Ecology, 94(7), 2013, pp. 1499–1509 Ó 2013 by the Ecological Society of America Predicting invertebrate herbivory from plant traits: Polycultures show strong nonadditive effects JESSY LORANGER,1,2,6 SEBASTIAN T. MEYER,1,7 BILL SHIPLEY,2 JENS KATTGE,3 HANNAH LORANGER,1,8 CHRISTIANE ROSCHER,4 CHRISTIAN WIRTH,5 AND WOLFGANG W. WEISSER1,7 1 Institute of Ecology, Friedrich Schiller University of Jena, Dornburger Str. 159, 07749 Jena, Germany De´partement de biologie and Centre des e´tudes de la forêt, Université de Sherbrooke, Sherbrooke, QC J1K 2R9 Canada 3 Max Planck Institute for Biogeochemistry, Hans-Knöll Str10, 07745 Jena, Germany 4 UFZ, Helmholtz Centre for Environmental Research, Department of Community Ecology, Theodor-Lieser-Str. 4, 06120 Halle, Germany 5 Department of Special Botany and Functional Biodiversity, Universität Leipzig, Ritterstr. 26, 04109 Leipzig, Germany 2 Abstract. Plant functional traits affect the capacity of herbivores to find, choose, and consume plants. However, in a community composed of different plant species, it is unclear what proportion of herbivory on a focal plant is explained by its own traits and which is explained by the characteristics of the surrounding vegetation (i.e., nonadditive effects). Moreover, nonadditive effects could be positive or negative, and it is not known if they are related to community properties such as diversity. To quantify nonadditive effects, we developed four different additive models based on monoculture herbivory rates or plant traits and combined them with measurements of standing invertebrate herbivore damage along an experimental plant diversity gradient ranging from monocultures to 60-species mixtures. In all four models, positive nonadditive effects were detected, i.e., herbivory levels were higher in polycultures than what was expected from monoculture data, and these effects contributed up to 25% of the observed variance in herbivory. Importantly, the nonadditive effects, which were defined as the deviance of the models’ predictions from the observed herbivory, were positively correlated with the communities’ plant species richness. Consequently, interspecific interactions appear to have an important impact on the levels of herbivory of a community. Identifying those community properties that capture the effects of these interactions is a next important challenge for our understanding of how the environment interacts with plant traits to drive levels of herbivory. Key words: community-weighted traits; consumers; diversity; grassland; insects; interactions; invertebrate herbivory; Jena Experiment; monocultures; polycultures; Saale River floodplain, Thuringia, Germany. INTRODUCTION Herbivory is a major selective pressure affecting plant communities (Allan and Crawley 2011) and the capacity of plants to avoid, resist, or tolerate herbivory is mediated by their functional traits. Consequently, plant species differing in such traits can drastically differ in rates of herbivory (Coley and Barone 1996, Rasmann and Agrawal 2009). Moreover, the rate of herbivory on a focal plant in a multispecies community depends on its own traits, but could also depend on the characteristics Manuscript received 27 November 2012; revised 7 February 2013; accepted 13 February 2013. Corresponding Editor: N. J. Sanders. 6 E-mail: [email protected] 7 Present address: Technische Universität München, Terrestrial Ecology Group, Department of Ecology and Ecosystem Management, Center for Food and Life Sciences Weihenstephan, Hans-Carl-von-Carlowitz-Platz 2, 85350 Freising, Germany. 8 Present address: Institute of Biology and Environmental Sciences, University of Oldenburg, Functional Ecology of Plants, 26111 Oldenburg, Germany. of the surrounding vegetation, such as its taxonomic and functional composition. Consider first the null hypothesis H0, which assumes that the decision of the herbivore to consume tissue of a given plant depends only on the functional traits of the focal species, irrespective of the surrounding vegetation. If this is true, then the herbivory experienced by a species is the same in multispecies vegetation as in a monoculture, assuming that the trait values of the species do not change in the two types of vegetation. Consequently, the total rate of herbivory of the entire vegetation is the sum of the monoculture rates of herbivory (measured or predicted from traits) of each species in the vegetation weighted by its relative abundance. This is called ‘‘additive scaling’’ of herbivory (H0) because (1) the total rate is an additive function of the monoculture rates and (2) these monoculture rates are scaled up to multispecies mixtures. Under the alternative hypothesis H1 (‘‘nonadditive scaling’’), the probability that a herbivore will consume tissue of a given plant depends on properties of the surrounding vegetation in addition to the functional 1499 1500 JESSY LORANGER ET AL. traits of the focal species. The scaling from monocultures to mixtures is ‘‘nonadditive’’ because one must include some factor beyond an additive function of the monoculture rates. For instance, herbivory on the focal plant could increase relative to monoculture if herbivores are attracted to it by the surrounding vegetation; the inverse mechanism would result in decreased herbivory. Because some studies (Huntly 1991, Hambäck et al. 2000, Finch and Collier 2012 ) have demonstrated a decreased herbivory that is mediated by the presence of another neighboring species, it is likely that nonadditive effects do occur in mixed communities. However, we do not know their strength relative to the additive scenario nor whether such nonadditive effects are consistently positive or negative. Although statistically additive models based on community-weighted traits, but not scaling up from monoculture herbivory, do not test the additive scaling hypothesis, they can still be used to quantify the size and direction of nonadditive effects by calculating the deviation between community herbivory predicted from the models and herbivory measured in mixed-species communities. In addition, comparing communities that differ in the magnitude of this deviation can give insights into which community properties determine the strength of nonadditive effects. In this study, we developed four models (M1 to M4) to predict herbivory measured in experimental herbaceous grassland communities, ranging from monocultures to 60-species mixes (Jena Experiment; Roscher et al. 2004) that are exposed to a natural community of invertebrate herbivores. M1 and M2 test the additive scaling hypothesis (H0) because they were directly based on monoculture herbivory data. M3 and M4 