Disentangling drivers of the abundance of coral reef fishes in the Western Indian Ocean

Abstract Aim Understanding the drivers of the structure of coral reef fish assemblages is vital for their future conservation. Quantifying the separate roles of natural drivers from the increasing influence of anthropogenic factors, such as fishing and climate change, is a key component of this understanding. It follows that the intrinsic role of historical biogeographical and geomorphological factors must be accounted for when trying to understand the effects of contemporary disturbances such as fishing. Location Comoros, Madagascar, Mozambique and Tanzania, Western Indian Ocean (WIO). Methods We modeled patterns in the density and biomass of an assemblage of reef‐associated fish species from 11 families, and their association with 16 biophysical variables. Results Canonical analysis of principal coordinates revealed strong country affiliations of reef fish assemblages and distance‐based linear modeling confirmed geographic location and reef geomorphology were the most significant correlates, explaining 32% of the observed variation in fish assemblage structure. Another 6%–8% of variation was explained by productivity gradients (chl_a), and reef exposure or slope. Where spatial effects were not significant between mainland continental locations, fishing effects became evident explaining 6% of the variation in data. No correlation with live coral was detected. Only 37 species, predominantly lower trophic level taxa, were significant in explaining differences in assemblages between sites. Main Conclusions Spatial and geomorphological histories remain a major influence on the structure of reef fish assemblages in the WIO. Reef geomorphology was closely linked to standing biomass, with “ocean‐exposed” fringing reefs supporting high average biomass of ~1,000 kg/ha, while “lagoon‐exposed fringing” reefs and “inner seas patch complex” reefs yielded substantially less at ~500kg/ha. Further, the results indicate the influence of benthic communities on fish assemblages is scale dependent. Such insights will be pivotal for managers seeking to balance long‐term sustainability of artisanal reef fisheries with conservation of coral reef systems.


| INTRODUC TI ON
Biodiversity is declining globally as a result of direct human impacts including overexploitation of natural resources which in turn threatens ecosystem functioning (Butchart et al., 2010;Mora et al., 2011).
Coral reefs are one of the most biodiverse ecosystems on the planet and of great importance for livelihoods and economies Moberg & Folke, 1999). These issues come together in the Western Indian Ocean (WIO) where coral reefs are associated with developing countries with artisanal fisheries of high socioeconomic value for poor coastal communities, though in many areas fisheries management measures are inadequate (Samoilys, Osuka, Maina, & Obura, 2017;Walmsley, Purvis, & Ninnes, 2006;Wells, Samoilys, Makoloweka, & Kalombo, 2010).
Biodiversity loss and human impacts on coral reefs are further exacerbated by climate change (Bellwood, Hughes, Folke, & Nyström, 2004;Hughes et al., 2007). Coral reefs are extremely vulnerable to rising sea surface temperatures resulting in mass coral bleaching (Ateweberhan, McClanahan, Graham, & Sheppard, 2011) and to ocean acidification, now driving some reefs into a state of net erosion . The concomitant impact of coral bleaching and mortality on reef fishes has been well studied Wilson et al., 2008). In addition, significant declines in reef fish biomass due to fishing in the Indo-Pacific have been reported (Friedlander & DeMartini, 2002;McClanahan, Maina, Graham, & Jones, 2016;Sandin, Smith, et al., 2008). However, teasing apart natural drivers, climate change impacts and fishing effects on the structure of reef fish assemblages have been less clearly examined.
In this study, we sought to assess the relative impacts of natural versus anthropogenic factors that are affecting reef fish assemblages in the WIO. We structured our approach through two relatively wellestablished hypotheses. The first was that fish assemblages will vary naturally in relation to a number of larger scale abiotic factors such as biogeography, reef structure, and oceanic nutrient levels, all either previously recognized or likely to have some influence in structuring coral reef communities (Heenan, Hoey, Williams, & Williams, 2016;Pinca et al., 2012;Taylor, Lindfield, & Choat, 2014). But at the smaller scale of reef habitat, the cover of hermatypic corals, algae, rubble, and rugosity also plays a significant role in structuring fish assemblages (Chabanet, Ralambondrainy, Amanieu, Faure, & Galzin, 1997;Halford, Cheal, Ryan, & Williams, 2004;Samoilys, Roche, Koldewey, & Turner, 2018). Our second hypothesis was that the abundance and biomass of coral reef fishes will vary in relation to protective management and fishing, which has been widely demonstrated in studies across the Indo-Pacific (D'agata et al., 2014;Edgar et al., 2014;McClanahan, Graham, Calnan, & MacNeil, 2007). We maximized the range of depths of the surveys to be able to characterize the fish assemblage at each reef (Wedding & Friedlander, 2008) rather than select a small depth range to minimize data variance. Anthropogenic stressors, both positive (management) and negative (extraction through fishing), were represented by human population density, fishing pressure, and an index of management protection.
One of our primary objectives was to clarify significant drivers of fish assemblage structure while also paying due regard to the role of spatial autocorrelation in the misinterpretation of multiple regression models (Hawkins, 2012;Legendre, 1993). Spatial autocorrelation is a well-known manifestation of community data which, if not explicitly accounted for in models, will inflate the significance of other terms in the model (Peres-Neto & Legendre, 2010). The Western Indian Ocean (WIO) biogeographic region (Spalding et al., 2007) represents a region of highly variable coral reefs (Sheppard, 2000), within an oceanic context of the South Equatorial Current and the East African Coastal Current (Schott, Xie, & McCreary, 2009). This provided an ideal study area for enabling the collation of 16 explanatory variables to explore drivers of the structure of coral reef fish assemblages.

