Disentangling the relative roles of climate and land cover change in driving the long-term population trends of European migratory birds

Aim: Global declines in the populations of migratory species have been attributed largely to climate change and anthropogenic habitat change. However, the relative contribution of these factors on species’ breeding and non-breeding ranges is un-clear. Here, we present the first large-scale assessment of the relative importance of climatic conditions and land cover on both the breeding and non-breeding grounds in driving the long-term population trends of migratory species. Location: Europe and Africa. Methods: We use data on


| INTRODUC TI ON
Increasingly, climate change is implicated in an emerging global trend whereby migratory species are experiencing more rapid rates of population decline than their resident counterparts (Runge et al., 2015;Studds et al., 2017;Wilcove & Wikelski, 2008). With large numbers of the world's vertebrates making long-distance movements (Somveille, Rodrigues, & Manica, 2018), the decline of migrant species poses a serious threat to global biodiversity and associated ecosystem services (Bairlein, 2016;Vickery et al., 2014).
Understanding the causes of these declines is complicated by the dependence of migrants on multiple habitats, including on their breeding and non-breeding grounds, as well as at stopover sites (Robinson et al., 2009;Zurell, Graham, Gallien, Thuiller, & Zimmermann, 2018).
The dependence of migrants on conditions in multiple areas, and on the phenology of events in these areas, renders them more vulnerable to environmental changes than their resident counterparts (Finch, Pearce-Higgins, Leech, & Evans, 2014;Runge, Martini, Possingham, Willis, & Fuller, 2014). Elucidating drivers of their population change is further complicated by the need to understand in which of the various parts of the annual life cycle the population is critically limited (Ockendon, Hewson, Johnston, & Atkinson, 2012).
Afro-Palaearctic long-distance migratory birds, which breed across Europe and spend the non-breeding season in sub-Saharan Africa, are declining significantly faster than European breeding-resident and short-distance migrant birds (Cresswell, 2014;Gregory, Skorpilova, Vorisek, & Butler, 2019;Sanderson, Donald, Pain, Burfield, & van Bommel, 2006;Vickery et al., 2014). Many of these long-distance migrants have shown continent-wide patterns of sustained, often severe, decline since circa 1970. By contrast, many of their resident counterparts have undergone little or no decline during the same period Vickery et al., 2014). Detailed monitoring in western Europe has provided a clear understanding of how changes in climate affect the populations of European breeding birds (Gregory et al., 2007;Thaxter, Joys, Gregory, Baillie, & Noble, 2010). Moreover, long-term trends in population sizes of birds in Europe have been related to climate trends on their breeding grounds (Stephens et al., 2016). Population declines in migratory species may be a consequence of asynchronous responses to changes in climate, resulting in phenological mismatch between, for example, the timing of breeding of insectivorous birds and their prey availability (Both et al., 2010;Møller, Rubolini, & Lehikoinen, 2008). Additionally, warmer winters may have improved the overwinter survival rates of resident species, increasing competition for breeding resources, to the detriment of those migratory species that arrive at the breeding grounds last (Berthold, Fiedler, Schlenker, & Querner, 1998;Pearce-Higgins, Eglington, Martay, & Chamberlain, 2015). Climatic conditions on species' non-breeding grounds have also been linked to long-term trends in the populations of migrant birds. In particular, rainfall across the Sahel region of Africa has been shown to affect overwinter survival (Peach, Baillie, & Underhill, 1991) and to influence the conditions of European breeding birds when they depart on spring migration, with potential carry-over effects on reproductive success (Ockendon, Leech, & Pearce-Higgins, 2013;Saino et al., 2011;Zwarts, Bijlsma, & van der kamp, & Wymenga, 2009). Land use change is also recognized as an important driver of population trends for some bird species (Sanderson et al., , 2016Sullivan, Newson, & Pearce-Higgins, 2015). For example, across Europe, birds associated with farmland habitats have declined steeply since the 1950s, which has been largely attributed to the intensification of agricultural practices Gregory et al., 2019). However, little is currently known about the role of land use changes at non-breeding grounds on population trends (Beresford et al., 2019;Cresswell, Wilson, Vickery, Jones, & Holt, 2007;Wilson & Cresswell, 2006). It has been suggested that between 1970 and 1990, Afro-Palaearctic migrants dependent on dry, open habitats on their non-breeding grounds declined significantly more than other migratory birds (Atkinson et al., 2014;Sanderson et al., 2006). This was linked to an extended drought across the Sahel during this period, which probably impacted the shallow-rooted vegetation that dominates these open habitats (Atkinson et al., 2014).
The similarities in population trends among species with shared habitat affinities imply that changes to land cover can drive population declines. However, the relative effects on the populations of European long-distance migrants of long-term climatic and land cover changes on breeding and non-breeding grounds remain uncertain (Ockendon et al., 2012;Vickery et al., 2014).
Here, we explore the potential contribution of long-term trends in (1) climate suitability and (2) land cover, on both breeding and non-breeding grounds, with the population trends of European breeding birds with different migratory strategies and habitat affinities. First, we associate the breeding and non-breeding distributions of these birds, at a continental scale, with long-term mean climate data. Based on those associations, we identify temporal trends in climate suitability for individual species, separately across their breeding and non-breeding ranges. We then calculated the trend in the extent of suitable land cover for each species on their breeding and non-breeding ranges. In order to determine the degree to which trends in breeding bird population sizes across Europe are directly attributable to changes in climate and land cover, we analyse the relationship between these climate suitability trends (CSTs), and land cover suitability trends (LCSTs), with long-term avian population K E Y W O R D S Africa, avian migration, breeding, climate, Europe, habitat, non-breeding, population trends trends. Within this analysis, we control for species' habitat affinities because similar trends among species with shared habitat associations may indicate that processes affecting the condition of specific habitats that are not captured within the LCSTs, for example agricultural intensification, are also affecting populations. By analysing these effects separately for long-and short-distance migrants, we evaluate whether the drivers of population changes differ for different types of migrant birds.