were based on community-weighted traits and therefore do not scale up from the levels of herbivory measured in monocultures. As a result, these models cannot test the additive (or nonadditive) scaling hypothesis, but are rather designed to investigate how the nonadditive effects of multispecies communities on herbivory are mediated by plant traits. The specific goals of this study were (1) to test the models’ ability to predict herbivory in polycultures correctly (all models) and to see whether results from monocultures can be scaled up to polycultures (M1 and M2), (2) to quantify if there are nonadditive effects and determine their magnitude and direction (all models), and (3) to analyze if any nonadditive effects are related to community properties of the vegetation (all models). METHODS Study site The field site was located on the floodplain of the Saale River at the northern edge of Jena (50855 0 N, 11835 0 E; altitude 130 m a.s.l.), Thuringia, Germany, on a Eutric Fluvisol (1997 update of FAO 1988). The Jena Experiment, established in 2002, is a biodiversity experiment consisting of 80 large plots (originally 20 3 20 m, now 5 3 7 m) in a randomized-block design, Ecology, Vol. 94, No. 7 containing 1, 2, 4, 8, 16, or 60 plant species with 14, 16, 16, 16, 14, and 4 replicates of these levels of species richness, respectively. The pool of 60 species (Appendix A) consists of herbaceous plant species commonly occurring in seminatural, mesophilic grasslands in Central Europe: Molinio–Arrhenatheretea meadows, Arrhenatherion community (Ellenberg 1996). Furthermore, there is one small monoculture plot (1 3 1 m) of each of the 60 species on the experimental site. All plots are mown twice a year and weeded 2–3 times annually, keeping only the target species in each plot. A detailed description of the setup of the experiment is given in Roscher et al. (2004). Biomass and herbivory measurements In May and August 2010, biomass inside a 20 3 50 cm frame was cut 3 cm above the ground from a randomized position in each large plot. Unsown species were separated and the remaining biomass was sorted to species, oven-dried (at 788C for 48 h), and weighed. Before drying, the same samples were used to measure invertebrate leaf standing herbivore damage (hereafter ‘‘herbivory’’) as follows. For each species per plot, 30 leaves were chosen randomly. In some large plots, less than 30 leaves were available for rare species (minimum ¼ 1 leaf per species; mean ¼ 21 leaves per species; on average, 65% of the initially sown species per plot occurred in the samples). The damaged surface area of all leaves (in square millimeters) was estimated visually by comparing the damaged leaf area to a series of circular and square templates ranging in size from 1 mm2 to 500 mm2. The total damaged area per leaf included four types of herbivory: chewing, rasping, sucking, and mining. The remaining surface area of each leaf was measured using a LI-3000C Area Meter (LICOR, Lincoln, Nebraska, USA). The potential undamaged leaf area before herbivory was estimated by adding the proportion of the area lost to chewing damage (a factor previously estimated for all species) to the measured area. Herbivory was calculated as the total damaged area divided by the potential undamaged leaf area. Following the same protocol, herbivory was estimated for leaf samples from the small-area monocultures (1 plot per species, 30 leaves per plot), hereafter simply referred to as monocultures; see Loranger et al. (2012) for details on sampling in monocultures. From the sampled large plots, a total of 10 were excluded because of missing herbivory measurements (six plots) or because they contained more than 15% biomass of one or more of the nine species for which reliable herbivory measurements in monoculture were not available (four plots; see Appendix A). When these species accounted for ,15% of the biomass in a plot, they were excluded (also for the calculation of community-weighted traits as we will detail), but the plot was kept in the analyses with the remaining species. For each large plot (hereafter referred to as ‘‘communities,’’ including 12 large-area monoculture plots), a commu- July 2013 COMMUNITY PROPERTIES AFFECT HERBIVORY 1501 nity-level estimate of herbivory was calculated by summing up herbivory per species in that plot multiplied by the respective relative biomass of each species. Values for the two harvests (May and August) were averaged, giving 70 values of measured community herbivory. is the abundance-weighted mean of the species-specific monoculture herbivory, as follows: Predicting herbivory in monocultures where hi is the herbivory experienced by an average individual of plant species i when growing in a monoculture, and raij is the relative abundance of species i in a community j composed of S species. The nonadditive scaling hypothesis (H1) assumes that the herbivory suffered by species i in monoculture changes by an amount di when growing in mixture. This implies that the predicted total herbivory experienced by the entire plant community j deviates from the sum of the predicted herbivory experienced by each species in monoculture by an amount Dj : As an initial step to predict herbivory in monocultures (Loranger et al. 2012), we assembled a data set of 42 plant traits, including physiological, morphological, phenological, and herbivore-related traits, from data collected in the Jena Experiment (Roscher et al. 2004, 2011a, b, Gubsch et al. 2011; M. Gubsch, A. Lipowsky, and C. Roscher, unpublished data), and from international plant trait databases: TRY (Kattge et al. 2011, including the following main references: Kuhn et al. 2004, Garnier et al. 2007, Pakeman et al. 2008, 2009, Fortunel et al. 2009), LEDA (Kleyer et al. 2008), and Biolflor (Klotz et al. 2002). Appendix B, Table B1 gives a detailed description of the traits and a full list of references. In the next step, the method of random forests (RF; Breiman 2001) was applied, which uses a series of regression trees to derive importance scores that indicate the most important traits for predicting herbivory among a large number of traits (see Appendix C for more details on the methodology of the random forests selection technique). Following the random forests, a multiple regression with model simplification via stepwise selection was done and seven of the 42 initial plant traits were identified as being significant predictors of herbivory: leaf nitrogen concentration (loge-transformed), leaf lignin concentration, number of coleopteran and hemipteran (excluding aphids) herbivores potentially feeding on the plants (logetransformed), leaf life span, stem