| Sites
Reef fish population abundance and benthic cover were measured at 53 coral reef sites across Tanzania (Table 1). Due to the extent of coral reefs in these countries, the study sites are not representative of each country as a whole. Reef geomorphology was categorized on a 3-level hierarchical typology of coral reef types defined for the WIO based on 6 geological, 7 geomorphological, and 6 reef types (Table 1).

| Explanatory variables
Environmental variables that may influence patterns in reef fish assemblages were measured at each site and included in situ estimates of exposure to waves and trade winds, minimum and maximum depth, rugosity, and reef slope (Table 2). Other variables defined per site included chlorophyll_a, as a proxy for nutrient levels in the water, obtained from the ocean color CCI web GIS portal.
Chl_a readings for each site were calculated from monthly averages between 1997 and 2013 from three pixels closest to each K E Y W O R D S biogeography, biomass, climate, coral reef, fish, geomorphology site. Two indices for anthropogenic variables were also defined per site: (a) human population density adjacent to the survey sites (sensu Taylor et al., 2014) and (b) an overfishing and destructive fishing threat measure (adapted from Burke, Reytar, Spalding, & Perry, 2011). Gridded human population data sourced at 4-km 2 pixel resolution (CIESEN, 2015;Doxsey-Whitfield et al., 2015) were converted to a vector point layer using ArcGIS 10.3 and then projected onto geo-referenced survey sites and the shortest distance between the two vector points taken to assign human population value per site. The fishing pressure threat index was derived from the World Resources Institute's evaluations of coastal population density and extent of fishing areas, with adjustments for increased fishing demand due to proximity to large populations and market centers. This adjustment involved all reef sites that were within 200 km of ≥500,000 people but were initially rated as low threat based on local population size. Such sites were reclassified to medium threat (sensu Burke et al., 2011). The threats ranged from low (0) to high (1,000). Areas where destructive fishing occurs (with explosives or poisons, MS pers. obs) were added.
A protection index was assigned for each site based on the effectiveness of management determined from the literature, personal knowledge, and communicating with managers (Supporting Information Table S3).

| Benthic surveys
Benthic surveys estimated % cover of live hard coral, three types of algae and rubble (Supporting Information Table S2) using 25 × 1 m transects (sensu Sandin, Smith, et al., 2008), with generally two transects per site, at 47 of the 53 sites.

