Using additive modelling to quantify the effect of chemicals on phytoplankton diversity and biomass

Full text

Using additive modelling to quantify the effect of chemicals on phytoplankton
diversity and biomass
K.P.J. Viaene
a,
⁎, F. De Laender
a
, P.J. Van den Brink
b,c,1
, C.R. Janssen
a
a
Laboratory of Environmental Toxicity and Aquatic Ecology, Ghent University, Plateaustraat 22, 9000 Ghent, Belgium
b
Department of Aquatic Ecology and Water Quality Management, Wageningen University, PO Box 47, 6700 AA Wageningen The Netherlands
c
Alterra, PO Box 47, 6700 AA Wageningen, The Netherlands
HIGHLIGHTS
►Richness and evenness decreased while dominance increased after linuron addition.
►Richness was affected at lower linuron concentrations and showed slower recovery.
►Initial biodiversity was negatively correlated with subsequent biodiversity.
►Primary production was unaffected by linuron.
►Biodiversity was affected at higher concentrations than individual species.
abstractarticle info
Article history:
Received 17 October 2012
Received in revised form 19 December 2012
Accepted 11 January 2013
Available online xxxx
Keywords:
Biodiversity
Pesticides
Algal communities
Additive modelling
Environmental authorities require the protection of biodiversity and other ecosystem properties such as bio-
mass production. However, the endpoints listed in available ecotoxicological datasets generally do not con-
tain these two ecosystem descriptors. Inferring the effects of chemicals on such descriptors from micro- or
mesocosm experiments is often hampered by inherent differences in the initial biodiversity levels between
experimental units or by delayed community responses. Here we introduce additive modelling to establish
the effects of a chronic application of the herbicide linuron on 10 biodiversity indices and phytoplankton bio-
mass in microcosms. We found that communities with a low (high) initial biodiversity subsequently became
more (less) diverse, indicating an equilibriumbiodiversity status in the communities considered here. Linuron ad-
versely affected richness and evenness while dominance increased but no biodiversity indices were different from
the control treatment at linuron concentrations below 2.4 μg/L. Richness-related indices changed at lower linuron
concentrations (effects noticeable from 2.4 μg/L) than other biodiversity indices (effects noticeable from 14.4 μg/L)
and, in contrast to the other indices, showed no signs of recovery following chronic exposure. Phytoplankton bio-
mass was unaffected by linuron due to functional redundancy within the phytoplankton community. Comparing
thresholds forbiodiversity withconventional toxicity testresults showed that standard ecological risk assessments
also protect biodiversity in the case of linuron.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
Ecological risk assessment aims to protect the structure and function-
ing of ecosystems (De Laender et al., 2008a; Forbes et al., 2009). Biodiver-
sity is generally considered a useful descriptor of ecosystem structure
and its role in ecosystem productivity and stability is generally accepted
in the ecological literature (Hooper et al., 2005, 2012). From an ecotoxi-
cological point of view, biodiversity sensu Hooper et al. (2005) is pro-
posed to be important because it insures against declines in ecosystem
functioning when exposed to (toxic) stress, a statement referred to as
the insurance hypothesis (Yachi and Loreau, 1999). More diverse com-
munities should thus better ensure ecosystem function maintenance
when facing perturbations by chemicals than less diverse communities.
These insights, which emerged in the early 1990s, led to the inclusion
of biodiversity in the environmental protection goals set by different reg-
ulating organisations. For example, the Convention on Biological Diversi-
ty by the United Nations states that “by 2020 the extinction of known
threatened species has been prevented and their conservation status,
particularly of those most in decline, has been improved and sustained”
(United Nations, 1992). Another example is the Water Framework Direc-
tive (WFD) that commits all European surface and groundwater bodies
to have a good ecological status by 2015 (EU, 2000).
To assess the potential effects of a chemical on ecosystems, one
typically extrapolates the results of single species toxicity tests to
Science of the Total Environment 449 (2013) 71–80
⁎Corresponding author. Tel.: +32 9 264 3779; fax: +32 9 264 4199.
E-mail addresses: karel.viaene@ugent.be (K.P.J. Viaene),
frederik.delaender@ugent.be (F. De Laender), Paul.vandenBrink@wur.nl
(P.J. Van den Brink), colin.janssen@ugent.be (C.R. Janssen).
1
Tel.: +31 317 481615; fax: +31 419000.
0048-9697/$ –see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.scitotenv.2013.01.046
Contents lists available at SciVerse ScienceDirect
Science of the Total Environment
journal homepage: www.elsevier.com/locate/scitotenv

