-
1448
Ecological Applications, 14(5), 2004, pp. 14481465q 2004 by the
Ecological Society of America
IDENTIFYING ECOLOGICAL CHANGE AND ITS CAUSES:A CASE STUDY ON
CORAL REEFS
KATHARINA E. FABRICIUS1 AND GLENN DEATHAustralian Institute of
Marine Science, PMB No. 3, Townsville MC, Qld 4810, Australia
Abstract. The successful management of ecosystems depends on
early detection ofchange and identification of factors causing such
change. Determination of change andcausality in ecosystems is
difficult, both philosophically and practically, and these
diffi-culties increase with the scale and complexity of ecosystems.
Management also dependson the communication of scientific results
to the broader public, and this can fail if theevidence of change
and causality is not synthesized in a transparent manner. We
developeda framework to address these problems when assessing the
effects of agricultural runoffon coral reefs of the Australian
Great Barrier Reef (GBR). The framework is based onimproved methods
of statistical estimation (rejecting the use of statistical tests
to detectchange), and the use of epidemiological causal criteria
that are both scientifically rigorousand understood by
nonspecialists. Many inshore reefs of the GBR are exposed to
terrestrialrunoff from agriculture. However, detecting change and
attributing it to the increasing loadsof nutrients, sediments, and
pesticides is complicated by the large spatial scale, presenceof
additional disturbances, and lack of historical data. Three groups
of ecological attributes,namely, benthos cover, octocoral richness,
and community structure, were used to discrim-inate between
potential causes of change. Ecological surveys were conducted along
waterquality gradients in two regions: one that receives river
flood plumes from agriculturalareas and one exposed to runoff from
catchments with little or no agriculture. The surveysshowed
increasing macroalgal cover and decreasing octocoral biodiversity
along the gra-dients within each of the regions, and low hard coral
and octocoral cover in the regionexposed to terrestrial runoff.
Effects were strong and ecologically relevant, occurred
in-dependently in different populations, agreed with known
biological facts of organism re-sponses to pollution, and were
consistent with pollution effects found in other parts of theworld.
The framework enabled us to maximize the information derived from
observationaldata and other sources, weigh the evidence of changes
across potential causes, make de-cisions in a coherent and
transparent manner, and communicate information and conclusionsto
the broader public. The framework is applicable to a wide range of
ecological assessments.
Key words: Bayesian analysis; biodiversity; bootstrap;
causality; community structure; environ-mental impact;
epidemiology; Great Barrier Reef; model averaging; model selection;
pollution; ter-restrial runoff.
INTRODUCTION
Early detection of ecological change and identifi-cation of
factors causing such change are essential forsuccessful ecosystem
management. Despite the avail-ability of the best scientific data,
interested parties oftendisagree about the existence of ecological
change andits causes. There are many reasons for this, including(1)
the selective use of scientific data and other infor-mation by
interested parties to support individualclaims and objectives, (2)
the misinterpretation andabuse of technical concepts such as
probability andcausality (Newman and Evans 2002), (3) the
complex-ities of large-scale ecosystems which cannot be
simplyexplained or reliably predicted, and (4) the exploitationof
disagreement amongst scientists by stakeholders.The resulting lack
of consensus can lead successively
Manuscript received 15 October 2003; accepted 23 February2004.
Corresponding Editor: P. K. Dayton.
1 E-mail: [email protected]
to conflict, confusion over policy development, gov-ernment
inaction and environmental degradation.
To overcome these problems, we developed a frame-work based on
(1) the use of improved methods fordetermining change through
estimation of effect sizes,as opposed to the usual use of
hypothesis tests(McCullagh and Nelder 1989, Nelder 1999), and
(2)the use of epidemiological criteria to attribute causality.The
framework can synthesize and evaluate scientificand other data
according to criteria that are both sci-entifically rigorous and
widely accepted. Applicationof the framework is simple and
transparent in order toeffectively communicate scientific evidence
to decisionmakers and the public. This enables the detection
ofchange and judgments about causality to be made in arigorous,
structured, and open manner, and thus theagreement among
stakeholders, necessary for success-ful implementation of
management strategies, can beobtained. An application of the
framework follows ina case study on the effects of water pollution
on coralreef benthos in the Great Barrier Reef, Australia.
-
October 2004 1449IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
Statistical issues in the detection of change
The statistical significance test does not tell us whatwe want
to know, and we so much want to knowwhat we want to know that, out
of desperation, wenevertheless believe that it does!
Cohen (1994)Most studies of environmental change (or
impacts)
adopt a falsification perspective; that is, they assumeno change
has occurred and assess the level of evidenceagainst this premise.
If the evidence against nochange is strong, then they accept change
has oc-curred, whereas if the evidence against no changeis weak,
then the initial position is retained. Evidenceagainst the null
hypothesis is almost invariably basedon a frequentist statistical
significance test of a point(precise) null hypothesis. Since
ecosystems are con-stantly changing through time and space such
hypoth-eses are a priori false; i.e., they are not plausible
(Ber-ger and Sellke 1987). Despite the no change premiseseeming
indefensible, the majority of studies continueto adopt it as a
starting point for investigations of tem-poral and spatial change.
Why is this so? As a basisfor management decisions, hypothesis
tests are inad-equate, and it can be argued that
decision-makingshould not be a part of impact studies, which
shouldinform, not decide (Stewart-Oaten 1996a). Even whena point
null hypothesis is plausible, frequentist testsare problematic
(Berger and Sellke 1987, Berger et al.1997). They are remarkably
uninformative (rejector fail to reject) and can be misleading when
im-properly interpreted, e.g., by describing a failure toreject as
evidence of no change or by misinterpre-tation of P values. Such
misunderstandings and therepeated use of tests of low power can
also lead tofalse knowledge as the null hypothesis becomes
ac-cepted as knowledge. Basing decisions on the result oftests also
conflicts with the precautionary principle(Bodansky 1991) and can
be hazardous, e.g., requiringa positive test as evidence of
population decline for arare species with high natural variation
can lead to localextinction. The problem of lack of information as
basisfor accepting the null hypothesis has led to the proposalfor
reversing the burden of proof (Dayton 1998), e.g.,it must be shown
that proposed actions will not resultin environmental damage,
rather than allowing all ac-tions which cannot be shown to result
in damage. Fre-quentists have argued that power analysis
(typicallypost hoc) offers protection against this problem
(e.g.,Cohen 1988). However, this approach has been criti-cized, and
equivalence testing suggested as a formalalternative (Hoenig and
Heisey 2001). Despite suchshortcomings of frequentist hypothesis
tests being re-peatedly noted (Berger and Sellke 1987, Raftery
1995,Stewart-Oaten 1996a, Johnson 1999), the use of testsstill
prevails. Various alternatives have been suggested,primarily
through the use of estimation of parameters
with confidence [credibility] intervals (Stewart-Oaten1996a,
Burnham and Anderson 1998, Burnham et al.2000) or Bayesian methods
(Ellison 1996, Berger etal. 1997, Berger and Pericchi 2001). By
abandoninghypothesis tests in studies of change, immense benefitcan
be derived, irrespective of whether frequentist orBayesian methods
are used. The principal objectivethen becomes one of how to
quantify (model) changeand obtain accurate estimates of parameters
represent-ing the quantities of interest.
When testing for significance or estimating the mag-nitude of
change, typically one or more parameters ofa chosen model represent
the change process. It is oftenforgotten, or not realized, that all
inference, be it hy-pothesis tests or parameter estimation, is
conditionalon the selected model. Thus, the validity of
conclusionsbased on results of significance tests and
parameterestimates are always contingent on the model being
anaccurate representation of reality, i.e., that the modelis true,
or is at least a good approximation. Often,many models are compared
before one is chosen, butthe uncertainty involved in model
selection is seldomtaken into account (Burnham and Anderson 1998),
andthis can result in biased and/or over-precise estimates,and
probability values from hypothesis tests that aretoo small. In
well-designed experiments or surveys, themodel is largely
determined by the study design, andwe can be relatively confident
that it is a reasonablerepresentation of reality. However, when
studies arenot as well controlled, or many variables are
involved,then the choices of model can be vast. This
furtherincreases if we consider interactions, transformationsof
variables, and alternative error structures. If all pos-sible
explanatory variables are included in a model, thepower to detect
change (equivalently, the precision ofthe estimate of change) may
be severely reduced. Con-versely, if important variables are
omitted, then esti-mates of change are likely to be biased. The
problemsof model selection are particularly difficult for
smallsample sizes with weak relationships between the re-sponse(s)
and predictors, and large data sets with manypredictors.
The use of hypothesis tests for model selection hasreceived
intense scrutiny in recent times, and based onsimulations, the use
of tests has been shown to be sub-optimal for identifying true
models (Freedman 1983,Draper 1995, Burnham and Anderson 1998). The
useof selection criteria such as Akaikes information cri-teria
(AIC; Sakamoto et al. 1986, Bozdogan 1987) andBayes information
criteria (BIC; Schwarz 1978, Ma-digan and Raftery 1994) have been
advocated, partic-ularly in the ecological literature (Burnham and
An-derson 1998). These criteria outperform the use of hy-pothesis
tests in determining true models, however nosingle criteria will
universally find the true or best ap-proximating model. Based on
AIC or BIC, weights ofevidence can be calculated for each of the
competingmodels. These weights quantify the uncertainty of mod-
-
1450 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
el selection and can be treated as relative probabilitiesof the
models. They do not, however, represent theprobability that any
particular model is the true model,or best approximating model of
all possible models,since all of the proposed models may be
deficient. Ifnone of the models are true (or good
approximations),then inferences based on the best model are also
likelyto be biased.
One way to overcome the issue of choosing a singlemodel is
simply to avoid it by model averaging. Thisalso has the attractive
property of generating more ac-curate estimates and predictions
(Hastie et al. 2001).Model averaging involves the fitting of
several plau-sible models to the data and averaging the results
(ei-ther parameter estimates or predictions) over all of themodels.
The averaging is usually weighted by somemeasure of the relative
probability or predictive ac-curacy of each model, with more likely
models re-ceiving proportionally more weight (Raftery 1988,Burnham
and Anderson 1998). In this way, poor mod-els receive little weight
and have a negligible influenceon the final model. Criteria such as
AIC (or variationsthereof) and BIC are often used for this purpose.
