Study sites were located on the eastern slopes of the Andes near the city of Chulumani, Bolivia (16° 24' 29'' S, 67° 31' 15'' W; 1.750 m a.s.l.). The mean annual temperature is 20.8 °C [26]. Mean annual precipitation is 1.366 mm and can be restricted to less than 50 mm per month during the dry season from May to July [26].

Two remnants of primary forests, Reserva Ecológica Apa-Apa and Bosque de Cala Cala, were investigated in this study (Figure 1). In each forest remnant, four sites were set up at the forest borders. Selection criteria for these sites were: (i) The fire-degraded sites were located adjacent to a forest fragment. (ii) The adjacent parts of these forests covered more than 600 m in diameter. (iii) The homogeneous fire-cleared site covered more than 150 m in diameter. All investigated sites were situated at 1940–2500 m a.s.l.

At each site, we established one spatial block alongside a previously established transect, perpendicular to the forest border. Each spatial block comprised three habitat types: (1) Forest interior, situated 120 m distant from the forest border, (2) Forest edge, situated at the border between forest and fire-degraded habitat, and (3) Fire-degraded area, 120 m distant from the forest border. At a single site, the fire-degraded habitat could not be sampled because it burned down during the time of the study.

In contrast to forest habitats, the fire-degraded and now fern-dominated habitats lack any trees higher than 5 m, but comprise patches of insect-pollinated pioneer shrubs, e.g., Borreria capitata (Ruiz & Pev.) DC. (Rubiaceae), Collaea speciosa (Loisel.) DC. (Fabaceae) or Raulinoreitzia crenulata (Spreng.) R.M. King & H. Rob. (Asteraceae). Abiotic conditions show higher fluctuations during the course of a day in fire-degraded habitats than in forest habitats, with particularly high temperatures and arid conditions during times of sun exposure (D. Lippok and S. Gallegos Ayala, unpublished data).

The study took place from February to May 2011. Bee and butterfly/moth individuals were caught with pan traps [27,28]. Although they may differ in their attraction among different bee species, pan traps are a frequently used, standardized passive sampling technique [27,28], often used in biodiversity assessments [19,29]. In order to attract a wide spectrum of potential pollinators [28,30], commercially available white plastic dishes were additionally painted in yellow and blue. These traps were filled with 300 mL of water and a dash of odorless soluble. Three pan traps, one of each color (white, yellow, blue), were placed in breast height. Thus, our study design compares pollinator diversity in the lower vegetation strata. In every habitat type at each site, three pan trap bundles were installed in a distance of 10 meters from each other.

First, bee species were collected for 24 h at every site. In a second sampling, bee and butterfly/moth species were collected for 48 h at every site. At a single study site, individuals were accidentally collected for five days instead of two days. Assuming a linear relationship between time and number of caught individuals, we applied a rarefaction approach to create a randomly rarefied community sample that was used in further analyses for this particular site. All collected specimen were stored in 60% ethanol and later identified to morphospecies at the Universidad Mayor de San Andrés, La Paz/Bolivia. For each morphospecies, we also measured its body length (i.e., distance from caput to abdomen) as a measure of body size.

We acknowledge that the sampling effort at each site was too low for estimating the true species richness in these diverse tropical pollinator communities. However, the aim of this rapid diversity assessment was to compare relative differences in pollinator diversity among the three habitat types. Because sampling effort and design were highly standardized in all habitat types, potential differences in species richness and community composition among habitat types are very unlikely to arise from sampling bias and are instead indicative of changes in the respective pollinator communities.

All statistical analyses were conducted with R, version 2.14.0 [31] and dedicated packages: vegan [32], nlme [33] and ade4 [34].

For each habitat type and study site, we determined the cumulative species richness and species abundance across all samplings and compiled a community matrix including species abundances, i.e., the cumulative individual number per morphospecies. Species richness and abundance were square-root transformed to meet normality assumptions. Bee and butterfly/moth richness and abundance were compared between different habitat types, using linear mixed-effect models with habitat as predictor variable and site as random factor to account for the spatial block design of sampling. We defined two orthogonal contrasts in order to identify differences in potential pollinator richness and abundance between (i) forest habitats (interior and edge) and fire-degraded habitats, and between (ii) forest interior and forest edge. Significances of both contrasts were tested using t-statistics.

Because shifts in community composition from habitat loss may be masked when only species richness is measured [35], changes in species composition were analyzed with a detrended correspondence analysis (DCA), based on the quantitative community matrix. The influence of rare species was down-weighted and a single outlying community from a fire-degraded habitat had to be excluded from the ordination analysis. We tested whether species compositions correlated with (i) habitat identity and (ii) altitude of the study sites. We also calculated the proportions of bee and butterfly/moth species in pollinator communities for each habitat type and study site and related their proportions to the gradient in species turn-over. Significances of predictor variables on community composition were tested with a permutation test (1,000 permutations). Effects of habitat type were tested with a stratified randomization approach within sites.

In order to test for differences in body sizes of bee and butterfly/moth guilds among habitats, we applied fourth-corner analyses to detect significant associations between species traits and habitat characteristics [36,37]. For both potential pollinator guilds, we compared body sizes with two approaches. First, we compared body sizes within each pollinator guild; bee body sizes had to be log-transformed. Second, we assigned species to be 'large' or 'small', separated by the median in body size in each guild (median for bees: 6.5 mm; median for butterflies/moths: 14.5 mm). By contrasting small and large size classes, we were able to test whether changes in body sizes were driven by responses of the small- or the large-bodied pollinators. In order to test for significances in correlations between species traits and environment, Legendre et al. [36] proposed four permutation models. According to our study questions, we applied model 1 and 2. In model 1, values for each species are permuted independently, assuming that individual species respond to the habitat gradient. In model 2, entire rows are permuted, assuming that fixed assemblages of species prefer environmentally distinct habitats. We conducted permutation tests using model 1 and model 2 with 999 permutations. Since p-values from both permutation models were qualitatively identical, we only report p-values from permutation model 1.

