Applying Species-sensitivity Distributions in Ecological Risk Assessment: Assumptions of Distribution Type and Sufficient Numbers of Species

Species-sensitivity distribution methods assemble single-species toxicity data to predict hazardous concentrations (HCps) affecting a certain percentage (p) of species in a community. The fit of the lognormal model and required number of individual species values were evaluated with 30 published data sets. The increasingly common assumption that a lognormal model best fits these data was not supported. Fifteen data sets failed a formal test of conformity to a lognormal distribution; other distributions often provided better fit to the data than the lognormal distribution. An alternate bootstrap method provided accurate estimates of HCp without the assumption of a specific distribution. Approximate sample sizes producing HC5 estimates with minimal variance ranged from 15 to 55, and had a median of 30 species-sensitivity values. These sample sizes are higher than those suggested in recent regulatory documents. A bootstrap method is recommended that predicts with 95% confidence the concentration affecting 5% or fewer species. Keywords—Risk assessment Species sensitivity Bootstrap Statistics Sample size INTRODUCTION Concentration–effect data applied to ecological risk assessment usually come from single-species toxicity tests measuring effects to individuals. However, populations, communities, and ecosystems are generally the entities to be protected. To resolve this incongruity between individual-based data and the complex biological entities addressed in ecological risk assessment, an evaluation of species-sensitivity distributions has been proposed [1]. Single-species test data are combined to predict concentrations affecting only a certain percentage of species in a community. Single-species data (e.g., median lethal concentration [LC50] or no-observed-effect concentration [NOEC] values) for many species are fit to a distribution such as the lognormal or log-logistic. From this distribution of species sensitivities, a hazardous concentration (HCp) is identified at which a certain percentage (p) of all species is assumed to be affected. The most conservative form of this approach uses the lower 95% tolerance limit of the estimated percentage to ensure that the specified level of protection is achieved [1–3]. Species-sensitivity distribution or extrapolation methods are being incorporated into assessments of ecological risk [4,5] and into recommendations for pesticide registration [6]. A general species-sensitivity distribution approach is detailed in the new U.S. Environmental Protection Agency guidelines for ecological risk assessment [7]. The variants most recently espoused specify a lognormal distribution for species sensitivities regardless of whether the modeled effect metrics were median effective concentration (EC50), LC50, NOEC, or maximum acceptable toxicant concentration values [1–5]. The species-sensitivity distribution innovation carries as* To whom correspondence may be addressed ([email protected]). Contribution 2223 of the College of William and Mary’s Virginia Institute of Marine Science and School of Marine Science. sumptions needing scrutiny [1]. The following are seven of the most important concerns. 1. The LC50, EC50, NOEC, and maximum acceptable toxicant concentration have very significant deficiencies as measures of effect to field populations and communities [8]. Any secondary metric based on such compromised metrics possesses the same deficiencies. 2. Whether any species loss is acceptable is equivocal. Proponents of the species-sensitivity distribution approach argue that enough redundancy exists in communities to allow some loss [9]. The counterpoint to this redundant species hypothesis combines the rivet popper hypothesis (community integrity is reduced by each loss of a species) with the argument that a conservative stance is best when faced with uncertainty in ecological risk assessment [10]. The rivet popper hypothesis is supported by the few studies directly addressing the question of redundancy [10–14]. Ecosystem productivity, stability, sustainability, and nutrient retention decrease as species diversity decreases, although species loss up to a certain level seemed to have little influence on drought resistance or resilience of prairie grassland communities [12–14]. 3. If performed without thought, species-sensitivity distribution methods could discount the importance of maintaining dominant and keystone species [15], and the influence of species interactions [8]. 4. In situ exposure differs among species because of dissimilarities in behavior, feeding habits, life histories, life stages, and microhabitats. Exposure differences proscribe the direct application of concentration–effect data derived by singlespecies laboratory testing [16] in this or other risk assessment procedures. 5. A bias exists toward mortality data despite the plausibility of nonlethal effects being at least as important in determining local population extinction. Also, a bias exists toBootstrap analysis of species-sensitivity data Environ. Toxicol. Chem. 19, 2000 509 ward a subset of standard species that are amenable to laboratory culture and manipulation. However, these shortcomings are shared by other risk assessment applications of ecotoxicity data. 6. The assumption of a specific distribution may not be justifiable for many data sets, which, if plotted alongside values predicted from the distribution, show clear deviations from the assumed lognormal distribution. Because data for diverse taxonomic groups are pooled, even the basic assumption of a unimodal distribution is dubious. Although the present movement toward using specific taxonomic groups in this speciessensitivity distribution procedure may reduce the magnitude of this problem, the ambiguity associated with selecting the lognormal distribution will remain. 