Theuse of BIC in this manner is an approximation to Bayes-ian model
averaging, a process which is technicallycomplex and
computationally demanding, though theneed to use approximations has
declined through theuse of increased computational power, and Monte
Carlosimulations can give more accurate results comparedto BIC. As
is the case with selection of a single model,we need some (or at
least one) of the models to be agood approximation to reality for
inferences to be un-biased.
The shift from hypothesis testing to estimation, anduse of model
averaging to better manage the uncer-tainty of model selection,
does not negate the impor-tance of good sampling design in
environmental studies(Schmitt and Osenberg 1996). Indeed, the
additionaluncertainty of model selection is incorporated into
theestimation of effect sizes and thus these estimates willbe less
precise, but hopefully more honest, than thosebased on a single
model.
Attribution of causalityThe concept of causality has a long and
complex
history and it has many meanings. Its everyday usageis often
straightforward, but as a philosophical or sci-entific concept, its
definition and use are often conten-tious. At this level, at least
three notions of causalityare supported (Pearl 2000, Gillies 2001).
First, somedeny the existence of causality or view it as
scientifi-cally unnecessary. Second, some acknowledge causal-ity as
a useful concept, but do not give it a central rolein their models.
Causality is thus seen as a useful wayof explaining some aspects of
empirical laws. Third,some advocate causality as a fundamental
construct.Cause is assumed rather than demonstrated, and in
theformulation of mathematical models, causality takes
precedence over probability. Of course, there is no con-clusive
argument giving universal support to any oneof the three views, and
different phenomena can beused to support differing views, e.g.,
our everyday ex-periences support the concept of causality as
somethingfundamental, but few mathematical models require it.
Statistics is widely accepted within ecology as a pri-mary
empirical methodology, yet statisticians are typ-ically not strong
advocates of causality as a funda-mental construct, favoring
probability models instead.There are statistical approaches that do
promote causeover probability (Pearl 2000), but equally there
arewarnings against the causal interpretations (Speed1990). The
randomized experiment is often invoked asone method for
unambiguously determining causality,but even that is questionable
when outcomes are sto-chastic and we rely on statistical
interpretation.
Causal arguments are needed in ecosystem manage-ment in order to
convince interested parties that man-agement actions should be
implemented and will beeffective. These arguments need to balance
scientificrigor with ease of communication to nonscientists.
Thissituation is not novel and we can borrow from epide-miology,
which deals with issues of comparable com-plexity to the ecological
and environmental sciencesand also has similar requirements of
scientific rigor andcommunication. Epidemiologists developed
criteria toassess causality as part of the research into the
linkbetween cigarette smoking and lung cancer. This linkwas
accepted by the Surgeon General after decades ofresearch (U.S.
Department of Health, Education, andWelfare 1964) when evidence
compiled from multiplesources of information and numerous studies
fulfilleda set of criteria (Hill 1965), the main ones being:
1) The relationship between the dose (the putativecause) and the
response should be monotonic.
2) The association between the dose and the responseshould be
strong.
3) The response should be specific to the cause.4) There should
be a logical time sequence of events;
i.e., the response should occur after the dose has
beenapplied.
5) There should be consistency both across popu-lations within a
study, and with results from other stud-ies.
6) The observations should agree with known bio-logical
facts.
These criteria (or subsets, extended sets, or redefinedsets of
them) are routinely used in epidemiology tojudge whether or not an
association is causal. None ofthe criteria are taken as indicative
by themselves, butequally, none are seen as absolutely necessary to
eval-uate causal significance of associations (Roth et al.1982).
The more criteria that are satisfied and the stron-ger the
association, the more confidence we shouldhave in our judgment that
the association is causal.Similar criteria have been proposed for
ecoepidemiol-ogical studies (Fox 1991) and impact assessment
stud-
-
October 2004 1451IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
FIG. 1. Maps of the northern Great Barrier Reef and the study
regions. (a) The risk of exposure to agricultural runoff isshown in
four shades of gray indicating (from dark to light) high, moderate,
small, and minimal risk (from Devlin et al.2002). Boxes surround
the two study regions (PC, Princess Charlotte Bay; WT, wet
tropics). Also shown are the locationsand names of inshore target
reefs (black circles) in (b) PC and (c) WT, and locations of
additional reef surveys across thecontinental shelf (smaller gray
circles), and estuaries of the main rivers affecting the
regions.
ies (Schroeter et al. 1993, Stewart-Oaten 1996b,
U.S.Environmental Protection Agency 1998). The criteriamay need to
be adapted or interpreted for particularstudies; e.g., the
doseresponse relationship may benon-monotonic due to toxicity, and
the strength of therelationship could be expressed in many ways
depen-dent on how dose and response are measured. In thiswork, we
formalize these procedures and extend its useto multiple possible
causes.
We have argued that a shift from the use of hypoth-esis tests to
estimation of parameters, and adoption ofbetter model selection
processes or model averagingcan lead to more informative analyses
of the detectionof change. We have also suggested that causal
criteriacan be used to rigorously yet transparently
attributecausality. Finally, by selecting combinations of
eco-logical attributes that are complementary with respectto
possible causes of change, we can better discriminatebetween likely
agents of change, and reduce the like-lihood of confounding that
may lead to spurious find-ings. The chosen attributes may be
aspects of the phys-ical-chemical environment (either measured
directly oras proxies), abundances and biodiversity of key
speciesgroups, or ecological processes. These processesim-proved
statistical analysis, use of epidemiological caus-al criteria, and
selection of combinations of comple-mentary ecological
attributescan lead to more effec-tive ecological assessments, and
we illustrate this inthe following case study.
THE CASE STUDYIncreasing terrestrial runoff of nutrients,
sediments
and pesticides is a major management issue facing theAustralian
Great Barrier Reef (GBR), but the presenceof measurable effects of
runoff on inshore areas has
been controversial (Bell 1991, Larcombe and Woolfe1999, van
Woesik et al. 1999, Haynes and Michalek-Wagner 2000, Brodie et al.
2001, Devlin et al. 2001,Furnas 2003, McCulloch et al. 2003). River
dischargesare the principal source of nutrients and sediments
forthe shallow continental shelf waters (Furnas and Mitch-ell
2001), and land clearing for agriculture, removal ofsoil-retaining
wetlands, and intensive use of agricul-tural chemicals have
increased nutrients and sedimentsin these discharges three- to
11-fold since Europeansettlement in 1850 (Furnas 2003). Discharges
from the423 000-km2 catchment area contained 1114 3 106 Mgof
sediment in 2002 compared with 14.4 3 106 Mgbefore 1850, and trends
of increasing soil erosion arerecorded in coral cores (McCulloch et
al. 2003). Fur-ther, 100 000 Mg of nitrogen and 20 000 Mg of
phos-phorus fertilizers are now applied to the catchmentsannually,
though how much of it eventually enters themarine system is unknown
(Furnas 2003). A regionalong the 200 km long wet-tropical coastline
betweenTully and Port Douglas containing 60 coral reefs within20 km
from the coast (latitude 188009 to 168209 S,longitude 1468109 to
1458309 E; Fig. 1) has been iden-tified as the area of greatest
risk from agricultural run-off (Devlin et al. 2002). However,
causal links betweenpollution and reef degradation in this region
have beendifficult to demonstrate. This is due to factors such asa
lack of historical data, the large spatial scale, thepresence of
natural cross-shelf gradients in communitystructure (Dinesen 1983)
and suspended particulatematter (Furnas 2003), and the presence of
other typesof disturbance. As in many environmental impact
stud-ies, there is no spatial replication for impacted
andnonimpacted regions. Furthermore, the spatial and tem-poral
variability of the flood-related episodic river dis-
-
1452 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
PLATE 1. A naturally turbid but highly di-verse inshore reef in
Princess Charlotte Bay.Photo credit: K. Fabricius.
charges of several pollutants is high, and the fate ofpollutants
while undergoing dilution, biological up-take, sediment burial, and
repeated resuspension duringthe transport from river to reef is
little understood.Finally, ecological responses (linear or
threshold re-lationships, synergistic responses) vary greatly
acrossthe multitude of organisms that characterize the
highlydiverse coral reef ecosystem.
Around well-defined point sources such as sewageoutfalls or
coastal developments, increased sedimentsand nutrients are known to
cause local reduction incoral recruitment, increase mortality, and
shift the dom-inance from hard corals to non-reef building
organisms(Smith et al. 1981, Wittenberg and Hunte 1992, Hunterand
Evans 1995). Results from laboratory and fieldexperiments also
demonstrate detrimental effects ofsedimentation and pesticides on
individual organismgroups and life stages (Rogers 1990, Dubinsky
andStambler 1996, Jones et al. 2003, Philipps and Fabri-cius 2003).
While enhanced concentrations of inorgan-ic nutrients appear to
have no direct effects on coralhealth (Szmant 2002), they can
affect coral populationsindirectly, e.g., by shifting competitive
advantages to-ward otherwise nutrient-limited algae when
grazingpressure is low (McCook 1999), or by the formationof marine
snow (Fabricius et al. 2003). Thus, whilecausal links between
pollution and reef degradationhave been difficult to demonstrate at
regional scalessuch as the inshore reefs of the GBR, pollution
impactsare well documented and accepted at local scales andunder
controlled conditions.
In this study, we assess possible associations be-tween the
state of some GBR inshore reefs that areexposed to terrestrial
runoff, but have also been ex-posed to a number of other
disturbances. This casestudy has a high political and environmental
profile,both locally due to large economic interests in a
healthyreef through tourism revenues, and globally since land-based
pollution and coastal development put 22% and30%, respectively, of
coral reefs on Earth at risk (Bry-ant et al. 1998).
METHODS
Field data
The study is based on water quality analyses, andecological
surveys of benthic cover, biodiversity, andoctocoral community
structure. Only summaries of therelevant water quality and
ecological surveys are pre-sented here; other laboratory and field
studies havebeen or will be described in greater detail
elsewhere(Fabricius et al. 2003, Phillips and Fabricius
2003,Diaz-Pulido and McCook 2003; K. E. Fabricius, G.Death, E.
Turak, and D. Williams, unpublished man-uscript).
