The premise that the intensity of biotic interactions decreases with increasing latitudes and elevations is broadly accepted; however, whether these geographical patterns can be explained within a common theoretical framework remains unclear. Our goal was to identify the general pattern of elevational changes in trophic interactions and to explore the sources of variation among the outcomes of individual studies. Meta-analysis of 226 effect sizes calculated from 136 publications demonstrated a significant but interaction-specific decrease in the intensity of herbivory, carnivory and parasitism with increasing elevation. Nevertheless, this decrease was not significant at high latitudes and for interactions involving endothermic organisms, for herbivore outbreaks or for herbivores living within plant tissues. Herbivory similarly declined with increases in latitude and elevation, whereas carnivory showed a fivefold stronger decrease with elevation than with latitude and parasitism increased with latitude but decreased with elevation. Thus, although these gradients share a general pattern and several sources of variation in trophic interaction intensity, we discovered important dissimilarities, indicating that elevational and latitudinal changes in these interactions are partly driven by different factors. We conclude that the scope of the latitudinal biotic interaction hypothesis cannot be extended to incorporate elevational gradients.

Search for and processing of studies

We focused our meta-analysis on trophic interactions in terrestrial ecosystems that are broadly defined as predation (i.e. as consumption of one organism by another organism) and classified into herbivory, carnivory and parasitism, as in previously conducted the meta-analysis of latitudinal changes in trophic interactions (Zvereva & Kozlov 2021). We extracted references from earlier reviews of elevational changes in trophic interactions (Hodkinson 2005; Andrew et al. 2012; Péré et al. 2013; Sundqvist et al. 2013; Moreira et al. 2018; Carmona et al. 2011, 2020), and we then searched for additional publications in the ISI Web of Science using the keywords ‘elevation*’, ‘geographic*’, ‘biotic interactions’, ‘herbivor*’, ‘predat*’, ‘carnivor*’ and ‘parasit*’. The search was completed on 11 January 2022. We did not use unpublished data or grey literature.

To enable comparisons between elevational and latitudinal patterns, we followed the methodology used in our previous meta-analysis (Zvereva & Kozlov 2021). We considered studies containing direct quantitative estimates of the intensity of herbivory (the percentage of plant biomass or leaf area lost to herbivores or the proportion of damaged leaves, shoots, flowers, seeds or plants), carnivory (the mortality of prey or the predator attack rates) or parasitism (prevalence; i.e. the percentage of infected hosts). We did not include studies where the interaction intensity was deduced from the abundance of herbivores or predators, because abundance may show variable relationships with the intensity of their impact on plants and on prey (Bito et al. 2011; Tela et al. 2021). We also excluded studies that did not contain data collected from individual elevational gradients but had quantified elevational changes in biotic interactions by combining data obtained in different mountain regions (e.g. Roslin et al. 2017).

We extracted information from studies that fit the following criteria: (i) the data were collected from natural ecosystems, (ii) the data were collected from at least two study sites with elevation differences of at least 100 m within the same mountain region, and (iii) the magnitude of the effect could be calculated from the data or statistics presented in the publication or provided by the authors. From multiyear studies, we extracted the combined result for all years if it was presented in the publication. If the data collected in different years were not combined by the authors, then we selected the year with the highest average value of the character under study. If the study employed some manipulations (e.g. enemy exclusion, insecticide treatment, water treatment), then we selected a control treatment.

Classificatory variables

Whenever possible, herbivores and carnivores were divided into ectotherms (invertebrates only; we found no data on carnivory by amphibians or reptiles) and endotherms (birds and mammals). All parasites were ectotherms; therefore, they were excluded from this comparison. We classified host plants and prey into natural, permanently inhabiting the study areas, and standardized, i.e. introduced to all study sites by the researchers (e.g. sunflower seeds or artificial prey).

The herbivory level was considered to be background unless the authors explicitly mentioned an outbreak. Herbivory was divided into folivory (consumption of leaves, sometimes with their supporting branches) and granivory (i.e. seed predation). Folivory was divided into mammalian grazing and invertebrate folivory, whereas granivory was divided into pre- and post-dispersal seed predation. Invertebrate herbivores were divided into exophagous (defoliators and sap-feeders) and endophagous (miners, gallers, borers) feeders; the latter group also included pre-dispersal seed feeders.

We divided studies of herbivory into those reporting elevational changes in herbivory on all plant species in a site or on several dominant species (community-wide herbivory hereafter) and those reporting herbivory in a certain plant species or genus (species-specific herbivory hereafter). Host plants were classified as woody or herbaceous, and woody plants were classified as evergreen (with foliage that remains green and functional through more than one growing season) or deciduous.

Elevational gradients were classified as those located entirely below or above the tree line and those crossing the tree line. We calculated the elevational span of a gradient as the difference in elevation between the highest and lowest sites. The geographical coordinates of study areas (to the nearest degree of latitude and longitude) were extracted from publications or searched on the internet based on information provided in the publication. The attribution of elevational gradients to climate zones (tropical, including subtropics; temperate; and polar, including boreal forests) was based on climate and vegetation at the foot of the mountain and was performed in the same way as in the meta-analysis of latitudinal patterns (Zvereva & Kozlov 2021).

Meta-analysis

We quantified the strength of the elevational gradients by the z-transformed correlation between elevation and the intensity of the interactions (z_r). We estimated the variation in effect sizes (ES) within groups by calculating the heterogeneity index (Q_t). To compare ES among different groups of studies, we calculated the between-group heterogeneity (Q_B) using a random effects model, and we tested Q_B against the χ² distribution with the number of groups minus one degree of freedom (Koricheva et al. 2013).

We used three approaches for the ES calculation. When a study reported the data from two or three sites, we calculated Hedge's d based on data from the lowest and highest sites. We then converted Hedge's d into a Pearson linear correlation coefficient (r) following an equation given by Lajeunesse (2013) and calculated the variance of z_r based on sample size (Rosenberg et al. 2013). When the number of sites was four or more, we extracted or calculated the Pearson linear correlation coefficient between the intensity of the trophic interaction and elevation and converted r into a z_r value. If the authors provided the F statistic, we transformed it into z_r using a Metawin calculator (Rosenberg et al. 2000). The ES calculated in different ways were of similar magnitudes (Q_B = 1.09, df = 2, P = 0.58), thereby justifying the combination of ES calculated by these methods in our analyses.

We explored the effects of latitude and gradient span on elevational changes in the intensity of trophic interactions by means of a meta-regression. We searched for publication bias by calculating the Kendall τ correlation between the standardised ES and sample size (Rosenberg et al. 2000); the significant correlation was interpreted as the presence of a small study effect hinting at publication bias (Jennions et al. 2013). Finally, we calculated Rosenthal’s fail-safe number, which shows the number of insignificant studies that are required to turn the significant mean ES into an insignificant one. The fail-safe numbers exceeding 5n (where n is the number of studies included in the meta-analysis) are considered as proof of robustness of the analysis against the insignificant results (Møller & Jennions 2001).