| Species data
Annual indices of abundance for 133 species of European migratory breeding birds, as defined by Snow, Perrins, Hillcoat, Gillmor, and Roselaar (1997), were taken from the Pan-European Common Bird Monitoring Scheme (PECBMS, http://www.ebcc.info/pecbm.html) for the period 1980-2016. Of these, we excluded species for which population monitoring commenced post-2000, along with any species for which there had been a period of limited geographical coverage in population monitoring. We further excluded species whose European breeding populations migrate primarily to non-breeding areas outside of our Europe-Africa study area (i.e. Asia). This left 100 migratory species with long-term population trends calculated from breeding ground monitoring across 28 European countries (see Appendix S1 in the Supporting Information, Table S1.1).
Multinational population trends were calculated as the coefficients of log-linear Poisson regressions of the annual index of the size of a species' European population across the 28 countries against calendar year (Gregory et al., 2005). To account for potential sampling error in the estimation of species' population trends in subsequent analyses, we also took the standard error of these regression coefficient values. Additional details on the calculation of population trends are provided in Appendix S1 in the supporting information. In addition, we calculated a multispecies population index separately for both short-and long-distance migrants (see below for definitions). Using the annual indices of population size, we took a geometric mean for each year as the annual population index for all species within each group, giving each species an equal weight in the analysis (Gregory et al., 2005).
Trait data for the 100 species were obtained principally from Gregory et al., (2009), who derived logged mean mass data from Cramp (1977Cramp ( -1994 and migratory strategy from Snow et al. (1997).
Species were classified into two groups according to their migratory behaviour: (1) short-distance migrants, which migrate principally to different parts of Europe and North Africa in the non-breeding season; (2) and long-distance migrants, which spend the non-breeding season entirely in sub-Saharan Africa (see Appendix S1 in the Supporting Information, Table S1.2 for classifications). Each species' principal breeding range habitat was classified as farmland, woodland, wetland or other, following the classification of PECBMS (http://www.ebcc.info/pecbm.html). Non-breeding range habitat associations were taken principally from Atkinson et al. (2014), and for omitted species from Barshep, Erni, Underhill, and Altwegg (2017) and Cramp (1977Cramp ( -1994, and were classified as open country, shrub/woodland, wetland or other (see Table S1.2 for classifications and their source). We also calculated a continuous measure of migratory dispersion for each species following the methods of Gilroy, Gill, Butchart, Jones, and Franco (2016). Migratory dispersion is measured as the relative difference in size between a species' breeding and non-breeding ranges and, previously, has been linked to population declines in migratory species (Koleček, Procházka, Ieronymidou, Burfield, & Reif, 2018). This measure indicates the ability of a species to utilize a wider range of areas during the non-breeding season, which may confer increased population-scale resilience to area-specific threats (Gilroy et al., 2016;Runge et al., 2014).
Range extent data (used to fit species distribution models [SDMs]-see below) for breeding and non-breeding distributions for each species were obtained from BirdLife International and NatureServe (2016) for the 100 species. The distribution maps of each species' breeding range were overlaid with a 0.5° x 0.5° grid covering both Europe and the part of Africa north of 20°N. A species was considered present in a 0.5° grid cell if ≥10% of the cell intersected with the species' breeding distribution. The non-breeding ranges of the 39 long-distance migrants were similarly overlaid with a 0.5° grid and converted to presence-absence data for the part of Africa south of 20°N. For the 61 species of short-distance migrants (the non-breeding ranges for some of which extend into Africa), the non-breeding ranges were gridded for both Europe and all of Africa.

