Elsevier

Chemosphere

Volume 60, Issue 7, August 2005, Pages 898-906
Chemosphere

Handling of dioxin measurement data in the presence of non-detectable values: Overview of available methods and their application in the Seveso chloracne study

https://doi.org/10.1016/j.chemosphere.2005.01.055Get rights and content

Abstract

Exposure measurements of concentrations that are non-detectable or near the detection limit (DL) are common in environmental research. Proper statistical treatment of non-detects is critical to avoid bias and unnecessary loss of information. In the present work, we present an overview of possible statistical strategies for handling non-detectable values, including deletion, simple substitution, distributional methods, and distribution-based imputation. Simple substitution methods (e.g., substituting 0, DL/2, DL/√2, or DL for the non-detects) are the most commonly applied, even though the EPA Guidance for Data Quality Assessment discouraged their use when the percentage of non-detects is >15%. Distribution-based multiple imputation methods, also known as robust or “fill-in” procedures, may produce dependable results even when 50–70% of the observations are non-detects and can be performed using commonly available statistical software. Any statistical analysis can be conducted on the imputed datasets. Results properly reflect the presence of non-detectable values and produce valid statistical inference. We describe the use of distribution-based multiple imputation in a recent investigation conducted on subjects from the Seveso population exposed to 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD), in which 55.6% of plasma TCDD measurements were non-detects. We suggest that distribution-based multiple imputation be the preferred method to analyze environmental data when substantial proportions of observations are non-detects.

Introduction

Environmental research frequently relies on measurements of chemical, physical or biological agents performed to evaluate low-level contamination of air, soil, water or food, and to quantify the exposure of wildlife and human individuals. In spite of extensive efforts to develop high-sensitivity assays, often a substantial proportion of samples have such low concentrations to border on the detection limit (DL) defined by the sampling and analytical methods. Uncertainty deriving from levels that are non-detectable may impair the capability of drawing conclusions functional to regulatory decision making (Currie, 2000). Dioxins, which may pose a threat to human health and the environment even at very low concentrations, often challenge investigators with exposure measurements including high proportions of non-detects (Currie, 2000, Singh and Nocerino, 2002).

Also in recent environmental investigations, percentages of non-detectable levels in environmental and biological samples have often been large, as presence of more than 40% of non-detects for at least one of the analytes investigated has been far from being a rare occurrence (Acquavella et al., 2004, Barra et al., 2004, Berkowitz et al., 2004, Caserini et al., 2004, Kato et al., 2004, Liu and Mou, 2004, Quandt et al., 2004, Roots et al., 2004, Silva et al., 2004, Sinkkonen et al., 2004, Toro et al., 2004). However, in spite of intense debate and extensive theoretical research activity on the topic, environmental research has often tolerated the loss of information and potential bias arising from improper or inadequate treatment of non-detects and rarely taken advantage of available statistical techniques to limit these problems.

In the present work, we discuss possible strategies for handling exposure data including non-detects. We recommend the use of a multiple imputation method based on distribution-based estimation of non-detectable values. Performances of this method have been previously assessed through data simulation studies (Helsel, 1990, Huybrechts et al., 2002, Lubin et al., 2004). We show its application in estimating mean levels of dioxin in subjects sampled from the Seveso population exposed to 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD). In this recent study, 55.6% of the subjects had non-detectable plasma TCDD levels.

Section snippets

Non-detects: what they are and how to handle them

Measurements including non-detects are “left-censored data”. A left-censored data sampling distribution is one for which the only information in some of the samples is that the true measurement is less than a given censoring value (Hornung and Reed, 1990). The censoring value, i.e., the DL, may be constant for all the observations in the dataset (singly censored data) or may vary between observations (multiply censored data). There have been considerable differences of opinion about how to

Study background

In 1976, the Seveso accident exposed a large residential population to TCDD, the most toxic dioxin congener. The exposure produced a large outbreak of chloracne, mostly among children (Baccarelli et al., in press). Chloracne is a skin intoxication similar in appearance to acne vulgaris, but characterized by pale-yellow keratin cysts and larger and prominent comedones. After the accident, the area was divided in four zones of decreasing contamination: zone A, where TCDD soil concentration was

Conclusions

Correct handling of non-detectable values is critical in environmental research, particularly when the range of analytes in the study is close to DL, as often occurs for persistent organic pollutants and dioxins. In our application on chloracne, the multiple-imputation means were considered as the best obtainable estimates (Huybrechts et al., 2002) and set as reference in the comparison with the other procedures, whose relative bias varied between 22.8% and 329.6%. Independently of the

Acknowledgment

We are indebted to Jay Lubin, Ph.D. for the original inspiration of the present work, advice, and critical reading of the manuscript.

References (32)

  • J.F. Acquavella et al.

    Glyphosate biomonitoring for farmers and their families: results from the Farm Family Exposure Study

    Environ. Health Perspect.

    (2004)
  • N.H. Adams

    Quality assurance reviews: how they differ from peer reviews

    Qual. Assur.

    (1998)
  • Baccarelli, A., Pesatori, A.C., Consonni, D., Mocarelli, P., Patterson, D., Caporaso, N.E., et al., in press. Health...
  • G.S. Berkowitz et al.

    In utero pesticide exposure, maternal paraoxonase activity, and head circumference

    Environ. Health Perspect.

    (2004)
  • L.A. Currie

    Detection and quantification capabilities and the evaluation of low-level data: some international perspectives and continuing challenges

    J. Radioanal. Nucl. Chem.

    (2000)
  • EPA, U.S., 2000. Guidance for data quality assessment. Practical methods for data analysis. Office of Environmental...
  • Cited by (0)

    View full text