Tim Newbold, Drew W. Purves are affiliated to 'Computational Science Laboratory, Microsoft Research Cambridge'. There are no patents, products in development or marketed products derived from the work presented in this paper. This does not alter the authors' adherence to all the PLoS ONE policies on sharing data and materials. All the other authors have declared they have no competing interests.
Conceived and designed the experiments: TN JPWS DWP. Performed the experiments: TN. Analyzed the data: TN. Contributed reagents/materials/analysis tools: SHMB CHS. Wrote the paper: TN JPWS SHMB CHS DWP.
Efforts to quantify the composition of biological communities increasingly focus on functional traits. The composition of communities in terms of traits can be summarized in several ways. Ecologists are beginning to map the geographic distribution of trait-based metrics from various sources of data, but the maps have not been tested against independent data. Using data for birds of the Western Hemisphere, we test for the first time the most commonly used method for mapping community trait composition – overlaying range maps, which assumes that the local abundance of a given species is unrelated to the traits in question – and three new methods that as well as the range maps include varying degrees of information about interspecific and geographic variation in abundance. For each method, and for four traits (body mass, generation length, migratory behaviour, diet) we calculated community-weighted mean of trait values, functional richness and functional divergence. The maps based on species ranges and limited abundance data were compared with independent data on community species composition from the American Christmas Bird Count (CBC) scheme coupled with data on traits. The correspondence with observed community composition at the CBC sites was mostly positive (62/73 correlations) but varied widely depending on the metric of community composition and method used (R2: 5.6×10−7 to 0.82, with a median of 0.12). Importantly, the commonly-used range-overlap method resulted in the best fit (21/22 correlations positive; R2: 0.004 to 0.8, with a median of 0.33). Given the paucity of data on the local abundance of species, overlaying range maps appears to be the best available method for estimating patterns of community composition, but the poor fit for some metrics suggests that local abundance data are urgently needed to allow more accurate estimates of the composition of communities.
Efforts to describe the composition of communities have focused largely on the occurrence or abundance of species (e.g.
A trait-based approach to community and ecosystem ecology is appealing for a number of reasons. First, the way that species respond to environmental changes is often related to certain traits
In response to this interest in traits, ecologists are beginning to generate maps of the trait composition of communities across large geographical areas
There are a number of challenges to mapping the trait composition of communities. One of the major challenges is accounting for the differing relative abundances of species in any one place. Previous attempts to describe biological communities in terms of the traits represented have often used species rather than individuals as the unit of analysis, thus ignoring differences in abundance among species (e.g.
Here, we investigate whether the accuracy and precision of maps of the trait composition of communities can be improved by incorporating estimates of local abundance interpolated from readily available abundance observations and estimates. We test for the first time the accuracy of maps of trait composition against observations. We focus on birds in the Western Hemisphere because good data on the abundance, distribution and traits of these species are available. Trait composition of communities was calculated from observed abundance of species from sites in the American Christmas Bird Count (CBC) scheme. For metrics of the trait composition of communities that include differences in the abundance of species, estimates made without abundance data will only be accurate if species with particular combinations of traits are not systematically more or less common than other species. We tested whether this was the case.
We inferred the non-breeding distributions of 4064 bird species that occur in the terrestrial Western Hemisphere (95% of the species known to occur in the region
We focused on four traits of bird species: body mass, generation length (the average age of breeding individuals), migratory behaviour and diet. These are likely to be functionally important: for example, bird diet is related to ecosystem functions such as pollination and seed dispersal
Data on mass, generation length and migratory behaviour were taken from BirdLife International's World Bird Database. The data on body mass therein were compiled from various sources, primarily
Species were assigned to four migratory classes: non-migrants, nomads, altitudinal migrants, and latitudinal/longitudinal migrants. Nomadic species move in response to resources that are sporadic and unpredictable in distribution and timing, and may congregate, but not predictably in terms of location and timing. Altitudinal migrants regularly or seasonally make cyclical movements to higher or lower elevations with predictable timing and destinations. Latitudinal/longitudinal migrants are species for which a substantial proportion of the global or regional population makes regular or seasonal cyclical movements beyond the breeding range, with predictable timing and destinations. This includes species that may be migratory only in part of their range/population, short-distance migrants and migrants that occasionally respond to unusual conditions in a semi-nomadic way.
