The implementation of rare events logistic regression to predict the distribution of mesophotic hard corals across the main Hawaiian Islands

Organismal and environmental data

The Hawai‘i Undersea Research Laboratory (HURL) and the Pacific Islands Fisheries Science Center (PIFSC) provided video and photo records from MCEs in the Hawaiian Islands for our analyses. This imagery came from 19 dives conducted using submersibles, remotely operated vehicles (ROVs), autonomous underwater vehicles (AUVs), and tethered optical assessment devices (TOADs) in the ‘Au‘au Channel, Maui (13 dives) and two other geographically distinct regions: south O‘ahu (5 dives) and southeast Kaua‘i (1 dive). These dives were conducted between 2001–2013. We analyzed dive video using the Coral Point Count with Excel extensions (CPCe) tool (Kohler & Gill, 2006) in combination with a modified PIFSC 2011 mapping protocol (Pacific Islands Benthic Habitat Monitoring Center (PIBHMC), 2015). PIFSC has used this type of combined analysis, referred to as the random five point overlay method (RFPOM), to process coral reef ecosystem benthic imagery throughout the U.S. Pacific Islands Region since August 2011, and our use of it ensures database consistency with regions processed prior to this study. The CPCe software placed five points randomly on each snapshot, which we then assessed for coral presence. If any of the five points was on coral, that observation was recorded as a “presence.” In an effort to evaluate the accuracy of RFPOM, we counted all corals in 200 randomly selected screengrabs and found that this method misses 2.4% of coral observations recorded in these images. We categorized corals by genus, because both Montipora (Forsman et al., 2010) and Leptoseris (Luck et al., 2013) contain species complexes that remain the subject of taxonomic uncertainty which prevent us from being able to reliably identify corals to the species level from photographs.

We recorded snapshots every 30 s for the duration of each dive video. In addition to an existing database of 40,193 records from dives in the ‘Au‘au Channel, 3517 new snapshots were collected from the additional dives across south O‘ahu and Kaua‘i (Fig. 1). Of these 43,710 total images, 20,980 were discarded because either: (1) crucial observational data were absent; (2) they were redundant due to an extended stationary period; or (3) they fell outside the study depth range of 30–180 m. Of the remaining 22,714 records, we analyzed 2,757 unprocessed images using the RFPOM (Table 1).

We selected our environmental covariates, listed in Table 2, based on the sufficiency of the data and the potential significance of each variable as an indicator of hard coral habitat suitability (e.g., Dolan et al., 2008; Rooney et al., 2010; Costa et al., 2012). We defined summer and winter seasons as May–September and October–April, respectively (Kay, 1994; Rooney et al., 2010). We delineated significant wave height estimates and mean current velocities by season and direction. We extracted and averaged significant wave height data from 144 days per season of twenty-four hour PacIOOS Simulating WAves Nearshore (SWAN) regional wave models estimated values for 2011–2015 (see website: http://oos.soest.Hawaii.edu/las/). Mean current velocity values were available from 0:00–21:00 every three hours for all months from 2013–2015; for each season and direction, 48 mean current velocity values were extracted and averaged from the PacIOOS Regional Ocean Modeling System (see website: http://oos.soest.Hawaii.edu/las/). This model has a 4 km horizontal resolution with 30 vertical levels across seafloor terrain. We sourced monthly MODIS Aqua Chlorophyll a averages for the year 2012 from the NOAA PIFSC OceanWatch Live Access Server (see website: http://oceanwatch.pifsc.noaa.gov/). Using the Morel (2007) method, we applied the following cubic polynomial equation to obtain logged euphotic depth: (1)${\displaystyle {log}_{10}{Z}_{eu}=1.524-0.436x-0.0145x^2+0.0186x^3,}$ where x represents the measured Chlorophyll a concentrations (mg/m³) at sea surface. Euphotic depth is the depth at which the level of photosynthetically active radiation (PAR), a limiting factor for many heterotrophic mesophotic corals, is at 1% of surface PAR. In total, we used 14 environmental predictor variables to shape our model (Table 2) (Figs. S1–S5).