Study sites and survey methods.The field researchwas carried out
in two regions within the Great BarrierReef (GBR), with one-off
surveys characterizing 54reef sites across the whole continental
shelf, and tar-geted research on 13 inshore reefs (Fig. 1). The
wettropics (WT) lies between Tully and Port Douglas, andinshore
reefs experience local river plumes from ag-ricultural catchments
on an almost annual basis, andlarge plumes from the distant
Burdekin River on a de-cadal basis (Furnas 2003). This region has
the highestexposure to runoff from agricultural areas within theGBR
(Fig. 1a; Devlin et al. 2002). The second regionlies north of
Princess Charlotte Bay (PC) and ;400km north of WT, and the inshore
reefs are exposed torunoff from sparsely populated catchments that
havereceived little or no fertilizer and pesticides to date,but
have low-density cattle grazing in some parts (seePlate 1). Both
regions contain turbid inshore reefs insimilar geophysical
settings, located within 20 km ofthe coast at 1218 m depth of the
surrounding sea floor,and protected by a barrier of mid- and
outer-shelf reefsup to 40 km offshore (Fig. 1b and 1c). In WT,
mostdischarged material is eventually transported north-ward away
from the reefs, whereas the large north-facing PC creates
anticyclonic eddies which result intrapped and deposited sediment
(Torgersen et al. 1983).The research on the inshore target reefs
was conductedbetween 2000 and 2002. An additional 40 mid- and
-
October 2004 1453IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
outer-shelf reef sites were visited within the two re-gions for
cross-shelf one-off surveys of benthic coverand octocoral
communities.
The detailed disturbance histories of individual reefsare
largely unknown. In the WT, some reefs experi-enced outbreaks of
the coral-eating crown-of-thornsseastar Acanthaster planci in the
late 1990s, tropicalcyclones in 1986 and 1990, and mortality
through coralbleaching (the expulsion of symbiotic algae from
thecoral tissue, primarily caused by high temperatures) in1998,
with bleaching estimated as moderate to ex-treme on most reefs
(Berkelmans and Oliver 1999).In PC, no data exist for
crown-of-thorns seastar (highnumbers were observed on a reef
neighboring the in-shore target reefs in 1991), but four tropical
cycloneshave passed through the region within the last two
de-cades. It is likely that the PC inshore target reefs didnot
bleach in 1998 since satellite-based estimates ofsea surface
temperatures were near-normal, but somereefs did suffer severe
bleaching mortality in early2002 after the surveys were completed
and during thecoral settlement experiment. The data on
crown-of-thorns, cyclone, and bleaching disturbance historywere
insufficient to assess or attribute effects on thescale of
individual reefs. In order to distinguish be-tween the potential
causes of change, a combination ofecological attributes with
contrasting responses tobleaching, crown-of-thorns seastar
predation, runoff,and cyclones were chosen for the study. The
chosenattributes were:
1) Benthic cover of hard corals, octocorals, and ma-croalgae.
These are the main groups of organisms usu-ally measured in the
assessment of coral reefs, and wereexpected to respond to changing
environmental con-ditions and disturbances in contrasting ways.
2) Taxonomic richness of zooxanthellate and azoox-anthellate
octocorals (Anthozoa, Octocorallia; com-monly termed soft corals
and sea fans). Thisgroup contains genera with and without symbiotic
al-gae (called zooxanthellae) in their tissue, the formergroup
depending on water clarity and light for pho-tosynthetic nutrition,
whereas the latter group is in-dependent of water clarity
(Fabricius and Death 2001).Octocorals were chosen as indicators for
ecologicalattributes because of their abundance, and because
theyrespond more specifically to water quality than hardcorals;
azooxanthellate octocorals (which constituteabout half of the
genera) do not suffer from coralbleaching, while zooxanthellate
octocorals respondstrongly to turbidity, probably because of low
photo-synthetic efficiency (Fabricius and Klumpp 1995). Oc-tocorals
are also rarely eaten by crown-of-thorns seas-tar (Death and Moran
1998).
3) Community structure of octocorals on both theinshore target
reefs and along the cross-shelf chloro-phyll gradient. This measure
was chosen because com-munities are known to respond more strongly
to en-
vironmental conditions than abundances of the maingroups (K. E.
Fabricius, unpublished data).
One-off rapid ecological assessment surveys (Fa-bricius and
Death 2001, Fabricius and Alderslade2001) were used to characterize
the ecological condi-tion of 54 reef sites across the continental
shelf in bothregions, and 13 inshore target reefs in both
regions.Surveys were conducted on two sites per reef (wind-ward and
leeward sides) at five depth zones per site(018 m); each survey at
each depth zone covered;500 m2 of reef area. Survey data were
collected onpercentage cover of the main benthos groups (hard
cor-al, octocoral, macro algae, turf algae, coralline algae,sand
and rubble) and taxonomic inventories and abun-dance estimates
(rating 05) of all genera of octocorals.
Water quality data.Two sets of water quality datawere available
from the two regions. First, a 10-yr dataset of chlorophyll
concentrations at sites across thecontinental shelf on the GBR in
both regions (J. Brodie,G. Death, M. Skuza, and M. Furnas,
unpublished man-uscript). The chlorophyll measurements were
sampledup to 12 times a year at each site. Second, water
qualitydata were collected around the inshore target
reefs.Concentrations of water quality parameters (suspendedsolids,
particulate nitrogen and phosphorus, nitrate, ni-trite and
ammonium, phosphate, total dissolved nitro-gen and total dissolved
phosphorus, chlorophyll andphaeopigments, salinity and silicate)
were determinedfrom water samples taken at each of the inshore
targetreefs during nine visits between December 2000 andApril 2002.
Water analyses followed standard proce-dures (Furnas and Mitchell
1996). Only a short sectionof the chlorophyll gradient is
represented in the inshorewater quality samples, as the innermost
reefs in PCwere avoided due to the likely presence of
saltwatercrocodiles (Crocodylus porosus). All nutrient data ex-cept
salinity were highly correlated and they were stan-dardized (z
scores) and summed to form a water qual-ity index for each reef.
Low index values correspondto water with low nutrients,
chlorophyll, and suspendedparticles.
Statistical methodsThe analytical methods of this study did not
include
the use of hypothesis tests for the reasons argued inthe
Introduction. Instead, we estimated effect sizes andpredicted
values with interval coverages. Inferenceswere based on model
averaging, cross-validatedsmoothing splines and bootstrap
estimation. For eachof these methods, the numerical results can be
inter-preted from either frequentist or Bayesian perspectives,but
of course the interpretations differ in each case.All data analyses
used S-Plus (Statistical Sciences1999).
The relationships between chlorophyll concentra-tions and
octocoral richness (zooxanthellate and azoo-xanthellate) were
modeled for each region (WT andPC) as a function of relative
cross-shelf distance (de-
-
1454 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
fined as the distance of a site from the coast dividedby sum of
distances from the coast and the edge of theouter continental
shelf). The relationships of the re-sponses with relative
cross-shelf distance were nonlin-ear, and smoothing splines were
used with the degreeof smoothing estimated by cross-validation
(Hastie andTibshirani 1990).
The water quality data and chlorophyll data were pre-analyzed
prior to inclusion in models relating the ben-thic variables to
gradients. To investigate the relation-ships between the inshore
water quality variables, aprincipal components analysis was used. A
water qual-ity index was then calculated as the sum of all
stan-dardized (z scores) variables other than salinity, andscores
on this index were used as measures of waterquality for each reef
in subsequent analyses. The eco-logical survey data and the
long-term chlorophyll datawere not recorded at identical sites,
hence chlorophylllevels at the survey sites were estimated by the
weight-ed mean of nearest neighbors from the chlorophyllsites.
Three sets of ecological attributes, namely (1) ben-thic cover
(hard corals, octocorals, and macroalgae),(2) taxonomic richness
(zooxanthellate and azooxan-thellate octocorals), and (3) community
composition(all octocorals) were related to gradients (either
chlo-rophyll or inshore water quality) and regional (WT vs.PC)
differences. For both the chlorophyll and waterquality gradient
analyses of benthic cover and richness,log-linear regression models
with linear gradient ef-fects and categorical regional effects were
used sincevariation increased with the mean and the implicit
logtransformation helped linearize gradient effects. Foreach
response, five models were fitted (1) differentslopes (gradient
effects) within each region and dif-ferent intercepts (region
effects), (2) same slope forboth regions, but different intercepts
(region and gra-dient effects), (3) single gradient common to both
re-gions, (4) no gradient effect but region effects, and (5)no
gradient or region effects. The regional effects wereincluded to
account for biological differences due toregion, which were partly
confounded with the gradi-ents. The data sets were small with ;50
and 1320observational units for the chlorophyll or inshore
waterquality data respectively.
Preliminary analyses indicated relatively weak as-sociations
between the responses and explanatory var-iables for the smaller
data sets, and suggested that con-clusions based on hypothesis
tests may not adequatelyreflect seemingly consistent patterns
across the chlo-rophyll and inshore water quality data. Thus, in
orderto select an optimum form of analysis, we conductedsimulations
based on estimates of effect sizes and errorobtained from
preliminary analyses of benthic coverand richness gradient data.
These simulations showedBIC (Schwarz 1978) to be marginally better
than bothAIC and AICc (Burnham and Anderson 1998) for mod-el
selection, and much better than hypothesis tests (see
Appendix). However, none of these methods reliablyselected the
true model, and model averaging (Raftery1988, 1995) gave slightly
better predictions than singlebest models. Hence, we have used
model averaging forall gradient analyses of benthic cover and
richness.Confidence (credibility) intervals were obtained
bybootstrapping (Efron and Tibshirani 1993, Davison andHinkley
1997).
Redundancy analyses (RDA; Rao 1964, Jongman etal. 1995) were
used to assess the dependence of oc-tocoral communities on regional
differences (WT andPC) and on both cross-shelf chlorophyll data and
in-shore water quality data. The abundances of octocoralswere
fourth-root transformed to downweight dominanttaxa, and reef
averaged over depths and sites. For com-munity analyses involving
the chlorophyll gradient,relative distance across the shelf was
also included.Ecological gradients were relatively short
(moderatespecies turnover) thereby justifying the use of RDA,which
requires linear changes of species along gradi-ents. The strengths
of regional and gradient effectswere quantified by bootstrapping
(Efron and Tibshirani1993, Davison and Hinkley 1997) the pseudo-F
statistic(ter Braak 1992).
