Environmental Toxicology and Chemistry, Vol. 33, No. 3, pp. 688–695, 2014 # 2013 SETAC Printed in the USA
Hazard/Risk Assessment AUGMENTING AQUATIC SPECIES SENSITIVITY DISTRIBUTIONS WITH INTERSPECIES TOXICITY ESTIMATION MODELS JILL A. AWKERMAN,* SANDY RAIMONDO, CRYSTAL R. JACKSON, and MACE G. BARRON Gulf Ecology Division, US Environmental Protection Agency, Gulf Breeze, Florida (Submitted 13 March 2013; Returned for Revision 10 April 2013; Accepted 1 November 2013) Abstract: Species sensitivity distributions (SSDs) are cumulative distribution functions of species toxicity values. The SSD approach is being used increasingly in ecological risk assessment but is often limited by available toxicity data needed for diverse species representation. In the present study, the authors evaluate augmenting aquatic species databases limited to standard test species using toxicity values extrapolated from interspecies correlation estimation (ICE) models for SSD development. The authors compared hazard concentrations at the 5th centile (HC5) of SSDs developed using limited measured data augmented with ICE toxicity values (augmented SSDs) with those estimated using larger measured toxicity datasets of diverse species (reference SSDs). When SSDs had similar species composition to the reference SSDs, 0.76 of the HC5 estimates were closer to the reference HC5; however, the proportion of augmented HC5s that were within 5-fold of the reference HC5s was 0.94, compared with 0.96 when predicted SSDs had random species assemblages. The range of toxicity values among represented species in all SSDs also depended on a chemical’s mode of action. Predicted HC5 estimations for acetylcholinesterase inhibitors showed the greatest discrepancies from the reference HC5 when SSDs were limited to commonly tested species. The results of the present study indicate that ICE models used to augment datasets for SSDs do not greatly affect HC5 uncertainty. Uncertainty analysis of risk assessments using SSD hazard concentrations should address species composition, especially for chemicals with known taxa-specific differences in toxicological effects. Environ Toxicol Chem 2014;33:688–695. # 2013 SETAC. This article is a US Government work and is in the public domain in the USA. Keywords: Aquatic toxicology
Species sensitivity distributions
Species extrapolation
Risk assessment
tradeoff of increasing the sample size while introducing uncertainty in the toxicity prediction has not been evaluated for aquatic species. For datasets with limited experimental data, extrapolated toxicity values can increase sample sizes while reducing statistical uncertainty in mammalian and bird species [11]. In the present study, we examine the utility of augmenting SSDs for aquatic species with ICE-estimated toxicity values. We compare HC5 estimation using minimal datasets of commonly tested species’ toxicity data without extrapolated values to those supplemented with ICE-estimated values. Meaningful application of SSDs for ecological risk assessment depends on identifying assumptions and establishing guidelines for developing SSDs. One assumption of the SSD is that species are a random assemblage of interchangeable species; indeed, this is seldom the case when using available data. Commonly tested toxicological species can be less sensitive than other species with a similar size and other characteristics [10]. Although taxonomic relatedness is a good indication of representative toxicity values for a species [7], even across large geographic distances [12], all species are not interchangeable within the SSD [13]. The assemblage of species can affect HC5 estimation, such that large sample sizes do not necessarily produce consistent toxicity ranges for modeling SSDs [14]. The present study explores the potential tradeoff between increasing the sample size of an SSD and propagating error by including extrapolated toxicity values. We determine the variability of HC5s estimated with inclusion of ICE-extrapolated toxicity data within an SSD. We examine the effect on HC5 of including extrapolated toxicity values in SSDs to increase taxonomic diversity and sample size. Through a comparative modeling approach, we identify SSD characteristics that contribute to uncertainty in estimating HC5, and these results offer general rules to consider when using extrapolated data in
INTRODUCTION
Species sensitivity distributions (SSDs) are used in ecological risk assessments to model the cumulative proportion of species’ sensitivity to a specific compound or environmental stressor [1,2]. In application, commonly used hazard levels of SSDs estimate the concentration affecting the specified proportion of the species (e.g., the hazard concentration that affects 5% of included species [HC5]). The SSD approach uses a set of toxicity data that represents species assemblages or various taxonomic groups of interest; notably, however, available toxicity data are often limited to a small number of commonly tested species. Consequently, statistical or regulatory criteria, such as minimum sample size or minimum data requirements for taxonomic diversity are rarely met [3–5], particularly for new or emerging compounds. Inclusion of extrapolated toxicity values offers a means to increase both sample size and taxa diversity in SSDs. However, extrapolated values with toxicity estimation error subsequently will add uncertainty to estimates of HC5 values [6]. Interspecies correlation estimation (ICE) models are log-linear least squares regressions of the sensitivity relationship between 2 species and use toxicity values of species with test data to estimate the toxicity to species without data for a chemical of interest [7]. Estimates of HC5 from SSDs of only ICE-estimated toxicities have been shown to be comparable to those from SSDs with only measured toxicity data for both aquatic [8] and terrestrial species [9,10]. Interspecies correlation estimation models have been used to augment SSDs for wildlife [11], but the statistical All Supplemental Data may be found in the online version of this article. * Address correspondence to [email protected]. Published online 8 November 2013 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/etc.2456 688
Uncertainty in extrapolation-augmented SSDs
SSDs. We also describe the effect of error in extrapolating single toxicity values on HC5 estimation through a local sensitivity analysis. We use both measured and extrapolated aquatic species toxicity data to understand the key sources of uncertainty in HC5 estimation. An additional goal is to identify the proportion of extrapolation data in an augmented dataset that optimizes taxa diversity while minimizing statistical uncertainty of the SSD. In doing so, we provide guidelines for applying ICE models when augmenting empirical datasets for ecological risk assessments. METHODS
Development of reference species sensitivity distributions
All analyses used toxicity data and models from the Web-ICE modeling application (http://epa.gov/ceampubl/fchain/webice/; [15]). Interspecies correlation estimation models are log-linear least squares regressions between 2 species that have been tested for the same chemicals
Environ Toxicol Chem 33, 2014
689
species. If multiple core surrogate models were available for a predicted species, we used the model with the lowest mean square error to estimate toxicity and added this value to the SSD dataset. Extrapolated toxicity values used to augment minimal datasets were restricted to the same species composition as the reference SSD. Because ICE models were not always available for every species in the reference SSD, augmented SSDs sometimes had fewer species. The augmented SSDs had from 1 to 29 more species than SSDs developed from the minimal dataset alone. Fold difference from the reference HC5 (HC5ref) was calculated for both the minimal toxicity dataset and the extrapolation-augmented dataset as follows ðHC5x HC5ref Þ HC5ref where HC5x is the hazard concentration derived from SSDs using either minimal or augmented datasets. SSD augmentation with variable species composition
log10 ðpredicted toxicityÞ ¼ slope log10 ðsurrogate toxicityÞ þ intercept The model is used to estimate toxicity to the predicted species from the surrogate when the toxicity for the latter is known. We used a subset of the Web-ICE database for 49 chemicals with experimental acute toxicity values (median lethal concentration [LC50] and median effective concentration [EC50] in the case of invertebrates with an endpoint equivalent to ecological mortality) for 10 or more species of aquatic invertebrates and vertebrates. The Web-ICE database follows strict data standardization criteria, which are described in detail by Raimondo et al. [7] to reduce extraneous sources of variability in estimating toxicity. For each of these chemicals, reference SSDs were created from all available experimental data (Supplemental Data, Table S1). Species sensitivity distributions were developed using S-plus [16] to fit a log-logistic cumulative distribution function fit to toxicity data as Y ¼ 1=f1 þ exp½ða XÞ=bg Maximum likelihood estimation was used to derive the intercept, a, and scale parameter, b, to toxicity values X at cumulative proportion Y. Model fit was evaluated by a Kolmogorov Smirnov goodness of fit test. The 5th centile of this distribution was used in all analyses below as the reference HC5. SSD augmentation with same species composition
For each chemical, we developed an SSD using a subset of commonly tested core species as an initial minimal dataset. This minimal SSD represents the scenario in which only standard test species data are available. These commonly tested core species were Americamysis bahia, Chironomus plumosus, Crassostrea virginica, Daphnia magna, Gammarus fasciatus, Ictalurus punctatus, Lepomis macrochirus, Micropterus salmoides, Oncorhynchus clarkii, Oncorhynchus mykiss, and Pimephales promelas. Toxicity data for all of these species were not available for all chemicals; therefore, SSDs were developed using as many of these species as possible, with a minimum of 5 (Table 1) [4]. Next, we augmented the minimal dataset for each chemical with extrapolated toxicity values calculated from ICE models. We used only the core species included in the minimum dataset for each chemical as surrogates to predict toxicity to other
