Ecological Applications, 22(4), 2012, pp. 1213–1223
Ó 2012 by the Ecological Society of America
Interaction networks in coastal soft-sediments highlight the potential
for change in ecological resilience
S. F. THRUSH,1 J. E. HEWITT,
AND
A. M. LOHRER
National Institute of Water and Atmospheric Research, P.O. Box 11-115, Hamilton 3251 New Zealand
Abstract. Recent studies emphasize the role of indirect relationships and feedback loops
in maintaining ecosystem resilience. Environmental changes that impact on the organisms
involved in these processes have the potential to initiate threshold responses and
fundamentally shift the interactions within an ecosystem. However, empirical studies are
hindered by the difficulty of designing appropriate manipulative experiments to capture this
complexity. Here we employ structural equation modeling to define and test the architecture of
ecosystem interaction networks. Using survey data from 19 estuaries we investigate the
interactions between biological (abundance of large bioturbating macrofauna, microphytobenthos, and detrital matter) and physical (sediment grain size) processes. We assess
the potential for abrupt changes in the architecture of the network and the strength of
interactions to occur across environmental gradients. Our analysis identified a potential
threshold in the relationship between sediment mud content and benthic chlorophyll a, at ;12
lg/g, using quantile regression. Below this threshold, the interaction network involved
different variables and fewer feedbacks than above. This approach has potential to improve
our empirical understanding of thresholds in ecological systems and our ability to design
manipulative experiments that test how and when a threshold will be passed. It can also be
used to indicate to resource managers that a particular system has the potential to exhibit
threshold responses to environmental change, emphasizing precautionary management and
facilitating a better understanding of how persistent multiple stressors threaten the resilience
and long-term use of natural ecosystems.
Key words: facilitation; feedbacks; macrofauna; microphytobenthos; mud; multiple stressors; New
Zealand.
INTRODUCTION
Ecological processes that involve feedbacks or indirect
relationships between biota and their environment have
important consequences for ecosystem function and
dynamics (Scheffer et al. 2001, Rietkerk and van de
Koppel 2008, Thrush et al. 2009). This can involve the
formation of distinct spatial patterning within the
landscape or temporal trajectories that produce abrupt
change into functionally different states. These selforganized patterns often result from interactions that
operate at the ecosystem level, involving biological,
physical, and chemical processes. Except under extreme
environmental forcing, such self-organized interaction
networks can play important roles in maintaining
resilience. The problem is that we do not know if the
architecture of the web of interactions that defines a
specific system will exhibit intrinsic dynamics or simply
track environmental forcing. Given the potential for
changes in the strength or direction of at least some
types of feedbacks to generate threshold responses, it is
Manuscript received 8 August 2011; revised 5 December
2011; accepted 16 January 2012. Corresponding Editor: S. B.
Baines.
1 E-mail: [email protected]
imperative that we develop empirical techniques to
assess the potential for these types of interactions to
exist in different ecosystems and to evaluate the
potential for changes in the architecture of interaction
webs across environmental gradients.
There is a long history of documented change in the
ecological state of marine ecosystems because of major
disturbance events (Dayton et al. 1984, Dayton and
Tegner 1984). However, chronic and cumulative impacts
on the organisms involved in feedback processes may be
especially important because of the potential to initiate
threshold responses and fundamentally shift ecosystem
function without extreme environmental forcing. Multiple stressors may have differential effects on these
interaction networks, reducing our ability to quantify
the risk of change to ecosystem function or resilience.
These are major challenges to our understanding of
ecosystem dynamics. From an applied research and
management perspective, threshold responses are even
more worrying because they imply the potential for
cumulative effects where the same amount of stress or
disturbance can result in very different ecosystem
responses depending on whether a threshold is tripped
or not. Moreover, many environmental management,
restoration, and conservation policies and strategies
focus on maintaining or enhancing resilience (McLeod
1213
1214
S. F. THRUSH ET AL.
and Leslie 2009, Stagg and Mendelssohn 2011). Thus,
we need to develop empirical methods to assess the
potential for such interaction networks to exist in
different ecosystems and, if they do, to identify how
the networks are modified by environmental change.
However, defining feedback responses in nature usually
requires intensive knowledge of process interactions
across scales of space and time. Similarly, determining
chronic multiple stressor impacts is usually beyond the
scope of simple impact assessments or experimental
manipulations.
As the scope of empirical research is expanded,
positive (facilitatory) processes are increasingly recognized as important in ecology (Bruno et al. 2003,
Norkko et al. 2006, Stachowicz et al. 2008, Bulleri
2009). Positive feedback mechanisms can provide
resilience by buffering the capacity of a system to
respond to change, and this can make it hard to detect
system responses until the feedbacks are broken. As
stated by Carpenter et al. (2001) the only sure way to
know you have a threshold is to cross it. Theoretically,
increased variance prior to a transition is predicted, and
there are signs of these phenomena in long-term
monitoring data (Hewitt and Thrush 2010). Unfortunately, such data are often not available. Other
approaches search for changes in the spatial distribution
of patches (Rietkerk and van de Koppel 2008) or a
slowing down in recovery rates of a system as it nears a
critical threshold (van Nes and Scheffer 2007). The
description of pattern in nature is highly scale dependent, and recovery rates are often context dependent
with degraded, low-diversity systems dominated by
opportunistic species likely to exhibit fast recovery.
Thus, there is not yet a clear path forward, particularly
for empirical research, to inform policy and management. One method of assessing the potential for regime
shifts to occur may be to identify possible feedback
processes from field data and then identify if the strength
of these interactions varies across environmental gradients. Recently van der Heide et al. (2011) used structural
equation modeling (SEM) to show the potential for
positive feedback processes to exist in the relationships
linking seagrass density, light attenuation, and sediment
grain size across a number of northern European
estuaries. SEM is ideally suited to exploring whether
theoretically driven conceptual models fit empirical
data, defining the architecture and strength of interactions within complex interaction networks and testing
the merits of competing models (Grace et al. 2010, Kline
2011). Here we use SEM to investigate the architecture
of interaction networks derived from survey data
collected from intertidal flats. We then consider whether
the architecture and strength of interactions within
networks changes at potential transition points along
environmental gradients using factor–ceiling relationships. SEM represents a family of related statistical
procedures, including path analysis. Without appropriate manipulative experiments, statistical techniques such
Ecological Applications
Vol. 22, No. 4
as SEM cannot provide strong evidence of causality, but
are useful in defining the degree of empirical support for
specified conceptual models. In this paper, we use the
procedure to demonstrate the presence of, and changes
in, interaction networks, both as an aid to future
experimental design and as an indicator to managers
of the presence of nonlinear dynamics and feedback
loops.
