Robust Statistical Approach to Analysis of Data Containing Non-detects

Oral Presentation

Prepared by B. Nott1, D. Helsel2
1 - Electric Power Research Institute, 3412 Hillview Avenue, Palo Alto, CA, 94304, United States
2 - Practical Stats, 2838 Mashie Circle, Castle Rock, CO, 80109, United States


Contact Information: bnott@epri.com; 650-855-7946


ABSTRACT

Improvements in analytical methodology and instrumentation will result in enhanced capabilities to measure concentrations of chemicals at lower and lower levels. Regulatory limits for many chemicals can be expected to become increasingly stringent, driven by health/risk considerations. The confluence of these factors results in many situations where measurements are made at or below reporting levels. The resulting data sets contain measurements above reporting levels as well as some non-detect values, i.e., concentrations below reporting levels. Simplistic approaches (such as setting non-detects to zero, ½ of the detection level) continue to be used to conduct statistical analyses of datasets containing non-detects. Many studies have shown that these approaches distort the real underlying data distribution, resulting in inaccurate conclusions regarding the concentration levels of chemicals of interest in the stream being monitored. The authors have investigated alternative, statistically sound and less arbitrary, approaches to make better sense of the data and to draw meaningful conclusions. The authors will present application of the statistical approaches to subsets of data on measurements of selected Hazardous Air Pollutants (HAPs) in electric utility power plant emissions collected as part of United States Environmental Protection Agency (USEPA) Information Collection Request (ICR) for the recent Mercury and Air Toxics Standards (MATS) rule.