ). TABLE 10 Mean, geometric standard deviation, range, model residual and correlation of the fixed-effects and mixed-effects models compared to the model building dataset. Model Dataset Fixed-Effects Model Mixed-Effects Model Geometric Mean Wood Dust ^ ^  Q 73 Concentration (mg/m3) 0.73 Geometric Standard Deviation 2.98 1.96 1.92 Range of Wood Dust Concentrations (mg/m3) 0.03-43.1 0.12-5.37 0.12-5.42 Model residual (unexplained variance) 0.759 0.763 Pearson Correlation (r) of predicted versus observed 0.617 0.619 values (log-transformed). FIGURE 9 Predicted wood dust concentrations (mg/m3) of the fixed-effects model compared to the predicted wood dust concentrations (mg/m3) of the mixed-effects model. D 0) XJ r\ w w 5 • 0 it • 0 • LU XJ 4 • CD X LL • • • c 3-O CO • • jncent to Predicted Cc Predicted Cc O Welder Predicted Cc D All other jobs c ) 1 2 3 4 5 e Predicted Concentration — Mixed-Effects Model 46 o TJ O E 0) sz E O <D re c £ $ V O o ® " TJ S E u £ a ™ *! I I re **-CO J C TJ re g E «= £ «g a) re E re o §>!• o | -5 3 re n o UJ O 0 TJ O "S LU « TJ iS CD CO >< Q LL • • O c/> o CD LU "D CU X •4 -4—1 -4—1 -4—I Ml r H anoavi dDNVaiO Aini±n CO CD CO c 'c ro D_ M3AMVS1IV1 cd3NlH9IVyiS yaiyosdoaa yanods uaA/vwsau do Mvsavno U3AMVS QV3H yax/vwsDNVo do yaoaa d o » o a a s o n U S A M V S ddoino do U 3 » y v a yaaaaa yaddiHO E 5 ro CZ) co I 2 3 i/MOoa Q . !c O £ o o CD o CN CNJ CO CD T f (gLU/Bui) U B 9 | / \ | 3M)3UI089 O 00 CD 1^" CM o T—• O O CD O O 47 TJ O E E _ . O a) J= TJ «2 | o  C £ i2 © o l£ § i 5 1 = I O ^ o « s> » cn TJ - £ £ i2 TJ ° « ^ CO i C TJ ro a 0) X E *= E w ® 5 E co o XJ s? ol ~» 3 £ - ° O 11 oi o 3 3 (O -A TJ ro >< Q LL ^ I B f H h 1-h*4 I—•-I—« -— i i — * i — • -o io o in o in co CN c\i o do U3i/\ii/\iiyi NVWA11V1 y a i a N V H diais do yaxovis dO SS3Hd 33V>OVd u a u v a a d d o i / \noyivd a a i a o s yva-r l d r a a o d dO 3NVU0 usAiaa aaiuuvo Nvi/wocna OINVHOaiftl yai i idMvs yaiidMvs NvwyaaNiao NviftiHONaa y a a i a M a a n o mom/vrmi/v NVI0iai03"13 ysiNSdavo Q-t r o CO o o -a ro > JD E 3 H i 1 ro CO 0 o c ro a) c 'ro 2 (£W/6lU) UB9|/ \ | 3 U ) 3 L U 0 3 0 48 5.4 External Validation of Models The predictive models developed using the model dataset were validated using a cross-validation technique (Hornung et al, 1991) against the Research-3 Dataset that was reserved for external validation. The models' coefficients were used to calculate predicted values for the validation dataset, such that the models' bias and precision could be estimated. Both predictive models underestimated the mean exposure levels in the validation dataset by 0.53 mg/m 3 (both predictive models, G M =1.12 mg/m 3; validation data, G M = 1.65 mg/m 3). (Table 11). The mixed-effects model was marginally less biased (-0.383 versus -0.390) and was slightly more precise (0.896 versus 0.915) than the fixed-effects model. The geometric standard deviation of bias (a measure of the variability in the bias where larger values represent greater variability) was 2.50 and 2.45 for the fixed-effects model and mixed-effects model, respectively. The predicted values for the validation dataset were only moderately correlated with the measured values: Pearson r = 0.394 and 0.420 (on log-transformed concentration) for the fixed-effects and mixed-effects models, respectively. While overall the predictive models underestimated the validation dataset's exposures, the magnitude of the underestimation varied on a job-by-job basis (Table 12, Figure 11). The predictive models underestimated outdoor jobs by 0.8 mg/m and indoor jobs by 0.3 mg/m . Both models estimated the same 5 jobs (of 18 jobs with a minimum of 5 measurements) within ± 0 . 2 mg/m of the validation dataset job geometric means and both models underestimated the same 6 jobs by over 1.5 mg/m . The precision of the predicted estimates was poor, with geometric standard deviations on the bias ranging from 1.27 to 9.98 by job, indicating large variability in the bias estimates. The two models performed similarly for predicting exposures on a job-by-job basis, with the predicted values (transformed to mg/m 3) from the two models nearly perfectly correlated (Pearson r = 0.985). The geometric job means of the two predicted models were withiri 0.1 49 mg/m 3 of each other for 16 of the 18 jobs (with a minimum of 5 measurements). The models' predicted means of the other two jobs, Cat Operator and Janitor, were different by 0.3 mg/m 3 . TABLE 1 1 Predicted dataset geometric mean, bias, precision and correlation of the fixed-effects model and the mixed-effects model compared to the validation dataset. Fixed-Effects Model Mixed-Effects Model Predicted Geometric Mean (mg/m3) 1.12 1.12 BiasA -0.390 -0.383 Equivalent Bias in mg/m3 B -0.53 -0.53 Precision*3 0.915 0.896 Geometric standard deviation of bias0 2.50 2.45 Pearson Correlation (r) of predicted versus observed values (log-transformed). 0.394 0.420 A Bias=mean of residuals based on log-transformed values (see Section 3.4.4 for equation); units=ln(mg/m3). B Equivalent Bias in mg/m3 = exp[Ln(geometric mean) + Bias] - Geometric Mean c Precision=standard deviation of residuals of log-transformed values (see Section 3.4.4); units=ln(mg/m3). D Geometric standard deviation of bias = exp(precision). 