Species richness of bees and butterfly/moth communities changed along the human-induced habitat gradient. Notably, both pollinator guilds were differently affected by altered habitat conditions. The high structural diversity at the contact zone between fire-degraded and forest habitats did not lead to an increase in species richness, neither in bee nor in butterfly/moth communities. Bee species richness was highest in fire-degraded habitats, whereas butterfly/moth richness significantly decreased in fire-degraded, now fern-dominated habitats. For bee species, Oteri and Sandino [38] also reported a higher abundance of bees in open farmland habitats than in tropical forests but, different from our study, forests still accounted for highest species richness. Contrarily, Liow et al. [39] described lower species richness but higher abundance in primary forests when compared to disturbed forest ecosystems. Diversity and abundance of Euglossine bees were reported to be unaffected by habitat alterations [40]. Thus, responses of bee pollinators to human-induced habitat disturbance appear to be context-dependent and may strongly depend on the type of habitat modification [29] and the surrounding landscape matrix [29].

In our study, we collected 32 individuals of the invasive honeybee [41], Apis mellifera L., in fire-degraded habitats and at forest edges. Honeybees contributed strongly to the increase in bee species richness in disturbed habitats and were strictly bound to disturbed habitats, a phenomenon also found in other studies [42,43]. Further research on the question whether honeybees indeed provide pollination functions and connect forest edges with the surrounding habitat matrix or whether they compete with native pollinators for resources [44] are crucial for assessing the potential value of honeybees for forest restoration.

The observed decline in butterfly/moth species diversity in fire-degraded habitats is consistent with previous studies [13,45,46]. Apart from climate and topography, human disturbance was reported to be one of the most important determinants of butterfly diversity [40]. In most studies, species diversity of Lepidoptera was negatively associated with the intensity of human disturbance [13,45,46], leading to reductions in butterfly richness and abundance of up to 80% in a temperate system [47]. On a larger scale, however, small-scale forest disturbances could even increase Lepidoptera richness [48].

We acknowledge that not only spatial, but also temporal variations in biotic and abiotic conditions influence species diversity [48,49]. Repeated samplings in different seasons could give a more comprehensive insight into differences in species diversity between forest and fire-degraded habitats in our study system. For instance, tropical Hymenoptera exhibited higher temporal species turnover in forest habitats compared to disturbed agricultural ecosystems [50].

At a small spatial scale of 240 m, our study revealed a complete species turnover between forest understorey and fire-degraded habitats. This demonstrates that bee and butterfly/moth species, albeit long distance foragers [51], exhibit strong habitat dependences. In agreement with this view, Powell and Powell [52] reported that male Euglossine bees, known as mobile pollinators [53], failed to cross 100 m wide cleared areas. This was likely because altered microclimatic conditions constrained species movements and local species distributions [52]. For moth species, temperature is assumed to be the most important abiotic determinant of species turnover in montane rain forests [54]. Because our study was conducted at the end of the wet season, lower frequencies of rainfall during the dry season likely result in stronger differences in humidity and temperature between forest and fire-degraded habitats in the dry than in the wet season. The strong differences in microclimatic conditions between forest habitats and adjacent fire-degraded habitats are a likely explanation for the observed species turn-over in our study system. Opposite responses of bee and Lepidoptera communities to altered microclimatic conditions are thus a plausible explanation for their divergent diversity patterns.

Another factor that could affect differences in pollinator communities between habitats is flower availability [38]. In our study system, flower abundances were, however, similarly high in all habitat types (mean number of flowers per 100 m²: 1,100 flowers; habitat differences: F = 0.41, p = 0.671) [55]. In addition to resource quantities, future studies should test whether changes in resource qualities, i.e., floral traits, from forest interior to degraded habitats are associated with the species turn-over in pollinator communities, particularly with the observed opposite trends in bee and Lepidoptera communities.

We could show that even on a small spatial scale bees and butterfly/moth communities varied in species' body sizes. As bee body size determines the thoracic temperature necessary for initiating flight [20], large-bodied bee species start foraging earlier during the day and in cooler conditions than small-bodied bee species [21]. Small-bodied species, on the other hand, can tolerate higher temperatures without overheating [21]. As temperatures were found to rise twice as high in fire-degraded habitats than in forest interior habitats [D. Lippok and S. Gallegos Ayala, unpublished data], we expected to collect smaller bee species in open fire-degraded habitats. Because we actually found more large-bodied species in fire-degraded habitats than in forest interiors, we suppose that other factors drive the increase in bee body size with human disturbance in the study area. For instance, colonization ability was reported as positively correlated with bee body size [56], possibly because body size is linked to bee foraging ranges [57]. Therefore, large-bodied species could be less affected by disturbance and habitat loss than small-bodied species [58].

In contrast to bees, body sizes of butterfly/moth species decreased in degraded habitats. This could be related to the fact that body temperatures of large butterfly species are close to ambient temperatures [58] and thus need to heat up less in closed forest habitats. Schulze et al. [22] also pointed out that the forest understorey offers higher structural diversity and thus better protection against predators than open habitats. Because large-bodied butterflies move more slowly and are therefore easier to catch than small-bodied species [59], it is plausible that large-bodied butterflies are mostly found in the forest interior where they are better protected against predators.

We highlight that different pollinator guilds responded differently to the human-induced habitat gradient in our study area and that physiological differences between the studied pollinator guilds may translate into opposing trends in pollinator traits. It might be an interesting field of future research to test whether different selection pressures on different pollinator guilds result in diverging community structures also at larger spatial scales.