Spatial extent of neighboring plants influences the strength of associational effects on mammal herbivory. Insights from a meta-analysis

There is high variability in the level of herbivory between individual plants from the same species with potential effects on population dynamics, community composition, and ecosystem structure and function. This variability can be partly explained by associational effects i.e. the impact of the presence of neighboring plants on the level of herbivory experienced by a focal plant, but it is still unclear how the spatial scale of plant neighborhood modulates foraging choice of herbivores; an inherently spatial process in itself. Using a meta-analysis, we investigated how spatial scale modifies associational effects on the susceptibility to browsing by herbivores with movement capacities similar to deer. From 2496 articles found in literature databases, we selected 46 studies providing a total of 168 differences of means in damage by herbivores or survival to woody plants (mostly) with and without neighboring plants. Spatial scales were reported as distance between plants or as plot size. We estimated the relationships between the effect sizes and spatial scale, type of associational effects and nature of the experiment using meta-analysis mixed models. The strength of associational effects declined with increasing plot size, regardless of the type of associational effects. Associational defences (i.e. decrease in herbivory for focal plants associated with unpalatable neighbors) had stronger magnitude than associational susceptibilities. The high remaining heterogeneity among studies suggests that untested factors modulate associational effects, such as nutritional quality of focal and neighboring plants, density of herbivores, timing of browsing, etc. Associational effects are already considered in multiple restoration contexts worldwide, but a better understanding of these relationships could improve their use in conservation, restoration and forest exploitation when browsing is a concern. This study is the first to investigate spatial patterns of associational effects across species and ecosystems, an issue that is essential to determine differential herbivory damages among plants.

effect sizes and spatial scale, type of associational effects and nature of the experiment using 23 meta-analysis mixed models. The strength of associational effects declined with increasing plot 24 size, regardless of the type of associational effects. Associational defences (i.e. decrease in 25 herbivory for focal plants associated with unpalatable neighbors) had stronger magnitude than 26 associational susceptibilities. The high remaining heterogeneity among studies suggests that 27 untested factors modulate associational effects, such as nutritional quality of focal and 28 neighboring plants, density of herbivores, timing of browsing, etc. Associational effects are 29 already considered in multiple restoration contexts worldwide, but a better understanding of 30 these relationships could improve their use in conservation, restoration and forest exploitation 31 when browsing is a concern. This study is the first to investigate spatial patterns of associational 32 Introduction 37 Herbivory can modify the composition, structure and functions of ecosystems (Hester et al. 38 2006). There is high variability in the susceptibility of different plant species and individuals to 39 herbivory. This variability is driven by forage selection, whom in itself is determined by the 40 nutritional requirements of herbivores (Pyke et al. 1977), intrinsic (e.g. nutritive quality, Pyke et 41 al. 1977), and extrinsic characteristics of both the plants and the environment (e.g. neighboring 42 plants, Atsatt and O'Dowd 1976). Multiple studies have demonstrated the influence of 43 neighboring plants on forage selection, a process named neighboring or associational effects 44 (Milchunas and Noy-Meir 2002, Barbosa et al. 2009), yet the conditions in which a specific 45 neighborhood will increase or reduce herbivory are not fully understood. The distance between 46 neighboring plants could explain part of the residual variability observed in associational effects 47 susceptibility and associational defence) or the "contrast" type (neighbor contrast defence and 106 susceptibility), according to the conceptual framework provided by Bergvall et al. (2006). We 107 thus predicted an interaction between distance and associational effect type (Figure 1b) where 108 associational susceptibility or defence would be more frequent at larger spatial scales (home 109 ranges, patches) when herbivore select resources based on the relative abundance of resources, 110 while "neighbor contrast" would be more frequent once herbivores are feeding within a patch 111 and selecting individual plant species. This study is the first to investigate how spatial scale 112 drives associational effects across herbivore species and ecosystems, an issue essential for 113 understanding variations in the level of herbivory incurred by individuals within a population 114 (Barbosa et al. 2009. 115