50 I ! far I cj I Q 2 <« CQ o S O •s"s 03 6JD| 5 5 03 OQ S 2 CJ M c o 09 g CQ o -J o i-s 3 CN rt CN " rt rt rt rt CN c^  CN " ON rt CO rt <->") c o ^ c > < ^ c N r n i n c N ' ' - > c j \ r ~ i n o o o ^ © - ' t f , r ~ r - -d d d d d d d d d d d d d c N d r t d r t o oo CN d o d o rt rt O CN <N C N co vo ON ^ O N > 0 ON CN —< o oo O O O O O O rt f M O r t - r t - r t T f - ^ O N O O O v N O s - ^ - o m t ^ i - ^ d d d d d d o d d d q ' r t d r t O r t ' r t ^ ; o v r - o t N r t r t c N o o c N C N O o o m r ^ c o r ^ rt rt (Nl (N rt cs O ^ - C S t ^ 0 0 T t \ 0 i o > ^ l t> — cn rt ro m ' ^ t O s O r t c < - l r ~ o o ' o O N t - ~ i n o o o v O ' ^ - r ~ t ~ -o o c N ^ v o ^ ^ c N ^ t ^ r ^ c n r ^ N q c o i o c N v o c o d d d d d d d d d d d d d c N d r t ' d r t r- r- r i oo o o o C N C N t-~ <*-> O O O vo m m cs rt o d d d d d P P P P P P P rt rt r- o o o q C N q iri C N C N rtCN00ONVot^\000\0^tv0>or»i>riCN'!t>^>rt ^ C N r t O O p P — t " " ! ^ . ^ r i ° ^ , < ^ 0 ° r r 5 o . 1 d d d d d q q q p p p " T ' P ' V P r 7 r T ' T c^  rt oo oo o CQ co CO CO CO CO ! - i 1~ W M. O o o o o s O O O O % TD TD TD TD C S3 C C => o o T3 o o TD O O TJ o o o o T3 TD a a rt rt H H r t r t r t Q Q r t Q Q Q r t Q r t r t 2 o to fc o 1 £ b S S3 CO ^ B -1 Pa-rt w u CD o 2 * « 5 o CO U PH O CD J A C O cn 5 rt Q CQ TO •*-; t-H 03 <D V -O § ^ O CM o to CD CL O CU T 3 CO o hJ 60 O —I o t/5 © IT) 00 o IT, ci cs ro d co i n <N CU CO CS CO s > "So J3 II c o CD a CO CD 2 8 > I C ^ •P CD SP 6 c o TD CU CO CO CO 3 T 3 CQ CQ < m 51 CD "D O ^ 0 <D TJ CO O TO ~> W W Q O O •4= LU ~ CD 05 > LL • • CO 4—' o £ LLI T J CD X t dO d3l/\ll/\lldl U3IONVH diyis dO U3>10V±S dO SS3Ud 39V*0Vd N i^ny id uva-r u s a v a o d3iyosdOda dO d 3 9 a 3 dO AAVSNdlHO U3NV"ld dO N i l * ldl~l>IHOd dO d 3 a v o i 901 U3AAAVS d d o i n o do d3>iava y o i i N v r dDNV310 dOlVO C D O O N C D i n t C O C M ' - O (eui/Biii) uea|A| oujeuioao 52 6. Discussion 6.1 Interpretation of the Model Parameters in the Fixed-Effects and Mixed-Effects Models Both the fixed-effects and mixed-effects models were similar in form, with only one variable (wood condition in sorting and packaging jobs) included in the fixed-effects model that was not significant in the mixed-effects model. M i l l location (coastal/interior), number of employees, production per m 2 , and process group were the major contributors in explaining the models, as judged by the partial R ' s . Enclosure, individual jobs, job location (indoors/outdoors), wood condition, department, and presence of planermill each contributed less than half the above variables' influence. The major differences between predictive models were the estimates of the effect of each process group. Because the mixed-effects model entered all jobs as a random effect and provided an adjustment factor for each job, the process group coefficients were substantially different than in the fixed-effects model; however, the resulting predicted values from the two models were nearly perfectly correlated (Pearson r=0.975). In the fixed-effects model, only jobs with geometric means that were significantly different than the process group means (13 jobs) stayed in the model, reducing the number of variables and reducing the degrees of freedom in the model. Only 5 jobs were significant as random effects in the mixed-effects model. Exposures are expected to vary by mill, dependent on many mill and time period specific factors, such as the mill location, physical size of the mills, the mill's ventilation systems, the amount of lumber milled, the type and number of saws used, the level of the mill's technology, and other factors. Because not all of these factors were available for modeling and since the available factors tended to be highly correlated with each other, the variables "mill location", "number of employees" and "production per m " were used to capture the mill-specific differences. Both models predicted interior mills to be substantially higher exposed than coastal mills; the magnitude of the difference between coastal and interior was just over 3 times the difference 53 found by Hall et al (2001). Hall et al (2001) took theoretical shift production into account, but did not take number of employees or the mill size, so these additional explanatory variables may account for the difference in magnitude. The models predicted that mills and time periods with more employees had higher exposures. The number of employees was highly correlated with mill size but not production levels or production per m 2 ; consequently, the variable 'number of employees' might account for mill size, while also measuring the degree of automation, as fewer employees are needed when the mill is more technologically advanced. Production per m was expected to provide a measure of mill density. In both models, an increase in production per m 2 resulted in a decrease of exposures. At first look, this would seem counter-intuitive, but the results suggest that mills/time periods with high production levels per m 2 may be more efficient, having more automated systems that allow its workers to operate the machinery from a distance. Lower production levels per m 2 might suggest that older, dustier technology is in use, thereby increasing exposures. While production levels per m and the number of employees were not correlated in this dataset, both variables may be attempting to explain some of the same mill differences, such as size and technology, hence their differing directions of effect. The use of process group in the models allowed similarly exposed jobs to be grouped, but at the same time complicated the use of other variables, forcing variables such as job location (indoors/outdoors and mill department), whether a planermill was on site, and wood condition to be nested within specific process groups. Nesting the variables was necessary because the variables were highly correlated with specific process groups or because the variable was not relevant to all process groups, requiring a priori decisions on nesting. For instance, wood condition and the presence of a planer mill were expected to affect jobs that received lumber downstream from the kiln or the planermill, respectively, so these variables were de facto nested only in the applicable process groups. While nesting the variables complicated the interpretation of the coefficients, it allowed variables that would otherwise have fallen out of the model to stay in the model to provide additional explanatory power. 