To examine the uncertainty of HC5 estimation as it relates to sample size and species composition, extrapolated toxicity values were added to the minimal dataset as described in the section SSD augmentation with same species composition but without restricting species composition to that of the reference SSD. In addition to creating SSDs with as many extrapolated toxicity values as possible, we replicated augmented SSDs for each chemical by adding randomly chosen extrapolated toxicity values to all available core toxicity values (from 3 to 9 depending on data availability). Combinations of core and extrapolated values were developed for sample sizes of up to 30 species, with 10 replicate SSDs per sample size per chemical. Although ICE models were available to extrapolate toxicity for 61 to 74 species for each chemical, we limited our replicate SSDs to a maximum of 30 species to maintain variability in the subset data used in SSDs and reduce possible pseudoreplication as a result of our replicate subsampling approach. All SSDs were fit to toxicity data as described in the section Development of reference species sensitivity distributions. For each chemical, HC5 was estimated for all replicate SSDs, and the fold difference from the reference HC5 was determined. Using a multimodel inference approach, we compared a candidate list of linear mixed models that included various attributes of the replicate augmented SSDs as independent variables and the absolute fold difference of HC5s estimated from augmented and reference SSDs as the dependent variable. This approach ranks the candidate models using Akaike’s Information Criterion corrected for small sample sizes (AICc) to find the best model fit [17,18]—in this case, the combination of SSD attributes that best accounted for variability of HC5 relative to the reference HC5. Models fit to the log-transformed absolute fold difference in HC5s included single effects or interactions of the following independent variables for each extrapolationaugmented SSD: a measure of taxonomic diversity represented as the number of taxonomic orders relative to number of species in the SSD (DIVERSE); the proportion of vertebrates (VERT); extrapolation SSD goodness of fit as represented by pseudo R2 (R2); the broad mode of action of the chemical (MOA); and the proportion of extrapolated values in the SSD (EXTRAP). Random effects of individual chemical datasets were also included in each model (CHEM). For each model, we present AICc, AICc difference, log likelihood, and Akaike weight. Models were developed using package lme4 [19]. Each model was based on a priori expectations about variables affecting the
690
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
Table 1. Mode of action (MOA), hazard concentrations at the 5th centile (HC5) values, and fold difference (FD) between reference and augmented species sensitivity distributions (SSDs) including all available toxicity data (reference HC5) for 49 chemicalsa Specific MOA
Broad MOA
Chemical
No. ref
Reference HC5 (mg/L)
N Coreb
Core FD
No. Same Extr
Same Extr FD
No. Full Extr
Full Extr FD
13 10 12 12 10 14 11 38 17 11 15 14 15 18 23 11 16 24 31 17 22 11 13 13 11 10 16 13 10 29 18 10 21 25 12 10 22 10 23 11
7.7 18 815 6.2 328 3298 91 6398 103 8911 17 77 42 0.27 0.28 0.02 0.15 1.4 0.06 1.2 1.6 1.9 6.0 0.56 0.11 1.2 6.2 0.04 2.3 0.18 0.66 5.9 3.2 0.07 0.83 0.79 0.22 1.2 0.03 0.2 0
7 6 6 7 5 — — 8 5 — 7 8 — 9 8 6 8 6 8 6 9 9 8 8 6 8 7 5 6 7 5 — 6 7 5 5 7 5 6 —
–0.26 0.01 –0.41 –0.31 –0.53 — — –0.91 –0.02 — 0.22 0.13 — 4.59 0.67 20.87 –0.9 –0.74 –0.71 –0.22 –0.91 –0.07 0.05 4.9 –0.92 0.19 1.37 –0.96 1.61 –0.26 0.81 — 2.26 2.67 7.74 –0.63 1.29 –0.34 –0.39 —
12 9 12 11 8 13 11 37 17 10 15 13 12 18 21 11 14 20 28 16 19 10 11 13 11 10 13 10 10 27 17 7 17 23 11 9 20 7 20 8
–0.29 –0.08 –0.69 –0.71 –0.12 –0.02 –0.22 –0.38 –0.96 0.19 –0.18 0.22 0.15 0.13 0.69 6.96 –0.07 0.04 1.01 –0.08 0.15 0.36 0.54 –0.51 0.14 –0.25 –0.61 0.48 1.5 –0.37 0.04 –0.08 0.75 0 –0.04 –0.8 –0.29 –0.69 1.25 5.43
71 78 71 70 70 66 76 79 64 73 76 76 62 77 77 71 79 72 80 71 77 77 72 77 71 71 72 75 70 71 63 70 71 72 63 75 77 69 69 43
0.63 –0.9 –0.61 –0.41 –0.8 –0.78 –0.91 –0.95 –0.98 –0.32 –0.8 –0.49 66.19 2.68 8.09 0.12 –0.74 –0.23 –0.64 1.22 –0.89 1.62 0.26 –0.15 0.64 0.83 1.65 –0.99 2.5 –0.6 –0.66 –0.87 –0.74 0.42 0.55 –0.93 –0.7 0.02 –0.39 0.09
0.52
72
–0.58
Narcosis
Nonpolar
Narcosis
Polar
AChE inhibitor
Carbamates
AChE inhibitor
OP
Neurotoxicant
OC
Neurotoxicant
Pyrethroid
Oxidative Phosphorylation
Inhibitor
2,4-Dichlorophenoxy acetic acid Captan Diuron Molinate 4-Nonylphenol Phenol Carbaryl Mexacarbate Carbofuran Aminocarb Methomyl Trichlorfon Fenthion Parathion Coumaphos Dichlorvos Azinphos-Methyl Malathion Fenitrothion Methyl parathion Disulfoton Naled Ethion Phosmet ENP Chlorpyrifos Temephos Profenofos DDT Lindane Dieldrin Endrin Methoxychlor Heptachlor Endosulfan Toxaphene Fenvalerate Permethrin Antimycin A
16
14.3
5
2.6
14
Oxidative Phosphorylation Other Other
Uncoupler
20 ,50 -Dichloro-40 nitrosalicylanilide compd. with 2-aminoethanol (1:1) Pentachlorophenol
29
6.8
7
0.5
23
–0.3
75
–0.68
Other
Respiratory toxicity
Ammonia Mercury Cadmium Copper Zinc Chromium VI Sodium chloride
56 12 33 42 34 14 17
5251 1.8 1.4 4.9 81.9 270 1 741 225
7 — — 6 — — —
1.99 — — 0.78 — — —
27 10 20 30 19 11 9
–0.27 0.49 –0.58 0.51 –0.63 –0.51 –0.04
82 67 77 78 71 60 76
–0.88 –0.6 –0.95 0.4 –0.8 –0.94 –0.98
Ammonia stress Metallic stress
a
Results for comparison include fold difference between the reference HC5 and HC5s using minimal data (Core FD), HC5 for extrapolation-augmented SSDs limited to the same species as reference (Same Extr FD) and HC5 for SSDs augmented with all extrapolation models (Full Extr FD). b Core datasets with fewer than 5 species were not used to develop SSD. Sample size is listed for each (No. Ref, No. Core, No. Same Extr, and No. Full Extr). AChE ¼ acetylcholinesterase; OP ¼ organophosphate; OC ¼ organochlorine.
HC5 estimation (taxonomic representation, model fit, chemical type, and extrapolation uncertainty [1,4,5,7]). The taxonomy of the aquatic species included in the SSDs was classified according to the Integrated Taxonomic Information System. We used a likelihood ratio test to determine if significant differences in model fit existed between 2 top-ranking models.
Sensitivity of HC5 to individual species toxicity values
A sensitivity analysis determined the influence of toxicity estimation error on the HC5. For each toxicity value in each of the 49 SSDs, we altered the value in 16 iterations within the range of 50% to 200% of the measured data (e.g., a toxicity value of 2 within the SSD would be exchanged with values ranging
Uncertainty in extrapolation-augmented SSDs
from 1 to 4). Estimates of HC5s were determined for SSDs created with each iteration of each data point. Finally, the fold difference between the reference HC5 and the manipulated HC5s determined for all iterations was calculated. RESULTS
Hazard dose estimation with same species composition
Surrogate data availability varied by chemical and MOA. Those SSDs that included only minimal core toxicity datasets ranged in sample size from 5 to 9 species (Table 1). The proportion of minimal core datasets that produced HC5s within 5-fold and 10-fold of the reference HC5 was 0.95 and 0.97, respectively, for 38 chemicals. For all 49 chemicals, including extrapolated toxicity values in SSDs limited to species in the reference SSD increased the sample sizes to 9 to 37 species. All same-species HC5 estimates were within 10-fold, and 0.96 were within 5-fold. Hazard dose estimation with variable species composition
The proportion of HC5s from SSDs augmented with all possible extrapolated values that were within 5-fold and 10-fold of the reference HC5 were 0.96 and 0.98, respectively (Figure 1). The proportion of chemicals with a smaller fold difference from the reference HC5 using a full complement of extrapolated values compared to the minimal core dataset was 0.45 (Table 1). Including all possible species in the extrapolated SSDs reduced fold difference from the reference HC5 for only 0.22 of the 49 chemicals when compared with extrapolated SSDs that were limited to the same species composition as reference SSDs. The HC5 and lower 95% confidence interval of the HC5 (HCL) estimates were lower than core data estimates for 23 of the 38 chemicals (0.62; Table 2). Organophosphates did not produce more conservative HC5 estimates with greater species representation for 12 of the 17 chemicals (0.71). Among the 12 740 SSDs produced with various quantities of extrapolated toxicity values, fold difference from the reference HC5 depended on chemical, sample size, species composition, and proportion of extrapolated values included. Estimates of HC5 for all augmented SSDs were within 10-fold of the reference HC5 for 98% of the SSDs developed, and 96% of the SSDs produced HC5 estimates within 5-fold of the reference SSD. The 2 top-ranking mixed effects models in our model set included an interaction effect between the proportion of vertebrates in the SSD and the MOA being modeled, as well as the random effect of the chemical (Figure 2, Table 3). The 2 models with a substantial model weight differed only in the inclusion of pseudo R2. A likelihood ratio test comparison of the top 2 models in our model set indicates no significant difference between them, suggesting that model fit has minimal influence on HC5 estimation (x2 ¼ 0.92, p ¼ 0.34). Fold difference from the reference HC5 is similar for proportions of the ICE models (Figure 1). Sensitivity of HC5 to individual species toxicity values
Variability in the toxicity of a single data point affected the HC5 differently depending on error bias and the sensitivity of the species. Underestimating toxicity values in the lower portion of the distribution increases the HC5 estimates, and these points decrease HC5 estimates when they are overestimated (Figure 3). Conversely, the toxicity values in the upper portion of the distribution decrease HC5 estimates when they are underestimated and increase HC5 estimates when they are overestimated.