The spatial resolution of the bathymetry data was 50 m × 50 m for all islands. We resampled the bathymetry raster to a cell size of 25 × 25 m consistent with a conservatively estimated ±25 m positioning error margin observed at a depth of ∼800 m. We estimated an average camera swath value of 3.24 m (range 2.45–4.54 m) based on previous measurements from 19 still image screenshots taken when the submersible was located at different heights above the seafloor. Our geopositional error for the images is ±5 m and we can expect that the location data are within a circle with a 10 m diameter. Our observation sampling area is projected out from the location area a distance of ≤5 m. Addition of a conservative 5 m observation area buffer to the location error area produces an observational data position of ±20 m from the given coordinates of a data point.

We removed all subsampling within cells due to slight variations in camera angles or vessel speed through a point-to-raster conversion. We categorized all cells with ≥1 presence observation as “present” cells and all cells with only absence observations as “absent” cells. This removal of multiple observations within the same 25 × 25 m pixel effectively eliminates pseudoreplication within the data. We used ArcToolbox and the Benthic Terrain Modeler Toolbox to calculate slope, curvature, rugosity, and aspect (compass direction) values (Wright et al., 2012). We performed a spatial join based on proximity to observation point data to assign values for surface Chlorophyll a concentration, mean current velocities, distance from shore, and significant wave heights.

Regression methods

In describing the relationship between a response variable and one or more predictor variables, we use a logistic regression model because the response variable is dichotomous (Hosmer & Lemeshow, 2004). The ordinary logistic regression (OLR) model is defined as: (2)${\displaystyle \theta =\text{expit}\left(\mathrm{\mu }\right)=\frac{1}{1+exp\left(-\mathrm{\mu }\right)},}$

where θ is the probability that the species of interest is present (y = 1), and 1 − θ is the probability it is absent (y = 0). The logit function is the inverse of the expit function, and (3)${\displaystyle \text{logit}\left(\theta \right)=\mathrm{\mu }={\beta }_0+{\beta }_1{x}_1+\cdots +{\beta }_n{x}_n}$

is the linear sum of predictor variables, x₁, x₂, …, x_n, with intercept β₀ andregression coefficients β₁, β₂, …, β_n. In the language of generalized linear models (GLM), OLR is said to have the logit function as its link function and the expit function as its inverse link function. Logistic regression provides a straightforward, meaningful interpretation of the relationship between a dichotomous dependent variable y and a set of predictor variables (Allison, 2001).

Despite the popularity of OLR, it may yield extremely biased results when an imbalance exists in the proportion of the response variable data (e.g., such as in our case, when y = 0 ≫ y = 1) (Van Den Eeckhaut et al., 2006). King & Zeng (2001) coined the term “rare events logistic regression” to describe their corrective methodology in dealing with unbalanced binary event data:

• 1.

  The first step requires the selection of a representative sample. Though researchers generally prefer to work with more uniform response data (e.g., Liu et al., 2005), selection of an unusually high proportion of the rare event (in this case, y = 1) to “balance” the dataset and increase θ estimates will yield nonsensical results. We divided the data in half to create our training and testing datasets and checked that each set of observations had an approximately equal proportion ($\overline{y}$) of presence observations to better reflect the “true state” of the full dataset.

• 2.

  The second step rectifies any bias that might be introduced when dividing the dataset. This prior correction on the intercept (β₀) can be calculated as: (4)${\displaystyle {\hat{\beta }}_0={\tilde{\beta }}_0-ln\left[\left(\frac{1-\tau }{\tau }\right)\left(\frac{\overline{y}}{1-\overline{y}}\right)\right];}$

  here, ${\hat{\beta }}_0$ is the corrected intercept, ${\tilde{\beta }}_0$ is the uncorrected intercept, τ is the true proportion of 1s in the population; and $\overline{y}$ is the observed proportion of 1s in the training sample.

• 3.

  The third step rectifies any underestimation of the probabilities of the independent variables β_1…n from the substitution of the intercept value, obtained as: (5)${\displaystyle P\left(y_i=1\right)={\tilde{\theta }}_i+C_i,}$

  where the correction factor C_i is given by: (6)${\displaystyle C_i=\left(0.5-{\tilde{\theta }}_i\right){\tilde{\theta }}_i\left(1-{\tilde{\theta }}_i\right)XV\left({\tilde{\beta }}_i\right)X^{\prime },}$