54 Hall et al (2001) and Teschke et al (1999a) both found that indoor jobs were more highly exposed. The two models predicted log processing jobs and department foremen as having higher exposures i f indoors. For cleaning and labouring jobs, the worker's assigned location influenced exposures, as workers assigned to the maintenance department and to the sawmill were predicted to be more highly exposed than those assigned to yard areas. The effects of indoors or assigned department was not seen in other process groups, as job location was considered as one aspect of defining the process groups. Exposure to green wood only, with no kiln-dried lumber exposure, decreased exposures for workers involved in cleaning, labouring, sorting, and packaging jobs. Teschke et al (1999a) found that planing kiln-dried wood was a highly exposed job task; however, in this model the wood condition in the planermill was not found to be significant, although the planing jobs were highly exposed regardless of wood condition. In the models, mills with a planermill had increased exposures for mill maintenance, sorting, packaging, and lumber yard jobs. Because planing wood results in high dust levels and finer dust (Hinds, 1988), it follows that the jobs that received the planed lumber were more highly exposed. When a planermill is on site, mill maintenance workers would be dividing their work time between the planermill and sawmill, rather than just the sawmill if no planermill was on site. As the planermill is dustier than the sawmill, the mill maintenance workers' average exposures will increase. 6.2 Predictive Ability of the Fixed-Effects and Mixed-Effects Models Compared to Model Building Dataset Both the fixed-effects and mixed-effects models explained 36% of the model building dataset variability. Other models developed for retrospective exposure assessment have explained their dataset's variability in the range of 36-43% (Hall et al, 2001; Burstyn et al, 2000), while another explained as much as 85% of the dataset's variability (Hornung et al, 1994). 55 The 5% improvement in predictive ability of the mixed-effects model, with job set as the random effect, over the fixed-effects model found by Hall (1999) was not observed between the two predictive models developed in this study as the two models were very similar. The two models had almost identical correlations between predicted and observed values (Pearson's correlation, r=0.617 and 0.619 for the fixed-effects and mixed-effects models, respectively) and their predicted values were almost perfectly correlated (Pearson's correlation, r=0.975). The mixed-effects model provided only a negligible improvement in internal precision (mixed-effects model: 0.857; fixed-effects model: 0.859). The unexplained residual was just slightly higher in the mixed-effects model (0.763) than the fixed-effects model (0.759). The job level bias in the fixed-effects model for the model building dataset was small, with 56 jobs within ± 0 . 2 mg/m 3 . The mixed-effects model * 3 performed somewhat poorer than the fixed-effects model, with 54 jobs within ± 0 . 2 mg/m . The similarity of the two models' predictive ability indicates that setting job as a random effect explained no more additional variability in the data than setting job as a fixed effect. Grouping jobs by process group appears to have reduced the influence of job in both models, such that only a few jobs (13 of 58 were significant in the fixed-effects model) that had mean exposures significantly different than their assigned process group's mean exposure provided any additional explanation of variability in the dataset. This may account for why the mixed-effects model provided no improvement in predictability over the fixed-effects model. The mixed-effects model did, however, provide valuable information regarding the within-job, between-job, within-process group, and between-process group variance components. Between-job differences explained an additional 11% of the variability in the dataset. When process group was entered as a random effect in the mixed-effects model, between-process group differences explained 33% of the variability. These variance components suggest that both between- and within-process group differences and within-job differences were significant, but that between-job differences were small when process group was in the model. The use of worker as a random effect has been employed in several models to explain variability in exposures that is not accounted for by the fixed effects by accounting for the 56 within- and between-worker variability (Burnstyn et al, 2000; Teschke et al, 1999a; Nylander-French et al, 1999; Rappaport etal, 1999). Calculating the between-and within-worker variance components was not possible for this dataset due to insufficient repeat measures on individual workers since the available measurements were collected for both compliance and research purposes and encompassed multiple mills and time periods. 