Ecography 36: 001–009, 2013

doi: 10.1111/j.1600-0587.2013.00291.x
© 2013 The Authors. Ecography © 2013 Nordic Society Oikos
Subject Editor: Thierry Oberdorff. Accepted 21 March 2013

Global patterns and predictors of tropical reef fish

species richness

V. Parravicini, M. Kulbicki, D. R. Bellwood, A. M. Friedlander, J. E. Arias-Gonzalez,

P. Chabanet, S. R. Floeter, R. Myers, L. Vigliola, S. D’Agata and D. Mouillot­
V. Parravicini, ([Link]@[Link]) and M. Kulbicki, IRD, Inst. de Recherche pour le Développement, UR 227 CoReUs, LABEX
Corail, Laboratoire Arago, BP 44, FR-66651 Banyuls/mer, France. VP also at: CESAB-FRB, Immeuble Henri Poincaré, Domaine du Petit
Arbois, FR-13857 Aix-en-Provence cedex 3, France. – D. R. Bellwood, School of Marine and Tropical Biology, James Cook Univ., Townsville,
QLD 4811, Australia. DRB also at: Australian Research Council, Center of Excellence for Coral Reef Studies, James Cook Univ., Townsville,
QLD 4811, Australia. – A. M. Friedlander, Hawaii Cooperative Fishery Research Unit and Univ. of Hawaii at Manoa, Honolulu, HI 096822,
USA. – J. E. Arias-Gonzalez, Laboratorio de Ecología de Ecosistemas de Arrecifes Coralinos, Depto de Recursos del Mar, Centro de Investigación
y de Estudios Avanzados del Inst. Politécnico Nacional, Unidad Mérida, Antigua Carretera a Progreso Km. 73 Cordemex, Mérida, Yucatán,
97310 Mexico. – P. Chabanet, IRD, Inst. de Recherche pour le Développement, UR 227 CoReUs, LABEX Corail, BP 50172, CS 41095,
FR-97495 Ste Clotilde cedex, Reunion Island, France. – S. R. Floeter, Depto de Ecologia e Zoologia, Univ. Federal de Santa Catarina, Flori-
anópolis, Santa Catarina 88010-970, Brazil. – R. Myers, Seaclicks – Coral Graphics, 9273 Silent Oak Circle, Wellington, FL 33411, USA. – L.
Vigliola and S. D’Agata, IRD, Inst. de Recherche pour le Développement, UR 227 CoReUs, LABEX Corail, BP A5 – 98848 Nouméa Cedex, New
Caledonia. SD also at: Laboratoire Ecologie des Systèmes Marins Côtiers UMR 5119, CNRS, IRD, IFREMER, UM2, UM1, cc 093, Place E.
Bataillon, FR-34095 Montpellier Cedex 5, France. – D. Mouillot, Australian Research Council, Center of Excellence for Coral Reef Studies,
James Cook Univ., Townsville, QLD 4811, Australia. DM also at: Laboratoire Ecologie des Systèmes Marins Côtiers UMR 5119, CNRS, IRD,
IFREMER, UM2, UM1, cc 093, Place E. Bataillon, FR-34095 Montpellier Cedex 5, France.­

In the marine realm, the tropics host an extraordinary diversity of taxa but the drivers underlying the global distribution
of marine organisms are still under scrutiny and we still lack an accurate global predictive model. Using a spatial data-
base for 6336 tropical reef fishes, we attempted to predict species richness according to geometric, biogeographical and
environmental explanatory variables. In particular, we aimed to evaluate and disentangle the predictive performances of
temperature, habitat area, connectivity, mid-domain effect and biogeographical region on reef fish species richness. We
used boosted regression trees, a flexible machine-learning technique, to build our predictive model and structural equation
modeling to test for potential ‘mediation effects’ among predictors. Our model proved to be accurate, explaining 80% of
the total deviance in fish richness using a cross-validated procedure. Coral reef area and biogeographical region were the
primary predictors of reef fish species richness, followed by coast length, connectivity, mid-domain effect and sea surface
temperature, with interactions between the region and other predictors. Important indirect effects of water temperature on
reef fish richness, mediated by coral reef area, were also identified. The relationship between environmental predictors and
species richness varied markedly among biogeographical regions. Our analysis revealed that a few easily accessible variables
can accurately predict reef fish species richness. They also highlight concerns regarding ongoing environmental declines,
with region-specific responses to variation in environmental conditions predicting a variable response to anthropogenic
impacts.

Understanding the mechanisms generating and maintaining vs. maintaining variation in species richness remains chal-
the global distribution of species richness is one of the major lenging, it is possible to identify, contrast and disentangle
goals of biogeography and ecology (Gaston 2000), but we those factors that best explain current-day large-scale pat-
still lack conclusive evidence to separate processes from terns (Hawkins et al. 2003, Jetz and Fine 2012).
patterns (Hawkins et al. 2003). Although mechanisms are In the marine realm, the tropics host an extraordinary
not well understood, strong relationships between species diversity of taxa upon which ecosystem goods and services
richness and environmental or geographical variables have depend (Worm et al. 2006). Tropical coastal zones, how-
been identified at large scales. Indeed, in the marine realm, ever, are presently in severe decline (Bellwood et al. 2004,
recurrent predictors of species richness are consistently sea Knowlton and Jackson 2008). Human pressures and global
surface temperature, habitat area, and the mid-domain environmental changes are inducing local biodiversity ero-
effect (Bellwood et al. 2005, Tittensor et al. 2010). This sion which may imperil ecosystem functioning (Bellwood
suggests that, even if understanding mechanisms producing et al. 2012). Identifying and ranking environmental or