Environ Toxicol Chem 33, 2014
691
Estimation error of species in the lower quartile of the distribution has a relatively greater influence on HC5 estimates, and error in a single toxicity estimate affects HC5 estimation more with fewer species in the SSD (Figure 4). DISCUSSION
The data needed to ensure SSDs represent diverse species assemblages and maintain a statistically robust sample size are difficult to obtain from existing toxicological data, which often represent only relatively few commonly tested species per chemical. Interspecies estimation models offer a way to supplement available data with extrapolated toxicity values to allow generating SSDs with broader taxonomic diversity and larger sample sizes. However, introducing uncertainty with extrapolated toxicity values is a concern when using extrapolation models to supplement measured data to estimate hazard dose levels. The current study’s simulations demonstrate that uncertainty associated with toxicity extrapolation has only a minimal effect on estimating HC5. Species sensitivity distributions are of greater utility in ecological risk assessment when their species composition includes several taxonomic classes, representing different functional groups of an ecological community. Large and diverse assemblages of species, however, produce variable ranges of toxicological effects that can differ depending on the chemical, according to physiology, behavior, and life history [14,20]. The taxonomic composition of SSDs affects HC5 estimates for aquatic species more than differences in geography or habitat of the assemblage [12,21]. Identifying biological and statistical variables that contribute to HC5 variance can help reduce uncertainty in risk assessments through the development of criteria for data selection and inclusion. The proportion of extrapolated values that augmented measured toxicity data was not identified as an influential parameter in HC5 estimation in the top models within our model set. Species composition is influential in HC5 estimation and should be representative of taxonomic groups or aquatic assemblages of interest, considering the variable impacts of some types of chemicals. The interaction between MOA and proportion of vertebrates in our top models further identifies the importance of species composition, particularly when discrepancies in sensitivity among taxonomic groups are known for different MOAs [22]. The MOA-specific differences in HC5 estimation demonstrate chemical-specific concerns that should be considered when developing and evaluating SSDs. Recognizing variable MOA effects emphasizes the need to include relevant taxa within the species composition of the SSD, ensuring that HC5s protect the most sensitive taxa. Risk assessment criteria specifying a minimum number of genera or a diverse group of representative species might be important for chemicals without a known discrepancy in effects among constituent taxonomic groups represented. For example, aquatic life criteria require 8 minimum data requirements aimed at including taxonomic diversity in SSDs [3]. The 49 chemicals selected for the present study were chosen because they had been tested on the greatest number of species. However, few chemicals met the requirements of taxa diversity specified by aquatic life criteria [3]. For chemicals tested on only a few species, the diversity of species models in the Web-ICE database allows broader taxonomic representation. The current ICE model set contains a higher proportion of vertebrate models. Fish are overrepresented in most toxicological databases and insects are often underrepresented, such that available toxicity data rarely represent true
692
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
Table 2. Hazard concentrations at the 5th centile (HC5) and lower 95% confidence intervals of HC5 (HCL) for 49 chemicals using minimal data (Core), extrapolation-augmented species sensitivity distributions (SSDs) limited to the same species as reference (Same Extr) and SSDs augmented with all extrapolation models (Full Extr)
Specific MOA Nonpolar
Polar Carbamates
OP
OC
Pyrethroid Inhibitor
Uncoupler Ammonia stress Metallic stress
Respiratory toxicity
Chemical 2,4-Dichlorophenoxy acetic acid Captan Diuron Molinate 4-Nonylphenol Phenol Carbaryl Mexacarbate Carbofuran Aminocarb Methomyl Trichlorfon Fenthion Parathion Coumaphos Dichlorvos Azinphos-methyl Malathion Fenitrothion Methyl parathion Disulfoton Naled Ethion Phosmet ENP Chlorpyrifos Temephos Profenofos DDT Lindane Dieldrin Endrin Methoxychlor Heptachlor Endosulfan Toxaphene Fenvalerate Permethrin Antimycin A 20 ,50 -Dichloro-40 nitrosalicylanilide compd. with 2-aminoethanol (1:1) Pentachlorophenol Ammonia Mercury Cadmium Copper Zinc Chromium VI Sodium chloride
Core HCL (mg/L)
Core HC5 (mg/L)
No. Same Extr
Same Extr HCL (mg/L)
Same Extr HC5 (mg/L)
No. Full Extr
Full Extr HCL (mg/L)
Full Extr HC5 (mg/L)
7 8 — 9 8 6 8 6 8 6 9 9 8 8 6 8 7 5 6 7 5
0.19 2340 0.02 17 113 — — 0.07 5230 — 2.09 3.2 — 0.019 2.5 10–5 0.0012 6.3 10–5 1.7 10–5 1.3 10–3 1.05 10–3 2.3 10–4 0.21 4.5 10–3 3.110–3 2.8 10–5 0.63 4.9 10–4 5.2 10–8 9.8 10–3 0.041 2.2
5.7 19 096 3.7 226 1565 — — 9.3 8703 — 95 47 — 1.6 0.039 3.4 0.14 0.015 0.33 1.3 0.17 5.6 0.58 0.63 0.093 7.3 0.10 0.080 0.47 0.49 10.7
6 7 5 5 7 5 6 — 5
0.060 0.41 1.4 7.5 10–4 1.2 6.5 10–5 2.5 10–3 — 9.1
0.22 3.03 6.9 0.082 2.8 0.017 0.12 — 51
12 9 12 11 8 13 11 37 17 10 15 13 12 18 21 11 14 20 28 16 19 10 11 13 11 10 13 10 10 27 17 7 17 23 11 9 20 7 20 8 14
0.85 3687 0.11 13 872 55 1466 21 97 5.2 8.9 10 0.0084 0.018 1.6 10–3 0.034 0.029 3.3 10–3 0.32 0.090 0.13 0.43 0.036 8.510–4 0.043 0.48 3.6 10–4 0.022 0.032 0.18 2.5 0.54 0.043 0.24 0.12 1.0 10–3 0.39 1.1 10–4 0.15 1.0 10–3 7.5
5.4 17 222 1.9 95 2917 89 5014 63 352 21 64 51 0.31 0.32 0.039 1.2 1.3 0.059 2.3 1.5 2.1 8.1 0.85 0.052 1.4 4.6 0.017 3.3 0.45 0.41 6.1 2.9 0.12 0.83 0.76 0.045 0.87 0.008 0.45 0.027 22
71 78 71 70 70 66 76 79 64 73 76 76 62 77 77 71 79 72 80 71 77 77 72 77 71 71 72 75 70 71 63 70 71 72 63 75 77 69 69 43 72
4.5 661 0.43 75 280 10 239 1.2 72 4.7 3.9 7.8 5.4 0.17 0.031 0.028 0.047 7.4 10–3 0.066 0.88 0.022 5.2 0.16 0.011 0.48 5.5 0.022 6.2 10–4 0.25 0.10 0.90 0.12 5.6 10–3 0.51 0.54 2.6 10–3 0.15 8.5 10–3 0.043 1.0 10–3 2.3
13 1965 2.4 193 655 20 557 5.5 187 12 15 21 18 1.04 0.21 0.17 0.35 0.044 0.42 3.6 0.21 16 0.70 0.091 2.0 11 0.11 0.013 0.63 0.26 2.02 0.42 0.017 1.18 1.22 0.02 0.37 0.027 0.12 4.6 10–3 6.03
7 7 —
1.3 9722 —
10 15 694 —
6 — — —
0.55 — — —
8.7 — — —
23 27 10 20 30 19 11 9
1.6 1943 0.23 0.045 3.1 5.07 9.03 630 649
4.8 3842 2.7 0.57 7.4 29.9 132 1 675 747
75 82 67 77 78 71 60 76
0.64 212 0.18 7.8 10–3 2.35 5.43 3.37 6263
2.2 617 0.72 0.066 6.8 17 17 30 068
No. Core 7 6 6 7 5 — — 8 5
MOA ¼ mode of action; OP ¼ organophosphate; OC ¼ organochlorine.
ecological communities [23]. Inclusion of a greater proportion of vertebrates in most SSDs can increase the range of toxicity values with more insensitive species. For chemicals that are more toxic to invertebrates, species composition of the SSD will have a greater effect on hazard concentration estimations and risk assessment. The uncertainty associated with ICE toxicity estimation is quantified for each value as modeled confidence interval around the prediction, and some degree of professional judgement is required to evaluate these values [15]. Despite recommendations
presented here on the importance of SSD species composition, recognizing where large uncertainty exists around an ICE toxicity value and the relative sensitivity of that species are also important considerations when using extrapolated values to augment datasets. The sensitivity analysis identified relatively greater influence on HC5 variability of a single toxicity value when species were in the more sensitive range of the SSD. Inclusion of models with large confidence intervals for extrapolated values that lie in the lower quartile of the SSD is not recommended, particularly when datasets are small.
