Study region

Mount Kilimanjaro was selected as the study system. It is located in Northern Tanzania at 3.1°S 37.4°E. It is the highest free-standing mountain in the world and covers an area of approx. 4000 km². The elevation gradient ranges from the lowlands at 800 m a.s.l. to the peak at 5892 m a.s.l. Precipitation values range from 550 mm•a^-1 to 3600 mm•a^-1, with the highest amount of rainfall occurring at mid-elevations in the forest belt and the lowest amounts in the alpine zone and the plains surrounding the mountain [41]. Anthropogenic disturbance is expressed differently upon the mountain. Fires, occasional timber extraction, and small-scale collection of other forest products affect higher elevations, while agriculture involving fertilizers and herbicide application is practiced in lowland areas.

Data collection

All necessary permits were granted from the Tanzanian Commission for Science and Technology, the Tanzanian Wildlife Authority, and Tanzania National Parks (340-ER-NA-96-44, TNO/HQ/C.10/13/VOL.III).

Collection of plant functional traits and data on animal taxa took place on 60 plots, each of 0.25 ha, in the twelve major vegetation types at the mountain between August 2010 and November 2012 (S1 Table). Plots were distributed equally among vegetation types, five plots belonging to each type.

Undisturbed plots at low elevations around 1000 m a.s.l. were dominated by annual grasses and drought-tolerant trees, whereas weeds characterize cultivated areas. At 1600 m, coffee plantations and agroforestry systems harbored mostly dicotyledonous weeds. Grasslands were dominated by several taxa of Poaceae. Woody life forms accounted for the largest part of plant biomass from middle to upper elevations, with a transition from rainforest to cloud forest characterized by small trees and high lichen abundance at around 3000 m. Alpine vegetation occurred up to 4500 m and was mainly composed of shrubs, perennial herbs, and grasses.

Mean temperature was derived from several years of continuous measurements with automatic data loggers covering the time period of our data collection [42]. Annual precipitation data was derived from the Kilimanjaro rainfall model [43]. Disturbance was calculated as a composite metric including the effects of land use at local and landscape scales ([44], S1 File).

Plant functional traits were collected from the most abundant vascular plant species, accounting for 80% of total plant biomass on each plot. Abundance was defined as percent cover determined from vegetation surveys. Cover was estimated for each stratum depending on the vegetation structure of the plot. Cover of all strata was summed up for each species and standardized to a plot total of 100%. Plant functional traits were chosen to indicate vegetative growth, persistence, and reproductive characteristics. Fifteen individuals per species were sampled from different plots within the elevational distribution range for specific leaf area (SLA), leaf dry matter content (LDMC), stem specific density (SSD), leaf nitrogen content (leaf N_mass), leaf phosphorus content (leaf P_mass), fruiting frequency, fruit number, and fruit size. Additionally, the total plant biomass per plot was calculated with allometric equations using complete tree and undergrowth inventories (details in [45]). Sampling and processing of plant material followed Kleyer et al. [46].

Percentage cover values from vegetation surveys were used to calculate community-weighted mean trait values for each plot (CWM, [47]). To describe leaf palatability for moths, the first axis of a principal components analysis using SLA, LDMC, SSD, leaf N_mass and leaf P_mass was extracted and termed "leaf economics" (for correlations of variables see S2 Table). The traits used are related to the worldwide leaf economics spectrum [32, 48], reflecting a gradient from plants with fast resource use and nutrient turnover to those with slow and persistent growth. The former are characterized by soft tissues with high nitrogen content and are generally preferred by herbivores, due to their easier digestion and higher nutritional value [32]. All data were collected on the same plots within a common time frame to avoid confounding effects of spatial and temporal variability.