the ecosystem level (Forbes and Calow, 2002), using e.g. statistical
distributions such as the species sensitivity distribution (Posthuma
et al., 2002). This practice has been criticized for being based on un-
realistic assumptions (De Laender et al., 2008b; Forbes et al., 2001).
In addition, many of the extrapolation approaches that are based on
single species toxicity data alone neglect essential ecological concepts
such as species interactions, the resulting indirect effects (Fleeger et al.,
2003; Relyea and Hoverman, 2006), functional redundancy (De Laender
et al., 2011b) and/or recovery (Relyea and Hoverman, 2006). Instead,
ecological modelling and multi-species experiments using micro-
or mesocosms have been proposed as useful approaches to enhance
the ecological relevance of effect assessments (Van den Brink, 2008).
Microcosm and mesocosm experiments allow a high degree of con-
trol and replication while they can be used to study the effects of
stress on higher levels of biological organisation, such as populations
and communities (Van den Brink, 2006). Likewise, mathematical
models that simulate population- (Preuss et al., 2010) and ecosystem-
level effects of chemicals (De Laender et al., 2008b) are useful tools for
testing hypotheses underlying current risk assessment paradigms (De
Laender et al., 2008a, 2010). Ideally, modelling approaches are com-
bined with data from micro- and mesocosm studies to validate theoret-
ical models (Dueri et al., 2009) or obtain new information on the
sensitivity of the ecosystem's functional aspects, e.g. De Laender
et al. (2011b). Despite the possibilities that models and experiments
offer, biodiversity and functioning of ecosystems stressed with chemicals
have rarely been studied. A notable exception is the simulation study of
de Vries et al. (2010) which examined the direct chemical effects on bio-
diversity indices in marine communities. These authors predicted that
SSD-derived HC
5
-values are protective for biodiversity, as quantified by
species richness or the Shannon–Wiener index, for 99.6% of the sub-
stances evaluated (de Vries et al., 2010). However, this was based purely
on simulated data and neglected indirect chemical effects, making its ap-
plicability to real-world situations limited. A recent literature analysis on
the effects of chemicals on the biodiversity of fluvial communities con-
cluded that the relationship between biodiversity and chemical exposure
is not always easy to interpret and that the reliance on a single biodiver-
sity index is inadequate to fully comprehend the effects of chemicals on
biodiversity (Ricciardi et al., 2009). This indicates that techniques for
analysing the relationship between chemicals and biodiversity should
be able to cope with this complexity.
In this paper, we introduce additive mixed modelling (Zuur et al.,
2009) to investigate if and to what extent the biodiversity of experi-
mental multispecies systems is altered when exposed to a chemical
stressor. Additive modelling can deal with non-linear relationships
between predictor variables (e.g. chemical concentration) and re-
sponse variables (e.g. species richness), not requiring the a priori
definition of any functional relationship. Also, additive models can
account for confounding variables that may blur the focal relation-
ships and, when combined with mixed models, for autocorrelation
that is typical for time series data (Zuur et al., 2007). We demon-
strate this technique for the case of phytoplankton communities in
macrophyte-dominated, aquatic microcosms (Van den Brink et al.,
1997) exposed to the photosynthesis-inhibiting herbicide linuron.
The relationships between 14 biodiversity indices, one ecosystem
property (phytoplankton biomass) and a chemical stressor (the her-
bicide linuron) are quantified, taking into account temporal dynam-
ics and the initial values of biodiversity and ecosystem properties.
2. Materials and methods
2.1. Microcosm experiment
To assess the effects of a chemical stressor on phytoplankton bio-
diversity (BD), we used phytoplankton abundance data from a previ-
ously conducted microcosm experiment with linuron (Van den Brink
et al., 1997). We limited our study to phytoplankton biodiversity and
did not include other primary producers (macrophytes and periphyton)
because phytoplankton was sampled to the species-level, allowing bio-
diversity reconstruction. Also, the dominant primary producer group in
the microcosms was limited to one species –the macrophyte Elodea
nuttallii –and the effects of linuron on its biomass production have
been studied in the original paper. The abundances of phytoplankton
species were measured from 1 week before (week 0: pre-treatment pe-
riod) until 11 weeks after (weeks 1–5: treatment and weeks 6–12:
post-treatment period) linuron addition to 600 L microcosms. These
microcosms were exposed to six different chronic linuron treatments
(0, 0.5, 5, 15, 50 and 150 μg/L), eachreplicatedtwice. The actual linuron
concentration was maintained constant for 4 weeks by measuring it
twice per week and by adding more linuron to compensate for losses
(Fig. S1). Further details can be found in the original paper (Van den
Brink et al., 1997).
2.2. Calculation of biodiversity and phytoplankton biomass
Based on the phytoplankton species abundances, 14 biodiversity
(BD) indices were calculated using the software packages vegan,
vegetarian and BiodiversityR in the statistical program R (version
2.14.0). These BD indices were: species richness, rarefied richness,
Margalef index, α-parameter of Fisher's log-series, Shannon–Wiener
index, Shannon evenness, Lloyd and Ghelardi evenness, Simpson diver-
sity, Berger Parker index, McNaughton dominance, McIntosh diversity,
McIntosh evenness, α-diversity and Menhinick index (formulas and
references: Table S1). These BD indices take into account different as-
pects of biodiversity e.g. some are calculated based on species richness
(e.g. Fisher's α), some on the dominance of species (e.g. Berger Parker
index), some on the evenness i.e. how the abundances of the different
species are distributed (e.g. Shannon evenness) and some on a combi-
nation of these aspects (e.g. Shannon Wiener index). Higher BD index
values always denote a higher biodiversity, except for the dominance-
related indices. The α-parameter of Fisher's log-series and the Lloyd
and Ghelardi evenness are calculated based on different species–
abundance relationships, i.e. the broken stick model and the log series,
respectively. To test the applicability of these two BD indices, the fitof
the phytoplankton community to the broken stick and log series model
was investigated. The species–abundance relationship showed a bad fit
for both species–abundance models (Fig. S22 for a typical example)
and both were thus omitted from further analyses. Phytoplankton bio-
mass was calculated by multiplying the species abundances with the
corresponding cell weights listed in De Laender et al. (2011a).
2.3. Data exploration
An initial data exploration showed large pre-treatment differences in
BD indices between the microcosms. For example, the species richness
before linuron application varied over almost the same range (between
7 and 15) as the species richness after linuron application (between 4
and 16). Because these differences could confound the potential linuron
effects picked up by the statistical modelling, each BD index was normal-
ized by subtracting the pre-treatment value from the treatment and
post-treatment values. BD index values given in this study are thus
normalized to the pre-treatment value, i.e. negative values indicate
a decrease of the BD index relative to its pre-treatment value, posi-
tive values indicate the opposite. Phytoplankton biomass showed
no large pre-treatment differences: pre-treatment phytoplankton bio-
mass ranged between 4 and 29 ng C/L while phytoplankton biomass
in the weeks after linuron application ranged between 3 and 275 ng
C/L. Therefore, phytoplankton biomass was not normalized.
2.4. Model construction
For each BD index, we examined if time and the linuron concen-
tration significantly contributed to the variability of the index values.
72 K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

In addition to the time and linuron effects, we also tested if the non-
normalized initial (pre-treatment) biodiversity (BD
0
)significantly
influenced the normalized biodiversity after linuron application. To
this end, generalized additive mixed models (GAMMs; (Wood, 2006))
were used in a time series approach (Zuur et al., 2007). GAMMs de-
scribe non-parametric relationships between the predictor and the re-
sponse variable(s) by smoothing functions. The advantage GAMMs
offer over parametric regression techniques is that no prior assump-
tions on the shape of the relation between response and predictor
are needed; instead these relations are obtained by fitting the GAMM
to the data. In addition, GAMMs can account for temporal autocorrela-
tion of model residuals, a common feature of time series data leading
to a violation of the independence assumption on which conventional
regression techniques are based.
Candidate predictor variables for the GAMMs were time (t), linuron
concentration (C) and initial (pre-treatment) biodiversity index (BD
0
).
Although we normalized the BD indices relative to their initial values
to compareBD among linuron treatments, we also added BD
0
as explan-
atory variable. We did so to allow effects of the initial BD on the subse-
quent responses of BD to the herbicide. For example, a community with
a high richness may be more resistant to chemical treatments than a
species poor community. Linuron concentrations were measured two
times a week, while abundances (and thus BD indices) were assessed
once a week. To obtain estimates of the linuron concentrations for
every BD value, linear interpolation between the measured linuron con-
centrations was used as described in Van den Brink et al. (1997).The
linuron concentrations C were transformed as log
10
(C+ 1). The GAMMs
were thus structured as:
EBD
tα;C;BD0
j¼αþf1t;log10 Cþ1ðÞ