growth form (percentage erection of the stem), and root architecture. The final model from Loranger et al. (2012), which explained 63% of the variation in herbivory measured in monocultures (Appendix B: Fig. B1) and was supported by cross-validation tests, is: lnðmeasured herbivoryÞ ’ 10:65 þ 1:75 3 lnðleaf½nitrogenÞ 0:07 3 leaf½lignin þ 0:55 3 lnðcoleopteran herbivoresÞ 0:41 3 lnðhemipteran herbivoresÞ þ 0:37 3 leaf life span 0:007 3 stem growth form þ 0:27 3 root architecture: ĥj ðH0 Þ ¼ S X ð2Þ raij hi i¼1 hj ðH1 Þ ¼ S X raij ðhi þ di Þ ¼ i¼1 S X i¼1 raij hi þ S X raij di i¼1 hj ðH1 Þ ¼ hj ðH0 Þ þ Dj : ð3Þ Note that the deviation term di does not represent a purely statistical error term (i.e., a random value from a distribution having a zero mean), but rather the deviation from the additive scaling hypothesis between measured and predicted herbivory (nonadditive effects). However, because hi contains sampling variation and measurement error, di also contains these errors. If the deviations from additive scaling (d i ) vary randomly and independently for each species in the plant community, and if they are equally likely to be positive or negative, then they will tend to cancel each other (D j ; 0) and additive scaling would be a reasonable approximation; otherwise nonadditive scaling occurs and Dj is an approximation of the strength of nonadditive effects in determining herbivory in plant communities. Models predicting herbivory in polycultures that are not based on monoculture data and thus do not scale up from monoculture, cannot be tests of the additive (or nonadditive) scaling hypothesis. Given this, theory and terms related to Eq. 3 cannot be applied to these models. However, if these models are statistically additive, i.e., they are a linear combination of predicting factors, an equivalent term for Dj can be calculated to separate additive from nonadditive effects. This term is also calculated in the same way as the deviation between measured and predicted herbivory and it will also be an approximation of the strength of nonadditive effects in the community. Testing the additive scaling hypothesis ð1Þ Additive vs. nonadditive scaling Given the additive scaling hypothesis (H0), the predicted herbivory of the plant community j (ĥ j(H0)) The additive scaling hypothesis was tested using two different models. First (M1), community herbivory was predicted (ĥ ) from community-weighted species-specific herbivory measured in monocultures (Eq. 2). Second (M2), predictions were based on plant functional traits and previous relationships between traits and herbivory 1502 JESSY LORANGER ET AL. (Loranger et al. 2012) obtained in monocultures. To do so, the species-specific trait values in the monoculture model (Eq. 1) were replaced by the associated community-weighted trait values, while keeping the intercept (b0) and partial slopes (bk) at the values estimated in the monoculture model: ĥj ðH0 Þ ¼ S X raij hi ¼ i¼1 S X rai ðb0 þ i¼1 T X bk tik Þ k¼1 ¼ b0 þ b1 t1 þ . . . þ bk tk : ð4Þ Predicted herbivory from M1 and M2 was regressed against measured herbivory for all 70 plant communities included in the study. If the additive scaling hypothesis is correct, the regression slopes will not be significantly different from unity. A paired t test was additionally used to verify if, on average, predicted values of community herbivory from these two models were significantly lower or higher than measured values. Testing the relative importance of plant traits in monocultures vs. polycultures Eq. 5 represents a model that assumes that the best traits predicting herbivory in monocultures and in polycultures are the same, but that the values of the parameters (b) can change by an unknown amount d. Thus (M3), the measured values of community herbivory were freely regressed against the communityweighted version of the same seven traits selected in the monoculture model, giving new partial slopes (bk0 ): ĥj ðH0 Þ ¼ S X raij hi0 i¼1 ¼ ðb0 þ d00 Þ þ ðb1 þ d10 Þt1 þ . . . þ ðbK þ dK0 Þtk ĥj ðH0 Þ ¼ b00 þ b10 t1 þ . . . þ bK0 tk : ð5Þ For this model, if the only difference between monocultures and multispecies communities is that the relative importance (i.e., regression slopes) of the different traits changes between monocultures and mixtures, an equivalent percentage of explained variation relative to the result in monocultures (i.e., ;63% variance explained) is to be expected. Because these regression slopes are now freely estimated from the multispecies communities, rather than being fixed by the monoculture values, M3 is not a true test of the additive scaling hypothesis, but is still an additive model that does retain the assumption that the same traits are important in determining herbivory levels in monocultures and mixtures. Finally, a new trait-based model (M4) was created using community-weighted values of all 42 traits by selecting traits of potential importance in a random forests analysis (see Appendix C) followed by a backward stepwise selection in a multiple regression until the partial slope of all traits in the model were significantly different from 0. M4 is not a test of the additive scaling hypothesis, but allows us to investigate Ecology, Vol. 94, No. 7 whether different traits are governing community herbivory compared to monoculture herbivory, while still maintaining the assumptions of statistical additive effects. In addition, we related the deviations (difference between measured and predicted herbivory) of models M1–M4 to a series of community properties. We initially calculated seven properties: sown species richness, community biomass, realized species richness, Shannon diversity, evenness (diversity divided by species richness), functional diversity via Rao’s quadratic entropy, and functional dispersion (Laliberté and Shipley 2011). However, sown species richness was selected as a surrogate for other community properties because (1) there were highly significant correlations between all of these community characteristics except for community biomass and evenness (see Appendix D), (2) sown species richness was the actual community property that was experimentally manipulated in the experiment, and (3) qualitative results were the same for all community properties. Besides simple sampling variation and ‘‘missing’’ variables causing variation in herbivory levels, deviations between the measured and predicted values in our four models can be caused by nonadditive effects. We will treat these deviations as an approximation of nonadditive effects and will call them so. In doing so we are assuming that measurement error and sampling variation are negligible, relative