To quantify nectar and fruit availability on the plots, pollination and dispersal syndromes were extracted from the Flora of Tropical East Africa [49]. Nectar availability is difficult to quantify, as nectar amounts per flower, flowering times, and the proportion of tissue invested in flowers are highly variable between species. This resource was approximated by the abundance-weighted proportion of insect-pollinated plant species. To assess food resources for frugivorous birds, average fruit numbers per individual plant sampled in the field (fn), average fruit size (fs), fruiting frequency (ff), and relative abundance (ra) were used to calculate the bird-dispersed fruit CWM (bdc) according to the following formula:

The sum is taken over all plant species in the respective plot. δ_bdc equals 1 for species producing fruits consumed by birds and 0 otherwise. Birds were observed through point counts, both in the dry and wet season to include temporal variation (see [50] for methodology). Bees were sampled with pan traps on the forest floor and in the canopy of woody vegetation. Sampling was repeated several times to account for temporal variation (see [44]). Moths were caught with an automatic light trap with a superactinic light tube (6 watt, Fritz Weber Entomologiebedarf, Stuttgart, Germany) as light source. This was repeated in dry and wet seasons. Body mass of birds was derived from Dunning Jr [51] and used as a proxy for body size [52]. Within animal groups of similar body structure, body length and size are highly correlated [53]. For bees, body size was approximated by the highly correlated intertegular distance (ITD, [54]). Moth body length was measured using a binocular Leica stereomicroscope with a calibrated ocular micrometer.

As with plant traits, animal body size was weighted by species abundance, yielding a body size CWM for each taxon or guild, to avoid giving rare species the same weighting as abundant ones [21]. For simplicity, in the following community-weighted mean body size is referred to as body size. Cumulative body sizes of moths and bees as food resources for insectivorous birds were calculated as the sum of the individual body sizes of all individuals of the respective groups sampled. For plot means and ranges of plant functional traits, animal body size, cumulative bee ITD, and cumulative moth body length see S3 Table. Analyses were also performed with community means not weighted by abundance, but as there were no substantial differences, only CWM results are presented.

Data analysis

Structural equation modeling with mixed effect models was applied to relate mean body size to environmental data and plant functional traits [55]. We used the directional-separation-test (d-sep, [56]) to confirm or reject the assumed relationships between variables in the structural equation model. d-sep works by testing individual regressions between all pairs of variables of the model. Direct relationships, i.e. variables connected by arrows, should turn out to be significant. Indirect relationships, i.e. variables connected by several arrows through other variables, should not be significant. However, for indirect relationships, not the variables themselves, but their residuals from regressions on their direct connections, are tested. This approach delivers P-values for the individual relationships between variables and an overall model fit. We applied mixed effect models for regression, because they offer the possibility of including plot-specific and vegetation type-specific random effects [55]. We assumed linear relationships between all parameters within the limited ranges of the variables investigated. Variables were standardized prior to analysis. To obtain correct P-values, we followed the recommendations of Barr et al. [57] and included random slopes and random intercepts. All calculations were done in R [58]. For mixed effect models, the lmer function in R package lme4 was used [59] with Gaussian error structure obtaining maximum likelihood estimates (option REML = F). Separate models were run for each animal group to keep total variable numbers low with regard to the number of observations. For each group, the hypothesized model including all predicted interdependencies was tested first. Then, an improved model omitting the non-significant terms at a P-value of 0.05 was run. Coefficients, P-values and R² values for both initial and "significant" models are reported. Table 1 lists the detailed hypotheses defining the structural equation models.

Environment – plant functional traits

Precipitation and disturbance were linked to increases in leaf economics. This variable aggregated a suite of correlated traits such as specific leaf area, leaf nitrogen and phosphorus content, leaf dry matter content, and stem specific density. It indicates a gradient ranging from relatively slow growth and conservation of resources to fast growth and rapid turnover of resources in leaves termed the leaf economics spectrum [32]. High values of leaf economics occurred in fast-growing plants with highly palatable tissues, while low values were found in slow-growing plants investing in long-lasting structural components. Whether traits indicating the leaf economics spectrum are linked to precipitation on global scales is still equivocal [60]. At Mount Kilimanjaro, precipitation was the strongest driver of total plant biomass and leaf economics, whereas temperature had almost no effect, in contrast to findings on a global scale [60]. The strong positive effect of disturbance (logging, grazing) on leaf economics is concordant with other studies worldwide (e.g. [31, 61]).

