Statistical analyses

Prey mortality in controls without predators was negligible (range 0–2% of initial prey) and prey mortality in controls without predators did not increase with prey density (GLM, F_1,68 = 2.07, p = 0.15). The data were thus not corrected for background mortality. We calculated per capita NCM strength as the ratio of dead uneaten prey density over initial prey density, divided by the number of predators. In addition, we also calculated an alternative measure of per capita NCM strength as the ratio of the density of dead uneaten prey over the density of eaten prey, divided by the number of predators. As the results were qualitatively similar, we do not present results for the latter NCM metric. We also tested the goodness of fit of our models using the Hosmer-Lemeshow test and verified that all models fitted the data well (P > 0.05). All model results are shown as mean ± 95% Wald confidence interval (CI).

We tested whether the per capita NCM strength (hereafter only NCM strength) is influenced by temperature, prey density, predator assemblage and their interactions using a GLM with a quasibinomial distribution to account for overdispersion⁶⁴. The most parsimonious model was determined by sequential deletion of the least significant explanatory parameters or interaction terms from the full model. Parameter significance was evaluated using F-tests from the analysis of deviance. The final model included only parameters with significant p-values, and post-hoc Tukey tests were used to assess significant differences among treatment means. Finally, we grouped predator assemblages by functional groups: predators (i.e., only dragonfly larvae), scavengers (i.e., only crayfish), and mixed treatment (one scavenger and one predator) and analysed the effect of temperature, prey density and functional group on NCM strength as described above.

We tested whether NCM strength in multiple-predator assemblages can be predicted using our experimental data from single predator treatments. For this purpose, we used the multiplicative risk model that often appears in studies investigating predation rate by multiple predators on a single prey species⁶⁵:

where NC _ab is the predicted NCM strength measured as the density of dead uneaten prey, N _p is the initial prey density, and P _a and P _b are NCM strengths measured as the respective proportions of dead uneaten prey in single predator a and b treatments.

To better understand the mechanisms underlying our results, we further tested the influence of temperature, prey density, predator assemblage and their interactions on the per capita (i.e., per predator) proportion of dead prey with and without visible attack marks using two GLMs (one for each dependent variable) with quasibinomial distribution. The most parsimonious model was determined by sequential deletion of the least significant explanatory parameters or interaction terms from the full model and parameter significance was evaluated using F-tests from the analysis of deviance.

Finally, we tested whether per capita NCM strength depended on predator density along with predator identity, temperature, prey density and their interactions. Only single predator treatments and treatments with predator pairs of the same size and species were used in this analysis. We again used GLMs (one for each dependent variable) with quasibinomial distribution and proceeded with model selection and evaluation of parameter significance as above. All analyses were implemented in R version 3.2.5⁶⁶.