Diet data were compiled by one of us (CHS) from the literature, primarily from the
Complete trait data were available for 3960 species, including species for which mass and generation length values were interpolated from values for congeners. 104 species with incomplete trait data were excluded from our analyses to ensure consistency in the calculations across all methods of calculation and across all measures of community trait composition. Excluded species represented 33 different bird families and thus are unlikely to constitute a functionally unique set of species.
Observed abundance data for 2715 species were taken from 2466 sites (
Black crosses are sites used in the generation of trait maps (n = 2398), whereas white circles are sites used for evaluating the maps (n = 68). In Behrmann cylindrical equal-area projection.
For one of the mapping methods, we also used estimates of the total global population size of bird species derived from BirdLife International's World Bird Database, available for 1324 of 3960 species. Estimates were taken from a wide variety of sources and were based on published estimates, surveys, censuses, inferences from distribution size, estimated population densities, and expert opinion. For species whose total population estimates consisted of minimum and maximum estimates, the mid-point of these values was taken. Because our study area was confined to the Western Hemisphere, we converted these estimates of global population size to estimates of total population size in the Western Hemisphere. To do this, we used a list of all countries in which each species was known to be resident or to overwinter. For each species, the global population estimate was reduced according to the ratio of the total area of these occupied countries in the Western Hemisphere to the total area of occupied countries in the whole world.
One of our mapping methods used three environmental variables: annual mean temperature and total annual precipitation, from the WorldClim dataset Version 1.4
Four methods were used to map species and trait composition based on the distribution maps and trait data (henceforth ‘distribution-based estimates’;
The maps were generated by combining i) trait data, ii) refined distribution maps and iii) various types of abundance data. We used four estimates of abundance (a–d) for generating the maps, based on three basic assumptions about the abundance of bird species: 1) that all species have an equal abundance (of one) in all grid cells (black text); 2) that species differ in abundance from one another, but with no spatial variation in abundance within species (blue text); and 3) that abundance varies both among species and spatially (red text). The maps were evaluated using iv) trait data and v) local abundance data from the Christmas Bird Count sites. Note (*) that the abundance data from the CBC sites were divided into a set for generating the maps (2398 sites) and a set for evaluating the resulting maps (68 sites). Note also (†) that the same trait data were used for generating and evaluating the maps.
Using each of the four methods, for each trait individually and for all traits together, we generated maps of community-weighted mean trait values, functional richness and functional divergence. For categorical traits (migratory behaviour and diet), community weighted mean trait values were calculated as the proportion of individuals (or species) in each class. Functional richness was calculated as the range of trait values present for continuous traits and as the number of classes present for categorical traits. To calculate functional richness for all traits together, we first reduced the trait data of species using principal coordinates analysis (using the dudi.pco function in the ade4 package
We chose to use the Rao index as a commonly used and easily calculated measure of functional divergence, rather than other available metrics such as those based on a ‘functional dendrogram’
We also generated a map of species richness, to test the ability of overlaid distribution maps to capture this basic measure of the composition of communities.
Values of each of the trait composition metrics were calculated from local observed abundances at the CBC evaluation sites, and compared to the distribution-based estimates of the metrics using linear regression (see also
Ignoring differences in the abundance of species will only be a problem for mapping trait metrics if species' abundances in grid cells are biased with respect to the traits in question. For example if small-bodied species are more abundant on average than large-bodied species, then estimates of the distribution of trait values will be biased if it is assumed that species are equally abundant. On the other hand, if variations in abundance are random with respect to trait values, then estimates of the distribution of trait values should be unbiased. To test for a relationship between trait values and abundance, for each CBC site, we correlated log-transformed species abundance with body mass and generation length, and tested differences in (log-transformed) abundance with respect to migratory behaviour and diet using analysis of variance.