The positive and negative relationships between leaf economics and the proportion of insect-pollinated plants and bird-dispersed fruit CWM, respectively, were expected due to the functional differences between undisturbed savanna and forests on the one hand and agricultural areas on the other. At Mount Kilimanjaro, grazed and ploughed habitats mostly harbor fast-growing dicotyledonous weeds with rather high leaf economics values. These weeds are pollinated by insects, but produce mostly wind-dispersed seeds, instead of bird-dispersed fruits. Seed dispersal by animals is often strongly affected by disturbance, but pollination is not [44, 62]. Large differences between marginal and conditional R² values were evident in total plant biomass, in the proportion of insect-pollinated plants, and in bird-dispersed fruit CWM. This could be caused by non-linear relationships of predictors and traits or factors not accounted for by the predictors [63].

Environment – animal body size

Bergmann's rule and its recent formulations in the framework of the metabolic theory of ecology [17] predict decreases in body size with increasing temperature. Moths and frugivorous birds were in accord with this expectation, in line with previous studies (e.g. [14, 20, 64]). Conversely, insectivorous bird size increased with temperature, contrary to most other studies on birds [65]. Unlike patterns of body size along elevational gradients in the Alps [66], mean body size of bees at Kilimanjaro did not follow Bergmann's rule. Bees of smallest body size occurred in the alpine Helichrysum shrubland situated at highest elevations, suggesting increased extinction probabilities for large-bodied bee species in low-energy habitats [67]. The only other variable found to increase bee body size was disturbance, probably because most wild bees that visit crops or ruderal plants in disturbed habitats nest in natural habitats [34] and isolated resources in disturbed habitats can better be exploited by large-bodied bees species with larger foraging ranges [35]. However, for social species, which accounted for a small proportion of bee species richness, but for a considerable part of the specimens, Kaspari and Vargo [68] noted that colony size rather than individual body size should follow Bergmann's rule. Nevertheless, this could not be tested, as colony size data were not available.

Plant functional traits – animal body size

The food availability hypothesis [21] and the textural discontinuity hypothesis [10] justified the expectation of a strong influence of plant functional traits and total plant biomass on animal body size. However, we did not find any significant relationship between body size and CWMs of functional traits that indicate resource availability. Assuming the "resource availability hypothesis" were correct, a shortcoming of the approach presented here might be that the traits used were not precise enough indicators of the "true" food resources. We used leaf and stem traits, that indicated a gradient from carbon-rich, hard tissue to soft, nutrient-rich tissue as an indicator for moth food availability [69]. This, however, may not adequately reflect the amount of palatable tissue for moths, as Lepidoptera tend to be specialized to particular host plants [70]. In moths, a correlation between body size distributions and size distributions of floral resources has been observed [71], indicating that nectar quantity may have predictive potential for body size. Still, the CWM of insect-pollinated plants may be a too coarse estimate of the nectar actually available for bees and moths, as flower rewards, inflorescence sizes and nectar contents can vary strongly across plant species [72].

Total plant biomass had a significant effect on the body size of bees. Rainforests were characterized by the highest structural complexity, indicating that larger distances to food sources select for larger bee species [28, 44]. Moths however, did not respond in the same way, possibly because the caterpillar life stage exerts a strong influence on body size and leaves are a rather abundant resource, which do not force caterpillars to forage long distances in both open landscapes and forests.

Insectivorous bird body size was not significantly related to resource availability, probably because the latter was approximated by the cumulative body sizes of all moth and bee individuals per plot. These measures should be good proxies of overall insect abundance in most vegetation types, but probably not at high elevations. Bee abundance in the alpine zone was high, but other insects were rather rare (pers. obs.). For leaf herbivores, this was evidenced by the virtual absence of damaged foliage at high elevations. Including other insect taxa in our study would likely change our estimate of the food availability pattern for insectivorous birds. Thus, a more precise measurement of cumulative insect body size should also include other relevant groups, e.g. Diptera and Coleoptera.