þf2BD0
ðÞ:
ð1Þ
With E[BD
t
|α,C,BD
0
] the expected value of BD on time t, given a
set of parameter values α,C,BD
0
. The two smoothing functions f
1
and
f
2
denote the combined effect of time and C and the effect of BD
0
, re-
spectively. One smoothing function f
1
(t, log
10
(C+ 1)) included time
and the log-transformed linuron concentration to account for possi-
ble interactions between the two predictor variables in their effect
on biodiversity, while still allowing both predictor variables to inde-
pendently affect biodiversity. The reason for this strategy was that
the effect of linuron was only noticeable after a few weeks i.e. effects
of linuron were time-dependent. This was explained by Van den Brink
et al. (1997) as species surviving for a few weeks on internal energy
storages, suggesting the effect of linuron changed over time. The resid-
uals of this model (ε
tij
)canbewrittenasρ·ε
t−1,ij
+γ
tij
(Auto-regressive
model of order 1 auto-correlation structure; i =replicate; j =treatment;
t=time, ε
t
=residual at time t; ρ= correlation parameter; ε
t−1
=
residual at time t−1; γ=noise; see Zuur et al. (2007) for further
details).
Prior to fitting the model (Eq. (1)) to the data for the 14 biodi-
versity indices, correlations and variance inflation factors (VIFs)
between the three predictor variables (time, linuron and initial
biodiversity) were examined, so as to avoid problems of collinear-
ity. Although the VIFs were never higher then 2.7, high correlation
coefficientsrangingfrom0.43to0.74(absolutevaluesofPearson
correlation) were found between measured linuron concentra-
tions and non-normalized initial values of most biodiversity indi-
ces. Therefore, the model (Eq. (1))couldnotbefitted to all the
data in one effort. To solve this, we divided the data set into two
subsets: the data at the low linuron treatments (≤5μg/L) and the
data at the high linuron treatments (>5 μg/L). By doing so, the correla-
tions between the initial BD indices and the linuron concentrations
were lowered to values ≤0.62 except for the Shannon–Wiener (0.74)
and alpha diversity indices (0.77), which were consequently omitted
from further analyses. For each of the 10 retained BD indices, the
resulting final model, after creating two subsets from the original
dataset, is given as:
EhBDtα;C;BD0


i¼αþf1t;log10 Cþ1
ðÞ

Lþf2BD0
ðÞ
L
þf3t;log10 Cþ1ðÞðÞHþf4BD0
ðÞHþεtij ð2Þ
with all symbols as in (Eq. (1)), and L= low linuron treatments; H=
high linuron treatments. If the linuron concentration is lower than or
equal to 5 μg/L, then L= 1 and H = 0. If the linuron concentration is
higher than 5 μg/L, then L =0 and H =1.
Phytoplankton biomass was modelled in the same way as biodiver-
sity. The GAMM for phytoplankton biomass (expressed as biomass, BM)
was given as:
Elog
10 BMt
ðÞα;C;BM0


i¼αþf1t;log10 Cþ1ðÞ

þf2BM0
ðÞþεtij
hð3Þ
with all symbols as in Eq. (1),andBM
0
=initial biomass.
2.5. Model fitting
The final model was fitted separately for each of the BD indices
(Eq. (2)) and for the phytoplankton biomass (Eq. (3)) using the R pack-
age ‘mgcv’(Wood, 2011). After fitting, we inspected if the approximate
p-values of the three predictor variables were below a critical threshold
(0.05) to judge their significance. Non-significant terms were dropped
from the model. In addition to Eq. (2), we tested if more complex
GAMMs with additional smoothing functions of other predictors or com-
binations of predictors, e.g. f(log
10
(C+ 1), BD
0
), better fitted the data
using the Akaike information criterion (AIC) and the Bayesian informa-
tion criterion (BIC). The AIC is a goodness-of-fit measure that rewards
proximity of a model to the data but penalizes model complexity, thus
protecting against overfitting. The BIC is a similar goodness-of fitmea-
sure but uses a higher penalty than the AIC for model complexity. The re-
siduals (the differences between observations and predictions) of this
final model were inspected: we evaluated relations between residuals
and predictor variables; the normality of the residuals was tested using
a QQ-plot and plots with the predicted versus observed values were
checked. Models were only considered reliable –and thus retained –if
no such relations were visually observed, residuals were normally dis-
tributed and when no clear deviations of the 1:1 relationship were ob-
served for predicted and observed values.
Using the fitted models for which residual diagnostics suggested
model reliability, biodiversity and phytoplankton biomass were pre-
dicted for a number of combinations of time, linuron concentration
and initialbiodiversity or phytoplankton biomass. The goal of this exer-
cise was twofold: (1) extensive inspection of model fit and (2) the iso-
lation of the linuron and time effect from potential effects of initial
diversity or phytoplankton biomass. To inspect the model fits for the
BD indices, predictions were first made for the same combinations of
time, linuron concentration and BD
0
as in the originalmicrocosm design
(Van den Brink et al., 1997). The means of the two replicate BD
0
index
values per treatment were used for the predictions. The resulting pre-
dictions were compared to the original values to assess the model fit.
To isolate the effects of BD
0
and BM
0
from the effects of linuron and
time, predictions were made using one BD
0
value across all treatments.
Because statistical models can only be used in the range of the data to
which they are fitted, the mean of the 5 and 15 μg/L linuron treatments
BD
0
indices was taken as BD
0
for all treatments. By doing so,we selected
aBD
0
index within range of both subsets that thus allowed reliable pre-
dictions. Next, these mean BD
0
indices were used to predict the time
trend of the BD indices for the range of 0 to 150 μg/L linuron. These
BD
0
indices were also used to determine at what concentration and at
what point in time a BD index was significantly affected. Differences
between and general patterns among BD indices were studied and
the overall effect of linuron on the biodiversity of the phytoplankton
73K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