to the range of variation in these values. To estimate the relative importance of additive and nonadditive effects in determining community herbivory in all models, the log ratio of their absolute contributions was calculated. Because our four models are statistically additive, the additive contribution is the value predicted by each model and the nonadditive contribution is the difference between measured and predicted herbivory. This difference can be positive or negative, and the log ratio of the absolute values quantifies their relative importance (Appendix E). All statistical analyses were done in R version 2.10.0 (R Development Core Team 2009). The values of herbivory, deviation of predicted herbivory from measured herbivory, sown species richness, and of several plant traits (see Appendix B) were loge-transformed. We present R 2 values adjusted for bias due to differing numbers of predictor variables, as provided by the ‘‘lm’’ function in R. This measures the proportion of the biologically relevant variation that is explained by the model and is defined as R 2(N 1)/(N k 1), where N is the number of observations and k is the number of predictors of each regression. The relative importance of the partial slopes, or magnitude of effect, was calculated by multiplying the absolute value of each partial slope by the range (maximum minus minimum) of its associated community-weighted trait. In addition, the goodness of fit of the models (M1–M4) was compared using the corrected version of the Akaike information July 2013 COMMUNITY PROPERTIES AFFECT HERBIVORY criterion for small ‘‘N’’ size (AICc), as defined by Burnham and Anderson (2010). RESULTS Testing the hypothesis of additive scaling The additive scaling hypothesis (H0) asserts that the predicted herbivory of multispecies vegetation is the abundance-weighted average of the species-specific monoculture herbivory values. Predicted herbivory based on this relation (M1) underestimated measured herbivory (slope 0.43 6 0.09) and explained only 23% (AICc ¼ 84.84) of the variation in measured community herbivory (Fig. 1A, Table 1). The actual level of herbivory in the communities, on average, was greater than expected from monoculture (t test: t ¼ 2.39; df ¼ 69, P ¼ 0.020) and this positive bias increased significantly with sown species richness of the communities (Fig. 1B, Table 1). Thus, we rejected the additive scaling hypothesis. The community-weighted version (M2) of a previously published model linking herbivory in monoculture to species-specific traits (Eq. 4) explained only 6% (Table 1) of the variation in measured community herbivory and the fit was worse than for M1 (AICc ¼ 70.70). Again, the model underestimated actual levels of herbivory (Fig. 1C, Table 1). The deviation between measured and predicted values, on average, was higher and more significantly differed from zero (t ¼ 3.31, df ¼ 69, P ¼ 0.001) than with predictions based on measured monoculture herbivory. Deviations increased with increasing sown species richness of the plant communities (Fig. 1D, Table 1). Again, the additive scaling hypothesis was rejected. Testing the relative importance of plant traits in monocultures vs. in polycultures For our third model (M3), we regressed the measured levels of community herbivory against the same seven traits as in the monoculture model, but estimated new regression slopes rather than fixing them to the values of the initial model. In this multiple regression, the seven traits explained 25% (Table 2) of the variation in community herbivory. Thus, the predictive power was weaker than for the model in monocultures (R 2 ¼ 0.63; Appendix B: Fig. B1), but was better than with M1 and M2 (Fig. 1E, Table 1). In contrast, the goodness of fit of the model was in between those of the first two models (AICc ¼ 72.37), because the increase in explained variation was counterbalanced by many more parameters that M3 had to estimate to obtain a higher R 2. Although this model was highly significant, only the slopes associated with coleopteran herbivores and leaf life span were (marginally) significant and the relative importance of the traits changed considerably in comparison to the model for monocultures (Table 2). Again, the deviation between measured and predicted herbivory correlated significantly and positively with sown species richness (Fig. 1F, Table 1). 1503 Given that the results of M3 indicated a change from monocultures to multispecies communities in the relative importance of plant traits in predicting herbivory, we used a new random forests analysis on the 42 initial traits to select a new set of traits that best predicted community herbivory. In total, 20 traits were selected (Table 3; Appendix C). From those 20 traits, a backward stepwise selection in a multiple regression identified five traits to significantly predict herbivory in communities (M4; Table 3). These five traits explained 55% of the variation in measured community herbivory, which was still less than that explained by the monoculture model. Only the trait ‘‘coleopteran herbivores’’ was selected in both the community and monoculture models. However, ‘‘height summer’’ (selected in the communities) is naturally related to, and correlated with (r ¼ 0.43, P , 0.001) stem growth form (chosen in the original monoculture model). Similarly, leaf primary fiber concentration (selected in communities) is correlated (r ¼ 0.72, P , 0.001) with leaf lignin concentration (selected in monocultures). Predicted herbivory in M4 correlated more closely with measured herbivory than did predictions from the first three models (Fig. 1G, Table 1; AICc ¼ 113.39), yet the deviation between measured and predicted herbivory correlated even more significantly and more positively with sown species richness (Fig. 1H, Table 1). Quantifying the strength and direction of nonadditive effects The deviance of all four models increased with diversity, ranging from an average deviance close to zero with some strongly negative or positive values at low levels of species richness, to a clearly positive deviance at highest diversity; i.e., all models increasingly underestimated levels of herbivory with increasing diversity (Fig. 1). In general, measured herbivory was higher than expected from the four different models (Fig. 1, Appendix E: Fig. E1; solid circles). To quantify how much additive, compared to nonadditive, effects contributed to the observed levels of herbivory, the measured herbivory needs to be partitioned into these two components. Therefore, we calculated the log ratio of the predicted herbivory (additive effect) and the absolute value of the difference between measured