Synthesis
To successfully use the causal criteria, both the eco-logical
attributes and the criteria have to be defined foreach study
individually. Our ecological attributes werecover of hard corals,
octocorals, and macroalgae, spe-cies richness of zooxanthellate and
azooxanthellate oc-tocorals, and community structure, across the
shelfalong the chlorophyll gradient and on the inshore targetreefs
along the water quality gradient. The criteria wereslightly
modified from Hills initial list (Hill 1965) tobest reflect the
nature of the case study, using the fol-lowing definitions:
1) Doseresponse relationship was satisfied if theprobability of
gradient effect was .0.99. This corre-sponds to a Bayes factor of
100:1 and can be taken asstrong evidence for a relationship between
dose andresponse (Raftery 1995).
2) Strength of association was defined by effect size.A strong
effect was defined as .100% increase (or 50%decrease) along 80% of
the length (to exclude extremevalues) of the inshore water quality
gradient within WT.
3) Logical time sequence indicates that the changedid not
precede exposure to the disturbance. An as-sessment of this
criterion relies on historic data fromthe study regions, which were
sparse in our case (nodata exist from the remote PC, and only a few
obser-vations exist from WT; Ayling and Ayling 2002; AIMSLong-Term
Monitoring Program, unpublished data).
4) Consistency across populations was defined asconsistency with
other studies (i.e., responses re-corded in our study were similar
to those reported fromother regions in independent published
studies).
-
October 2004 1455IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
TABLE 1. Comparison of water quality values around inshore reefs
of the wet tropics (WT) and Princess Charlotte Bay(PC).
Inshore water qualityPC
Mean SE
WTMean SE Ratio 95% CI P
Suspended solids (SS; mg/L)Chlorophyll (Chl; mg/L)Phaeopigments
(phae; mg/L)Particulate nitrogen (PN; mmol/L)Nitrate (NO3;
mmol/L)Ammonium (NH4; mmol/L)
1.400.400.191.410.0240.16
0.130.040.030.130.010.02
3.770.560.322.550.140.24
0.840.090.060.380.070.08
2.691.401.681.815.831.50
1.49, 3.990.93, 1.990.97, 2.731.27, 2.490.71, 27.40.55, 2.67
.0.990.930.960.990.880.83
Nitrite (NO2; mmol/L)Total dissolved nitrogen (TDN;
mmol/L)Particulate phosphorus (PP; mmol/L)Total dissolved
phosphorus (TDP; mmol/L)Silicate (Sil; mmol/L)Salinity (Sal;
g/L)
0.0128.460.0960.434.04
31.8
0.0020.590.0120.061.190.48
0.0198.380.160.569.46
29.1
0.0031.010.030.062.591.54
1.580.991.671.302.340.93
0.99, 2.460.75, 1.341.01, 2.550.95, 1.921.08, 6.380.82, 1.01
0.960.490.970.930.990.02
Notes: The table lists means and standard errors for each region
together with the ratio for WT/PC, 95% confidence(credibility)
intervals based on bootstrap resampling, and the probability that
the WT/PC ratio is .1. The intervals can betreated as confidence
intervals or credibility intervals with a non-informative
prior.
5) Specificity referred to responses that were known(from
published literature) to be caused by runoff andwas unlikely to be
caused by another disturbance type.
6) Agreement with biological facts was scored bycomparing the
responses found in our field surveyswith the results of studies
where relationships weredirectly assessed through manipulative
experiments.
By applying each of Hills criteria to each of theecological
attributes, we created a matrix that was usedto determine to which
extent the observed types orchanges in the inshore coral reefs
might have beencaused by exposure to runoff. Criteria 13
(dosere-sponse relationship, strength of association, and
logicaltime sequence) were assessed based on the relation-ships
between the ecological attributes and the twowater quality data
sets. Criteria 46 (consistency withother studies, specificity of
response, and agreementwith biological facts) were scored by
comparing theresponses of the ecological attributes with results
fromother independent sets of published studies
(laboratoryexperiments on pollution effects, and data from
pol-luted locations in other regions). For ease of commu-nication,
we summarized all results in the matrix byscoring each cell in one
of three possible ways (cellsthat could not be addressed due to the
lack of data weremarked with na): agreement of the response of
theattribute with the criterion (11), weak or inconclu-sive
response due to inconsistent results or weaknessin the study design
(0), responses that are in dis-agreement with the criterion (21).
The overall evi-dence for a causal association with water quality
wasthen assessed as the sum over all cells for each of
theecological attributes, expressed as proportion of thetotal
number of cells for each attribute for which datawere
available.
RESULTSField data
Regional differences in water quality and ecologicalattributes
on inshore target reefs.Mean concentra-
tions of suspended solids, particulate nitrogen, and
par-ticulate phosphorus were 170270% higher in the wateraround the
WT reefs than in the water around PC reefs,and mean nitrate levels
were 580% higher in WT thanin PC (Table 1). Concentrations of other
dissolved andparticulate nutrients were also at least as high or
higherin WT than in PC. All water quality variables exceptsalinity
were highly correlated, and most of the inshoretarget reefs in WT
were exposed to higher nutrient andsediment concentrations than in
PC (Fig. 2).
The ecological attributes of WT also differed sub-stantially
from those of PC. On WT reefs 67.7% 65.6% (mean 6 1 SE) of space
was covered in algae(turf, coralline, and macroalgae). In contrast,
algae oc-cupied 39.6% 6 5.4% of space on PC reefs. Coral coverwas
lower in WT than in PC (mean hard coral cover,15.1% 6 2.4% vs.
43.4% 6 1.0%; octocorals, 2.8% 60.8% vs. 5.2% 6 1.5%). The richness
of octocoralswas also lower in WT than in PC (25.5 6 2.4 vs. 35.66
2.3 genera per reef). In contrast, dead coral coverin WT was higher
than in PC (17.2% 6 6.0% vs. 3.8%6 0.4%).
Changes in ecological attributes along the chloro-phyll gradient
across the continental shelf.The gra-dients in mean water column
chlorophyll concentra-tions across the continental shelf differed
between PCand WT (Fig. 3). In WT, chlorophyll increased
steeplytowards the coast, and was up to three times higher inthe
innermost 20 km of the region than offshore or inPC. In PC,
chlorophyll remained constant and rela-tively low across the shelf.
Changes in the taxonomicrichness of octocorals across the
continental shelf alsoshowed clear patterns. Richness of
zooxanthellate oc-tocorals (genera that contain symbiotic algae in
theirtissue and depend on water clarity and light for
theirnutrition) declined by 30% in the WT inshore regionwithin the
innermost 20 km of the shelf where chlo-rophyll was high (Fig. 3).
In contrast, no clear cross-shelf changes in richness of the
zooxanthellate genera
-
1456 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
FIG. 2. Principal components biplot of water quality
data(log-transformed and z-score-transformed) at inshore
targetreefs in the two study regions (gray, Princess Charlotte
Bay;black, wet tropics). The first two components accounted
for78.6% of water quality variation. Each symbol represents atarget
reef, for which data were averaged over within-reeflocations and
sampling times; the extent of fill of the symbolsrepresents
octocoral richness, with lowest and highest rich-ness displayed as
open and completely filled symbols, re-spectively. Reefs from the
same region are surrounded by apolygon. Vectors of the water
quality variables and water qual-ity index point at reefs with
highest concentrations. Abbre-viations of the water quality
variables are listed in Table 1.
were apparent in PC where chlorophyll was stableacross the
shelf. The richness of octocorals withoutzooxanthellae (which do
not require light for their nu-trition) varied little across the
continental shelf in WTand PC and appeared unrelated to the
chlorophyll gra-dient.
Some of the ecological attributes varied systemati-cally along
the chlorophyll gradient in both regions(Fig. 4, Table 2).
Macroalgal cover increased with in-creasing chlorophyll
concentrations in both regions.While hard coral cover was highly
variable and unre-lated to chlorophyll and region, octocoral cover
de-clined with chlorophyll, reaching lowest cover at achlorophyll
concentration of .0.3 mg/L. The richnessof zooxanthellate
octocorals decreased with increasingchlorophyll, whereas the
richness of azooxanthellateoctocorals was unrelated to
chlorophyll.
Octocoral communities showed strong cross-shelf(pseudo-F 5
9.52), chlorophyll (pseudo-F 5 5.02), andregional effects (pseudo-F
5 3.13; Fig. 5, Table 3).Community composition varied most along
the chlo-
rophyll gradient within the WT region, with highestchlorophyll
levels associated with lowest richness. Fewgenera occurred in areas
of high chlorophyll concen-trations, and these conditions were only
found on someof the WT inshore reefs. While mid-shelf reefs
hadhighest richness, some wave-exposed outer-shelf reefsalso had
relatively low richness, but were characterizedby a very different
suite of species than the low-di-versity high-chlorophyll
near-shore reefs.
Changes in ecological attributes along the waterquality
gradients on the inshore target reefs.Changesin ecological
attributes along the inshore water qualitygradient were weaker than
those along the chlorophyllgradient across the shelf, most likely
because of therelative shortness of the inshore gradient and the
lowernumber of inshore reefs (1320). Nonetheless, three ofthe
ecological attributes varied along the water qualitygradient in
similar ways to the cross-shelf variationalong the chlorophyll
gradient (Fig. 4, Table 2). Onlyone of the three effects would have
been detected hadfrequentist tests been used for the analyses
(Table 2).Macroalgal cover increased along the water quality
gra-dient in WT around threefold, but was variable in PC.Hard coral
cover and the richness of azooxanthellateoctocorals both showed
strong regional differences, butwere unrelated to the water quality
gradient. Octocoralcover was similar in both regions and also
appearedunrelated to water quality. In contrast, the richness
ofzooxanthellate octocorals declined with increasing nu-trients
along the water quality gradient in both regions.
The octocoral communities varied most stronglyalong the water
quality gradient (pseudo-F 5 3.39,Table 3) and to a lesser degree
between regions (1.25).Generic richness was highest in clear water
PC reefs,and most genera were absent or occurred in low num-bers in
WT water of low water quality (Fig. 6). Withineach region, reefs
associated with the highest nutrientand sediment levels were those
with lowest richnessand vice versa. Only two octocorals species
(the en-crusting Briareum sp. and Clavularia koellikeri)
wereassociated with high nutrients, whereas a large pro-portion of
genera were strongly associated with thelow-nutrient PC reefs.
