Mesocosm experiments in the approval procedure for pesticides
5 PLS (partial least squares)
PLS is a sort of a "regression technique" that is used to describe the relationship between two sets of variables (X and Y matrices). The method is widely used within QSAR (quantitative structure-activity relationships) to compare the toxic effects of different substances on different test organisms (Y matrix) with the physico-chemical characteristics of the substances (X matrix) (Eriksson et al., 1995). Each substance thus makes up an observation, and the various physical-chemical characteristics and the toxic effects on the various test organisms function as individual variables. Each observation thus includes several variables, and the PLS-technique is therefore a so-called multivariate technique.
The advantage of PLS and other multivariate techniques is that the complexity of a data set, consisting of several variables, often can be reduced to a much lower number of dimensions by which the essential relationships between the two matrices can be elucidated. Being a multivariate technique (and not a multiple regression technique) PLS can include variables that are either true independent (e.g. depth of the mesocosm and logKow) or interrelated (e.g. logKow and logKD). For a number of pesticides, the PLS technique in this report has been used to predict the toxic response of various organisms (Y matrix) in model ecosystems on the basis of system volume, depth, location (latitude, longitude) and the toxicological and physico-chemical characteristics of the pesticide (X matrix). Initially, the "field half-life" was included in the X matrix. However, as we were only able to obtain data for 11 pesticides, this parameter was omitted.
Compared to other multivariate techniques the main advantages of the PLS analysis are, that it allows to analyse data sets consisting of more variables than observations, and that the method can handle observations where data on one or more variables are missing. The handling of such missing observations is based on an iterative procedure by which the missing values are estimated. As a rule of thumb the PLS method may thus handle data sets with up to 20% missing data, provided that these data are randomly distributed throughout the data set (Eriksson et al. 1995). In addition the PLS routine allows estimations of confidence intervals around the predicted values.
5.1 Development of PLS models
Due to the limited number of experiments analysing effects in more than 2 major groups of organisms (see Fig. 6) separate PLS models were developed for communities of macroinvertebrates, zooplankton and micro algae (periphyton and phytoplankton). For the remaining groups of organisms, such as micro-organisms (bacteria, ciliates) and macrophytes (vascular plants) it was not possible to develop PLS models due to shortage of data. The raw data, extracted from the database, for the PLS models is shown in Table 3, 8 and 11. For all communities the lowest effect concentrations observed (significant positive or negative deviations from controls) for each functional or taxonomic group in each mesocosm experiment were used as Y variables, expressing the toxic response of the organisms in the mesocosms. It should be noticed that by including only LOECs (but tested for significance, see above) the PLS analysis do not explicitly take account of recovery of populations.
In less than 25 experiments were the pesticide concentration monitored and described in such a detail that an observed effect could be ascribed to a specific pesticide concentration. Therefore, nominal (added) effect concentrations were used throughout. In the case of several dosings, the total (accumulated) concentration was used as the nominal effect concentration. The effect of dosing mode of the pesticides (number of doses, interval between doses) was entered as independent variables (X-matrix) and thus accounted for specifically in the analysis. All analyses were based on log-transformed data, e.g. log Kow, log Depth.
Initially, a series of scenarios were defined and a PLS model was fitted using the auto-fit routine of the program SIMCA-P 8.1 and examined for each scenario. However, due to the limited amount of data it was not possible to examine all scenarios. The different scenarios selected are shown in Table 6. The final choice of PLS models were based on successively narrowing the mesocosm characteristics (e.g. including only mesocosms with a sediment compartment). For each PLS model possible outliers were identified using plots of residuals (normal distributed) and predicted values against observed values. For the models chosen the experimental set-up in the outliers was examined in detail to explain their deviance.
After removing outliers PLS models with the highest predictability were selected for further interpretation.
5.2 Interpretation of PLS models
To interpret the PLS analyses, the following procedure was used:
- For each data set, the number of significant PLS axes were found.
- For each of the significant PLS axes, the importance of the different variables was determined by means of so-called loadings or weights that are a measure of how much each variable contributes to the axis in question. The weights can be both positive and negative, and weights with opposite signs can be interpreted as being negatively correlated, whilst weights with identical signs can be interpreted as being positively correlated.
- To reduce complexity the interpretations were limited to variables with an overall significant contribution to the PLS model (roughly equivalent to loadings larger then 0.2 or lower than -0.2; see Fig. 7).
To obtain an adequate amount of data for a PLS analysis, it was necessary to group the observations into different data sets. In the following section, the results of each of these analyses will be described followed by a general discussion including recommendations. An overview of the abbreviations used can be found in Table 2.