community was assessed. To determine when predicted BD indices
showed significant differences, the procedure described by Schenker
and Gentleman (2001) was used.
To compare the effects of linuron on the biodiversity of the micro-
cosms with effects indicated by single species toxicity data from the
literature, lowest effect concentrations (LECs) were calculated for
each BD index. LECs were defined as the lowest linuron concentration
at which the predicted BD index differed significantly at the 0.05 level
from the predicted BD index of the control treatment. Significant dif-
ferences were tested using the method described by Schenker and
Gentleman (2001).
A similar approach was adopted for phytoplankton biomass. In short,
predictions based on the original data were made to assess model fitand
the effects of linuron and time on BM were studied. When BM
0
was sig-
nificant, predictions with a mean BM
0
were made to rule out confounding
effects of BM
0
. Finally, LECs for phytoplankton biomass were calculated as
described above.
3. Results
3.1. Model selection
All 10 biodiversity (BD) indices were best predicted using time,
linuron concentration and initial biodiversity (BD
0
)aspredictorvari-
ables. Phytoplankton biomass could be predicted using time and linuron
concentration only. The AIC and BIC indicated that more complex model
structures were not supported by the data. For all the models, the model
assumptions were met: there were no patterns in predictor variables
versus residuals plots (Figs. S2–S12), residuals were normally distribut-
ed as indicated by QQ-plots (e.g. for species richness, Fig. 1A) and plots of
the observed versus fitted values showed no consistent deviations from
a 1:1 line (e.g. for species richness, Fig. 1B). In addition, R
2
-values (0.74–
0.85) indicated that the optimal model could explain most of the ob-
served variation. Plots combining the original data with the model pre-
dictions, e.g. for the Simpson diversity index (Fig. 2, other biodiversity
indices: Figs. 3 and S13–20), confirmed the good model fit.
3.2. Effects of initial biodiversity on subsequent biodiversity
Initial BD had a significant negative effect on subsequent biodiver-
sity, regardless of the linuron treatment. This was found for all the BD
indices except for the Simpson and McIntosh diversity indices, where
this relationship was significant in the low linuron treatments only
(Table S2). The negative relationship between BD
0
and subsequent
biodiversity was always linear, as indicated by the estimated degrees
of freedom (1) for this smoothing function (e.g. Fig. S21).
3.3. Effects of linuron and time on biodiversity
During the first 2 weeks of linuron exposure, no significant differ-
ences were found between biodiversity in the linuron treatments and
biodiversity in the control treatment, except in the 5 μg/L linuron treat-
ment (Table S3). In the 5 μg/L linuron treatment, species richness was
negatively affected from week 1 till week 10 and the Margalef index
from week 2 till week 9. A significant effect of linuron in the two highest
linuron treatments (50 and 150 μg/L linuron) was only predicted from
week 3 onwards and in the three highest linuron treatments (15, 50
and 150 μg/L linuron) from week 4 on wards. Independent of the linuron
concentration, all the BD indices stabilized i.e. showed no significant dif-
ferences with the corresponding BD index from the week before, during
the post-exposure period (in week 7, e.g. Figs. 2–3).
The 10 biodiversity indices could be grouped according to their re-
sponse to linuron. A first group of BD indices (species richness and the
Margalef index) showed no significant change with time at the three
lowest linuron concentrations (≤5μg/L, Figs. 3A–C, S15). However, at
the three highest linuron concentrations (≥15 μg/L, Figs. 3D–F, S15)
there was a decline in BD index until week 6 (thus 5 weeks after
the first linuron addition). In the following weeks, the BD indices sta-
bilized. A second group of BD indices (the rarefied richness, Simpson
diversity, Shannon evenness, McIntosh diversity and McIntosh even-
ness) remained stable during the first 4 to 5 weeks, declined during
2 to 3 weeks thereafter and then stabilized in the four lowest con-
centrations (e.g. Figs. 2A–D, S16–19). The two highest treatments
remained stable for 1 week only, after which a decline started until
week 7, followed by a stabilization period (e.g. Figs. 2E–F, S18–21).
A third group (the dominance-related Berger Parker and McNaugh-
ton indices) exhibited a mirrored pattern to the second group where
the BD indices showed an increasing instead of a decreasing pattern
(Figs. S13 and S20). The Menhinick index could not be allocated to
any of the aforementioned groups as it showed a linear decrease
from the beginning of the linuron addition until week 7, with a short
stable period in weeks 3 and 4 for the three lowest linuron concentra-
tions (Fig. S14).
3.4. Time trends of linuron effects on overall biodiversity
To demonstrate the effects of linuron at various points in time on
BD while ruling out potential confounding effects of BD
0
on subse-
quent BD, BD
0
was set to a default value. As default BD
0
value we
chose the mean of the BD
0
of the 5 and 15 μg/L treatments and
thus a BD
0
within the range of the initial BD
0
values for both subsets.
By doing so, we avoided uncertainty resulting from using the models
to predict out of the range of initial values. The resulting predictions
indicated that at the two lowest linuron concentrations (0 and
0.5 μg/L), none of the BD indices were adversely affected at any
point in time (Fig. 4). In the 5 μg/L linuron treatment, 10–20% of
the BD indices were affected from week 1 till week 9. For the three
highest linuron concentrations (15, 50 and 150 μg/L), no effects on
BD could be noted during the first 2 weeks (Fig. 4). From week 3 on-
wards, 10 to 100% of the BD indices were affected at the two highest
linuron concentrations. At the highest linuron concentration (150 μg/L),
all 10 BD indices were significantly affected in weeks 4, 5 and 6 (Fig. 4,
Table S3). Among the affected BD indices were both richness-,
dominance- and evenness-related indices. From week 7 onwards,
2 weeks after the last linuron addition, the number of affected
BD indices started to decrease, indicating recovery of the phyto-
plankton diversity. However, three to four BD indices –species rich-
ness, rarefied richness, the Margalef and the Menhinick index –of the
two highest linuron treatments continued to be significantly lower
than those of the controls, even 6 weeks after the last linuron addition
(Table S3). Richness-related indices were thus affected longer by linuron
than dominance- and evenness-based indices.
3.5. Relationship between single-species toxicity and effects on biodiversity
To depict the location and variability of available toxicity data for
a given chemical, species sensitivity distributions (SSDs) can be
constructed. For a herbicide, only the target group i.e. primary pro-
ducers should be included in the SSD (Van den Brink et al., 2006).
We did so using available chronic toxicity data for linuron (Van
den Brink et al., 2006) and compared this toxicity range with the cal-
culated LECs of BD for two points in time: week 5, when effects were
highest, and week 11, when most BD indices had recovered (Fig. 5;
Table S4). Note that no attempt was made to fit any statistical distri-
bution to the LECs as the latter are not independent data points but
were derived from the same abundance data. Some general conclu-
sions on the relation between single species toxicity and the effects
of linuron on BD can be drawn from this comparison for week 5. First,
linuron concentrations that affect BD appear to be higher than linuron
concentrations that affect individual species. The lowest LEC (2.4 μg/L)
is more than a factor five higher than the most sensitive species-level
endpoint (0.45 μg/L) and a factor four higher than the HC
5
based on
74 K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

