*2.5. Statistical Analyses*

All analyses were conducted in R version 4.0.4 [92] and PRIMER7 [93,94]. Three-way repeated measures permutational multivariate analysis of variance (PERMANOVA) with 9999 permutations were conducted to test for significant effects between sites (Rural-N, Rural-E, Urban-W, Urban-E), depths (5 m, 10 m), and years (2015, 2017) on a Bray–Curtis similarity matrix of benthic cover categories, prevalence of disease and compromised health, and a Bray–Curtis similarity matrix of the number of colonies per coral genera (the count of coral genera per belt transect), all square root transformed [93,94]. All PERMANOVAs were also tested for homogeneity of dispersion akin to the homogeneity of variance in univariate tests [93]. Principal Coordinates Analysis (PCoA) was run on the same similarity matrix of coral genera to visualize coral community structure. Repeated measures analysis of variance (ANOVA; Anova in the car, emmeans, and nlme R packages) [95–97] was used to test the transformed hard coral abundance, the categories of disease and compromised health (only the bleached category was transformed), and the transformed Shannon diversity index of coral genera for significant effects between sites, depths, and years. All transformations were square root. A repeated measures ANOVA was also conducted on the log-transformed number of acroporids per transect between site, depth, and year. Normality was visually inspected (hist, qqplot, qqnorm, and leveneTest in the car package).

Nutrient data were only collected in 2015, with which a two-way ANOVA with factors site, and depth was performed on the seawater nutrient data, including DIN (transformations: log—NH4 <sup>+</sup> and DIN; square root—NO3 − + NO2 −). A two-way ANOVA was used to test for significant differences in δ15N for each of the two genera of algae, *Halimeda* spp. and *Chlorodesmis* spp., with the factors site and depth. Only three samples of *Chlorodesmis* spp. were collected on a single transect at Rural-E and these were removed from the analysis. Variables were visually inspected for normality and tested for homogeneity of variance using Levene's test (leveneTest). Percent nitrogen was log-transformed for *Halimeda* spp. Post hoc tests were conducted (multcomp and emmeans R packages) for *Halimeda* spp. and *Chlorodesmis* spp.

The monthly mean temperature was calculated from the 24 h daily maximum temperature obtained from both datasets (in situ temperature logger data and remotely sensed CRWTL data). A one-way repeated measures ANOVA was used to test temperature logger data (pooled by site) to test for differences in the monthly means between sites. A two-way repeated ANOVA assessed seasonal (summer January–March, fall April–June, winter July–September, spring October–December) and methodological (in situ loggers, remotely sensed SST from CRWTL) differences between the monthly temperature means. Both analyses employed random intercept models with residual autocorrelation structures to account for temporal autocorrelation of individual temperature loggers and CRWTL measurements (nlme R package). Post hoc tests were conducted using the emmeans R package. To assess levels of thermal stress, remotely sensed DHWs were retrieved from CRWTL online [71].