and predicted herbivory, i.e., the absolute deviance (nonadditive effect). In contrast to the deviance, the log ratio of nonadditive to additive effects did not change significantly with species richness (Appendix E: Table E1), even if levels for the highest diversity level were much higher for all models. On average, the additive effects estimated by the models M1–M3 exceeded 2.7 times the nonadditive effects. In other words, more than onequarter of the measured herbivory was determined by nonadditive effects (assuming that measurement error was small and there were no missing variables). For M4, which was designed to explain as much variation as 1504 JESSY LORANGER ET AL. Ecology, Vol. 94, No. 7 FIG. 1. Herbivory (percentage leaf standing herbivore damage for whole plant communities) as predicted from different models regressed against measured herbivory at the field site of the Jena Experiment, Thuringia, Germany, and the relationship between deviance of the models (difference between measured and predicted herbivory) and species richness (number of plant species sown) of the plant communities. Note the logarithmic axis for herbivory and species richness. Dashed lines give expected relationships (a slope of 1 between predicted and measured herbivory in the left-hand panels and a constant average deviance of 0 in the right-hand panels). Solid lines are best-fit lines of significant additive models, given in detail in Table 1. (A, B) In the upper row of panels, predictions are based on herbivory levels of the species in monocultures (M1, Eq. 2). (C, D) In the second row, predictions are from a model, based on seven plant functional traits, that was developed to predict species-specific herbivory in plant monocultures and that was used with community-weighted traits for the different plant communities (M2, Eqs. 1 and 4). (E, F) In the third row, predictions are based on a trait model in which new partial slopes were estimated for the same seven community-weighted traits directly for the herbivory measured in communities (M3; the resulting model is given in Table 2). (G, H) In the bottom row, predictions are based on a trait-based model for which a new trait selection identified five community-weighted traits out of a set of 42 plant traits to be the best predictors of herbivory measured in communities (M4; the resulting model is given in Table 3). July 2013 COMMUNITY PROPERTIES AFFECT HERBIVORY 1505 TABLE 1. Statistics for models presented in Fig. 1 on the relationship between (1) herbivory predicted from four different models and herbivory measured in plant communities of differing diversity and (2) the deviance of these models and diversity of the plant communities. Intercept 6 SE Slope 6 SE r2 F P M1, predicted from monocultures log(HpM1) ; log(Hm) log(DevM1 þ 15) ; log(sowndiv) 0.15 6 0.07 2.70 6 0.02 0.43 6 0.09 0.02 6 0.01 0.23 0.07 21.7 6.06 ,0.001 0.016 M2, predicted from monoculture trait-based model log(HpM2) ; log(Hm) log(DevM2 þ 10) ; log(sowndiv) 0.16 6 0.08 2.29 6 0.03 0.22 6 0.10 0.04 6 0.01 0.06 0.08 5.28 7.10 0.025 0.010 M3, predicted from relaxed trait-based model log(HpM3) ; log(Hm) log(DevM3 þ 3) ; log(sowndiv) 0.37 6 0.05 0.98 6 0.07 0.33 6 0.06 0.08 6 0.03 0.32 0.07 33.5 6.33 ,0.001 0.014 M4, predicted from new trait-based model log(HpM4 ) ; log(Hm) log(DevM4 þ 3) ; log(sowndiv) 0.23 6 0.05 0.95 6 0.06 0.59 6 0.06 0.09 6 0.03 0.58 0.12 96.0 10.7 ,0.001 0.002 Model for predicting herbivory Notes: Variable definitions: HpM1 is community herbivory predicted from monoculture herbivory; Hm is measured community herbivory; DevM1 ¼ Hm HpM1; sowndiv ¼ sown species richness; HpM2 is community herbivory predicted from a trait-based monoculture model; DevM2 ¼ Hm HpM2; HpM3 is community herbivory predicted from a trait-based model; DevM3 ¼ Hm HpM3; HpM4 is community herbivory predicted from a new trait-based model; DevM4 ¼ Hm HpM4. For all F statistics, df ¼ 1, 68. possible with community-weighted traits, i.e., to minimize the ratio of nonadditive to additive effects, nonadditive effects still accounted for ;10% of the observed herbivory in polycultures (Appendix E: Fig. E1, Table E1). For all four models, the percentage of observed herbivory attributable to nonadditive effects did not change with diversity, but this contribution became increasingly positive as diversity in the communities increased. DISCUSSION Additive scaling assumes that the amount of herbivory experienced by an average individual of a given plant species will be the same, irrespective of what other species occur with it in the community. If so, then the total herbivory of the plant community is an abundanceweighted sum of the monoculture levels. Our results contradict this hypothesis, with nonadditive effects accounting for up to 25% of the measured herbivory in communities in the tests of additive scaling (i.e., M1 and M2). Although the percentage of additive effects was always larger than that of nonadditive effects, the percentage of measured herbivory accounted for by nonadditive effects was surprisingly high. Even for M4, designed to explain as much variation as possible by additive effects of plant traits, nonadditive effects still accounted for 10% of the variation in herbivory, and the deviance from the additive models increased with species richness, indicating an unexplained increase in observed herbivory in more diverse plant communities. Furthermore, the results of models M3 and M4 suggest that the relative importance of plant traits changes depending on which combinations of species occur in a community. Therefore, all four models confirm the occurrence of important nonadditive effects, which is in line with other studies in experimental and seminatural grassland systems showing that community properties such as diversity, evenness, or species richness were correlated with levels of invertebrate herbivory (Scherber et al. 2006a, b, 2010b, Unsicker et al. 2006, Stein et al. 2010, Allan and Crawley 2011). It seems that the choice to feed TABLE 2. Multiple regression model of the loge-transformed leaf standing herbivore damage measured in 70 plant communities at the field site of the Jena Experiment, Germany. Trait Intercept Leaf nitrogen concentration Leaf lignin concentration Coleopteran herbivores Hemipteran herbivores Leaf lifespan Stem growth form Root architecture Regression coefficient 6 SE 4.202 0.381 0.042 0.356 0.037 0.222 0.005 0.227 6 6 6 6 6 6 6 6 1.212 0.327 0.034 0.185 0.130 0.124 0.004 0.153 P Magnitude ,0.001 0.248 0.216 0.059 0.776 0.078 0.194 0.143 0.35 0.75 0.85 0.07 0.44 0.47 0.45 Notes: Damage was determined by seven community-weighted traits that are important to predict herbivory in monocultures (M3). The regression parameters were freely estimated. The regression coefficient shows the intercept or partial slope (for the complete model, R 2 ¼ 0.254, P , 0.001). ‘‘Magnitude’’ is the magnitude of the maximal effect of a trait on herbivory: the absolute value of the partial slope multiplied by the range (maximum–minimum) of the trait. Traits in the table are in decreasing order of magnitude score from the monoculture results (Loranger et al. 2012). 