Apparent competition drives community-wide parasitism rates and changes in host abundance across ecosystem boundaries

Species have strong indirect effects on others, and predicting these effects is a central challenge in ecology. Prey species sharing an enemy (predator or parasitoid) can be linked by apparent competition, but it is unknown whether this process is strong enough to be a community-wide structuring mechanism that could be used to predict future states of diverse food webs. Whether species abundances are spatially coupled by enemy movement across different habitats is also untested. Here, using a field experiment, we show that predicted apparent competitive effects between species, mediated via shared parasitoids, can significantly explain future parasitism rates and herbivore abundances. These predictions are successful even across edges between natural and managed forests, following experimental reduction of herbivore densities by aerial spraying of insecticide over 20 hectares. This result shows that trophic indirect effects propagate across networks and habitats in important, predictable ways, with implications for landscape planning, invasion biology and biological control.

would be in adult form, and thus not killed along with their hosts during the herbivore reduction spray treatment. That is, the adult parasitoids from the first generation would have emerged, and would be mating and ovipositing at the time of the herbivore reduction treatment. We present flight intercept trap data collected during this study but analyzed elsewhere 1 for adult parasitoid wasps (Ichneumonidae and Braconidae) of the same species as were reared from caterpillars in this study. Flight intercept traps at control sites were used to monitor parasitoid wasp movement across the edge between plantation and native forest at each site 1 . Our herbivore reduction spray occurred at the point in the summer when adult parasitoid wasp activity was highest (i.e. most had emerged as adults and were moving around mating and ovipositing), and thus should have Trap collection date Nov-10 early Jan-11 late Jan-11 Feb-11 Mar-11 Apr-11 killed relatively few parasitoid larvae of the first or second generations. Our caterpillar collections to obtain quantitative food web data at the validation sites at time t and t+1 occurred before and after the herbivore reduction spray respectively, as marked below the x-axis in red, with lines representing the time periods over which each collection occurred.

Supplementary Figure 2.
Raw data corresponding to Fig. 5a,b. Here we present the raw data for which statistical models showed that (a) expected parasitism rate significantly predicts observed parasitism rate, but (b) initial parasitism rate does not. Fig. 5a,b presents the same data, but with a random effect removing the variation among sites. Each point represents a species within a site that was collected and successfully reared at both time steps, and was parasitized in the first time step (see Eq. 3, Fig. 1i).   The model was a generalized linear mixed model with a Poisson error distribution (including the canonical log link function) and caterpillar abundance as the response variable. The predictors were collection (two levels: with the two before-herbivore-reduction collections pooled into one "before" sample vs. the two after-herbivore-reduction collections pooled into one "after" sample), treatment (herbivore reduction vs. control) and the collection x treatment interaction as fixed effects. Site, nested within block were included as random factors to account for the nonindependence of the two treatments within each site, and the repeated measures at the before and after time steps at each site. Bold p-values indicate significance at α = 0.05. The significant interaction between herbivore reduction treatment and collection (p = 0.0015) suggests that caterpillar abundance was initially greater at herbivore reduction sites (Treatment main effect) by chance (reduction vs. control treatments were allocated randomly), but after the herbivore reduction treatment, caterpillar abundance was significantly lower at treated sites than at control sites. The model was a generalized linear mixed model with a binomial error distribution (including the canonical logit link function) and parasitism rate in native forest as the response variable. The predictors were collection (two levels: with the two before-herbivore-reduction collections pooled into one "before" sample vs. the two after-herbivore-reduction collections pooled into one "after" sample), treatment (herbivore reduction vs. control) and the collection x treatment interaction as fixed effects. Site nested within block were included as random factors to account for the non-independence of the two treatments within each site, and the repeated measures at the before and after time steps at each site. Bold p-values indicate significance at α = 0.05. The best model contained only collection; parasitism rates at both herbivore reduction and control sites were higher at the 'after' time step, though collection explained only 2.7% of the variation in parasitism rate in native forest (R 2 GLMM(m) = 0.027).