5.3 Using the PLS models in a standardised evaluation procedure of pesticides
When appropriate PLS models have been developed it is possible to use the models for prediction of effect concentrations for the organisms in the mesocosms and to associate the predicted effect concentrations with for instance a 95 % confidence interval. The PLS models do even allow to predict the effect concentrations with associated confidence interval for experiments where toxicity data for certain groups of organisms were missing. Since the PLS models are based on all appropriate data in the database it is thus possible to develop an evaluation procedure taking all the available information into account, rather than on a restricted use of a single or a few mesocosms experiment for each pesticide. Effects of pesticides in nature will depend on a suite of factors including the direct toxic effects of a pesticide, the physical-chemical conditions in the environment and the biological structure and interactions within the environment. Several of these conditions not related to the direct toxic effects may be as important for the actual effects as the pesticide concentration and moreover they may modify the effect of different pesticides in a more or less uniformly way. Thus with the aid of the PLS models it is possible to evaluate all mesocosm experiments with pesticides on a common basis.
Throughout the following the lowest observed effect concentrations (LOEC) obtained for the various groups of organisms in the mesocosm experiments were used as Y-variables in the PLS analysis. Thus the predicted effect concentrations are conceptually comparable to HC5,50 concentrations estimated on the basis of single species toxicity data and the extrapolation methods of Wagner and Løkke (1991). Similarly the lower limit of the confidence interval might be considered as equivalent to a HC5,95 concentration estimated with the aid of the extrapolation procedure of Wagner and Løkke (1991).
However, with the aid of PLS models it is possible to take the information from other mesocosm experiments into account, whereby the critical extrapolation from lower levels of biological organisation (single species level) to higher levels of biological organisation (ecosystem community) is avoided. In effect, by applying the PLS technique in a steadily growing data base the hazards of new pesticides at ecosystem level can be evaluated by interpolation instead of by extrapolation. Furthermore, the importance of pesticide properties, such as log Kow and log KD, and system properties, such as the volume and size of the mesocosm, are taken into account.
Table 2.
List of abbreviations used in PLS figures and in Tables 3, 8, 11.
Common X matrix for all PLS models |
|
Variable |
Explanation/remark |
Day number |
Refers to julian day of first pesticide dosage. Sinusiodal distribution with mid summer (21 June) as the highest day number (log(183)). |
Latitude |
|
Longitude |
|
Log KD |
Particle sorption coefficient |
Log KOW |
Distribution coefficient between octanol and water |
Dosing interval |
Time in days between addition of pesticides |
Number of additions |
Number of additions of a pesticide |
Volume |
Volume in litre |
Depth |
Average depth in m. |
HC5,50 |
Extrapolated HC5,50 concentration (Wagner and Løkke 1991) |
OECD |
EC50 (algae) or LC50 for the most sensitive organism divided by 10 (OECD 1991) |
Y matrix for macroinvertebrates |
|
Variable |
Explanation/remark |
Non_pred |
Lowest effect concentration for non predatory organisms |
Pred |
Lowest effect concentration for predatory organisms |
Epi_fauna |
Lowest effect concentration for epifauna organisms |
In_fauna |
Lowest effect concentration for infauna organisms |
Y matrix for zooplankton communities |
|
Variable |
Explanation/remark |
Cladocea |
Lowest effect concentration for Cladocera abundance |
Copepod |
Lowest effect concentration for Copepod abundance |
Rotifer |
Lowest effect concentration for rotifer abundance |
Y matrix for microalgae |
|
Variable |
Explanation/remark |
Micr_algae |
Lowest effect concentration for microalgae abundance |
5.4 PLS models for macroinvertebrates
An overview of raw data from the database used in the PLS models developed for macroinvertebrates is shown in Table 3. The analyses are based on data from 17 different experiments with a total of 9 different pesticides.
Table 3.
Overview of raw data from the database for the PLS models developed for
macroinvertebrates. See Table 2 for abbreviations used. Sediment 1 refers to sediment
present in mesocosm, 0 to no sediment; Macrophytes: 1 = present, 0 = no macrophytes; N= no
information given on presence of marcophytes Field/lab: 1 = Field study, 0 = Laboratory
study. For all macroinvertebrate groups the lowest effect concentrations observed for each
functional group in each mesocosm experiment were used as Y variables, expressing the
toxic response of the organisms in the mesocosms (values shown in bold). L = Laboratory
study at controlled temperature and light availability (hence latitude and longitude not
relevant); F = flow-through study; - = no data. See Annex B for literature references.
