2025-08-18 SMA-03-000007 Testimony Maki Morinoue

Home

Article pollution,urban runoff,nutrient loading,phytoplankton biomass and P(y.<j) peak rainfall.A fourth-root transformation was applied to sediment. In P(yt>j) -Cj+Blzil+"'+Bkzik To reduce model overfitting,Pearson's correlation coefficients were calculated among all predictors(Supplementary Fig.5),removing one of each pair of highly correlated(r>0.7)predictors.To further Here,i indexes each of N observations,with categories yi,and the strive for model parsimony,we a priori excluded human population left-hand side of the equation is the logit of the probability of a density from the model-fitting process as it was a poor indicator of reef-bull der category ofjor lower,forj=1(high)or2(moderate).Reefs human-driven land-to-sea impacts on local scales(Figs.1c and 2d and with low reef-builder cover contributed to the regression through cal- Extended Data Fig.3).We also excluded browser biomass as they rep- culation ofthe log odds.Each Cjis an MLE-computed model intercept, resented less than 10%on averageoftotal herbivore biomassacross all and each Bk is the MLE coefficient corresponding to the standardized reefs before,during and postdisturbance.This resulted in thefollowing independent variable zik,for k=1 through n,where n is the variable predictors included in the models(correlated predictors in parentheses number of predictors used in a given candidate model,hence the were removed):total fish biomass,biomass of scrapers,biomass of ellipsis(...).Afundamental component of this model isthe assumption grazers(total herbivore biomass),DHW(SST mean and variability), of proportional odds,or parallel regression,which indicates that Bk wastewater pollution,nutrient input,urban runoff,sediment inputand values are independent of the logitlevelj.The validity of this parallel peak rainfall(annual rainfall correlated with both),wave power,phy- regression assumption was ascertained using Brant's Wald test97,as toplankton biomass(irradiance),fishinggear restrictions and depth. well as a likelihood ratio test(a=0.05). The decision of which correlated predictors to retain was based on a We then calculated the following predictors based on current litera- hypothesis-driven approach,in part whether the given predictor had ture and our hypotheses of the principal factors that drive changes in the potential to directly(for example,sediment input)rather than reef-builder cover across space and time following a major thermal indirectly(for example,annual rainfall driving sediment input)affect disturbance:fish biomass metrics,wastewater pollution,nutrient heat-driven coral loss. loading,urban runoff,annual rainfall,peak rainfall,wave exposure, We incorporated a random spatial factor to account for the possible phytoplankton biomass and irradiance:the mean of all data from 2016 influence of a change in an underlying variable along the coastline not to 2019;sediment was measured as the mean of top three events over quantified in this study.This was done by breaking the coastline up the 2006-2019 time period;SST mean and SST variability:mean of all into discrete 10 km sections running north to south.Section size was data from 2000 to 2018.Notethat2019 was excluded in SST mean and determined using hierarchical clustering based on pairwise Euclid- SSTvariability owing to the marine heatwave that affected Hawai`i21,but can distances between reefs and identifying an inflection point in the occurred after our 2019 fish and benthic surveys;fishing gear restric- intragroupvariance24(SupplementaryFig.7).WefittedGAMMsfor tions involved the marine managed area designation in 2016 and depth all possible candidate models(unique combinations of the predictor was assessed by in-water diver-assessed values. variables)using the UGamm wrapper function,in combination with We used the same process as in the GAMManalysis to removeoutliers the dredge function in the MuMln package94.Nonlinear smoothness in in our predictor variables(above).We then square-root transformed the models was determined using penalized cubic regression splines, the following predictors to down-weightthe influenceof values atthe with the numberof knots(limited to fourto reduce overfitting)spread extreme ends of their distributions:total fish biomass,wastewater evenly throughout each covariate.All possible candidate modelswere pollution,sediment inputand nutrient loading.Pearson's correlation computed(unique combinations of the predictor variables)but limit- coefficients were calculated among all predictors(Supplementary ingthe total number of predictors in anygiven candidate model to five Fig.8),removing highly correlated(r>0.7)predictors.For the rea- to reduce overfitting.We used Akaike's information criterion with abias sons outlined in our GAMM analysis and for continuity,we a priori correction for small sample sizes95(AICc)for model comparison and excluded human population densityand thebiomass ofbrowsers from all models within OAlCc<-2ofthetop model(DAICc=0)are presented the model-fitting process.This resulted in the following predictors in Extended Data Table 2.To visualize the effectof predictor terms on included in the models(correlated predictors in parentheses were coral cover change,we averaged the coefficients fromthe top models removed):total fish biomass,biomass of scrapers,biomass ofgrazers (that is,AAICc<_2)to generate a predicted dataset and set all other (total herbivore biomass),wastewater pollution,nutrient input,sedi- predictor terms to their median value.Finally,we calculated a meas- ment input,urban runoff(phytoplankton biomass),wave exposure, ure of predictor variable relative importance within each candidate fishing gear restrictions and depth.The decision of which correlated model by calculating the sum of AlCc model weightsforeach predictor predictors to retain followed the same logic asour GAMManalysis.The (that is,the sum of model weights across all models containing each mean and variability in SST were excluded given the negligible range predictor;Fig.3). of values among reefs(0.1 and 0.025°C,respectively).All possible candidate models were computed while limiting the total number of Coral reefs four years postdisturbance.Our goal was to assess the predictors in anygiven candidate model to four(to reduce overfitting local land-sea human impacts and environmental factors that best and to account for the lower response variable replication compared explained variations in the cover of reef-building organisms four years to our GAMM analysis).Models were computed using the multino- following the marine heatwave.The cover of reef-building organisms mial logistic regression function mnrfitinMATLAB.We again used for reefs surveyed in 2019(n=SS)were parsed into three categories AICcfor model comparison and all modelswithin DAICc<-2 of thetop on the basis of thefollowing percentiles:low,less than or equal to the model(DAICc=0)arepresented in Extended Data Table3.McFadden's 25th;moderate,more than 25th and lessthan75th;and high,morethan pseudo-R2 was computed for the highest ranked models and ranged or equal to the 75th.We then performed ordinal logistic regression96 from 0.21 to 0.22.Unlike traditional R2 values,McFadden's pseudo-R2 to determine the probability of a given reef having high,moderate or of more than 0.2 represents an excellent fit98.Models Within OAICC<<-2 lowcoverof reef-building organisms onthebasisof the prevailing local of model 1 in Extended Data Table3 demonstrated comparable levels of human impacts and environmental factors(that is,predictor variables; goodness of fit andparsimony""O Many ofthe parameter coefficients Extended Data Table 1).Logit models are multivariate extensions of within these models were sensitive to the underlying variability in the generalized linear regression models that provide parameter estimates dataandtheirestimates did notdiffer significantly from zero(P<0.05). by means of maximum likelihood estimation(MLE)to model the rela- Thetop model contained parameters with covariate estimates signifi- tive log odds of observing a reef-builder cover category or less versus cantly different from zero,namely scraper biomass and wastewater observing the remaining higher categories: pollution.Using model 1,we examined changes in the probability of a