| Climate data
Data for three climatic variables, mean monthly temperature, precipitation and percentage cloud cover from 1951 to 2000 (the period during which the majority of data underlying the species' range extent maps were collected), were obtained from the CRU TS 3.25 0.5° dataset (Harris, Jones, Osborn, & Lister, 2014). These data were used in conjunction with soil water capacity data and formulations, both from Prentice et al., (1992), to calculate five bioclimatic variables, at 0.5° resolution, for all of Europe and Africa. Separate sets of bioclimatic variables were calculated for the breeding and non-breeding seasons. Mean temperature of the warmest month (MTWA), mean temperature of the coldest month (MTCO) and total precipitation (TP) were assessed over months chosen to coincide as closely as possible with the times that the majority of species spend on their breeding (March-September) and non-breeding (August-February) ranges (Ponti, Arcones, Ferrer, & Vieites, 2019). We also included annual measures of seasonality of temperature (ST) and precipitation (SP). This was to account for the role of significant variations in climate that may affect underlying habitat conditions, for example the periodicity of precipitation across the Sahel or seasonal variability across Europe. These variables have been shown to be informative in describing both breeding and non-breeding range extents (Doswald et al., 2009;Thuiller, Araujo, & Lavorel, 2004) and abundance patterns (Green et al., 2008;Gregory et al., 2009;Howard, Stephens, Pearce-Higgins, Gregory, & Willis, 2014 of European breeding birds. We calculated the mean values of these bioclimatic variables for the period 1950-2000, as well as annual values from 1980 to 2016, to match the species' population monitoring period.

| Modelling species' distributions
To model the relationship between the 1950 and 2000 mean bioclimatic variables and species distributions, we used an ensemble modelling framework, combining four widely applied techniques: generalized linear models (GLMs), generalized additive models (GAMs), generalized boosted regression models (GBMs) and random forests (RFs). These methods have all been shown to produce models that perform well when used in an ensemble SDM approach (Bagchi et al., 2013;Elith et al., 2006). Separate SDMs were built for a species' breeding range and non-breeding range using the rangespecific bioclimatic variables. To ensure that model evaluation is robust to potential spatial autocorrelation (SAC), we used a "blocking" method (Bagchi et al., 2013), whereby we split the data into ten sampling blocks based on ecoregions (Olson et al., 2001; http://www. world wildl ife.org/scien ce/data; Bagchi et al., 2013). We fitted the models to nine of the ten sampling blocks and tested model performance on the omitted block. We repeated this ten times for each of the four SDM approaches to produce 40 models for both the breeding and non-breeding ranges for each of the 100 species. Model fit was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC) plot (Brotons, Thuiller, Araujo, & Hirzel, 2004;Manel, Williams, & Ormerod, 2001). Additional details of the four SDM approaches, the methods utilized to account for SAC and our approach to assessing model fit can be found in Appendix S1 in the supporting information.

| Calculating trends in climate suitability
We applied the 40 SDMs for each species (10 block models x 4 modelling techniques) to the annual bioclimatic data from the CRU TS 3.25 0.5° dataset (Harris et al., 2014) to project climate suitability for each species for the years 1980 to 2016, for both breeding and non-breeding ranges. We then calculated an annual median projected climate suitability for each species for both the breeding and non-breeding ranges, using the projections from all 40 SDMs. Given that all SDMs performed well, with AUC > 0.7 for all species (see results), all projections were included in the calculations of median climate suitability. For the breeding range, this meant suitability from the 40 models was calculated across all cells of a species' breeding range within the area covered by PECBMS (the region for which we have population trend data; Table S1.1). For some species, there is evidence of a high level of dispersal during the non-breeding season (Finch, Butler, Franco, & Cresswell, 2017). Therefore, when evaluating climate suitability across the non-breeding ranges, we took the median climate suitability across species' entire European and African non-breeding ranges.
After calculating median annual climate suitability for each species, we used a generalized linear regression model, with a binomial error structure and a logit link, to regress climate suitability against year. This was restricted to the same time period for which population index data were available for each species. The slope from this regression was termed the climate suitability trend (CST: following Stephens et al., 2016), which indicates the overall trend in climate suitability for a species. CST was evaluated separately for the breeding and non-breeding grounds for each species, hereafter referred to as CST b and CST nb , respectively.
As with the population trends, we calculated the mean climate suitability for both short-and long-distance migrants. We first standardized CST values to have a value of 100 in 1980. We then took the geometric mean of all annual climate suitability values for each group of species, again giving each species an equal weight within the analysis. We calculated this metric separately for species' breeding and non-breeding grounds.

