1
2 BAYESIAN AND TIME-INDEPENDENT SPECIES SENSITIVITY
3 DISTRIBUTIONS FOR RISK ASSESSMENT OF CHEMICALS
4
5
6
7
8 ERIC P. M. GRIST§*, ANTHONY O'HAGAN‡, MARK CRANE+, NEAL
9 SOROKIN||, IAN SIMS|| and PAUL WHITEHOUSE±
10
§
11 CSIRO Marine and Atmospheric Research, GPO Box 1538, Hobart, Tasmania 7001,
12 Australia
13
‡
14 Department of Probability and Statistics, University of Sheffield, Hicks Building,
15 Sheffield, S3 7RH, UK
16
17 +Watts & Crane Associates, Faringdon, Oxfordshire, SN7 7AG, UK
18
||
19 WRc, Henley Road, Medmenham, Marlow, Buckinghamshire, SL7 2HD, UK
20
±
21 Environment Agency, Wallingford, Oxfordshire, OX10 8BD, UK
22
23
24
25 RUNNING TITLE: CHLORPYRIFOS SPECIES SENSITIVITY DISTRIBUTIONS
26
27
28
*
29 Author to whom correspondence should be addressed.
30 Email: eric.grist@csiro.au
31 tel +61 3 6232 5354.
32
1
33 Abstract
34
35 Species Sensitivity Distributions (SSDs) are increasingly used to analyse toxicity
36 data, but have been criticised for a lack of consistency in data inputs, lack of
37 relevance to the real environment and a lack of transparency in implementation. This
38 paper shows how the Bayesian approach addresses concerns arising from frequentist
39 SSD estimation. Bayesian methodologies are used to estimate SSDs and compare
40 results obtained with time-dependent (LC50) and time-independent (Predicted No
41 Observed Effect Concentration) endpoints for the insecticide chlorpyrifos.
42 Uncertainty in the estimation of each SSD is obtained either in the form of a point-
43 wise percentile confidence interval computed by bootstrap regression or an associated
44 credible interval. We demonstrate that uncertainty in SSD estimation can be reduced
45 by applying a Bayesian approach which incorporates expert knowledge, and that use
46 of Bayesian methodology permits estimation of an SSD which is more robust to
47 variations in data. The results suggest that even with sparse data sets, theoretical
48 criticisms of the SSD approach can be overcome.
49
50
51
52
53
54
55
56
57
58 Keywords - Species sensitivity distribution, risk assessment, chlorpyrifos, Bayesian
59 statistics, expert judgment, bootstrap regression, time-to-event analysis
2
60 INTRODUCTION
61
62 Species Sensitivity Distributions (SSDs) are increasingly employed in ecological risk
1-3
63 assessment procedures and are usually constructed by fitting a statistical
64 distribution, typically log-normal4 or log-logistic5, to toxicity data. One aim of an
65 SSD analysis is to determine a chemical concentration protective for most species in
66 the environment, usually by calculating an HC5 (Hazardous Concentration for 5% of
67 species)6. In Europe, risk assessment methods for new and existing chemicals are
68 described in the Technical Guidance Document (TGD)7. This includes a methodology
69 for the use of statistical extrapolation methods with SSDs, and there has also been
70 considerable interest in the use of SSDs to assess pesticide risks8.
71
72 Despite the increased use of SSDs, the approach potentially suffers from some
73 important flaws. For example, Forbes and Calow 9 have the following main criticisms,
74 which are also discussed in Posthuma et al.10:
75
76 1. SSDs are usually constructed from what they describe as a haphazard collection of
77 values that have no direct relevance to site-specific assemblages of organisms.
78 This may be a particular problem if site-specific exposure concentrations are then
79 compared with these generic SSDs.
80
81 2. The endpoints used to construct SSDs are not demographically relevant, i.e. they
82 are collections of lethal or sublethal threshold concentrations, such as median
83 lethal effects (LC50s) or sublethal no observed effect concentrations (NOECs).
84
85 3. The technical aspects of constructing an SSD, such as model choice, selection of
86 appropriate confidence intervals, definition of the minimum number, quality and
87 representativeness of data points required, and selection of summary statistics
88 such as the HC5 are often opaque and may not clearly relate to environmental
89 protection goals.
90
91 In this paper we begin to address these criticisms, using the organophosphorus
92 insecticide chlorpyrifos as a model. Chlorpyrifos is used worldwide to control pests in
93 arable crops and orchards, and in homes and gardens11. Chlorpyrifos toxicity is rapid
3
94 and intense for susceptible species, but relatively rapid degradation in the
95 environment means that cumulative toxicity is unlikely. A comprehensive ecological
96 risk assessment for chlorpyrifos has been performed for North American aquatic
97 environments10 and mesocosm data are also available, providing a useful comparison
98 for results in the present study.
99
100 The main objectives of this paper are to address the valid criticisms of SSDs by
101 Forbes and Calow9 by,
102
103 1. Collating available data on chlorpyrifos and selecting only those that are relevant
104 to UK species for inclusion in an SSD.
105
106 2. Eliciting information on organism tolerances to chlorpyrifos from UK field
107 biologists, to counteract bias in the toxicity database due to data being available
108 only for species which are suitable for laboratory tests.
109
110 3. Developing scenarios for three different aquatic habitat types containing different
111 generic organism assemblages, in a preliminary effort to understand differences
112 between site-specific assemblages.
113
114 4. Generating time-independent toxicity data that more closely relate to the
115 protection aims of the risk assessment process.