Table 3
cont.
The PLS models examined for macroinvertebrates are shown in Table 4.
Table 4
Predictability of the examined PLS models for macroinvertebrates. Q2(cum)
denotes the cumulative predictability (both significant axis included). See Annex B for
literature references.
Q2(cum) |
||
All available data for stagnant water with sediment |
none |
0.481 |
Mesocosm data for stagnant water with sediment |
experiment 76tll1) and 125flm2) |
0.599 |
Mesocosm data for experiments with macrophytes and sediment |
experiment 76tll and 125flm |
0.625 |
1)
Mesocosm experiment 76tll consisted of 450 m3 cosms dosed with Lambda-cyhalothrin every 14 days during a period of 147 days. Along with experiment 107tll (2,4-D) the exposure scheme was by far the most extensive in terms of length of exposure period and number of additions.2)
Mesocosm experiment 125flm consisted of 635 m3 cosms pulse-exposed to Trahalomethrin 5 times during 65 days. Because of rather high through-flow 90% of the pecticide was washed-out within 24 h after each dosage. Hence, the calculated total exposure concentration (=sum of each dosage) inevitably will grossly overestimate the actual concentrations (i.e. the experiment may not qualify for a true stagnant water experiment).As shown in Table 4 the highest predictability (Q2(cum)) was obtained for the PLS model based on mesocosm experiments where both sediment and macrophytes were present in the test system. However, an almost as high Q2(cum) were obtained for the PLS models for mesocosm experiments with sediment but without macrophytes in the test system. On the other hand a much lower predictability (Q2 = 0.481) was obtained when the PLS model was developed including all experiments carried out in stagnant water (i.e. including laboratory experiments). Hence, the larger similarity in the constituents of the different mesocosms (and closer resemblance to natural conditions) the higher is the predictability of toxic effects based on the various physical, chemical and toxicological properties of an experiment (see Table 2). Therefore, a PLS model for macroinvertebrates with an acceptable predictability needs to be based on mesocosm experiments containing sediments and preferentially macrophytes in the test system. The interpretation of the PLS model is therefore based on the PLS model for mesocosm data with both macrophytes and sediment present in the test systems, but excluding 2 experiments (76tll and 125flm) i.e. a total of 9 experiments.
5.4.1 Interpretation of the PLS model for macroinvertebrates
An overview of the model selected in the previous section is shown in Table 5.
Table 5.
Predicted variation of the significant axis of the PLS model selected for
macroinvertebrates. Q2: Variation in the Y matrix predicted from the variation
in the X matrix by the current axis. Q2(cum): Cumulative variation in the Y
matrix predicted from the variation in the X matrix.
PLS axis number |
Q2 |
Q2(cum) |
1 |
0.527 |
|
2 |
0.207 |
0.625 |
As shown in Table 5 the first PLS axis predicts 52.7 % of the variation in the Y matrix (i.e. the toxic response) from the variation in the X matrix (system characteristics and toxicological and physico-chemical characteristics of the pesticide). The second PLS axis predicts 20.7% of the variation in the Y matrix from the variation in the X matrix. The fact that Q2(cum) is lower than the sum of Q2 for the two axis (i.e. 0.527 + 0.207) indicates some overlap between the predictions of the first and second PLS axis.
Figure 8.
Weights (loadings) of variables contributing to the first PLS axis for
macroinvertebrates. Day number through OECD represent variables in the X-matrix while the
responses (LOEC) of the different macroinvertebrate groups are shown at right. Weights
with opposite signs can be interpreted as being negatively correlated (e.g. INTERVAL
between pesticide doses and toxic response of either macroinvertebrate group), while
weights with identical signs can be interpreted as being positively correlated (e.g.
mesocosm Depth, Number of dosings and toxic response of macroinvertebrates).
Figure 9.
Weights (loadings) of variables contributing to the second PLS axis for
macroinvertebrates.
As described in Section 4.2 the PLS axis was interpreted on the basis of the weights (loadings) of the variables to each PLS axis. The loading plots are shown in Figs 8 & 9.
