Background: For chemicals with high within-subject temporal variability, assessing exposure biomarkers in a spot biospecimen poorly estimates average levels over long periods. The objective is to characterize the ability of within-subject pooling of biospecimens to reduce bias due to exposure misclassification when within-subject variability in biomarker concentrations is high.
Methods: We considered chemicals with intraclass correlation coefficients of 0.6 and 0.2. In a simulation study, we hypothesized that the chemical urinary concentrations averaged over a given time period were associated with a health outcome and estimated the bias of studies assessing exposure that collected 1 to 50 random biospecimens per subject. We assumed a classical type error. We studied associations using a within-subject pooling approach and two measurement error models (simulation extrapolation and regression calibration), the latter requiring the assay of more than one biospecimen per subject.
Results: For both continuous and binary outcomes, using one sample led to attenuation bias of 40% and 80% for compounds with intraclass correlation coefficients of 0.6 and 0.2, respectively. For a compound with an intraclass correlation coefficient of 0.6, the numbers of biospecimens required to limit bias to less than 10% were 6, 2, and 2 biospecimens with the pooling, simulation extrapolation, and regression calibration methods (these values were, respectively, 35, 8, and 2 for a compound with an intraclass correlation coefficient of 0.2). Compared with pooling, these methods did not improve power.
Conclusion: Within-subject pooling limits attenuation bias without increasing assay costs. Simulation extrapolation and regression calibration further limit bias, compared with the pooling approach, but increase assay costs.