Study design and bumblebee survey
Løken (1973) surveyed bumblebees at a large number of sites in Norway during the mid-1900s, and compiled distribution records based on her own and other collections. The records from Løken’s collections have later been digitized, and are avaliable though Artskart (https://artskart.artsdatabanken.no/). It is not possible to identify the exact sampling effort from Løken’s publication. Instead, we extracted the dates that bumblebee specimens had been collected, from the digitized records in Artskart. This gives us the minimum number of days that a certain site was visited (Supplementary material, Appendix S1). In a few sites (Øyer high, Øyer low, Lom low) that had been visited extensively as a part of a study focusing on flower visitation patterns rather than systematically surveying for bumblebee species in general (Løken 1949), we excluded days when no new bumblebee species had been collected. The majority (10) of sites had been visited in a single year, but eight sites had been visited more than one year (Appendix S1). The median number of days a certain site had been visited was 3.
We selected 18 sites to be re-surveyed in 2012 (Fig. 1). A criterion for selection was the presence of the plant Aconitum septentrionale, which was the subject of a parallel study, and occurs mainly in nutrient-rich, wet meadows and sub-alpine forests with high soil pH. However, since we surveyed bumblebees within a large area with a 1 km radius (see below), this specific criterion is unlikely to have any larger impact on the bumblebee species composition except for the occurrence of Bombus consobrinus which is specialized on A. sepentrionale. We selected sites in order to cover a large latitudinal and longitudinal gradient (Lat.: 60.42–62.62°, ca. 240 km; long.: 7.24–10.61°, ca. 180 km), and to form 9 pairs of sites where each pair consisted of one site at a high altitude—approximately 1000 m.a.s.l. (min = 730 m, max = 1000 m) and one site at a low altitude—approximately 500 m.a.s.l. (min = 440 m, max = 765 m) in each pair (Fig. 1). We also made sure that each site contained a variety of flower-rich habitats such as different types of grassland, road verges, semi-open forest etc. The median distance between two sites in a pair was 12.3 km (min = 2.6 km, max = 44.4 km), and the median altitudinal difference was 352 m (min = 235 m, max = 470 m). The high-altitude sites were located within the upper part of the subalpine forest zone, close to the forest line which is around 1000 m.a.s.l. in this region (Moen 1999). We identified the locations and checked the suitability of sites to be re-surveyed in the field in July 2011.
In some cases, it was difficult to identify the exact location of the site for the historical records, which typically consisted of the name of a farm or village, rather than any exact coordinates. The identification of sites was however facilitated by the fact that the altitude was often stated. To account for this uncertainty, we took a conservative approach, and surveyed bumblebees in multiple flower-rich habitats within a circular area with a radius of 1 km around the most likely location for the historical records. This means that we are less likely to under-estimate current bumblebee species richness than the historical surveys. Prior to the field visits, potentially flower-rich habitats were identified from aerial photos and by traveling along roads within the 1 km circle by car. Each site was then visited twice between July 6 and July 27 2012. In the field, two experienced field entomologists visited as many different flower-rich habitats within the 1 km circle as possible during 2 h (excluding time for transport between habitats) and recorded the abundance of all bumblebee species encountered. The bumblebee surveys were carried out by experienced entomologists with expertise not only in bumblebee identification but also with large field experience of their ecology. This ensured that the search efforts were as efficient as possible, as these persons were trained to identify suitable habitats of different bumblebee species. In addition, flower-visiting bumblebees on Aconitum sepentrionale was observed during 45 min in each site, during the same day as the surveys. The data from the flower visitation and the 2-h surveys were pooled for the analyses. Most species could be identified in the field, but in a few cases bumblebee specimens were collected and identified later in the laboratory. To allow comparison with historical data, we discarded abundances in further analyses and used species occurrences only. Data are available from the Figshare repository: https://doi.org/10.6084/m9.figshare.6833531.v1.
Many of the rarest species of the historical data were not observed in 2012 (see “Results” section and Table 1). This might potentially be caused by a failure to detect those species due to an inadequate sampling protocol relative to their low detection probability. To account for this potential bias, we repeated all analyses (see below) after removal of the seven species that were present in less than three sites in the historical data, therefore focusing only on the common species. Moreover, because simple counts of the number of observed species typically underestimate the true number of species, we aimed to test whether our results were consistent when using a non-parametric estimator of species richness instead of our raw count of species richness. Therefore, we repeated all our analyses of species richness replacing, in each site, the number of species encountered in the present survey by the Chao 1 abundance-based estimator of species richness (Chao 1987) computed with the “vegan” R package (Oksanen et al. 2015). Note that we could not derive richness estimators from the historical data because they require either abundance data or multiple surveys. In addition, we also computed accumulation curves of total species richness for historical and present survey, using the “iNEXT” R package (Hsieh et al. 2016), in order to ensure that our survey protocol captured the entire species richness in the region.