Early View (EV): 1-EV

geographical variables maintaining species richness is thus species richness declines nearly uniformly with increasing
critical to addressing the worldwide biodiversity crisis. distance from the IAA, which is close to the mid-domain
Here we built a predictive model of reef fish species rich- of the region (Bellwood et al. 2012). In contrast, the bio­
ness at a global scale and we disentangled the relative con- diversity hotspot in the Atlantic Ocean (i.e. the Caribbean) is
tribution of environmental (i.e. temperature, habitat area, distant from the mid-domain and comparatively less
connectivity) and geographical variables (i.e. mid-domain connected to surrounding areas because of several bio­
effect, biogeographical region) in explaining observed global geographical barriers (Luiz et al. 2012).
diversity patterns. Previous works on reef fishes focused on the Contrary to the pattern observed in the tropical Indo-
Indo-Pacific (Bellwood et al. 2005) or Tropical Eastern Pacific Pacific (Bellwood et al. 2005), at a broader latitudinal scale in
(Mora and Robertson 2005). These studies identified habitat the Atlantic Ocean, Macpherson (2002) found a major role
area (Bellwood and Hughes 2001), the mid-domain effect of temperature and a negligible effect of area in explaining
(Connolly et al. 2003), connectivity to the Indo-Australian the distribution of species richness for several marine taxa,
Archipelago (IAA; Mora et al. 2003), and a combination of including fishes. Complex interactions between environ­
mid-domain effect and habitat area (Bellwood et al. 2005) mental, geographic and historical variables may thus blur
as the main predictors of reef fish richness. The mid-domain the patterns and weaken the predictive power of models at
effect considers the spatial variation of species richness a global scale.
due to bounding constraints (Colwell and Hurtt 1994). To overcome these constraints and to provide a robust
If the geographical ranges of species are randomly placed predictive model of fish species richness in the tropics, we
within a bounded domain, the resultant richness gradient accounted for the biogeographical region in our analysis.
will form a peak in the middle of the domain (i.e. the mid- Our goal is to test whether an accurate predictive model can
domain). Following Stehli, Wells, Potts and Rosen (reviewed be proposed for reef fishes at global scale based on geometric,
by Bellwood et al. 2012) we propose a connectivity hypo­­ biogeographical and environmental factors and accounting
thesis that postulates that areas with low geographical dis- for their interactions. After construction of the predictive
tances between habitat patches will sustain comparatively model, we then used structural equation models to further
higher species richness as it allows for potentially more explain the complex pathways between explanatory factors
interconnected meta-communities. At smaller scales, this and fish richness.
hypothesis is supported by experimental evidences docu-
menting higher fish species richness along continuous reefs
than in isolated reef patches (Belmaker et al. 2011). The Methods
area hypothesis predicts that the greater the habitat area,
the higher the number of species that can be maintained Fish species lists
(Gray 2001). Finally, the energy hypothesis predicts that
higher temperature supports high species richness through We obtained estimates of reef fish species richness at 163
its effects on the biochemical kinetics of metabolism (Allen locations worldwide (Fig. 1). We examined almost 500 ref-
et al. 2002). erences and extracted information from published works,
Although major predictors of reef fish richness have regional checklists, monographs on specific families or gen-
been identified, a more complex scenario emerges when era, and reports (Supplementary material Appendix 1). We
attempting to disentangle their relative influence at a global restricted data collection to the oceans portion showing a
scale. Different oceanic basins have markedly different minimum monthly sea surface temperature (hereafter SST)
evolutionary histories (Floeter et al. 2008, Cowman and of 17°C. Although the present study focused on the tropics,
Bellwood 2013). The Indo-Pacific region, where most previ- whose SST limit is usually set at 20°C, we decided to use
ous research was focused, shows a unique pattern of varia- a broader definition of ‘tropical’ oceans, thereby including
tion in species richness which is markedly different from locations where species of tropical affinity are present. The
that observed in the Atlantic Ocean. In the Indo-Pacific, inclusion of those locations between 17°C and 20°C broadens

Figure 1. Geographical position of the 163 locations (black circles) for which information on the presence–absence of tropical reef-
associated fish was available. Dotted lines indicates known physical or physiological barriers considered as boundaries of the domains. Stars
represents the position of the mid-domains. EPB: Eastern Pacific Barrier; MAB: Mid-Atlantic Barrier; AOB: Amazon-Orinoco Barrier.