SynthesisWe combined the results of our field study with re-
sults from other regions and from laboratory experi-ments, and
used Hills causality criteria to evaluate thepotential link between
water quality and the conditionof the inshore coral reefs. The
results were expressedconcisely in matrix format with the rows of
the matrix(Table 4) defined by the criteria, and the columns bythe
ecological attributes, namely, cover of the mainbenthos groups,
richness of zooxanthellate and azoo-xanthellate octocorals, and
community structure.
Doseresponse relationship.There was strong ev-idence for
doseresponse relationships for three of thefive ecological
attributes both along the cross-shelf
-
October 2004 1457IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
FIG. 3. Cross-shelf gradients in mean chlorophyll
concentrations, based on 10 years of chlorophyll monitoring data
andtaxonomic richness in octocorals with and without zooxanthellae
(PC, Princess Charlotte Bay; WT, wet tropics). Thick solidlines
indicate cross-validated smoothing splines; thinner lines are 95%
confidence intervals. Vertical light gray bands indicatethe
locations of the inshore target reefs across the shelf.
chlorophyll gradient and along the water quality gra-dient on
the inshore target reefs (Fig. 4, Tables 2 and3). Doseresponse
relationships were established formacroalgal cover, octocoral
cover, generic richness ofzooxanthellate octocorals, and octocoral
communitystructure. In contrast, hard coral cover, although
muchlower in WT than in PC, was unrelated to the chlo-rophyll and
inshore water quality gradients, possiblydue to the effect of other
disturbances. The richness ofazooxanthellate octocorals was also
unrelated to waterquality, as expected from their biological
requirements.In all of the relationships, the directions of change
(i.e.,increase or decrease along the water quality gradient,or
differences between regions) agreed with those ex-pected from
existing biological knowledge.
Strength of association.Some of the effects alongthe water
quality gradient across the shelf and withinthe WT inshore region
were large and ecologically sig-nificant (Fig. 4, Table 2).
Macroalgal cover was 3.1times higher on the reefs with highest
nutrient and par-ticle loads compared with those in clearest water
inWT. The number of octocoral genera was 2.4 timeshigher on WT
reefs with clearest water compared withthose in the least clear
water. Associations betweencommunity structure and the water
quality and chlo-rophyll gradients were stronger than the
differencesbetween the regions (Table 3), with depauperate
com-munities recorded in the reefs with highest nutrient
andparticle loads.
Logical time sequence.The assessment of this cri-terion is
limited by the scarcity of historic data. Fewdata exist to compare
the ecological attributes on anyof the inshore target reefs with
those from the past, andsome differences between the two regions in
biodiver-sity are likely to have always existed due to the
naturaldecline in biodiversity with increasing latitudes. Theonly
historic data available are for one reef in the WTregion where hard
coral cover decreased from ;80%in 1989 to the present state due to
a number of dis-turbances (Ayling and Ayling 2002). The logical
timesequence criterion of water quality effects on coralreefs has
however been met in other places, e.g., inHawaii, where coral cover
increased and algal coverdecreased following offshore diversion of
a coastalsewage outfall site (Smith et al. 1981, Hunter and
Evans1995).
Consistency across different studies.The findingsof increased
macroalgal cover and low hard coral coverare consistent with
previous work in other coral reefregions exposed to terrestrial
runoff (e.g., Tomascik etal. 1993, Edinger et al. 2000, Hodgson and
Yau 1997,van Woesik et al. 1999, West and Van Woesik 2001).Few
studies have included octocorals, but species rich-ness in response
to disturbance have been reported forother groups of reef organisms
(Edinger et al. 1998).The response of zooxanthellate octocorals,
and the ab-sence of a response in azooxanthellate octocorals
areconsistent with a study on larger scale on the relation-
-
1458 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
FIG. 4. Relationship between ecological reef attributes and
chlorophyll and water quality. Gray and black lines and
circlesindicate reefs within the PC and WT regions, respectively.
Solid lines indicate the model-averaged predictions from log-linear
model fits, dashed lines indicate 95% confidence (credibility)
intervals derived by bootsrapping, and the (pseudo-) R2are the
explained deviance of the model-averaged fit divided by the
deviance of the null model. The left column of plotsshows variation
of ecological attributes along the chlorophyll gradient across the
continental shelf, and the right columnsimilarly for the water
quality index, calculated as the sum of the z scores of all water
quality variables excepting salinityof each reef at inshore target
reefs (high index values, high nutrient concentrations; low values,
cleaner water). Ecologicalattributes are: HC, hard corals; OC,
octocorals; MA, macroalgae; zoox, zooxanthellate; azoox,
azooxanthellate. Richness isdefined as the number of genera of
octocorals per reef encountered during swim surveys.
-
October 2004 1459IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
TABLE 2. Analyses of models relating ecological responses to (a)
chlorophyll gradients and cross-shelf regions and (b)inshore water
quality gradients and regions.
Variable R 3 G R 1 G G R 1 G exp(G ) 95% CI Pa) Cross-shelf
(chlorophyll)
MA cover (%)HC cover (%)OC cover (%)OC richness (zoox)OC
richness (azoox)
0.030.000.190.100.07
0.170.020.440.150.40
0.80*0.100.37*0.75*0.00
0.000.140.000.000.52*
0.000.73*0.000.000.01
0.9120.0121.4920.75
0.29
2.480.990.230.471.34
1.93, 3.180.92, 1.070.15, 0.340.38, 0.600.95, 1.95
.0.990.61
.0.99
.0.990.04
b) Inshore (water quality index)MA cover (%)HC cover (%)OC cover
(%)OC richness (zoox)OC richness (azoox)
0.390.110.240.050.09
0.210.270.280.190.29
0.240.000.210.72*0.26
0.060.62*0.25*0.020.36*
0.10*0.000.020.010.00
1.120.03
20.0120.3420.24
3.061.030.990.710.79
1.61, 6.710.82, 1.320.60, 1.610.59, 0.870.55, 1.13
.0.990.390.51
.0.990.91
Notes: Abbreviations: MA, macroalgae; HC, hard coral; OC,
octocoral; zoox, zooxanthellate; azoox, azooxanthellate. Log-linear
models were used for all analyses. In the models, the gradient (G)
effects are linear, and the region (R) effects arecategorical. For
each ecological response, five models are compared: R 3 G,
different gradient effects (slopes) within eachregion and different
region effects (intercepts); R 1 G, same slope but different region
effects; G, single slope common toboth regions; R, no gradient
effect but regional effects; and 1, no gradient or regional
effects. The best model (most likelyaccording to Bayes information
criteria [BIC]) is denoted by boldface text, and the best model
selected by backward eliminationof nonsignificant terms (P . 0.05)
is indicated by an asterisk. For four of the 10 models, the most
likely model was notselected by backward elimination. For four of
the models, the relative probability for the most likely model is
low (,0.5).Models were averaged, weighted by the relative
probabilities, and the estimated gradient (G ), proportional change
(exp(G )),and 95% confidence (credibility) interval (CI) were
estimated. For macroalgal cover (expected to increase along the
gradient),the probability (P) that proportional change was .1 is
shown; for other responses, the probabilities of decline are
shown.For each response, the predicted values weighted by the
relative probabilities and averaged across models are shown in Fig.
3.
FIG. 5. Redundancy analysis biplot of oc-tocoral coral
communities, showing the depen-dence on chlorophyll, relative
distance acrossthe continental shelf (across) and region(gray,
Princess Charlotte Bay; black, wet trop-ics). The first two
components shown in thebiplot accounted for 78.6% of the explained
var-iation in octocoral abundances. Each symbolrepresents a reef,
and the fill of the symbol rep-resents the generic richness in
octocorals of thatreef. Chlorophyll and distance of the reefsacross
the shelf are represented as vectors andpoint in the directions of
highest chlorophyllloads and offshore reefs, respectively.
ship between octocoral richness and water clarity (Fa-bricius
and Death 2001).
The response is specific for the cause, thus an as-sociation
should be stronger if there are few ratherthan many causal
factors.Some, but not all of theresponses were specific to water
quality. In particular,total hard coral cover is a nonspecific
response, as itmay be low after a bleaching or crown-of-thorns
dis-turbance, or it may be high even in chronically
adverseconditions due to the asexual spread of a few
resistantspecies. Macroalgal growth increases with nutrients,and
unlike corals is not known to be affected by periodsof high
temperatures. Macroalgal cover remains lowdespite high nutrient
levels at high grazing pressure(McCook 1999), however macroalgal
cover is rarelyhigh in low-nutrient environments, and
macroalgaehave established dominance on reefs in some areas of
eutrophication (Smith et al. 1981). Octocoral cover ismore
specific than hard coral cover, as octocorals arerarely eaten by
the crown-of-thorns seastar (Death andMoran 1998), or indeed any
other major group of pred-ators, and azooxanthellate octocorals are
unaffected bycoral bleaching. A few octocorals of the families
Al-cyoniidae, Clavulariidae, and Briareidae may increasein cover in
areas of high nutrient loads, if other physicalenvironmental
conditions such as currents and light arefavorable. However,
bleaching-susceptible zooxanthel-late genera tend to be more
abundant than azooxan-thellate genera, and thus octocoral cover may
be re-duced by bleaching. The decrease in octocoral richnessappears
to be specific to water quality. This conclusionis supported by the
rarity of species with low toleranceof poor water quality, the
abundance of some moretolerant taxa (e.g., Briareum sp.) and the
lack of re-
-
1460 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
TABLE 3. ANOVA of community composition of octocoral genera
showing the effects of (a)relative distance across the shelf, the
chlorophyll cross-shelf gradient and regions (WT andPC), and (b)
the inshore water quality gradient and regions.
Effect df SS MS SS (%) Pseudo-F 95% CI Pa) Cross-shelf
AcrossChlorophyllRegionResiduals
111
51
53.328.117.5
285.5
53.328.117.5
5.6
13.97.34.5
74.3
9.525.023.13
6.77, 12.823.47, 7.441.24, 4.92
.0.99
.0.990.99
b) InshoreWater quality indexRegionResiduals
11
10
16.9696.285
50.048
16.96.35.0
23.18.6
68.3
3.391.25
1.32, 6.660.45, 5.56
.0.990.61
Notes: For panel (a) 66 genera were analyzed, and for panel (b)
57 genera were analyzed;the sequential sums of squares were summed
over the responses to give ANOVA tables. Thepseudo-F statistic was
bootstrapped and bias-adjusted to give 95% intervals. P denotes
theprobability that the true value of pseudo-F was .1 (the expected
value of pseudo-F, giventhat a variable has no effect). For (a)
there were strong effects of relative distance across theshelf and
chlorophyll, and weaker regional differences. For (b) there was a
strong effect ofwater quality and no effect of region. Reversing
the order of inclusion of chlorophyll andregion, and water quality
and region, slightly increased the strength of the region effects
(;30%)and correspondingly reduced the chlorophyll and water
gradient effects.
