Statistical procedures for addressing nondetect results in environmental analyses

Censor data sets, those containing observations that are less than the analytical chemistry limit of detection (LOD) value, are becoming more common as regulatory agencies and dischargers strive to reduce concentrations of trace organic contaminants in treated effluent. The recently reissued Technical Support Document For Water quality Based Toxics Control,'' a frequently used reference during NPDES permit development, states that the mean and variance of censored data sets may be determined by assuming the data conform to what EPA has termed the delta log-normal distribution'' (D-LOG). The US EPA D-LOG statistical procedure along with two commonly used alternatives (Cohen's MLE and regression of normal order scores) are described, illustrated with example calculations and evaluated using Monte Carlo simulation experiments. Using minimal absolute bias as the evaluation criteria, Cohen's method had superior predictive capability overall, particularly for real world data sets. For mean and standard deviation estimators, Cohen's method was at least 2.5 and 1.2 times less biased than those of the EPA recommended methodology, respectively.