Analyzing model outputs
We first assessed various food web properties of the 30,000 networks (“original webs”) to check whether the generated food webs were reasonable. Then, we identified the linked and unlinked food webs that persisted for 1000 growing seasons (i.e., years) and passed the abovementioned criteria (Fig. 3), and retained only persisting taxa in these webs (“ATN-filtered webs”). We then examined food web properties of the ATN-filtered food webs using the last 100 years of the simulations. We calculated the mean of the CVs of biomasses of fish stages to characterize dynamical stability of food webs and the number of nodes and link density to characterize food web complexity. We also compared the unlinked and the linked webs for total fish biomass, the mean body masses of fishes weighted by relative abundance, the number of fish stages, and the maximum trophic levels and median predator-prey body mass (PPMR) ratios for fish stages. We computed a surrogate of total interaction strengths by summing up all energy flows into fish stages from their resources, CVs and skewness of the energy flows, which were individually normalized by the total gain of the recipient fish , and the slopes of biomass spectra by fitting a linear model to estimate biomass at trophic levels as a function of body mass . The presence of weak interactions is theorized to stabilize food webs . Less negative slopes of biomass spectra indicate less bottom-heavy food webs (Treblico et al. 2013), which tend to be less stable than more bottom-heavy food webs . Some measurements were log10 transformed for interpretability. We computed 10000 bootstrap estimates of the mean effect sizes of a stage structure (linked – unlinked) for these metrics by taking the differences between the means of resampled values from persisting linked and unliked food webs.
We checked the sensitivity of the model outputs to major input parameters (Table 1). We ran 10000 simulations for each parameter variation (16 sets in total) using the same food web topologies as the baseline simulations. We computed 1000 bootstrap estimates of the mean effect sizes of a stage structure (linked – unlinked) for these metrics by taking the differences between the means of resampled values from persisting linked and unliked food webs and compared the mean effect sizes with those from the baseline simulations.