Literature review 117
We obtained 2496 peer-reviewed publications using the search strategy presented in Appendix A 118 in ISI Web of Science, Biosis preview and BioOne (in July 2013), and through citations found in 119 these publications. We searched for studies involving herbivores with movement capacities 120 similar to deer from the smallest to the largest deer species; the smallest herbivore in our dataset 121 is European roe deer and the largest is the European bison (Bison bonasus). Studies reported data 122 on damage or survival of plants (hereafter called the focal plants) with and without the presence 123 of a neighboring plant (hereafter called the neighbor plant). Damage was inferred from counts of 124 browsed twigs or leaves, or biomass removal and did not include measures of growth or 125 regrowth following herbivory. We included studies using feeding trials in controlled or natural 126 environments, transplantation/removal of neighbors and observations in natural environments.
We established the criteria regarding acceptance or rejection of a study prior to conducting the 128 meta-analysis using a PRISMA inspired protocol (see process in Appendix A, Moher et al. 2009). 129 The criteria were the presence of a control treatment (herbivory without neighboring plant), a 130 palatable plant in the focal-neighbor group, and a difference in palatability between plants. To 131 evaluate the effect of spatial scale, each study needed to clearly state the size of the plot where 132 data were recorded or the distance between the focal and neighboring plant. We rejected data on 133

Statistical analyses 187
We tested the impact of independent variables on the standardized difference of mean (d) in 188 three meta-analysis mixed models using the function rma of the metafor package (Viechtbauer 189 2010) in R 3.1.2 (R Core Team 2013). For our first objective, we used the complete dataset to 190 test the variation in effect size depending on the direction of the association (susceptibility, 191 defence; figure 1a), type of association ("classic": associational defence/associational 192 susceptibility, "contrast": neighbor contrast defence/neighbor contrast susceptibility; figure 1a) 193 and interaction between direction and type of association. We also included the nature of the 194 experiment (feeding trial, observation study, transplantation or removal experiments) since effect 195 sizes from controlled experiments such as feeding trials could be stronger than results of 196 observational studies where foraging by herbivores would be influenced by uncontrolled factors. 197 The conversion of OR and r in d could have generated a bias in the values of the effect sizes; we 198 tested this supposition in a simple model with effect size class (d, r or OR) as an independent 199 variable. Since effect size class did not influence the value of d (d-class compared to OR-class: z 200 = -0.2, p = 0.8; compared to r-class: z = -0.5, p = 0.6), we did not include it in our final model. 201 For our second objective, we tested the effect of spatial scale on associational effect strength for 202 plot-based and distance-based studies separately. We log-transformed plot size to control for its 203 large dispersion (Bland and Altman 1996). For both models, together with the variables 204 describing the linear and quadratic parameters for the spatial scale (log plot size or linear 205 distance), we included the type of association and their interactions to test for predictions of 206 higher frequency of "classic" interaction at a finer scales and higher frequency of "contrast" 207 interaction at a larger scales (Figure 1b). Both models also included the nature of the experiment 208 as an independent variable to control for differences in effect sizes from different experiments. 209 The function rma weights effect sizes using the inverse-variance method for mixed models studies. The extracted data were equally distributed between decreased and increased herbivory 237 with neighboring plant, but "classical" types (associational defence and associational 238 susceptibility, n = 104) were more frequent than "contrast" types (neighbor contrast defence and 239 neighbor contrast susceptibility, n = 47). Most effect sizes resulted from feeding trials (n = 71), where various assemblages were proposed to herbivores, but 54 came from observational studies 241 and 38 from transplantation experiments. Removal experiments were rarely used (n = 5). 242 Additional summary data can be found in Appendix B. 243 The first model using the complete dataset explained 23% of the heterogeneity between effect 244 sizes (omnibus test for independent variables: Q df = 8 = 50.0, p < 0.0001) and the pseudo-R 2 for 245 the model reached 23.0%. There was, however, a high residual heterogeneity in the model (test 246 for residual heterogeneity: Q df = 159 = 1047.0, p <0.0001). Effect sizes for defence associational 247 effects (associational defence and neighbor contrast defence) had a greater magnitude than 248 susceptibility associational effects (associational susceptibility and neighbor contrast 249 susceptibility; Figure 2). Classic associational effects also had a greater value than contrast 250 associational effects (Figure 2). Except for the contrast level of associational effects, all I 2 were 251 above 70%, indicating the presence of untested variables (Figure 2). Transplantation experiments 252 presented the strongest and more variable values of d, while feeding trials found consistently 253 small associational effects ( Figure 2); values for observational studies were intermediate ( Figure  254 2) . 255 The model of the effect of plot size on associational effects explained 68% of the heterogeneity 256 (omnibus test for independent variables Q df= 9 = 28.5, p = 0.0008, pseudo-R 2 = 19.6 %) but also 257 presented high remaining heterogeneity (Q df = 86 = 312.9, p <0.0001). As the log-plot size 258 increased, there was a linear decrease in the strength of associational effects (Figure 3a found some evidence of potential publication bias in funnel plots for the entire dataset and used 282 the trim and fill method to test the robustness of the overall mean effect size (Appendix C). The 283 trim and fill method identifies and correct the asymmetry by imputing smaller effect sizes around 284 an estimated true center (Viechtbauer 2010). For the entire dataset, the trim and fill method 285 generated more values of associational susceptibilities, suggesting either a publication bias in the analyses or a naturally higher frequency of associational defences (Appendix C). In addition, our 287 analyses revealed potential publication bias among the effect sizes calculated as difference of the 288 means (effect size of class d) and in observational studies (Appendix C). Even with input values, 289 the d-class subgroup mean is similar to the r and OR-class subgroups and thus should not modify 290 our conclusions. The trim and fill method suggests more associational susceptibilities in the 291 observational studies subgroup, but this asymmetry could also result from the higher natural 292 occurrence of associational defences. We found no evidence of a temporal trend (Appendix C). 293