FIG. 6. Redundancy analysis biplot of octocoral coral
communities at the inshore target reefs in PC (gray symbols) andWT
(black symbols). The first two components shown in the biplot
accounted for all of the explained variation in
octocoralabundances. Each symbol represents a reef, and the fill of
the symbol represents the generic richness of octocorals at
thatreef. The water quality index vector points in the direction of
highest nutrient loads.
sponse in azooxanthellate taxa in WT. Similarly, thegradients in
octocoral community structure, with com-munities being
progressively depauperate with increas-ing nutrients and sediments,
could not be explained byother major disturbances (cyclones,
bleaching, andcrown-of-thorns).
Association agrees with known biological facts.Increased supply
of limiting nutrients is known to in-crease macroalgal growth rates
in the absence of otherlimiting factors (Schaffelke and Klumpp
1998). Sedi-mentation affects recruitment and increases adult
mor-tality in hard and octocorals, thus potentially reducingcoral
cover (Riegl and Branch 1995, Fabricius et al.2003, Philipps and
Fabricius 2003). A reduction in bio-diversity is likely due to
variable tolerances of speciesto sedimentation, turbidity, and
bleaching. Octocoralrichness is known to decline by one genus per
meterreduction in visibility on the Great Barrier Reef in
areaswhere visibility is ,10 m (Fabricius and Death 2001),a finding
that matches the decline in richness along thewater quality and
chlorophyll gradients in this study.
DISCUSSION
Identifying anthropogenic causes of ecologicalchange and
distinguishing such change from naturaldynamics is a nontrivial
task. However causal attri-bution profoundly enhances the ability
of scientists tocontribute to environmental management, and
increas-es the effectiveness of management action. For ex-ample, in
coral reef ecology the cause(s) of outbreaksof the crown-of-thorns
seastar Acanthaster planci arestill being debated by scientists 40
years after the firstobservations of outbreaks. Costly local
eradication pro-grams are now in place to protect some tourism
sites;funding that might have been spent on preventativemeasures if
causes had been identified with a reason-able level of certainty. A
contrasting example is thathigh sea surface temperatures are now
accepted bymost scientists as the major cause of coral
bleaching(Strong et al. 1997), and agreement on the cause
ofpredicted massive ecosystem changes by coral bleach-ing is adding
momentum to the call for political action
-
October 2004 1461IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
TABLE 4. Synthesis matrix defined by causality criteria and
ecological attributes, summarizingecological evidence of
terrestrial runoff effects on reefs within the wet tropical region
of theGreat Barrier Reef (Fig. 1).
CriteriaBenthic cover
HC OC MARichness
OC Z OC ACommunitiesOC C OC I
A1) Biological gradient: cross-shelfB1) Effect size:
cross-shelfA2) Biological gradient: inshoreB2) Effect size:
inshore
2222
1122
1111
1111
2220
1111
1111
C) Logical time sequenceD) Consistency with other studiesE)
Specificity of responseF) Agreement with biological facts
1021
NA101
NA111
NA111
NA111
NANANA1
NANANA1
Score 23/8 2/7 7/7 7/7 0/8 5/5 5/5Notes: Data from this region
are compared with those from reefs at low risk from runoff in
the Princess Charlotte Bay. Criteria A and B address responses
observed in this study (Figs.46, Tables 23). Criteria CF address
the agreement between responses in this study andresults from other
runoff-exposed regions and from laboratory studies on pollution
effects.Symbols in the matrix indicate the following: agreement
with the criterion (1), inconclusive(0, inconsistent or weak),
disagreement (2), and not addressed due to the lack of data
(NA).Abbreviations: HC, hard coral; OC, octocorals; Z,
zooxanthellate; A, azooxanthellate; C, cross-shelf; I, inshore. The
final row is the sum of scores as the proportion of all scores for
eachattribute, scoring 1 as 11, 2 as 21, and ignoring NA.
to combat greenhouse gas emission. While enormousinternational
commitments are required to halt or re-verse global climate change,
in our case study the re-duction of soil and fertilizer loss
through terrestrialrunoff is achievable through integrated river
basin man-agement and long-term government support on
regionalscales (Brodie 2003).
The use of a framework based on the estimation ofeffect sizes to
determine change, an established set ofcriteria to attribute
causality, and a rigorous choice ofecological attributes with
contrasting environmental re-sponses, represents a shift in
philosophy and practicefrom some of the existing ecological
techniques. Forsimple ecological analyses, methodologies such as
con-trol-impact studies and ecological indicators can ef-fectively
detect change, and meta-analysis of multiplestudies can help relate
change to potential causes on abroader scale (Schmitt and Osenberg
1996). However,for large or complex ecosystems, existing methods
areoften insufficient to detect change and attribute cau-sality
(Stewart-Oaten 1996a, b). Many forms of evi-dence of varying
strengths are to be considered, e.g.,field studies with notoriously
imperfect controls andlaboratory experiments that oversimplify
natural sys-tems. It is therefore not surprising that simple
hypoth-esis tests are unable to resolve such complex
questions.After all, it took decades of extensive and
expensiveresearch for epidemiologists to assemble sufficient
ev-idence to identify cigarette smoking as a cause of lungcancera
link that is obvious in hindsight. Hills meth-od of causal
attribution has been recommended by theEnvironmental Protection
Agency of the United States(U.S. Environmental Protection Agency
1998), but ap-plications, both on single-factor single-cause
analyses,or on ecosystem analyses, have so far rarely been
at-tempted in the ecological literature.
In our study, we found the inshore reefs in the WTregion to be
in a disturbed state. The use of Hills causalcriteria helped us to
synthesize and assess a body ofdata that quantified the ecological
attributes and theenvironmental conditions in the two regions, and
alongwater quality gradients within the regions. The twobiological
gradients, the large effect sizes, the consis-tency across
populations and with other studies, andthe agreement with
biological facts, all contributed toenhancing the belief that water
quality had an adverseeffect on aspects of the ecological state of
these reefs.These data indicated the existence of causal links
be-tween high nutrient loads and high macroalgal coveras well as
low octocoral biodiversity. Hard coral coveralso differed more than
two-fold between the two re-gions. Since the two regions contrasted
not only inwater quality but also in exposure to other
disturbances(coral bleaching and coral predation by crown-of-thorns
seastar), these strong regional differences couldnot be attributed
exclusively to water quality, althoughresponses to these multiple
disturbances including wa-ter quality may have been cumulative or
synergistic.We found the matrix to be helpful in synthesizing
thedata, attributing causality to some attributes while ex-cluding
others, and communicating the results to otherscientists and the
public.
The scientific understanding of coastal marine sys-tems in the
tropics is far less developed than in tem-perate systems such as
Chesapeake Bay or the Northand Baltic Seas, where considerable
effort has beeninvested into understanding the effects of nutrient
en-richment and toxic chemicals. In our case study, theavailable
data are comparatively sparse. Due to the highbiological diversity
of the Great Barrier Reef, we fo-cused on responses of relatively
broad ecologicalgroups rather than of individual key species. It
became
-
1462 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
apparent that community structure responded more sen-sitively to
water quality than summary parameters suchas benthic cover,
indicating that substantial informationis contained in the
responses of individual species thatwill require follow-up
laboratory exposure-effects ex-periments. It also became apparent
that there was littlespecies replacement in octocorals from clean
to turbidwater, rather species increasingly dropped out towardshigh
nutrient conditions, and only two species appearedto grow well in
the waters with the least clean water.Although octocorals do not
contribute to the formationof coral reefs, they are the second-most
abundant groupof macrobenthos in coral reefs. The disappearance ofa
large number of octocoral genera in areas of highloads of nutrients
and suspended particles may be takenas an indicator for severe
deterioration of environmen-tal conditions, with potential
consequences for otherorganism groups that were not included in the
study.
What remained unresolved in our study is the originof the high
nutrient levels in the water around the in-shore target reefs, and
thus the direct link betweenterrestrial runoff and nutrient
enrichment. It may beargued that nutrient, chlorophyll, and
sediment levelsin the marine inshore region of the WT may have
al-ways been high. Rivers discharging from catchmentswith heavily
fertilized sugar crops into the wet tropicalmarine system carry
levels of nutrients, sediments andpesticides that are several times
higher than estimatedlevels in pre-European times and than rivers
discharg-ing into Princess Charlotte Bay (Furnas 2003).
Therelatively high levels of nitrate measured around theWT reefs
(Table 1) indicate fresh supply of this nutri-ent, as
concentrations of dissolved inorganic nitrogenare rapidly taken up
by nutrient-limited algae and bac-teria in tropical marine systems.
This supply is likelyto have come from the rivers, which are the
largestsource of new nutrients in the GBR lagoon away fromupwelling
areas (Furnas et al. 1997, Mitchell et al.2001). Combined with
flood-plume dispersal models(Devlin et al. 2002), the limited
existing data thereforesuggest that modern agrochemicals do reach
the targetinshore reefs of the WT where they may contribute tothe
measurably higher concentrations of water nutri-ents. Similarly, it
may be argued that coral cover andbiodiversity on the WT reefs may
have always beenlow. However, the widths of the reef flats around
theislands in both regions indicate positive reef accretionover
extended periods in earlier times (Hopley et al.1983). Observations
on one WT reef group (the Rus-sell-Normanby Islands) show declining
coral coversince 1990 due to a series of disturbances by
bleaching,cyclones, and a crown-of thorns seastar outbreak (Ayl-ing
and Ayling 2002), and it appears that the reefs havefailed to
recover from these disturbances. Coral coverhas also been
monotonously declining on a number ofother WT inshore reefs since
1985 to ;10% cover in2002 (AIMS Long-Term Monitoring Program,
unpub-lished data). Coral cover can increase by 10% per year
on some inshore reefs after disturbance (Ayling andAyling 2002),
and anecdotal reports indicated that coralcover on the
Russell-Normanby Islands did recover af-ter tropical cyclones in
1975 and 1977, and reachedpre-cyclone levels in 1990 after a severe
tropical cy-clone in 1986 (Devantier 1994). This suggests that
mul-tiple disturbances have reduced coral cover on WTreefs, while
exposure to terrestrial runoff appears toretarding coral recovery
after the recent disturbances.In the absence of new disturbances,
coral cover mayrecover within one to two decades despite
continuingexposure to high levels of nutrients and
sediments(probably through the re-establishment of
fast-growingMontipora and Acropora), whereas the processes andtime
frames leading to restoration of coral richness andpresence of more
sensitive or slower-growing speciesare unknown. Importantly,
although changes betweenpast and present conditions both in water
quality andin the ecological properties of the WT inshore reefscan
not be established with certainty, the ecologicalgradients along
the water quality gradient (as demon-strated in this study) can
serve to predict future changesin macroalgal cover and octocoral
biodiversity, shouldthe water quality deteriorate or improve as
land usepractices change.