| Fish surveys
The density and size classes of all reef fish species from 11 families and 12 trophic groups (Table 3; sensu Green & Bellwood, 2009) were counted using SCUBA-based underwater visual census (Samoilys & Carlos, 2000)

| Data compilation
The complete fish dataset consisted of density and biomass for 156 species across 53 sites. Preliminary explorations identified two extreme outlier sites (Fernau Vloso & Lalane, Mozambique), which were therefore removed. To further reduce variance, any species occurring <3 times across all sites were removed. The final dataset for analyses consisted of 45 sites (Comoros-8, Madagascar-9, Mozambique-16, and Tanzania-12) for 123 coral reef-associated fish species/taxa (see Supporting Information: Table S1).

Zanzibar
Continental Island Inner seas-exposed fringing reef Forereef 1 Ocean-exposed fringing reef Forereef 1 Numbers in parentheses sum total number of sites per geomorphology type and reef type across all countries, illustrating most sites were forereefs.
where necessary, followed by testing for collinearity through the use of variance inflation factors (VIF, Supporting Information Appendix S1). The presence of spatial autocorrelation was tested at two scales: across all locations and across locations within the eastern African mainland countries (Mozambique and Tanzania). See Supporting Information Appendix S1 for further details.

| Analyses
Spatial patterns of the fish assemblages across sites were first examined through unconstrained clustering of Bray-Curtis similarities calculated on square-root-transformed abundance data using Ward's clustering algorithm. An IndVal analysis (De Caceres & Legendre, 2009) was performed to identify which species were significant delineators of the observed groupings and the distribution of species was displayed via a heat map. Further examination of spatial patterns was performed using canonical analysis of principal coordinates (CAP), to maximize differences between groups, using square-root-transformed abundance and log (x + 1) transformed biomass data. One-way PERMANOVAs confirmed the significance of country differences for both abundance and biomass at (P(Perm) < 0.001). Sites were therefore coded by country a priori for the CAP. We identified which fish species were most influential in describing differences between countries by superimposing vectors representing Pearson correlations (|>0.45|) of individual species with the CAP axes (Anderson, Gorley, & Clarke, 2008). A complementary SIMPER analysis (Clarke & Gorley, 2015) was run to provide more quantitative information on species that were most influential in describing differences between countries.
Distance-based linear models (DistLM) were run to investigate potential drivers of the observed reef fish assemblage patterns (sensu Legendre & Anderson, 1999), incorporating the 16 explanatory variables. Marginal tests were done on all variables to investigate the range of variation that could be explained. The BEST procedure was used for building the models with the best models chosen through AIC c and BIC selection criteria (see Supporting Information Appendix S1). All analyses were performed using the TA B L E 2 Final list of driver variables tested for influence on fish species' population density and biomass

Geographic location
Latitude and longitude of reef site Geography well-known driver of reef fish assemblages in the I-P Mora (2015)  Reef slope Visual estimate of slope in degrees following Sandin, Smith, et al. (2008) Reef fish population abundances are known to vary with the degree of reef slope Wedding and Friedlander (2008) Depth Two measures of depth at site were tested: minimum and depth range Reef fish population abundance is known to vary with depth Wedding and Friedlander (2008) Rugosity Visual estimate on site High rugosity associated with higher biomass of reef fishes Samoilys et al. (2018) and Heenan et al. (2016) Reef benthos % cover of five key benthic types measured in situ: hard coral, fleshy macroalgae, turf algae, CCA, and rubble, following Sandin, Smith, et al. (2008) Reef benthos known to drive fish populations in multiple relationships Chabanet et al. (1997)  Note. All taxa were recorded to species level (not all species are listed here). Those split by body size are species that change diet with size. Trophic categories and feeding information based on Choat and Clements (1998), Choat, Clements and, Robbins (2002), Samoilys and Carlos (2000), Green and Bellwood (2009) and Clements, German, Piche, Tribollet, and Choat (2016).