| Land cover suitability trend (LCST)
To quantify trends in the amount of suitable land cover within both species' breeding and non-breeding ranges, we first obtained information on species' habitat preferences from BirdLife International.
For each species, we identified the suitability of nine broad habitat categories (shrubland, grassland, cropland, broadleaved forest, needle-leaved forest, wetlands and water bodies, urban, bare areas and sparsely vegetated), for both the breeding and non-breeding grounds (see Appendix S2 in the Supporting Information, Tables S2.1 and S2.2). We then calculated the area of land cover within a species' range classified as suitable. To do this, we obtained global land cover data from the European Space Agency Climate Change Initiative (2017) (ESA CCI: https://www.esala ndcov er-cci. org/?q=node/1). These data are available at a spatial resolution of 300 m and consist of 24 annual maps of land cover from 1992 to 2015, with each map comprising 22 land cover classes. We aggregated these land cover classes to align with the nine broad habitat categories used to classify species habitat preferences (Tables S2.1 and S2.2). Then, for each year, we calculated the proportion of land cover within a species' range classified as suitable. Finally, we used a generalized linear regression model, with a binomial error structure and a logit link, to regress the proportion of suitable land cover against year. The slope from this regression was termed the land cover suitability trend (LCST) and indicates the overall trend in the amount of suitable land cover within a species range. We evaluated LCST separately for both species' breeding and non-breeding grounds (LCST b and LCST nb , respectively). For species, for which population index data were only available after 1992, the calculation of LCST was restricted to the same time period. Although these annual measures of land cover do not fully encompass the period of interest, they are, as far as we are aware, the best available at this scale. To assess the potential impact on our results of evaluating LCSTs over a shorter time period, we repeated all following analyses using trend data restricted to the period between 1992 and 2015. Finally, we calculated mean land cover suitability values for both short-and long-distance migrants. We first standardized LCST values to have a value of 100 in 1992. We then took the geometric mean of all annual land cover suitability values for each group of species each year, giving each species an equal weight within the analysis. We evaluated this metric separately for species' breeding and non-breeding grounds.

| Accounting for variable sampling effort
Given the inevitable discrepancy between recording efforts in many infrequently visited non-breeding regions and the well-monitored breeding grounds, there is the potential for bias between models based on breeding and non-breeding grounds due to more complete occurrence datasets from breeding grounds. To control for this, we developed an additional series of analyses, incorporating a random bootstrapped sampling approach to control for potential discrepancies in recording effort. We used the total number of point occurrence records from the non-breeding range for a given species from the Global Biodiversity Information Facility database (GBIF: GBIF. org, 2018; using all records up to 2016) to guide the size of a subsample of cells to utilize in calculating CSTs and LCSTs for both the breeding and non-breeding ranges. Using the number of point occurrence records from across a species' non-breeding range as our sample size (x), we randomly sampled x cells from each of a species' breeding and non-breeding range. We then took the median climate suitability across the subsampled cells to recalculate CST b and CST nb .
We also recalculated LCST b and LCST nb using the subsampled cells.
We repeated this process 100 times to give 100 bootstrapped estimates of CST b , CST nb , LCST b and LCST nb for each species, which we then used to repeat our assessment of the relationship between species' population trends and both CSTs and LCSTs, described below.

| Assessing correlates of species' population trends
We used Markov Chain Monte Carlo generalized linear mixed models from the "MCMCglmm" R package (Hadfield, 2010) to assess the relationship between species' population trends and potential explanatory variables. This analysis was performed at a pan-European scale because, despite having national indices of abundance during the breeding season, we currently lack data on where different European populations spend the non-breeding season. MCMCglmm takes a Bayesian approach to fitting generalized linear mixed models.
It can account for the non-independence between species that can arise from common ancestry, by including a phylogenetic variancecovariance matrix as a random effect. To account for sampling error in the estimation of species' population trends, we also included the standard error of the trend estimates as an additional random effect.
In addition to the random effects of phylogeny and population trend standard error, fixed effects within the model included terms for each species to control for differences in body size (logged mean mass of a species), migratory strategy (short-or long-distance), migratory dispersion and primary habitat association. As species' primary breeding and non-breeding range habitat associations were significantly associated with each other for both short-(Χ 2 = 157.2, df = 9, p < .01) and long-distance migrants (Χ 2 = 31.6, df = 9, p < .01), we only included species' primary breeding habitat associations as an explanatory variable. We included interactions between CST b , CST nb , LCST b and LCST nb , and migratory strategy, as we were interested in the potential for differing contributions of climate and land cover suitability to the population trends of short-and long-distance migrants. We assumed that the response variable had a Gaussian error distribution and used non-informative priors with an inverse Wishart distribution (V = 1, ν = 0.002). Model outcomes were insensitive to the specification of the non-informative priors. We ran the model for 220,000 iterations, with a burn-in period of 20,000 and a sampling interval of 20.
Approximately 1,000 independent samples were generated for each model. We used Gelman-Rubin statistics and diagnostic plots to check for convergence of model chains and the independence of samples.
Rather than using one phylogenetic tree, and assuming this tree was error-free, we instead randomly selected 100 trees from birdtree.org (Jetz, Thomas, Joy, Hartmann, & Mooers, 2012) and fitted the model to each of these trees. We then combined the posterior outputs of the resulting 100 MCMCglmms to provide estimates of model coefficients that incorporate phylogenetic uncertainty. For each model, we calculated the percentage of variance explained by each of the fixed effects and assessed model performance using marginal R 2 , following the methods described in Nakagawa and Schielzeth (2013). All analyses were performed in R version 3.5.0 (R Core Team, 2019).