2-EV
the range of variables, potentially increasing the accuracy of in the same analysis. Thus, we implemented the approach
our predictive model. Overall, we obtained information on developed by Bellwood et al. (2005) calculating the distance
the distribution of 6336 reef fishes from 37 locations in the from the MD relative to the domain size (i.e. the norma-
Atlantic Ocean, 39 in the Indian Ocean, 70 in the Pacific lised distance to the MD) for each location. We operatively
Ocean, and 17 in the Tropical Eastern Pacific. defined the domains as those areas within well-known physi-
cal or physiological barriers that are rarely crossed by reef
fishes. We defined five such domains in all (Fig. 1). In the
Explanatory variables Indo-Pacific we considered the area from the western Indian
Ocean to Easter Island and Salas y Gómez as separated
Explanatory variables were selected according to several from the Tropical Eastern Pacific by the East Pacific Barrier
hypotheses explaining the variation in species richness (Robertson et al. 2004, Robertson and Cramer 2009). In the
based on previous works and general ecological theories. Atlantic Ocean, we considered two distinct domains along
In particular, we considered the following hypotheses: 1) the west coast, respectively separated by the Amazon River
energy; 2) area; 3) mid-domain (hereafter MD) effect and barrier and one domain in the east coast, east and west coasts
4) connectivity. of the Atlantic being separated by the Mid-Atlantic Barrier
Estimates for each explanatory variable were obtained for (Floeter et al. 2008, Luiz et al. 2012).
each location using two geographical definitions of location. In order to account for the connectivity hypothesis, con-
First, according to maps and descriptions in the original nectivity was defined as the relative proximity of each loca-
literature used to estimate species richness, we defined for tion to patches of reef fish habitat. Connectivity estimates
each location, the area of the shelf (sea-bottom down to 200 were calculated using a nearest neighbour approach, which
m depth) to which the species list pertained. The shelf was is typically employed in ecological research (Moilanen and
defined using SRTM30_PLUS bathymetry (Shuttle Radar Nieminen 2002). In particular, using a Behrmann projection,
Topography Mission) available at [Link] the world was divided into cells (200  200 km at the equa-
WWW_html/srtm30_plus.html. In this way, we obtained tor) from which only cells containing reef fish habitat (i.e.
polygons of different size according to the locations, which coast or coral reef ) were retained. These cells represent habi-
were later used for the extraction of explanatory variables tat patches and for each location we computed the inverse
from raster and vector spatial files. Second, we estimated of the mean distance from the location to the 10 nearest
explanatory variables within a standard radius of 600 km patches. We arbitrarily used 10 nearest neighbours instead of
around each location. The 600 km radius was chosen because 1 in order to limit the strong dependency of the variable to a
it was already successfully employed by previous research single neighbour, which may not always be highly connected
conducted in the Indo-Pacific (Bellwood et al. 2005) and to the most proximate reefs because of current flows (Mora
encompasses the probable larval dispersal envelope of reef et al. 2012). Finally, we considered the different evolution-
fishes (Swearer et al. 2002). All spatial analyses were con- ary histories characterising each biogeographical region
ducted using a global equal area Behrmann projection with (Cowman and Bellwood 2013). We therefore included a
the prime meridian set at 145° longitude. ‘region’ factor as a predictor variable. We used four bio­
For the energy hypothesis, we evaluated SST (i.e. mini- geographical regions: the Atlantic Ocean, Indian Ocean, and
mum, mean, maximum and the range of annual values of Pacific Ocean and the Tropical Eastern Pacific.
sea surface temperature). This information was obtained
from the Bio-ORACLE database at a resolution of 5 arcmin
(Tyberghein et al. 2012) available at [Link]. Data analysis
be. SST estimates were based on monthly mean values
from 2002 to 2009. In order to build a predictive model for global patterns of
For the area hypothesis we considered the total coral reef species richness for tropical reef fishes we implemented
area (km2), the shelf area (km2) and the coast length (km). boosted regression trees (hereafter BRT, Friedman 2001).
Shelf area was considered as a relevant proxy of the historical BRT differ from traditional regression methods that produce
habitat availability as it provides an approximate estimate of a single ‘best’ model in that they combine large numbers of
the coastal waters during the Pleistocene low sea level stands simple trees to optimize predictive performance (Elith et al.
(Bellwood et al. 2005). Coast length was used as an estimate 2008). This approach was preferred over linear or additive
of the present extent of coastline used by shallow water reef models for two main reasons: 1) BRT can handle complex
fish. Both coast length and shelf area were obtained using relationships and interactions among predictor variables that
SRTM30_PLUS bathymetry (Shuttle Radar Topography were expected to emerge at a global scale; 2) BRT are consid-
Mission). Estimates of coral reef area were obtained from ered as a robust technique which can handle outliers and non-
data based on the Coral Reef Millennium Census project linearity (Friedman 2001, Elith et al. 2008). Although BRT
(Andréfouët et al. 2006) and available at: [Link] were also considered relatively robust to collinearity among
[Link]/. predictor variables, it has been recently shown that removing
The potential effect of MD on species richness was also collinearity from explanatory variables showing a Pearson cor-
considered. Typically it is assessed using null-models based relation coefficient higher than |0.7| is a desirable practice also
on resampling of species with defined geographical ranges with techniques based on boosting (Dormann et al. 2012).
(Currie and Kerr 2008). However, there is a substantial lack After testing for collinearity between explanatory vari-
of agreement on the most appropriate null-model and this ables, we decided to remove maximum SST, minimum
approach does not easily account for multiple predictors SST and the extent of the continental shelf (Supplementary