6.3 Predictive Ability of the Fixed-Effects and Mixed-Effects Models Compared to the Validation Dataset The fixed-effects model and the mixed-effects model underestimated the validation mill's geometric mean exposure level by 0.5 mg/m 3 . The individual predicted and measured values (log-transformed) for the validation dataset were only moderately correlated (Pearson correlation coefficient, r = 0.39-0.42) compared to the stronger correlation seen between the individual predicted and measured values within the model building dataset (Pearson correlation coefficient, r = 0.62). The two models underestimated outdoor jobs by an average of 0.8 mg/m 3 and indoor jobs by an average of 0.3 mg/m 3 . A l l outdoor jobs were underestimated by a minimum of 0.25 mg/m 3 suggesting that the models poorly predict factors influencing outdoor exposures. The validation mill was a particularly dusty mill with an unpaved yard that was sampled during the summer. A n unpaved yard, which was not considered as an exposure determinant, is likely a considerable source of outdoor dust. The magnitude of the bias also varied on a job-by-job basis as job level biases ranged from +0.4 to -4.8 mg/m 3 , with 5 (of 18) jobs predicted within ± 0 . 2 mg/m 3 of the validation dataset job means and 6 jobs underestimated by over 1.5 mg/m . Janitors, underestimated by over 4 mg/m 3 , are one group of workers difficult to group, as their job tasks were quite variable from mill to mill. The models also poorly estimated planing and some sawing jobs. Strip handlers at the validation mill worked in an open doorway of the planermill. Their proximity to both yard dust and planermill dust may account for some of their additional exposures unexplained by the models. The models' poor predictive ability with respect to several jobs 57 suggest that there may be significant differences in job tasks between mills that have not been captured by the job groupings or by model parameters, such as the job's proximity to specific saws, time spent performing specific tasks such as using compressed air for cleaning, or the percentage of booth enclosure. Indeed, the high within-job variance in the mixed-effects model suggest that this problem is not solely related to the validation dataset, but that the predictive models are unable to account for within-job exposure determinants. Since within-job exposures typically have a high degree of day-to-day variability (Rappaport et al, 1993), the precision is expected to be poor. Within the model building dataset, the geometric standard deviation (GSD) of the bias (exp(precision)) was 2.36 for both the mixed-effects and fixed-effects models. The precision in the external dataset was slightly less precise than within the model building dataset, with GSD's of bias of 2.50 and 2.45 for the fixed-effects model and mixed-effects model, respectively, suggesting that the within-job variability is similar between the model building and validation datasets. Few external validations of predictive models exist, particularly for models developed for retrospective exposure assessment. Burstyn et al (2001) used a similar validation procedure using exposure measurements from other countries (bitumen fume, n=98; benzo(a)pyrene, n=339) to assess the predictive ability of two predictive models developed for retrospective exposure assessment for the bitumen paving industry. Burstyn et al (2001) found that his models underestimated an external dataset by 51% and 70%, for the bitumen fume and benzo(a)pyrene models, respectively, and that the predicted and observed values were only weakly correlated (Pearson correlation coefficient, r = 0.28). Burstyn et al (2001) also found poor precision due to large day-to-day variability, although direct comparisons of precision are difficult as the magnitude of precision is dependent on the unit of exposure and its magnitude as it is calculated as the standard deviation of bias. A statistical model for predicting ethylene oxide developed for use in a retrospective exposure assessment was also validated by a cross-validation procedure using measurements (n=46) reserved from 6 of the 20 facilities studied (Hornung et al, 1994). Hornung et al found that the predictive model underestimated the validation dataset exposures with a bias 58 of 1.1 ppm (compared to the arithmetic mean, 3.50 ppm, of the reserved dataset) and a precision of 3.7 ppm. Using a non-randomiy selected validation dataset, such as was conducted here and in the two studies mentioned above, is a more severe test of a model's predictive ability and its generalizability beyond the model dataset (Burstyn & Teschke, 1999). The overall bias of the models developed here was approximately -32% (bias in mg/m 3 divided by validation dataset geometric mean), on the same order of magnitude (-31%) as found by Hornung et al, (1994), and significantly less than the -51% and -70% biases found by Burstyn et al (2001). 