Appendix S5. Hypothetical causal relationship between grasshopper traits and their impact on plant biomass The confirmatory multilevel path analysis (Shipley, 2009) is based on directed acyclic graphs (i.e. a box-and-arrow diagrams without feedback loops). The graphs are used to specify the direct, indirect and independent relationships between the variables of the system implied by each competing hypothesis (Fig. S5). The validity of each path model is tested by deriving from each graph the set of independence claims (Table S5). Using multilevel/mixed effect models the probabilities pi of each of the k independence claims are obtained, which are then combined into a C statistic: k C 2 ln( p i ) (1) i 1 The resulting value is compared to a χ2 distribution with 2k degrees of freedom (Shipley 2013). If the value of the C-statistic is lower than the specified significance level (here, α = 0.05) the path model is rejected, as the data depart significantly from what expected under the tested causal model. To test the independence claims, we used linear mixed models with block as random effect, using the lmer function of the lme4 package version 1.1-6. (Bates et al. 2014) within the R language and software environment for statistical computing version 3.0.2, following Shipley (2013). Model assumptions were tested by inspecting the residuals Pinheiro & Bates (2000). All variables were standardised (mean-centred and divided by the standard deviation) and standardised path coefficients were obtained. Direct, indirect, and total effects were computed using standardised path coefficients following Grace & Bollen (2005). We tested here two alternative hypotheses using confirmatory path analysis (Shipley 2013) see main text: (i) direct effect of grasshopper traits on plant community biomass; (ii) indirect effect mediated by grasshopper feeding niche (see main text for detailed hypotheses on each arrow). Fig. S5 describes the two a priori hypotheses for all links (arrows) leading to endogenous variables. Table S5 details the conditional independence between variables tested in the analyses. For each hypothesis, we also tested alternative models which included one grasshopper trait only or the two traits. Hypothesis 1 was rejected while Hypothesis 2 was accepted (Table S5; Fig. 5 in the main text). Table S5 Conditional independence tests in the basis sets implied by the hypothesized path models (Fig. S5). D-sep claim of independence Hypothesis 1 : Direct effect (BS,Predicted LNRR)|{Ø} (IS, Predicted LNRR)|{Ø} (Survival, Predicted LNRR)|{Ø} (Predicted LNRR, Observed LNRR)|{IS,Survival} (BS, IS)|{Ø} (BS, Survival)|{Ø} (IS, Survival)|{Ø} (BS, LNRRobs)|{IS,Survival} Model formula Ho P value Predicted LNRR ~ BS Predicted LNRR ~ IS Predicted LNRR ~ Survival Observed LNRR ~ Predicted LNRR + BS + IS + Survival IS ~ BS Survival ~ BS Survival ~ IS Observed LNRR ~ BS + IS + Survival BS=0 IS=0 Survival=0 Predicted LNRR=0 BS=0 BS=0 IS=0 BS=0 0.25 <0.0001 0.02 0.07 0.45 0.60 0.11 0.045 IS ~ BS Survival ~ BS + Predicted LNRR Survival ~ IS + Predicted LNRR Observed LNRR ~ BS + Survival + Predicted LNRR Observed LNRR ~ IS + Survival + Predicted LNRR Predicted LNRR ~ BS + IS BS=0 BS=0 IS=0 BS=0 IS=0 BS=0 0.45 0. 94 0. 78 0.07 0.36 0.40 C statistic : 60.28 P value of C (DF) : < 0.0001 (16) Hypothesis 2 : Mediated effect (BS,IS)|{Ø} (BS, Survival)|{ Predicted LNRR } (IS, Survival)|{ Predicted LNRR } (BS, Observed LNRR)|{Survival, Predicted LNRR} (IS, Observed LNRR)|{Survival, Predicted LNRR} (BS, Predicted LNRR)|{IS} C statistic : 11.30 P value of C (DF) : 0.5 (12) Notes: Key to variables: BS = grasshopper body size, IS = grasshopper incisive strength, Survival = grasshopper survival within each cage at the end of the experiment, Observed LNRR and predicted LNRR referred to equations 1 and 6 respectively in the main text. (Xi,Xj)|{Xk} means that variables Xi and Xj are independent conditional on variable Xk (thus variation in Xi does not imply variation in Xj if Xk is held constant,).† The associated mixed model regression for each d-sep claim using the lmer function in the lme4 package in R to test the independence claims. * The p-value is obtained by comparing the value of the C statistic for each hypothesis to a chi-square distribution with the same degrees of freedom – note that a model is rejected if the C statistic is significantly different from the χ2 value.