 | The first PLS axis showed positive loadings of almost equal magnitude in all groups of macroinvertebrates, which means that the four groups are equally sensitive to pesticides and the conditions in the mesocosms. |
| For the second PLS axis a much lower positive loading is obtained for the infauna (i.e. borrowing animals) than for the other functional groups. Thus the two PLS axis seems to be indicative of different toxic responses of the macroinvertebrates according to their habitat. |
| For the X variable longitude positive loadings are seen for both PLS axis, whereas for latitude, negative loadings are seen for both PLS axis. Thus all macroinvertebrate groups in the mesocosms seem to be most sensitive when the experiments are conducted at high latitudes and low longitudes in mesocosms. Therefore, toxic effects at lower concentrations are expected with increasing distance from Equator. A likely explanation is probably related to a slower turn-over of populations at high latitudes, i.e. fewer generations each year at lower temperatures. Therefore, recovery of populations affected by pesticide exposure takes longer time at northern latitudes. The positive loading for longitude suggests that organisms in experiments conducted in USA are less sensitive than the macroinvertebrates in experiments conducted in Europe. A possible explanation could be that most mesocosm experiments in USA, but not in Europe, are carried out with fish present in enclosures, which may override or mask the effect of pesticides. |
| For the first axis the "Interval" between pesticide dosings correlated negatively with LOEC of the macroinvertebrates. Thus by increasing the interval between dosings a lower LOEC will result. This may be due to the relativly long generation time of most macroinvertebrates. Hence, recovery will be hampered if pesticides are dosed at intervals close to the generation time. Interestingly, the related variable "Number of pesticide Dosings" correlated positively with LOEC’s (especially on the second PLS axis), meaning that a low but persistent pesticide concentration will have a lower effect on the macroinvertebrates than a high but temporary pesticide concentration. |
| High positive loading for the variable Depth on both axis could be related the fate of pesticides in mesococms. At decreasing mesocosm depth a larger fraction of the pesticides will end up in the sediment compartment and thus increase the exposure to the sediment living macroinvertebrates. |
| For the X variables Log KD, Log KOW and Interval (between dosings) significant negative loadings are obtained for the first PLS axis. The loadings of these variables to the second axis were considered as insignificant. Thus the toxic response of the macroinvertebrates expressed by the first PLS axis are most pronounced for hydrophobic, adsorbable (high log KD) substances added to the mesocosms over a long time period (Interval). In effect, the toxic response of the macroinvertebrates expressed by the first PLS axis might be considered as a long term response probably involving sorption of the pesticides to particles, sedimentation of the particles and a subsequent exposure of the organisms to pesticides adsorbed to sediment particles. |
| High positive loadings to the second PLS axis are obtained for the X variables hazard concentrations (HC5,50 and LC50/10 (i.e. OECD procedure)), while their significance on the 1. axis are considered insignificant. As stated above the loadings of the infauna to the second axis is insignificant (see Fig. 9). Hence, the toxic response of the macroinvertebrates expressed by the second axis can be considered as a short-term response attributable to a direct exposure through the water phase, which consequently do not affect the macroinvertebrates living within the sediment. As the hazard concentrations are calculated from standardised short term (48-96 h) toxicity tests the loadings (correlations) are expected. |
| The variables day number (i.e. season) and volume are considered as insignificant (low loadings and of opposite sign). |
A summary of the effect of experimental mesocosm and pesticide characteristics is shown in Table 6.
Table 6.
Summary of influences of mesocosm and pesticide characteristics and toxicology
(extrapolated effect concentrations) on toxic response on macroinvertebrates. Ý Ý = major decrease in toxicity; Ý = minor decrease in toxicity (i.e. higher LOEC); ß ß = major increase in toxicity; ß = minor increase in toxicity (i.e. lower LOEC); - = no effect. See
Table 2 for an explanation of system variables.
Macroin- |
Sea- son |
Lati- |
Longi- |
Log |
Log |
Inter- |
# of doses |
Depth |
HC5, |
LOEC |
Non- pred. |
- |
ß ß |
Ý Ý |
ß |
ß |
ß |
Ý Ý |
Ý Ý |
ß |
ß |
Preda- tory |
- |
ß ß |
Ý Ý |
ß |
ß |
ß |
Ý Ý |
Ý Ý |
ß |
ß |
Epi- fauna |
- |
ß ß |
Ý Ý |
ß |
ß |
ß |
Ý Ý |
Ý Ý |
ß |
ß |
In- fauna |
- |
ß |
Ý |
ß |
ß |
ß |
Ý |
Ý |
- |
- |
The arrows in Table 6 indicate if numeric increases in system variables (see Table 2) will decrease ( Ý ) or increase ( ß ) the toxic response in the different groups of macroinvertebrates. Double arrows denote that a system variable have the same significant influence in both PLS axes.
5.4.2 Predicting effect concentrations for the macroinvertebrates with the aid of the PLS models
The observed and predicted effect concentrations with associated 95 % confidence intervals for all the mesocosm experiments analysed by the PLS model appear in Table 7 and Figure 10. The asymmetric confidence interval is due to the logarithmic transformation of data before the PLS analysis. When the lower limit of the confidence intervals was below 0 (seemingly a bug occurring in the Simca program during log- and antilog transforming procedure) the lower limit of the confidence interval was set to 0.