In coastal and estuarine soft-sediment habitats,
benthic organisms play crucial roles in ecosystem
functioning. Feedback processes often involve interactions between organisms, biogeochemical processes, or
hydrodynamics (Lohrer et al. 2004, van de Koppel et al.
2005, Coco et al. 2006, Weerman et al. 2010b). Many
studies have identified simple cause–effect relationships
between macrofauna, sediment characteristics, and
microphytobenthos (Miller et al. 1996). These individual
cause–effect relationships can be linked into more
complex interaction networks (Fig. 1): a positive
relationship between microphytobenthos and sediment
muddiness, illustrated by arrow ‘‘i’’ (van de Koppel et al.
2001); a positive relationship between large depositfeeding macrofauna and microphytobenthos, arrow ‘‘ii’’
(Thrush et al. 2006b, Van Colen et al. 2008); fine
sediment binding by microphytobenthos mediated by
resident macrofaunal, arrow ‘‘iii’’ (Widdows et al. 2004);
a negative effect of fine sediment on abundance and
diversity of large bioturbators, arrow ‘‘iv’’ (Thrush et al.
2003, 2004); and a positive effect of shell hash (a major
contributor to coarse sediments) on macrofaunal size
and diversity, arrow ‘‘v’’ (Thrush et al. 2006a). Sediment
grain size influences microphytobenthic productivity,
arrows ‘‘vi’’ (Cahoon et al. 1999); grain size also affects
the quantity and processing of detritus, arrows ‘‘vii’’
(Lopez et al. 1989); and macrofauna both contribute to
and exploit detritus, arrows ‘‘viii’’ (Rhoads and Boyer
1982). The importance of this type of interaction
network on estuarine flats was illustrated by Weerman
et al. (2010b), who demonstrated that microphytobenthos affected sediment erosion thresholds to generate
hummock–hollow spatial patterns on Dutch mudflats,
i.e., local self-organizing processes generated spatial
patterns at the landscape level. These hummock–hollows
break down seasonally as the density of benthic grazers
increases and crop down the microphytobenthos (Weerman et al. 2010a). All of these studies provide
compelling insight into the ecological functioning and
geomorphology of intertidal flats, but it is difficult to
know how generalizable particular mechanisms are
across locations. We used these interactions to define
the architecture of our starting point for SEM (Fig. 1).
Specifically in our application all variables were
empirically measured (i.e., were manifest variables),
although it is also possible to use latent variables
representing hypothetical constructs, such as integrity or
health, in SEM (Kline 2011).
Estuaries and coasts are important transitional
environments connecting the land to the sea. They have
June 2012
INTERACTION WEBS AND RESILIENCE
1215
FIG. 1. Interaction network based on previously identified relationships between macrofauna, sediment chlorophyll a, detrital
organic matter content, and the proportion of coarse and muddy sediments in estuaries. Roman numerals relate to specific
relationships derived from the literature and specified in the Introduction.
high ecological, biodiversity–function values as well as
high and diverse cultural and economic values (Levin et
al. 2001). The multiple uses of these ecosystems have
subjected them to multiple stressors, with major
stressors derived from both land (e.g., sediment and
nutrient loading) and sea (e.g., ecosystem effects of
fishing and climate change). By definition, transitional
ecosystems have strong environmental gradients, thus
adding environmental variation as another factor
modifying stress responses (Thrush et al. 2008b).
However, this environmental variability also provides
an opportunity to sample relationships across a range of
conditions and identify where they change, indicating
the generality of ecological relationships and their
susceptibility to stress.
Changes in the relationship between biological or
environmental variables involved in interaction networks are often evidenced in bivariate scatter plots by
the presence of nonlinearities or step functions. However, often multiple factors are important in affecting the
magnitude of a response variable at any point along a
putative gradient. For example, shifts in sediment grain
size from mud to sand-sized particles frequently affect
the distribution and abundance of soft-sediment species,
although many other environmental and biotic factors
are also important (Snelgrove and Butman 1994, Thrush
et al. 2003). In these circumstances, it may be
informative to consider the envelope encompassing the
points in a scatter plot, using the concept of a factor
ceiling (Blackburn et al. 1992, Thomson et al. 1996). For
example, the upper bound of a scatterplot identifies
where the independent variable limits the possible
magnitude of the response variable, thus defining a
ceiling (e.g., Fig. 2a cf. b). Changes in the slope or shape
of these ceilings can thus reflect changes in the influence
of a particular factor, assuming a reasonably balanced
sampling. Quantile regression techniques are available
to model these boundaries (Scharf et al. 1998, Cade et al.
1999, Cade et al. 2005) and are routinely used for species
distribution modeling (Anderson 2008, Hothorn et al.
2011, Peters et al. 2011, Keeley et al. 2012). Unimodal
relationships and step changes in the factor ceilings (e.g.,
Fig. 2c, d) are used to indicate optimal conditions (e.g.,
sensitivity to mud; Anderson 2008) and threshold
responses (e.g., threshold below which iron concentrations are not important; Peters et al. 2011). They thus
have the potential to indicate critical transitions where
the architecture of networks and magnitude of interaction strength could fundamentally shift.
In this paper, we assess the efficacy of the interaction
network drawn from the literature (Fig. 1). Furthermore, by splitting the data at points indicated by abrupt
changes in factor ceiling relationships, we determine
whether such break points indicate important transitions
in the architecture of interaction networks. The purpose
of this two-phase approach is to indicate not only that
feedback mechanisms can be at work, implying the
potential for self-organization, but that these relationships may break down over gradients that reflect
environmental change (e.g., changes in sediment characteristics or primary productivity). The information
generated by such an approach can be used to indicate
the risk of abrupt change in a system or to design
manipulative experiments demonstrating causality and
confirming break points. This approach is likely to be
useful for resource managers faced with natural
ecosystems where nonlinear dynamics may lead to
surprises and thresholds. The implication is that simply
managing up to guideline concentrations, or other
limits, is risky (Thrush and Dayton 2010).