  where X is a 1 × (n + 1) vector of values for each independent variable β_i, X′ isthe transpose of X, and $V\left({\tilde{\beta }}_i\right)$ is the variance covariance matrix. We obtained the improved probability estimates through estimation of β_i via ${\tilde{\beta }}_i$ , thereby considered “mostly” Bayesian (King & Zeng, 2001). Our priors in this case would be uninformative, which means that we lack sufficient knowledge to estimate the probability distributions of our data and our parameter of interest, θ. This is often the case when working with sparse ecological datasets. As the uninformative prior for a regression coefficient with domain (∞, − ∞) is uniform, a full Bayesian estimation with uninformative priors is equivalent to a traditional logistic regression. Therefore, this correction is effectively a correction to the approximate Bayesian estimator, and its addition improves our regression by lowering the mean squared error of our estimates. We implemented this rare events logistic regression using the ‘Zelig’ package run in R (Imai, King & Lau, 2008; Choirat et al., 2015).

We constructed a correlation scatterplot matrix per coral genus to observe correlation levels between all variables. In choosing which highly correlated variables to exclude from the analyses, we followed the criteria outlined by Dancey & Reidy (2004) and Tabachnick & Fidell (1996), who suggest a cutoff correlation value of 0.7. Only mean significant wave height parsed by season consistently overreached this threshold; the covariate that was least correlated with the response variable was removed. We excluded predictors that lacked a clear distribution pattern or correlated minimally (<0.05) with the response variable.

One of the more studied habitat preferences of Leptoseris and Montipora is the influence of depth on their distribution (Rooney et al., 2010; Costa et al., 2012; Kahng et al., 2010). Increasing depths often correlate with greater distance from shore. The inclusion of squared terms (e.g., $x_2=x_1^2$) in our regression equation expit(θ) = β₀ + β₁x₁ + ⋯ + β_nx_n permits the logistic curve to reflect the bell curve shape expected in plotting the distribution of these animals across a range of depths or distance from shore. In order to account for these trends, we added Depth Squared and Distance Squared as potential variables for consideration in our final model. As depth or distance increases, its square increases even more rapidly, allowing the squared term to eventually dominate and “pull down” the probability curve.

We withheld 50% of our information per genus as testing (i.e., validation) data. Using the remaining 50% (our training data), we performed the rare events corrected logistic regression described above. Using an exhaustive iterative algorithm (Calcagno, Mazancourt & Claire, 2010), we modeled all possible combinations of included covariates. We ranked models using the corrected Akaike information criterion (AICc) (Hurvich & Tsai, 1989), which is considered an excellent comparative measurement of model strength, especially for sparse datasets. For both genera, the models with the lowest (lowest = best) AICc scores were lower than the “second best” AICc scores by at least 2 (i.e., Δ AICc ≥2), indicating strong preference for the best model (e.g., (Hayward et al., 2007)).

In an ideal and unrealistic study, all biotic and abiotic components of a model would be homogenous and evenly distributed across a sampling space. Our sampling design includes overlapping submarine dive tracks and the inherent heterogeneity of the marine environment, which could problematically violate our model’s underlying assumption regarding the independence of our biological and environmental data. We removed all instances of pseudoreplication (multiple observations in one grid cell) when we assigned each grid cell to a category of “corals present” or “corals absent.” After we removed subsampling within our observational data, we checked for the presence of clustering, or spatial autocorrelation, within these data. Uncorrected spatial autocorrelation between observational data points confounds and undermines any biological inferences drawn from model predictions.

We checked small-scale, local spatial autocorrelation using Geary’s C statistic (Geary, 1954), based on the deviations in the responses of observation points with one another: (7)${\displaystyle C=\frac{n-1}{2S_0}\frac{\sum _i\sum _j{w}_{ij}{\left(x_i-x_j\right)}^2}{\sum _i{\left(x_i-\overline{x}\right)}^2}.}$

Here, x is the variable of interest, i and j are locations (where i ≠ j), w_ijrepresents the components of the weight matrix, and S₀ is the sum of the components of the weight matrix. Geary’s C ranges from 0 (maximal positive autocorrelation) to 2 for high negative autocorrelation. In the absence of autocorrelation, its expectation is 1 (Sokal & Oden, 1978).

We also examined global spatial autocorrelation using Moran’s I statistic, which measures cross-products of deviations from the mean (Moran, 1950): (8)${\displaystyle I=\frac{n}{S_0}\frac{\sum _i\sum _j{w}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{\sum _i{\left(x_i-\overline{x}\right)}^2}.}$

Moran’s I values generally range from –1 to 1, with 0 as the expectation when no spatial autocorrelation is present.