All maps were generated using custom-built C# code developed in Microsoft® Visual Studio 10.0 (code available from TN on request). The precision of the distribution-based estimates was measured using R2 and the accuracy of the estimates by assessing departures from a fitted slope of zero. To test that the results were not influenced by spatial autocorrelation in the values of the metrics among CBC sites, the same analyses were repeated using simultaneous autoregressive (SAR) models. To do this, spatial weights were calculated based on the longitude and latitude coordinates of the evaluation sites using the tri2nb and nb2listw functions in the spdep package
The total local abundance of bird communities, estimated using the three mapping methods that incorporated abundance data, correlated positively with observed total local abundance at the Christmas Bird Count (CBC) sites, although the fit of these relationships varied depending on the method used (R2 values were: 0.34 where local abundance was estimated as a species' estimated total population size divided by range size; 0.52 where local abundance was estimated as the average abundance across CBC sites; and 0.57 where local abundance was modelled against environmental variables to capture spatial variation; see
Trait | Metric | Method | Correlation | Departure from unity slope | |||
R2 | P | Slope | t | P | |||
Mass | CWM | Range maps |
|
|
0.3 | 11.9 | <0.001 |
Total population | 0.17 | <0.001 |
|
|
|
||
Mean of records | 0.19 | <0.001 | 0.33 | 8.04 | <0.001 | ||
GAM models | 0.17 | <0.001 | 0.4 | 5.31 | <0.001 | ||
FRICH | Range maps |
|
|
|
|
|
|
FDIV | Range maps |
|
|
−0.044 | 17.9 | <0.001 | |
Total population | 0.00048 | 0.86 |
|
|
|
||
Mean of records | 0.0013 | 0.77 | −0.028 | 10.8 | <0.001 | ||
GAM models | 0.0029 | 0.66 | −0.064 | 7.26 | <0.001 | ||
Generation length | CWM | Range maps |
|
|
0.42 | 7.42 | <0.001 |
Total population | 0.25 | <0.001 |
|
|
|
||
Mean of records | 0.17 | <0.001 | 0.35 | 6.68 | <0.001 | ||
GAM models | 0.17 | <0.001 | 0.34 | 7.14 | <0.001 | ||
FRICH | Range maps |
|
|
|
|
|
|
FDIV | Range maps† |
|
|
0.22 | 9.19 | <0.001 | |
Total population | 0.041 | 0.099 | 0.29 | 4.18 | <0.001 | ||
Mean of records† | 0.086 | 0.015 | 0.28 | 6.37 | <0.001 | ||
GAM models | 0.14 | 0.0019 |
|
|
|
||
Migratory behaviour | CWM (non-migratory) | Range maps |
|
|
|
|
|
Total population | 0.032 | 0.14 | 0.26 | 4.32 | <0.001 | ||
Mean of records | 0.035 | 0.12 | 0.2 | 6.12 | <0.001 | ||
GAM models† | 0.084 | 0.018 | −0.28 | 11.2 | <0.001 | ||
CWM (nomadic) | Range maps | 0.38 | <0.001 |
|
|
|
|
Total population |
|
|
0.42 | 15.9 | <0.001 | ||
Mean of records | 0.58 | <0.001 | 0.43 | 12.4 | <0.001 | ||
GAM models | 0.6 | <0.001 | 0.52 | 8.96 | <0.001 | ||
CWM (altitudinal migrants) | Range maps | 0.14 | 0.0016 |
|
|
|
|
Total population | 0.017 | 0.29 | 0.022 | 48 | <0.001 | ||
Mean of records |
|
|
0.56 | 3.4 | <0.001 | ||
GAM models | 0.11 | 0.0066 | 0.28 | 7.33 | <0.001 | ||
CWM (full migrants) | Range maps |
|
|
|
|
|
|
Total population | 0.036 | 0.12 | 0.27 | 4.34 | <0.001 | ||
Mean of records | 0.039 | 0.11 | 0.21 | 6.19 | <0.001 | ||
