Specimen Allocation in Longitudinal Biomarker Studies: Controlling Subject-Specific Effects by Design

Abstract

It is important to understand specimen allocation factors that may impact the validity and reliability of results in longitudinal studies examining within-person changes in biomarker levels. Using data from a randomized clinical trial of an exercise intervention in 136 postmenopausal women, we determined the effect of assaying the baseline and follow-up samples of some subjects in different batches on the intervention effect estimates for serum concentrations of estrone, estradiol, testosterone, androstenedione, and dehydroepiandrosterone. Twenty-five subjects had their baseline and 3-month follow-up samples and 50 subjects had their baseline and 12-month samples assayed in different batches; all other subjects had their baseline, 3-month, and 12-month samples assayed in the same batch. Subjects with split samples were reassayed with all samples in the same batch. We compared the estimated regression coefficient for the intervention effect using the split sample data with one estimated excluding the split sample data and one estimated replacing the split sample data with the reassayed data. The median percentage difference in the intervention effect estimate was 59.6% between using versus excluding the split sample data and 74.6% between using the split sample versus using the reassayed data. In general, the coefficients from the model including the split sample data were closer to zero and statistically less significant than those from the models excluding the split sample data or using the reassayed data. These results suggest that bias can be artificially introduced into intervention effect estimates of longitudinal studies if samples from a subject are not assayed in the same batch.