this SSD (0.6 μg/L). Second, among the four lowest LECs are three
richness-related indices: species richness (2.4 μg/L linuron), Margalef
index (2.9 μg/L linuron) and rarefied richness (18.8 μg/L linuron). This
implies that the richness aspect of BD is most prone to linuron stress.
As the other BD indices had already recovered by week 11, LECs
could only be calculated for five BD indices in week 11: species rich-
ness (6.6 μg/L), Margalef index (8.6 μg/L), rarefied richness (9 μg/L),
Menhinick index (52.2 μg/L) and the McNaughton index (125.6 μg/L).
Only richness-related indices except for the McNaughton index thus
showed significant differences with the control treatment for linuron
concentrations up to 150 μg/L. These LECs were similar to the LECs of
week 5 and were more than an order of magnitude higher than the low-
est NOEC for microcosms (0.5 μg/L).
3.6. Effects of linuron on phytoplankton biomass
All linuron treatments showed an increase in phytoplankton biomass
with time. The three lowest linuron concentrations (0, 0.5 and 5 μg/L)
showed a significant increase in phytoplankton biomass 2 weeks after
linuron addition (Fig. 6). The 15, 50 and 150 μg/L linuron treatments
showed a significant increase in phytoplankton biomass three, 3 and
5 weeks after linuron addition, respectively (Fig. 6). The difference
between the control and other linuron treatments was however
never significant (Table S3) and thus no LEC for phytoplankton bio-
mass could be calculated.
4. Discussion
4.1. Role of initial biodiversity
A higher initial biodiversity (BD
0
) resulted in BD decreases over
time and vice-versa. This possibly indicates that biodiversity (BD)
oscillates around an equilibrium level, with high and low BD
0
values
moving towards intermediate (equilibrium) values. Such oscilla-
tions have been observed e.g. for species richness in microbial com-
munities (McGrady-Steed and Morin, 2000). The importance of initial
BD was unexpected because the microcosms were connected until the
start of the experiment and earlier analysis showed no indications of
pre-treatment differences (Van den Brink et al., 1997). However, earlier
analyses were performed using the means of replicates while the present
study showed that the differences in initial biodiversity, both between
replicates and between treatments, were considerable and should be
Fig. 1. Observed versus predicted values and QQ-plots (with the 1:1 line) for species richness (A and B) and Simpson diversity (C and D). QQ-plots are used to evaluate normality of
model residuals; if the points approach the 1:1 line, residuals are normally distributed.
75K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

accounted for in our analysis. This has important implications and, if left
unaccounted for, could lead to misinterpretation of the combined
time and linuron effect. For example, the LEC for Simpson diversity
in week 5 is 9 μg/L when not accounting for initial species richness
versus 26 μg/L when accounting for initial species richness which
is almost a factor three difference.
Natural decay of BD in the isolated microcosms could possibly affect
the observed BD patterns. Natural decay was described by the variable
time in the GAMMs (e.g. Fig. 2A) and was observed for all BD indices ex-
cept for species richness and Margalef index (Figs. 3A and S15A). As nat-
ural decay was described by the time variable in the models, the
observed effects of initial BD –and of linuron –were independent of
any natural decay.
4.2. Effect of linuron on phytoplankton biodiversity
In general, the addition of linuron to the microcosms led to a de-
crease in richness- and evenness-related BD indices while domi-
nance indices increased. Wellman et al. (1998) observed decreases
in the Shannon–Wiener diversity and Shannon evenness of plankton
communities as exposure to methabenzthiazuron increased. More-
over, they observed an increased dominance of a few tolerant spe-
cies. Other studies observed the dominance of tolerant species as
sensitive species were lost by herbicide toxicity e.g. for fomesafen
(Caquet et al., 2005). Similarly, we observed increases in Berger
Parker and McNaughton indices after linuron application, indicating the
dominance of a few –more tolerant –species, especially Chlamydomonas
sp. (Van den Brink et al., 1997). The clear decrease of the dominant
macrophyte E. nuttallii and the reduced competition with sensitive
species offered opportunities for tolerant phytoplankton species to
increase in abundance. Also, the increase in available nutrients by
the die-off of more sensitive species, e.g. the observed increase in ni-
trate concentration (Cuppen et al., 1997), promoted the abundance
of tolerant species.
An outdoor mesocosm study investigating the effects of three
herbicides (atrazine, isoproturon, and diuron), both separately and
in a mixture, on phytoplankton taxa richness and Shannon–Wiener
diversity reports that biodiversity was not significantly lowered by
the pesticides (Knauert et al., 2009). Also, De Laender et al. (2012),
using palaeolimnological data and statistical modelling, found no
negative relations between sedimentary metal concentrations and
diatom richness and evenness. These findings do not correspond
with what is reported in the current paper. Many possible explanations
can be listed why this difference exists. First, the chemical concentra-
tions may have been too low in certain studies to elicit negative changes
in biodiversity. For example, in the mesocosms of Knauert et al. (2009)
exposure concentrations corresponded to the 30% hazardous concen-
trations of atrazine, isoproturon, and diuron, obtained from an SSD. In
the current paper the three highest linuron treatments used (15, 50
and 150 μg/L) correspond to, based on the SSD in Fig. 5, the 70%, 90%,
and 98% hazardous concentrations for linuron, respectively. Second,
the sensitivity of the dominant species in the control treatments will
Fig. 2. Modelled Simpson diversity (black line; 95% pointwise confidence interval as a grey zone) over time at various linuron concentrations for an average initial Simpson diversity
(0.56, 0.75, 0.77, 0.81, 0.82 and 0.85 for treatments 0 (A), 0.5 (B), 5 (C), 15 (D), 50 (E) and 150 (F) μg/L linuron respectively). Data are shown as points.
76 K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