3-EV
material Appendix 2). Regarding SST estimates, high col- effects (Mitchell 1992). This analysis was conducted because
linearity was detected between mean SST and minimum or ecological relationships between our predictor variables may
maximum SST, thus suggesting that mean SST alone con- blur direct relationships. Thus, disentangling the existence
tains most of information about temperature extremes. In of ‘mediation effects’ on fish richness may allow for a better
addition, we detected collinearity between shelf area and understanding of their interactions.
coast length and we decided to remove the shelf area from The modeling process in SEM requires the a priori speci-
further analyses because previous research evidenced a neg- fication of pathways considering the relationships between
ligible role of this variable on reef fish richness (Bellwood variables. In this study, we specified the conceptual model
et al. 2005). based on hypothesized mechanisms expected to operate in
BRT were fitted using the gbm library in R (Ridgeway tropical reef systems. First, we included all the direct rela-
2007) with additional functions provided by Elith et al. tionships between explanatory variables and reef fish rich-
(2008). This technique allows for the specification of four ness. Second, we introduced hypothesized relationships
main parameters: bag fraction (bf ), learning rate (lr), tree between pairs of explanatory variables, largely based on pre-
complexity (tc) and the number of trees (nt). bf is the pro- vious knowledge and general ecological theory. A potential
portion of samples used at each step, lr is the contribution of effect of temperature-related variables on coral reef area was
each fitted tree to the final model, tc is the number of nodes included in the model as temperature is known to be a limit-
of each fitted tree determining the extent to which statisti- ing factor for coral reef development and also relates to the
cal interactions were fitted, and nt represents the number aragonite saturation (Kleypas et al. 1999).
of trees corresponding to the number of boosting iterations. Moreover we hypothesized an effect of the region on several
Optimal setting of the parameters was chosen using 5-fold variables. In particular, we hypothesized a region effect on coral
cross validation (CV). The procedure provides a parsimoni- reef area, because coral reef area is asymmetrically distributed
ous estimate, CV–D2 (i.e. the cross validated proportion of across regions (Bellwood et al. 2005). Similarly, we considered
the deviance explained), representing the expected perfor- a regional effect on connectivity, as the Pacific, Indian and
mance of the model when fitted to new data (Elith et al. Atlantic Oceans differ markedly in their history of connectiv-
2008). Using CV, we explored different combinations of the ity (Cowman and Bellwood 2013). Other potential effects of
parameters to be set and retained the model showing the regions, namely on temperature and MD effect were excluded
highest CV–D2 (Supplementary material Appendix 3). because we forcedly considered the same temperature range in
We also explored the possibility of eliminating non- each region and the metric used to measure the MD influence
informative predictor variables to select the most parsimoni- was a priori standardized according to the domain size.
ous model. This simplification process uses CV procedure, SEM was fitted using the ‘lavaan’ library in R using the
progressively simplifying the model fitted to each fold, and Bollen–Stine bootstrapping technique which is adequate
using the average CV errors to decide which and how many when fitting SEMs on variables not showing normal distri-
variables can be removed from the original model without bution (Rosseel 2012).
decreasing its performance (see Appendix 2 in Elith et al. Spatial autocorrelation is a fundamental property of bio-
2008 for details on the simplification procedure). geographical data, with nearby locations are more related than
The relative influence of predictor variables in BRT mod- distant ones. However, the autocorrelation of model residu-
els was evaluated with the contribution to model fit attrib- als is a direct evidence that some unmeasured variables are
utable to each predictor, averaged across all the trees fitted missing to explain the spatial structure in the data (Hawkins
(Friedman 2001). In order to account for uncertainty around et al. 2007, Hawkins 2012). We thus tested for spatial auto-
these estimates, the relative influence of each explanatory correlation in the residuals of the final models (both BRT
variable was calculated on 1000 bootstrap replicates of the and SEM) by computing Moran’s I index for different spatial
original dataset in order to compute 95% confidence inter- lags. Moran’s I correlograms were computed using 5 classes
vals. Similarly, we evaluated the accuracy of our predictive of distance-based neighbors defined by increasing the spatial
model calculating and mapping 95% confidence intervals of lag by 500 km, from 2500 km (i.e. the minimum distance
model predictions at each location. needed for having at least one neighbor for each location)
We also identified the most important interactions to 5000 km which represents the largest spatial gap between
between each pair of predictors (Elith et al. 2008). The effect islands (Victor and Wellington 2000). The analysis was con-
of the most important predictors on fish richness was visu- ducted using the ‘spdep’ package in R (Bivand et al. 2013).
alized by means of partial dependence plots, obtained by Spatial patterns of the residuals were also mapped in order to
plotting predicted values in relation to major variables, while highlight locations where overestimation or underestimation
maintaining other variables constant. of reef fish richness occurred.
As BRT cannot provide a simple way to depict model
structure (i.e. no equation) and outputs, we illustrated BRT
by plotting the first regression tree fitted by the technique. Results
In addition to BRT, which was used with the main goal
of building a predictive model, structural equation models The highest values of reef fish richness were disproportionally
(SEM) were employed to disentangle direct vs indirect effects found in the Indo-Pacific, and especially in the IAA (Table 1;
of explanatory variables on fish richness by evaluating many Fig. 2) with 2016 species in the Philippines. From the IAA,
causal paths. SEM thus quantifies the relative direct and richness tends to decrease almost regularly both eastward
indirect contributions of predictors by incorporating par- and westward. An exception to this regular pattern occurred
tial regression models in order to disentangle pure variable in the western Indian Ocean where species richness between