In addition, the use of community-weighted means (CWMs) of body sizes may affect the result: While the unweighted body sizes of species may decrease, changes in abundances may result in increases of the body size CWMs [73]. This is evidenced by the changes in moth and bee cumulative body size with temperature and with total plant biomass, which were opposed to those of the corresponding CWMs. Thus, increases or decreases in average body size can be counterbalanced by changes in overall abundance of large-bodied and small-bodied species. This could be another reason for the positive relationship between insectivorous bird body size and temperature, meaning that at higher elevations, small species may have more individuals than large ones, as compared to lower elevations. In addition, body size CWMs do not account for the dispersion of body sizes in habitats. Dispersion, rather than means, may likely respond to increasing structural complexity of habitats or increasing discontinuity of resource availability across scales [74–75].

Body size distributions probably also depend on other factors not considered in our study [10]. Large body size may be caused by sexual selection or elevated fecundity of large individuals [76]. The interplay of sexual selection for larger individuals and natural selection for smaller ones can complicate patterns [77]. Blanckenhorn [76] proposed selective forces favoring small body size: Larger individuals need more resources, may be preferred by parasites and predators, and may have reduced fecundity if reproducing late. However, quantifying these parameters in a comparative empirical approach across multiple taxa, or guilds along multiple environmental gradients, is difficult.

We used proxies of plant traits to indicate resource availability as a predictor of body size. This was necessary because logistic constraints did not allow for quantifying true nectar availability, or insectivorous bird food supply, among others. The selection of proxies was not arbitrary; higher proportions of insect-pollinated plants should, on average, yield more nectar than lower ones. Higher cumulative body sizes of moths and bees should indicate higher availability of other prey insects eaten by birds. These relationships, while plausible, may not always hold: In certain habitats, chemically or physically well-defended insects could have a larger cumulative body size than the insects from other localities without increasing resource availability for birds. Abundant insect-pollinated plant species with heavy allocation to vegetative compared to reproductive biomass could provide less nectar than rare insect-pollinated species with stronger allocation to reproductive biomass. Eventually, this may lead to a divergence between the proxies and the actual traits measured. While this constraint needs to be considered, the test of hypotheses with proxies may lead to new insights, if the relationships of proxies and true variables hold. This has been seen in the use of "soft" instead of "hard" traits to predict plant photosynthesis and carbon allocation [78], or the common use of precipitation instead of soil water availability to predict vegetation patterns [79].

In summary, major effects of all abiotic predictors on plant functional traits could be confirmed. Temperature affected animal body sizes, albeit in different directions. The more-food hypothesis was, however, hardly supported by our data. If this hypothesis were correct, we should have found some coupling between leaf traits indicating palatability and body sizes of animals, particularly herbivorous guilds. There are several possible explanations. Firstly, the traits describing resource availability may not have been good enough proxies to reveal the true relationships. Secondly, using the community-weighted means (CWMs) of traits and body sizes may have lumped together many species that do not necessarily depend on the same resources. Moth caterpillars are known to be host-specific [70], and the relationship of plants and pollinators is characterized by a large variety of interactions from co-dependent plant-animal species pairs to generalist species [80]. Also, frugivorous and insectivorous birds are likely to be specialized to some degree, making the use of CWMs inappropriate.

The concept behind CWMs is that environmental conditions exert strong selective forces on a trait and filter only a small subset of the total trait expressions of the geographical species pool, leading to convergence of the trait values around a mean describing this subset [81–82]. The large environmental variation along the slopes of Mt. Kilimanjaro, resulting in habitats as diverse as alpine vegetation, rainforests, savanna and fields favored the filtering of plant traits as evidenced by our structural equation model. CWMs have become increasingly popular in plant functional trait studies, mainly because trait means per plot can be directly linked to environmental variables characterizing the plot, thus making statistical analyses relatively easy. Plant-animal interactions, however, seem to be too species-specific to allow the use of community-weighted means, both in plants and animals. For future work, it appears more promising to study pairwise matching of plant and animal traits in ecological networks [83–85] and to correlate these matches with environmental conditions.