LAUSR

This method was developed to demonstrate the application of probabilistic bias analysis to quantify and adjust for exposure misclassification in a historical cohort mortality study of New Zealand trichlorophenol workers where exposure to 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) was measured as a multi-level variable. Published exposure information available for this cohort of workers was used to specify the initial bias parameter distributions, which were then varied under 18 different scenarios to assess the potential impact of differing amounts of misclassification as well as both non-differential and differential exposure misclassification. For each scenario, each bias parameter distribution was sampled 50 000 times using Monte Carlo simulation techniques to generate adjusted counts of cases and non-cases of ischemic heart disease (IHD) by exposure group. These counts were then used to calculate odds ratios adjusted for exposure misclassification and the associated exposure misclassification error terms. Given the specified assumptions, the geometric mean (GM) adjusted odds ratio had a range of 2.89 to 5.05, and the GM error term ranged from 0.60 to 1.05. In all non-differential scenarios and scenarios in which non-cases had greater proportions of misclassification, the observed odds ratio of 3.05 was closer to the null (i.e., 1) than the GM adjusted odds ratio. For the differential simulations where cases had higher proportions of misclassification, the direction of the error was dependent on the level of misclassification error, with smaller proportions of misclassification resulting in the observed odds ratio being farther away from the null than the GM adjusted odds ratio. These findings demonstrate that probabilistic bias analysis of historical cohort mortality studies can be an effective tool for understanding trends in study error stemming from exposure misclassification and confirm the importance of quantifying potential sources of systematic error.