1506 JESSY LORANGER ET AL. Ecology, Vol. 94, No. 7 TABLE 3. Plant functional traits selected by a random forests (RF) approach to be important to predict loge-transformed leaf standing herbivore damage of 70 different communities at the field site of the Jena Experiment, Germany. Trait RF value P Intercept Leaf lignin concentrationa Leaf carbon concentration Coleopteran herbivores Stem growth forml Phosphorus leaf concentrationd Silica Ruderalc Aromatic compoundsj Aphid herbivores8 Period of seed sheddingg Orthopteran herbivorese Relative growth ratek Heighth Nitrogen leaf concentrationb Mollusc herbivoresi Lepidopteran herbivoresm Leaf primary fiber concentration SLAf Height summer Flowering phasen 71.8 62.0 52.2 41.8 40.4 38.8 36.4 35.9 31.2 28.9 28.8 28.5 26.7 26.5 26.3 25.5 24.4 24.1 23.4 21.0 ,0.001 0.984 ,0.001 ,0.001 0.357 0.854 ,0.001 0.779 0.325 0.080 0.414 0.797 0.434 0.555 0.837 0.342 0.367 ,0.001 0.784 ,0.001 0.075 Regression coefficient 6 SE 16.918 6 0.018 6 0.570 6 1.00 6 1.466 6 0.022 6 1.824 0.003 0.129 0.240 0.416 0.005 Magnitude 1.58 1.37 1.00 1.41 1.14 Notes: Traits that remained in the model after a stepwise backward selection (M4) are in boldface. RF value is the importance scores given by the RF for each trait. Superscript letters (a–o) indicate the order in which the traits have been removed in the backward stepwise selection. P values are for the intercept and partial slopes of the traits in the multiple regression of logetransformed herbivore damage in communities against each trait at the time it was removed by backward stepwise selection (P . 0.05) or in the final resulting model (P , 0.05). The regression coefficient shows intercept and partial slopes of the remaining traits, with standard error (in the final model, R 2 ¼ 0.553, P , 0.001). Magnitude is the magnitude of the maximal effect of an explanatory variable on the response variable calculated as the absolute value of the partial slope multiplied by the range (maximum – minimum) of the explanatory variable. Where cells are blank, data were not applicable. on a particular species depends both on traits of that species and on traits of surrounding species. Why would trait–herbivore relationships that were strong and robust predictors when herbivores were confronted with monocultures not apply as well when herbivores were confronted with multispecies vegetation? Why would predictions from additive models systematically underestimate actual levels of herbivory, and even more so as plant diversity increases? There are at least five possible explanations, the first three being related more directly to the herbivores’ reaction to increasing biodiversity and the last two being more related to effects of plant–plant interactions. First, it is likely that insect herbivores do not choose the vegetation to feed on at the plot scale (5 3 7 m2), but at a much finer scale, choosing individual plants from a heterogeneous mixture of food sources differing in quality. Consequently, the insect would not perceive a mixture of a high- and low-quality species in a plot as a ‘‘vegetation’’ of intermediate quality, as predicted by additive scaling hypothesis. In a monoculture of intermediate quality, levels of herbivory would also be intermediate. However, in the mixture of low- and high-quality species, herbivores could concentrate feeding on the high-quality species because of feeding preferences. Imagine that in monocultures containing 10 g of leaves, the herbivore eats 1 of 10 g of species A (the more palatable species) and 0.5 g of 10 g of species B. The predicted herbivory in a 1:1 mixture of the two species would be 0.5 3 10% þ 0.5 3 5% ¼ 7.5%. In the actual mixture, the herbivore concentrates its feeding entirely on the preferred species A, eating 1 g of the 5 g while ignoring the 5 g of leaves of species B. The observed herbivory would be 1 g of the total 10 g: thus 10%. If so, then levels of community herbivory would be close to the levels of the monocultures of the most attractive species. As a consequence, additive scaling models would underestimate herbivory in mixed communities, as we observed. Second, generalist herbivores can change their preference for food plants based on the composition of available plant species. For example, diverse communities also provide a diverse nutritional regime, thus potentially diluting deterrent chemicals or improving the nutrient balance of generalists (Bernays and Bright 1993). These effects can contribute to the decrease in importance of nutritional characteristics (leaf nitrogen and lignin concentration) in predicting herbivory in communities compared to herbivory in monocultures. Third, the community of generalist and specialist herbivores can change with plant composition, thus changing herbivore loads and levels of herbivory and their relations to plant traits. In fact, there is a positive correlation between higher diversity and abundance of July 2013 COMMUNITY PROPERTIES AFFECT HERBIVORY herbivores and plant diversity in the Jena Experiment (Scherber et al. 2010a). A likely mechanism is that increasing plant diversity leads via increased plant architectural complexity to more protection for more types of herbivores by providing more hiding and resting places (Southwood et al. 1979, Lawton 1983). The higher importance of architectural than nutritionalquality traits in M3 is consistent with this mechanism. In addition, higher plant diversity increased plant productivity in the Jena Experiment (Marquard et al. 2009) and elsewhere (Naeem et al. 1994, Tilman et al. 1996). Such a higher primary productivity has been shown to correlate with higher herbivore pressure (Haddad et al. 2001). Fourth, the effect of trait values of a given plant species can be modulated by traits of other surrounding plants. For example, the level of herbivory in a community could be partly driven by magnet species (or their absence) that initially attract herbivores in a similar way as documented for plant–pollinator interactions (Johnson et al. 2003). For instance, highly nutritious species could attract herbivores to the site. Once the preferred plant species has been consumed, it might