116
117 We apply both frequentist and Bayesian statistical approaches to construct SSDs and
118 associated confidence or credible intervals from time-independent PNOEC (Predicted
119 No Observed Effect Concentration) toxicity values, as well as the more usual 96-h
120 LC50 data.
121
122 First, we collated literature data on the toxicity of the insecticide chlorpyrifos and
123 estimated time-independent no effect concentrations. Then we elicited the opinions of
124 freshwater biologists on the relative sensitivities of different freshwater taxa to the
125 effects of this insecticide. Finally, species sensitivity distributions were constructed
126 with and without elicited opinions and compared with results from new toxicity
127 studies with previously untested species.
4
128
5
129 METHODS
130
131 LC50s from literature review
132
133 Acute lethal data for chlorpyrifos on the USEPA ECOTOX database
134 (http://www.epa.gov/ecotox/) were collated by Crane et al.12 and used in a simple
135 probabilistic risk assessment without regard to data quality or specific relevance to the
136 UK environment. The original papers were reviewed for the present study according
137 to the usual criteria for acceptability of data when setting UK Environmental Quality
138 Standards13. Those that contained information on species relevant to the UK
139 environment were selected because UK field biologists could not be expected to
140 possess expertise on the sensitivity to chlorpyrifos of non-native species during the
141 expert elicitation exercise described below. A total of 17 species for which 96-h
142 LC50s had been calculated were selected for use in SSD construction (Table 1).
143
144 [Table 1 here]
145
146 PNOECs based on time-to-event analysis
147 Raw survival data after 48-h and 96-h exposure to chlorpyrifos were kindly provided
148 by René van Wijngaarden14 of Alterra (Netherlands) for the following 8 species:
149 Chaoberus sp. (phantom midge larva), Cloeon dipterum (mayfly nymph), Corixa
150 punctata (water boatman), Simocephalus vetulus (waterflea), Daphnia longiseta
151 (waterflea) Gammarus pulex (shrimp) Gasterosteus aculeatus (3-spined stickleback),
152 and Pungitius pungitius (10-spined stickleback).
153
154 Mayer et al.15 describe a two-step linear regression method for estimating time-
155 independent PNOEC values from such raw survival data. We applied this approach to
156 estimate PNOEC values for each of the species in the van Wijngaarden raw survival
157 data. An inherent problem of the Mayer et al. technique is that it does not properly
158 account for the effects of individual variability whenever there is a decrease in
159 mortality with an increase in concentration. Thus, although mortality under any single
160 concentration can never decrease with time (the experiment is performed on one set of
6
161 individuals, with death being irreversible), it may sometimes decrease with increasing
162 concentration when measured at a fixed point in time because of the effects of
163 heterogeneity (individual variability in response). This problem occurred in the
164 PNOEC calculations for 1 species (Gasterosteus aculeatus) out of the 8 species in the
165 Wijngaarden raw survival data set. Biologically meaningful PNOECs were hence
166 calculable only for 7 species, hereafter referred to as the Wijngaarden PNOECs
167 (Table 2). These values were then subsequently used to construct time-independent
168 SSDs.
169
170 [Table 2 here]
171
172 Expert elicitation
173 The views of freshwater field biologists were sought on the sensitivity to chlorpyrifos
174 of 96 taxonomic groups found in UK freshwaters (Table 3). Seventeen biologists
175 employed by the Environment Agency or belonging to the Freshwater Biological
176 Association were asked to score the sensitivity to chlorpyrifos of each taxonomic
177 group on a scale from 1 (insensitive) to 8 (highly sensitive). They were also asked to
178 score their own knowledge of each taxonomic group from 0 (no knowledge) to 5
179 (high level of knowledge). The sensitivity scores for each taxon were then weighted
180 according to expertise. Experts all had the same information presented to them and
181 were given the same instructions for completing the exercise after a trial run with a
182 facilitator. The expert opinions on species sensitivities to chlorpyrifos, weighted
183 according to their assessment of individual knowledge, are shown in the final column
184 of Table 3.
185
186 [Table 3 here]
187
188 Generic aquatic assemblage scenarios
189 The taxonomic groups listed in Table 3 were associated with three generic UK
190 assemblages on the basis of information in Fitter and Manuel16. These assemblages
191 were a) a fast-flowing stream, b) a slow-flowing lowland river, and c) a static pond or
192 ditch.
193
194 SSDs based on frequentist analysis
7
195 SSDs were constructed from the 96-h LC50 data or time-independent PNOEC values
196 by nonlinear regression of a logistic model to the cumulative frequency plot of the
197 log-transformed data set5. Associated 95% point-wise percentile confidence intervals
198 were derived using bootstrap regression in which nonlinear regression is applied
199 repeatedly to re-samples of the cumulative frequency data plot3.
200
201 SSDs based on Bayesian analysis
202 A Bayesian statistical model was developed to construct SSDs from either 96-h LC50
203 data or time-independent PNOEC values, plus the results of the expert elicitation
204 exercise, and was implemented in the software package WINBUGS (http://www.mrc-
205 bsu.cam.ac.uk/bugs). Taxa mean log(96-h LC50) or log(PNOEC) values were
206 assumed to be normally distributed around a linear function of the experts' weighted
207 mean sensitivity values, with a precision representing experts' assessment errors. The
208 linear function was determined by least squares regression as the line of best fit to
209 measurements of the toxicity data for chlorpyrifos (relating to freshwater aquatic
210 species found in the UK in the literature) plotted against the corresponding average
211 assessment of sensitivity made by the experts (on a scale of 1 to 8).