To conclude, using the matrix in combination withmodern
statistical analyses, we have aimed at over-coming the problems
that often prevent the early de-tection of ecological change and
its causes. We havesynthesized multiple and complex sources of
infor-mation to avoid any misleading selective use of sci-entific
data, and used modern statistical methods toovercome abuse of
technical concepts such as proba-bility and causality. As
ecologists, we need to synthe-size multiple and complex sources of
information,weigh the evidence, quantify effect sizes, and
predictthe ecological consequences and socioeconomic costsof
alternative actions. The matrix framework can helpdoing all this.
By applying stringent, scientifically ro-bust, yet simple criteria
to well-chosen properties ofthe ecosystem, we can greatly improve
our capacity todetermine change and attribute causality. This
shouldlead to higher performance benchmarks for complexecological
studies, and a greater scientific contributionto sustainable
ecosystem management. It is then up toa better-informed society to
decide how much ecolog-ical change is acceptable.
ACKNOWLEDGMENTSWe gratefully acknowledge the contributions of
Jon Brodie
for providing the long-term chlorophyll monitoring data ofthe
Great Barrier Reef Marine Park Authority, and of MichelleSkuza,
Jane WuWon, and Margaret Wright for analyzing thewater quality
samples. We thank the Subject Editor Paul Day-ton, Simon Thrush,
and an anonymous reviewer for very help-ful reviews and comments.
Thanks also to Miles Furnas, JonBrodie, David Williams, Britta
Schaffelke, Lyndon Devantier,Janice Lough, and Malcom McCulloch for
comments on ear-lier versions of the manuscript. A number of
volunteers andthe crews of the AIMS Research Vessels were a great
help
-
October 2004 1463IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
in the field studies. The research was funded by the
Coop-erative Research Centre for the Great Barrier Reef
WorldHeritage Area (CRC Reef), and the Australian Institute
ofMarine Science (AIMS).
LITERATURE CITEDAyling, A. M., and A. L. Ayling. 2002. The
dynamics of
Cairns section fringing reefs: 2002 final report. Great Bar-rier
Reef Marine Park Authority, Townsville, Australia.
Bell, P. R. F. 1991. Status of eutrophication in the
GreatBarrier Reef Lagoon. Marine Pollution Bulletin 22:8993.
Berger, J. O., B. Boukai, and Y. Wang. 1997. Unified
fre-quentist and Bayesian testing of a precise hypothesis
(withdiscussion). Statistical Science 12:133160.
Berger, J. O., and L. Pericchi. 2001. Objective Bayesianmethods
for model selection: introduction and comparison(with discussion).
Pages 135207 in P. Lahiri, editor. Modelselection. Institute of
Mathematical Statistics LectureNotesMonograph Series. Institute of
Mathematical Sta-tistics, Beachwood, Ohio, USA.
Berger, J. O., and T. Sellke. 1987. Testing a point null
hy-pothesis: the irreconcilability of p-values and evidence.Journal
of the American Statistical Association 82:112139.
Berkelmans, R., and J. K. Oliver. 1999. Large-scale bleachingof
corals on the Great Barrier Reef. Coral Reefs 18:5560.
Bodansky, D. 1991. Law: scientific uncertainty and the
pre-cautionary principle. Environment 33:4344.
Bozdogan, H. 1987. Model selection and Akaikes infor-mation
criterion (AIC): the general theory and its analyticalextensions.
Psychometrika 52:345370.
Brodie, J. E. 2003. Keeping the wolf from the door:
managingland-based threats to the Great Barrier Reef. Pages 705714
in M. K. Kasim Moosa, S. Soemodihardjo, A. Nontji,A. Soegiarto, K.
Romimohtarto, Sukarno, and Suharsono,editors. Proceedings of the
Ninth International Coral ReefSymposium. Indonesian Institute of
Sciences in coopera-tion with the State Ministry for Environment,
Bali, Indo-nesia.
Brodie, J. E., C. Christie, M. Devlin, D. Haynes, S. Morris,M.
Ramsay, J. Waterhouse, and H. Yorkston. 2001. Catch-ment management
and the Great Barrier Reef. Water Sci-ence and Technology
43:203211.
Bryant, D., L. Burke, J. McManus, and M. Spalding. 1998.Reefs at
risk. A map-based indicator of threats to theworlds coral reefs.
World Resources Institute, Cambridge,UK.
Burnham, K. P., and D. R. Anderson. 1998. Model selectionand
inference: a practical information theoretic
approach.Springer-Verlag, New York, New York, USA.
Burnham, K. P., D. R. Anderson, and W. L. Thompson. 2000.Null
hypothesis testing: problems, prevalence, and an al-ternative.
Journal of Wildlife Management 64:912923.
Cohen, J. 1988. Statistical power analysis for the
behaviouralsciences. Erlbaum Publishing, Hillsdale, New York,
USA.
Cohen, J. 1994. The earth is round (p , .05).
AmericanPsychologist 49:9971003.
Davison, A. C., and D. V. Hinkley. 1997. Bootstrap methodsand
their application. Cambridge University Press, Cam-bridge, UK.
Dayton, P. K. 1998. Reversal of the burden of proof in
fish-eries management. Science 279:821822.
Death, G., and P. Moran. 1998. Factors affecting behavioursof
crown-of-thorns seastar (Acanthaster planci). 2:
Feedingpreferences. Journal of Experimental Marine Biology
andEcology 220:107126.
Devantier, L. M. 1994. The structure of assemblages of mas-sive
corals in the central Great Barrier Reef and assessmentof the
effects of predation by the crown-of-thorns seastarAcanthaster
planci. Dissertation. Queensland University,Brisbane,
Australia.
Devlin, M., J. Brodie, J. Waterhouse, A. Mitchell, D. Audas,and
D. Haynes. 2002. Exposure of Great Barrier Reef in-ner-shelf reefs
to river-borne contaminants. Proceedings ofthe Second National
Conference on Aquatic Environments:Sustaining Our Aquatic
EnvironmentsImplementing So-lutions, 2023 November 2001,
Townsville, Australia.
Devlin, M., J. Waterhouse, J. Taylor, and J. Brodie. 2001.Flood
plumes in the Great Barrier Reef: spatial and tem-poral patterns in
composition and distribution. Great Bar-rier Reef Marine Park
Authority, Townsville, Australia.
Diaz-Pulido, G., and L. J. McCook. 2003. Relative roles
ofherbivory and nutrients in the recruitment of coral-reef
sea-weeds. Ecology 84:20262033.
Dinesen, Z. D. 1983. Patterns in the distribution of soft
coralsacross the Central Great Barrier Reef. Coral Reefs
1:229236.
Draper, D. 1995. Assessment and propagation of model
un-certainty (with discussion). Journal of the Royal SocietySeries
B 57:4597.
Dubinsky, Z., and N. Stambler. 1996. Marine pollution andcoral
reefs. Global Change Biology 2:511526.
Edinger, E. N., J. Jompa, G. V. Limmon, W. Widjatmoko, andM.
Risk. 1998. Reef degradation and coral biodiversity inIndonesia:
effects of land-based pollution, destructive fish-ing practices and
changes over time. Marine Pollution Bul-letin 36:617630.
Edinger, E. N., G. V. Limmon, J. Jompa, W. Widjatmoko, J.M.
Heikoop, and M. Risk. 2000. Normal coral growth rateson dying
reefs: are coral growth rates good indicators ofreef health? Marine
Pollution Bulletin 40:404425.
Efron, B., and R. J. Tibshirani. 1993. An introduction to
thebootstrap. Chapman and Hall, San Francisco, California,USA.
Ellison, A. M. 1996. An introduction to Bayesian inferencefor
ecological research and environmental decision-mak-ing. Ecological
Applications 6:10361046.
Fabricius, K. E., and P. Alderslade. 2001. Soft corals and
seafans: a comprehensive guide to the tropical shallow watergenera
of the central-west Pacific, the Indian Ocean andthe Red Sea.
Australian Institute of Marine Science, Towns-ville, Australia.
Fabricius, K. E., and G. Death. 2001. Biodiversity on theGreat
Barrier Reef: large-scale patterns and turbidity-re-lated local
loss of soft coral taxa. Pages 127144 in E.Wolanski, editor.
Oceanographic processes of coral reefs:physical and biological
links in the Great Barrier Reef.CRC Press, London, UK.
Fabricius, K. E., and D. W. Klumpp. 1995. Widespread mix-otrophy
in reef-inhabiting soft corals: the influence ofdepth, and colony
expansion and contraction on photosyn-thesis. Marine Ecology
Progress Series 125:195204.
Fabricius, K. E., C. Wild, E. Wolanski, and D. Abele.
2003.Effects of transparent exopolymer particles (TEP) and mud-dy
terrigenous sediments on the survival of hard coral re-cruits.
Estuarine, Coastal and Shelf Science 57:613621.
Fox, G. A. 1991. Practical causal inference for
ecoepide-miologists. Journal of Toxicology and EnvironmentalHealth
33:359368.
Freedman, D. A. 1983. A note on screening regression equa-tions.
American Statistician 37:152155.
Furnas, M. J. 2003. Catchments and corals: terrestrial runoffto
the Great Barrier Reef. Australian Institute of MarineScience and
CRC Reef, Townsville, Australia.
Furnas, M. J., and A. W. Mitchell. 1996. Nutrient inputs intothe
central Great Barrier Reef (Australia) from subsurfaceintrusions of
Coral Sea waters: a two-dimensional displace-ment model.
Continental Shelf Research 16:11271148.