4-EV
Table 1. Species and genera richness in each of the four regions distance to the MD, mean SST and SST range (Fig. 3). Area
(right part of the table) and species and genera in common for each predictors (i.e. coral reef area and coast length) showed a
pair of regions (left part of the table). Grey cells represent informa-
tion at genus level while white ones are referred to species level. more or less complex positive asymptotic relationship with
species richness (Fig. 4). A positive relationship with fish
Species Genera richness was found for the mean SST and connectivity as
Atlantic Indian Pacific TEP
richness richness
well. The relationship between fish richness and the distance
Atlantic 159 148 161 1186 372 from MD was instead negative and the SST range was linked
Indian 84 678 137 3236 809
to fish richness by a complex hump-shaped relationship.
Pacific 74 2383 142 3919 815
TEP 65 103 120 634 281
The final BRT model showed a high predictive perfor-
mance, explaining 80% of the total deviance using a cross-
validated procedure. The relationship between observed and
Madagascar and Mauritius rises to approximately half the predicted species richness was remarkably high (R2  0.95 for
IAA with more than 950 species. The Atlantic Ocean and a regression of slope 1 and 0 intercept, Fig. 5). In addition,
the Tropical Eastern Pacific (hereafter TEP) showed a com- uncertainty around predictions was extremely low, generally
paratively lower richness in reef fishes. The richest area in being  2% of the observed richness at a given location (see
the TEP was Panama with 330 species, while the richest area maps in Supplementary material Appendix 5).
of the Atlantic Ocean was located between Cuba and the Interaction between predictors played an important role.
Virgin Islands with 461 species. BRTs considering interactions between predictors performed
The final models obtained using either locations defined much better than simpler BRTs (Supplementary material
as polygons of varying size or standard areas within a 600 km Appendix 6). Region, in particular, was the most ‘interact-
radius yielded almost identical results. Therefore, we report ing’ variable and had relatively strong interactions especially
only the results obtained using polygons of varying sizes with coral reef area, distance from MD, and coast length,
for each location because they better reflect the basis of our thus highlighting that these factors have different effects in
fish data, while the results obtained using standard areas are each region (Supplementary material Appendix 6, Fig. 4).
reported in Supplementary material Appendix 4. In general individual effects in the Atlantic Ocean and TEP
The final model accounted for all seven predictor vari- appeared to be less pronounced than in the Pacific and
ables, none of which was removed by the simplification Indian oceans. This is partially due to an ‘intercept’ effect,
procedure. Their relative influences are in descending order: since the diversity of the Indian and Pacific oceans is much
coral reef area, region, coast length, connectivity, normalized higher than that of the other regions.

Figure 2. (a) Map of the observed richness of reef-associated fishes; (b) map of the richness of reef-associated fishes predicted by the BRT
model. Colors and circle size represent 10 richness classes indentified based on the observed richness. The same scheme was later applied
on the map of predicted richness.

5-EV
Figure 3. (a) Relative influence of the predictor variables as ranked by the final BRT model. Error bars represent 95% confidence intervals
obtained from 1000 bootstrap samples of the original dataset. (b) First tree fitted by the BRT. The end points of the tree (nodes) represent
mutually exclusive combinations of independent variables. At each node, a mean estimate of the dependent variable is calculated from the
data points falling within that node. The final BRT model contains 1700 trees.

The tests for detecting spatial autocorrelation in the residu- material Appendix 7). Since the violation of spatial inde-
als were not significant for all the spatial lags considered (at 1st pendence tends to inflate type I errors (Legendre 1993),
lag: Moran’s I  0.0528, p  0.068, Supplementary material we limited the interpretation of the SEM model to the
Appendix 7), indicating that a spatial structure in overestima- path coefficients, without reporting information on the
tion or underestimation of species richness was not found. significance of the effects (Kissling et al. 2008). SEM
The structural equation model was highly significant results were generally consistent with the results obtained
(c2  214, DF  12, p  0.0001) and explained 54% of with the BRT, particularly the ranking of direct predictor
the total variation in fish richness. However, we detected effects (Fig. 6). Path coefficients highlighted that region
significant spatial autocorrelation in SEM residuals (at played mainly a direct role on species richness; its effect
1st lag: Moran’s I  0.1708, p  0.0001, Supplementary on coral reef and connectivity was comparatively lower.

Figure 4. BRT model predictions of reef-associated fish richness according to several explanatory variables for the different regions. Values
for predictors other than those on the x-axis are held constant at their mean. Black lines: Pacific Ocean; grey lines: Indian Ocean; dotted
black lines: Atlantic Ocean; dotted grey lines: Tropical Eastern Pacific.