| Spatial patterns in fish assemblages
The cluster analysis on species density data identified five groups (similarity <0.7) and illustrated that sites in Comoros and Madagascar strongly separated from those in the adjacent mainland countries of Mozambique and Tanzania (Supporting Information Figure S1)   Figure 2). Both Madagascar and Comoros also differed due to several species that occurred in very low numbers or were missing, most notably the caesionids (Caesio spp.) which were found in most sites throughout Tanzania and Mozambique, and were highly abundant in some of these sites ( Figure 2). Other species that occurred in low densities in Madagascar and Comoros were the pennant butterflyfish Heniochus acuminatus and the emperor Lethrinus erythracanthus. Note that for ease, sites F I G U R E 2 Heat map illustrating the spatial distribution and abundance (square-root-transformed) of the 37 species/taxa found to be significant indicators of the five cluster groupings from a total species list of 123 species. Significance derived from IndVal analysis (Dufrene & Legendre 1997). The 3rd and 4th groups illustrate the most diverse assemblages, while species that are ubiquitous spanning most sites are illustrated in horizontal bands      (Table 4).
These mirror patterns in the heat map ( Figure 2).
Significant species that explained regional patterns in fish biomass ( Figure 3b) included many of the same species in the analyses for fish density (Table 4). However, some new species appear which were larger bodied, such as the omnivore Lutjanus fulviflamma and the piscivorous grouper Cephalopholis argus, while others such as Ctenochaetus striatus were no longer significant (Figure 3b). The SIMPER analysis with biomass data revealed other significant species, such as the omnivore Plectorhinchus gaterinus, the scraper Scarus tricolor, the grazer-detritivore Acanthurus blochii, the planktivore Heniochus acuminatus, and the corallivore Chaetodon lineatus (Table 4). Six species that strongly delineate differences in both density and biomass of species assemblages between the countries are illustrated in Figures 4 and 5.

| Abiotic and biotic factors affecting species density and biomass
The most parsimonious DistLM selected four significant variables which explained 39.3% of the observed variance in the fish species density dataset. Space (composite of latitude and longitude) and reef geomorphology were the most significant, explaining 12.8% and 18.9% of the variation in the dataset, respectively.
Reef exposure (4.1%) and chlorophyll_a (3.5%) were also significant, but at a reduced level. The direction of these variables on the species' abundance matrix is shown as vectors in the CAP ordination ( Figure 3a). Space clearly separates the relative density of species in the assemblages of the two island countries, and from the mainland countries. The vectors show that the 7 reef geomorphology types (Table 1) correlated strongly with fish assemblages at different sites and countries. For example, "Coastal barrier reef complex" (geo_cbrc) reefs were typical of Mozambique, both Mozambique and Tanzania shared fish assemblages characteristic of "Inner seas patch reef complex" (geo-isprc) reefs, while "Ocean-exposed fringing reef" (geo_oefr) correlated most with sites in Comoros. The most exposed F I G U R E 4 CAP ordinations of density (left) and biomass (right) of fish assemblages with bubble plots representing density indiv./1,000 m 2 and biomass (kg/1,000 m 2 ) of the excavating sites (exposure) were found in Comoros and sites with the highest chlorophyll_a were in Mozambique (Figure 3).
The best-fit DistLM identified similar clear separation of fish assemblages between countries based on biomass data (Figure 3b), explaining 40.4% of the variation in the data and identified three of the same variables as in the density data. Space and geomorphology explained 13.2% and 18.9% of the variation in the dataset, respectively, and Chlorophyll_a explained 5.0% of the variation. However, reef slope rather than exposure was significant, explaining 3.3% of the variation in the data. In summary, the CAP analyses and DistLM showed that space (latitude/longitude), reef geomorphology, chloro-phyll_a, exposure, and slope were valid predictors of the structure of the fish assemblages across the region. Notably, none of the benthic variables, including live coral cover, were significant.