| Species distribution models
SDMs fitted to the breeding and non-breeding ranges of all short-(61 species) and long-distance (39 species) migrants produced robust models for all species using all four SDM approaches. Model fit was very good across the breeding ranges (AUC on validation data: median = 0.98, IQR = 0.85 -0.99; see Appendix S1 in the Supporting Information, Table S1.3 for individual model technique results). SDMs for species' non-breeding ranges were also good, though median fit was slightly lower than for the breeding range models (AUC on validation data: median = 0.96, IQR = 0.78 -0.99; Table S1.3). There was no significant difference in SDM performance between short-and longdistance migrants for either the breeding (t test, t 96.5 = 0.07, p = .94) or non-breeding grounds (t test, t 73.60 = 1.30, p = .20). The mean annual projections from these models were used to calculate CST values for both the breeding and non-breeding grounds for all 100 species.

| Predictors of population trends
On average, the populations of short-distance migrants have remained relatively stable since 1980, while those of long-distance migrants have declined steadily over the same period (Figure 1a). There was no difference in average climate suitability between short-and long-distance migrants on either their breeding or non-breeding grounds (Figure 1b and c). However, while climate suitability within species' breeding grounds has declined steadily since 1980, the same cannot be said for species' non-breeding grounds, with climate suitability remaining relatively constant over time (Figure 1b and c). For both short-and long-distance migrants, the extent of suitable land cover across their breeding ranges has increased steadily since 1992 ( Figure 1d). Across species non-breeding ranges, however, while F I G U R E 1 Changes in populations and climatic suitability for 61 species of short-distance (yellow lines) and 39 species of long-distance (blue lines) migratory European breeding birds between 1980 and 2016. Changes in Pan-European population indices are shown in (a), while changes in climate suitability are shown for both breeding grounds (b) and non-breeding grounds (c), and changes in the extent of suitable land cover are shown for both breeding grounds (d) and non-breeding grounds (e). For all panels, solid lines indicate the geometric mean, and shading indicates the standard deviation. For both groups of species in panels a-c, values for 1980 are arbitrarily set to 100. In panels d and e, for both groups of species values for 1992 are arbitrarily set to 100. The horizontal dashed lines at values of 100 show the expectation if there is no trend there have been increases in land cover suitability for short-distance migrants, the suitability of land cover for long-distance migrants has steadily declined (Figure 1e).
Phylogenetically informed linear mixed models explained the population trends of the 100 European breeding migratory birds reasonably well (marginal R 2 = .40, S.D. ±0.02). When considered separately, the model-averaged parameter estimates explained the population trends of long-distance migrants less well than those of short-distance migrant species (R 2 = .15 and R 2 = .37, respectively).
Our models indicated that higher CST b and LCST nb , greater body mass and increased dispersion during the non-breeding season were all positively related to species' population trends (Figure 2). In contrast, an association with farmland and open habitats had a negative effect on species' population trends (Figure 2). There was a significant interaction between CST b and migratory strategy (Figure 2). CST b was positively related to the population trends of short-distance migrants with a coefficient that was significantly greater than zero (β = 0.48, 95% credible intervals = 0.18 -0.79, Figure 3). The relationship between CST b and the population trend of long-distance migrants, however, was non-significant (with a coefficient that overlapped zero; β = −0.03, 95% CI = −0.33 -0.28, Figure 3). None of CST nb , LCST b or the interactions between CST nb , LCST b and LCST nb , with migratory strategy, were identified as significant predictors of species' population trends. The results from the analysis using 100 bootstrap estimates of CST b , CST nb , LCST b and LCST nb were qualitatively similar to those of the main analysis (see Appendix S1, Figure   S1.1), but the model had less explanatory power (marginal R 2 = .29, S.D. ±0.01). The results from the analysis using data restricted to the period between 1992 and 2015 were also qualitatively similar to those in the main analysis (see Appendix S1, Figure S1.2), but again, the model had less explanatory power (marginal R 2 = .31, S.D. ±0.07).