Table 7.
Comparison of observed and predicted LOEC (µg l-1) with associated 95 %
confidence interval for the mesocosm experiment calculated with the PLS model for
macroinvertebrates. Macroinvertebrates have been grouped according to mode of feeding
(non-predatory/predatory) and habitat (epifauna = macroinvertebrates living on sediment
surface; infauna = macroinvertebrates living within the sediment).
Exp. |
Pesticide |
Non- pred |
Confidence |
Preda- tory |
Confidence |
||||
|
|
Observ |
Pred. |
Low |
Upp |
Observ |
Pred. |
Low |
Upp |
83tll |
cyfluthr |
2.625 |
2.581 |
0.386 |
5.947 |
4.116 |
2.302 |
0.266 |
5.490 |
84tll |
cyfluthr |
2.625 |
2.577 |
0.572 |
5.516 |
2.625 |
2.252 |
0.428 |
4.977 |
42flm |
chlorpyr |
0.010 |
0.035 |
0.001 |
0.093 |
0.100 |
0.091 |
0.000 |
0.249 |
57flm |
esfenval |
0.160 |
0.041 |
0.004 |
0.100 |
0.160 |
0.116 |
0.007 |
0.294 |
123fl |
lamb_cyh |
0.068 |
0.200 |
0.071 |
0.374 |
0.068 |
0.343 |
0.113 |
0.661 |
57tll |
Diazinon |
36.800 |
30.70 |
0.000 |
97.13 |
9.600 |
11.35 |
0.000 |
37.61 |
110tll |
carbofur |
5.000 |
1.982 |
0.000 |
6.083 |
-- |
1.516 |
0.000 |
4.865 |
47flm |
esfenval |
1.279 |
2.206 |
0.091 |
5.684 |
1.279 |
2.119 |
0.010 |
5.675 |
60flm |
lamb_cyh |
0.068 |
0.045 |
0.007 |
0.102 |
-- |
0.113 |
0.014 |
0.267 |
Exp. |
Pesticide |
Epifauna |
Confidence |
Infauna |
Confidence |
||||
|
|
Observ |
Pred. |
Low |
Upp |
Observ |
Pred. |
Low |
Upp |
83tll |
cyfluthr |
2.625 |
1.789 |
0.436 |
3.747 |
2.625 |
2.910 |
1.390 |
4.843 |
84tll |
cyfluthr |
2.625 |
1.782 |
0.553 |
3.495 |
2.625 |
3.274 |
1.723 |
5.197 |
42flm |
chlorpyr |
0.010 |
0.024 |
0.003 |
0.056 |
0.100 |
0.194 |
0.076 |
0.352 |
57flm |
esfenval |
0.020 |
0.028 |
0.005 |
0.061 |
0.160 |
0.101 |
0.044 |
0.174 |
123fl |
lamb_cyh |
0.068 |
0.136 |
0.059 |
0.237 |
0.680 |
0.504 |
0.310 |
0.735 |
57tll |
Diazinon |
9.600 |
21.13 |
0.000 |
58.89 |
88.00 |
64.90 |
17.32 |
132.8 |
110tll |
carbofur |
5.000 |
1.342 |
0.000 |
3.639 |
-- |
8.725 |
2.504 |
17.50 |
47flm |
esfenval |
1.279 |
1.532 |
0.221 |
3.550 |
-- |
2.148 |
0.866 |
3.844 |
60flm |
lamb_cyh |
0.068 |
0.030 |
0.008 |
0.063 |
-- |
0.179 |
0.087 |
0.296 |
Figure 10.
Plots between observed (µg l-1) and predicted LOEC by PLS models for
macroinvertebrates allocated to different modes of feeding and their habitat. X=Y
(stippled line) shown.
As depicted in Fig. 10 the LOEC predicted by the PLS models were in excellent agreement with the observed LOEC, irrespective of the chosen grouping of macroinvertebrates.
5.5 PLS models for zooplankton
An overview of raw data from the database used in the PLS models developed for zooplankton is shown in Table 8. The analyses are based on data from 31 different experiments with a total of 14 different pesticides.
The PLS models examined for zooplankton are summarised in Table 9.