6.4 Limitations of the Predictive Models The predictive ability of the models were limited by the use of non-specific dust sampling methods to estimate wood dust exposure, by the use of conversion factors to transpose the various sampling methods used for the measurements to inhalable dust equivalents, and by amalgamating several datasets that were collected for different sampling purposes, i.e. compliance or research. The exposures predicted by the models are inhalable non-specific dust exposures and not purely wood dust exposures due to the non-specific particulate sampling methods used. Only exposure determinants expected to influence wood dust exposure were considered for inclusion in the models, so the models were limited in their ability to explain variability in the dataset that was contributed by other dust sources, such as vehicle exhaust, metal filings, and welding particulates. To use these models in a retrospective exposure assessment based on exposures to inhalable wood dust, the contribution of non-wood dust sources to the combined dust exposure would have to be accounted for prior to using the predicted exposures from the models. One approach for accounting for non-wood dust sources was taken by Demers et al (2000), who categorized jobs by their dust sources and then applied weighting factors derived from ratios comparing resin acid levels to dust concentration to adjust for non-wood dust sources. 59 Using sampling method conversion factors has been suggested by some to be a feasible way to include more dust measurements to increase the modeling power (Heederik & Attfield, 2000; Gao et al, 2000; Burstyn, et al, 2000; Martin & Zalk, 1998; Tsai et al, 1995; Hinds, 1988); however, consistency in the underlying particulate size distribution and the homogeneity of the particulate shape are critical assumptions for their use. The inter-sampler ratios (ratios of measurements by two different sampling trains) applied to the sawmill dust measurements were obtained from two sources specific to wood dust environments: a side-by-side comparison of sampling methods collected at four of the cohort mills (Davies et al, 1999) and a comparison of 37mm open- and closed-face cassette samplers in a wood dust environment (Beaulieu, 1980). These ratios were similar in magnitude to other inter-sampler ratios calculated by other researchers for wood dust and other dust environments (Martin & Zalk, 1998; Liden et al, 1995; Hinds, 1988). However, the inter-sampler ratios were found to be highly variable, in particular at low dust concentrations. These ratios may depend on type of saw, ventilation systems, and other factors that could result in variability in the ratios both between and within mills, so the use of a single ratio for each sampling method may be a source of exposure misclassification in the predictive model. As the aim of the predictive model was to calculate the average exposure of each job, using an averaged inter-sampler ratio is reasonable considering the naturally occurring high within-job exposure variability. Another approach to adjusting for inter-sampler differences would be to use the sampling method as an independent variable in the model, such as used by Burstyn et al (2000). This approach was not considered feasible for the Sawmill Cohort models as sampling method was highly correlated with mill and time period specific variables, such as number of employees and mill size. Also, the datasets were unbalanced with regard to year, job, source of dataset, and sampling method, so any estimate of effect of the sampling methods would be difficult to separate from these factors or the mill/time period specific variables. By using these inter-sampler ratios, over 1700 measurements collected at fourteen B . C . sawmills were available for use in model development and validation, rather than only 411 measurements collected at four mills with the GSP sampler. Modeling only the 411 GSP 60 sampled measurements would likely result in a more predictive model with respect to the model building dataset. However, the resulting model would lack the variability with respect to time, mill, and job specific variables that was seen in the larger dataset, thereby reducing the model's ability to be generalizable beyond the four mills to all fourteen cohort sawmills. The datasets of B . C . sawmill wood dust measurement data used in developing the models were collected for various purposes, including retrospective exposure assessment (Research-1, Research-2), cross-sectional respiratory health epidemiological studies (Research-3, subset of Compliance-2), determinants of exposure studies (Research-1, Research-2, Research-3), and for assessing compliance with regulations (Compliance-1, bulk of Compliance-2). In a study comparing epidemiological and compliance sampling strategies, Hall et al (2001) concluded that while compliance sampling strategies focus on the more highly exposed jobs, they do not result in an overestimation of mean exposure levels at the job level. The absence of a difference in mean exposure and measurement distribution between compliance and epidemiological sampling strategies suggest that regulatory exposure databases may be used in exposure assessment in epidemiology when job is considered. The challenge with developing predictive models for retrospective exposure assessment is the time frame involved, often including time periods before any hygiene samplings were available. By also including multiple plants in the studies, the challenge becomes accounting for mill, job, and time differences. Seldom are measurements or exposure determinants available for all jobs, mills, and time periods in a study, as was the case in this study. To avoid exposure misclassification, only variables that could be determined with a high level of confidence in their accuracy were chosen. While several mill level exposure