| D ISCUSS I ON
Until now, the relative contribution of changes in climate and land cover on the breeding and non-breeding grounds to the declines of European breeding long-distance migrants has been poorly understood Vickery et al., 2014). Here, we discuss our results in the light of three key findings: 1) the long-term population trends of migratory birds are more closely associated F I G U R E 2 Standardized coefficients (a) and percentage of variance explained (b) from MCMC generalized linear mixed models of the population trends of 100 European migratory birds since 1980 (marginal R 2 = 0.40, S.D. ±0.02). In (a), the centre point indicates the mean, the thick bars indicate the posterior standard deviations, and the thin lines indicate the 95% credible intervals of the coefficient values produced by averaging 100 separate MCMCglmms. To standardize coefficient values, all predictors were z-transformed. The reference level for categorical variables includes "short-distance" for migratory strategy and "other" for both breeding and non-breeding habitat association. Overlap of CIs with the dashed line at zero indicates that the effect of the parameter is not statistically significant (i.e. p > .05). In (b), bars indicate the mean percentage of variance explained that is attributable to each variable, while thin lines indicate the standard deviation. Significant terms are also indicated in bold on the shared y-axis with changes in climate than changes in land cover on their breeding grounds, but the converse is true on their non-breeding grounds; 2) climate trends on the breeding grounds explain more variation in long-term population trends for short-distance migrants than for long-distance migrants; and 3) our ability to explain species' population trends depends on their migratory strategy but nonetheless still only explains about 40% of observed variation in population trends.
We discuss these three key findings in turn below.

