Correction to “Predicting intraspecific trait variation among California's grasses”

Abstract

Sandel, B., Pavelka, C., Hayashi, T., et al. (2021) Predicting intraspecific trait variation among California's grasses. Journal of Ecology, 109, 2662–2677. https://doi.org/10.1111/1365-2745.13673.

In the paper by Sandel et al. (2021), an error has been identified in the code.

The error was in generating the testing data subset for assessing random forest fit, causing it to not be independent of the training dataset. This affects Figures 3-5, and the corrected versions of these are included below. Table S3 has been updated in the article. The updated text referring to these figures in the section ‘3.2 Modelling ITV’ is also included below. The changes do not fundamentally alter the message of the paper.

3.2 Modelling ITV

Across all specifications of the random forest models, performance scores were very similar on the training and testing data subsets (on average, differing by <0.09, Table S3), suggesting little overfitting. When applied to the testing dataset, random forests containing all five predictor groups predicted values that were well correlated with the observed trait values (for delta-SLA: 0.74, SLA: 0.82, delta-Height: 0.67, Height = 0.88, delta-LA: 0.72, LA: 0.89, Table S3). Across all subsets of variable groups, other local traits (values of the non-focal trait from the local population, e.g. when predicting SLA, the Height of the plants) and climate were the most important groups for model performance (Figure 3). Species mean traits, name and phylogeny had smaller contributions to model fit. The performance of one such random forest, excluding the species predictor variable, is shown in Figure 4. The correlation between observed and predicted values is strong for both training and testing datasets. However, the observed–predicted relationships deviated somewhat from the 1:1 line, particularly for the delta-trait predictions. Standard major axis (SMA) regression slopes were less than 1, ranging from 0.65 to 0.73 for delta-trait models and 0.82 and 0.92 for the final local trait predictions. These deviations indicate that these models tend to predict less extreme values for the most extreme trait observations.

A model including other local trait measurements and species names would be of limited use for predicting trait values of a plant in an unmeasured location. In contrast, the climate of that location is readily available, and phylogenetic relationships are known for most species. Thus, we focused on a reduced model including just these two variable groups and two species-level traits: the species mean value for the focal trait and its life span. Removing species names from the model had little impact (Table S3), but removing other local traits reduced model performance (Figure 5). For example, predicted-observed correlations for SLA, height and LA dropped to 0.79, 0.88 and 0.89. This likely reflects the fact that other local trait measurements can provide insight into local conditions that are not captured by our two broad climate predictors. Despite this modest reduction, model performance for this simplified model was still fairly high.

We apologise for this error.