7. Most discussions of appropriate sample size and sample representativeness of the community are focused on practical issues of producing an agreeable number of observations for regulatory agencies [6]. The exception was the original work of Kooijman [1], who introduced the approach of applying parametric methods to formally address the sample size question. Assessment of these issues of sufficient sample size and representativeness is difficult in most applications. Most studies either do not report confidence limits along with estimates of HCp, fail to calculate minimal sample sizes, or do not discuss the completeness of the data set relative to the community at risk. Bootstrap estimation provides a partial answer to two of the above issues, ambiguity in selecting a specific distribution and estimation of the approximate number of species needed to precisely estimate HCp [17]. Bootstrap estimation alleviates the difficulty of selecting an appropriate distribution because calculations do not require an explicit distribution. The only requirement is a random sample of species sensitivities from the universe of possible species sensitivities. The approximate number of species needed to minimize variation around the HCp estimate can also be obtained by utilizing bootstrap methods. In the present study, we examined 30 diverse data sets to test the appropriateness of the now prevailing preference for a lognormal distribution for fitting species-sensitivity data; to illustrate an alternate, bootstrap method that imposes no distributional requirements; and to estimate the approximate number of species-sensitivity values required to precisely estimate HC5, HCC10, and HC20 values. MATERIALS AND METHODS Data sources No-observed-effect-concentration and EC50–LC50 data were assessed because both are analyzed with species-sensitivity distribution methods. The NOEC data (potassium bichromate; sodium bromide; tetrapropylene benzene sulfonate; 2,4-dichloroaniline; p-nitrotoluene; 2,4-dinitro-o-cresol; dimethoate; and pentachlorophenol) were those tabulated in Slooff and Canton [18] with supplemental values from the AQUIRE database [19]. Atrazine EC50–LC50 data were provided by K. Solomon and applied after culling unspecified species from the data set. Diazinon data from Giddings et al. [20] were compiled similarly. Aldrin, copper, chlordane, chromium, dieldrin, endosulfan, and heptachlor data were taken from water-quality criteria documents [21–26] and augmented with the AQUIRE database [19]. Also, data entries for unspecified species were removed from chromium and copper data. Mercury data came from the AQUIRE database except one datum from Khangarot and Ray [27]. Chlorpyrifos data were derived from the AQUIRE database. Methyl parathion, guthion, and malathion data came from the AQUIRE database with expansion from Brandt et al. [28], Cripe [29], Cripe et al. [30], Key et al. [31], Morton et al. [32], and Verschueren [33]. Geometric means were used in analyses if multiple entries for species were found within a data set. Parametric statistics We determined if the lognormal model was generally applicable for fitting species-sensitivity information for each of the 30 data sets (Table 1). After log transformation of effect concentrations, a Shapiro–Wilk’s test was performed to ascertain how often the null hypothesis of a (log) normal distribution was rejected (a 5 0.05) (SASt Version 6.12, Proc Univariate [34]). Pearson x2 statistics were calculated after the 30 data sets were fitted by maximum likelihood methods to lognormal, log-logistic, and Gompertz models (SAS Version 6.12, Proc Probit [34]). Resulting x2 statistics allowed comparison of data goodness-of-fit for these candidate models: the model with the lowest x2 statistic was the best. Plots of regression residuals from the candidate models (lognormal, loglogistic, and Gompertz models fit with Proc Probit) were also examined for systematic deviations from the model. Finally, HC5 values and associated 95% confidence limits were generated with the lognormal model. Bootstrap statistics Alternatives to conventional parametric statistics appeared soon after affordable, powerful computers became widely available [35]. These computationally intensive methods possess several advantages relative to conventional statistics. A relevant advantage in this case is the capacity to generate univariate statistics (i.e., the HC5) and an associated confidence interval without assuming a specific distribution [17]. This method also provides a straightforward way to estimate the number of observations required to minimize the variation about an estimate such as the HC5 [17,36]. Bootstrap estimates of the HC5 and its 95% confidence limits were produced for the 30 data sets with Resampling Stats Version 4 [37]. For each of the 30 sets, the available data were sampled randomly with replacement to create a resampling data set of 100 observations. These observations were ranked from smallest to largest and the value ranked at the fifth percentile was selected as the HC5. Here and elsewhere, linear interpolation between ranked values was used as required in simulations resampling low numbers of observations. The resampling was repeated to produce 10,000 estimates of HC5. These 10,000 estimates were ranked, and the value corresponding to 50% was taken as the best estimate of the HC5. The estimates corresponding to 2.5 and 97.5% were used as the 95% bootstrap confidence limits. The value ranked at 5% estimated the concentration protecting 95% of the species with 95% certainty, that is, only 5% or fewer species will be affected 19 out of 20 times at this concentration. Notice that none of these procedures required a specific distribution. The bootstrap procedure was modified to estimate the number of species-sensitivity values needed to approach the point of minimal variation about a HC5 estimate. The above process was repeated with resampling data sets of size 5 to 100 in increments of 5. The resulting 2.5, 50, and 97.5% values were plotted against sample size. The confidence interval about the HC5 estimate decreased as sample size increased until the first point of minimal improvement was reached. The sample size at that point approximated the number of observations needed 510 Environ. Toxicol. Chem. 19, 2000 M.C. Newman et al. T ab le 1. A na ly si s of 30 da ta se ts fo r es ti m at io n of ha za rd ou s co nc en tr at io n (m g/ L ) af fe ct in g 5% of sp ec ie s (H C 5)