The PLS model with the highest predictability (0.736) for zooplankton was obtained when the pesticides were added as single addition and the analysis was restricted to insecticides only (see Table 9). However, an almost as high predictability was obtained when the analysis included both insecticides and herbicides (0.655). For the remaining PLS models examined lower and more inconsistent predictabilities (Q2(cum)) were obtained. The interpretation of the PLS models was therefore restricted to the PLS models for mesocosm experiments with a single addition of insecticides (i.e. a total of 11 experiments).
Table 8.
Overview of raw data from the database for the PLS models developed for zooplankton.
See Table 2 for abbreviations used. Sediment 1 refers to sediment present in mesocosm, 0
to no sediment; Macrophytes: 1 = present, 0 = no macrophytes, N= no information given on
presence of marcophytes; Field/lab: 1 = Field study, 0 = Laboratory study. For all
zooplankton groups the lowest effect concentrations observed for each taxonomic group in
each mesocosm experiment were used as Y variables, expressing the toxic response of the
organisms in the mesocosms (values shown in bold). L = Laboratory study at controlled
temperature and light availability (hence latitude and longitude not relevant); F =
flow-through study; - = no data. See Annex B for literature references.
Table 9.
Predictability (Q2(cum)) of the examined PLS models for zooplankton. The
procedure for removal of outliers is explained in Section 5.1.
Data included |
Outliers |
Q2(cum) |
All experiments for stagnant water |
experiment 38tll |
0.346 |
All experiments with sediment |
experiment 38tll |
0.239 |
All experiments with macrophytes |
none |
0.596 |
All experiments with insecticides |
none |
0.206 |
All experiments with a single addition of pesticides |
experiment 120tll, 64flm and 118tll |
0.233 |
All experiments with sediment and a single addition of pesticides |
experiment 64flm |
0.424 |
All mesocosms experiments |
experiment 38tll |
0.454 |
All mesocosms experiments with sediment |
experiment 57flm, 38tll, 106tll and 64flm |
0.450 |
All mesocosms experiments with insecticides |
experiment 106tll and 64flm |
0.184 |
All mesocosms experiments with single addition |
none |
0.655 |
All mesocosms experiments with single addition restricted to insecticides |
none |
0.736 |
5.5.1 Interpretation of the PLS model for zooplankton.
The prediction of the model selected in the previous section is shown Table 10.
Table 10.
Predicted variation of the significant axis of the PLS model selected for zooplankton.
Q2: Variation in the Y matrix predicted from the variation in the X matrix by
the current axis. Q2(cum): Cumulative variation in the Y matrix predicted from
the variation in the X matrix.
PLS axis number |
Q2 |
Q2(cum) |
1 |
0.736 |
0.736 |
2 |
-0.053 |
0.736 |
As shown in Table 10 the first PLS axis predicts 73.6 % of the variation in the Y matrix from the variation in the X matrix. The second PLS axis predicts 0 % of the variation in the Y matrix from the variation in the X matrix. Hence, the second axis does not contribute to the overall predictability of the model and an interpretation of the second axis was therefore not carried out.
Figure 11.
Weights (loadings) of variables contributing to the first PLS axis for zooplankton.
Day number through OECD represent variables in the X-matrix while the responses (LOEC) of
the different zooplankton groups are shown at right.
From the loadings of the different variables (see Fig. 11) to the PLS axis it appears:
The axis primarily represents a "traditional" toxicity axis with positive correlations between hazard concentrations (HC5,50 and LC50/10 = OECD) and LOEC obtained in the mesocosms. In specific:
| Loadings were most positive for the cladocerans, least positive for the rotifers and intermediate for the copepods. Hence, cladocerans semingly are the most sensitive zooplankters to insecticides followed by copepods and rotifers. |
| Positive loadings were obtained for the variables Day number and latitude, whereas a negative loading was obtained for longitude. Thus insecticides seem to be less toxic to the zooplankton (i.e. high LOEC) if experiments are conducted in cold climates (high latitude) and/or during in the summer (high Day #). The effect of latitude is in contradiction to the effect of climate on macroinvertebrates, but could be due to a higher activity of zooplankters and thus exposure to pesticide at higher temperatures. On the other hand, the negative correlation between Day# and LOEC, does not support such relationship. The negative loading for Longitude suggests that zooplankton in experiments conducted in USA are more sensitive than the zooplankton in experiments conducted in Europe. This is in contradiction to the response of macroinvertebrates. As for macroinvertebrates the deviation between European studies and studies carried out in USA could be related to the stocking of fish in enclosures in USA. |
| For the variables expressing hazard concentrations (HC5,50 and EC50/10 = OECD) positive loadings were obtained, whereas a negative loading was obtained for log KOW. Hence, as expected hydrophobic substances characterised by high single species toxicity seems to be most toxic to the zooplankters in the mesocosmos. |
A summary of the effect of experimental mesocosm and pesticide characteristics on response of zooplankton is shown in Table 11.