1216
S. F. THRUSH ET AL.
Ecological Applications
Vol. 22, No. 4
FIG. 2. Types of factor–ceiling relationships that can be derived using quantile regressions: (a) no relationship, (b) a simple
linear decrease, (c) a threshold with no relationship below a certain value (e.g., Peters et al. 2011), and (d) a unimodal relationship
indicating a threshold of change (e.g., Anderson 2008). The graph is theoretical to illustrate the value of factor ceiling relationships
and quantile regression, but the scale numbers are of value as they indicate empirical measurement, rather than a purely theoretical
concept.
METHODS
We used data on macrofaunal communities, sediment
grain size, benthic chlorophyll a, and detrital organic
matter concentrations from a survey of 19 estuaries;
details of the sites and the sampling strategy are
presented in Thrush et al. (2003). The estuaries (Ahuriri,
Bowentown, Kawhia, Koutou, Manaia, Mangawhai,
Manukau, Matakana, Mangonui, Ngunguru, Ohiwa,
Otamatea, Porangahau, Te Puna, Waiheke, Waiotahi,
Whangape, Waikopua, and Whitianga) were scattered
around the North Island of New Zealand. All were
coastal-ocean-dominated systems with low freshwater
inputs relative to tidal exchange. In each estuary, we
focused our sampling along a transect over the mud-tosand interface on one intertidal flat. Transects typically
started in soft muddy sediment, often near the edge of
mangrove habitats, and ended on firm sand. Start
positions of each transect were haphazardly selected,
but transects ran parallel to the shore to reduce
confounding by changes in tidal elevation. Each transect
was divided into 12 strata that encompassed the mud-tosand transition. All sites were sampled October–December 2000.
In each stratum, we took one core (1.5 cm diameter 3
2 cm deep) for analyzing sediment particle size and
organic matter and another for benthic chlorophyll a.
To assess macrofauna, two cores (13 cm diameter and 15
cm deep) separated by ;1.0 m were collected. To
determine particle size, sediments were digested in 6%
hydrogen peroxide for 48 h to remove organic matter.
Wet sieving was used to measure cumulative percentage
masses of coarse shell and sand, medium sand, fine sand,
and mud sediment fractions (i.e., particle sizes .0.5,
0.5–0.25, 0.25–0.063, ,0.063 mm, respectively). Organic
content was measured as loss on ignition in 5.5 h at
4008C, after drying the samples at 608C for 48 h.
Sediment samples for chlorophyll a analyses were kept
chilled and in the dark while in the field and frozen as
soon as possible and then freeze-dried on return to the
laboratory. Chlorophyll a was extracted from sediments
by boiling in 95% ethanol, and the extract analyzed
using a spectrophotometer (UV1800; Shimadzu, Kyoto,
Japan). An acidification step was used to separate
degradation products from chlorophyll a (Sartory 1982).
Summary statistics for these environmental parameters
are presented in Thrush et al. (2003). Benthic chlorophyll a varied from 0.7 to 49.4 lg/g; mud from 1.5% to
June 2012
INTERACTION WEBS AND RESILIENCE
1217
TABLE 1. Species defined as large bioturbating fauna (LBF) from 19 estuaries, North Island, New Zealand.
LBF group
Mollusks
Polychaetes
Crustaceans
Other
Species
Austrovenus stutchburyi, Macomona liliana, Paphies australis, Diloma subrostrata, Musculista senhousia, Amalda
sp., Amphibola crenata
Pectinaria australis, Scolecolepides benhami, Nereididae, Scoloplos cylindifer, Glycera sp., Macroclymenella
stewartensis, Anaitides sp., Asychis theodori, Orbinia papillosa, Scoloplos ohlini, Lepidodontidae, Travisia
olens, Aglaophamus macroura
Austrohelice crassa, Hemigrapsus crenulatus, Hemiplax hirtipes, Cirolandiae, Calianassa filholi, Alpheus sp.
Sipunculida
88%; coarse shell and sand from 0% to 55%; and organic
content from 0.4% to 8.1%. Organic matter was not well
correlated with chlorophyll a content (Pearson’s r ¼
0.62). Note that one organic matter sample from
Matakana and two chlorophyll a samples from Mangonui were compromised during sample processing, and
all data from these sampling positions were removed
from the following analyses.
Macrofauna retained on 1.0-mm mesh sieves were
preserved in 70% isopropyl alcohol, stained with 0.2%
Rose Bengal, sorted, and identified to the lowest
taxonomic level practicable. An important mechanistic
link between macrofauna and microphytobenthos production in non-eutrophic systems is associated with
efflux of ammonia from the sediment (e.g., Lohrer et al.
2004, Thrush et al. 2006b). This is a result of the
activities of the macrofauna within the sediment,
excreting nitrogenous waste products and influencing
water and particle flows affecting the activity of the
resident microbial communities. Thus, many biological
traits will influence this role for an individual species,
although size and mobility within the sediment are
particularly important. In a previous experiment we
excluded large macrofauna from patches of sandflats,
and this emphasized that we can aggregate the density of
individual species into a functional category of large
bioturbating (mixing of sediment by biological activity)
fauna (LBF; Thrush et al. 2006b). We selected species to
include in this group based on size and mobility traits
(Table 1) and calculated overall abundance of large
bioturbating fauna for use in the models as the sum over
the two cores taken at each sampling position (varied
from 1 to 104 individuals per position).