Uncertainty in extrapolation-augmented SSDs
Environ Toxicol Chem 33, 2014
693
Table 3. Candidate model set evaluating the absolute fold-difference between hazard concentrations at the 5th centile (HC5) estimates of extrapolation-augmented species sensitivity distributions (SSDa) and HC5 estimates using measured toxicity values Modelb VERT MOA þ CHEM VERT MOA þ R2 þ CHEM VERT MOA þ R2 þ EXTRAP þ CHEM VERT MOA þ R2 þ SPECIES þ CHEM VERT þ CHEM VERT þ EXTRAP þ R2 þ CHEM VERT þ SPECIES þ R2 þ CHEM R2 þ CHEM DIVERSE þ CHEM EXTRAP þ CHEM DIVERSE þ R2 þ CHEM EXTRAP þ R2 þ CHEM EXTRAP þ DIVERSE þ CHEM DIVERSE þ EXTRAP þ R2 þ CHEM SPECIES þ CHEM SPECIES þ DIVERSE þ CHEM SPECIES þ R2 þ CHEM DIVERSE þ SPECIES þ R2 þ CHEM MOA þ CHEM SPECIES þ EXTRAP þ CHEM DIVERSE MOA þ CHEM DIVERSE MOA þ R2 þ CHEM DIVERSE MOA þ R2 þ EXTRAP þ CHEM DIVERSE MOA þ R2 þ SPECIES þ CHEM
log likelihood
AICc
DAICc
AICc weights
–6463.65 –6463.19 –6465.36 –6469.82 –6725.78 –6727.46 –6732.42 –6827.91 –6829.62 –6830.03 –6829.28 –6830.22 –6830.29 –6830.50 –6834.13 –6833.29 –6834.48 –6833.69 –6826.90 –6835.11 –6819.35 –6819.27 –6820.32 –6823.46
12 975.39 12 976.48 12 982.83 12 991.75 13 459.56 13 466.92 13 476.85 13 663.82 13 667.25 13 668.07 13 668.56 13 670.45 13 670.59 13 673.00 13 676.25 13 676.59 13 678.96 13 679.40 13 679.83 13 680.23 13 686.80 13 688.64 13 692.75 13 699.03
0.00 1.09 7.44 16.36 484.17 491.53 501.46 688.43 691.86 692.68 693.17 695.07 695.20 697.61 700.86 701.20 703.57 704.01 704.44 704.84 711.41 713.25 717.36 723.64
0.62 0.36 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
a
Extrapolation-augmented SSDs were developed with 10 replicated subsets of up to 30 randomly chosen species. The variables included in models were sample size (SPECIES), the proportion of extrapolated values (EXTRAP), the proportion of vertebrates in the SSD (VERT), a measure of taxonomic diversity that included the number of orders represented per species (DIVERSE), the mode of action of a chemical (MOA), and the SSD fit (R2). Each model also included random effects of chemical (CHEM), which determined surrogate and models available for subset SSD included in analyses. Models are presented in ranked order of lowest Akaike’s Information Criterion corrected for small sample sizes (AICc) and DAICc. b
However, differences in HC5 estimation were relatively small within the ranges of toxicity characterized in the present study. The sensitivity analysis also demonstrated how the HC5 changes with uncertainty in the values in the upper and lower quartiles. Overestimating the toxicity values (or using the predicted upper confidence interval value) results in a lower, more conservative
HC5 value when the toxicity value is in the more sensitive range of the SSD. Conversely, applying the lower confidence interval of less sensitive toxicity values results in a more conservative hazard concentration. Including extrapolated values with larger uncertainty for species that lie in the median of the SSD has little effect on HC5 estimation. Using multiple surrogates to develop
Figure 1. Log-transformed fold difference from reference hazard concentrations at the 5th centile (HC5) with increasing proportion of extrapolated toxicity values included in the species sensitivity distribution. Boxes indicate the median within upper and lower quartiles, and whiskers denote 90% of estimates.
Figure 2. Variance in hazard concentrations at the 5th centile (HC5) estimation, as noted by the log fold difference, by mode of action of chemical for replicate species sensitivity distributions using various combinations of experimental and extrapolated toxicity values. Boxes indicate the median within upper and lower quartiles, and whiskers denote 90% of estimates.
694
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
chemical. Hazard concentrations depend on the dataset with which the SSD was developed, a result underscored by the analyses presented here. Reference SSDs are representative of data available to a risk assessor for broadly tested chemicals. Comparing reference SSDs to smaller subsets of data, representing species composition more like that available for most chemicals, demonstrates that while minimal datasets often produce HC5 estimates similar to larger datasets (0.95 within 5-fold of reference), many minimal dataset HC5 estimations (0.61) are improved when extrapolated values are included. Variable results as sample sizes are maximally increased emphasize the importance of considering species composition, surrogate selection, and relative sensitivity of species when developing extrapolation-augmented SSDs for risk assessment. SUPPLEMENTAL DATA
Table S1. (31 KB DOCX).
Figure 3. Effect of underestimating (1) toxicity values on fold difference from reference hazard concentrations at the 5th centile (HC5s) ([predicted HC5-reference HC5]/reference HC5). Toxicity values are color-coded to represent their relative sensitivity by quartile within the species sensitivity distribution from sensitive to insensitive (black ¼ < 0.25; red ¼ > 0.25, 0.5, 0.75).
augmented SSD is recommended, when possible, to minimize uncertainty in the models selected. Previous studies also support inclusion of ICE models for which surrogate and predicted species are closely related to reduce uncertainty [8]. These general trends can help guide professional judgement for accepting extrapolated values to augment SSDs. The present study evaluated difference in HC5 from a reference HC5 derived using all available measured values within our database. Reference SSDs are not intended to represent real communities and are not necessarily intended to be interpreted as definitive hazard concentration values for a given
Figure 4. Diminishing effects of underestimated (green) or overestimated (yellow) single toxicity point, independent of location within the species sensitivity distribution (SSD), with increasing sample size of SSD. Fold difference from reference hazard concentrations at the 5th centile (HC5) is (predicted HC5-reference HC5)/reference HC5.
Acknowledgment—For database development and maintenance, we thank B. Montague (US Environmental Protection Agency [USEPA], Office of Pesticide Programs) and P. Mineau, A. Baril, and B. Collins (Environment Canada). We are also grateful to support personnel: D. Vivian, M. Marchetto, A. DiGirolamo, C. Chancy, N. Lemoine, N. Allard, and C. McGill (USEPA, Gulf Ecology Division). Two anonymous reviewers provided helpful guidance to an earlier draft. Support in preparing this manuscript was provided by the USEPA, National Health and Environmental Effects Research Laboratory. Although this manuscript was submitted for internal review within USEPA and approved for publication, it does not reflect the views of the agency or constitute an endorsement of any trade names or commercial products mentioned.
REFERENCES 1. Posthuma L, Traas TP, Suter GW. 2002. General introduction to species sensitivity distributions. In Posthuma L, Suter GW, Trass TP, eds, Species Sensitivity Distributions in Ecotoxicology. CRC Press, Boca Raton, FL, USA, pp 3–10. 2. De Vries P, Tamis JE, Murk AJ, Smit MGD. 2008. Development and application of a species sensitivity distribution for temperature-induced mortality in the aquatic environment. Environ Toxicol Chem 27:2591– 2598. 3. Stephan CE, Mount DI, Hansen DJ, Gentile JH, Chapman GA, Brungs WA. 1985. Guidelines for deriving numerical National Water Quality Criteria for the protection of aquatic organisms and their uses. NTIS PB– 85-227049. Technical Report. National Technical Information Services. Springfield, VA, USA. 4. Newman MC, Ownby DR, Mezin LCA, Powell DC, Christensen TRL, Lerberg SB, Anderson B-A. 2000. Applying species-sensitivity distributions in ecological risk assessment: Assumptions of distribution type and sufficient numbers of species. Environ Toxicol Chem 19:508–515. 5. Wheeler JR, Grist EPM, Leung KMY, Morritt D, Crane M. 2002. Species sensitivity distributions: Data and model choice. Mar Pollut Bull 45:192–202. 6. Suter GW, Traas TP, Posthuma L. 2002. Issues and practices in the derivation and use of species sensitivity distributions. In Posthuma L, Suter GW, Traas TP, eds, Species Sensitivity Distributions in Ecotoxicology. CRC Press, Boca Raton, FL, USA, pp 437-474. 7. Raimondo S, Jackson CR, Barron MG. 2010. Influence of taxonomic relatedness and chemical mode of action in acute interspecies estimation models for aquatic species. Environ Sci Technol 44:7711–7716. 8. Dyer SD, Versteeg DJ, Belanger SE, Chaney JG, Raimondo S, Barron MG. 2008. Comparison of species sensitivity distributions derived from interspecies correlation models to distributions used to derive water quality criteria. Environ Sci Technol 42:3076–3083. 9. Awkerman JA, Raimondo S, Barron MG. 2008. Development of species sensitivity distributions for wildlife using interspecies toxicity correlation models. Environ Sci Technol 42:3447–3452. 10. Awkerman JA, Raimondo S, Barron MG. 2009. Estimation of wildlife hazard levels using interspecies correlation models and standard laboratory rodent toxicity data. J Toxicol Env Heal A, 72:1604–1609. 11. Golsteijn L, Hendriks HWM, van Zelm R, Ragas AMJ, Huijbregts MAJ. 2012. Do interspecies correlation estimations increase the reliability of toxicity estimates for wildlife? Ecotox Environ Safe 80:238–243.
Uncertainty in extrapolation-augmented SSDs 12. Hose GC, Van den Brink PJ. 2004. Confirming the species-sensitivity distribution concept for endosulfan using laboratory, mesocosm, and field data. Arch Environ Contam Toxicol 47:511–520. 13. Craig PS, Hickey GL, Luttik R, Hart A. 2011. Species nonexchangeability inprobabilistic ecotoxicological risk assessment. J Roy Stat Soc A Sta 175:243–262. 14. Forbes VE, Calow P. 2002. Species sensitivity distributions revisited: A critical appraisal. Hum Ecol Risk Assess 8:473–492. 15. Insightful Corporation. 2001. S-plus Programmer’s Guide. Seattle, WA, USA. 16. Raimondo S, Vivian D, Barron MG. 2007. Web-Based Interspecies Correlation Estimation (Web-ICE) for Acute Toxicity: User Manual, Ver 3.1. Technical Report. EPA 600/R-12/603. US Environmental Protection Agency, Gulf Breeze, FL. 17. Burnham KP, Anderson DR. 2002. Model Selection and Multimodel Inference: A practical Information-Theoretic Approach, 2nd ed, Springer –Verlag, New York, USA.
Environ Toxicol Chem 33, 2014
695
18. Anderson DR. 2008. Model Based Inference in the Life Sciences: A Primer on Evidence. Springer, New York, NY, USA. 19. R Development Core Team. 2010. R: A Language and Environment for Statistical Computing. [cited 2013 November 25]. Available from: http://www.R-project.org. 20. Van Straalen NM. 1994. Biodiversity of ecotoxicological responses in animals. Netherlands J Zool 441/442:112–129. 21. Raimondo S, Vivian DN, Delos C, Barron MG. 2008. Protectiveness of species sensitivity distribution hazard concentrations for acute toxicity used in endangered species risk assessment. Environ Toxicol Chem 27:2599–2607. 22. Maltby L, Blake N, Brock TCM, Van den Brink PJ. 2005. Insecticide species sensitivity distributions: Importance of test species selection and relevance to aquatic ecosystems. Environ Toxicol Chem 24:379– 388. 23. Baird DJ, Van den Brink PJ. 2007. Using biological traits to predict species sensitivity to toxic substances. Ecotox Environ Safe 67:296–301.