We also verified the spatial independence of our observational point data using a semivariogram, which is a graphical method of quantifying spatial correlation in a set of points (Figs. 2–3). We selected our theoretical semivariogram to fit the empirical semivariance using the ordinary least squares (OLS) method (Jian, Olea & Yu, 1996; Kendall et al., 2005). The spherical model had the best quantitative fit based on OLS estimates (Table 3). For each dataset, the low thresholds at which semivariance stopped increasing indicated the almost complete absence of spatial autocorrelation for each genus.

Model assessment

Evaluation of the rare events logistic regression model output is more complicated than for the typical linear model. For example, R² values, although calculated, have little applicability to logistic regressions and are therefore ignored (Menard, 2000; Peng, Lee & Ingersoll, 2002). Sample size and selected threshold largely influence the results of the Hosmer and Lemeshow goodness-of-fit test (Hosmer et al., 1997). Accordingly, we use model classification accuracy as a second measure of goodness-of-fit (in addition to ΔAICc). We want to maximize true positives (TP) and true negatives (TN) while minimizing false positives (FP) and false negatives (FN). The sensitivity-specificity sum maximization approach (Cantor et al., 1999) therefore maximizes (9)${\displaystyle S{S}_{\max }=\frac{TP}{TP+FN}+\frac{TN}{TN+FP},}$

which is equivalent to finding the point on the ROC (receiver operating characteristics) curve at which the tangent slope is 1, indicating the optimal cutoff point at which “cost” (here, the number of FN and FP) and “benefit” (the number of TN and TP) is balanced. We chose this technique because we aim to identify regions devoid of hard corals as well as regions deemed potentially suitable for habitation.

ROC curves plot the true positive test rate against the false positive test rate across different theta cutoff points (Hadley & McNeil, 1982). We calculated values for sensitivity and specificity for threshold increments of 0.005 ± 1 standard deviation of the rounded mean for each model. Because each theta threshold value varied based on the genus and model, the threshold-independent area under the curve (AUC) test statistic best reflects the predictive accuracy of the model.

In addition to creating ROC curves, we also took into account the overall prediction success of each model, given as: (10)${\displaystyle OPS=\frac{TP+TN}{TP+TN+FP+FN}.}$

Overall prediction success is a measure of total correct classification of both present and absent observations. While this is a good final assessment of model classification error, consideration of the prediction success alone is not a viable evaluation method when binary data is highly imbalanced, as a value given by this method may primarily represent model success in identifying the most common observation type (Fielding & Bell, 1997). We plotted our sensitivity and specificity values on a ROC curve to show how each model performed relative to chance (Fig. 4). All models fall in the range 0.7 ≤ AUC < 0.9, which indicates good discrimination and reliability of model predictions (Hosmer & Lemeshow, 2004).

We also created maps of individual and summed predicted occurrence probabilities of both coral genera across the main Hawaiian Islands and ran a hotspot analysis using the ArcGIS Getis-Ord $G_i^{\ast }$ Hotspot Analysis tool. We constructed a polygon fishnet composed of 1 × 1 km cells which encompassed all islands. We summed each 25 × 25 m raster cell value for probability of Leptoseris occurrence and probability of Montipora occurrence. We performed a spatial join of raster cell values within each polygon for an average value of summed probabilities. The Getis-Ord $G_i^{\ast }$ statistic identifies clusters within these polygons that display values higher in magnitude than random chance would permit. The Getis-Ord local statistic is given as: (11)${\displaystyle G_i^{\ast }=\frac{\sum _{j=1}^n{w}_{i,j}{x}_j-\overline{X}\sum _{j=1}^n{w}_{i,j}}{S\sqrt{\frac{1}{n-1}\left[n\sum _{j=1}^n{w}_{i,j}^2-{\left(\sum _{j=1}^n{w}_{i,j}\right)}^2\right]}}.}$

Here, w_i,j represents the spatial weights between features i and j; n represents the total number of features;x_j is the attribute value for feature j; $\overline{X}=\frac{1}{n}{\sum }_{j=1}^n{x}_j$; and $S=\sqrt{\frac{1}{n}{\sum }_{j=1}^n{x}_j^2-{\left(\overline{X}\right)}^2}$.