The arrows in Table 11 indicate if numeric increases in system variables (see Table 2) will decrease ( Ý = high LOEC) or increase ( ß = low LOEC) the toxic response in the different groups of zooplankton.
Table 11.
Summary of influences of mesocosm characteristics, pesticide characteristics and
toxicology (extrapolated effect concentrations) on toxic response on zooplankton. Ý = decrease in toxicity (i.e. higher LOEC); ß
= increase in toxicity (i.e. lower LOEC);
- = no effect. See Table 2 for an explanation of system variables.
Zooplankt. |
Day# |
Lati- tude |
Longi- tude |
Log |
Log |
Volu- |
Depth |
HC5,50 |
LC50 |
Cladocera |
Ý |
Ý |
ß |
- |
ß |
(Ý ) |
- |
Ý |
Ý |
Copepoda |
Ý |
Ý |
ß |
- |
ß |
(Ý ) |
- |
Ý |
Ý |
Rotifera |
Ý |
Ý |
ß |
- |
ß |
(Ý ) |
- |
Ý |
Ý |
5.5.2 Predicting the effect concentrations for the zooplankton with the aid of the PLS model
The observed and predicted effect concentrations with associated 95 % confidence interval for the mesocosmos experiment calculated with the PLS model for zooplankton are shown in Table 12 and Figure 12. When the lower limit of the confidence intervals was below 0 the lower limit of the confidence interval was set to 0 (see section 5.4.2).
Figure 12.
Plots between observed LOEC (µg l-1) and LOEC predicted by PLS model for
zooplankton (Cladocera, Copepoda & Rotifera). X=Y (stippled line) shown.
Table 12.
Comparison of observed and predicted effect concentrations with associated 95 %
confidence interval for the mesocosmos experiment calculated with the PLS model for
zooplankton. -- = no observation. See Annex B for references.
Generally, within the observed interval the effect concentrations predicted by the PLS model were in excellent agreement with the observed effect concentrations for both cladocerans and copepods, while the PLS model tended to over-estimate effect concentrations for rotifers. Such deviation could be expected, however, as effects of insecticides on rotifers primarily was of indirect nature (i.e. increases in abundance due to reduced competition from crustecean zooplankters).
5.6 PLS models for microalgae
An overview of raw data from the database used in the PLS models developed for microalgae is shown in Table 13. In total only 9 mesocosm experiments with toxicity data for microalgae were available from the data base. Thus it was only possible to consider the scenarios including either all experiments or all mesocosm experiments carried out in the field. Of these two scenarios the highest predictability (0.721) was obtained for the scenario of field mesocosm experiments (Table 14).
Table 14.
Predictability of the examined PLS models for microalgae
Data included |
Outliers |
Q2(cum) |
All experiments |
none |
0.671 |
Field Mesocosm experiments |
none |
0.721 |
For the selected PLS model with micro algae only one significant PLS axis was present (Table 15).
Table 15.
Predicted variation of the significant axis of the PLS model selected for micro algae.
Q2: Variation in the Y matrix predicted from the variation in the X matrix by
the current axis. Q2(cum): Cumulative variation in the Y matrix predicted from
the variation in the X matrix.
PLS axis number |
Q2 |
Q2(cum) |
1 |
0.721 |
0.721 |
From the plots of loadings and of variable importance the following interpretation of the PLS axis of the PLS model for micro algae is conducted:
| The highest (and positive) loading was found for the X variable Interval (between pesticide dosings). Thus, pesticides added over a short period are most toxic to the algae in the mesocosmos. This is probably related to the short generation time of microalgae: frequent dosings will prevent microalgae to recover, while one or less frequent dosings will allow the microalgae to recover, when the pesticide dissipates. |
| Positive loadings were seen for the X variables expressing the extrapolated hazard concentrations (HC5,50 and EC50/10 = OECD procedure), whereas negative loadings were seen for the variables log KD and log KOW. Hence, hydrophobic and adsorpable substances with high single species toxicity were most toxic to the micro algae in the mesoscosmos. |
Figure 13.
Weights (loadings) of variables contributing to the PLS axis for microalgae.
The observed and predicted effect concentrations with associated 95 % confidence interval for the mesocosmos experiment handled with the PLS model for micro algae appear is shown in Table 16 and Fig.14. When the lower limit of the confidence intervals was below 0 the lower limit of the confidence interval was set to 0 (see section 5.4.2).