determine if dominance patterns change under pesticide stress. If domi-
nant species in the control are sensitive, diversity can be expected to
change more when facing pesticide stress than if the dominant spe-
cies are tolerant. Indeed, the dominant species in the Knauert et al.
(2009) control microcosms (Chroomonas acuta,Cryptomonas erosa
et ovata and Katablepharis ovalis) were more sensitive than the rare
species, causing biodiversity to increase when exposed to the herbicides.
In contrast, in the current study, the most abundant species in the control
(Cocconeis sp. and Phormidium sp.) were among the most sensitive spe-
cies, causing large changes in community composition when exposed to
sufficiently high concentrations of linuron. Lastly, the isolation of most
experimental systems, hampering immigration and thus re-colonisation
of sensitive species, will influence if and how the diversity of aquatic com-
munities responds to chemical stress (Caquet et al., 2007). Mismatches
between results from field studies (De Laender et al., 2012)andexperi-
mental studies with isolated communities (the current study) may thus
be attributed to inherent differences in the immigration probabilities of
the species making up the focal community, given the physical con-
straints to dispersal.
The effects of linuron at the higher treatments became noticeable
2 weeks after linuron addition. A delay in the response of abundance
was observed in Van den Brink et al. (1997) for e.g. Cocconeis sp. The
proposed hypothesis for this phenomenon, i.e. a quiescent phase, has
been confirmed for different algal groups (von Dassow and Montresor,
2011) and is a possible explanation for the observed time lag in biodiver-
sity decrease. Other studies on the effects of linuron in tropical (Daam
et al., 2009) and in temperate microcosms (Slijkerman et al., 2005)also
showed a 2 weeks delay in the response of the phytoplankton communi-
ty. The effect of linuron on biodiversity is clearly not constant through
time and the here used additive models are ideally suited for the analysis
of such data. The here constructed GAMMs could identify the non-linear
time course of BD and the time-dependent effect of linuron on BD,
allowing to assess the effects of linuron at different points in time.
4.3. Recovery of biodiversity after linuron addition
The longer recovery time for richness than for dominance and
evenness indicates the lower resilience of richness compared to the
other aspects of BD in this study. The longer recovery time of richness
could be the result of the experimental design where immigration is
negligible and locally extinct species can therefore not recover unless
these species can survive unsuitable environmental conditions as
cysts, spores or by strongly lowering the metabolic activity of their
vegetative cells (von Dassow and Montresor, 2011). Whether or not
these species would have recovered if the experiment had continued
thus depends on their ability to form such life stages. For example, the
lower reported biovolume per cell of the initially abundant Cocconeis
after linuron application (Van den Brink et al., 1997) could be indica-
tive of vegetative cells with strongly decreased metabolic activity. In
addition, the difficulty of sampling rare species may contribute to an
underestimation of recovery rate of species richness. This is indicated
by the momentarily disappearance of taxa from the samples, e.g. for
Fig. 3. Modelled species richness (black line; 95% pointwise confidence interval as a grey zone) over time at various linuron concentrations for an average initial species richness
(8.5, 10.5, 11.0, 12.0, 13.5 and 13.5 for treatments 0 (A), 0.5 (B), 5 (C), 15 (D), 50 (E) and 150 (F) μg/L linuron respectively). Data are shown as points.
77K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

Phormidium foveolarum (Van den Brink et al., 1997). Lastly, one could
argue that –from a theoretical point of view –effects on other BD indi-
ces than richness can be compensated for by changes in the abundances
of the species already present, i.e. immigration is not needed to let, for
example, evenness recover in a community made less rich by chemical
toxicity. The differential response of these two main types of BD indices
emphasizes the importance of dispersal for the recovery of biodiversity
after stress (Trekels et al., 2011).
4.4. Link between biodiversity and ecosystem properties
The overall functioning of the plankton community in this contribu-
tion, which is representative for natural communities in macrophyte-
dominated ditches, has been shown to be unaffected by linuron because
it consists of functionally redundant species (De Laender et al., 2011b). In
this study, we found phytoplankton biomass to be unaffected by linuron.
Given that linuron is a photosynthesis inhibitor, one would expect a
decrease in phytoplankton biomass after linuron application be-
cause phytoplankton primary production –the process contributing
to phytoplankton biomass production –has been found to decrease
after the application of other herbicides e.g. Wellman et al. (1998).
In this study, Chlamydomonas sp. was found to be very tolerant to
linuron. This tolerance, combined with reduced competition by macro-
phytes and increased nutrient availability (Cuppen et al., 1997), allowed
Chlamydomonas sp. to become highly dominant in the highest linuron
treatments and as such compensate for the decreased biomass produc-
tion of more sensitive phytoplankton species. Chlamydomonas reinhardtii
has been found to be tolerant to different herbicides with a similar mode
of action as linuron –photosystem II inhibition –and this has been at-
tributed to a mutation in one of the chloroplast genes (Erickson et al.,
1984). Possibly, this tolerance combined with a competitive advantage
provided by mixotrophic capabilities as reported for C. humicola
(Lalibertè and de la Noüie, 1993).
The functional redundancy of phytoplankton communities found
here has also been reported for other herbicides, e.g. for fomesafen
(Caquet et al., 2005). Both initial community structure, indirect ef-
fects (e.g. the reduced competition by macrophyte growth inhibition
in the current study) and nutrient dynamics (Pannard et al., 2009)
seem to play a role in determining the degree of functional redundan-
cy in phytoplankton communities exposed to herbicides. For exam-
ple, the reduced growth of macrophytes (Van den Brink et al., 1997)
in the macrophyte-dominated microcosms increased the availability
of and lowered the competition for nutrients and thus offered oppor-
tunities for tolerant phytoplankton species (i.e. Chlamydomonas sp.)
to sharply increase in abundance.
Our findings indicate that the relationship between biodiversity
and ecosystem functioning in stressed ecosystems may deviate from
the often described positive relation (Hooper et al., 2005) because
Fig. 4. Proportion of biodiversity indices negatively affected for different combinations of
time and linuron concentration. Biodiversity indices were calculated with a mean initial
biodiversity index for all treatments. The size of a circle is proportional to the number of
biodiversity indices affected. Proportions are shown for the log
10
(C+ 1)-transformed
linuron concentrations 0, 0.5, 5, 15, 50 and 150.
Fig. 5. LECs for biodiversity indices at week 5 (↓) and week 11 (↑). Exact values for the LECs are given in Table S4. A SSD (n =11) based on chronic NOECs is given (black triangles,
[1]).
78 K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