Climate and land-cover change
We obtained total daily precipitation and mean daily temperature from the Norwegian Meteorological Institute as interpolated raster grids with a 1 km × 1 km spatial resolution, based on the national network of weather stations (www.met.no, data available at: ftp://ftp.met.no/projects/klimagrid). Daily values from 1960 to 2012 were extracted at each survey site. As measures of climate change, we used the slopes of the linear trends of the mean annual temperature and the mean annual precipitation during this period. Temperature increased in all sites [mean slope = 0.026 °C/year (+ 1.46 °C over the entire period), min. = 0.003 °C/year (+ 0.17 °C), max. = 0.040 °C/year (+ 2.24 °C)], while precipitation increased in all sites except one [mean slope = 0.225 mm/year (+ 12.6 mm), min. = − 0.0007 mm/year (− 0.04 mm), max. = 0.560 mm/year (+ 31.4 mm)].
To assess the direction and intensity of land-cover change that occurred during the ca. 50 years separating the two surveys, we used aerial photographs from the Norwegian Mapping Authority taken in 1958—1968 and 2008–2013. Using ArcGIS 10.3 (® ESRI), we defined a buffer of 1 km radius around each survey site and extracted for each period the proportion of land in the buffer covered by: arable land, forest, grassland, wetland, built-up area, clear-cut forest or open (heath or scrub) area above the tree-line. There was high variability in land-cover among sites (Supplementary material, Appendix S2-A). Some were predominantly covered by forest (e.g.: Oppdal, high altitude, in the 1960s: 75% forest), others by grassland (e.g.: Geilo, high altitude, in the 1960s: 95% grassland) or had a more even distribution of land-cover types (e.g.: Lom, low altitude, nowadays: 40% arable land, 21% forest, 31% grassland). We calculated an index of dissimilarity, based on the Euclidian distance, between the historical and present land-cover. This index could thus range from 0 for sites whose land-cover did not change to 1 for sites whose land-cover has been entirely replaced (mean land-cover change across all surveyed sites: 0.23, min. = 0.023, max. = 0.55). This index therefore represented the magnitude of land-cover change between the two periods irrespective of the direction of this change. Additionally, we computed a principal component analysis (PCA) based on the proportion of each land-cover category (Appendix S2-B). For each site, we quantified the direction of land-cover change by calculating the difference along the two first PCA axes between the historical and present periods. These two principal components account for 48.8% of the total variation in land-cover.
From the survey data, we calculated for each site and for the historical and present surveys the species richness (the number of bumblebee species identified) and the community temperature index (CTI). CTI is a measure of the relative proportion of cold- and warm-adapted species in a community and has been commonly used to describe the species turnover that occurs as a result of climate change (Devictor et al. 2012). CTI was calculated as the mean species temperature index (STI) of all co-occurring species (Devictor et al. 2008). STI represents the average temperature experienced by a species within its geographical range. STI for all bumblebee species was extracted from Rasmont et al. (2015), where temperature data were based on 1970-2000 climate. We compared the estimated annual rate of CTI change—assuming that our historical data represent community composition in 1960—to the observed temperature trend to assess whether communities track or lag behind the actual climate change. In addition, for each species, we averaged the altitude, latitude and longitude of the sites where the species has been encountered in the historical and present surveys. This way, we aimed to describe shifts in species’ distributions that may have occurred over the 50-year period. We tested whether these mean values differed significantly between surveys using Welch’s t tests. Note, however, that since we recorded the same limited number of sites in both surveys, this analysis is unable to reveal broad patterns of range shifts and is only relevant to describe the distribution of species among our 18 surveyed sites.
In a first step, we described how species richness and CTI varied geographically and how they changed between the historical and present surveys. We used linear mixed models with either species richness or CTI as response variables, and the period of survey (historical or present), the altitude (low or high), the longitude and latitude of sites and the interaction between altitude and the period of survey as explanatory variables. Since surveyed sites were grouped by pairs of high and low altitude sites, we included the identity of site pairs as a random effect. Mixed models were computed in R using the “lme4″ package (Bates et al. 2015), with the significance of fixed effects assessed using the “lmerTest” package (Kuznetsova et al. 2017).
In a second step, we aimed to test whether the physical characteristics of the sites (altitude, magnitude of land-cover and climate change) could explain the magnitude and direction of the change in species richness and CTI between historical and present surveys. For this purpose, we computed linear models with, for each surveyed site, the net change between surveys in species richness (log-transformed) or in CTI as response variable. Here, we log-transformed species richness to standardise its temporal change whatever the level of species richness observed in the site. Explanatory variables were the magnitude of land-cover change (Euclidian distance), land-cover change along the first and second principal components, temperature trend, precipitation trend, and the altitude (low or high). Due to the relatively low number of sites, we could not test the interaction between land-cover change, climate change and altitude. In addition, in order to control for spatial autocorrelation between sites, we included a spatial correlation structure as a function of longitude and latitude. We tested several types of correlation structures and selected a Gaussian correlation structure because it yielded the lowest AICc. Models with correlation structures were fitted using generalized least squares with the “nlme” R package (Pinheiro et al. 2018). We additionally performed a post hoc analysis with the “lsmeans” R package (Lenth 2018) to test if the changes in species richness at low and high altitude were significantly different from 0.