Using Exposure Data to Identify Priority Substances Under the European Water Framework Directive: The Quest to Reflect Uncertainties

The European Water Framework Directive (European Commission 2000) identifies “priority substances” as those posing a significant risk to or via the aquatic environment. These substances are then subject to development of an environmental quality standard (EQS) and potential reduction of discharges and emissions. A set of new candidate priority substances is under consideration by the European Commission, with selection criteria including a risk assessment based on regulatory monitoring datasets of measured concentrations in fresh and marine waters. Summary statistics from these exposure data (primarily the 95th percentile) are then compared with the EQS to assess the presence of a potential European Union-wide risk. If such a risk is confirmed, the substances are then selected for a legally binding EQS. Both reliable exposure data and a reliable predicted no-effect concentration (PNEC) are necessary for a robust risk assessment. Unfortunately, the exposure datasets used to evaluate candidate priority substances contain a large proportion of undetected or unquantified samples (Table 1). This highlights the left-censored characteristics (i.e., having a high proportion of unquantified results) of many of these datasets. Without appropriate consideration of these nondetects, the risk assessment becomes meaningless, because it does not provide objective evidence that exposure occurs above the relevant PNEC over wide areas or at a significant number of sites. Importantly, these exposure datasets cannot be scrutinized by Member State experts or stakeholders, because they are not publicly available. The approach taken by the European Commission to deal with the “less than” values in these datasets is to assume that each of them is one-half the limit of detection (LOD). This substitution approach reflects a recommendation in a European Directive (European Commission 2009) for the calculation of mean values from compliance datasets. However, the Directive was not intended for datasets with high levels of censorship, such as those shown in Table 1. The uncertainty introduced by this substitution approach is never discussed, and the summary statistics and percentiles from these datasets are presented in dossiers without associated measures of uncertainty. It has long been established that highly left-censored datasets require specific statistical treatment to deliver robust summary statistics before they can be used in regulatory risk assessment. The US Environmental Protection Agency (2006) provides guidance for practitioners managing environmental data to assist in robustly estimating summary statistics, with the preferred method linked to the degree of censorship in the dataset. In general, it is accepted that when the censoring rate is higher than 70 to 80%, irrespective of method used, statistical estimates are not reliable (Gardner 2011). We are concerned that use of half the LOD can lead to biased estimation of environmental exposure when used in conjunction with highly censored datasets. The extent of the bias depends on the LOD value, how it is defined, and the proportion of the monitoring data falling below it. The implications of this bias increase when the LOD is at or above the PNEC value. In addition, for highly censored datasets, the bias is worse because the 95th percentile is likely underestimated, and the mean is potentially overestimated. Especially under these circumstances, the substitution method has been demonstrated to be ineffective compared with other approaches based on firmer statistical principles (e.g., maximum likelihood estimation) and lacks a robust scientific basis (Antweiler and Taylor 2008). Recommendations and guidelines for better practice in the handling of “less than” values in environmental monitoring data are readily available (e.g., Helsel 2006; Gardner 2011). It is critical to interpreting the outcomes of the risk assessment to provide a clear statement of how “less than” values have been dealt with and to provide a detailed narrative on the type and size of the errors that are likely to have been introduced by the treatment used. Any summary statistic used in the reporting of monitoring data must be accompanied by a measure of its associated variability and uncertainty because such summary statistics should not be treated as absolute values. For the datasets summarized by the European Commission, this could be achieved through application of a resampling method such as bootstrapping, to examine known variability around summary statistics (e.g., approximately a 95th percentile) and the effect of uncertainty (e.g., ignorance of the true location of a value below the LOD). In an environmental risk assessment, exposure data represent half of the input for the risk characterization, yet arguably receive a great deal less attention and scientific scrutiny compared with the hazard data that drive the PNEC (and EQS). The processing of the exposure data in evaluating candidate priority substances is highly relevant: European countries will need to routinely measure all confirmed priority substances across their waterbodies, expending time and resources on monitoring a chemical substance that may or may not be adversely affecting ecosystems. Consequently, it is important that the substances identified as priority substances are not only potential problems at a small number of specific local sites (where they should still, of course, be controlled) but also represent a truly widespread risk that requires appropriate European Union-wide management. There is a real danger that the wrong priority substances will be selected due to inappropriate data handling practices and the subsequent arrival at spurious conclusions, leading to wasted time and resources without delivering any environmental benefit. The authors acknowledge that they work for industry, regulators, and governments, but have not received funding for the preparation of this manuscript. Associated data are not publicly available. For questions please contact the corresponding author ([email protected]).