GAM models† | 0.075 | 0.024 | −0.26 | 11.2 | <0.001 | ||
FRICH | Range maps |
|
|
|
|
|
|
FDIV | Range maps |
|
|
|
|
|
|
Total population† | 0.07 | 0.029 | 11.2 | <0.001 | |||
Mean of records | 0.004 | 0.61 | 8.9 | <0.001 | |||
GAM models | 0.025 | 0.2 | 10.6 | <0.001 | |||
Diet | CWM (fruit) | Range maps |
|
|
0.29 | 12.3 | <0.001 |
Total population | 0.16 | <0.001 | 0.23 | 11.7 | <0.001 | ||
Mean of records | 0.089 | 0.014 |
|
|
|
||
GAM models | 0.013 | 0.36 | 0.15 | 5.06 | <0.001 | ||
CWM (nectar) | Range maps | 0.8 | <0.001 |
|
|
|
|
Total population | 0.69 | <0.001 | 0.69 | 5.52 | <0.001 | ||
Mean of records |
|
|
0.79 | 4.46 | <0.001 | ||
GAM models | 0.77 | <0.001 | 0.78 | 4.25 | <0.001 | ||
CWM (other plant material) | Range maps |
|
|
0.25 | 19.5 | <0.001 | |
Total population† | 0.068 | 0.032 | 0.27 | 5.96 | <0.001 | ||
Mean of records | 0.12 | 0.0044 | 0.2 | 11.6 | <0.001 | ||
GAM models | 0.23 | <0.001 |
|
|
|
||
CWM (invertebrates) | Range maps† | 0.3 | <0.001 | 0.23 | 18 | <0.001 | |
Total population | 0.079 | 0.021 | 0.26 | 6.68 | <0.001 | ||
Mean of records | 0.091 | 0.012 | 0.22 | 9.1 | <0.001 | ||
GAM models |
|
|
|
|
|
||
CWM (vertebrates) | Range maps | 0.12 | 0.0034 | 0.17 | 14.6 | <0.001 | |
Total population | 0.011 | 0.39 | 0.1 | 7.84 | <0.001 | ||
Mean of records | 0.45 | <0.001 | 0.44 | 9.49 | <0.001 | ||
GAM models |
|
|
|
|
|
||
CWM (mixed) | Range maps | 0.0036 | 0.63 | 0.024 | 19.6 | <0.001 | |
Total population | 0.0065 | 0.51 | 7.71 | <0.001 | |||
Mean of records | 0.0064 | 0.52 | 0.048 | 12.8 | <0.001 | ||
GAM models† |
|
|
|
|
|
||
FRICH | Range maps |
|
|
|
|
|
|
FDIV | Range maps | 0.011 | 0.4 | 0.032 | 25.9 | <0.001 | |
Total population | 0.00055 | 0.85 | 0.015 | 12.4 | <0.001 | ||
Mean of records | 0.00059 | 0.84 | 0.012 | 15.8 | <0.001 | ||
GAM models |
|
|
|
|
|
||
All traits | FRICH | Range maps |
|
|
|
|
|
FDIV | Range maps |
|
|
|
|
|
|
Total population | 0.0037 | 0.62 | −0.054 | 9.71 | <0.001 | ||
Mean of records | 5.7E-07 | 1 | 0.00052 | 11.8 | <0.001 | ||
GAM models | 0.024 | 0.21 | 0.1 | 10.8 | <0.001 | ||
Species richness | Range maps |
|
|
|
|
0.37 | |
Total abundance | Total population | 0.41 | <0.001 | 0.22 | 24.1 | <0.001 | |
Mean of records | 0.34 | <0.001 | 0.35 | 10.9 | <0.001 | ||
GAM models |
|
|
|
|
|
For each of the four traits considered (mean mass, generation length, migratory behaviour and diet) and for all traits together, we calculated community-weighted mean trait values (CWM), functional richness (FRICH) and functional divergence (FDIV). The overall fit of the relationship between distribution-based estimates and observed community composition at the Christmas Bird Count sites was measured using a correlation test (R2 and P-values reported). Departures from a fitted relationship of y = x (i.e. a slope of 1) were assessed using a t-test (slope, t and P-values reported). For each metric the results of the strongest correlation and the smallest departure from one are shown in bold. †s indicate comparisons that were significant for these non-spatial models but that were non-significant for the spatial autoregressive models (for full spatial model results see Appendix S1).
