Dataset: RI Functional Trait Data

Generalized additive model fitting
We fit generalized additive models (GAMs) to the 13 functional traits (response variables) with species or length as predictor variables to hindcast the functional traits of fish from past surveys collected by the Rhode Island Department of Environmental Management. GAMs model a response variable (e.g., functional trait) as the sum of nonlinear functions from different predictor variables such as species and fork length (Hastie and Tibshirani 1990). We utilized thin-plate penalized regression splines, which adds a penalty to the smoother function to avoid overfitting (Wood and Augustin 2002). Penalty weights were optimized using the generalized cross validation score which minimizes the root mean square error of the model fit to the data and balances model complexity with goodness-of-fit. To run all models, we used the function ‘gam’ of the mgcv R package (R Development Core Team 2013). We used the GAMs fit to each functional trait (Appendix S3 for model fits) to hindcast the functional traits of 30 fish species for which we had fork length data across 6 years and 6 ponds of survey data. This generated functional trait information for a total of 81,337 fish.

Intra- vs. interspecific trait distance analysis
We measured the Bray-Curtis dissimilarity within (intraspecific) and across (interspecific) species within functional trait space by constructing a dissimilarity matrix across every species-length combination using the function ‘vegdist’ within the vegan R package. We then tested for mean differences in intra- vs. interspecific distances across the total community using a two-sample T- test and differences in their distributions via a Kolmogorov-Smirnov (ks) test.

Functional diversity analysis
We examined three well-studied functional diversity metrics which measure diversity in multivariate trait space and can accommodate both continuous and qualitative traits (Villéger et al. 2008): Functional Richness (FRic), Functional Evenness (FEve), and Functional Divergence (FDiv). For each community, all species were plotted in functional trait space. FRic measures the total trait space filled by the community and is the area of the ‘convex hull’, or minimal polygon drawn to encompass all species and is scaled by total FRic across all communities. FEve is the evenness of abundance across trait space and is measured as the linearized distance of minimum spanning tree, which is a tree that connects all species in trait space with the minimum sum of branch length. Lastly, FDiv measures how community’s abundances are distributed in the volume of trait space occupied. This metric is calculated as the deviations from the mean circle drawn though the species in trait space (which represents the mean distance to the centroid for each community). The higher abundance that is greater than the mean, the larger the FDiv. We used the Distance-based Functional Diversity Indices, ‘dbFD’, function in the FD R package (R Development Core Team 2013) to calculate all three metrics across all pond communities in two ways: (1) with traits averaged across species (as is most common in the literature; Mouillot et al. 2007, 2013, Villéger et al. 2010, 2012, Albouy et al. 2011, McLean et al. 2019) and (2) with intraspecific variation included. When using species averages, the functional trait by species matrix and the abundance by site matrix had a row and a column, respectively, for the species average. In contrast, when intraspecific variation was explicitly included, the functional trait matrix and the abundance matrix had a row and column, respectively, for each species-length combination.

For each functional diversity metric, we used simple regression to determine whether the slope relating the functional diversity metric with vs. without intraspecific trait information was significantly different from 1. Specifically, we ran a forced linear regression (intercept of zero) between the functional diversity metrics for each community calculated with vs. without intraspecific information and then determined whether the slope of the regression differed from 1 using a one-sample t-test:  t- slope-1SEslope  . A significant slope means that functional diversity computed with intraspecific variation differs from that computed without intraspecific variation (i.e., including intraspecific trait information significantly alters the functional diversity metric).

Ecosystem function analysis
We calculated ecosystem function as the total biomass of the fish community within each pond and year. As with the functional trait dataset, we constructed a GAM with species and length as explanatory variables to calculate biomass from past survey data, based on the biomass (g) that was measured at the same time as the functional trait data collections. We then summed total biomass for each unique pond-year combination to generate an estimate of ecosystem function for each pond community through time. Using simple regression, we tested for relationships between FRic with or without intraspecific trait information and ecosystem function comparing the fit and slope between the two models.

BCO-DMO processing description:
- Converted dates to YYYY-MM-DD format
- Adjusted field/parameter names to comply with BCO-DMO naming conventions
- Missing data identifier ‘NA’  and '#DIV/0!' replaced with 'nd' (BCO-DMO's default missing data identifier) 
- Added a conventional header with dataset name, PI names, version date
- Replaced positive "Narrow River" longitude values with negative values to correct for error in original spreadsheet