| Tanzania and Mozambique datasets
At the smaller spatial scale of Tanzania and Mozambique, the DistLM found geomorphology explained the highest variation in the density data at 23.3%, with exposure and fishing pressure also significant explanatory variables, at 6.5% and 6.0%, respectively. The same variables were also significant correlates of the biomass data with geomorphology, exposure, and fishing pressure explaining 22.7%, 6.5%, and 6.5% of the variance, respectively. Thus, at this smaller spatial scale, fishing pressure became a significant variable for both fish density and biomass. However, again, none of the benthic variables were significant.

| Total fish biomass
Total fish biomass, based on all 12 families, was low to moderate in all sites in the Comoros at <558 kg/ha and highly variable within the other three countries depending on the location of the survey site ( Figure 6, Supporting Information  Table S4). A similar wide range of values was also seen in Tanzania   (Supporting Information Table S4).

| Geographic patterns in fish assemblages
Geographic location and reef geomorphology were the most significant drivers of the observed patterns in the structure of reef fish assemblages across the WIO, based on density and biomass data Notes. Species rankings were ordered by the ratio of species dissimilarity/standard deviation to highlight those species which best highlight the differences between countries. Numbers in parentheses are the percentage contribution of each species to the overall dissimilarity, which is presented in bold type. B: big; S: small.

TA B L E 4 (Continued)
F I G U R E 5 Bubble plots of those species identified as significant delineators of country differences across both abundance (left plots) and biomass (right plots) CAP ordinations. Size of bubbles are comparable within each column and represent squareroot abundances (individ./1,000 m 2 ) and log(x + 1) biomass (kg/1,000 m 2 ), respectively. + = zero count found human influences were more significant drivers than biogeographic variables on parrotfish functional and phylogenetic diversities. We show that the structure of fish assemblages of Comoros, NE Madagascar, northern Mozambique, and Tanzania aligns with ocean-exposed fringing reefs, lagoon and barrier banks, coastal barrier reef complexes, and inner-seas-exposed fringing reefs. Three other biophysical variables were also significant, albeit to a lesser degree: chlorophyll_a; reef exposure and reef slope. Since the WIO represents a biogeographic province of similarity in fish species distributions (Bellwood & Wainwright, 2002;Kulbicki, Parravacini, & Mouillot, 2015), the results provide evidence that a reef's geographic location, structure, and surrounding environmental conditions are key variables influencing patterns in reef fish assemblages in the WIO.

| Large-scale environmental drivers of reef fishes
Biogeographic drivers, such as the mid-domain effect (MDE), reef location, isolation, and connectivity (Mora, 2015;Parravicini et al., 2013), may help explain why the fish assemblages in Comoros and Madagascar are distinctly different from mainland eastern Africa.
Other key explanatory variables these studies report include sea surface temperature (SST), coast length, and reef area (see also Bellwood, Hughes, Connolly, & Tanner, 2005), based on predictors of reef fish species richness. While these studies conclude that productivity (chlorophyll_a) is not a key factor (Mora, 2015), our results show that chlorophyll_a is a significant variable in driving the F I G U R E 6 Pie charts representing total biomass (kg/ha) allocated into seven biomass categories based on number of sites per country. Unmarked charts are from 13 families of reef-associated fishes from this study; other charts represent other studies from the same countries in approximately the same time period. Chagos and the Hawaiian Islands included for comparisons of unfished or lightly fished reefs. *Study's values include sharks observed patterns in fish assemblages. Higher productivity on the mainland African coast may support higher densities and biomass of reef fishes. This finding is consistent with a large-scale study across the Pacific which found higher oceanic productivity was associated with over double the biomass of all reef fishes (sharks and trevally not included), and was notably significant for planktivores and piscivores (Williams et al., 2015). fore, SST may also be a contributing factor to the regional patterns detected. Coast length and reef area, both associated with higher species richness and fish size (Bellwood et al., 2005;Kulbicki et al., 2015;Parravicini et al., 2013), may explain differences between Comoros and the other three countries, since Comoros' reef area ranges from only 6% to 9% of that of the other three countries, and Comoros' coast length is between 9% and 34% of the other countries (UNEP-WCMC, 2010; Wessel & Smith, 1996). For both these variables, Madagascar has the largest reef area and longest coast length; therefore, MDE, isolation, and connectivity may be more important in explaining the strong separation of Madagascan fish assemblages from Tanzania and Mozambique.