Copyright of Environmental Toxicology & Chemistry is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
Hazard/Risk Assessment AUGMENTING AQUATIC SPECIES SENSITIVITY DISTRIBUTIONS WITH INTERSPECIES TOXICITY ESTIMATION MODELS JILL A. AWKERMAN,* SANDY RAIMONDO, CRYSTAL R. JACKSON, and MACE G. BARRON Gulf Ecology Division, US Environmental Protection Agency, Gulf Breeze, Florida (Submitted 13 March 2013; Returned for Revision 10 April 2013; Accepted 1 November 2013) Abstract: Species sensitivity distributions (SSDs) are cumulative distribution functions of species toxicity values. The SSD approach is being used increasingly in ecological risk assessment but is often limited by available toxicity data needed for diverse species representation. In the present study, the authors evaluate augmenting aquatic species databases limited to standard test species using toxicity values extrapolated from interspecies correlation estimation (ICE) models for SSD development. The authors compared hazard concentrations at the 5th centile (HC5) of SSDs developed using limited measured data augmented with ICE toxicity values (augmented SSDs) with those estimated using larger measured toxicity datasets of diverse species (reference SSDs). When SSDs had similar species composition to the reference SSDs, 0.76 of the HC5 estimates were closer to the reference HC5; however, the proportion of augmented HC5s that were within 5-fold of the reference HC5s was 0.94, compared with 0.96 when predicted SSDs had random species assemblages. The range of toxicity values among represented species in all SSDs also depended on a chemical’s mode of action. Predicted HC5 estimations for acetylcholinesterase inhibitors showed the greatest discrepancies from the reference HC5 when SSDs were limited to commonly tested species. The results of the present study indicate that ICE models used to augment datasets for SSDs do not greatly affect HC5 uncertainty. Uncertainty analysis of risk assessments using SSD hazard concentrations should address species composition, especially for chemicals with known taxa-specific differences in toxicological effects. Environ Toxicol Chem 2014;33:688–695. # 2013 SETAC. This article is a US Government work and is in the public domain in the USA. Keywords: Aquatic toxicology
Species sensitivity distributions
Species extrapolation
Risk assessment
tradeoff of increasing the sample size while introducing uncertainty in the toxicity prediction has not been evaluated for aquatic species. For datasets with limited experimental data, extrapolated toxicity values can increase sample sizes while reducing statistical uncertainty in mammalian and bird species [11]. In the present study, we examine the utility of augmenting SSDs for aquatic species with ICE-estimated toxicity values. We compare HC5 estimation using minimal datasets of commonly tested species’ toxicity data without extrapolated values to those supplemented with ICE-estimated values. Meaningful application of SSDs for ecological risk assessment depends on identifying assumptions and establishing guidelines for developing SSDs. One assumption of the SSD is that species are a random assemblage of interchangeable species; indeed, this is seldom the case when using available data. Commonly tested toxicological species can be less sensitive than other species with a similar size and other characteristics [10]. Although taxonomic relatedness is a good indication of representative toxicity values for a species [7], even across large geographic distances [12], all species are not interchangeable within the SSD [13]. The assemblage of species can affect HC5 estimation, such that large sample sizes do not necessarily produce consistent toxicity ranges for modeling SSDs [14]. The present study explores the potential tradeoff between increasing the sample size of an SSD and propagating error by including extrapolated toxicity values. We determine the variability of HC5s estimated with inclusion of ICE-extrapolated toxicity data within an SSD. We examine the effect on HC5 of including extrapolated toxicity values in SSDs to increase taxonomic diversity and sample size. Through a comparative modeling approach, we identify SSD characteristics that contribute to uncertainty in estimating HC5, and these results offer general rules to consider when using extrapolated data in
INTRODUCTION
Species sensitivity distributions (SSDs) are used in ecological risk assessments to model the cumulative proportion of species’ sensitivity to a specific compound or environmental stressor [1,2]. In application, commonly used hazard levels of SSDs estimate the concentration affecting the specified proportion of the species (e.g., the hazard concentration that affects 5% of included species [HC5]). The SSD approach uses a set of toxicity data that represents species assemblages or various taxonomic groups of interest; notably, however, available toxicity data are often limited to a small number of commonly tested species. Consequently, statistical or regulatory criteria, such as minimum sample size or minimum data requirements for taxonomic diversity are rarely met [3–5], particularly for new or emerging compounds. Inclusion of extrapolated toxicity values offers a means to increase both sample size and taxa diversity in SSDs. However, extrapolated values with toxicity estimation error subsequently will add uncertainty to estimates of HC5 values [6]. Interspecies correlation estimation (ICE) models are log-linear least squares regressions of the sensitivity relationship between 2 species and use toxicity values of species with test data to estimate the toxicity to species without data for a chemical of interest [7]. Estimates of HC5 from SSDs of only ICE-estimated toxicities have been shown to be comparable to those from SSDs with only measured toxicity data for both aquatic [8] and terrestrial species [9,10]. Interspecies correlation estimation models have been used to augment SSDs for wildlife [11], but the statistical All Supplemental Data may be found in the online version of this article. * Address correspondence to [email protected]. Published online 8 November 2013 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/etc.2456 688
Uncertainty in extrapolation-augmented SSDs
SSDs. We also describe the effect of error in extrapolating single toxicity values on HC5 estimation through a local sensitivity analysis. We use both measured and extrapolated aquatic species toxicity data to understand the key sources of uncertainty in HC5 estimation. An additional goal is to identify the proportion of extrapolation data in an augmented dataset that optimizes taxa diversity while minimizing statistical uncertainty of the SSD. In doing so, we provide guidelines for applying ICE models when augmenting empirical datasets for ecological risk assessments. METHODS
Development of reference species sensitivity distributions
All analyses used toxicity data and models from the Web-ICE modeling application (http://epa.gov/ceampubl/fchain/webice/; [15]). Interspecies correlation estimation models are log-linear least squares regressions between 2 species that have been tested for the same chemicals
Environ Toxicol Chem 33, 2014
689
species. If multiple core surrogate models were available for a predicted species, we used the model with the lowest mean square error to estimate toxicity and added this value to the SSD dataset. Extrapolated toxicity values used to augment minimal datasets were restricted to the same species composition as the reference SSD. Because ICE models were not always available for every species in the reference SSD, augmented SSDs sometimes had fewer species. The augmented SSDs had from 1 to 29 more species than SSDs developed from the minimal dataset alone. Fold difference from the reference HC5 (HC5ref) was calculated for both the minimal toxicity dataset and the extrapolation-augmented dataset as follows ðHC5x HC5ref Þ HC5ref where HC5x is the hazard concentration derived from SSDs using either minimal or augmented datasets. SSD augmentation with variable species composition
log10 ðpredicted toxicityÞ ¼ slope log10 ðsurrogate toxicityÞ þ intercept The model is used to estimate toxicity to the predicted species from the surrogate when the toxicity for the latter is known. We used a subset of the Web-ICE database for 49 chemicals with experimental acute toxicity values (median lethal concentration [LC50] and median effective concentration [EC50] in the case of invertebrates with an endpoint equivalent to ecological mortality) for 10 or more species of aquatic invertebrates and vertebrates. The Web-ICE database follows strict data standardization criteria, which are described in detail by Raimondo et al. [7] to reduce extraneous sources of variability in estimating toxicity. For each of these chemicals, reference SSDs were created from all available experimental data (Supplemental Data, Table S1). Species sensitivity distributions were developed using S-plus [16] to fit a log-logistic cumulative distribution function fit to toxicity data as Y ¼ 1=f1 þ exp½ða XÞ=bg Maximum likelihood estimation was used to derive the intercept, a, and scale parameter, b, to toxicity values X at cumulative proportion Y. Model fit was evaluated by a Kolmogorov Smirnov goodness of fit test. The 5th centile of this distribution was used in all analyses below as the reference HC5. SSD augmentation with same species composition
For each chemical, we developed an SSD using a subset of commonly tested core species as an initial minimal dataset. This minimal SSD represents the scenario in which only standard test species data are available. These commonly tested core species were Americamysis bahia, Chironomus plumosus, Crassostrea virginica, Daphnia magna, Gammarus fasciatus, Ictalurus punctatus, Lepomis macrochirus, Micropterus salmoides, Oncorhynchus clarkii, Oncorhynchus mykiss, and Pimephales promelas. Toxicity data for all of these species were not available for all chemicals; therefore, SSDs were developed using as many of these species as possible, with a minimum of 5 (Table 1) [4]. Next, we augmented the minimal dataset for each chemical with extrapolated toxicity values calculated from ICE models. We used only the core species included in the minimum dataset for each chemical as surrogates to predict toxicity to other