To identify potential thresholds that could have
implications for the breakdown of interactions between
large bioturbating fauna, benthic chlorophyll a and
organic matter, and sediment mud and coarse sand, we
investigated ‘‘factor–ceiling’’ relationships in scatter
plots (Thomson et al. 1996). Initially, bivariate scatter
plots of all combinations of two variables (Fig. 3) were
examined for indications that any of the variables were
acting as a limiting factor for the other variables, and,
more importantly, whether this relationship was monotonic or exhibited thresholds. Different types of relationships were observed (Fig. 3) from increasing
variance (e.g., organic matter with mud [a]), a decreasing
upper factor ceiling (e.g., LBF with mud [d]), and a
random pattern (e.g., LBF with chlorophyll a [f ]). Only
the mud and chlorophyll a scatter plot suggested the
potential for a threshold in the factor–ceiling relationship (Fig. 3j, k). The most likely benthic chlorophyll a
concentration to be the break point was identified using
LAD regression (least absolute deviation) of the 90th
percentile and a LOESS smoother (procedure Quantreg
in R; R Development Core Team 2011). This value was
used as a start point in an iterative procedure
minimizing the mean square error (procedure NLIN in
SAS; SAS Institute 2010) that identified a benthic
chlorophyll a concentration of 11.6 lg/g as the break
point. The break point for mud content, identified by a
similar process, was 19.9%. Two data sets were then
created around both potential break points. Above and
below chlorophyll a concentration of 11.6 lg/g (n ¼ 107
and 118, respectively) and above and below mud content
of 19.9% (n ¼ 92 and 133, respectively), resulting in four
data sets. Note that in general, samples from each
estuary were found in both subsets of the data with five
exceptions for the chlorophyll subsets (Ahuriri, Whangape, and Waikopua, all samples ,11.6 lg/g and
Bowentown and Te Puna, all samples .11.6 lg/g) and
three for the mud subsets (Mangawhai, Mangonui, and
Whitford, all samples ,19.9% mud). The four created
data sets, together with the initial full data set, were used
to test the model structure (Fig. 1).
Structural equation modeling (SEM) was conducted
(using the linesqs option within PROC CALIS; SAS
Institute 2010), using measured variables (Grace et al.
2010, van der Heide et al. 2011). Simple types of SEMs
include multiple regression and path analysis. Models
that are more complicated allow the overall statistical
assessment of a number of multiple regression models
and path analyses with many feedback loops and twoway interactions. Initially we had analyzed the data by
path analysis, but moved to SEM to gain the overall
statistical assessment that simplified comparison of the
different network architectures. The model tested was
non-recursive (i.e., it included feedback loops and
allowed two-way interactions). In summary, sedimentary characteristics (coarse shell and sand and mud
content) were predicted to affect LBF. LBF was
predicted to affect organic matter and chlorophyll a
content, and be affected by them in turn. Similarly,
chlorophyll a content and organic matter were predicted
to affect mud content and be affected by it. As the
1218
S. F. THRUSH ET AL.
Ecological Applications
Vol. 22, No. 4
FIG. 3. Pairwise scatter plots of study variables showing different types of factor ceilings. Abbreviations: OM, detrital organic
matter; mud, mud content; Cs, coarse sand content; Chl a, chlorophyll a content; LBF, abundance of large bioturbating
macrofauna. For plots ( j) and (k) the thick vertical line indicates the change point in ‘‘factor ceiling response’’ (thin black lines) at
11.6 lg/g chlorophyll a or at mud content of 19.9%, respectively.
June 2012
INTERACTION WEBS AND RESILIENCE
1219
TABLE 2. Structural equation modeling (SEM) chi-square test results for the best model obtained
for each of the five data sets.
Data set
v2
P
GFI
RMSEA
AIC
Full data set
Mud , 19.9%
Mud . 19.9%
Chl a , 11.6 lg/g
Chl a . 11.6 lg/g
13.44
14.86
1.22
4.24
2.84
0.0012
0.0003
0.5432
0.1201
0.2413
0.99
0.98
0.99
0.00
0.09
0.03
2.78
0.23
1.16
Notes: Where the test that the model fits the data is not rejected (P , 0.05), other model fit
statistics are given; df ¼ 2. Abbreviations: GFI, goodness of fit; RMSEA, root mean square error of
approximation; AIC, Akaike’s Information Criterion; Chl a, chlorophyll a; Mud, mud content.
degrees of freedom for the null hypothesis (the model
fitted the data) were only 15 (based on N ¼ 5 variables,
df ¼ N(N þ 1)/2), all relationships could not be fitted at
once, and a number of iterations were conducted to
determine the best, and most stable, model.
The process of iteration began with variables and
interactions from the conceptual model selected randomly; the model fit was recorded and a new model was
developed replacing nonsignificant interactions with
ones not present in the previous model. This process
continued until all variables and interactions had been
included and tested for significance. Lack of strong
correlation between variables made this a simple
process, with no changes in direction of relationships
dependent on model structure. Model fit was determined
based on v2 values (where P , 0.05 indicates poor model
fit), RMSEA (root mean square error of approximation;
values close to zero indicate a good model fit), GFI
(goodness of fit; values close to 1 indicate a good model
fit), and AIC (Akaike’s Information Criterion; lower
values indicate a better model fit) (see Kline 2011). In
some cases, similar to most multiple regression analyses,
this meant that relationships whose estimates were not
significant at the 0.05 level were retained as important in
the model.
The presence of outliers, nonnormality, and nonhomogeneity of variances was assessed graphically
(Kline 2011). No outliers were detected, but transformations were required for mud .19.9% and chlorophyll
a .11.6 lg/g. We used a loge-transformation for LBF
and mud and a square root transformation for
chlorophyll a, coarse sand, and organic matter.
Comparisons between the models based on the five
data sets were initially focused on whether better fits
could be obtained from any of the four data subsets
relative to the full data set. Following this, we compared
the fit of the models derived from splitting the data set
based on mud vs. chlorophyll a. Further analysis was
then based on the splitting variable that produced the
best-fit statistics for one model. Due to the marked
difference in model fits and model architecture, further
comparisons between the models were based on these
factors. It was not considered necessary to present actual
path coefficients.
RESULTS
No satisfactory models (i.e., P . 0.05) were developed
for the full data set (Table 2), although individual
correlations and multiple regressions were significantly
different from zero (at the P ¼ 0.05 level). No
satisfactory models were developed for the data subset
where mud content was ,19.9% (Table 2). Satisfactory
models were developed for the other data subsets,
although the model developed from the chlorophyll a
.11.6 lg/g data subset had the lowest AIC (Table 2).
The rest of the results presented focus on differences
between the models obtained from the two chlorophyll a
driven data subsets.
Analysis of the relationship when chlorophyll a
concentrations were .11.6 lg/g showed a complicated
network architecture with positive feedbacks between
chlorophyll a, LBF, and mud (Fig. 4). However, at low
benthic chlorophyll a concentrations (,11.6 lg/g),
which also coincided with the highest range in sediment
mud content, a simpler network was observed with no
positive feedbacks and based mainly around sediment
organic matter (Fig. 4). In order to ensure that these two
different architectures were not simply a result of
comparing best-fit models, we also tested both architectures in each data set. Using the complex network
architecture when chlorophyll a concentrations were
,11.6 lg/g, resulted in a very poor model fit (P ¼ 0.003).