| Importance of environmental conditions on the breeding and non-breeding grounds
This is the first large-scale assessment of the relative importance of climate and habitat, as measured by land cover on both the breeding and non-breeding grounds for populations of migratory birds.
We have demonstrated that the variables of greatest importance for driving the long-term population trends of migratory species differ between their breeding and non-breeding grounds. Our results highlight the importance of climatic conditions on species' breeding grounds, with a positive relationship between population trend and breeding ground CST, even after accounting for phylogeny, migratory behaviour, body mass and habitat. These results corroborate previous research demonstrating that interspecific variation in the recent population trends of European birds is correlated with climatic trends on species' breeding grounds (Green et al., 2008;Stephens et al., 2016). By contrast, population trends on the breeding grounds were not related to land cover change, at least at the scale as which we assessed this (see below). This is despite the fact that since the 1950s, there have been large-scale changes in both the extent and management intensity of land use across Europe. Particularly notable have been the increases in the extent of forested and urban areas alongside intensification of agricultural practices (Kuemmerle et al., 2016). By contrast, our results suggest that on species' nonbreeding grounds, large-scale land cover change is of greater importance. Our finding that there has been an approximately 5% reduction in the extent of suitable land cover across long-distance migrants, sub-Saharan non-breeding ranges, but an approximately 2.5% increase for short-distance migrants, indicates that this results may be partially driven by extensive land cover change across Africa.
Since the 1970s, there have been extensive changes in land management and land cover across Africa, but in particular the Sahel region (Cour, 2001), with large-scale expansion and intensification of agriculture, loss of wetlands as a consequence of dam construction and irrigation, and deforestation from clearance for agriculture, wood fuel and grazing (Adams, Small, & Vickery, 2014). Previously, it has been widely assumed that these large-scale changes in land cover are linked to the population declines of Afro-Palaearctic migrant birds on their breeding grounds Vickery et al., 2014). Until now, however, evidence for this has been mixed and largely limited to smaller-scale field studies .
Our results provide evidence for a link between large-scale changes F I G U R E 3 Relationship between the annual percentage change in population size and trend in climate suitability (CST) across the breeding ground ranges of 61 short-distance migrant species and 39 long-distance migrant species. Lines indicate the mean non-standardized effect size from across the 100 separate MCMC generalized linear mixed models, while shaded areas indicate the 95% credible intervals. Solid line indicates a relationship that is significantly different from zero (i.e. p < .05), while dashed line indicates a relationship non-significantly different from zero in land cover across migratory species' non-breeding ranges and their breeding population trends.
The apparent low importance of land cover suitability on species' breeding grounds may be a consequence of the high spatial and temporal heterogeneity in large-scale changes in land cover across Europe. Changes in the extent of broad habitat categories across Europe show distinct spatial patterns with, for example, hotspots of cropland decline occurring mostly in Eastern Europe but cropland expansion occurring mostly in North-West Europe (Kuemmerle et al., 2016). When assessed across a large scale, these contrasting patterns may counteract one another to reduce the apparent importance of large-scale land cover change in driving Europe-wide species' population trends. The spatial heterogeneity in the population monitoring periods, with most Eastern European countries having shorter monitoring periods than countries in the west, may further exacerbate these effects. In addition, our measure of land cover suitability only measures large-scale changes in broad habitat classifications and not changes in the intensity of land use management. Despite, as mentioned above, some large-scale changes in the extent of some land cover types, since the 1950s, land use change in Europe has predominantly occurred along intensification gradients (Kuemmerle et al., 2016;Rounsevell et al., 2012); these changes would not be reflected in our measure of land cover suitability.
The lack of a climate signal at non-breeding grounds on breeding populations may be influenced, in part, by the fact that some species in our analysis are somewhat itinerant during the non-breeding season, potentially overwintering in multiple discrete areas during different periods of the non-breeding season. For example, some sub-Saharan migrants, such as willow warblers (Phylloscopus trochilus), move southwards through the non-breeding range over the course of the European winter (Cresswell, Boyd, & Stevens, 2008;Salewski & Jones, 2006). With such species only using non-breeding ranges transiently, mean climatic conditions across non-breeding areas for the entire period may provide misleading estimates of climate suitability. Further data on the intra-seasonal use of the non-breeding range by individual species are required to develop a more nuanced measure of changing climatic suitability in dynamic non-breeding ranges. Such data are not currently available for most species. Migratory connectivity, the extent to which migrant populations spread out and mix during the non-breeding season, may also have influenced this result. Low migratory connectivity, where individual populations of a species spread out and mix over a larger area during the non-breeding season, is common in long-distance migrants (Lemke et al., 2013). Low connectivity should provide greater resilience to climate change by facilitating rapid range shifts (Gilroy et al., 2016). However, if the extent of suitable habitat declines, the proportion of populations that spread out over a large area during the non-breeding season, and that reach the remaining suitable habitat, will also decline (Finch et al., 2017). This may, in turn, decouple the relationship between climate suitability trends on the non-breeding ranges and population trends. Currently, we lack the data required to model the precise relationship between population trends and migratory connectivity (Finch et al., 2017).
The weak evidence for a link between trends in non-breeding climate suitability and long-term population trends emerges despite the wealth of published evidence linking fluctuations in non-breeding grounds' weather conditions, particularly rainfall, to overwinter survival (Johnston et al., 2016) and population size (Ockendon, Johnston, & Baillie, 2014). Although events such as droughts have substantial impacts on species, their irregular occurrence means that the overall relationship between long-term climatic conditions on the non-breeding grounds and migrant population trends is weak (Nevoux et al., 2008a(Nevoux et al., ,2008b. In addition, given that we focus on breeding ground population trends, the relationship with conditions on the breeding ranges is likely to be stronger than with conditions on non-breeding ranges. Non-breeding ranges may encompass areas utilized by populations that breed outside of the area covered by PECBMS, and so, climate suitability evaluated across the entire non-breeding range may be less directly linked to trends in breeding populations. This relationship may be further diluted for species containing populations that demonstrate different migratory behaviours. In such cases, the likely stronger relationship between population trends and conditions on the breeding grounds for more resident populations may mask the relationship with conditions on the non-breeding grounds of migratory populations. This result may also, in part, be driven by our use of gridded species distribution data. Species' range maps are often derived from point observations and delineate broad range boundaries (extent of occurrence, sensu Gaston, 2003) rather than maps of occupancy. Given the discrepancies in recording efforts between Europe and Africa, there is an increased probability of false positives in species' gridded data across non-breeding grounds. However, by analysing our data at a relatively coarse scale, our characterizations of species' distributions are less prone to false positives (Hurlbert & Jetz, 2007). Furthermore, our results were also robust to the inclusion of sampling effort, suggesting that the weak relationship between non-breeding ground CST and the long-term population trends of European breeding migratory birds was not a consequence of variable sampling effort across species' ranges.
Conditions across the Sahel, a key environment in the non-breeding season for many long-distance trans-Saharan migrants, are intricately linked to seasonal precipitation (Ockendon et al., 2014;Vickery et al., 2014). When rainfall in the Sahel is higher, resources for migrant birds are more plentiful. From the 1960s, drought conditions predominated in the Sahel (Nicholson, 2000). Although rainfall across some of this region has increased since the 1990s (Fontaine, Roucou, Gaetani, & Marteau, 2011), there are strong regional differences in the extent and direction of this "re-greening" trend (Kaptué, Prihodko, & Hanan, 2015). Habitat changes resulting from the lengthy period of drought have been widely linked to population declines of long-distance migrant birds, observed since the 1970s . Improving conditions since the 1990s may have led to the stabilization of, and increases in, some populations Nevoux et al., 2008aNevoux et al., ,2008b. In addition, spatial heterogeneity in changes in climate across Africa may have resulted in contrasting directional population trends between species that overwinter in different regions. For example, it has been shown that species that winter in the humid tropical zone have recently declined, while migrants that winter in the arid Sahelian zone show more stable population trends (Ockendon et al., 2012;Thaxter et al., 2010).
These contrasting patterns may counteract one another to weaken the apparent relationship between the population trends of migratory birds and long-term trends in climate on their non-breeding grounds. Thus, while there is strong evidence that annual fluctuations or short-term trends in the abundance of migrants are linked to precipitation across the non-breeding grounds, our results suggest that the long-term population trends of most long-distance migrants are more strongly associated with long-term trends in climate suitability on their breeding ranges than on their wintering grounds.