212
213 [FIGURE 1 here]
214
215
216 Figure 1 shows the line of best fit together with corresponding data points. As
217 expected, the line has negative slope reflecting that more sensitive species were
218 broadly judged by the experts to be associated with more sensitive taxa. However, the
219 scattering in the plot reveals that expert judgements are not strongly correlated with
220 true toxicity and hence that the data are unable to support a more complex relationship
221 being fitted. Values for individual species were assumed to be normally distributed
222 around their taxa means. SSDs were constructed for each of the three generic
223 assemblages by predicting the proportion of species in each taxon with 96-h LC50 or
224 PNOEC values below a given concentration, and then averaging over the taxa present
225 in the assemblage. These were constructed both with and without use of expert
226 opinion from the elicitation exercise, in the latter case by forcing the linear regression
227 of taxa means on expert weighted means to have zero slope. Further details of the
8
228 methodology are available in O'Hagan et al.17 and at
229 http://www.shef.ac.uk/~st1ao/pub.html.
230
231 Empirical testing
232 Using the expert opinions of taxa sensitivity obtained in the elicitation exercise, taxa
233 of different predicted sensitivity for which empirical data were not available were
234 identified. Four species were then exposed to chlorpyrifos to examine whether the
235 experts had accurately predicted their sensitivity. These species, with the common
236 name and experts' assessment of their relative sensitivity in parentheses were
237 Ephemerella sp. (mayfly, 6th), Brachycentrus subnubilis (caddis fly, 8th), Leuctra sp.
238 (stonefly, 17th) and Hirudo medicinalis (leech, 47th). Test organisms were collected
239 from the wild (mayflies, caddisflies and stoneflies) or acquired commercially
240 (leeches) and exposed to a range of concentrations of technical grade chlorpyrifos for
241 96h in semi-static test systems, with test medium renewal every 24h. Stock solutions
242 were chemically analysed to verify exposure concentrations. A 96h LC50 was not
243 calculable for H. medicinalis, because no mortality occurred in that exposure
244 experiment but the observation thereby gave an empirical lower bound on the LC50
245 value. In summary, three additional 96h LC50s (hereafter referred to as the laboratory
246 96-h LC50 data) were obtained, together with a lower bound on the LC50 for H.
247 medicinalis which was used as a fourth additional data point in the Bayesian analysis.
248
249 Application of the Mayer et al.14 technique to the mortality data collected for these
250 three species yielded only 1 further time-independent PNOEC, for Ephemerella sp.,
251 hereafter referred to as the laboratory PNOEC, because of the problem caused by high
252 levels of individual variability in mortality response (discussed earlier). Hence finally,
253 two combined data sets consisting of literature and laboratory 96-h LC50 data (as
254 either 20 species (=17+3) for frequentist or 21 species (=17+4)) for the Bayesian
255 analyses) or Wijngaarden and laboratory PNOECs (8=7+1 species) were obtained.
256
257 The frequentist and Bayesian approaches described above were then run again to
258 generate new 96-h LC50 and time independent SSDs using these two newly combined
259 data sets.
260
261
9
262
263
10
264 RESULTS
265
266 [FIGURE 2 here]
267
268 Figure 2(a) shows the SSD together with 95% point-wise percentile confidence
269 intervals constructed by the frequentist approach using the 96-h LC50 literature data
270 (17 species). Figure 2(b) compares the Bayesian SSD constructed from the same data
271 for each of the three generic habitats and associated assemblages using expert
272 opinion, together with the single SSD constructed without expert opinion. If no expert
273 opinion is incorporated, the SSD for each generic assemblage is identical because
274 there is simply no information to specify how they should differ. Only by using expert
275 knowledge can the assemblages be distinguished and an SSD thus be tailored to fit a
276 given assemblage, unless there are large numbers of empirical data for each
277 assemblage type.
278
279 [FIGURE 3 here]
280
281 In general, incorporation of expert opinion produces a leftward shift of the Bayesian
282 SSD corresponding to each assemblage. Figure 3 illustrates this effect by comparing
283 the Bayesian SSD obtained both with and without expert opinion for the fast-flowing
284 stream assemblage, which incurred the greatest shift through inclusion of expert
285 opinion. This is likely to be because of the experts’ perception, probably reflecting a
286 more general view, that fast-flowing streams are characterised by sensitive species
287 such as stoneflies and mayflies, while ponds and ditches are characterised by less
288 sensitive species such as snails and leeches.
289
290 For all three habitats the results show that the use of expert opinion within the
291 Bayesian analysis produces lower (i.e., more sensitive) estimates of the respective
292 SSD and its 95% credible interval, plus summary statistics such as the HC5. This is
293 largely because the taxa that experts regarded as being particularly sensitive to
294 chlorpyrifos were not well represented in the empirical toxicity dataset (that is, the
295 LC50s obtained from either the literature or the laboratory tests). For example, the
296 HC5 was approximately 0.02 μg l-1 with, and 0.05 μg l-1 without, expert opinion for
297 both fast-flowing streams and slow-flowing rivers. The HC5 estimate turned out to be
11
298 slightly higher for the static pond, reflecting the fact that taxa found in this habitat and
299 not the others were judged by the experts to be generally less sensitive. The curves for
300 the different assemblages are strongly correlated because they have many species in
301 common. Hence, although the 95% credible intervals are generally wide and
302 overlapped substantially for all SSDs, there is evidence that the SSD for the fast-
303 flowing stream lies to the left of that for the slow-flowing river, which in turn lies to
304 the left of the SSD for the static pond.
305
306 [FIGURE 4 here]
307
308 Figure 4 compares time-independent SSDs constructed by either the frequentist (4a
309 and 4c) or Bayesian (4b and 4d) methodologies using the Wijngaarden PNOECs
310 derived from time-to-event analyses. The HC5 values for these PNOECs are all
311 considerably lower than those calculated previously from 96-h LC50 data, as would
312 be expected from extrapolation of short-term lethal concentrations to time-
313 independent no effects concentrations. Bayesian incorporation of expert opinion with
314 the Wijngaarden data produced a similar trend to the 96-h LC50 analyses, with lower
315 estimates of HC5 for fast-flowing stream and slow-flowing river habitats when
316 compared with a frequentist analysis, and a higher estimated HC5 for static pond
317 habitats.