Fitting the simpler network architecture when chlorophyll a concentrations were .11.6 lg/g was not possible,
as convergence could not be obtained. In fact, for this
data set, no model containing effects of other variables
on organic matter could be resolved.
Thus, based on model stability, model significance,
and a number of goodness-of-fit statistics, we are
reasonably sure that the shifts in relationships illustrated
in Fig. 4 are real and highlight how feedbacks and
network relationships may break down as we move
across environmental gradients, or conversely, as
environmental conditions at a site change.
DISCUSSION
The major focus of this study was to develop a
method for identifying potential changes in the architecture of ecosystem interaction networks associated
with shifts in environmental conditions. From our
1220
S. F. THRUSH ET AL.
Ecological Applications
Vol. 22, No. 4
FIG. 4. Interaction networks derived above and below chlorophyll a ¼ 11.6 lg/g. Arrows show direction of relationship, and
symbols indicate whether the relationship is positive or negative. Coefficients not significant at P ¼ 0.05 are represented by dotted
lines. See Fig. 3 for abbreviations. R 2 values are those for the simple multiple regression models buried within the overall structural
equation modeling, SEM (e.g., R 2 ¼ 0.26 beside LBF is related to the multiple regression of Cs, OM, Chl a, and mud on LBF).
survey of 19 estuaries, we were unable to develop a
single overall model that covered all environmental
conditions. Conversely, we did demonstrate a specific
threshold in benthic chlorophyll a concentration that
changed the architecture of the interaction web. We see
this as an important step allowing scientists to refine
hypotheses and experimental tests that can be used to
more accurately inform environmental managers of
potential risk of threshold responses due to changes in
intrinsic dynamics. We consider the latter to be
particularly important because the potential for threshold responses due to the interaction of intrinsic
dynamics and environmental drivers should lead resource managers to assess environmental risks in very
different ways. To do this they will need the support of
empirical analysis from natural ecosystems.
We developed a plausible interaction network, involving processes operating on intertidal flats based on
previous studies reported in the literature from a wide
variety of locations. Many of these studies involved
manipulative experiments. The occurrence of feedbacks
and indirect relationships in such networks suggests that
simple cause–effect experiments are unlikely to fully
capture the dynamical processes in ecosystems until they
can be synthesized into broader interaction networks.
When another process in fact mediates observed
bivariate relations, this may account for contradictory
effects in the literature. For example, microphytobenthic
gross primary productivity is considered usually higher
in sheltered muddy habitats (MacIntyre et al. 1996),
whereas Yallop et al. (1994) reported greater values in
non-cohesive sediments. Identifying interaction networks is particularly important where processes are
associated with cross-scale interactions and positive
feedbacks because they can predispose the ecosystem
to rapid shifts in function. Our approach highlights that
changes in environmental characteristics can modify the
architecture of interaction networks and produce
information that can be used to design future experiments to confirm both linkages and break points. For
example, from our analysis we would want to focus
experiments on factors likely to modify benthic chlorophyll a concentrations around the 11.6 lg/g threshold.
We would also want to further investigate mechanistic
links occurring as mud content increased from 10% to
30%.
The key result of our study was the change in the
architecture of the interaction networks around the
benthic chlorophyll threshold as revealed by SEM (Fig.
4). Benthic chlorophyll a concentrations can be expected
to vary with changes in temperature and light levels, but
the data collected to build these networks were all
collected in spring over daytime low tides. The inherent
variability in this and the other variables included in the
SEM adds to the variability below the factor ceiling
defined by the quantile regression but does not prevent
us being able to observe these important interactions.
The aim of creating Fig. 4 is to show that interaction
networks can abruptly change. From this, we infer that
the large bioturbators on the intertidal flats change their
interactions with a range of environmental factors
associated with a threshold shift in the role of benthic
chlorophyll a. We cannot infer the time scale of these
different interactions, but the results imply that as
benthic chlorophyll a concentrations drop, for example
a result of elevated turbidity affecting light levels, the
role of microphytobenthos as a food resource, as an
influence on sediment stability and as a mediator of
bioturbation induced nutrient regeneration all shift. In
the context of our survey analysis, we have not measured
nutrient flux or the sediment erosion potential, and we
June 2012
INTERACTION WEBS AND RESILIENCE
invoke these mechanisms based on local experimental
studies and broader evidence from the literature.
Designing the experiments necessary to confirm mechanisms and test the role of environmental change on
network functioning is not a simple task. The potential
importance of a large number of environmental factors
that may modify the network leads to a bewildering
combination of pairwise contrasts (Wootton 1994), and
the processes involved in these interactions often operate
over different space or time scales (Hewitt et al. 2007). The
feedback processes that can generate the dynamical
properties of the ecosystem also complicate the experimental elucidation of mechanisms because of the need to
tease apart cause and effect relationships within feedback
loops. We seek to emphasize the merits of combined and
iterative approaches using both experimental and observational studies. We draw on a long history of experimental
marine ecology to demonstrate the value of observational
studies as a route to providing context, generality, and
mechanisms to scale-up our results (Dayton 2003). Impacts
on species or environmental characteristics that change
feedbacks can have profound implications for resilience.
Thus, it is imperative that methods are developed that can
identify where and how such relationships may change.
Our analysis also helps place experimental results into
a broader environmental context. Ecological field
studies are notoriously context dependent (Lawton
1999). Nevertheless, studying how broadscale factors
influence ecology can be exceedingly informative
(Thrush et al. 2000, Norkko et al. 2006). Our analysis
demonstrates that synthesizing the results of different
studies, addressing components of the macrofauna–
sediment–microphytobenthos interaction network, can
form the basis for quantitatively defining the architecture of interaction networks and provide clues to the
generality of site-specific mechanistic studies (e.g., van
de Koppel et al. 2001, Widdows et al. 2004, Thrush et al.
2006b, Van Colen et al. 2008, Weerman et al. 2010b).
Differences in the importance of processes within the
interaction network may account for major differences
in experimental studies conducted in different locations.
