Environ. Sci. Technol. 2008, 42, 3447–3452
Development of Species Sensitivity Distributions for Wildlife using Interspecies Toxicity Correlation Models JILL A. AWKERMAN,* SANDY RAIMONDO, AND MACE G. BARRON U.S. Environmental Protection Agency, Gulf Ecology Division, 1 Sabine Island Drive, Gulf Breeze, Florida 32561
Received November 13, 2007. Revised manuscript received February 15, 2008. Accepted February 18, 2008.
Species sensitivity distributions (SSD) are probability distributions of chemical toxicity of multiple species and have had limited application in wildlife risk assessment because of relatively small data sets of wildlife toxicity values. Interspecies correlation estimation (ICE) models predict the acute toxicity to untested taxa from known toxicity of a single surrogate species. ICE models were used to predict toxicity values to wildlife species and generate SSDs for 23 chemicals using four avian surrogates. The hazard levels associated with the fifth percentile of the distribution (HD5) were compared for ICE SSDs and independent SSDs created with measured data. SSDs were composed of either avian only or avian and mammalian taxa. ICE HD5s were within 5-fold of 90% of measured HD5s and were generally higher than measured HD5s. The first percentile of the distribution (HD1) and the fifth percentile of the lower confidence limit (HDL) of ICE SSDs produced values that were not significantly different from measured HD5s. Using a bird surrogate to predict toxicity to birds and the Norway rat to predict toxicity to mammals improved some estimates of ICE HD5s compared with those generated using only bird surrogates. These results indicate that ICE models can be used to generate SSDs comparable to those derived from measured wildlife toxicity data and provide robust estimates of the HD5.
Introduction Species sensitivity distributions (SSDs) are probability distribution functions from a sample of toxicity values that can be used to represent variation in chemical tolerance among species and determine prescribed levels of hazard. For example, toxicity values lower than the LD50 of 95 or 99% of wildlife species (HD5 and HD1, respectively) can be determined from the fifth or first percentile of an SSD generated from LD50 values of all available species with data for a specific chemical. SSDs have been widely used to assess risks and to determine protective levels for aquatic species (1) but have had limited application in wildlife risk assessment because of the need for large toxicity data sets that include a diversity of wildlife species. Interspecies correlation estimation (ICE) models are leastsquares regressions of the sensitivity relationship of two species and can be used to extrapolate acute toxicity of a chemical from the known toxicity of one species to another * Corresponding author phone: 850.934.9230; fax 850.934.2402; e-mail:
[email protected]. 10.1021/es702861u
Not subject to U.S. Copyright. Publ. 2008 Am. Chem. Soc.
Published on Web 03/26/2008
species with known uncertainty (2). In practice, a leastsquares model II regression is developed from the known LD50 values of two species tested for the same group of chemicals. The model is then used to approximate toxicity of one species to a chemical from the known toxicity of the other species (the surrogate). ICE models for wildlife have been determined to be most accurate for two species within the same order, but may be used to estimate toxicity from birds to mammals (2). Confidence intervals calculated with ICE estimates are based on the linear relationship and mean model error and may be used to assess the robustness of model predictions. In aquatic species, acute toxicity values generated from ICE models have been used to populate SSDs and have been recommended for generating reasonable estimates of hazard levels when toxicity data are limited (1, 3). The application of ICE to generate SSDs and the accuracy of ICE HD5s for wildlife species has not been investigated. While SSDs are useful in ecological risk assessment and have been used in the development of water quality criteria (4), they are dependent upon available data sets and can differ in distribution, taxonomic diversity, and sample size (5). Recommendations of minimal sample sizes necessary for meaningful estimation have varied with the data set and estimation methodology. With adequate representation of diverse taxa, eight species have been considered a sufficient number of observations (5). Whereas 10-15 species have reduced variance of parameters in some studies (6), others have suggested that up to 55 species may be required to minimize variance (7). In the present study, the effect of sample size on HD5 prediction was examined using random subsets of measured data to generate SSDs of varying sample sizes. Parameters of logistic SSDs of log10-transformed values are frequently determined through maximum likelihood estimation; however, inverse estimation of HD5 independently is an alternative with fewer computational requirements. These two different methods of SSD parameter estimation were also compared to test the assumption that the two methods would produce similar estimates. The HD5s of SSDs derived from measured LD50 values were compared to those of SSDs created with toxicity values predicted from ICE models. Four bird species ((Japanese quail (Coturnix japonica), mallard (Anas platyrhynchos), northern bobwhite (Colinus virginianus), and red-winged blackbird (Agelaius phoeniceus)) were selected as surrogate species for ICE models because of the high number of robust models previously developed for these species (2). HD5s determined from predicted toxicity values using these four surrogate species were compared to test the null hypothesis that they would not differ in their prediction of HD5 values for bird species despite differences in body weight and life history traits. The effect of species composition of SSDs was also examined by comparing HD5s of measured and predicted toxicity values for the same set of species. A second analysis of ICE surrogates compared SSDs generated from the mallard or Japanese quail to SSDs generated using the same avian surrogate to predict bird LD50 values and the Norway rat (Rattus norvegicus) to predict mammal LD50 values. Because ICE model confidence is related to taxonomic distance (2), this analysis further tested the relationship of ICE HD5 accuracy and surrogate species selection. The objectives of this study were to (1) identify adequate sample size for SSD development, (2) determine if HD5s developed from ICE-generated SSDs are similar to measured HD5s, (3) find the best potential bird surrogate for estimation of ICE HD5, (4) determine if ICE SSDs generated using both bird and mammal surrogates improved HD5 accuracy VOL. 42, NO. 9, 2008 / ENVIRONMENTAL SCIENCE & TECHNOLOGY 9 3447
compared to ICE HD5s derived from bird-only surrogacy, (5) identify differences in HD5s that might result from different species composition in SSDs, and (6) detect any differences in SSD parameter estimation using inverse estimation as an alternative to maximum likelihood approximation. ICE models for wildlife were used to predict toxicity values and generate SSDs for 23 chemicals using four avian surrogates. The hazard levels associated with the fifth percentile of the distribution (HD5) were compared between ICE SSDs and SSDs created with measured data. The results of this investigation are intended to provide risk assessors with an approach for generating SSDs for wildlife species when measured values are limited.
