2.6. Statistical Analyses
Distribution of seagrass and environmental variables were explored using multi-panel scatterplots and rank correlations to evaluate outliers, correlations, and skewness. Strong co-linearity (± 0.7 - 0.9) was identified between shoot and leaf densities, leaf length and width, and temperature and pH. These variables were excluded from multivariate analyses to avoid confounding the true degree of variability (Lipovetsky and Conklin 2001).
Diversity indices were assessed visually using histograms and q-q plots. Variances in the data were tested for normality (Shapiro-Wilk’s test) and homogeneity (Levene’s Test) and natural log transformed where necessary to meet assumptions for parametric analyses. Two-way factorial ANOVAs were used to test for main and interactive effects of site (random factor: five levels) and season (fixed factor: four levels) on each index: abundance (N ), richness (d) and diversity (H’) followed by Tukey’s HSD tests. Univariate analyses were performed using Sigma Stat (IBM SPSS Statistics 29).
Species abundance across seasons (fixed) and sites (random factor nested within season) was compared using multivariate data testing in PRIMER 6 (Clarke & Gorley, 2006). Abundance was considered an adequate metric on which to assess the effects of seagrass structure, epiphyte biomass and environmental factors since epifauna were largely comprised of grazers whose densities were important in controlling epiphytes (Hovel et al., 2002; Vonk et al., 2010). Data were 4th-root transformed to down-weight the influence of large variances. Visual assessments of relationships between variables in ordination space were made using non-metric multi-dimensional scaling (MDS) and cluster dendograms based on Bray-Curtis similarity matrices. Non-parametric multivariate analyses of variance were explored using PERMANOVA (Permutational Multivariate Analysis of Variance) (Anderson et al. 2008). Within group differences as well as comparisons of species composition between groups (i.e., beta diversity, Anderson et al. 2006) were tested using Permutational Analysis of Multivariate Dispersions (PERMDISP (9999 permutations)). The relative contribution of individual taxa to the similarity among sites across seasons was evaluated using SIMPER (similarity percentage breakdown).
To determine the environmental factors (temperature, salinity, turbidity, pH, oxygen, chlorophyll a and exposure) and seagrass metrics (shoot density, leaf length, aboveground biomass, algal epiphyte biomass) that best explained epifaunal variability, a distance-based linear model (DistLM, selection procedure ‘best’ and criteria ‘AIC’) was used. DistLM calculates the proportion of variability contributed by each factor using multiple regression modelling to partition variation according to selected predictor variables (Anderson et al. 2008). After all possible model combinations were explored, the five best models that explain variability based on smallest AIC values and adjusted R2 derivatives are presented.
To partition the net effects of environmental variables (exogenous) and seagrass structure (endogenous) on epifaunal abundance (endogenous) into direct and indirect effects, maximum-likelihood estimated structural equation models (SEM) were created in AMOS (IBM SPSS Amos version 26 Graphics). Structural equation modelling is a multivariate analysis framework that encompasses techniques derived from factor and path analyses (Grace, 2009). This framework enables the examination of both direct and indirect relationships between observed and unobserved (latent) variables. These relationships are represented by paths that indicate the statistical dependency, and the associated parameters specify the magnitude of the effect (direct or indirect) of independent variables on dependent variables (observed or latent).
A full model was specified based on theory and hypothesised relationships that (i) exogenous environmental variables: temperature, salinity, pH and oxygen would positively or negatively influence endogenous variables: seagrass biomass, leaf length/width, shoot density, epiphyte biomass, chlorophyll a (chl a ) and epifaunal abundance; while (ii) turbidity and exposure would negatively influence all endogenous variables; and (iii) seagrass density, leaf length/width, epiphytes and chl a would positively influence epifaunal abundance. Stepwise model selection was performed in which variables with non-significant paths i.e. regression coefficientsp > 0.05 were removed until all remaining paths were significant (p < 0.05) (Grace, 2009). Model fit was assessed using the chi-square value and the root mean square error of approximation (RMSEA) as measures of goodness-of-fit and evaluated by ensuring observed and predicted covariance matrices were aligned. Models were then adjusted to produce a low chi-square value with a corresponding p value > 0.05 denoting observed and fitted models were not significantly different, and an RMSEA < 0.08 indicating acceptable fit (Grace, 2009). The final model output presents standardized coefficients to compare the strength of direct and indirect effects.