Important environmental covariates

Predicted Montipora presence peaks at about 60 m (median occurrence probability = 7.5%); Leptoseris presence peaks at about 100 m (median occurrence probability = 7.5%). These predictions are consistent with the inferences of Rooney et al. (2010), which separates mesophotic reefs into three distinct depth sections: upper (30–50 m), branching/plate dominated (50–80 m), and Leptoseris dominated (≥80 m). The depth at which suitability peaks for Leptoseris occurs at a range where steep ridges and drop-offs are plentiful in our study region, and therefore the mean preferred depth may be prone to slight overestimation.

In addition to depth, four environmental covariates appeared to influence the distribution of Leptoseris: rugosity, slope, summer mean current velocity (northward), and winter mean current velocity (eastward). Scleractinians easily colonize environments that are relatively calm and rugose due to the larger amount of available surface area, and this positive correlation was reflected in our model. Leptoseris habitat preference was also positively associated with slope, which was not observed for Montipora. Corals that inhabit the upper mesophotic zone may be more susceptible to damage from debris displaced by high wave energy, and are therefore less likely to colonize steep slopes (e.g., Harmelin-Vivien & Laboute, 1986; Bridge & Guinotte, 2013). The deeper distribution of Leptoseris may protect it from damage related to wave intensity, allowing it to colonize slopes (e.g., White et al., 2013). Another possibility is that the model is picking up drop-offs from masses accreted during the last glacial maximum. These steep drop-offs are present between 90–120 m in the Leptoseris-dominated lower mesophotic zone (Yokoyama et al., 2001; Webster et al., 2004).

Leptoseris also favors well-flushed areas exposed to year-round moderate current flow (i.e., up to 0.3 m/s). The plate-like morphology of Leptoseris corals effectively boosts sunlight capture by its symbiotic zooxanthellae and zooplankton capture by the corals themselves, but it also makes the coral vulnerable to smothering by sediment accumulation (Bak, Nieuwland & Meesters, 2005; Bongaerts et al., 2010; Marcelino et al., 2013). The success of Leptoseris corals in areas of moderate current flow may be related to the improbability of sediment settlement and accumulation. While the model did not capture the same effect of current flow on Montipora distribution, we recognize that the morphology of some Montipora species is extremely similar to that of Leptoseris. We do not expect either genus to readily colonize highly turbid regions, especially given that certain species of heterotrophic Montipora are thought to exploit strong currents to meet their energy requirements (Grottoli, Rodrigues & Palardy, 2006; Rooney et al., 2010).

Substrate hardness, a variable known to influence coral colonization, was notably absent from each model. Substrate hardness values were derived from acoustic backscatter imagery readings. The base resolution of these readings (50 m × 50 m) was not sufficiently detailed for purposes of this analysis. We noted plentiful coral colonization along larger surfaces like lava fingers, the hardness of which would be detectable by backscatter surveys, as well as across small rock fragments strewn across a sand flat, which would be obscured by the softness of the surrounding benthos. We can conclude that measurements of benthic hardness are not detailed enough for predictive modeling purposes at a 25 × 25 m resolution.

We emphasize that the purpose of this study was to build a pan-Hawai‘i predictive habitat map for two dominant coral genera within the mesophotic zone. Because the scope of this study included all main Hawaiian Islands, we were constrained by the coarseness of available full-coverage environmental data. As we build on this analysis, we plan to use our maps to identify areas of interest for further study at higher resolution and to include additional variables currently only available in certain regions, such as light intensity and temperature at depth. For example, our predictive and hotspot maps identify Penguin Bank (southwest Moloka‘i) as particularly suitable for Leptoseris colonization, which has not been verified by video or photo records. High resolution backscatter data (1 × 1 m) exist for this region, and incorporation of these data into new analyses of subsets of our study area may refine our conclusions.

Error sources and model reliability

We examined two types of error (false negatives and false positives) and analyzed our models without giving preference to either one. This approach is widely accepted as the best method of overall error minimization (e.g., Liu et al., 2005; Fielding & Bell, 1997). Rare events corrected models for both Leptoseris and Montipora achieved levels of specificity and sensitivity well above the null, indicating good predictive power. Additionally, both models attained about 74% overall prediction success. We assumed coral detectability was constant across the study region and that we can therefore consider the true absence observations to be reliable indicators of a potentially unsuitable habitat for corals. For each genus, the model tended to slightly over-predict presence observations; large numbers of false positives lowered sensitivity values. This is inevitable in the analysis of severely imbalanced or sparse binary data; the ongoing addition of presence observations to the dataset will improve overall model classification accuracy.