determinants were available and could be determined with confidence, few job level factors were available. Only whether the job occurred in a booth and whether the job was indoors or outdoors could be determined for the entire range of the cohort. Other within-job exposure determinants that had been collected for some of the data sources, such as proximity to saws or percentage of time in booth, were not collected by all sources and the missing gaps were not possible to extrapolate with any degree of confidence, 61 so these within-job variables were excluded due to the high likelihood of misclassification. Misclassifying exposure determinants in the modeling process would attenuate the effect of the entered variable and reduce the significance of that factor in the model. The use of variables in a model that are not available or interpretable for the entire study's scope considerably limits the usefulness of the model for exposure assessment. 6.5 Conclusions The results of this study indicate that multiple sources of dust measurements and exposure determinants can be combined to develop a predictive model for use in retrospective exposure assessment. The predictive ability of the models were 36% and external validation of the models demonstrated the models could be expanded to another B.C. sawmill with an overall bias of 0.5 mg/m . Several jobs in the validation mill were predicted within the range of normal day-to-day variability, but a few jobs, specifically outdoor jobs, were significantly underestimated, suggesting that the models are only generalizable to mills of similar size, level of technology, and building/yard conditions. Improving the model's predictability would require a large number of measurements that use the same sampling method over a wide time span and the availability of more within-day job-specific exposure determinants. For retrospective exposure assessment for health outcomes with long lag times, the aim is to predict long-term average exposures, not daily exposures. So improving predictability of the models with respect to within-job parameters is likely unnecessary. The use of mixed-effects modeling provided no improvement in explaining either the model building dataset or the validation set over the more traditional modeling of fixed effects. Given the similar predictive ability of the two models, the fixed-effects model is a more flexible model for assigning exposures for the purpose of retrospective exposure assessment. The mixed-effects model only predicts exposures for specific jobs, while the fixed-effects model predicts exposures based upon process group, with modifying factors for jobs that may vary from that group estimate. In the context of a retrospective study, there are often jobs 62 where no air sampling measurements are available, but whose task characteristics are known; consequently, the fixed-effects model provides greater flexibility in assigning exposures. While the two models were very similar in predictive ability with respect to the datasets used in this project, the mixed-effects model may be the more appropriate model in other situations, such as when the mixed model provides greater precision and predictive ability by accounting for variability within job and/or within worker. 63 References American Conference of Governmental Industrial Hygienists (ACGIH). 2001 T L V s ® and B E I s ® . Cincinnati: A C G I H . 2001. Beaulieu H.J . , A . V . Fidino, K . L . B . Arlington, and R . M . Buchan. A comparison of aerosol sampling techniques: "open" versus "closed-face" filter cassettes. A I H A J . 41(10):758-765. 1980. Big Book - The Buyer's & Seller's Directory of the Forest Products Industry. Eugene, OR: Random Lengths Publications. 1992, 1996. British Columbia Ministry of Forests. Major Primary Timber Processing Facilities in B . C . Victoria, B . C . : Timber Management Branch. 1984 - 1996. Burstyn I., P. Boffetta, G . A . Burr, et al. Validity of Empirical Models of Exposure in Asphalt Paving. Occupational and Environmental Medicine. (Submitted) Burstyn I., H . Kromhout, T. Kauppinen, et al. Statistical Modelling of the Determinants of Historical Exposure to Bitumen and Polycyclic Aromatic Hydrocarbons Among Paving Workers. Annals of Occupational Hygiene. 44(l):43-56. 2000. Burstyn I., and K . Teschke. Studying the determinants of exposure: a review of methods. American Industrial Hygiene Association Journal. 60(l):57-72. 1999. Davies H .W. , K . Teschke, and P.A. Demers. A Field Comparison of Inhalable and Thoracic Size Selective Sampling Techniques. Annals of Occupational Hygiene. 43(6):381-392. 1999. Dement J . M . , R . L . Harris, M . J . Symonds, and C M . Shy. Exposures and Mortality Among ' Chrysotile Asbestos Workers. Part I: Exposure Estimates. American Journal of Industrial Medicine. 4:399-419. 1983. Demers P.A. , K . Teschke, H . W . Davies, et al. Exposure to dust, resin acids, and monoterpenes in softwood lumber mills. American Industrial Hygiene Association Journal. 61(4):521-28. 2000. Demers P.A. and P. Boffetta. Cancer Risk from Occupational Exposure to Wood Dust: A Pooled Analysis of Epidemiological Studies. International Agency for Research on Cancer World Health Organization. I A R C Technical Report No. 30. Lyon, 1998. Directory of the Forest Products Industry. Portland, OR: M . Freeman. 1970 - 1999. Douwes J. , D . McLean, E . van der Maarl, et al. Worker Exposures to Airborne Dust, Endotoxin and p(l,3)-Glucan in Two New Zealand Sawmills. American Journal of Industrial Medicine. 