Modeling repeated measurement data for occupational exposure assessment and epidemiology

Abstract

Repeated measurements designs, occur frequently in the assessment of exposure to toxic chemicals. This thesis deals with the possibilities of using mixed effects models for occupational exposure assessment and in the analysis of exposure response relationships. The model enables simultaneous estimation of both the variance components of exposure (between- and within-subject) and the unbiased regression coefficients for determinants of exposure. Implications are relevant to grouping strategies, hazard control , risk assessment and over-exposure assessment. In an Israeli cohort of industry workers sampled over a year, the geometric standard deviations representing variation of exposure between workers (after adjustment for agent and factory) and within workers were 3.1 and 3.0, respectively. These values may be used, as crude estimates of exposure variability to obtain an interval estimate of mean exposure. In studies among Dutch rubber manifactoring workers exposed to inhalable particulate and rubber fumes and pig farmers exposed to bacterial endotoxins, exposure determinants reduced the random between-worker variance estimators between 59-100%. Interestingly, the random within-worker variability was reduced only in the pig farmer data set, by 25%, when specific work activities that varied over time were accounted for. Results of linear regression and mixed models were compared. In rubber manufacturing, coefficients were similar, but fewer factors affecting exposure were statistically significant due to the high correlation between repeated measurements. In a cohort of benzene-workers, both time related factors and a non-time related factor (e.g. job task) were found to affect the mean exposure significantly. The random between workers variance was highly affected by job task. Time related factors (warm month, “pay day”, day of the week), were found to be responsible for the high random within-worker variance, which was more than two times higher than the between-worker variance. Grouping strategies in occupational health should result in a small between-worker variance. The within-worker variance often varies greatly. Consequently, in simulated data based on real data we found that it is common to obtain a zero or negative ANOVA estimate of the between-worker variance. We evaluated an approach proposed earlier to use an upper confidence bound when the estimate is negative, and found that this method has three main disadvantages: the estimator can remain negative , performs poorly with two repeated measures per worker, and the method can be extremely sensitive to small changes in the data. Our alternative estimator incorporates "plugging in" of an estimator in which the observed mean squares replaces the expected values and this offers a solution to these problems. Exposure assessment plays an important role in a valid exposure-response evaluation in epidemiology. In a study among Dutch bakers mixed modeling was used in two procedures: firstly to estimate exposure based on specific exposure detreminants, and secondly for exposure-response relationship where the estimated variance components were used as a scaling factor to avoid possible exposure-respons attenuation. The shape of the relationship between sensitization and exposure, was found to be a quadratic function.