For example, experimental defaunation on intertidal
flats in the Netherlands resulted in a major microphytobenthos bloom and changes in sediment topography associated with the accumulation of fine sediments
(Montserrat et al. 2008, Van Colen et al. 2008). Such
microphytobenthos blooms were not apparent in similar
defaunation experiments conducted in more oligotrophic New Zealand systems (Thrush et al. 1996, 2008a).
The ecosystem response is complicated but our analysis
suggests that it is likely that another threshold, based on
nutrient concentrations, could occur, beyond the scope
of our data. Below it, in more oligotrophic systems,
large bioturbating macrofauna play an important role,
positively affecting microphytobenthos, through nutrient regeneration (Lohrer et al. 2004, Thrush et al.
2006b). Above it, in more eutrophic systems, the positive
feedback between large bioturbating fauna and micro-
1221
phytobenthos diminishes in importance and the dominant role for macrofauna is grazing (Weerman et al.
2010a). Changes in these interactions could foreshadow
major shifts in functional role of benthic organisms
associated with increased eutrophication. Such early
warning signals are useful because maintaining the role
of benthic organisms in carbon and nutrient cycling are
important if estuarine systems are to be restored from
eutrophic conditions (Conely et al. 2007).
We observed a simplification of the interaction network
at low concentrations of sediment chlorophyll a where food
resources for the large bioturbators appear to switch from
high quality microphytobenthos to more refractory
organic detritus. These low-quality food resources can be
considered stressful for many benthic species. In this case,
our results could be considered at odds with the stressgradient hypothesis (Menge and Sutherland 1987, Bertness
and Callaway 1994), which predicts that facilitatory
interactions should dominate in more stressful conditions.
These models consider community level interactions (e.g.,
competition and predation) and how they are directly
influenced by environmental stress gradients. While the
nature of facilitatory interactions and the relative importance of physical and biotic stresses can modify these effects
(Bruno et al. 2003, Bulleri et al. 2011), this hypothesis does
not focus on how environmental change can influence selforganizing interactions. In soft-sediment habitats, many of
the ecosystem functions involve interactions between
animal behavior (e.g., bioturbation or bioirrigation) and
the local environmental characteristics (e.g., sediment
grain size, nutrient status), and these interactions can
generate highly context-dependent results affecting how
emergent processes change along gradients (Daleo and
Iribarne 2009, Needham et al. 2010).
The potential for shifts in the resilience of communities
and ecosystems represents a profound ecological challenge
in switching from hindsight to foresight (Thrush et al.
2009). Loss of resilience implies that an ecological system is
prone to rapidly shift into a different state, which is
frequently referred to as a regime shift. While some regime
shifts have been driven by extreme environmental forcing,
in many cases there is evidence that the interactions
between organisms and their physical and chemical
environment play an important role. The problem is that
with hindsight we frequently lack sufficient environmental
data to tease apart why systems have changed (Collie et al.
2004). However, such changes do happen and profoundly
affect the way we should be thinking about managing
ecosystems. Where an ecological system has the potential
to lose resilience and undergo a regime shift, we cannot
calibrate stress levels against ecological responses because
the system will not necessarily respond linearly. This
requires a shift in management perspective to focus on
maintaining adaptive capacity and ‘‘resilience thinking’’
(Cumming et al. 2005, Walker and Salt 2006). The ability to
demonstrate that the interaction networks operating
within a particular system can give rise to self-organization,
with network architecture changing abruptly across
1222
S. F. THRUSH ET AL.
environmental thresholds, adds weight to the implementation of precautionary management approaches rather
than relying on more traditional stress-response approaches. Extensive sampling across relevant environmental
gradients or anthropogenic activities increases our ability
to predict where such thresholds may occur. By learning
how the network of interactions within a particular type of
ecosystem is structured, we can begin to identify the types
of interactions that, when broken, are likely to lead to
major shifts. We suggest that this is a way forward in
understanding resilience in nature and an important part of
an iterative process of prediction and testing to inform the
design of manipulative experiments and adaptive management to test specific cause–effect relationships.
ACKNOWLEDGMENTS
We thank our colleagues who helped with the original
sample collection: Alf Norkko, Pip Nicholls, Greig Funnell,
and Jo Ellis. The manuscript was improved by comments from
Ivan Rodil, Ellen Weerman, and an anonymous reviewer. This
work was funded by FRST CO1X1005.
LITERATURE CITED
Anderson, M. 2008. Animal–sediment relationships re-visited:
characterising species’ distributions along an environmental
gradient using canonical analysis and quantile regression
splines. Journal of Experimental Marine Biology and
Ecology 366:16–27.
Bertness, M. D., and R. M. Callaway. 1994. Positive
interactions in communities. Trends in Ecology and Evolution 9:191–193.
Blackburn, T. M., J. H. Lawton, and J. N. Perry. 1992. A
method of estimating the slope of upper bounds of plots of
body size and abundance in natural animal assemblages.
Oikos 65:107–112.
Bruno, J. F., J. J. Stachowicz, and M. D. Bertness. 2003.
Inclusion of facilitation into ecological theory. Trends in
Ecology and Evolution 18:119–125.
Bulleri, F. 2009. Facilitation research in marine systems: state
of the art, emerging patterns and insights for future
developments. Journal of Ecology 97:1121–1130.
Bulleri, F., C. Cristaudo, T. Alestra, and L. Benedetti-Cecchi.
2011. Crossing gradients of consumer pressure and physical
stress on shallow rocky reefs: a test of the stress-gradient
hypothesis. Journal of Ecology 99:335–344.
Cade, B. S., B. R. Noon, and C. H. Flather. 2005. Quantile
regression reveals hidden bias and uncertainty in habitat
models. Ecology 86:786–800.
Cade, B. S., J. W. Terrell, and R. L. Schroeder. 1999.
Estimating effects of limiting factors with regression quantiles. Ecology 80:311–323.
Cahoon, L. B., J. E. Nearhoof, and C. L. Tilton. 1999.
Sediment grain size effect on benthic microalgal biomass in
shallow aquatic ecosystems. Estuaries 22:735–741.
Carpenter, S., B. Walker, J. M. Anderies, and N. Abel. 2001.
From metaphor to measurement: resilience of what to what?
Ecosystems 4:765–781.
Coco, G., S. F. Thrush, M. O. Green, and J. E. Hewitt. 2006.