reductions in BD do not induce similar reductions in biomass, the end
product of the ecosystem function ‘primary production’. We argue
that this deviation is caused by the experimental approaches that
have been followed in other studies to relate ecosystem functioning
to BD. These approaches typically do not explicitly include stressors
in the experimental design. Instead of having increasing stressor levels
creating a gradient of BD levels –as is the case in nature –experimenters
themselves created a BD gradient by randomly composing multi-species
assemblages with increasing BD levels (Balvanera et al., 2006). This ap-
proach implicitly assumes that the likelihood of a species to be removed
by a stressor is the same for all species, which is not necessarily the case
as species sensitivities determine if a species will be lost following expo-
sure or not. Steudel et al. (2012) evaluated how a stress gradient influ-
ences the biodiversity–ecosystem functioning relationship. The authors
concluded that the positive effect of biodiversity declines with increasing
stress and that “more diverse biotic communities are functionally less
susceptible to environmental stress”.
The phytoplankton community plays an essential role in the aquatic
ecosystem because it converts solar energy into biomass, which can be
consumed byhigher trophic levels. Previous studies have indicated that
ecosystemstructure is most likely more sensitive to chemicals than eco-
system functioning (De Laender et al., 2008b). The present study has
shown that the phytoplankton biodiversity, as a proxy for ecosystem
structure, is unaffected at linuron concentrations below 2.4 μg/L. At
this linuron concentration, the total phytoplankton biomass production
was also found to be unaffected. However, recent literature has shown
that taking into account only one ecosystem function is insufficient
(Isbell et al., 2011). The effects of linuron on multiple ecosystem
functions have been analysed for the here studied dataset using lin-
ear inverse modelling and the results largely confirm the functional
redundancy found here for the total phytoplankton biomass produc-
tion (De Laender et al., 2011b).
4.5. Applicability of traditionally derived environmental protective
concentrations
The lowest LEC calculated here for BD indices (2.4 μg/L) was more
than a factor five higher than the lowest microcosm derived NOECs
(0.45 μg/L; Van den Brink et al., 2006) and a LEC for phytoplankton
biomass could not be calculated. Thus, BD indices do not pick up the
effects of linuron on separate species but provide an overall picture
of the effects of linuron on the phytoplankton community. This is in
agreement with previous findings related to microcosm studies and
similarity indices (Van den Brink and Ter Braak, 1998). Using an alter-
native approach, a simulation study with a hypothetical toxicant and
different initial conditions showed that at the HC
5
96.6% of the BD in-
dices showed a change smaller than 5% (de Vries et al., 2010). In con-
clusion, although species are affected at linuron concentrations as low
as 0.45 μg/L (Van den Brink et al., 2006), BD indices and phytoplank-
ton biomass are more robust and thus, traditionally derived environ-
mental protective concentrations are protective for biodiversity and
ecosystem functioning in the case of linuron.
Fig. 6. Modelled phytoplankton biomass (black line; 95% confidence interval as a grey zone) over time at various linuron concentrations: 0 (A), 0.5 (B), 5 (C), 15 (D), 50 (E) and 150
(F)μg/L linuron. Data are shown as points.
79K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80