318
319 The results from toxicity tests with four additional species, designed to test the
320 predictions made by experts, suggested that the experts were broadly correct in their
321 assignment of sensitivities (Table 3). Ephemerella sp. (mayfly) was the most sensitive
322 species, followed by Brachycentrus subnubilus (caddis fly), Leuctra sp. (stonefly) and
323 Hirudo medicinalis (leech), in the order predicted by the weighted mean expert
324 ranking. All statistical analyses were performed again to include these new
325 experimental data and the results are summarized in Table 4 together with those
326 obtained from the original datasets. A comparison between the analyses performed
327 with the various data sets reveals two interesting aspects.
328
329 First, there is a greater difference between the SSDs constructed from the time-
330 independent PNOECs than those with the 96-h LC50 endpoints. In particular, Table 4
331 shows that the SSD constructed from PNOECs estimated for the third habitat (pond or
12
332 ditch) has a noticeably higher HC5 than that calculated for the other two habitats
333 (whether using just 7 or all 8 PNOECS). Table 4 also shows that the SSD constructed
334 from PNOECs estimated for ponds or ditches has a more than two-fold higher HC5
335 than that calculated for the other two habitats, which suggests that lentic invertebrate
336 assemblages may be less susceptible to the toxicity of chlorpyrifos than lotic
337 assemblages.
338
339 [TABLE 4 here]
340
341 Second, a comparison between the time-independent SSDs associated with each
342 habitat when derived from the Wijngaarden values with that derived from the
343 Wijngaarden & laboratory values reveals a further important difference. This can be
344 seen by comparing respective HC5 values obtained in each case, and observing that
345 those derived from the SSDs constructed with Wijngaarden & laboratory values are in
346 general lower by a factor of about 5 (Table 4). Since the single laboratory PNOEC
347 (Ephemerella sp) has a very much lower value than any of those obtained by
348 Wijngaarden, an effect of this kind through its inclusion in SSD construction would
349 be expected. However, the Bayesian SSD analysis had access to expert knowledge,
350 which quite clearly believed that the Ephemeridae would be more sensitive than any
351 of the Wijngaarden species. So in principle this very low value should have been
352 discounted to an extent. In any event, the Bayesian SSDs would be expected to be less
353 affected by the singularly different laboratory data point than each of the respective
354 frequentist SSDs, which did not incorporate expert elicitation. A direct comparison of
355 the specific HC5s determined for each habitat exhibited in Table 4 shows this to
356 indeed be the case, as follows.
357
358 Using only the Wijngaarden PNOEC values to construct the Bayesian SSDs, the HC5
359 values (in μg l-1) were 1.34 x 10-4 for fast-flowing streams, 1.56 x 10-4 for slow-
360 flowing rivers and 3.89 x 10-4 for static ponds. Using the Wijngaarden & laboratory
361 values these respectively become 2.33 x 10-5, 2.92 x 10-5 and 9.94 x 10-5, which
362 represent leftward shifts by a factor of 5.75, 5.34 and 3.91 (corresponding to an
363 average factor of about 5, as stated above).
364
13
365 With the frequentist SSD construction approach, using the Wijngaarden values, the
366 HC5 is estimated as 1.74 x 10-4 μg l-1, whereas using the Wijngaarden & laboratory
367 values the HC5 estimate is 1.66 x 10-5. The leftward shift therefore now corresponds
368 to a factor of 10.48, which is twice that for the Bayesian SSDs.
369
370 This provides demonstrable evidence of the greater robustness to be expected in the
371 Bayesian approach to SSD construction.
372
373
374 DISCUSSION
375
376 The present study shows that for chlorpyrifos, use of a time-to-event approach to
377 estimate PNOEC values leads to estimates of toxicity with an HC5 generally less than
378 0.0004 μg l-1 (see Table 4). However, this estimate is fifty times lower than the 96-h
379 LC10 value of 0.02 μg l-1 estimated for Gammarus pulex, the most sensitive species
380 tested by Wijngaarden et al.13, and may therefore be an overestimate of likely effects
381 caused by transient environmental exposure. Wheeler et al.18 reported freshwater HC5
382 values for chlorpyrifos, based on acute data, of 0.086 µg/L (log-logistic model) and
383 0.063 µg/L (log-normal model), which are 215 and 158 times higher, respectively,
384 than the HC5 based on time-independent PNOEC values. However, Giesy et al.10
385 report acute-to-chronic ratios for chlorpyrifos of up to 181, so such a difference
386 between HC5 estimates based on acute or chronic summaries is at least plausible.
387
388 The use of Bayesian methodology to incorporate expert judgement of species
389 tolerance distributions and empirical data into SSDs produced lower estimates of HC5
390 than SSDs based on empirical data alone. This is because the available empirical data
391 were probably not fully representative of the whole population of data. In this study,
392 the differences were not large, but they could be for other substances if available data
393 are from studies that concentrate on highly sensitive or insensitive taxa. Use of
394 formalized expert opinion is therefore of considerable value in reducing any bias that
395 may be due to an unrepresentative selection of test species.
396
397 The experts were correct in their relative ranking of the four species tested as part of
398 this study, but did make some mistakes with other species. For example,
14
399 Chironomidae were judged by some experts to have low sensitivity to chlorpyrifos,
400 which clearly should not be the case for an insecticide. Such beliefs probably
401 originate from the 'sanitary water quality' bias of many field biologists in the UK, and
402 particularly those who work for the regulatory agencies. If the experts had been more
403 accurate in their relative assessments then the predictions of SSDs would have tighter
404 credible intervals. Hence, appropriate selection of experts is important and some
405 objective test of their true level of expertise, rather than reliance on self-assessment,
406 could usefully be incorporated into elicitation exercises such as these.