To examine the uncertainty of HC5 estimation as it relates to sample size and species composition, extrapolated toxicity values were added to the minimal dataset as described in the section SSD augmentation with same species composition but without restricting species composition to that of the reference SSD. In addition to creating SSDs with as many extrapolated toxicity values as possible, we replicated augmented SSDs for each chemical by adding randomly chosen extrapolated toxicity values to all available core toxicity values (from 3 to 9 depending on data availability). Combinations of core and extrapolated values were developed for sample sizes of up to 30 species, with 10 replicate SSDs per sample size per chemical. Although ICE models were available to extrapolate toxicity for 61 to 74 species for each chemical, we limited our replicate SSDs to a maximum of 30 species to maintain variability in the subset data used in SSDs and reduce possible pseudoreplication as a result of our replicate subsampling approach. All SSDs were fit to toxicity data as described in the section Development of reference species sensitivity distributions. For each chemical, HC5 was estimated for all replicate SSDs, and the fold difference from the reference HC5 was determined. Using a multimodel inference approach, we compared a candidate list of linear mixed models that included various attributes of the replicate augmented SSDs as independent variables and the absolute fold difference of HC5s estimated from augmented and reference SSDs as the dependent variable. This approach ranks the candidate models using Akaike’s Information Criterion corrected for small sample sizes (AICc) to find the best model fit [17,18]—in this case, the combination of SSD attributes that best accounted for variability of HC5 relative to the reference HC5. Models fit to the log-transformed absolute fold difference in HC5s included single effects or interactions of the following independent variables for each extrapolationaugmented SSD: a measure of taxonomic diversity represented as the number of taxonomic orders relative to number of species in the SSD (DIVERSE); the proportion of vertebrates (VERT); extrapolation SSD goodness of fit as represented by pseudo R2 (R2); the broad mode of action of the chemical (MOA); and the proportion of extrapolated values in the SSD (EXTRAP). Random effects of individual chemical datasets were also included in each model (CHEM). For each model, we present AICc, AICc difference, log likelihood, and Akaike weight. Models were developed using package lme4 [19]. Each model was based on a priori expectations about variables affecting the
690
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
Table 1. Mode of action (MOA), hazard concentrations at the 5th centile (HC5) values, and fold difference (FD) between reference and augmented species sensitivity distributions (SSDs) including all available toxicity data (reference HC5) for 49 chemicalsa Specific MOA
Broad MOA
Chemical
No. ref
Reference HC5 (mg/L)
N Coreb
Core FD
No. Same Extr
Same Extr FD
No. Full Extr
Full Extr FD
13 10 12 12 10 14 11 38 17 11 15 14 15 18 23 11 16 24 31 17 22 11 13 13 11 10 16 13 10 29 18 10 21 25 12 10 22 10 23 11
7.7 18 815 6.2 328 3298 91 6398 103 8911 17 77 42 0.27 0.28 0.02 0.15 1.4 0.06 1.2 1.6 1.9 6.0 0.56 0.11 1.2 6.2 0.04 2.3 0.18 0.66 5.9 3.2 0.07 0.83 0.79 0.22 1.2 0.03 0.2 0
7 6 6 7 5 — — 8 5 — 7 8 — 9 8 6 8 6 8 6 9 9 8 8 6 8 7 5 6 7 5 — 6 7 5 5 7 5 6 —
–0.26 0.01 –0.41 –0.31 –0.53 — — –0.91 –0.02 — 0.22 0.13 — 4.59 0.67 20.87 –0.9 –0.74 –0.71 –0.22 –0.91 –0.07 0.05 4.9 –0.92 0.19 1.37 –0.96 1.61 –0.26 0.81 — 2.26 2.67 7.74 –0.63 1.29 –0.34 –0.39 —
12 9 12 11 8 13 11 37 17 10 15 13 12 18 21 11 14 20 28 16 19 10 11 13 11 10 13 10 10 27 17 7 17 23 11 9 20 7 20 8
–0.29 –0.08 –0.69 –0.71 –0.12 –0.02 –0.22 –0.38 –0.96 0.19 –0.18 0.22 0.15 0.13 0.69 6.96 –0.07 0.04 1.01 –0.08 0.15 0.36 0.54 –0.51 0.14 –0.25 –0.61 0.48 1.5 –0.37 0.04 –0.08 0.75 0 –0.04 –0.8 –0.29 –0.69 1.25 5.43
71 78 71 70 70 66 76 79 64 73 76 76 62 77 77 71 79 72 80 71 77 77 72 77 71 71 72 75 70 71 63 70 71 72 63 75 77 69 69 43
0.63 –0.9 –0.61 –0.41 –0.8 –0.78 –0.91 –0.95 –0.98 –0.32 –0.8 –0.49 66.19 2.68 8.09 0.12 –0.74 –0.23 –0.64 1.22 –0.89 1.62 0.26 –0.15 0.64 0.83 1.65 –0.99 2.5 –0.6 –0.66 –0.87 –0.74 0.42 0.55 –0.93 –0.7 0.02 –0.39 0.09
0.52
72
–0.58
Narcosis
Nonpolar
Narcosis
Polar
AChE inhibitor
Carbamates
AChE inhibitor
OP
Neurotoxicant
OC
Neurotoxicant
Pyrethroid
Oxidative Phosphorylation
Inhibitor
2,4-Dichlorophenoxy acetic acid Captan Diuron Molinate 4-Nonylphenol Phenol Carbaryl Mexacarbate Carbofuran Aminocarb Methomyl Trichlorfon Fenthion Parathion Coumaphos Dichlorvos Azinphos-Methyl Malathion Fenitrothion Methyl parathion Disulfoton Naled Ethion Phosmet ENP Chlorpyrifos Temephos Profenofos DDT Lindane Dieldrin Endrin Methoxychlor Heptachlor Endosulfan Toxaphene Fenvalerate Permethrin Antimycin A
16
14.3
5
2.6
14
Oxidative Phosphorylation Other Other
Uncoupler
20 ,50 -Dichloro-40 nitrosalicylanilide compd. with 2-aminoethanol (1:1) Pentachlorophenol
29
6.8
7
0.5
23
–0.3
75
–0.68
Other
Respiratory toxicity
Ammonia Mercury Cadmium Copper Zinc Chromium VI Sodium chloride
56 12 33 42 34 14 17
5251 1.8 1.4 4.9 81.9 270 1 741 225
7 — — 6 — — —
1.99 — — 0.78 — — —
27 10 20 30 19 11 9
–0.27 0.49 –0.58 0.51 –0.63 –0.51 –0.04
82 67 77 78 71 60 76
–0.88 –0.6 –0.95 0.4 –0.8 –0.94 –0.98
Ammonia stress Metallic stress
a
Results for comparison include fold difference between the reference HC5 and HC5s using minimal data (Core FD), HC5 for extrapolation-augmented SSDs limited to the same species as reference (Same Extr FD) and HC5 for SSDs augmented with all extrapolation models (Full Extr FD). b Core datasets with fewer than 5 species were not used to develop SSD. Sample size is listed for each (No. Ref, No. Core, No. Same Extr, and No. Full Extr). AChE ¼ acetylcholinesterase; OP ¼ organophosphate; OC ¼ organochlorine.
HC5 estimation (taxonomic representation, model fit, chemical type, and extrapolation uncertainty [1,4,5,7]). The taxonomy of the aquatic species included in the SSDs was classified according to the Integrated Taxonomic Information System. We used a likelihood ratio test to determine if significant differences in model fit existed between 2 top-ranking models.
Sensitivity of HC5 to individual species toxicity values
A sensitivity analysis determined the influence of toxicity estimation error on the HC5. For each toxicity value in each of the 49 SSDs, we altered the value in 16 iterations within the range of 50% to 200% of the measured data (e.g., a toxicity value of 2 within the SSD would be exchanged with values ranging
Uncertainty in extrapolation-augmented SSDs
from 1 to 4). Estimates of HC5s were determined for SSDs created with each iteration of each data point. Finally, the fold difference between the reference HC5 and the manipulated HC5s determined for all iterations was calculated. RESULTS
Hazard dose estimation with same species composition
Surrogate data availability varied by chemical and MOA. Those SSDs that included only minimal core toxicity datasets ranged in sample size from 5 to 9 species (Table 1). The proportion of minimal core datasets that produced HC5s within 5-fold and 10-fold of the reference HC5 was 0.95 and 0.97, respectively, for 38 chemicals. For all 49 chemicals, including extrapolated toxicity values in SSDs limited to species in the reference SSD increased the sample sizes to 9 to 37 species. All same-species HC5 estimates were within 10-fold, and 0.96 were within 5-fold. Hazard dose estimation with variable species composition
The proportion of HC5s from SSDs augmented with all possible extrapolated values that were within 5-fold and 10-fold of the reference HC5 were 0.96 and 0.98, respectively (Figure 1). The proportion of chemicals with a smaller fold difference from the reference HC5 using a full complement of extrapolated values compared to the minimal core dataset was 0.45 (Table 1). Including all possible species in the extrapolated SSDs reduced fold difference from the reference HC5 for only 0.22 of the 49 chemicals when compared with extrapolated SSDs that were limited to the same species composition as reference SSDs. The HC5 and lower 95% confidence interval of the HC5 (HCL) estimates were lower than core data estimates for 23 of the 38 chemicals (0.62; Table 2). Organophosphates did not produce more conservative HC5 estimates with greater species representation for 12 of the 17 chemicals (0.71). Among the 12 740 SSDs produced with various quantities of extrapolated toxicity values, fold difference from the reference HC5 depended on chemical, sample size, species composition, and proportion of extrapolated values included. Estimates of HC5 for all augmented SSDs were within 10-fold of the reference HC5 for 98% of the SSDs developed, and 96% of the SSDs produced HC5 estimates within 5-fold of the reference SSD. The 2 top-ranking mixed effects models in our model set included an interaction effect between the proportion of vertebrates in the SSD and the MOA being modeled, as well as the random effect of the chemical (Figure 2, Table 3). The 2 models with a substantial model weight differed only in the inclusion of pseudo R2. A likelihood ratio test comparison of the top 2 models in our model set indicates no significant difference between them, suggesting that model fit has minimal influence on HC5 estimation (x2 ¼ 0.92, p ¼ 0.34). Fold difference from the reference HC5 is similar for proportions of the ICE models (Figure 1). Sensitivity of HC5 to individual species toxicity values
Variability in the toxicity of a single data point affected the HC5 differently depending on error bias and the sensitivity of the species. Underestimating toxicity values in the lower portion of the distribution increases the HC5 estimates, and these points decrease HC5 estimates when they are overestimated (Figure 3). Conversely, the toxicity values in the upper portion of the distribution decrease HC5 estimates when they are underestimated and increase HC5 estimates when they are overestimated.