Feedbacks between bivalve density, flow, suspended sediment
concentration on patch stable states. Ecology 87:2862–2870.
Collie, J. S., K. Richardson, and J. H. Steele. 2004. Regime
shifts: Can ecological theory illuminate the mechanisms?
Progress in Oceanography 60:281–302.
Conely, D. J., J. Carstensen, G. Aertebjerg, P. B. Christensen,
T. Dalsgaard, J. L. S. Hansen, and A. B. Josefson. 2007.
Long-term changes and impacts of hypoxia in Danish coastal
waters. Ecological Applications 17:S165–S184.
Ecological Applications
Vol. 22, No. 4
Cumming, G. S., G. Barnes, S. Perz, M. Schmink, K. E. Sieving,
J. Southworth, M. Binford, R. D. Holt, C. Stickler, and T. Van
Hol. 2005. An exploratory framework for the empirical
measurement of resilience. Ecosystems 8:975–987.
Daleo, P., and O. Iribarne. 2009. Beyond competition: the
stress-gradient hypothesis tested in plant–herbivore interactions. Ecology 90:2368–2374.
Dayton, P. K. 2003. The importance of the natural sciences to
conservation. American Naturalist 162:1–13.
Dayton, P. K., V. Currie, T. Gerrodette, B. D. Keller, R.
Rosenthal, and D. Ven Tresca. 1984. Patch dynamics and
stability of some Californian kelp communities. Ecological
Monographs 54:253–289.
Dayton, P. K., and M. J. Tegner. 1984. Catastrophic storms, El
Nino, and patch stability in a Southern California kelp
community. Science 224:283–285.
Grace, J. B., T. M. Anderson, H. Olff, and S. M. Scheiner.
2010. On the specification of structural equation models for
ecological systems. Ecological Monographs 80:67–87.
Hewitt, J. E., and S. F. Thrush. 2010. Empirical evidence of an
approaching alternate state produced by intrinsic community
dynamics, climatic variability and management actions.
Marine Ecology Progress Series 413:267–276.
Hewitt, J. E., S. F. Thrush, P. K. Dayton, and E. Bonsdorff.
2007. The effect of spatial and temporal heterogeneity on the
design and analysis of empirical studies of scale-dependent
systems. American Naturalist 169:398–408.
Hothorn, T., J. Muller, B. Schroder, T. Kneib, and R. Brandl.
2011. Decomposing environmental, spatial, and spatiotemporal components of species distributions. Ecological Monographs 81:329–347.
Keeley, N. B., C. K. Macleod, and B. M. Forrest. 2012.
Combining best professional judgement and quantile regression splines to improve characterisation of macrofaunal
responses to enrichment. Ecological Indicators 12:154–166.
Kline, R. B. 2011. Principles and practice of structural equation
modeling. Third edition. Guilford, New York, New York,
USA.
Lawton, J. 1999. Are there general laws in ecology? Oikos
84:177–192.
Levin, L. A., D. F. Boesch, A. Covich, C. Dahm, C. Erséus,
K. C. Ewel, R. T. Kneib, A. Moldenke, M. A. Palmer, P. A.
Snelgrove, D. Strayer, and J. M. Weslawski. 2001. The
function of marine critical transition zones and the importance of sediment biodiversity. Ecosystems 4:430–451.
Lohrer, A. M., S. F. Thrush, and M. M. Gibbs. 2004.
Bioturbators enhance ecosystem function through complex
biogeochemical interactions. Nature 431:1092–1095.
Lopez, G., G. Taghon, and J. Levinton. 1989. Ecology of
marine deposit feeders. Springer-Verlag, New York, New
York, USA.
MacIntyre, H. L., R. J. Geider, and D. C. Miller. 1996.
Microphytobenthos: the ecological role of the ‘‘Secret
Garden’’ of unvegetated, shallow-water marine habitats. I.
Distribution, abundance and primary production. Estuaries
19:186–201.
McLeod, K., and H. Leslie, editors. 2009. Ecosystem-based
management for the oceans. Island Press, Washington, D.C.,
USA.
Menge, B. A., and J. P. Sutherland. 1987. Community
regulation: variation in disturbance, competition, and predation in relation to environmental stress and recruitment.
American Naturalist 110:351–369.
Miller, D. C., R. J. Geider, and H. L. MacIntyre. 1996.
Microphytobenthos: the ecological role of the ‘‘Secret
Garden’’ of unvegetated, shallow-water marine habitats. II.
Role in sediment stability and shallow-water food webs.
Estuaries 19:202–212.
Montserrat, F., C. Van Colen, S. Degraer, T. Ysebaert, and
P. M. J. Herman. 2008. Benthic community-mediated
June 2012
INTERACTION WEBS AND RESILIENCE
sediment dynamics. Marine Ecology Progress Series 372:43–
59.
Needham, H. R., C. A. Pilditch, A. M. Lohrer, and S. F. Thrush.
2010. Habitat dependence in the functional traits of Austrohelice
crassa, a key bioturbating species. Marine Ecology Progress
Series 414:179–193.
Norkko, A., J. E. Hewitt, S. F. Thrush, and G. A. Funnell.
2006. Conditional outcomes of facilitation by a habitatmodifying subtidal bivalve. Ecology 87:226–234.
Peters, A., M. Crane, and W. Adams. 2011. Effects of iron on
benthic macroinvertebrate communities in the field. Bulletin
of Environmental Contamination and Toxicology 86:591–
595.
R Development Core Team. 2011. A language and environment
for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. http://www.R-project.org
Rhoads, D. C., and L. F. Boyer. 1982. The effects of marine
benthos on physical properties of sediments: a successional
perspective. Pages 3–52 in P. L. McCall and M. J. S. Tevesz,
editors. Animal–sediment relations. Plenum, New York, New
York, USA.
Rietkerk, M., and J. van de Koppel. 2008. Regular pattern
formation in real ecosystems. Trends in Ecology and
Evolution 23:169–175.
Sartory, D. P. 1982. Spectrophotometric analysis of chlorophyll
a in freshwater phytoplankton. Hydrological Research
Institute Technical Report TR 115. Pretoria, South Africa.
SAS Institute. 2010. Version 9.1.3. SAS Institute, Cary, North
Carolina, USA.