407
408 There were insufficient data points to be able to test any assumptions on taxa
409 heterogeneity and so for reasons of parsimony the same variance in every taxon was
410 incorporated into the Bayesian model. In principle, the experts could have been asked
411 to estimate such heterogeneity, but this would have raised questions about what they
412 could tell us reliably. Unlike their assessments of sensitivity, such assessments of
413 heterogeneity would not be accessible to meaningful validation. Hence it was not
414 possible to develop a more realistic model.
415
416 The choice of organisms used to produce the three generic assemblages in this study
417 is open to debate. There was substantial species overlap between the three
418 assemblages, with most phyla represented in all three scenarios, and this will have
419 contributed to the relatively small differences in estimated SSDs. Forbes and Calow9
420 suggest that risk assessments based upon SSDs should be relevant to specific sites.
421 However, there is a question over whether we should seek to protect what is currently
422 present at a site, or whether we should protect what could be present at a site. A more
423 sophisticated treatment of site-specific assemblages is certainly achievable for UK
424 lotic systems, by using RIVPACS19 to predict site-specific assemblages under pristine
425 conditions for subsequent SSD construction.
426
427 In general, it seems from other studies that LC/EC values for chlorpyrifos can be
428 broadly predictive of longer-term toxic effects, and do not appear to over- or under-
429 estimate them greatly. Crane et al.11 concluded that chlorpyrifos is highly toxic to
430 arthropods, with the water flea Ceriodaphnia dubia the most sensitive species on the
431 ECOTOX database with a 96-h LC50 of 0.057 μg l-1. These data compare well
432 Bayesian estimates of HC5 for 96-h LC50 values calculated for the three generic
15
433 assemblages, with expert judgment, which ranged from 0.02 to 0.032 μg l-1. This
434 suggests that chlorpyrifos concentrations should be less than 0.057 μg l-1 and may
435 need to be less than 0.02 μg l-1 to protect all aquatic systems from harm.
436
437 Mesocosm results may help in 'ground-truthing' laboratory estimates, although even
438 these systems cannot fully represent the range of natural water bodies and taxa that
439 could potentially be adversely affected in the natural environment20. Giesy et al. 10
440 reviewed available mesocosm data for chlorpyrifos and concluded that effects on
441 invertebrates could be reliably measured at concentrations of chlorpyrifos >0.2 μg l -1,
442 with recovery of most populations within 2-8 weeks, and that effects on fish occurred
443 at concentrations >0.5 μg l-1 21. These values are an order of magnitude higher than
444 those estimated in this study, which may reflect overly conservative estimates based
445 on laboratory studies, or problems in detecting low levels of effect on a wide variety
446 of organisms in variable mesocosm experiments. If laboratory-to-field extrapolation
447 factors are required to take lower field sensitivity into account then it would be better
448 to apply them to time-independent no effects concentrations rather than LC50 values
449 estimated at an arbitrary multiple of 24 hours
450
451 This study has shown that valid theoretical criticisms of the SSD approach can be
452 overcome. We have demonstrated some strategies for constructing and using SSDs
453 that would help to minimise current deficiencies and make SSDs more
454 environmentally relevant, technically robust, and useful for both site-specific and
455 more generic risk assessments. In this investigation, the PNOECs of assemblages
456 representing different freshwater habitats did not differ in sensitivity to chlorpyrifos
457 by more than a factor of 4.3. In this case it may therefore be that risk management
458 decisions could be based on a generic species assemblage without the need to
459 consider different habitats. However, this may not always be the case, and the use of
460 expert elicitation in the construction of SSDs can therefore help with two quite
461 different aspects. First, expert knowledge allows us to account for differences
462 between habitats. Second, it allows us to take account of the non-representativeness
463 of the highly selective data that may be available. Finally, we have also been able to
464 show how more meaningful toxicological endpoints may be estimated from existing
16
465 datasets through the use of time-to-event approaches to estimate low levels of effect
466 which are, presumably, the protection goals of most regulatory frameworks.
467
468 In principle, similar approaches to those described in this paper may be adopted for
469 other chemicals. However, there are undoubtedly some practical constraints. These
470 include difficulties in soliciting opinions on species sensitivities to chemicals with
471 which field biologists are unfamiliar, and in obtaining raw data for a sufficiently wide
472 range of species from which more meaningful toxicological endpoints may be
473 calculated. Nevertheless the benefits seem sufficiently important to justify efforts to
474 address these deficiencies.
475
476 Acknowledgment - This work was funded by NERC Environmental Diagnostics Grant
477 GST/02/2062, DEFRA Grant PN0933, and the Environment Agency of England and
478 Wales. We thank Environment Agency staff and members of the Freshwater
479 Biological Association for help during the elicitation exercise. We also thank
480 colleagues, especially René van Wijngaarden, for providing access to raw data.
481
482 LITERATURE CITED
483
484 (1) Solomon, K. R.; Baker, D. B.; Richards, R. P.; Dixon, D. R.; Klaine, S. J.;
485 Lapoint, T. W.; Kendall, R. J.; Weisskopf, C. P.; Giddings, J. M.; Giesy, J. P.;
486 Hall, L. W.; Williams, W. M. Environ. Toxicol. Chem. 1996, 15, 31-74.
487
488 (2) Steen, R. J. C. A.; Leonards, P. E. G.; Brinkman, U. A. T.; Barcelo, D.;
489 Tronczynski, J.; Albanis, T. A.; Cofino, W. P. Environ. Toxicol. Chem. 1999,
490 18, 1574-1581.
491 (3) Grist, E. P. M.; Leung, K. M. Y.; Wheeler, J. R.; Crane, M. Environ. Toxicol.
