Assessing assay variability for field samples in environmental research is challenging, since a quantitative assay is typically constrained by a lower limit of detection. The purpose of this paper is to compare three parametric models for assessing assay variability using duplicate data subject to heavy left-censoring. Efron information criterion (EIC) and Bayesian information criterion (BIC) are used to aid in model selections. Distributional parameter estimates are obtained using maximum likelihood estimation for bivariate lognormal, bivariate zero-inflated lognormal, and bivariate 3-component mixture models. We illustrate a practical application using duplicate pesticide data from the Community Participatory Approach to Measuring Farmworker Pesticide Exposure (PACE3) study. Furthermore, a simulation study is conducted to empirically evaluate the performance of the three models. The results from PACE3 indicate that the bivariate zero-inflated lognormal model is fairly competitive based on EIC or BIC. Further, total variability for the lognormal component can be decomposed into between-subject and within-subject variance based on this model. Assay variability estimates such as within-subject coefficient variation, minimum detectable change, and probability of
$$k$$
-fold difference can be easily derived under the bivariate zero-inflated lognormal model. Additionally, the assay variability is rather large for the PACE3 data. Therefore, apparent longitudinal change in pesticide exposure should be examined cautiously in the context of substantial assay variability.

Supplementary materials accompanying this paper appear online.