Determinants of dust exposure in sawmills : a comparison of fixed-effects and mixed-effects predictive statistical models

• Main Author: Friesen, Melissa Charmaine
• Format: Others
• Language: English
• Published: 2009
• Online Access: http://hdl.handle.net/2429/11706

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; mixedeffects model: Pearson r = 0.619). The fixed-effects model predicted 56 of 58 jobs within ± 0.2 mg/m3 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/m³ . On average, outdoor jobs were underestimated by 0.8 mg/m³ and indoor jobs by 0.3 mg/m³ . 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. === Medicine, Faculty of === Population and Public Health (SPPH), School of === Graduate