492 Chem. 2002, 21, 1515-1524.
493 (4) Wagner, C.; Lokke, H. Water Res. 1991, 25, 1237-1242.
494 (5) Aldenberg, T.; Slob, W. Ecotoxicol. Environ. Safe. 1993, 25, 48-63.
17
495 (6) Van Straalen, N.M.; Van Rijn. J. P. Rev. Environ. Contam. Toxicol. 1998, 154,
496 85-141.
497 (7) EC Technical Guidance Document No 1488/94 on Risk Assessment for Existing
498 Substances and Directive 98/8/Ec of the European Parliament and the Council
499 Concerning the Placing of Biocidal Products on the Market. Part II. European
500 Commission Joint Research Centre, Eur 20418 En/2. 2003.
501 (8) Hart, A. Probabilistic Risk Assessment for Pesticides in Europe: Implementation
502 and Research Needs. Central Science Laboratory,York, UK. 2001.
503 (9) Forbes, V. E.; Calow, P. Hum. Ecol. Risk Assess. 2002, 8, 473-492.
504 (10) Posthuma L.; Suter, G.W. II; Traas, T.P.; Eds. Species Sensitivity Distributions
505 in Ecotoxicology, Lewis Publishers, Boca Raton, FL., 2002
506 (11) Giesy, J. P. ; Solomon, K. R.; Coats, J. R.; Dixon, K. R.; Giddings, J. M.;
507 Kenaga, E. E.. Rev. Environ. Contam. Toxicol. 1999, 160, 1-129.
508 (12) Crane, M.; Whitehouse, P.; Comber, S.; Watts, C.; Giddings, J.; Moore, D. R. J.;
509 Grist, E. Pest Manage. Sci. 2003, 59, 512-526.
510 (13) Whitehouse P.; Cartwright N. In Pollution Risk Assessment and Management: a
511 Structured Approach; Douben P.E.T., Ed., Wiley, Chichester, UK, 1998, pp
512 235-272.
513 (14) van Wijngaarden, R.; Leeuwangh, P.; Lucassen, W. G. H.; Romijn, K.; Ronday,
514 R.; Van Der Velde, R.; Willigenburg, W. Bull. Environ. Contam. Toxicol. 1993,
515 51, 716-723.
516 (15) Mayer, F. L.; Ellersieck, M. R.; Krause, G. F.; Sun, K.; Lee. G.; Buckler, D. R.
517 In Risk Assessment With Time to Event Models; Crane, M., Newman, M.C.,
518 Chapman P.F., Fenlon J., Eds., Lewis Publishers, Boca Raton, FL., 2002, pp 39-
519 67.
520 (16) Fitter, R.; Manuel, R. Lakes, Rivers, Streams and Ponds of Britain and North-
521 West Europe. Collins Photo Guide, Harper Collins, Hong Kong 1994.
18
522 (17) O’Hagan, A.; Crane, M.; Grist, E. P. M.; Whitehouse, P. Submitted to Appl.
523 Stat.
524 (18) Wheeler, J.R.; Leung, K.M.Y; Morritt, D; Whitehouse, P.; Sorokin, N.; Toy, R.;
525 Holt, M.; Crane, M. Environ. Toxicol. Chem. 2002, 21, 2459-2467.
526 (19) Wright, J. F.; Furse, M. T.; Armitage, P. D. In Water Quality and Stress
527 Indicators in Marine and Freshwater Systems; Sutcliffe DW Ed. Freshwater
528 Biological Association, Ambleside, UK, 15-34, 1994.
529 (20) Crane, M. Hydrobiologia 1997, 346, 149-155.
530 (21) Giddings, J. M.; Biever, R. C.; Racke, K. D. Environ. Toxicol. Chem. 1997, 16,
531 2353-2362.
19
532 Table legends
533
534 Table 1
535 96-h LC50s obtained from the literature (17 species) and laboratory experiments (3 species). Values were computable for 3 laboratory
536 species only, because no mortality occurred in the exposure experiment conducted on Hirudo medicinalis (a leech).
537
538 Table 2
539 Time-independent Predicted No Effect Concentrations (PNOECs) calculated by the Mayer et al.14 technique. Values were calculable for
540 only 7 species in the Wijngaarden data set and 1 species (Ephemerella sp) in the laboratory data because of the high levels of individual
541 variability in mortality response.
542
543 Table 3
544 Taxonomic groups considered during expert elicitation and their likely habitats. The 6th column gives the expert weighted mean
545 sensitivity for each taxon as assessed by the experts on a scale of 1 to 8
546
547 Table 4
548 Comparison of HC5 values determined from species sensitivity distributions (SSDs) constructed by frequentist and Bayesian
549 methodologies, using the 96-h LC50 data sets and time independent PNOEC values. L =lower 2.5%, U=upper 97.5 percentiles, for
550 frequentist point-wise confidence or Bayesian credible intervals. # Table 4 Foot Note: 17 +4 species were used in the Bayesian analysis
551
552
553
20
554
555 Figure legends
556 Figure 1 Plot showing the 17 measurements of the 96 hour LC50 obtained for chlorpyrifos from the literature for freshwater aquatic
557 species found in the UK plotted against the average assessment of the corresponding taxa sensitivity made by the experts.
558 The y-axis is the log LC50 and the x-axis is the average assessment of sensitivity by the experts (on a scale of 1 to 8), with
559 rings around data points to identify those species which are in the same taxon.
560
561 Figure 2 Chlorpyrifos species sensitivity distributions (SSDs) estimated from the 96-h LC50 literature data (open circles, 17
562 species)
563 (a) Frequentist using a log-logistic model with bootstrap regression. The 50% percentile (bold) is shown together with
564 lower (2.5%) and upper (97.5%) point-wise percentile confidence intervals (thin). Number of re-samples =2000.