Discussion 294
Using a meta-analysis based on 46 studies and 168 data points on associational effects of 295 neighboring plants on the level of herbivory, we found a decrease in associational effect strength 296 with spatial scale. In contradiction with our hypothesis, the decrease was independent of the type 297 of associational effect (i.e. "classic" or "contrast" type). We also found that associational 298 defences had stronger effects than associational susceptibilities. There is a common agreement 299 that hierarchical forage selection has been overlooked in associational effect studies (Barbosa et  The descriptors of spatial scale, i.e. presence of neighbors in a plot or distance between focal and 303 neighbor, highly influenced the relation between scale and associational effects. Distance 304 between plants is a one-dimensional measure, mostly used when studying the relationships 305 between two individual plants (e.g. nurse plant studies or in feeding trials). This is reflected by 306 the small range of distances in our dataset. When considering those simple interactions, 307 associational effects declined quickly with increased distance between the plants. Typical We did not find support for the predictions that "classic" effects should influence patch choice 324 by herbivores while "contrast" effects should affect within patch selection (Bergvall et al 2006). 325 Because few associational effects reported were measured in large patches, the model could have 326 been unable to detect an interaction between type of association and distance. Every type of 327 effects could also be seen at all scales because of the additive effects of herbivore selection at 328 multiple scales (Miller et al. 2006 scales (between patches > between feeding stations > within feeding stations) with squirrels 337 (Sciurus spp.), and found that both neighbor contrast susceptibility and associational defence 338 occur among patches and among feeding stations. At a larger scale, they found only associational 339 defence; high palatability seeds were less susceptible to be consumed in low palatability patches. 340 The study of associational effects could be greatly improved by more experimentation with 341 varying patch size and distance between neighbors, which could test the extent of associational 342 susceptibilities and defences such as the study by Oom and Hester (1999). 343 Associational defences had stronger effects than associational susceptibilities, thereby suggesting 344 stronger effects of facilitation. Facilitation between plants is known to be common in stressful 345 environments, such as those with high herbivory pressure (Callaway and Walker 1997). High 346 herbivory pressure, however, can also reduce the impact of associational defences, as herbivores 347 could become less selective when competition between individuals increases (Baraza et al. 2006). 348 Some studies have demonstrated a relation between herbivory pressure and associational effects 349 (Aerts et al. 2007, Graff et al. 2007, Smit et al. 2007), but the heterogeneity in reporting 350 herbivore pressure prevented us to test this factor. "Classic" type of associational effects also 351 presented stronger effects than "contrast" type. Although Atsatt and O'Dowd (1976) introduced 352 the attractant-decoy hypothesis 40 years ago, interest in contrast associational effects is more 353 recent (see Bergvall et al. 2006) and they might be understudied; only 47 of our effect sizes 354 concerned "contrast" interactions. standardized difference of means) separated by the independent variable levels tested, with 95% 585 CI and I 2 , the percentage of total variability due to heterogeneity among d's. A higher d indicates 586 a higher associational effect of the neighboring plant on the focal plant's herbivory level. 587 Numbers to the right of the data points are the number of effect sizes in each summary effect. 588 We used a meta-analysis mixed model to test the impact of variables on the standardized 589 difference of means. 590