6-EV
 history, Jetz and Fine (2012) demonstrate how jointly con-
sidering historical and contemporary environmental pre-
dictors accounts for more than 80% of the variation in the
diversity of terrestrial vertebrates at a global scale.
The analysis of interaction strength between explanatory
variables highlighted that the region (used here as a proxy of
evolutionary history) established strong interactions, espe-
cially with the MD effect, coral reef area, coast length and
SST range. The effect of evolutionary and biogeographical
history on reef fish distribution patterns has been docu-
mented worldwide (Floeter et al. 2008, Bellwood et al.
2012, Cowman and Bellwood 2013) and has underpinned
major shifts in species richness distribution during the past
50 million yr (Renema et al. 2008). Our results and the
interactions detected between region and other predictors
are consistent with the hypothesis that a combined effect
of phylogenetic and biogeographical histories, and present
environmental characteristics is likely to shape the present
Figure 5. Relationship between observed species richness at each geographical pattern of reef fish richness. In this regard, it
location and species richness predicted by the BRT model. Empty is worth mentioning that the relative importance of vari-
black circles: Atlantic Ocean; empty grey circles: Tropical Eastern ables obtained with the BRT model includes most of their
Pacific (TEP); filled black circles: Pacific Ocean; filled grey circles:
Indian Ocean. interactions. This also explains the higher influence of the
region factor recorded by BRT compared to SEM models.
While BRT accounts for interactions between predictors
In contrast, mean SST showed a strong positive indirect but not for mediation effects, SEM does the opposite and
effect on species richness, mediated by coral reef area. their results should be seen as complementary. The spatial
autocorrelation detected in SEM residuals but not in BRT
residuals is likely due to the higher predictive performance
Discussion of the latter models which show a low proportion of
We analyzed the relative contribution of a range of unexplained variation.
variables in explaining current-day tropical reef fish rich- Besides history, our analyses evidence a primary role of
ness at a global scale. Our analysis revealed the potential area variables and, in particular, of coral reef area. This is
for biogeographical history to interact with present envi- in agreement with the findings of Bellwood et al. (2005),
ronmental variables to shape the distribution of reef fish who identified coral reef area and MD effect as the key vari-
richness. Environmental predictors, in fact, showed differ- ables explaining most variation in reef fish diversity patterns
ent effects according to the biogeographical region. This across the Indo-Pacific. However, contrary to their analyses,
has been previously documented for terrestrial environ- we detect only a minor role of the MD effect. This is, at least
ments, where the role of present environmental variables in part, due to the global scale of our work, embracing five
has been shown to differentially shape amphibians rich- separate domains, whereas Bellwood et al. (2005) focused on
ness in different biogeographical realms (Buckley and Jetz the Indo-Pacific domain. Partial dependence plots confirm
2007). Likewise, with deeper insights into evolutionary an important role of the MD in the Pacific Ocean while its
effect is less influential in the Indian Ocean and especially in
the TEP and the Atlantic Ocean (Fig. 4). The comparatively
lower influence of the MD effect in the Indian Ocean is
likely due to the high richness of the south-west coast of the
basin reaching a fish species richness almost half that of the
IAA. This secondary hotspot has already been documented
to be the cause of a main deviation from the null expecta-
tion depicted by the sole MD effect (Connolly et al. 2003).
Despite the high number of species that the Indian Ocean
shares with the Pacific, our findings suggest that the Indian
Ocean may act, at least in part, as a separate domain. The
negligible role of the MD effect in the Atlantic Ocean, is
probably due to an asymmetric distribution of shallow water
or reefal habitats, low diversity faunas, limited geographical
Figure 6. Results of the structural equation model performed to extents, and extensive extinction (O’Dea et al. 2007, Floeter
assess direct and indirect effects of explanatory variables on reef fish
species richness. Numbers represent standardized path coefficients.
et al. 2008).
Arrows thickness is proportional to path coefficients. Since we Our analyses also evidence an important role of connec-
detected spatial autocorrelation for SEM residuals potentially tivity, especially in the Indian and Pacific Oceans. This result
inflating type I error, we reported all the path coefficients without is highly consistent with Cowman and Bellwood (2013) who
considering p-values. highlight the evolutionary role of dispersal from the IAA

7-EV
in the development of Indian and Pacific Ocean reef fish contrast, the geographically restricted Caribbean, Atlantic and
faunas. Our results are also in agreement with the findings of TEP have had a long history of isolation, with the resultant
Mora et al.(2003) who explained the low richness character- extinction of marine taxa and limited diversification within
izing isolated Pacific islands as the result of a filtering from the remaining fish lineages (Bellwood and Wainwright 2002,
the IAA (cf. the center of origin hypothesis; Bellwood et al. O’Dea et al. 2007, Cowman and Bellwood 2013). This evo-
2012). Similarly, although working at smaller scales, Sandin lutionary filtering of the fish faunas may, at least in part, help
et al. (2008) report an important role of connectivity on spe- explain why Caribbean fish assemblages have much stronger
cies richness in the Caribbean. associations with non-reef habitats (rocky reefs, mangroves,
Our results highlighted a limited influence of SST related seagrasses etc) than their Indo-Pacific counterparts (Bellwood
