Methodology 2
Wednesday 8 October 2025, 13.30 – 14.30 Polar
Chairs: Neil Pearce, Wenxin Wan
Assessing the effectiveness of automatically assigning occupational exposure modules in a multi-center hospital-based case-control study in Asia
Melissa Friesen (presenter)
Sarah J Locke, Pabitra Josse, Calvin Ge, Jun Xu, Tai Hing Lam, Yok Lam Kwong, Bryan Bassig, Wei Hu, Lin Fritschi, Troy Sadkowsky, Mark P. Purdue, Raymond HS Liang, John K.C. Chan, Kexin Chen, Caigang Xu, Yu-Chieh Su, Brian C.H. Chiu, Dean Hosgood, Nat Rothman, Qing Lan, Roel Vermeulen
Abstract
Objective: Exposure-oriented questionnaire modules provide crucial task information for occupational exposure assessment. We evaluated the effectiveness of an algorithm for automatically assigning modules during interviews in a hospital-based case-control study in Asia. Methods: We used a rule-based expert system based on a keyword search of occupational history responses and screening questions pertaining to paints/stains, solvents/glues/degreasers, and engineered woods to assign jobs to one of 23 modules. If no module was identified, we assigned a Work Location module that redirected some participants to more detailed modules. Post interview, each job was coded to standardized occupation and industry classifications. For each job and industry code we determined the ‘ideal’ module. We reviewed each ‘ideal’ vs. ‘assigned’ module combination and characterized potential information loss: none, task loss, and industry loss. We evaluated the screening questions’ positive and negative predictive value (PPV, NPV) compared to task responses for jobs assigned the Solvent module. For those redirected from the Work Location to the Solvent module, we calculated the task prevalence. Results: The algorithm assignment for 26,608 jobs was based on keywords for 55.4% and screening questions for 8.6%; the remainder received the work location module. Potential information loss was identified for 8.8% of jobs (7.2% task loss; 1.6% industry loss). For the 5,847 jobs completing the Solvent module, the overall PPV and NPV of the screening questions was 82.4% and 70.0%, respectively, with higher NPVs for engineered woods and paints/stains than for solvents/glues/degreasers. The Work Location re-directed 663 jobs to the Solvent module; the most frequently-reported activities for these jobs were cleaning hands with solvents (14%), paint bystander (8.0%), cleaning/degreasing bystander (5.1%) and glues/adhesives bystander (5.0%). Conclusions: Overall, our automated approach resulted in excellent capture of tasks of interest. Jobs identified with potential information loss will be prioritized for additional review during exposure assessment efforts.
Bayesian model accounting for exposure variability and uncertainty in inhalation exposure-response models for occupational health
Caroline Groth (presenter)
Sijin Wen; Elizabeth Stuart; Timothy Nurkiewicz; Stella Hines; Mohammed Abbas Virji
Abstract
Objective: Bayesian methods are available to model inhalation exposures accounting for measurements below the LOD or missing completely. Individual or job-group-based exposure estimates obtained from these models are often assigned to individual workers or to job-groups of workers, respectively, enrolled in epidemiologic studies. This strategy, however, does not account for variability in exposures around the group means, or exposure uncertainty stemming from missing or censored data. The objective of this study is to present a new model that combines Bayesian models for exposure with exposure-response models for a standard continuous health outcome, e.g., spirometry parameters, accounting for uncertainty and variability in exposure. Materials and Methods: Statistical simulation studies were set up to resemble standard exposure-response scenarios wherein the true value of all exposure and health outcome parameters are known. The standard approach which does not incorporate measurement error was compared to this new approach using different statistical criteria such as bias (distance between estimate and truth), coverage (percentage of uncertainty intervals contain truth), and posterior predictive accuracy (the model’s ability to predict the data). Results: Compared to the standard approach, the new method had the lowest bias, more optimal coverage, lower uncertainty interval widths, and slightly better posterior predictive accuracy across scenarios. Conclusion: Results suggested that the new approach performed more optimally compared to the standard approach across most statistical criteria. This new approach will enable better understanding of the exposure-response relationships while accounting for exposure variability and uncertainty. Future work will apply this model to a sample of welding workers to understand the relationships between welding fume exposures and spirometry measures. Funding: NIGMS: 5U54GM104942-08.
Impact of interventions to prevent asbestos-related respiratory disease in an exposed worker registry using a simplified G-computation
Nathan DeBono (presenter)
Louis Everest, David B. Richardson, Colin Berriault, Ryann E. Yeo, Maya A. Meeds, Victoria Arrandale, Paul A. Demers
Abstract
Background: The Ontario Asbestos Workers Registry is a regulatory exposure registry obligating employers to report the number of work hours with asbestos containing materials for each of their workers. Currently, each worker is notified of the need for a medical examination once they have accrued 2,000 reported hours of work with asbestos. However, recent research showed a substantial excess of asbestos-related respiratory disease among workers in the registry indicating that the notification policy requires revision. We sought to evaluate the impact on disease prevention of alternative policies limiting asbestos work hours among registry participants using a novel application of parametric G-computation we term ‘G-POSH’. Methods: A cohort of 26,164 asbestos registry workers were followed for cancer and non-malignant disease diagnoses in Ontario from 1986 through 2019. G-POSH was used to estimate decreases in disease risk under five hypothetical exposure reduction interventions. Inverse probability of censoring weighting was used to adjust for time-varying confounding due to healthy worker survivor bias. Results: Standard Poisson regression analyses of the association between cumulative asbestos work hours and respiratory disease incidence rates showed substantially elevated rates well before reaching 2,000 asbestos work hours. Using G-POSH, limiting cumulative asbestos work hours to 100 hours would have reduced the risk of asbestosis by half (Risk Ratio, RR: 0.52, 95% CI 0.43-0.63) and lung cancer by 15% (RR: 0.85, 95% CI 0.78-0.92) compared to the observed natural course in the cohort. Limiting exposure to 2,000 asbestos work hours had a smaller but still substantial impact on prevention of asbestosis (RR 0.77, 95% CI: 0.70-0.86). Inverse probability weighted estimates showed a minor influence of healthy worker bias. Conclusion: G-POSH is a simplified tool for estimating intervention effects in occupational cohorts. Regulatory agencies should intervene sooner to prevent respiratory disease among workers in the registry.
A less detailed job axis in a quantitative job-exposure matrix results in a similar exposure-response association
Hans Kromhout (presenter)
Susan Peters
Abstract
Introduction: Quantitative job-exposure matrices (JEMs) have been developed to assign exposure using International Standard Classification of Occupations (ISCO)-68 coded job information. For extended compatibility with the less detailed ISCO-88 coding, a quantitative JEM using the same underlying model was developed. We compared exposure-response relationships between cumulative respirable crystalline silica (RCS) and lung cancer risk using a quantitative JEM based on ISCO-88 (88-JEM) and ISCO-68 (68-JEM). Methods: Based on a common set of approximately 15 000 RCS measurements, job-specific, region-specific and time-specific exposure levels were estimated for the 88-JEM and the 68-JEM and linked to participants’ job histories. Exposure-response relationships in an international lung cancer case-control study were analysed by logistic regression and generalised additive models. Results: The 88-JEM and the 68-JEM yielded similar RCS-lung cancer associations, with elevated lung cancer risks across each cumulative exposure quartile. The 88-JEM exhibited a minor not statistically significant upward bend in the exposure-response curve at higher exposures. Conclusion: To accurately detect associations between disease risk and occupational exposure, quantitative JEMs can be applied in community-based studies that provide job histories in either ISCO-88 or ISCO-68.