Environ Toxicol Chem 33, 2014
691
Estimation error of species in the lower quartile of the distribution has a relatively greater influence on HC5 estimates, and error in a single toxicity estimate affects HC5 estimation more with fewer species in the SSD (Figure 4). DISCUSSION
The data needed to ensure SSDs represent diverse species assemblages and maintain a statistically robust sample size are difficult to obtain from existing toxicological data, which often represent only relatively few commonly tested species per chemical. Interspecies estimation models offer a way to supplement available data with extrapolated toxicity values to allow generating SSDs with broader taxonomic diversity and larger sample sizes. However, introducing uncertainty with extrapolated toxicity values is a concern when using extrapolation models to supplement measured data to estimate hazard dose levels. The current study’s simulations demonstrate that uncertainty associated with toxicity extrapolation has only a minimal effect on estimating HC5. Species sensitivity distributions are of greater utility in ecological risk assessment when their species composition includes several taxonomic classes, representing different functional groups of an ecological community. Large and diverse assemblages of species, however, produce variable ranges of toxicological effects that can differ depending on the chemical, according to physiology, behavior, and life history [14,20]. The taxonomic composition of SSDs affects HC5 estimates for aquatic species more than differences in geography or habitat of the assemblage [12,21]. Identifying biological and statistical variables that contribute to HC5 variance can help reduce uncertainty in risk assessments through the development of criteria for data selection and inclusion. The proportion of extrapolated values that augmented measured toxicity data was not identified as an influential parameter in HC5 estimation in the top models within our model set. Species composition is influential in HC5 estimation and should be representative of taxonomic groups or aquatic assemblages of interest, considering the variable impacts of some types of chemicals. The interaction between MOA and proportion of vertebrates in our top models further identifies the importance of species composition, particularly when discrepancies in sensitivity among taxonomic groups are known for different MOAs [22]. The MOA-specific differences in HC5 estimation demonstrate chemical-specific concerns that should be considered when developing and evaluating SSDs. Recognizing variable MOA effects emphasizes the need to include relevant taxa within the species composition of the SSD, ensuring that HC5s protect the most sensitive taxa. Risk assessment criteria specifying a minimum number of genera or a diverse group of representative species might be important for chemicals without a known discrepancy in effects among constituent taxonomic groups represented. For example, aquatic life criteria require 8 minimum data requirements aimed at including taxonomic diversity in SSDs [3]. The 49 chemicals selected for the present study were chosen because they had been tested on the greatest number of species. However, few chemicals met the requirements of taxa diversity specified by aquatic life criteria [3]. For chemicals tested on only a few species, the diversity of species models in the Web-ICE database allows broader taxonomic representation. The current ICE model set contains a higher proportion of vertebrate models. Fish are overrepresented in most toxicological databases and insects are often underrepresented, such that available toxicity data rarely represent true
692
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
Table 2. Hazard concentrations at the 5th centile (HC5) and lower 95% confidence intervals of HC5 (HCL) for 49 chemicals using minimal data (Core), extrapolation-augmented species sensitivity distributions (SSDs) limited to the same species as reference (Same Extr) and SSDs augmented with all extrapolation models (Full Extr)
Specific MOA Nonpolar
Polar Carbamates
OP
OC
Pyrethroid Inhibitor
Uncoupler Ammonia stress Metallic stress
Respiratory toxicity
Chemical 2,4-Dichlorophenoxy acetic acid Captan Diuron Molinate 4-Nonylphenol Phenol Carbaryl Mexacarbate Carbofuran Aminocarb Methomyl Trichlorfon Fenthion Parathion Coumaphos Dichlorvos Azinphos-methyl Malathion Fenitrothion Methyl parathion Disulfoton Naled Ethion Phosmet ENP Chlorpyrifos Temephos Profenofos DDT Lindane Dieldrin Endrin Methoxychlor Heptachlor Endosulfan Toxaphene Fenvalerate Permethrin Antimycin A 20 ,50 -Dichloro-40 nitrosalicylanilide compd. with 2-aminoethanol (1:1) Pentachlorophenol Ammonia Mercury Cadmium Copper Zinc Chromium VI Sodium chloride
Core HCL (mg/L)
Core HC5 (mg/L)
No. Same Extr
Same Extr HCL (mg/L)
Same Extr HC5 (mg/L)
No. Full Extr
Full Extr HCL (mg/L)
Full Extr HC5 (mg/L)
7 8 — 9 8 6 8 6 8 6 9 9 8 8 6 8 7 5 6 7 5
0.19 2340 0.02 17 113 — — 0.07 5230 — 2.09 3.2 — 0.019 2.5 10–5 0.0012 6.3 10–5 1.7 10–5 1.3 10–3 1.05 10–3 2.3 10–4 0.21 4.5 10–3 3.110–3 2.8 10–5 0.63 4.9 10–4 5.2 10–8 9.8 10–3 0.041 2.2
5.7 19 096 3.7 226 1565 — — 9.3 8703 — 95 47 — 1.6 0.039 3.4 0.14 0.015 0.33 1.3 0.17 5.6 0.58 0.63 0.093 7.3 0.10 0.080 0.47 0.49 10.7
6 7 5 5 7 5 6 — 5
0.060 0.41 1.4 7.5 10–4 1.2 6.5 10–5 2.5 10–3 — 9.1
0.22 3.03 6.9 0.082 2.8 0.017 0.12 — 51
12 9 12 11 8 13 11 37 17 10 15 13 12 18 21 11 14 20 28 16 19 10 11 13 11 10 13 10 10 27 17 7 17 23 11 9 20 7 20 8 14
0.85 3687 0.11 13 872 55 1466 21 97 5.2 8.9 10 0.0084 0.018 1.6 10–3 0.034 0.029 3.3 10–3 0.32 0.090 0.13 0.43 0.036 8.510–4 0.043 0.48 3.6 10–4 0.022 0.032 0.18 2.5 0.54 0.043 0.24 0.12 1.0 10–3 0.39 1.1 10–4 0.15 1.0 10–3 7.5
5.4 17 222 1.9 95 2917 89 5014 63 352 21 64 51 0.31 0.32 0.039 1.2 1.3 0.059 2.3 1.5 2.1 8.1 0.85 0.052 1.4 4.6 0.017 3.3 0.45 0.41 6.1 2.9 0.12 0.83 0.76 0.045 0.87 0.008 0.45 0.027 22
71 78 71 70 70 66 76 79 64 73 76 76 62 77 77 71 79 72 80 71 77 77 72 77 71 71 72 75 70 71 63 70 71 72 63 75 77 69 69 43 72
4.5 661 0.43 75 280 10 239 1.2 72 4.7 3.9 7.8 5.4 0.17 0.031 0.028 0.047 7.4 10–3 0.066 0.88 0.022 5.2 0.16 0.011 0.48 5.5 0.022 6.2 10–4 0.25 0.10 0.90 0.12 5.6 10–3 0.51 0.54 2.6 10–3 0.15 8.5 10–3 0.043 1.0 10–3 2.3
13 1965 2.4 193 655 20 557 5.5 187 12 15 21 18 1.04 0.21 0.17 0.35 0.044 0.42 3.6 0.21 16 0.70 0.091 2.0 11 0.11 0.013 0.63 0.26 2.02 0.42 0.017 1.18 1.22 0.02 0.37 0.027 0.12 4.6 10–3 6.03
7 7 —
1.3 9722 —
10 15 694 —
6 — — —
0.55 — — —
8.7 — — —
23 27 10 20 30 19 11 9
1.6 1943 0.23 0.045 3.1 5.07 9.03 630 649
4.8 3842 2.7 0.57 7.4 29.9 132 1 675 747
75 82 67 77 78 71 60 76
0.64 212 0.18 7.8 10–3 2.35 5.43 3.37 6263
2.2 617 0.72 0.066 6.8 17 17 30 068
No. Core 7 6 6 7 5 — — 8 5
MOA ¼ mode of action; OP ¼ organophosphate; OC ¼ organochlorine.
ecological communities [23]. Inclusion of a greater proportion of vertebrates in most SSDs can increase the range of toxicity values with more insensitive species. For chemicals that are more toxic to invertebrates, species composition of the SSD will have a greater effect on hazard concentration estimations and risk assessment. The uncertainty associated with ICE toxicity estimation is quantified for each value as modeled confidence interval around the prediction, and some degree of professional judgement is required to evaluate these values [15]. Despite recommendations
presented here on the importance of SSD species composition, recognizing where large uncertainty exists around an ICE toxicity value and the relative sensitivity of that species are also important considerations when using extrapolated values to augment datasets. The sensitivity analysis identified relatively greater influence on HC5 variability of a single toxicity value when species were in the more sensitive range of the SSD. Inclusion of models with large confidence intervals for extrapolated values that lie in the lower quartile of the SSD is not recommended, particularly when datasets are small.
