Computational protocol: Predicting the establishment success of introduced target species in grassland restoration by functional traits

Similar protocols

Protocol publication

[…] For each site, we calculated the relative introduction rate of target species as the number of species found during the whole observation period divided by the total number of introduced species in 2009 to show the overall potential of the applied restoration methods. Spearman‐correlation coefficients were used to test for intercorrelations among plant traits and for highly correlated traits such as plant height and releasing height just one was used. To test the hypothesis that plant species traits affect the establishment success of the species, we performed a two‐step approach. In the first step, we modeled the occurrence of the species as a proxy for their establishment success with linear mixed effect models across all years at the plot level, using the presence of each species in each of the 12 plots per site as response variable and a logit link function with binomial error distribution. We used the GLIMMIX procedure of SAS (SAS Institute Inc., Cary, NC, USA), approximating maximum likelihood with Laplace's method and deriving the degrees of freedom of the error terms with the containment method. Fixed factors were (1) restoration method and year, as well as their interaction, while the species, block, plot, and site were included as random effects (no‐trait model). Year entered the model as continuous variable. Afterward, to omit noninformative traits, we included trait by trait (2) species traits as further fixed factor, keeping all other components of the preceding model (1) (single‐trait models), except replacing species as random factor with species × trait interaction, which has identical levels as species because traits did not vary within species. However, we used the species × trait interaction instead of species only to inform SAS that traits have to use species as error term, thus making sure that the correct degrees of freedom were obtained in the subsequent ANOVA tables. However, there was no difference in the amount of variance explained by species × trait interaction and species. We also calculated models with design‐specific variables that described (3) the presence of a species at the donor site, or (4) the presence in the seed mixture and treated this variable like a trait, also keeping all other components of the preceding model (2). The significance of the predictors was tested with type III ANOVA. All models were ranked by Akaike's information criterion (AIC) and all models of (2)–(4) performing better than that of (1) (the no‐trait model) were subsequently examined for effects.In a second step, we analyzed the combined effect of traits on establishment at the species level. Principal component analysis (PCA) was employed to identify the main gradients of trait states and trait values in all species. Post hoc correlations were used to relate PCA axis scores to the establishment rates of every species averaged across restoration methods for each of the observation years. The analyses were performed using the vegan package (Oksanen et al., ) in R. We also used the averaged establishment rates to analyze the combined effect of several species traits. Across all treatments and blocks (i.e., all 12 plots per site), we counted the number of plots in which each species occurred and the number of plots in which they were not present, separately for all years and applied generalized linear models with logit link function to relate establishment to the traits. For those species that occurred only at the donor site or in the seed mixture of one of the two study sites, the number of plots in which a species could be potentially present was 12. For those species that occurred at the donor sites or in the seed mixtures of both study sites, the number of potentially colonizable plots was 24. We employed a two‐step model selection procedure, separately for all years. In a first selection step, we disregarded variable interactions and included all two‐variable combinations for those variables that showed a smaller AIC in step 1 than the no‐trait model and ranked all models by AICc (Akaike's Information Criterion corrected for small sample sizes). We then identified the seven variables that were included in one of the top models, starting with the best model and stopping when the number of seven variables was reached. In a second step, we additionally included variable interactions of all these seven variables and then identified the best three‐predictors model (multitrait model) according to AICc, with either three different variables or two variables and their interaction. The coefficient of determination as a measure of model fit was calculated as one minus residual deviance divided by null deviance (1‐residual deviance/null deviance). This procedure was conducted separately for all years and applied to models containing traits. The analyses were performed using the glm procedure in the stats package (R Core Team ) and the dredge procedure in the MuMIn package (Bartoń, ). […]

Pipeline specifications

Software tools vegan, MuMIn
Application Phylogenetics