Scharf, F. S., F. Juanes, and M. Sutherland. 1998. Inferring
ecological relationships from the edges of scatter diagrams:
comparison of regression techniques. Ecology 79:448–460.
Scheffer, M., S. Carpenter, J. A. Foley, C. Folke, and B. Walker.
2001. Catastrophic shifts in ecosystems. Nature 413:591–596.
Snelgrove, P. V. R., and C. A. Butman. 1994. Animal–sediment
relationships revisited: cause versus effect. Oceanography
and Marine Biology Annual Review 32:111–177.
Stachowicz, J. J., R. J. Best, M. E. S. Bracken, and M. H. Graham.
2008. Complementarity in marine biodiversity manipulations:
reconciling divergent evidence from field and mesocosm
experiments. Proceedings of the National Academy of Sciences
USA 105:18842–18847.
Stagg, C. L., and I. A. Mendelssohn. 2011. Controls on
resilience and stability in a sediment-subsidized salt marsh.
Ecological Applications 21:1731–1744.
Thomson, J. D., G. Weiblen, B. A. Thomson, S. Alfaro, and P.
Legendre. 1996. Untangling multiple factors in spatial
distributions: lilies, gophers, and rocks. Ecology 77:1698–
1715.
Thrush, S. F., and P. K. Dayton. 2010. What can ecology
contribute to ecosystem-based management? Annual Review
of Marine Science 2:419–441.
Thrush, S. F., J. S. Gray, J. E. Hewitt, and K. I. Ugland. 2006a.
Predicting the effects of habitat homogenization on marine
biodiversity. Ecological Applications 16:1636–1642.
Thrush, S. F., J. Halliday, J. E. Hewitt, and A. M. Lohrer.
2008a. Cumulative degradation in estuaries: the effects of
habitat, loss fragmentation and community homogenization
on resilience. Ecological Applications 18:12–21.
Thrush, S. F., J. E. Hewitt, V. Cummings, J. I. Ellis, C. Hatton,
A. Lohrer, and A. Norkko. 2004. Muddy waters: elevating
sediment input to coastal and estuarine habitats. Frontiers in
Ecology and the Environment 2:299–306.
Thrush, S. F., J. E. Hewitt, V. J. Cummings, M. O. Green,
G. A. Funnell, and M. R. Wilkinson. 2000. The generality of
1223
field experiments: interactions between local and broad-scale
processes. Ecology 81:399–415.
Thrush, S. F., J. E. Hewitt, P. K. Dayton, G. Coco, A. M. Lohrer,
A. Norkko, J. Norkko, and M. Chiantore. 2009. Forecasting the
limits of resilience: integrating empirical research with theory.
Proceedings of the Royal Society B 276:3209–3217.
Thrush, S. F., J. E. Hewitt, M. Gibbs, C. Lundquist, and A.
Norkko. 2006b. Functional role of large organisms in
intertidal communities: community effects and ecosystem
function. Ecosystems 9:1029–1040.
Thrush, S. F., J. E. Hewitt, C. W. Hickey, and S. Kelly. 2008b.
Multiple stressor effects identified from species abundance
distributions: interactions between urban contaminants and
species habitat relationships. Journal of Experimental Marine
Biology and Ecology 366:160–168.
Thrush, S. F., J. E. Hewitt, A. Norkko, P. E. Nicholls, G. A.
Funnell, and J. I. Ellis. 2003. Habitat change in estuaries:
predicting broad-scale responses of intertidal macrofauna to
sediment mud content. Marine Ecology Progress Series
263:101–112.
Thrush, S. F., R. B. Whitlatch, R. D. Pridmore, J. E. Hewitt,
V. J. Cummings, and M. Maskery. 1996. Scale-dependent
recolonization: the role of sediment stability in a dynamic
sandflat habitat. Ecology 77:2472–2487.
Van Colen, C., F. Montserrat, M. Vincx, P. M. J. Herman, T.
Ysebaert, and S. Degraer. 2008. Macrobenthic recovery from
hypoxia in an estuarine tidal mudflat. Marine Ecology
Progress Series 372:31–42.
van de Koppel, J., P. M. J. Herman, P. Thoolen, and C. H. R.
Heip. 2001. Do alternate stable states occur in natural
ecosystems? Evidence from a tidal flat. Ecology 82:3449–
3461.
van de Koppel, J., D. van der Wal, J. P. Bakker, and P. M. J.
Herman. 2005. Self-organization and vegetation collapse in
salt marsh ecosystems. American Naturalist 165:E1–E12.
van der Heide, T., E. H. van Nes, M. M. van Katwijk, H. Olff,
and A. J. P. Smolders. 2011. Positive feedbacks in seagrass
ecosystems—evidence from large-scale empirical data. PLoS
ONE 6:e16504.
van Nes, E. H., and M. Scheffer. 2007. Slow recovery from
perturbations as a generic indicator of a nearby catastrophic
shift. American Naturalist 169:738–747.
Walker, B., and D. Salt. 2006. Resilience thinking: sustaining
ecosystems and people in a changing world. Island Press,
Washington, D.C., USA.
Weerman, E. J., P. M. J. Herman, and J. Van de Koppel. 2010a.
Top-down control inhibits spatial self-organization of a
patterned landscape. Ecology 92:487–495.
Weerman, E. J., J. van de Koppel, M. B. Eppinga, F.
Montserrat, Q.-X. Liu, and P. M. J. Herman. 2010b. Spatial
self-organisation on intertidal mudflats through biophysical
stress divergence. American Naturalist 176:E15–E32.
Widdows, J., A. Blauw, C. H. R. Heip, P. M. J. Herman, C. H.
Lucas, J. J. Middelburg, S. Schmidt, M. D. Brinsley, F.
Twisk, and H. Verbeek. 2004. Role of physical and biological
processes in sediment dynamics of a tidal flat in Westerschelde Estuary, SW Netherlands. Marine Ecology Progress Series 274:41–56.
Wootton, J. T. 1994. Predicting direct and indirect effects: an
integrated approach using experiments and path analysis.
Ecology 75:151–165.
Yallop, M. L., B. de Winder, D. M. Paterson, and L. J. Stal.
1994. Comparative structure, primary production and
biogenic stabilization of cohesive and non-cohesive marine
sediments inhabited by microphytobenthos. Estuarine,
Coastal and Shelf Science 39:565–582.