be advantageous for the herbivore to feed on surrounding less nutritious species rather than leaving the site and using time to search for a new site or accepting risks associated with moving to a new site. This spillover effect (White and Whitham 2000) has been found in the Jena Experiment, where the presence of legumes increased the rates of invertebrate herbivory on the other species (Scherber et al. 2006b). An associational resistance between plant species can also occur, leading to a decrease of herbivore damage on attractive plant species due to interference in detection by associated non-host plant species (Tahvanainen and Root 1972, Hambäck et al. 2000, Finch and Collier 2012). Fifth, the trait values expressed by a species can change depending on the other species that are growing in close proximity. For example, where legume species are present in the Jena Experiment, leaf nitrogen concentration in grasses may increase (Gubsch et al. 2011). It has also been shown in a case study with Plantago lanceolata in the Jena Experiment that the allocation to chemical defense compounds may change with increasing diversity (Mraja et al. 2011). Although trait variation certainly is an important mechanism causing the observed nonadditive effects, the values of several categorical traits taken from the literature that were important in monocultures do not change across communities (e.g., root architecture or leaf life span). Consequently, trait variation cannot be the only mechanisms causing nonadditive effects. All five of these mechanisms can also contribute to the observed differences in the choice of traits that were selected in the models. The most striking difference is that leaf nitrogen concentration, which was the most important predictor in monocultures, was not selected in multispecies communities. This could be caused by 1507 nutritional quality being generally less important in mixed vegetation compared to a single plant species, due to selection of attractive plants and mixing of different resources as previously explained. A similar argument can explain why leaf life span was selected in the monoculture model (where it quantifies the seasonal availability of foliage) but was not selected in mixtures where temporal complementarity between species means that there is always foliage available. As for root architecture, it had a positive effect on herbivory in monoculture. However, its communityweighted value in mixtures was related to the relative abundance of grasses, which can decrease herbivory (H. Loranger et al., unpublished manuscript). These contrary effects may have canceled each other in communities. The effect of grasses on herbivory in mixtures might also account for silica being selected as a significant trait, as Poaceae is the main plant family defended by silica. CONCLUSION The hypothesis of additive scaling of herbivory was rejected and nonadditive effects were detected in each model. This indicates that complex plant–insect interactions are of importance in determining the levels of herbivory in multispecies communities. Which of the proposed potential mechanisms causes predictions of community herbivory to deviate from our models remains to be investigated, concentrating on (1) the variation of relevant traits, (2) changes in the herbivore community along the diversity gradient, and (3) direct investigations on feeding preferences and behavior of important herbivore groups. To do this, some of the unmeasured potential causes described here (or traits associated with them) should be included in future models: vegetation properties such as canopy height, biomass, diversity, and/or the distinctiveness (uniqueness) of the target species. Such an approach can also help one to understand how environmental conditions (abiotic and biotic) interact with the traits governing the level of herbivory experienced by a plant. ACKNOWLEDGMENTS We thank Anne Ebeling, the gardeners, and technical staff who have worked on the Jena Experiment for maintaining the site. The Jena Experiment was funded by the Deutsche Forschungsgemeinschaft (FOR 1451). We thank Enrica de Luca, who provided biomass data, and Annett Lipowsky and Marlén Gubsch, who provided some plant trait data. This study was funded by the Natural Sciences and Engineering Research Council of Canada (NSERC), the Fonds Québécois de Recherche sur la Nature et les Technologies (FQRNT), and the AquaDiva@Jena project financed by the state of Thuringia. The study has also been supported by the TRY initiative on plant traits, and we thank all the contributors who have provided trait data via the TRY database (www.trydb.org). TRY is/has been supported by DIVERSITAS, IGBP, the Global Land Project, the U.K. Natural Environment Research Council (NERC) through its program QUEST (Quantifying and Understanding the Earth System), the French Foundation for Biodiversity Research (FRB), and GIS ‘‘Climat, Environnement et Société’’ France. 1508 JESSY LORANGER ET AL. LITERATURE CITED Allan, E., and M. J. Crawley. 2011. Contrasting effects of insect and molluscan herbivores on plant diversity in a long-term field experiment. Ecology Letters 14:1246–1253. Bernays, E. A., and K. L. Bright. 1993. Mechanisms of dietary mixing in grasshoppers—a review. Comparative Biochemistry and Physiology A Physiology 104:125–131. Breiman, L. 2001. Random forests. Machine Learning 45:5–32. Burnham, K. D., and D. R. Anderson. 2010. Model selection and multimodel inference. A practical information-theoretic approach. Second edition. Springer, New York, New York, USA. Coley, P. D., and J. A. Barone. 1996. Herbivory and plant defenses in tropical forests. Annual Review of Ecology and Systematics 27:305–335. Ellenberg, H. 1996. Vegetation Mitteleuropas mit den Alpen in ökologischer, dynamischer und historischer Sicht. Fifth edition. Verlag Eugen Ulmer, Stuttgart, Germany. FAO. 1988. FAO/UNESCO soil map of the world. Revised legend with corrections and updates. World Soil Resources Report 60, FAO, Rome, Italy. Reprinted in 1997 with updates as Technical Paper 20, ISRIC [International Soil Reference and Information Centre], Wageningen, The Netherlands. Finch, S., and R. H. Collier. 2012. The influence of host and non-host companion plants on the behaviour of pest insects in field crops. Entomologia Experimentalis et Applicata 142: 87–96. Fortunel, C., et al. 2009. Leaf traits capture the effects of land use changes and climate on litter decomposability of grasslands across Europe. Ecology 90:598–611. Garnier, E., et al. 2007. Assessing the effects of land-use change on plant traits, communities and ecosystem functioning in grasslands: A standardized methodology and lessons from an application to 11 European sites. Annals of Botany 99:967– 985. Gubsch, M., N. Buchmann, B. Schmid, E. D. Schulze, A. Lipowsky, and C. Roscher. 