variables. This is in agreement with previous research con- and Wainwright 2002, Barnes et al. 2012); because they had
ducted in the Indo-Pacific (Bellwood et al. 2005), but in to survive during periods with limited reef area. It may also
contrast with other studies conducted in the Atlantic Ocean explain why the composition of Caribbean and TEP reef fish
or at a global scale (Macpherson 2002, Tittensor et al. 2010). faunas have a stronger resemblance to those of non-tropical
This is probably due to the extent of the latter works which coastal faunas, such as that of New Zealand, rather than low-
explored a broader latitudinal gradient, including multiple latitude coral reef assemblages (Bellwood and Wainwright
taxa in temperate and polar environments. From this per- 2002). It is this history that may explain, at least in part, the
spective, the relationship between species richness and tem- limited response of TEP and Atlantic fish faunas to reef area,
perature may be not linear, thereby exerting a significant role temperature or other reef-associated variables.
at the interface between cold-temperate and tropical waters, Our results have important implications in terms of reef
but a minor effect once tropical temperatures (i.e. higher fish conservation. Coral reef area was the most influential vari-
than 17°C in our case) are reached. able according to both BRT and SEM analyses. However coral
Structural equation models provided further insight into reefs are presently at risk and under multiple human pressures,
the way the predictors of species richness may interact. In including climate change, habitat destruction, pollution, over-
this regard, we expected history to play mainly indirectly on fishing, and ocean acidification (Pandolfi et al. 2011) lead-
reef fish richness, most isolated islands and coral reefs being ing to the reduction of live coral cover and suitable habitat.
asymmetrically located in the Pacific Ocean. Our results, In addition, our findings show an important positive effect of
however, suggest that the region factor has a major direct mean SST on reef fish richness, mediated by coral reef area.
effect on species richness and help to disentangle the rela- Anthropogenic climate change has already triggered profound
tionship between fish richness and the other predictors for responses of coral reefs due to increasing temperature and posi-
different regions. SST, instead, has a major indirect role on tive thermal peaks that have resulted in massive die-off of coral
reef fish richness mediated by coral reef area. This partially species and a significant decrease of coral cover worldwide
clarifies the minor role of temperature on reef species richness (Goreau et al. 2000) and their associated fishes (Wilson et al.
documented by previous research (Bellwood et al. 2005) and 2006). Human pressure is thus acting on the main predic-
from the BRT models in the present work. Coral reef devel- tors of reef fish species richness both directly by local impacts,
opment is known to be severely limited by SST (Kleypas and indirectly by global warming. Given the strong interac-
et al. 1999) and, according to the results obtained by SEM, tions identified, our results suggest that different regions will
its effect on fish richness may be interpreted as linked to hab- respond differently to ongoing anthropogenic pressures.
itat provision more than to the energy hypothesis itself. Our work demonstrates that a few and readily available
One of the most striking results was the among-region variables are sufficient to accurately predict the pattern of
variation in the response of fish diversity to environmental variation in reef fish species richness at global scale. We found
variables. Previous work has reported variation among major a very strong relationship between observed and predicted
marine groups in their response to environmental variables. fish species richness and low uncertainty around prediction
Macroalgae, for example, have been reported to be more sus- estimates. The high proportion of cross-validated deviance
ceptible than fishes to SST (Kerswell 2006). Variation within highlights the applicability of our model to new locations,
a taxon is, in some ways, counterintuitive. One may expect thereby representing an important tool for attempting fill-
fishes to react to temperature in a similar manner, regardless ing the gaps in those areas where information on fish species
of the region. The observed variation among regions, how- richness is still unavailable or inaccessible.­­­­­­
ever, may have both methodological and historical expla-
nations. In terms of methodology, the regions vary greatly Acknowledgements – This work was part of the GASPAR program in
in the magnitude of the gradients observed. For example, CESAB, funded by FRB (Foundation for Research on Biodiversity).
the area of reef habitat in the Pacific (within 600 km radius DRB was supported by the Australian Research Council. DM was
areas) varies from 0 km2 (e.g. Sala Y Gomez) to  10 000 km2 supported by a Marie Curie International Outgoing Fellowship
(Great Barrier Reef ), while in the TEP it ranges from (FISHECO) with agreement number IOF-GA-2009-236316. AMF
0 to about 100 km2 (Costa Rica), as the entire TEP has was supported by the Univ. of Hawaii Dept of Biology with Contri-
 200 km2 of reef area. Thus the gradients, and the capacity bution no. 2013-010. We are thankful to Mariana Bender (Univ.
of fishes to respond, are region-specific. Federal do Paraná) and Derek Tittensor (UNEP-WCMC) for useful
suggestions on early versions of the manuscript. Thanks are due also
Historically, the various regions have experienced mark- to Camilo Mora (Univ. of Hawaii) for sharing ideas about reef fish
edly different evolutionary histories. The Indian and Pacific species richness, to Serge Andréfouët (IRD) for his help on the man-
Oceans have experienced extended periods of connectivity, agement of spatial information regarding coral reefs and to Doug
and within these oceans reef fish lineages have exhibited Hoese (Australian Museum) for his help in collecting information
expansion and dispersal (Cowman and Bellwood 2013). In on reef fish species richness around Australia.