Acknowledgements
F.D.L. is a postdoctoral research fellow from the Fund for Scientific
Research (FWO, Flanders).
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://
dx.doi.org/10.1016/j.scitotenv.2013.01.046.
References
Balvanera P, Pfisterer AB, Buchmann N, He JS, Nakashizuka T, Raffaelli D, et al. Quanti-
fying the evidence for biodiversity effects on ecosystem functioning and services.
Ecol Lett 2006;9:1146–56.
Caquet T, Deydier-Stephan L, Lacroix G, Le Rouzic B, Lescher-Moutoue F. Effects of
fomesafen, alone and in combination with an adjuvant, on plankton communities
in freshwater outdoor pond mesocosms. Environ Toxicol Chem 2005;24:1116–24.
Caquet T, Hanson ML, Roucaute M, Graham DW, Lagadic L. Influence of isolation on the
recovery of pond mesocosms from the application of an insecticide. II. Benthic
macroinvertebrate responses. Environ Toxicol Chem 2007;26:1280–90.
Cuppen JGM, Van den Brink PJ, Van der Woude H, Zwaardemaker N, Brock TCM. Sensi-
tivity of macrophyte-dominated freshwater microcosms to chronic levels of the
herbicide linuron. Ecotoxicol Environ Saf 1997;38:25–35.
Daam MA, Rodrigues AMF, Van den Brink PJ, Nogueira AJA. Ecological effects of the her-
bicide linuron in tropical freshwater microcosms. Ecotoxicol Environ Saf 2009;72:
410–23.
De Laender F, De Schamphelaere KAC, Vanrolleghem PA, Janssen CR. Do we have to in-
corporate ecological interactions in the sensitivity assessment of ecosystems? An
examination of a theoretical assumption underlying species sensitivity distribution
models. Environ Int 2008a;34:390–6.
De Laender F, De Schamphelaere KAC, Vanrolleghem PA, Janssen CR. Is ecosystem struc-
ture the target of concern in ecological effect assessments? Water Res 2008b;42:
2395–402.
De Laender F, Soetaert K, De Schamphelaere KAC, Middelburg JJ, Janssen CR. Ecological
significance of hazardous concentrations in a planktonic food web. Ecotoxicol En-
viron Saf 2010;73:247–53.
De Laender F, Van den Brink PJ, Janssen CR. Functional redundancy and food web func-
tioning in linuron-exposed ecosystems. Environ Pollut 2011a;159:3009–17.
De Laender F, Van den Brink PJ, Janssen CR. Functional redundancy and food web func-
tioning in linuron-exposed ecosystems. Environ Pollut 2011b;159(10):3009–17.
De Laender F, Verschuren D, Bindler R, Thas O, Janssen CR. Biodiversity of freshwater
diatom communities during 1000 years of metal mining, land use, and climate
change in central Sweden. Environ Sci Technol 2012;46:9097–105.
de Vries P, Smit MGD, van Dalfsen JA, De Laender F, Karmant CC. Consequences of
stressor-induced changes in species assemblage for biodiversity indicators. Envi-
ron Toxicol Chem 2010;29:1868–76.
Dueri S, Dahllof I, Hjorth M, Marinov D, Zaldivar JM. Modeling the combined effect of
nutrients and pyrene on the plankton population: validation using mesocosm ex-
periment data and scenario analysis. Ecol Model 2009;220:2060–7.
Erickson JM, Rahire M, Bennoun P, Delepelaire P, Diner B, Rochaix J-D. Herbicide resis-
tance in Chlamydomonas rei nhardtii results from a mutation in the chloroplast gene
for the 32-kilodalton protein of photosystem II. Proc Natl Acad Sci 1984;81:3617–21.
EU. Directive 2000/60/EC of the European parliament and of the council of 23 October
2000 establishing a framework for community action in the field of water policy.
Off J Eur Communities 2000;L327:1-72.
Fleeger JW, Carman KR, Nisbet RM. Indirect effects of contaminants in aquatic ecosys-
tems. Sci Total Environ 2003;317:207–33.
Forbes VE, Calow P. Species sensitivity distributions revisited: a critical appraisal. Hum
Ecol Risk Assess 2002;8:473–92.
Forbes VE, Calow P, Sibly RM. Are current species extrapolation models a good basis for
ecological risk assessment? Environ Toxicol Chem 2001;20:442–7.
Forbes VE, Calow P, Grimm V, Hayashi T, Jager T, Palmqvist A, et al. Integrating population
modeling into ecological risk assessment. Integr Environ Assess Manag 2009;6:191–3.
Hooper DU, Chapin FS, Ewel JJ, Hector A, Inchausti P, Lavorel S, et al. Effects of biodiver-
sity on ecosystem functioning: a consensus of current knowledge. Ecol Monogr
2005;75:3-35.
Hooper DU, Adair EC, Cardinale BJ, Byrnes JEK, Hungate BA, Matulich KL, et al. A global
synthesis reveals biodiversity loss as a major driver of ecosystem change. Nature
2012;486:105–8.
Isbell F, Calcagno V, Hector A, Connolly J, Harpole WS, Reich PB, et al. High plant diver-
sity is needed to maintain ecosystem services. Nature 2011;477:199–202.
Knauert S, DawoU, Hollender J, Hommen U, KnauerK. Effects of photosystemII inhibitors
and their mixture on freshwater phytoplankton succession in outdoor mesocosms.
Environ Toxicol Chem 2009;28:836–45.
Lalibertè G, de la Noüie J. Auto-, hetero-, and mixotrophic growth of Chlamydomonas
humicola (cmloroimiyckak) on acetate1. J Phycol 1993;29:612–20.
McGrady-Steed J, Morin PJ. Biodiversity, density compensation, and the dynamics of
populations and functional groups. Ecology 2000;81:361–73.
Pannard A, Le Rouzic B, Binet F. Response of phytoplankton community to low-dose at-
razine exposure combined with phosphorus fluctuations. Arch Environ Contam
Toxicol 2009;57:50–9.
Posthuma L, Suter II GW, Traas TP. Species sensitivity distributions in ecotoxicology.
Boca Raton: CRC Press; 2002.
Preuss TG, Hammers-Wirtz M, Ratte HT. The potential of individual based population
models to extrapolate effects measured at standardized test conditions to relevant
environmental conditions—an example for 3,4-dichloroaniline on Daphnia magna.
J Environ Monit 2010;12:2070–9.
Relyea R, Hoverman J. Assessing the ecology in ecotoxicology: a review and synthesis
in freshwater systems. Ecol Lett 2006;9:1157–71.
Ricciardi F, Bonnineau C, Faggiano L, Geiszinger A, Guasch H, Lopez-Doval J, et al. Is
chemical contamination linked to the diversity of biological communities in rivers?
TrAC Trends Anal Chem 2009;28:592–602.
Schenker N, Gentleman JF. On judging the significance of differences by examining the
overlap between confidence intervals. Am Stat 2001;55:182–6.
Slijkerman DME, Moreira-Santos M, Jak RG, Ribeiro R, Soares AMVM, Van Straalen NM.
Functional and structural impact of linuron on a freshwater community of primary
producers: the use of immobilized algae. Environ Toxicol Chem 2005;24:2477–85.
Steudel B, Hector A, Friedl T, Löfke C, Lorenz M, Wesche M, et al. Biodiversity effects on
ecosystem functioning change along environmental stress gradients. Ecol Lett
2012;15:1397–405.
Trekels H, Van de Meutter F, Stoks R. Habitat isolation shapes the recovery of aquatic
insect communities from a pesticide pulse. J Appl Ecol 2011;48:1480–9.
United Nations. Convention on biological diversity; 1992.
Van den Brink PJ. Letter to the editor: response to recent criticism on aquatic semifield
experiments: opportunities for new developments in ecological risk assessment of
pesticides. Integr Environ Assess Manag 2006;2:202–3.
Van den Brink PJ. Ecological risk assessment: from book-keeping to chemical stress
ecology. Environ Sci Technol 2008;42:8999–9004.
Van den Brink P, Ter Braak C. Multivariate analysis of stress in experimental ecosys-
tems by principal response curves and similarity analysis. Aquat Ecol 1998;32:
163–78.
Van den Brink PJ, Hartgers EM, Fettweis U, Crum SJH, VanDonk E, Brock TCM. Sensitiv-
ity of macrophyte-dominated freshwater microcosms to chronic levels of the her-
bicide linuron.1. Primary producers. Ecotoxicol Environ Saf 1997;38:13–24.
Van den Brink PJ, Blake N, Brock TCM, Maltby L. Predictive value of species sensitivity
distributions for effects of herbicides in freshwater ecosystems. Hum Ecol Risk As-
sess 2006;12:645–74.
von Dassow P, Montresor M. Unveiling the mysteries of phytoplankton life cycles: pat-
terns and opportunities behind complexity. J Plankton Res 2011;33:3-12.
Wellman P, Ratte H, Heimbach F. Primary and secondary effects of methabenzthiazuron on
plankton communities in aquatic outdoor microcosms. Aquat Ecol 1998;32:125–34.
Wood SN. Generalized additive models: an introduction with R. Boca Raton, Florida:
Chapman & Hall/CRC; 2006.
Wood SN. Fast stable restricted maximum likelihood and marginal likelihood estima-
tion of semiparametric generalized linear models. J R Stat Soc 2011;73:3-36.
Yachi S, Loreau M. Biodiversity and ecosystem productivity in a fluctuating environ-
ment: the insurance hypothesis. Proc Natl Acad Sci U S A 1999;96:1463–8.
Zuur AF, Ieno EN, Smith GM, Reed J, Elphick C, Zuur A, et al. Time series analysis of Hawaiian
waterbirds. Analysing ecological data. New York: Springer; 2007. p. 615–31.
Zuur AF, Elena NI, Walker NJ, Saveliev AA, Smith GM. Mixed effects models and exten-
sions in ecology with R. New York: Springer Science +Business Media; 2009.
80 K.P.J. Viaene et al. / Science of the Total Environment 449 (2013) 71–80