565 (b) Bayesian using expert opinion from the elicitation exercise for the three generic assemblages of fast-flowing stream
566 (dashes), slow-flowing river (dots) or static pond or ditch (dot-dash). With no expert opinion (bold), the three SSDs are
567 identical. In each case the 50% credible percentile is shown [together with lower (2.5%) and upper (97.5%) credible
568 percentiles (thin) with no expert opinion].
569 Figure 3 The leftward shift that occurs through incorporation of expert opinion in the Bayesian-estimated SSD, illustrated for the
570 fast-flowing stream habitat using the 96-h LC50 literature data (open circles, 17 species). Respective 50% credible
571 intervals with (dashes) and without (bold) expert opinion are compared together with lower (2.5%) and upper (97.5%)
572 95% credible limits (dots and solid respectively).
21
573 Figure 4 Chlorpyrifos species sensitivity distributions (SSDs) estimated from time-independent PNOEC values (open circles).
574 Respective 50% credible intervals (bold) are shown in each case together with lower (2.5%) and upper (97.5%) credible
575 limits (thin).
576 (a) Frequentist with a log-logistic model and bootstrap regression using the Wijngaarden PNOECs (7 species). Number of
577 re-samples =2000.
578 (b) Bayesian for the fast-flowing stream habitat using the Wijngaarden PNOECs (7 species).
579 (c) Frequentist with a log-logistic model and bootstrap regression using the combined Wijngaarden and laboratory
580 PNOECs (8=7+1 species). Number of re-samples =2000
581 (d) Bayesian for the fast-flowing stream habitat using the combined Wijngaarden and laboratory PNOECs (8=7+1
582 species).
583
584
585
586
587
588
589
590
591
22
592 Table 1
Species Taxon 96-h LC50 (μg/L)
Literature (17 species)
Anguilla anguilla Anguillidae 540
Asellus aquaticus Asellidae 2.7
Caenis horaria Caenidae 0.5
Chironomus tentatus Chironomidae 0.47
Chironomus thummi Chironomidae 0.2
Corixa punctata Coroxidae 2
Rutilus rutilus Cyprinidae 120
Gammarus lacustris Gammaridae 0.11
Gammarus pulex Gammaridae 0.07
Gammarus fasciatus Gammaridae 0.32
Gammarus pseudolimnaeus Gammaridae 0.18
Gasterosteus aculeatus Gasterosteidae 8.54
Pungitus pungitus Gasterosteidae 4.7
Peltodytes sp. Haliplidae 0.8
Leptocerida sp. Leptoceridae 0.77
Oncorhynchus mykiss Salmonidae 7.1
Oncorhynchus clarki Salmonidae 14.42
Laboratory (3 species)
23
Ephemerella sp. Baetidae 0.035
Leuctra sp. Leuctridae 0.87
Brachycentrus subnubilus Brachycentridae 0.45
Hirudo medicinalis Hirudidae *
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
24
612 Table 2
613
Species Taxon 48-h LC0 96-h LC0 PNOEC
(μg/L) (μg/L) (μg/L)
Wijngaarden (7 species)
Chaoberus obscuripes Chaoboridae 1.77 x 10-1 1.10 x 10-1 6.81 x 10-2
Cloeon dipterum Baetidae 1.56 x 10-1 1.45 x 10-2 1.35 x 10-3
Corixa punctata Corixidae 5.66 x 10-1 2.62 x 10-1 1.21 x 10-1
Daphnia longiseta Daphniidae 6.31 x 10-2 5.56 x 10-2 4.89 x 10-2
Gammarus pulex Gammaridae 6.35 x 10-3 2.66 x 10-3 1.12 x 10-3
Pungitus pungitus Gasterosteidae 7.50 x 10-1 6.48 x 10-1 5.59 x 10-1
Simocephalus vetulus Daphniidae 9.40 x 10-2 9.26 x 10-2 9.12 x 10-2
Gasterosteus aculeatus Gasterosteidae 3.27 x 10-1 4.06 x 10-1 *
Laboratory (1 species)
Cloeon dipterum Baetidae 1.09 x 10-4 4.63 x 10-5 1.25 x 10-5
Leuctra geniculata Leuctridae 0.0000 1.41 x10-5 *
Brachycentrus subnubilus Brachycentridae 1.99 x 10-3 6.72 x 10-3 *
Hirudo medicinalis Hirudidae * * *
614
615
616
617
25
618
619
620 Table 3
621
Taxon Type of organism Fast-flowing Slow-flowing Pond or Expert weighted
stream river Ditch mean sensitivity
Aeshnidae Dragonflies 5.56
Ancylidae Limpets 4.29
Anguillidae Eels 4.17
Aphelocheiridae Saucer bugs 5.00
Asellidae Water hoglice 4.08
Astacidae Crayfish 6.25
Baetidae Mayflies 5.79
Beraeidae Caddis flies 6.20
Brachycentridae Caddis flies 6.27
Bufonidae Toads 5.20
Caenidae Mayflies 5.86
Calopterygidae Demoiselles 5.33
Capniidae Stoneflies 5.74
Chironomidae Midges 3.56
Chloroperlidae Stoneflies 6.13