8-EV
References Jetz, W. and Fine. P. V. A. 2012. Global gradients in vertebrate
diversity predicted by historical area-productivity dynamics
Allen, A. P. et al. 2002. global biodiversity, biochemical kinetics, and contemporary environment. – PLoS Biol. 10: e1001292.
and the energetic-equivalence rule. – Science 297: 1545–1548. Kerswell, A.P. 2006. Global biodiversity patterns of benthic marine
Andréfouët, S. et al. 2006. Global assessment of modern coral reef algae. – Ecology 87: 2479–2488.
extent and diversity for regional science and management Kissling, W. D. et al. 2008. Spatial patterns of woody plant and
applications: a view from space. – Proceedings of the Interna- bird diversity: functional relationships or environmental effects?
tional Coral Reef Symposium. – Global Ecol. Biogeogr. 17: 327–339.
Barnes, L. et al. 2012. The use of clear-water non-estuarine man- Kleypas, J. A. et al. 1999. Environmental limits to coral reef devel-
groves by reef fishes on the Great Barrier Reef. – Mar. Biol. opment: where do we draw. – Am. Zool. 159: 146–159.
159: 211–220. Knowlton, N. and Jackson, J. B. C. 2008. Shifting baselines, local
Bellwood, D. R. and Hughes, T. P. 2001. Regional-scale assembly impacts, and global change on coral reefs. – PLoS Biol. 6: e54.
rules and biodiversity of coral reefs. – Science 292: 1532–1534. Legendre, P. 1993. Spatial autocorrelation: trouble or new para-
Bellwood, D. R. and Wainwright, P. C. 2002. The history and digm? – Ecology 74: 1659–1673.
biogeography of fishes on coral reefs. – In: Sale, P. F. (ed.), Coral Luiz, O. J. et al. 2012. Ecological traits influencing range expansion
reef fishes. Dynamics and diversity on a complex ecosystem. across large oceanic dispersal barriers: insights from tropical
Academic Press, pp. 5–32. Atlantic reef fishes. – Proc. Biol. Sci. 279: 1033–1040.
Bellwood, D. R. et al. 2004. Confronting the coral reef crisis. Macpherson, E. 2002. Large-scale species-richness gradients in the
– Nature 429: 827–833. Atlantic Ocean. – Proc. Biol. Sci. 269: 1715–1720.
Bellwood, D. R. et al. 2005. Environmental and geometric constraints Mitchell, R. J. 1992. Testing evolutionary equation modelling path
on Indo-Pacific coral reef biodiversity. – Ecol. Lett. 8: 643–651. analysis and structural equation modelling. – Funct. Ecol. 6:
Bellwood, D. R. et al. 2012. Biodiversity hotspots, evolution and 123–129.
coral reef biogeography: a review. – In: Gower, D. et al. (eds), Moilanen, A. and Nieminen, M. 2002. Simple connectivity meas-
Biotic evolution and environmental change in southeast Asia. ures in spatial ecology. – Ecology 83: 1131–1145.
Cambridge Univ. Press, pp. 2–32. Mora, C. and Robertson, D. R. 2005. Causes of latitudinal gradi-
Belmaker, J. et al. 2011. The influence of connectivity on richness and ent in species richness: a test with fishes of the Tropical Eastern
temporal variation of reef fishes. – Landscape Ecol. 76: 113–133. Pacific. – Ecology 86: 1771–1782.
Bivand, R. et al. 2013. spdep: spatial dependence: weighting Mora, C. et al. 2003. Patterns and processes in reef fish diversity.
schemes, statistics and models. – R package ver. 0.5-56, – Nature 421: 933–936.
[Link]  spdep. Mora, C. et al. 2012. High connectivity among habitats precludes
Buckley, L. B. and Jetz, W. 2007. Environmental and historical the relationship between dispersal and range size in tropical reef
constraints on global patterns of amphibian richness. – Proc. fishes. – Ecography 35: 89–96.
Biol. Sci. 274: 1167–1173. O’Dea, A. et al. 2007. Environmental change preceded Caribbean
Colwell, R. K. and Hurtt, G. C. 1994. Nonbiological gradients in extinction by 2 million years. – Proc. Natl Acad. Sci. USA 104:
species richness and a spurious Rapoport effect. – Am. Nat. 5501–5506.
144: 570–595. Pandolfi, J. M. et al. 2011. Projecting coral reef futures under global
Connolly, S. R. et al. 2003. Indo-Pacific biodiversity of coral reefs: warming and ocean acidification. – Science 333: 418–422.
deviations from a mid-domain model. – Ecology 84: 2178–2190. Renema, W. et al. 2008. Hopping hotspots: global shifts in marine
Cowman, P. F. and Bellwood, D. R. 2013. The historical bio­ biodiversity. – Science 321: 654–657.
geography of coral reef fishes: global patterns of origination and Ridgeway, G. 2007. gbm: generalized boosted regression
dispersal. – J. Biogeogr. 40: 209–224. models. – R package ver. 1.6-3, [Link]
Currie, D. J. and Kerr, J. T. 2008. Tests of the mid-domain hypoth- web/packages/[Link].
esis: a review of the evidence of the mid-domain. – Ecol. Robertson, D. R. and Cramer, K. L. 2009. Shore fishes and bio-
Monogr. 78: 3–18. geographic subdivisions of the Tropical Eastern Pacific. – Mar.
Dormann, C. F. et al. 2012. Collinearity: a review of methods to Ecol. Prog. Ser. 380: 1–17.
deal with it and a simulation study evaluating their perform- Robertson, D. R. et al. 2004. Tropical transpacific shore fishes.
ance. – Ecography 35: 1–20. – Pac. Sci. 58: 507–565.
Elith, J. et al. 2008. A working guide to boosted regression trees. Rosseel, Y. 2012. lavaan: an R package for structural equation
– J. Anim. Ecol. 77: 802–813. modeling. – J. Stat. Softw. 48: 1–36.
Floeter, S. R. et al. 2008. Atlantic reef fish biogeography and Sandin, S. et al. 2008. Island biogeography of Caribbean coral reef
evolution. – J. Biogeogr. 35: 22–47. fish. – Global Ecol. Biogeogr. 17: 770–777.
Friedman, J. H. 2001. Greedy function approximation: a gradient Swearer, S. E. et al. 2002. Evidence of self-recruitment in demersal
boosting machine. – Ann. Stat. 29: 1189–1232. marine populations. – Bull. Mar. Sci. 70: 251–271.
Gaston, K. J. 2000. Global patterns in biodiversity. – Nature 405: Tittensor, D. P. et al. 2010. Global patterns and predictors of
220–227. marine biodiversity across taxa. – Nature 466: 1098–1101.
Goreau, T. et al. 2000. Conservation of coral reefs after the 1998 Tyberghein, L. et al. 2012. Bio-ORACLE: a global environmental
global bleaching event. – Conserv. Biol. 14: 5–15. dataset for marine species distribution modelling. – Global
Gray, J. S. 2001. Marine diversity: the paradigms in patterns of Ecol. Biogeogr. 21: 272–281.
species richness examined. – Sci. Mar. 65: 41–56. Victor, B. C. and Wellington, G. M. 2000. Endemism and the
Hawkins, B. A. 2012. Eight (and a half ) deadly sins of spatial pelagic larval duration of reef fishes in the eastern Pacific
analysis. – J. Biogeogr. 39: 1–9. Ocean. – Mar. Ecol. Prog. Ser. 205: 241–248.
Hawkins, B. A. et al. 2003. Energy, water, and broad-scale geo- Wilson, S. K. et al. 2006. Multiple disturbances and the global
graphic patterns of species richness. – Ecology 84: 3105–3117. degradation of coral reefs: are reef fishes at risk or resilient?
Hawkins, B. A. et al. 2007. Red herrings revisited: spatial auto­ – Global Change Biol. 12: 2220–2234.
correlation and parameter estimation in geographical ecology. Worm, B. et al. 2006. Impacts of biodiversity loss on ocean eco-
– Ecography 30: 375–384. system services. – Science 314: 787–790.

Supplementary material (Appendix ECOG-00291 at www.

[Link]/appendix). Appendix 1–7.

9-EV