| Effects of fishing on fish assemblages
Fishing has been shown to precipitate top-down and bottom-up trophic cascades on coral reefs Graham et al., 2017;Sandin & Zgliczynski, 2015). We did not find a significant fishing effect on the fish assemblages we surveyed at the largest spatial scale, despite including two uncorrelated measures of fishing effects: human population density (Cinner, et al., 2009;Taylor et al., 2014) and an index of fishing pressure (Burke et al., 2011). Nor did we detect a MPA protection effect. However, when tested at a smaller spatial scale (Tanzania and Mozambique only), we found fishing pressure explained 6% of the variation in fish assemblages. Nevertheless, at this smaller spatial scale, reef geomorphology remained the overwhelming driver of differences between sites. Two factors may help explain the apparent lack of fishing effects. Firstly, by controlling for spatial correlation, we extracted the most unconfounded fishing signal possible.
Secondly, we integrated our surveys over a much greater depth range (0.5-33 m) than other studies (often <15 m) which increased the available biomass in our calculations. If reef productivity is to be measured in fish biomass/per area, we believe it is logical to include the full extent of the reef. In addition, these depths are accessible to fishing by coastal communities. Further, the ~30 m limit imposed by SCUBA diving already restricts accurate measures of target fishery species' biomass (Lindfield, Harvey, Halford, & McIlwain, 2016

| Fish biomass as an indicator of reef productivity
Total fish biomass is regularly used as an index of productivity on coral reefs and is a more sensitive indicator of fishing effects than density (Graham et al., 2017;McClanahan et al., 2016). However, not all reefs are equal in their ability to sustain high levels of biomass (Williams et al., 2015). Williams and co-authors attribute large differences in fish biomass in remote unfished reefs in the US Line Islands and NW Hawaii to variable oceanic productivity among locations.
In the WIO, we found ocean-exposed fringing reefs had total fish biomass values of 900-1,100 kg/ha, while lagoon-exposed fringing reefs and inner seas patch reefs yield moderate total biomass at ~500 kg/ha (Table 5). It is possible that these latter more weatherprotected reefs, and hence more accessible to artisanal fishers, may have reduced biomass due to higher fishing pressure. Smaller spatial scale comparisons are needed to separate these effects.
Nevertheless, very high biomass of >1,500 kg/ha was recorded at individual sites in Tanzania, Mozambique, and Madagascar ( Figure 6), including sites where there is fishing. These latter values are on a par with some sites in the Chagos Archipelago which is uninhabited and represents close to "pristine" biomass for the WIO (Graham, Pratchett, Mcclanahan, & Wilson, 2013;Samoilys et al., 2018). Our results suggest that the productivity of reefs in the WIO in terms of fish biomass depends on their geomorphology, exposure, and nutrient levels.
Comparing total reef fish biomass between studies can be problematic (Chabanet et al., 2016). For example, studies that include sharks will substantially inflate biomass values up, as illustrated by measures from two studies in Chagos ( Figure 6). We suggest the biomass of Epinephelinae, Lutjanidae, Lethrinidae, and Haemulidae (piscivores and omnivores) may be a useful metric, as these families contain widely exploited target fishery species in coral reef fisheries (Samoilys & Carlos, 2000) and have been regularly surveyed in UVC surveys in the WIO over the last 20 years . We recorded the highest biomass of piscivores/omnivores on ocean-exposed fringing reefs in Tanzania, some of which were protected from fishing within the Mafia Island National Marine Park, while similar reef types that were fished in Mozambique yielded half this value (Table 5). We propose that a ratio of piscivores/omnivores to total biomass may provide a useful metric of fishery exploitation and that values ~0.3 represent naturally productive reefs that are fished, and that higher values of ~0.4-0.5 might be achieved through protection within MPAs (Table 5). The latter compare with 0.44-0.52 recorded on atoll rim and lagoon sites, respectively, in the unfished Chagos Archipelago (Samoilys et al., 2018).

