Survey data
To assess impacts of land-use change on bird diversity, we began by collating surveys from the PREDICTS database, a repository of species occurrence and abundance data sampled across multiple land-use types70. We removed 30 datasets because they lacked abundance data (n = 12) or were incomplete (n = 18), usually because sampling was limited to particular guilds or methods, such as camera traps (Supplementary Information). This process produced a baseline of 72 datasets that we augmented by conducting a systematic literature review using Web of Science to identify further published bird surveys that targeted land-use gradients (Supplementary Information). After contacting authors for data, we received 29 suitable datasets that we added to the PREDICTS database. Our sample in this study contains data from these 29 surveys, along with 5 additional datasets released in the latest version of PREDICTS71. Finally, to improve sampling in megadiverse regions, we integrated further independent datasets generated by intensive surveys in Bornean72 and Amazonian rainforests73. Geographical location and sources for all published surveys used in our analyses are presented in Fig. 2, Supplementary Table 1 and Supplementary Data 1.
Most survey data in the PREDICTS database are organized as a hierarchy. Survey sites are nested in study blocks, and blocks are nested in study landscapes. Study blocks are spatially segregated but not always temporally defined70. We found 16 datasets in PREDICTS that contain surveys sampled in different years or seasons; therefore, we subdivided the data into separate study blocks partitioned by location and time of survey. To ensure consistency, we also collapsed 17 surveys in PREDICTS into 7 studies by combining all data extracted from the same original source publications. We then partitioned these seven studies into distinct study blocks representing geographical and temporal subsets. After restructuring, our final dataset consisted of 98 study landscapes (Fig. 2), each with numerous survey sites clustered into study blocks. Subdividing our data in this way enabled us to account for spatial, annual and seasonal effects across studies by including study block as a random effect in our models (Supplementary Information).
We converted survey data into species assemblages to enable comparisons across land-use types. In some study landscapes, species assemblages reflect the total number of species identified in a study site, usually pooled across a series of transects or point-counts conducted at intervals between dawn and midday. Other published studies focused at finer resolution, sometimes defining each point-count as a separate survey site. Single point-counts generally undersample species richness, and neighbouring sites may be very close together, which caused problems for our analyses. In these cases, to facilitate calculation of functional metrics and to minimize the risk of pseudo-replication, we grouped species into larger assemblages by aggregating survey sites with similar land-use types in the same study block (Supplementary Information). Species assemblages were therefore defined as all species encountered in a restricted, spatially and temporally segregated area, largely confined to the same land-use type. We do not use the term community because we do not have direct data confirming species interactions74.
Land-use classification
To classify land-use types for each survey site, we used PREDICTS data to estimate the predominant type and stage of vegetation and the intensity of human use. First, we assigned sites to one of six vegetation classes: primary vegetation, secondary vegetation, plantation forests, pasture, cropland and urban. We then classified lightly or intensively used primary vegetation sites as disturbed primary vegetation, whereas minimal-use primary vegetation sites were classified as pristine primary vegetation (a proxy for undisturbed natural vegetation). Based on previous analyses showing reduced avian FD in intensely urbanized areas55, we also split minimal-use urban sites from sites with more intensive urbanization. Finally, to account for the effects of vegetation structure at different successional stages75, we partitioned secondary vegetation according to age class (mature, intermediate, young, indeterminate; Supplementary Information). Indeterminate age secondary vegetation was removed from our dataset.
Our final dataset consisted of 98 study landscapes distributed across 6 continents (Fig. 2a and Supplementary Table 1), representing a total of 1,281 avian assemblages in 10 distinct land-use types: pristine primary vegetation (n = 177); disturbed primary vegetation (n = 281); mature secondary vegetation (n = 44); intermediate age secondary vegetation (n = 77); young secondary vegetation (n = 86); plantation forest (n = 218), pasture (n = 184); cropland (n = 107); and urban, including both minimal-use (n = 46) and intense-use (n = 61) urban landscapes.
Functional trait data
Species traits can provide information about sensitivity to perturbations (response traits) and the impacts of species presence or absence on ecological function (effect traits). In both cases, data availability is often patchy for major taxonomic groups at a global scale76. We obtained morphometric measurements for all 3,696 species reported in our study landscapes from the AVONET trait database44. Species means were compiled for seven traits: beak length (culmen), beak length (tip-to-nares distance), beak depth, beak width, tail length, tarsus length and wing length (Supplementary Data 1). These traits have been shown to predict a range of key ecological niche axes, including diet and foraging strategy16 (Supplementary Table 3). We also included data on the hand–wing index (HWI), a metric of wing elongation that predicts aerial lifestyle and dispersal distance in birds77. HWI is widely used as a proxy for dispersal ability78. Species mean values for all traits used in our study were calculated from an average of 11 individuals per species (41,515 individual birds measured in total).
Avian morphological traits are often intercorrelated because of an underlying association with body size16. Accordingly, all traits in our dataset were strongly correlated with the body size axis (Extended Data Fig. 2a), apart from HWI (R = 0.22). Following previous studies40,79, we removed the association with body size through a two-step PCA, which reduced our seven linear morphometric traits into three niche axes related to ecological functions (Extended Data Fig. 2b). We performed two separate PCAs on trophic traits (related to beak morphology) and locomotory traits using all species in our dataset. In both cases, the first principal component (PC) was strongly correlated with body size; therefore, we used the second PC to represent the dominant axis of variation, which is effectively independent from body size (Extended Data Fig. 2c). We then performed a third PCA on the first PC scores from both the trophic and locomotory PCAs, taking the resultant first PC to represent the body size axis.
... continue reading