2011. Differential effects of plant diversity on functional trait variation of grass species. Annals of Botany 107:157–169. Haddad, N. M., D. Tilman, J. Haarstad, M. Ritchie, and J. M. H. Knops. 2001. Contrasting effects of plant richness and composition on insect communities: a field experiment. American Naturalist 158:17–35. Hambäck, P. A., J. Agren, and L. Ericson. 2000. Associational resistance: insect damage to purple loosestrife reduced in thickets of sweet gale. Ecology 81:1784–1794. Huntly, N. 1991. Herbivores and the dynamics of communities and ecosystems. Annual Review of Ecology and Systematics 22:477–503. Johnson, S., C. Peter, and L. Nilsson. 2003. Pollination success in a deceptive orchid is enhanced by co-occurring rewarding magnet plants. Ecology 84:2919–2927. Kattge, J., et al. 2011. TRY—a global database of plant traits. Global Change Biology 17:2905–2935. Kleyer, M., et al. 2008. The LEDA traitbase: a database of lifehistory traits of the Northwest European flora. Journal of Ecology 96:1266–1274. Klotz, S., I. Kühn, and W. Durka. 2002. BIOLFLOR—Eine Datenbank mit biologisch-ökologischen Merkmalen zur Flora von Deutschland. Schriftenreihe für Vegetationskunde 38:1–334. Kuhn, I., W. Durka, and S. Klotz. 2004. BiolFlor—A new plant-trait database as a tool for plant invasion ecology. Diversity and Distribution 10:363–365. Laliberté, E., and B. Shipley. 2011. Package ‘‘FD.’’ Measuring functional diversity (FD) from multiple traits, and other tools for functional ecology. R package version 1.0-11. Ecology, Vol. 94, No. 7 Lawton, J. H. 1983. Plant architecture and the diversity of phytophagous insects. Annual Review of Entomology 28:23– 39. Loranger, J., S. T. Meyer, B. Shipley, J. Kattge, H. Loranger, C. Roscher, and W. W. Weisser. 2012. Predicting invertebrate herbivory from plant traits: evidence from 51 grassland species in experimental monocultures. Ecology 93:2674– 2682. Marquard, E., A. Weigelt, C. Roscher, M. Gubsch, A. Lipowsky, and B. Schmid. 2009. Positive biodiversity– productivity relationship due to increased plant density. Journal of Ecology 97:696–704. Mraja, A., S. B. Unsicker, M. Reichelt, J. Gershenzon, and C. Roscher. 2011. Plant community diversity influences allocation to direct chemical defence in Plantago lanceolata. PloS One 6:e28055. Naeem, S., L. J. Thompson, S. P. Lawler, J. H. Lawton, and R. M. Woodfin. 1994. Declining biodiversity can alter the performance of ecosystems. Nature 368:734–737. Pakeman, R. J., et al. 2008. Impact of abundance weighting on the response of seed traits to climate and land use. Journal of Ecology 96:355–366. Pakeman, R. J., J. Leps, M. Kleyer, S. Lavorel, and E. Garnier. 2009. Relative climatic, edaphic and management controls of plant functional trait signatures. Journal of Vegetation Science 20:148–159. R Development Core Team. 2009. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www. r-project.org Rasmann, S., and A. Agrawal. 2009. Plant defense against herbivory: progress in identifying synergism, redundancy, and antagonism between resistance traits. Current Opinion in Plant Biology 12:473–478. Roscher, C., B. Schmid, N. Buchmann, A. Weigelt, and E. D. Schulze. 2011a. Legume species differ in the responses of their functional traits to plant diversity. Oecologia 165:437– 452. Roscher, C., J. Schumacher, J. Baade, W. Wilcke, G. Gleixner, W. W. Weisser, B. Schmid, and E. D. Schulze. 2004. The role of biodiversity for element cycling and trophic interactions: an experimental approach in a grassland community. Basic and Applied Ecology 5:107–121. Roscher, C., S. Thein, A. Weigelt, V. M. Temperton, N. Buchmann, and E. D. Schulze. 2011b. N2 fixation and performance of 12 legumes species in a 6-year grassland biodiversity experiment. Plant and Soil 341:333–348. Scherber, C., et al. 2010a. Bottom-up effects of plant diversity on multitrophic interactions in a biodiversity experiment. Nature 468:553–556. Scherber, C., J. Heimann, G. Kohler, N. Mitschunas, and W. W. Weisser. 2010b. Functional identity versus species richness: herbivory resistance in plant communities. Oecologia 163:707–717. Scherber, C., A. Milcu, S. Partsch, S. Scheu, and W. W. Weisser. 2006a. The effects of plant diversity and insect herbivory on performance of individual plant species in experimental grassland. Journal of Ecology 94:922–931. Scherber, C., P. N. Mwangi, V. M. Temperton, C. Roscher, J. Schumacher, B. Schmid, and W. W. Weisser. 2006b. Effects of plant diversity on invertebrate herbivory in experimental grassland. Oecologia 147:489–500. Southwood, T. R. E., V. K. Brown, and P. M. Reader. 1979. Relationships of plant and insect diversities in succession. Biological Journal of the Linnean Society 12:327–348. Stein, C., S. B. Unsicker, A. Kahmen, M. Wagner, V. Audorff, H. Auge, D. Prati, and W. W. Weisser. 2010. Impact of invertebrate herbivory in grasslands depends on plant species diversity. Ecology 91:1639–1650. July 2013 COMMUNITY PROPERTIES AFFECT HERBIVORY Tahvanainen, J., and R. B. Root. 1972. The influence of vegetational diversity on the population ecology of a specialized herbivore, Phyllotreta cruciferae (Coleoptera: Chrysomelidae). Oecologia 10:321–346. Tilman, D., D. Wedin, and J. Knops. 1996. Productivity and sustainability influenced by biodiversity in grassland ecosystems. Nature 379:718–720. 1509 Unsicker, S. B., N. Baer, A. Kahmen, M. Wagner, N. Buchmann, and W. W. Weisser. 2006. Invertebrate herbivory along a gradient of plant species diversity in extensively managed grasslands. Oecologia 150:233–246. White, J. A., and T. G. Whitham. 2000. Associational susceptibility of cottonwood to a box elder herbivore. Ecology 81:1795–1803. SUPPLEMENTAL MATERIAL Appendix A Species pool of the Jena Experiment with inclusion status in the analyses (Ecological Archives E094-136-A1). Appendix B Detailed list of the 42 traits initially considered to predict leaf standing herbivore damage and graph of predicted against measured values of herbivory in monocultures (Ecological Archives E094-136-A2). Appendix C Description of the Random Forests (RF) analysis and graph of the results, explaining how traits were selected (Ecological Archives E094-136-A3). Appendix D Correlation matrix of the different community properties of the communities in the Jena Experiment (Ecological Archives E094-136-A4). Appendix E Log-response ratios for nonadditive and additive effects based on the different models (Ecological Archives E094-136-S1).
© Copyright 2026 Paperzz