38:426-30. 2000. Duchaine C , a. Meriaux, P.S. Thorne, and Y . Cormier. Assessment of Particulates and Bioaerosols in Eastern Canadian Sawmills. American Industrial Hygiene Association Journal. 61:727-32. 2000. Eisen E . , D. Smith, T. Wegmen, et al. Estimation of long-term dust exposures in the Vermont Granite Sheds. American Industrial Hygiene Association Journal. 45:89-94. 1984. 64 Fisher L . D . , G . van Belle. Biostatistics: A Methodology for the Health Sciences. 1st Ed. New York (NY): John Wiley and Sons. 1993. p. 277-278. Gao P., B .T . Chen, F.J. Hearl, et al. Estimating Factors to Convert Chinese "Total Dust" Measurements to A C G I H Respirable Concentrations in Metal Mines and Pottery Industries. Annals of Occupational Hygiene. 44(4):251-57. 2000. Goldberg M . , H . Kromhout, P. Guenel. Job Exposure Matrices in Industry. International Journal of Epidemiology. 22(6)Suppl.2:S10-S15. 1993. Hall, A . , K . Teschke, H . W . Davies, et al Exposure levels and determinants of softwood dust exposures in B . C . Lumber Mills, 1981 - 1997. American Industrial Hygiene Association Journal. (Submitted 2001) Hall, A . Broad Determinants of Softwood Dust Exposures: Investigation with a Mixed Effects Empirical Model. Master of Science Thesis. Wageningen Agricultural University. 1999. Heederik D. , and M . Attfield. Characterization of Dust Exposure for the Study of Chronic Occupational Lung Disease: A Comparison of Different Exposure Assessment Strategies. American Journal of Epidemiology. 151(10):982-990. 2000. Hinds W . C . Basis for Particle Size-Selective Sampling for Wood Dust. Applied Industrial Hygiene. 3(3):67-72. 1988. Hornung R.W. Statistical Evaluation of Exposure Assessment Strategies. Applied Occupational & Environmental Hygiene. 6(6):516-20. 1991. Hornung R.W. , A . L . Greife, L . T . Stayner, et al. Statistical Model for Prediction of Retrospective Exposure to Ethylene Oxide in an Occupational Mortality Study. American Journal of Medicine. 25:825-36. 1994. I A R C Working Group. Wood dust and formaldehyde (IARC monographs on the evaluation of the carcinogenic risk of chemicals to humans, vol 62). Lyon: Internal Agency for Research on Cancer, 1995. Kauppinen T. , and T. Partanen. Use of plant- and period-specific job-exposure matrices in studies on occupational cancer. Scandinavian Journal of Work Environment and Health. 14:161-67. 1988. Kauppinen T.P. Assessment of exposure in occupational epidemiology. Scandinavian Journal of Work & Environmental Health. 20(special issue): 19-29. 1994. Kauppinen T.P. Development of a Classification Strategy of Exposure for Industry-Based Studies. Applied Occupational and Environmental Hygiene. 6(6):482-87. 1991. Kauppinen T.P. , T.J . Partanen, S.G. Hernberg, et al. Chemical exposures and respiratory cancer among Finnish woodworkers. British Journal of Industrial Medicine. 50:143-48. 1993. Kromhout H . , P. Swuste, and J. Boleij. Empiral modeling of chemical exposure in the rubber manufacturing industry. Annals of Occupational Hygiene. 38(l):3-22. 1994. Liden G . , B. Melin, A . Lidblom, et al. Personal Sampling in Parallel with Open-Face Filter Cassettes and I O M Samplers for Inhalable Dust - Implications for Occupational 65 Exposure Limits. Applied Occupational and Environmental Hygiene. 15(3):263-76. 2000. Madison's Canadian Lumber Directory. Vancouver, B . C . : Madison. 1967 - 1996. Martin J.R., and D . M . Zalk. Comparison of Total Dust/Inhalable Dust Sampling Methods for the Evaluation of Airborne Wood Dust. Applied Occupational Environmental Hygiene. 13(3): 177-82. 1998. Nylander-French L . A . , L . L . Kupper, and S .M. Rappaport. A n Investigation of Factors Contributing to Styrene and Styrene-7,8-oxide Exposures in the Reinforced-Plastics Industry. Annals of Occupational Hygiene. 43(2):99-109. 1999. Random Lengths Buyer's & Seller's Guide. Eugene, OR: Random Lengths Publications. 1987. Rappaport S .M. Selection of the Measures of Exposure for Epidemiology Studies. Applied Occupational & Environmental Hygiene. 6(6):448-57. 1991. Rappaport S .M. , H . Kromhout, and E . Symanski. Variation of Exposure Between Workers in Homogeneous Exposure Groups. American Industrial Hygiene Association Journal. 54(11):654-662. 1993. Rappaport S .M. , M . Weaver, D. Taylor, L . Kupper, and P. Susi. Application of Mixed Models to Assess Exposures Monitored by Construction Workers During Hot Processes. Annals of Occupational Hygiene. 43(7):457-69. 1999. Seixas N.S. and H . Checkoway. Exposure assessment in industry specific retrospective occupational epidemiology studies. Occupational and Environmental Medicine. 52:625-33. 1995. Seixas N.S. and L . Sheppard. Maximizing accuracy and precision using individual and grouped exposure assessments. Scandinavian Journal of Work & Environmental Health. 22:94-101. 1996. Smith T.J . , S.K. Hammond, M . Hallock, and S.R. Woskie. Exposure Assessment for Epidemiology: Characteristics of Exposure. Applied Occupational & Environmental Hygiene. 6(6):441-47. 1991. Stewart P.A. and R.F. Herrick. Issues in Performing Retrospective Exposure Assessment. Applied Occupational & Environmental Hygiene. 6(6):421-27. 1991. Stewart P.A. , and M . Dosemeci. A Bibliography for Occupational Exposure Assessment for Epidemiologic Studies. American Industrial Hygiene Association Journal. 55(12):1178-87. 1994. Stewart P. A . , J .N. Zey, R. Hornung, et al. Exposure Assessment for a Study of Workers Exposed to Acrylonitrile. III. Evaluation of Exposure Assessment Methods. Applied Occupational & Environmental Hygiene. 