Table 16.
Comparison of observed and predicted LOEC’s (µg l-1) with associated
95 % confidence interval for the mesocosmos experiment calculated with the PLS model for
microalgae.
Exp. |
Pesticide |
Microalgae |
Confidence |
||
Observ |
Pred. |
Lower |
Upper |
||
16ank |
Atrazin |
225 |
13.134 |
0 |
37.23 |
59flm |
esfenval |
3.60 |
2.665 |
0 |
7.595 |
121flm |
fenpropi |
0.60 |
0.176 |
0 |
0.712 |
122flm |
fenpropi |
0.58 |
0.151 |
0 |
0.629 |
123flm |
lamb_cyh |
0.68 |
0.195 |
0 |
0.778 |
38tll |
gluf_amm |
2000 |
1954.7 |
0 |
11728 |
72tll |
Atrazin |
160 |
180.86 |
0 |
710.5 |
113tll |
esfenval |
0.035 |
0.5381 |
0 |
1.826 |
Table 13.
Overview of raw data from the database for the PLS models developed for microalgae
(phytoplankton & periphytes). See Table 2 for abbreviations used. For Sediment 1 refer
to sediment present in mesocosm, 0 to no sediment; Macrophytes: 1 = present, 0 = no
macrophytes; Field/lab: 1 = Field study, 0 = Laboratory study. For all microalgal groups
tested the lowest effect concentrations observed in each mesocosms experiment were used as
Y variables, expressing the toxic response of the organisms in the mesocosms (values shown
in bold). L = Laboratory study at controlled temperature and light availability (hence
latitude and longitude not relevant); - = no data. See Annex B for literature references.
Figure 14.
Plots between observed LOEC (µg l-1) and LOEC predicted by PLS model for
microalgae (phytoplankton + periphytes). X=Y (stippled line) and linear regression
equation shown.
Within the observed interval the effect concentrations predicted by the PLS model were in excellent agreement with the observed effect concentrations for microalgae.
5.7 Summary of PLS models
The amounts of data available for the different communities were quite variable and a direct comparison of PLS models should therefore be conducted with caution.
Macroinvertebrates
To obtain a PLS model with a reasonable predictability of the toxic effects to various macroinvertebrate groups, mesocosms should contain sediment and preferably macrophytes in the test system. Overall, the model developed was able to predict 63 % of the observed effects among macroinvertebrates.
In summary, the PLS analysis showed that
- All macroinvertebrate groups in the mesocosms seem to be most sensitive when the experiments are conducted at high latitudes. Therefore, toxic effects at lower concentrations are expected with increasing distance from Equator, which may be due to a slower turn-over of populations at high latitudes, i.e. fewer generations each year at lower temperatures. Therefore, recovery of populations affected by pesticide exposure takes longer time at northern latitudes.
- Macroinvertebrates living within the sediment (i.e. infauna) were less sensitive the pesticides than macroinvertebrates living on the sediment surface.
- At a given total dose the effect of pesticides decreases with number of pesticide additions. Therefore, a low but persistent pesticide concentration will have a lower effect on the macroinvertebrates than a high but temporary pesticide concentration.
- The toxic effects of pesticides are most pronounced in shallow mesocosms. At decreasing mesocosm depth a larger fraction of the pesticides will end up in the sediment compartment and thus increase the exposure to the sediment living macroinvertebrates. This interpretation is further reinforced by the inverse relation between Log KD of pesticides and toxicity to invertebrates.
Zooplankton
The PLS model with the highest predictability for zooplankton was obtained when the pesticides were applied as single addition and the analysis was restricted to insecticides only.
The PLS analysis showed that
| 3. | Hydrophobic insecticides with high single species toxicity were the most toxic to the zooplankters in the mesocosmos. |
| 4. | Cladocerans were the most sensitive group to insecticides followed by copepods and rotifers. |
| 2. | The effect of climate zone (latitude) and season was contradictory, as the highest sensitivity was obtained at low latitudes but outside the summer months. |
Microalgae
The highest predictability of pesticide effects to microalgae was obtained when only field mesocosm experiments were included in the analysis.
The PLS analysis showed that
| 3. | Hydrophobic and adsorpable pesticides with high single species toxicity were the most toxic to the micro algae in the mesoscosmos. |
| 4. | At a given total dose pesticides added over a short period were more toxic to the algae in mesocosmos than pesticides dosed at longer intervals. Frequent dosings will prevent microalgae to recover, while microalgae characterised by short generation times will be able to recover in between dosings applied at longer intervals. |