Materials and Methods Acute Toxicity Database. The wildlife database used for creation of ICE models comprised 4329 single dose LD50 values (mg/kg bodyweight) for 156 species and 939 chemicals. Details on source of information, toxicity testing methods, and quality control of the database are described elsewhere (2) and are available as Supporting Information. ICE Estimated Species Sensitivity. A separate set of ICE models was generated for each chemical that had been tested on a minimum of 15 species based on the sample size analysis discussed below. Predicted toxicities from ICE models were subsequently used in SSD development and estimation of predicted HD5 values. For each chemical, all toxicity values for that chemical were removed from the original database to create an independent data subset to compare ICE-derived hazard dose estimates with the hazard dose levels of removed measured values. The data subsets, each with one chemical excluded, were used to generate ICE models with four avian surrogates and one mammalian surrogate: Japanese quail, mallard, northern bobwhite, red-winged blackbird, and Norway rat. Surrogate species were selected on the basis of having the greatest number of previously developed robust ICE models (2), although the new models created for this exercise represented the best fit to the data subset with possibly one fewer chemical data point. ICE models were developed as the least-squares regression of measured toxicity values from each data subset, using each surrogate with all possible predicted species. Species used in model development shared known toxicity values with the surrogate species for at least three chemicals. The known toxicity of the surrogate species, which was removed from the data subset along with all other measured data for that chemical, was used as input for each model such that log10(predicted toxicity) ) intercept + slope × log10(surrogate toxicity) (1) estimated the unknown toxicity value of the other species. A detailed example is available as Supporting Information. Toxicity estimates where the 95% confidence intervals, as calculated using the method described by Zar (8), were greater than 5-fold of the predicted toxicity value and were not included in the analysis to exclude predicted toxicity values with high uncertainty. This exclusion criterion, based on the interlaboratory variability for acute toxicity tests (9), prevented extraneous variability associated with less robust toxicity estimates from influencing SSDs. Only significant ICE models (P < 0.05) were included in the analysis, and negatively correlated models were additionally excluded. Bird surrogates were used for prediction of toxicity values of bird and mammal species (50 species total), and the Norway rat was used to predict mammal toxicity for use in taxon-specific SSD development as described below. General SSD Model. Although many distributions have been used to generate SSDs (10–14), the log–logistic distribution often fits data sets best (11, 6) and was used here with the following equation (1): 3448
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 9, 2008
Y ) 1 ⁄ (1 + exp((R - X) ⁄ β))
(2)
where Y is the cumulative probability of species affected at that toxicity value and X is the log10-transformed LD50 value. X values included in the database were listed in increasing order, and the corresponding Y value was defined as (the order of the data point)/(1 + total number of data points). The location or intercept, R, and the slope or scale parameter, β, were derived using maximum likelihood estimation. The hazardous dose lower than acute toxicity values of 95% of all species (HD5) was then determined by calculating the LD50 at the fifth percentile of the distribution fit to toxicity values (Y ) 0.05). An example of SSDs generated in this study using measured and predicted toxicity values is available as Supporting Information. All modeling was executed using S-plus (15). Minimum SSD Sample Size. Seven chemicals with measured toxicity values for at least 25 species were selected to determine the effect of SSD sample size on the accuracy of the HD5. For each chemical, SSDs of various sample sizes were developed from the measured toxicity values. Sample sizes increased in intervals of two between five and 15 species and increased in intervals of five species thereafter (n ) 5, 7, 9, 11, 13, 15, 20, 25, . . .). Each sample size for each chemical had 100 replicates of a random subset of all species with measured toxicity values. From each replicate subset, an SSD was developed and HD5 calculated. The proportional difference between the HD5 calculated using all data for a chemical and HD5 determined from the subset of each sample size replicate was calculated as abs|HD5allHD5subset|/HD5all and is equivalent to the difference in log10transformed values of the two HD5s. The mean and 95% confidence interval of the proportional differences were plotted according to sample size for all chemicals. The point at which the difference from the complete data set was negligible was determined by identifying the interval where the increase in sample size resulted in