Clupeidae Shad 3.91
Cobitidae Loach 4.00
Coenagriidae Damselflies 5.56
Cordulegasteridae Dragonflies 5.59
Corduliidae Dragonflies 5.56
Coregonidae Whitefish 4.27
Corixidae Lesser waterboatmen 5.11
Corophiidae Shrimps 5.00
Cottidae Bullhead 4.40
Cyprinidae Carp 4.08
26
Dendrocoelidae Flatworms 5.00
Dryopidae Water beetles 4.57
Dytiscidae Diving beetles 3.91
Elminthidae Riffle beetles 5.20
Ephemerellidae Mayflies 6.50
Ephemeridae Mayflies 6.44
Erpobdellidae Leeches 4.80
Escocidae Pike 4.00
Gammaridae Shrimps 5.57
Gasterosteidae Sticklebacks 4.13
Gerridae Pond skaters 3.69
Glossiphoniidae Leeches 4.89
Goeridae Caddis flies 5.95
Gomphidae Dragonflies 5.65
Gyrinidae Whirligig beetles 4.89
Haliplidae Water beetles 5.00
Heptageniidae Mayflies 6.75
Hirudidae Leeches 5.21
Hydrobiidae Snails 4.89
Hydrometridae Water measurers 4.00
Hydrophiliidae Scavenger beetles 4.50
Hydropsychidae Caddis flies 5.92
Hydroptilidae Caddis flies 6.18
Hygrobiidae Screetch beetles 4.69
Lepidostomatidae Caddis flies 6.14
Leptoceridae Caddis flies 6.00
Leptophlebiidae Mayflies 6.57
Lestidae Damselflies 5.88
Leuctridae Stoneflies 6.04
Libellulidae Dragonflies 5.41
Limnephilidae Caddis flies 5.17
Lymnaeidae Snails 4.47
Mesoveliidae Water bugs 4.56
27
Molannidae Caddis flies 5.89
Nemouridae Stoneflies 6.00
Nepidae Water scorpions 4.81
Neritidae Snails 4.88
Notonectidae Water boatmen 4.72
Odontoceridae Caddis flies 6.00
Oligochaeta Worms 2.78
Percidae Perch 3.92
Perlidae Stoneflies 6.59
Perlodidae Stoneflies 6.59
Petromyzonidae River lamprey 4.23
Philopotamidae Caddis flies 6.05
Phryganeidae Caddis flies 5.90
Physidae Bladder snails 4.82
Piscicolidae Fish leeches 4.94
Planariidae Flatworms 5.00
Planorbidae Ramshorn snails 4.68
Platycnemidae Damselflies 4.86
Pleidae Lesser backswimmers 4.29
Polycentropidae Caddis flies 6.18
Potamanthidae Mayflies 5.75
Psychomyidae Caddis flies 5.76
Ranidae Frogs 5.10
Rhyacophilidae Caddis flies 5.86
Salmonidae Salmon 5.40
Scirtidae Beetles 4.87
Sericostomatidae Caddis flies 5.95
Sialidae Alderflies 5.33
Simuliidae Black-flies 5.50
Siphlonuridae Mayflies 6.53
Sphaeriidae Mussels 5.44
Taeniopterygidae Stoneflies 6.19
Thymallidae Grayling 4.00
28
Tipulidae Crane-flies 4.84
Triturus Newts 5.10
Unionidae Mussels 5.40
Valvatidae Snails 5.18
Viviparidae Snails 5.33
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
29
641 Table 4
SSD constructed using BAYESIAN FREQUENTIST
Habitat HC5 (μg/L) HC5 (μg/L)
96-h LC50 data Median L (2.5%) U (97.5%) Median L (2.5%) U (97.5%)
Literature (17 species) No expert opinion 4.95 x 10-2 1.11 x 10-2 2.41 x 10-1 2.53 x 10-2 4.60 x 10-3 1.10 x 10-1
Fast-flowing stream 1.90 x 10-2 1.86 x 10-4 1.60 x 10-1 * * *
Slow-flowing river 2.14 x 10-2 2.95 x 10-4 1.73 x 10-1 * * *
Static pond 3.29 x 10-2 5.39 x 10-4 2.34 x 10-1 * * *
# Literature & Laboratory (17 +3species) No expert opinion 3.17 x 10-2 1.38 x 10-3 1.47 x 10-1 2.40 x 10-2 5.40 x 10-3 9.45 x 10-1
Fast-flowing stream 2.03 x 10-2 9.10 x 10-4 1.03 x 10-1 * * *
Slow-flowing river 2.30 x 10-2 1.19 x 10-3 1.15 x 10-1 * * *
Static pond 3.23 x 10-2 2.01 x 10-3 1.53 x 10-1 * * *
Time independent PNOECs
Wijngaarden (7 species) No expert opinion 4.01 x 10-4 1.16 x 10-6 4.28 x 10-3 1.74 x10-4 8.20 x10-6 3.05 x 10-2
Fast-flowing stream 1.34 x 10-4 3.26 x 10-7 1.47 x 10-3 * * *
Slow-flowing river 1.56 x 10-4 4.42 x 10-7 1.75 x 10-3 * * *
Static pond 3.89 x 10-4 1.86 x 10-6 2.83 x 10-3 * * *
Wijngaarden & Laboratory (7 +1 species) No expert opinion 2.17 x 10-5 1.52 x 10-8 5.55 x 10-4 1.66 x 10-5 3.87 x 10-8 2.02 x 10-2
Fast-flowing stream 2.33 x 10-5 9,82 x 10-8 2.75 x 10-4 * * *
-7 -4
Slow-flowing river 2.92 x 10-5 1.49 x 10 3.51 x 10 * * *
Static pond 9.94 x 10-5 6.44 x 10-7 7.83 x 10-4 * * *
30
642 FIGURE 1
Regression Plot
x
y = 6.06098 - 1.12545 w
S = 2.33369 R-Sq = 14.7 % R-Sq(adj) = 9.0 %
7
6
5
4
3
yy 2
1
0
-1
-2
-3
4 5 6
x
w
643
31
644
645 FIGURE 2a
646
647
648
32
649 FIGURE 2b
650
651
33
652
653 Figure 3
654
34
655 Figure 4
656
657
35
658
659
36
660
37