Uncertainty in extrapolation-augmented SSDs
Environ Toxicol Chem 33, 2014
693
Table 3. Candidate model set evaluating the absolute fold-difference between hazard concentrations at the 5th centile (HC5) estimates of extrapolation-augmented species sensitivity distributions (SSDa) and HC5 estimates using measured toxicity values Modelb VERT MOA þ CHEM VERT MOA þ R2 þ CHEM VERT MOA þ R2 þ EXTRAP þ CHEM VERT MOA þ R2 þ SPECIES þ CHEM VERT þ CHEM VERT þ EXTRAP þ R2 þ CHEM VERT þ SPECIES þ R2 þ CHEM R2 þ CHEM DIVERSE þ CHEM EXTRAP þ CHEM DIVERSE þ R2 þ CHEM EXTRAP þ R2 þ CHEM EXTRAP þ DIVERSE þ CHEM DIVERSE þ EXTRAP þ R2 þ CHEM SPECIES þ CHEM SPECIES þ DIVERSE þ CHEM SPECIES þ R2 þ CHEM DIVERSE þ SPECIES þ R2 þ CHEM MOA þ CHEM SPECIES þ EXTRAP þ CHEM DIVERSE MOA þ CHEM DIVERSE MOA þ R2 þ CHEM DIVERSE MOA þ R2 þ EXTRAP þ CHEM DIVERSE MOA þ R2 þ SPECIES þ CHEM
log likelihood
AICc
DAICc
AICc weights
–6463.65 –6463.19 –6465.36 –6469.82 –6725.78 –6727.46 –6732.42 –6827.91 –6829.62 –6830.03 –6829.28 –6830.22 –6830.29 –6830.50 –6834.13 –6833.29 –6834.48 –6833.69 –6826.90 –6835.11 –6819.35 –6819.27 –6820.32 –6823.46
12 975.39 12 976.48 12 982.83 12 991.75 13 459.56 13 466.92 13 476.85 13 663.82 13 667.25 13 668.07 13 668.56 13 670.45 13 670.59 13 673.00 13 676.25 13 676.59 13 678.96 13 679.40 13 679.83 13 680.23 13 686.80 13 688.64 13 692.75 13 699.03
0.00 1.09 7.44 16.36 484.17 491.53 501.46 688.43 691.86 692.68 693.17 695.07 695.20 697.61 700.86 701.20 703.57 704.01 704.44 704.84 711.41 713.25 717.36 723.64
0.62 0.36 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
a
Extrapolation-augmented SSDs were developed with 10 replicated subsets of up to 30 randomly chosen species. The variables included in models were sample size (SPECIES), the proportion of extrapolated values (EXTRAP), the proportion of vertebrates in the SSD (VERT), a measure of taxonomic diversity that included the number of orders represented per species (DIVERSE), the mode of action of a chemical (MOA), and the SSD fit (R2). Each model also included random effects of chemical (CHEM), which determined surrogate and models available for subset SSD included in analyses. Models are presented in ranked order of lowest Akaike’s Information Criterion corrected for small sample sizes (AICc) and DAICc. b
However, differences in HC5 estimation were relatively small within the ranges of toxicity characterized in the present study. The sensitivity analysis also demonstrated how the HC5 changes with uncertainty in the values in the upper and lower quartiles. Overestimating the toxicity values (or using the predicted upper confidence interval value) results in a lower, more conservative
HC5 value when the toxicity value is in the more sensitive range of the SSD. Conversely, applying the lower confidence interval of less sensitive toxicity values results in a more conservative hazard concentration. Including extrapolated values with larger uncertainty for species that lie in the median of the SSD has little effect on HC5 estimation. Using multiple surrogates to develop
Figure 1. Log-transformed fold difference from reference hazard concentrations at the 5th centile (HC5) with increasing proportion of extrapolated toxicity values included in the species sensitivity distribution. Boxes indicate the median within upper and lower quartiles, and whiskers denote 90% of estimates.
Figure 2. Variance in hazard concentrations at the 5th centile (HC5) estimation, as noted by the log fold difference, by mode of action of chemical for replicate species sensitivity distributions using various combinations of experimental and extrapolated toxicity values. Boxes indicate the median within upper and lower quartiles, and whiskers denote 90% of estimates.
694
Environ Toxicol Chem 33, 2014
J.A. Awkerman et al.
chemical. Hazard concentrations depend on the dataset with which the SSD was developed, a result underscored by the analyses presented here. Reference SSDs are representative of data available to a risk assessor for broadly tested chemicals. Comparing reference SSDs to smaller subsets of data, representing species composition more like that available for most chemicals, demonstrates that while minimal datasets often produce HC5 estimates similar to larger datasets (0.95 within 5-fold of reference), many minimal dataset HC5 estimations (0.61) are improved when extrapolated values are included. Variable results as sample sizes are maximally increased emphasize the importance of considering species composition, surrogate selection, and relative sensitivity of species when developing extrapolation-augmented SSDs for risk assessment. SUPPLEMENTAL DATA
Table S1. (31 KB DOCX).
Figure 3. Effect of underestimating (1) toxicity values on fold difference from reference hazard concentrations at the 5th centile (HC5s) ([predicted HC5-reference HC5]/reference HC5). Toxicity values are color-coded to represent their relative sensitivity by quartile within the species sensitivity distribution from sensitive to insensitive (black ¼ < 0.25; red ¼ > 0.25, 0.5, 0.75).
augmented SSD is recommended, when possible, to minimize uncertainty in the models selected. Previous studies also support inclusion of ICE models for which surrogate and predicted species are closely related to reduce uncertainty [8]. These general trends can help guide professional judgement for accepting extrapolated values to augment SSDs. The present study evaluated difference in HC5 from a reference HC5 derived using all available measured values within our database. Reference SSDs are not intended to represent real communities and are not necessarily intended to be interpreted as definitive hazard concentration values for a given
Figure 4. Diminishing effects of underestimated (green) or overestimated (yellow) single toxicity point, independent of location within the species sensitivity distribution (SSD), with increasing sample size of SSD. Fold difference from reference hazard concentrations at the 5th centile (HC5) is (predicted HC5-reference HC5)/reference HC5.
Acknowledgment—For database development and maintenance, we thank B. Montague (US Environmental Protection Agency [USEPA], Office of Pesticide Programs) and P. Mineau, A. Baril, and B. Collins (Environment Canada). We are also grateful to support personnel: D. Vivian, M. Marchetto, A. DiGirolamo, C. Chancy, N. Lemoine, N. Allard, and C. McGill (USEPA, Gulf Ecology Division). Two anonymous reviewers provided helpful guidance to an earlier draft. Support in preparing this manuscript was provided by the USEPA, National Health and Environmental Effects Research Laboratory. Although this manuscript was submitted for internal review within USEPA and approved for publication, it does not reflect the views of the agency or constitute an endorsement of any trade names or commercial products mentioned.
REFERENCES 1. Posthuma L, Traas TP, Suter GW. 2002. General introduction to species sensitivity distributions. In Posthuma L, Suter GW, Trass TP, eds, Species Sensitivity Distributions in Ecotoxicology. CRC Press, Boca Raton, FL, USA, pp 3–10. 2. De Vries P, Tamis JE, Murk AJ, Smit MGD. 2008. Development and application of a species sensitivity distribution for temperature-induced mortality in the aquatic environment. Environ Toxicol Chem 27:2591– 2598. 3. Stephan CE, Mount DI, Hansen DJ, Gentile JH, Chapman GA, Brungs WA. 1985. Guidelines for deriving numerical National Water Quality Criteria for the protection of aquatic organisms and their uses. NTIS PB– 85-227049. Technical Report. National Technical Information Services. Springfield, VA, USA. 4. Newman MC, Ownby DR, Mezin LCA, Powell DC, Christensen TRL, Lerberg SB, Anderson B-A. 2000. Applying species-sensitivity distributions in ecological risk assessment: Assumptions of distribution type and sufficient numbers of species. Environ Toxicol Chem 19:508–515. 5. Wheeler JR, Grist EPM, Leung KMY, Morritt D, Crane M. 2002. Species sensitivity distributions: Data and model choice. Mar Pollut Bull 45:192–202. 6. Suter GW, Traas TP, Posthuma L. 2002. Issues and practices in the derivation and use of species sensitivity distributions. In Posthuma L, Suter GW, Traas TP, eds, Species Sensitivity Distributions in Ecotoxicology. CRC Press, Boca Raton, FL, USA, pp 437-474. 7. Raimondo S, Jackson CR, Barron MG. 2010. Influence of taxonomic relatedness and chemical mode of action in acute interspecies estimation models for aquatic species. Environ Sci Technol 44:7711–7716. 8. Dyer SD, Versteeg DJ, Belanger SE, Chaney JG, Raimondo S, Barron MG. 2008. Comparison of species sensitivity distributions derived from interspecies correlation models to distributions used to derive water quality criteria. Environ Sci Technol 42:3076–3083. 9. Awkerman JA, Raimondo S, Barron MG. 2008. Development of species sensitivity distributions for wildlife using interspecies toxicity correlation models. Environ Sci Technol 42:3447–3452. 10. Awkerman JA, Raimondo S, Barron MG. 2009. Estimation of wildlife hazard levels using interspecies correlation models and standard laboratory rodent toxicity data. J Toxicol Env Heal A, 72:1604–1609. 11. Golsteijn L, Hendriks HWM, van Zelm R, Ragas AMJ, Huijbregts MAJ. 2012. Do interspecies correlation estimations increase the reliability of toxicity estimates for wildlife? Ecotox Environ Safe 80:238–243.
Uncertainty in extrapolation-augmented SSDs 12. Hose GC, Van den Brink PJ. 2004. Confirming the species-sensitivity distribution concept for endosulfan using laboratory, mesocosm, and field data. Arch Environ Contam Toxicol 47:511–520. 13. Craig PS, Hickey GL, Luttik R, Hart A. 2011. Species nonexchangeability inprobabilistic ecotoxicological risk assessment. J Roy Stat Soc A Sta 175:243–262. 14. Forbes VE, Calow P. 2002. Species sensitivity distributions revisited: A critical appraisal. Hum Ecol Risk Assess 8:473–492. 15. Insightful Corporation. 2001. S-plus Programmer’s Guide. Seattle, WA, USA. 16. Raimondo S, Vivian D, Barron MG. 2007. Web-Based Interspecies Correlation Estimation (Web-ICE) for Acute Toxicity: User Manual, Ver 3.1. Technical Report. EPA 600/R-12/603. US Environmental Protection Agency, Gulf Breeze, FL. 17. Burnham KP, Anderson DR. 2002. Model Selection and Multimodel Inference: A practical Information-Theoretic Approach, 2nd ed, Springer –Verlag, New York, USA.
Environ Toxicol Chem 33, 2014
695
18. Anderson DR. 2008. Model Based Inference in the Life Sciences: A Primer on Evidence. Springer, New York, NY, USA. 19. R Development Core Team. 2010. R: A Language and Environment for Statistical Computing. [cited 2013 November 25]. Available from: http://www.R-project.org. 20. Van Straalen NM. 1994. Biodiversity of ecotoxicological responses in animals. Netherlands J Zool 441/442:112–129. 21. Raimondo S, Vivian DN, Delos C, Barron MG. 2008. Protectiveness of species sensitivity distribution hazard concentrations for acute toxicity used in endangered species risk assessment. Environ Toxicol Chem 27:2599–2607. 22. Maltby L, Blake N, Brock TCM, Van den Brink PJ. 2005. Insecticide species sensitivity distributions: Importance of test species selection and relevance to aquatic ecosystems. Environ Toxicol Chem 24:379– 388. 23. Baird DJ, Van den Brink PJ. 2007. Using biological traits to predict species sensitivity to toxic substances. Ecotox Environ Safe 67:296–301.
Copyright of Environmental Toxicology & Chemistry is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