Supplementary
The model was a generalized linear mixed effects model with a binomial error distribution (including the canonical logit link function), and block included as a random factor. Herbivore reduction treatment, forest, and all two-way and three-way interactions were removed from the full model during model selection, as were the forest within site nested random factors. AIC, Akaike Information Criterion; PCV, percentage change in variance; VC, variance components; R 2 GLMM(m) , marginal R 2 ; R 2 GLMM(c) , conditional R 2 . Bolded coefficients and P-values are significant at α = 0.05. The model was a generalized linear mixed effects model, with a binomial error distribution (including the canonical logit link function), and block included as a random factor. Herbivore reduction treatment, the two-way interactions between E ia(t+1) and treatment, and habitat A and treatment, and the three-way interaction were removed from the full model during model selection, as were the forest within site nested random factors. AIC, Akaike Information Criterion; PCV, proportion change in variance; VC, variance components; R 2 GLMM(m) , marginal R 2 ; R 2 GLMM(c) , conditional R 2 . Bolded coefficients and P-values are significant at α = 0.05. The model was a generalized linear mixed effects model, with a binomial error distribution (including the canonical logit link function), and forest was included as a random factor. Crosshabitat expected parasitism rate was removed during model selection, as were the site within block random factors. AIC, Akaike Information Criterion; PCV, proportion change in variance; VC, variance components; R 2 GLMM(m) , marginal R 2 ; R 2 GLMM(c) , conditional R 2 . Bolded coefficients and P-values are significant at α = 0.05. Table 7. Cross-habitat data were also sufficient for successful prediction.

Supplementary
Results of mixed-effects models testing whether expected parasitism rate based on apparent competition between hosts across habitats, Log( E iA(t+1) Cross), predicts observed parasitism rate.

Response variable:
Observed parasitism rate Null model Full model The model was a generalized linear mixed effects model, with a binomial error distribution (including the canonical logit link function), and forest was included as a random factor. AIC, Akaike Information Criterion; PCV, proportion change in variance; VC, variance components; R 2 GLMM(m) , marginal R 2 ; R 2 GLMM(c) , conditional R 2 . Bolded coefficients and P-values are significant at α = 0.05.

Supplementary Table 8. Predictions based on binary food web data worked less well.
Coefficients from the best-fitting models testing whether a) expected parasitism rate, b) withinand cross-habitat expected parasitism rates, and c) cross-habitat expected parasitism rate, calculated using binary d iAjB , could predict observed parasitism rate, O iA(t+1). Each model was a generalized linear mixed effects model with binomial error distribution (including the canonical logit link function), and block included as a random factor. P-values in bold are significant at α = 0.05. a) The forest-within-site nested random factors were removed during model selection. Expected parasitism rate, herbivore reduction treatment, and habitat A , as well as all interactions were included as fixed effects. Compare these binary results to the quantitative results in Table 3. b) Cross-habitat expected parasitism rate was removed during model selection, as were the forest-within-site random factors. Compare these binary results to the quantitative results in Supplementary Table 6. c) Compare these binary results to the quantitative results in Supplementary Table 7. The random factors (forest, site, and block) did not explain any of the variation in change in focal host abundance, so we used a linear model. The full model contained as predictors initial parasitism rate, forest, herbivore reduction treatment and all interactions, but none of the fixed factors improved model fit over the intercept model. Data for this analysis were from the regional metaweb (see Fig. 2), so sites were pooled, and habitat-specific species pairs were replicates (PN: n = 178; NN: n = 255; NP: n = 178; PP: n = 155).

Supplementary Table 12.
Food web metrics for habitat-specific food webs. Food web metrics for regional quantitative food web data from training sites, but with food webs from native and plantation forests separated to allow comparison between forest types. Quantitative food-web data were pooled over all collection dates and sites, but separated by forest type. Network structure was very similar between forest types. Shannon Diversity was calculated using the diversity function in the vegan package 2 in R. Network metrics were calculated using the networklevel function in the bipartite package 3 in R, using the 'sum' method for calculating interaction evenness.