While the consistent identification of southern coastal areas as suitable is reliable, the comparatively infrequent selection of northern coasts is likely due to the source of the model-building observations. The vast majority of mesophotic exploration has been along southern coastlines, which is often where waters are calmest in Hawai‘i. It is speculated that because mesophotic corals are more shielded from winter long-period wave energy than their shallow water counterparts, they are able to flourish at depth along northern coastlines (Grigg, 1998; Rooney et al., 2010). The addition of data sourced from northern expeditions would likely improve predictive power of the model across north-facing coastlines (Alin, 2010).

We acknowledge that the original data were not collected in a standardized fashion (e.g., variation in vessel traveling speed or differences in data collection vessel and/or quality). Our careful exclusion of overlapping observation points within each 25 × 25 m rectified this sampling design flaw as much as possible and eliminated pseudoreplication.

Distinctions between coral genera

Our Montipora model was simpler than the Leptoseris model in that the only variable included other than depth was winter significant wave height. Though uncertainty was highest at lower values of significant wave height, Montipora demonstrated a preference in colonizing habitats that experience lower significant wave height during winter. This preference contrasts with Montipora species in shallow waters that were more likely to be observed in higher wave height environments (Franklin, Jokiel & Donahue, 2013). This likely influenced the inability of the model to identify any suitable habitat around Ni‘ihau, where the average winter significant wave height equaled 1.78 m, almost double the mean significant wave height of our model training data (0.91 m). Though mesophotic corals are generally thought to be exempt from the growth limitations faced by shallow water corals in regions of high wave energy, prolonged wave intensity has been shown to negatively affect the colonization of upper mesophotic scleractinians, especially in sloping areas prone to debris avalanches (Bridge & Guinotte, 2013; Kahng, Copus & Wagner, 2014). Continuation of this work might include a more in-depth examination of the relationship of this coral genus with the combined effects of slope of available substrate and exposure to wave energy.

We found no records of Montipora presence when processing our O‘ahu dataset, which probably contributed to the very low predicted mean probability of Montipora occurrence there (0.1%). We believe this is due in part to the sampling pattern across south O‘ahu; we recorded 62.3% of all observations processed for this region at a depth of 75 m or greater. Montipora prevalence is greater in the upper-to-middle mesophotic zone, and the relative deepness of the O‘ahu dives likely influenced their nonappearance in this portion of the dataset. We emphasize that the dearth of Montipora observations around O‘ahu is an artifact of the dataset we used to construct our model; Montipora corals have been observed in mesophotic depths across O‘ahu (e.g., Fig. 4B, Rooney et al., 2010). The mean significant wave height across the mesophotic zone was lower across the southern and western coasts (1.50 m) than that observed across the northern and eastern coasts (2.37 m) of the island. As at Ni‘ihau, we assume that this high northern and eastern average height, coupled with the absence of Montipora presences in O‘ahu in the training dataset, greatly impacted our model’s ability to detect areas of suitable habitat around the island. The results of our Getis–Ord $G_i^{\ast }$ Hotspot Analysis corroborate the findings of Costa et al. (2015), who used Maximum Entropy software to predict the highest occurrence probability of Leptoseris and Montipora in the middle and mid-coastal ‘Au‘au Channel, respectively (Costa et al., 2015).

The factors influencing the distribution of coral species in shallow and mesophotic habitats differ. One of the fundamental drivers of the occurrence and abundance of coral species on shallow reefs in Hawaiian waters is wave stress (Dollar, 1982; Grigg, 1983; Franklin, Jokiel & Donahue, 2013). Given the depth range of MCEs, wave stress is unlikely to serve as a direct influence on coral occurrence but may provide secondary effects as wave events lead to debris reaching MCEs (Kahng, Copus & Wagner, 2014). Furthermore, the decoupled effects of environmental drivers on shallow and mesophotic zones extend between the islands. In shallow reef communities Montipora species become relatively more dominant from Hawaii Island to Ni‘ihau (Franklin, Jokiel & Donahue, 2013), but appear to peak in occurrence in the mesophotic zone of Maui Nui. While strong environmental drivers influence the distributions of shallow corals, the occurrence patterns of mesophotic corals may reflect a more stable environment with an increased influence of biotic factors such as interspecific competition in a habitat zone with limited light and space resources available.