| Patterns in species assemblages
A relatively small number of species were significant in explaining the differences in fish assemblages between countries. From a sampled assemblage of 123 reef-associated fishes, just over 30% were significant, including ten species/taxa ubiquitous across the region, though with varying densities: Chaetodon melannotus and
The predominance of lower trophic level species among these taxa suggests bottom-up influences of detritus, algae, and small invertebrates are important.
Other species significant in delineating differences in assemblage structure are highlighted here because they may serve as useful indicators for reef type and health. These included the detritivorous Ctenochaetus striatus, which was most the abundant species throughout the region and is known to prefer low sediment levels on reefs (Tebbett, Goatley, & Bellwood, 2017a, 2017b (11 families), piscivores/omnivores (pisci/omni), and ratio of pisci/omni: total. Reef types: bb = bank barrier, bl = bank lagoon, cbrc = coastal barrier reef complex, isefr = inner seas-exposed fringing reef, isprc = inner seas patch reef, lefr = lagoon-exposed fringing reef, oefr = ocean-exposed fringing reef. Numbers in parentheses are total number of transects (replicates) per geomorphology per country. See Supporting Information Table S2 for mean values per site.
in Mozambique, possibly because parrotfishes are not targeted in Comorian artisanal fisheries (Freed & Granek, 2014). The potential for the species discussed here to serve as candidate biodiversity indicator species for monitoring coral reef health, for example, by the Convention on Biological Diversity (Pereira et al., 2013) is an important avenue for further research.
Climate change-induced coral bleaching is unquestionably one of the primary impacts occurring on coral reefs today (Hughes et al., 2018) with concomitant impacts on reef fish assemblages due to the loss of live coral . There is a substantial body of work that demonstrates clear relationships between reef benthos and the structure of reef fish assemblages with the food and shelter provided by coral invoked to explain these relationships.
The most widely understood relationship is between obligate corallivore butterflyfish species and the extent of live coral (Munday, Jones, Pratchett, & Williams, 2008) and two corallivorous chaetodons, Chaetodon meyeri and C. lineolatus, were among the significant species in delineating the differences in assemblages between countries. It is therefore surprising that none of the six benthic variables in our study were significant in explaining patterns in fish assemblages, even at the smaller scale of mainland Africa. In contrast, significant associations between live coral cover, recently dead coral, and rugosity and the structure of fish assemblages were found in the Chagos Archipelago (Samoilys et al., 2018). The different outputs from these studies support the view that the influence of benthic variables on fish communities is scale dependent and that confounding variables such as geography, geomorphology, and fishing pressure need to be controlled.

| CON CLUS IONS
The observed differences in the structure of fish assemblages across the WIO support the growing understanding that comparing reef fish assemblages across large spatial scales has to first take into account reef geomorphology, other reef structural attributes, and larger scale environmental drivers such as nutrient levels (Heenan et al., 2016;Mora, 2015;Taylor et al., 2014;Williams et al., 2015). Understanding the effects of fishing or loss of live coral from bleaching on fish assemblages needs to be examined where these larger scale variables are explicitly controlled for, which is often at smaller scales. We show a subset of species responded strongly to environmental conditions, though a large number did not. The role these significant species may play as biodiversity indicators for coral reefs, and the trophic dynamics of these assemblages are important avenues for future research. We suggest that variation in fish assemblages caused by the extent of reef area and coastline (Bellwood et al., 2005;Parravicini et al., 2013) is still poorly understood, and the eastern African coastline provides an ideal location for further research.
However, our models revealed 60% of the variation in fish assemblages remains unexplained. This highlights an urgent need to develop more spatially structured and controlled monitoring programmes with better measures of fishing pressure if we are to properly understand the influence of anthropogenic effects on coral reef systems.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R CO NTR I B UTI O N S
MAS conceived the ideas. MAS collected the data. AH led the data analysis. MAS led the writing. AH contributed to the writing. KO contributed ideas, data analysis, and writing.