In terms of precision (R2 values), the correspondence between distribution-based and CBC-based estimates of community trait composition varied widely (R2 from 5.7×10−7 to 0.82, median of 0.12), although most (62/73) relationships were positive (
For each of the two continuous traits considered – body mass and generation length – maps were generated of community-weighted mean trait value (CWM; a, d, f, i), functional richness (FRICH; b, g) and functional divergence (FDIV; c, e, h, j). Observed values were calculated from recorded abundances at 68 Christmas Bird Count (CBC) evaluation sites. Distribution-based estimates of the metrics were generated using four methods, but only results from the best two methods are presented here: 1) overlaying range maps (black symbols); and 2) overlaying range maps with estimates of species abundance that vary among species and within species' ranges (red symbols). Abundance was estimated by modelling recorded abundances with respect to three environmental variables using generalized additive models. Lines represent y = x. Full results for all four methods are presented in
In terms of accuracy (approximation to a slope of one in the relationship between distribution-based and local estimates of the trait metrics), overlaying range maps was best in 7 out of 17 cases (mean slope estimate = 0.40), using total population size of species was best in 3 cases (mean slope estimate = 0.23), taking the average abundance of species at CBC sites was best in 1 case (mean slope estimate = 0.25), and using spatial models of local abundance for each species was best in 6 cases (mean slope estimate = 0.27;
For each of the two categorical traits considered – migratory behaviour and diet – maps were generated of community-weighted mean trait value (CWM; a, d, f, i), functional richness (FRICH; b, g) and functional divergence (FDIV; c, e, h, j). For the categorical traits, community-weighted mean was calculated as the proportion of birds in each of the trait classes. Observed values were calculated from recorded abundances at 68 evaluation Christmas Bird Count (CBC) sites. Distribution-based estimates of the metrics were generated using four methods, with the best two shown here, as in Fig. 3. Lines represent y = x. Full results for all four methods are presented in
Maps were generated of functional richness (FRICH; a) and functional divergence (FDIV; b, c). Functional divergence was measured using the Rao index. Observed values were calculated from recorded abundances at 68 evaluation Christmas Bird Count (CBC) sites. Distribution-based estimates of the metrics were generated using four methods, with the best two shown here, as in Fig. 3. Lines represent y = x. Full results for all four methods are presented in
a) community-weighted mean value of (log-transformed) body mass; b) functional richness based on all four functional traits (body mass, generation length, migratory behaviour and diet) measured as the volume of a convex hull enclosing all species positions in trait space; and c) functional divergence measured using the Rao index. Colour schemes for the rasters and for the points are the same. Displayed using the Behrmann cylindrical equal-area projection.
The map of species richness, generated by overlaying range maps, corresponded very closely with observed species richness at the CBC sites (R2 = 0.90;
Accounting for spatial autocorrelation by using SAR models did not alter the results substantially: in 8 out of 73 cases, the relationship was significant for the non-spatial models but non-significant for the spatial models (marked with a † in
As expected, among bird species observed at CBC sites, there were more smaller-bodied species than larger-bodied species (Figure S4a, Appendix S3). However, the average local abundance of species was not strongly related to body mass (mean Pearson correlation coefficient across all sites was −0.025±0.003). Similarly, more species had shorter than longer generation length (Figure S4b, Appendix S3), but average recorded local abundance was not related to generation length (mean Pearson correlation coefficient across all sites was −0.016±0.003). Finally, local abundance did not generally differ with migratory behaviour (across all 2378 sites with sufficient data, analyses of variance revealed a significant difference – P<0.05 – for 55 sites, mean F = 1.04) or with diet (across 2383 sites with sufficient data, analyses of variance were significant for 7 sites, mean F = 3.10).