Furnas, M., and A. Mitchell. 2001. Runoff of terrestrial
sed-iment and nutrients into the Great Barrier Reef World Her-itage
Area. Pages 3751 in E. Wolanski, editor. Oceano-
-
1464 KATHARINA E. FABRICIUS AND GLENN DEATH Ecological
ApplicationsVol. 14, No. 5
graphic processes of coral reefs: physical and biologicallinks
in the Great Barrier Reef. CRC Press, Boca Raton,Florida, USA.
Furnas, M., A. Mitchell, and M. Skuza. 1997. Shelf-scalenitrogen
and phosphorus budgets for the central Great Bar-rier Reef. Pages
809814 in H. A. Lessios and I. G. Ma-cintyre, editors. Proceedings
of the Eighth InternationalCoral Reef Symposium. Smithsonian
Tropical Research In-stitute, Balboa, Republic of Panama.
Gillies, D. 2001. Critical choice. British Journal of
Philos-ophy and Science 52:613622.
Hastie, T. J., and R. J. Tibshirani. 1990. Generalized
additivemodels. Chapman and Hall, London, UK.
Hastie, T. J., R. J. Tibshirani, and J. H. Friedman. 2001.
Theelements of statistical learning. Springer-Verlag, New York,New
York, USA.
Haynes, D., and K. Michalek-Wagner. 2000. Water qualityin the
Great Barrier World Heritage Area: past perspectives,current issues
and new research directions. Marine Pollu-tion Bulletin
41:428434.
Hill, A. B. 1965. The environment and disease: associationand
causation. Proceedings of the Royal Society of Med-icine
58:295300.
Hodgson, G. E., and P. M. Yau. 1997. Physical and
biologicalcontrols of coral communities in Hong Kong. Pages 459464
in H. A. Lessios and I. G. Macintyre, editors. Pro-ceedings of the
Eighth International Coral Reef Sympo-sium. Smithsonian Tropical
Research Institute, Balboa, Re-public of Panama.
Hoenig, J. M., and D. M. Heisey. 2001. The abuse of power:the
pervasive fallacy of power calculations for data anal-ysis.
American Statistician 55(1):1924.
Hopley, D., A. M. Slocombe, F. Muir, and C. Grant.
1983.Nearshore finging reefs in North Queensland. Coral
Reefs1:151160.
Hunter, C. L., and C. W. Evans. 1995. Coral reefs in KaneoheBay,
Hawaii: two centuries of western influence and twodecades of data.
Bulletin of Marine Science 57:501515.
Johnson, D. H. 1999. The insignificance of hypothesis test-ing.
Journal of Wildlife Management 63:763772.
Jones, R. J., J. Muller, D. Haynes, and U. Schreiber.
2003.Effects of herbicides diuron and atrazine on corals of
theGreat Barrier Reef, Australia. Marine Ecological ProgressSeries
251:153167.
Jongman, R. H. G., C. F. J. ter Braak, and O. F. R.
Tongeren.1995. Data analysis in community and landscape
ecology.Second edition. Cambridge University Press,
Cambridge,UK.
Larcombe, P., and K. Woolfe. 1999. Increased sediment sup-ply to
the Great Barrier Reef will not increase sedimentaccumulation at
most coral reefs. Coral Reefs 18:163169.
Madigan, D., and A. E. Raftery. 1994. Model selection
andaccounting for model uncertainty in graphical models usingOccams
window. Journal of the American Statistical As-sociation
89:15351546.
McCook, L. J. 1999. Macroalgae, nutrients and phase shiftson
coral reefs: scientific issues and management conse-quences for the
Great Barrier Reef. Coral Reefs 18:357367.
McCulloch, M., S. Fallon, T. Wyndham, E. Hendy, J. Lough,and D.
Barnes. 2003. Coral record of increased sedimentflux to the inner
Great Barrier Reef since European settle-ment. Nature
421:727730.
McCullagh, P., and J. A. Nelder. 1989. Generalized linearmodels.
Second edition. Chapman and Hall, London, UK.
Mitchell, A. J., R. Reghenzani, and M. Furnas. 2001. Nitro-gen
levels in the Tully Rivera long-term view. WaterScience and
Technology 43:99105.
Nelder, J. A. 1999. Statistics for the millennium: from
sta-tistics to statistical science. Statistician 48:257269.
Newman, M. C., and D. A. Evans. 2002. Enhancing beliefduring
causality assessments: cognitive idols or Bayesstheorem? Pages 7396
in M. C. Newman, M. H. Roberts,and R. C. Hale, editors. Coastal and
estuarine risk assess-ment. Lewis Publishers, New York, New York,
USA.
Pearl, J. 2000. Causality: models, reasoning, and
inference.Cambridge University Press, Cambridge, UK.
Philipps, E., and K. F. Fabricius. 2003.
Photophysiologicalstress in scleractinian corals in response to
short-term sed-imentation. Journal of Experimental Marine Biology
andEcology 287:5778.
Raftery, A. E. 1988. Approximate Bayes factors for gener-alized
linear models. Technical Report no. 121, Departmentof Statistics,
University of Washington, Seattle, Washing-ton, USA.
Raftery, A. E. 1995. Bayesian model selection in social
re-search (with Discussion). Pages 111196 in P. V. Marsden,editor.
Sociological methodology. Blackwells, Cambridge,Massachusetts,
USA.
Rao, C. D. 1964. The use and interpretation of
principalcomponents analysis in applied research. Sankhya A
26:329358.
Riegl, B., and G. M. Branch. 1995. Effects of sediment onthe
energy budgets of four scleractinian (Bourne 1900) andfive
alcyonacean (Lamouroux 1816) corals. Journal of Ex-perimental
Marine Biology and Ecology 186:259275.
Rogers, C. 1990. Responses of coral reefs and reef organismsto
sedimentation. Marine Ecology Progress Series 62:185202.
Roth, L. H., B. Selwyn, A. Holguin, and B. L. Christensen.1982.
Principles of epidemiology. Academic Press, NewYork, New York,
USA.
Sakamoto, Y., M. Ishiguro, and G. Kitagawa. 1986.
Akaikeinformation criterion statistics. D. Reidel Publishing
Com-pany, KTC Scientific Publishers, Tokyo, Japan.
Schaffelke, B., and D. W. Klumpp. 1998. Short-term
nutrientpulses enhance growth and photosynthesis of the coral
reefmacroalga Sargassum bacularia. Marine Ecology ProgressSeries
170:95105.
Schmitt, R. J., and C. W. Osenberg. 1996. Detecting ecolog-ical
impacts. Academic Press, San Diego, California, USA.
Schroeter, S. C., J. D. Dixon, J. Kastendiek, R. O. Smith, andJ.
R. Bence. 1993. Detecting the ecological effects of en-vironmental
impacts: a case study of kelp forest inverte-brates. Ecological
Applications 3:331350.
Schwarz, G. 1978. Estimating the dimension of a model.Annals of
Statistics 6:461464.
Smith, S. V., W. Kimmerer, E. Laws, R. Brock, and T. Walsh.1981.
Kaneohe Bay sewage diversion experiment: per-spectives on ecosystem
responses to nutritional perturba-tion. Pacific Science
35:270395.
Speed, T. 1990. Complexity, calibration and causality in
in-fluence diagrams. Pages 361384 in R. M. Oliver and J.Q. Smith,
editors. Influence diagrams, belief nets and de-cision analysis.
Wiley, Chichester, UK.
Statistical Sciences. 1999. S-PLUS, Version 2000 for Win-dows.
Mathsoft, Inc, Seattle, Washington, USA.
Stewart-Oaten, A. 1996a. Goals in environmental monitor-ing.
Pages 1728 in R. J. Schmitt and C. W. Osenberg,editors. Detecting
ecological impacts. Academic Press, SanDiego, California, USA.
Stewart-Oaten, A. 1996b. Problems in the analysis of
envi-ronmental monitoring data. Pages 109132 in R. J. Schmittand C.
W. Osenberg, editors. Detecting ecological impacts.Academic Press,
San Diego, California, USA.
Strong, A. E., C. B. Barrientos, C. Duda, and J. Sapper.
1997.Improved satellite techniques for monitoring coral
reefbleaching. Pages 14951498 in H. A. Lessios and I. G.Macintyre,
editors. Proceedings of the Eighth International
-
October 2004 1465IDENTIFYING ECOLOGICAL CHANGE AND CAUSE
Coral Reef Symposium. Smithsonian Tropical Research In-stitute,
Balboa, Republic of Panama.
Szmant, A. M. 2002. Nutrient enrichment on coral reefs: isit a
major cause of coral reef decline? Estuaries 25:753766.
ter Braak, C. J. F. 1992. Permutation versus bootstrap
sig-nificance tests in multiple regression and ANOVA. Pages7986 in
K. H. Jockel, G. Rothe, and W. Sendler, editors.Bootstrapping and
related techniques. Springer-Verlag,Berlin, Germany.
Tomascik, T., T. Suharsono, and A. Mah. 1993. Case histo-ries: a
historical perspective of the natural and anthropo-genic impacts in
the Indonesian archipelago with a focuson the Kepulauan Seribu,
Java Sea. Pages 2632 in R. N.Ginsburg, editor. Global aspects of
coral reefs. Universityof Miami Rosenstiel School of Marine and
AtmosphericSciences, Miami, Florida, USA.
Torgersen, T., A. R. Chivas, and A. Chapman. 1983. Chem-ical and
isotopic characterization and sedimentation ratesin Princess
Charlotte Bay, Queensland. Journal of Austra-lian Geology and
Geophysics 8:191200.
U.S. Department of Health, Education, and Welfare. 1964.Smoking
and health: report of the advisory committee tothe Surgeon General
of the Public Health Service. PublicHealth Service Publication. No.
1103. U.S. Department ofHealth, Education, and Welfare, Washington,
D.C., USA.
U.S. Environmental Protection Agency. 1998. Guidelines
forecological risk assessment. Federal Register 63(93):2684626
924.
Van Woesik, R., Y. Tomascik, and S. Blake. 1999.
Coralassemblages and physico-chemical characteristics of
theWhitsunday Islands: evidence of recent community chang-es.
Marine and Freshwater Research 50:427440.
West, K., and R. van Woesik. 2001. Spatial and temporalvariation
of river discharge in Okinawa (Japan): inferringthe temporal impact
on adjacent coral reefs. Marine Pol-lution Bulletin 42:864872.
Wittenb