Combining Decision Rules from Classification Tree Models and Expert Assessment to Estimate Occupational Exposure to Diesel Exhaust for a Case-Control Study.

OBJECTIVES: To efficiently and reproducibly assess occupational diesel exhaust exposure in a Spanish case-control study, we examined the utility of applying decision rules that had been extracted from expert estimates and questionnaire response patterns using classification tree (CT) models from a similar US study. METHODS: First, previously extracted CT decision rules were used to obtain initial ordinal (0-3) estimates of the probability, intensity, and frequency of occupational exposure to diesel exhaust for the 10 182 jobs reported in a Spanish case-control study of bladder cancer. Second, two experts reviewed the CT estimates for 350 jobs randomly selected from strata based on each CT rule's agreement with the expert ratings in the original study [agreement rate, from 0 (no agreement) to 1 (perfect agreement)]. Their agreement with each other and with the CT estimates was calculated using weighted kappa (κ w) and guided our choice of jobs for subsequent expert review. Third, an expert review comprised all jobs with lower confidence (low-to-moderate agreement rates or discordant assignments, n = 931) and a subset of jobs with a moderate to high CT probability rating and with moderately high agreement rates (n = 511). Logistic regression was used to examine the likelihood that an expert provided a different estimate than the CT estimate based on the CT rule agreement rates, the CT ordinal rating, and the availability of a module with diesel-related questions. RESULTS: Agreement between estimates made by two experts and between estimates made by each of the experts and the CT estimates was very high for jobs with estimates that were determined by rules with high CT agreement rates (κ w: 0.81-0.90). For jobs with estimates based on rules with lower agreement rates, moderate agreement was observed between the two experts (κ w: 0.42-0.67) and poor-to-moderate agreement was observed between the experts and the CT estimates (κ w: 0.09-0.57). In total, the expert review of 1442 jobs changed 156 probability estimates, 128 intensity estimates, and 614 frequency estimates. The expert was more likely to provide a different estimate when the CT rule agreement rate was <0.8, when the CT ordinal ratings were low to moderate, or when a module with diesel questions was available. CONCLUSIONS: Our reliability assessment provided important insight into where to prioritize additional expert review; as a result, only 14% of the jobs underwent expert review, substantially reducing the exposure assessment burden. Overall, we found that we could efficiently, reproducibly, and reliably apply CT decision rules from one study to assess exposure in another study.

Identifying gender differences in reported occupational information from three US population-based case-control studies.

OBJECTIVES: Growing evidence suggests that gender-blind assessment of exposure may introduce exposure misclassification, but few studies have characterised gender differences across occupations and industries. We pooled control responses to job-specific, industry-specific and exposure-specific questionnaires (modules) that asked detailed questions about work activities from three US population-based case-control studies to examine gender differences in work tasks and their frequencies. METHODS: We calculated the ratio of female-to-male controls that completed each module. For four job modules (assembly worker, machinist, health professional, janitor/cleaner) and for subgroups of jobs that completed those modules, we evaluated gender differences in task prevalence and frequency using χ(2) and Mann-Whitney U tests, respectively. RESULTS: The 1360 female and 2245 male controls reported 6033 and 12 083 jobs, respectively. Gender differences in female:male module completion ratios were observed for 39 of 45 modules completed by ≥20 controls. Gender differences in task prevalence varied in direction and magnitude. For example, female janitors were significantly more likely to polish furniture (79% vs 44%), while male janitors were more likely to strip floors (73% vs 50%). Women usually reported more time spent on tasks than men. For example, the median hours per week spent degreasing for production workers in product manufacturing industries was 6.3 for women and 3.0 for men. CONCLUSIONS: Observed gender differences may reflect actual differences in tasks performed or differences in recall, reporting or perception, all of which contribute to exposure misclassification and impact relative risk estimates. Our findings reinforce the need to capture subject-specific information on work tasks.