The apparent lack of correlations between the traits of species and their local abundances may reflect a failure to detect a relationship, arising from an effect of traits on the detectability of species and thus on estimated local abundance. This appeared to be the case to some extent: larger-bodied, migratory species with large range sizes were recorded at significantly more sites than small-bodied, non-migratory species with small range sizes. However, the effect of body mass on detectability was weak, with most of the variation in the number of sites at which a species was recorded being explained by range size and migratory behaviour (deviances explained: overall 46.7%; range size 19.7%; mass 3.9%; migratory behaviour 24.0%).
Previous attempts to map the distribution of traits and trait-based diversity metrics have involved overlaying distribution maps and have generally ignored differences in abundance among species
Importantly, for most metrics of community composition, the method of overlaying range maps produced nearly as good or better estimates of community composition, both in terms of accuracy and precision, than did the methods that included abundance, although even for this method the fit between distribution-based estimates and observed composition varied widely depending on the trait and metric considered. The fact that the methods that included abundance estimates failed to improve the estimated trait composition of communities was almost certainly because the methods used to estimate local abundance corresponded at best moderately with the abundance of bird communities observed in the Christmas Bird Counts (R2 values ranged from 0.34 to 0.52). In the following paragraphs, we will discuss the results for each of the community composition metrics separately.
Functional richness measures the range of different trait values present in a community and has been shown in some cases to correlate positively with measures of ecosystem functioning
The trait composition of communities can also be measured as the mean trait value of species in the community, weighted by those species’ relative abundances (community-weighted mean trait values; e.g.
The functional trait composition of communities can also be measured as the extent to which individuals in a community fill trait space (functional divergence). As with community-weighted mean trait values and functional richness, it has been suggested that functional divergence is related to the functioning of ecosystems and the delivery of ecosystem services
Given that in any one location species are known to differ greatly in abundance, it is perhaps surprising that overlaying distribution maps gave the maps of community trait composition that showed the best fit to observed trait composition. However, for the purposes of mapping trait distributions, ignoring differences in abundance will only be a problem if the average local abundance of species is related to the traits under consideration. For example, it has been shown that body mass often correlates with abundance
Finally, a metric that is not trait based, but will continue to be of interest to ecologists, is local species richness. We found that the maps of species richness produced by the distribution overlap method correlated strongly with observed species richness at the evaluation sites (
It is important to note that this study considered only one taxonomic group. Furthermore, the abundance data used covered the northern hemisphere more extensively than the southern hemisphere, and the range maps are likely to be more accurate for North America than for South America. Therefore, caution is needed in interpreting the estimates of community trait composition for the southern part of the study area considered here. Further work is needed to test whether our results apply to regions outside North America and for other taxonomic groups. Overall though, it would seem that, given more numerous and geographically wide-ranging data on the abundance of a greater proportion of species, it might be possible to improve the estimation and mapping of abundance and traits, but such data are lacking at present even for birds, which are among the best studied of the taxonomic groups. However, in the absence of detailed abundance data, overlaying distribution maps appears to be the best method available at present and produced apparently accurate maps for at least some of the metrics of community trait composition considered here.
(DOC)
(DOC)
(DOC)
We thank: Matthew Smith, Derek Tittensor, Greg McInerny and Mark Vanderwel for helpful comments on a draft of this manuscript; Fleming Skov for providing an Excel spreadsheet for calculating growing degree days; Bob Ridgely and Nature Serve for compiling the distribution maps of the birds of the Western Hemisphere; and the Audubon Society and the many volunteers who have collected and compiled the Christmas Bird Count data. CHS is grateful to Sherron Bullens, Debbie Fisher, David Hayes, Beth Karpas and especially Kathleen McMullen for their dedicated help with the world bird ecology database.