A sensitivity analysis of bias in relative risk estimates due to disinfection by-product exposure misclassication

We conducted a sensitivity analysis of relative risk estimates using local area mean disinfection by-product exposures. We used Monte Carlo simulations to generate data representing 100 towns, each with 100 births (n = 10,000). Each town was assigned a mean total trihalomethane (TTHM) exposure value (mean = 45, SD = 28) based on a variable number of sampling locations (range 2-10). True maternal TTHM exposure was randomly assigned from a lognormal distribution using that town's true mean value. We compared the effect of a 20 mg/1 increase in TTHM exposure on the risk of small-for-gestational age infancy using the true maternal exposure compared to various weighting measures of the town mean exposures. The exposure metrics included: (1) unweighted town mean, ( 2) town mean weighted by the inverse variance of the town mean, (3) town mean weighted by the inverse standard deviation of the town mean, (4) town mean weighted by 1-(standard deviation of sites per town/mean across all towns), and (5) a randomly selected value from one of the sites within the town of residence. To estimate the magnitude of misclassification bias from using the town mean concentrations, we compared the true exposure odds ratios (1.00, 1.20, 1.50, and 2.00) to the mean exposure odds ratios from the five exposure scenarios. Misclassification bias from the use of unweighted town mean exposures ranged from 19 to 39%, increasing in proportion to the size of the true effect estimates. Weighted town mean TTHM exposures were less biased than the unweighted estimates of maternal exposure, with bias ranging from 0 to 23%. The weighted town mean analyses showed that attenuation of the true effect of DBP exposure was diminished when town mean concentrations with large variability were downweighted. We observed a trade-off between bias and precision in the weighted exposure analyses, with the least biased effects estimates having the widest confidence intervals. Effect attenuation due to intrasystem variability was most evident in absolute and relative terms for larger odds ratios.