11(11):1312-1321. 1996(a). Stewart P. A . , P.S.J. Lees, and M . Francis. Quantification of historical exposures in occupational cohort studies. Scandinavian Journal of Work & Environmental Health. 22:405-14. 1996(b). 66 Teschke K . , A. Ostry, C . Hertzman, et al. Opportunities for a Broader Understanding of Work and Health: Multiple Uses of an Occupational Cohort Database. Revue Canadienne de Sante Publique. 89(2):132-136. 1998. Teschke K . , C . Hertzman, and B. Morrison. Level and Distribution of Employee Exposures to Total and Respirable Wood Dust in Two Canadian Sawmills. American Industrial Hygiene Association Journal. 55(3):245-50. 1994. Teschke K . , P. A . Demers, H . W . Davies, et al. Determinants of Exposure to Inhalable Particulate, Wood Dust, Resin Acids, and Monoterpenes in a Lumber M i l l Environment. Annals of Occupational Hygiene. 43(4):247-55. 1999(a). Teschke K . , S.A. Marion, T . L . Vaughan, et al. Exposures to Wood Dust in U.S. Industries and Occupations, 1979 to 1997. American Journal of Industrial Medicine. 35:581-89. 1999(b). Tielemans E . , L . L . Kupper, H . Kromhout, et al. Individual-based and Group-based Occupational Exposure Assessment: Some Equations to Evaluate Different Strategies. Annals of Occupational Hygiene. 42(2): 115-119. 1998. Tsai P.J., J .H . Vincent, G . Wahl, and G . Maldonado. Occupational exposure to inhalable and total aerosol in the primary nickel production industry. Occupational and Environmental Medicine. 52:793-99. 1995. Werner M . A . and M . D . Attfield. Effect of Different Grouping Strategies in Developing Estimates of Personal Exposures: Specificity Versus Precision. Applied Occupational and Environmental Hygiene. 15(l):21-25. 2000. Vedal, S., M . Chan-Yeung, D . Enarson, et al. Symptoms and pulmonary function in western red cedar workers related to duration of employment and dust exposure. Archives of Environmental Health. 41: 179-83. 1986. Y u R . C . , W . - Y . Tan, R . M . Mathew, et al. A Deterministic Mathematical Model for Quantitative Estimation of Historical Exposure. American Industrial Hygiene Association Journal. 51 (4): 194-201. 1990. 67 cu T3 O CO u <u i t 111 • -a X o E o u _ I < X T3 C CU OL Q . < m A CO CU CU >, o "ft s w <rt o S H CD X> fi 3 o in o o A fi rt o 3 -a o o o CD O NH o ft 3 O S H O CD O o k. PH IT _o ' C CD rt c c o o o rtl a , 3 O In a CD O c CD In CD al a o e CD O a o U Is c CD CD a c O O O T 3 c o "is cj o x> o o PH d AH, ON" O O + o o • o ,5 c o o o rtl X> o 60 c o u T3 O O £ '60 s 'SP « PM <% 60 C ti o O w CN CN © I c? CD CD OH 60 O o o T 3 e o T 3 O O 'So _ c "C 3 O x> Cd CD O c ed C CD 2 .fi CD fi ft CD Q 60 a "C 3 o ed . t c cd CD O O o + * H TS fi I c CD fi •c cd a. CD Q ^60 _C "C 3 o •8 *io S .fi -fi CD a a AH, PH rt 00 r-~ oo o CD s o CD w /—. 60 e *6D ed o ed PH 60 C o C J c>o d d PH PH. T * CN CD •a ° o C ed X SI CD >H I « § -2 <D C •le. - H PH CD TS + O I m m o d < = > PH AH + O T a -rn O) o © + + CO CD . ^ •5 rt?^ a a.rt CD ft i °-S © .fi ^ .s is v c2 CD 3 rt X3 .3 13 CO C C ed .3. > «^ 1 O <C c O CD CD <D o x : •S CD fc 0 0 Crt .2 rt CD ed" 3 O ° i h rt fi o C ed '—* CD \© CD CN fa — « O o .5 o ed r - . » U ^ fe C* ^ (30 d • S ° S s c« u co cd CD CD o o x i — c g Id 2 ^ cd ^ co cd CD ^ s •c fi « C H CD O i -CD rt _C CD •3 o rt O CD = CD > o X3 rt * CD •a a CD - H rt i -J CO * r t CD rt 3 H rt > -c X 3 cd o > cd . W w rt > CD JS ft ft rt c _CD c5 -> CD ft O ft u rt 68 cu 3 T3 a> u o o o o OL. Q_ co < CO co T J o CO •+-» o a> T3 X o E o Ll_ I CO >< T J tz co a . C L < o m co A co CD u "a. S w C+H o !_ CL) S X5 a o m o © A 1 o 3 X I O i -PH o o CQ cu o o cx 3 O CD o o CD c c o o o cx 3 O s-o CD O e CU L> ,<D U-t CD C O CD a c o U "St —1 '3 °» SO + ^ 2 CN oo a ° ^ ^ + O O > PH OH, O m CO • o o CN CD CD >-. o W C+H O S-c CD XI oo CN o o CN V IA CD CD >> O XS -*-» o o PQ o o e a, s S § 0 5 ^ H o O CN » 0 <N OS —' SO CN O d —H d CD cn d I rt o o —1 43 O <—> CO •ST CO O O -a c c o •{§ o o H J X! o a CD a CD L -O (JH d PH + o o X S e c o ••§ a o _) o 60 c CD O o i-PH 60 o —1 O o ' s£> d + c CD CD i-o a o X ) c o O xt o o CD o c CD + H g '3 c CD a csS CX CD Q "ec .g "C 3 O x> rt aS CD u 6 PH, VO OO + 6 I </> 4H e CD a cx CD Q 60 60 3 O XI rt _ l s as IT) 00 d I o rt c rt CD 05 CD C o CD 60 g '5b a rt PH *J 60 C ' £ o o PH as CO  o PH PH, O IT) CO d + CD >H S JZ G . 3 ^ 3 'c ^ O C • 2 CD g s S PH § X PH -a 'CD >H 0 u S w 1 i 2 CN .. ^ > O d + XI c« H CD CD C/3 X> O i—> XS o es W VH ,o c CD o C+H CD o 6 = P-1, + H ow a o x) + CO O o c co CD O CD . •S x>^ r a n. CD Q . i es * ^ ° - = CD .s 5 .a ^- -5 c2 CD 3 S .2 CD C wi c es X) H CD • • c C I S O CD CD O X! cj •<-• CD $ CD b S « M .2 S U CD es 2 U - —^» 2 >-a o s o c CD es — CD i ,2 C ( N fc — « o o .5 o « -<D <0 ^ fa. b H r? O ^ 60 d • B ° S 5 CO ^ co cS CD a> o o xi — 3 CD x: .a «s CD CO 1-CD X I 3 ^ rt 0 •S E j i c3 j - n > !s o n. i .2 o CD CS c T= 53 XJ 'o .9 CD . 3 > co o X) rt cu • ' •o a >> •2 j a .2 cx J3 . rt CD , - H — 3 LL .2 J > 5S . <4H CD CD ._: o eS J> > O 6 0) o o PH X3 o es w x>a CX CL cx^ e s m 69 CO a -a < u *H CD CD CD CD O u Ji • c a a U o CM GC a a fi i ) 13 co CZ3 < i—) u CD 4=1 CD -*-» CO *^  5 CO ,—1 CD CO 43 ca O co 43 4> T J rt ~ 'S rt > g O o | fi m "3 ^ E3 co C+H 'to 42 43 -4—* c3 !td co a u "Lo I * •H co r t c u co rt <D - g - 8? & rt rt o * "C o 73 oo 6 H cd o 70 

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:
https://iiif.library.ubc.ca/presentation/dsp.831.1-0090072/manifest

Comment

Related Items