| The effect of breeding ground climate suitability on trends of long-distance versus shortdistance migrants
The relationship between CST across migrant species' breeding grounds and their population trends is strongly dependent upon migratory strategy. While improving trends in climate suitability across the breeding grounds correspond with increasing population trends for short-distance migrants, this was not the case for longdistance migrants. A number of mechanisms could underlie this pattern. Long-distance migrants are likely to be more susceptible to the potential effects of phenological mismatch upon their reproductive success than short-distance migrants (Møller et al., 2008;Saether & Engen, 2010). Because of the later breeding, and potentially slower rates of nesting advancement in some long-distance migrants, an increasing proportion of individuals appear to be mis-timing their breeding relative to peak availability of key prey species, which has advanced in response to warming spring conditions on the breeding grounds (Both et al., 2006(Both et al., , 2010Mayor et al., 2017;Saino et al., 2011). However, the evidence for phenological mismatch driving large-scale declines in the breeding success of mismatched migrants is limited. This suggests that there is not a causal link between changes in migrant arrival times and long-term population trends or that any such link occurs through changes in post-breeding survival rates, rather than productivity (Franks et al., 2017). Climate change may also disrupt competitive relationships on the breeding grounds (Bohning-Gaese & Bauer, 1996;Vickery et al., 2014).
It has been suggested that warmer winters may improve overwinter survival of short-distance migrants that remain in Europe, while warmer springs may also boost their productivity . This may account for the close relationship between climate suitability trends on the breeding grounds and short-distance migrant population trends, particularly given that the breeding and non-breeding ranges of short-distance migrants are closer than their long-distance migrant counterparts. Increases in resident and short-distance migrant populations may lead to long-distant migrants facing greater competition for resources in breeding areas (Both & Visser, 2001).

| Explaining the population trends of short-and long-distance migrants
Breeding and non-breeding ground CST were better at explaining the population trends of short-distance migrants compared with long-distance migrants (our models explained only 15% of variation in trends for the latter). This may be suggestive of other factors driving population trends of long-distance migrants, or that our ability to detect the effect of CST on long-distance migrants was reduced by uncertainty over their winter distributions and their potentially more dispersed distribution at this time, as outlined above. Given the complexity of the annual cycle of long-distance migrants, it is perhaps unsurprising that CST and habitat only explained a small proportion of the variance in the population trends of these birds (Robinson et al., 2009). By using multiple landscapes throughout the annual cycle, long-distance migrants are more likely to be exposed to land use changes during at least part of the year than are other species (Runge et al., 2014). Furthermore, their use of stopover sites during migration can expose long-distance migrants to a greater number of location-specific threats, such as habitat degradation and hunting (Runge et al., 2014). Long-range movements made during extended migrations also increase the likelihood and associated risks of encountering new, unfamiliar environments (Rotics et al., 2017).
The enhanced energetic costs of long-distance migration may also have a detrimental effect on the survival of Afro-Palaearctic migrants, with possible carry-over effects (Ryan Norris & Marra, 2007).
Indeed, it has been suggested that there is an inverse relationship between energy expenditure and survival in birds (Sala, Wilson, & Quintana, 2015). Notably, some species, including Eurasian blackcap (Sylvia atricapilla) and common chiffchaff (Phylloscopus collybita), have shown changes in their migratory behaviours, with short-stopping and overwintering closer to breeding grounds becoming increasingly common (Elmberg, Hessel, Fox, & Dalby, 2014;Sutherland, 1998), often resulting in markedly improved population trends (Berthold & Terrill, 1988;Rotics et al., 2017;Visser, Perdeck, van Balen, & Both, 2009). It may also be that the variables included in our models are better proxies for the conditions experienced during migrations and on stopovers by short-rather than long-distance migrants. As noted above, the breeding and non-breeding ranges of short-distance migrants are closer together than those of longdistance migrants. Consequently, the conditions that short-distance migrants experience during migration may be, in part, captured by our measures of breeding and non-breeding ground climate suitability and land cover. In contrast, the same measures are less likely to encapsulate the conditions experienced by long-distance migrants during their migrations.

| Future directions
Detailed knowledge of the destinations of birds that migrate beyond Europe during the non-breeding season, and the specific habitats that they utilize, is currently lacking for the vast majority of passerine species included in our analysis, especially in terms of linking specific breeding populations to non-breeding localities.
Furthermore, we lack nuanced data on trends in the conditions of different habitat types, particularly on the non-breeding grounds.
Consequently, population-specific drivers of migrant declines are little studied at present (Cresswell, 2014). Recent advances in tracking, remote sensing technology and the development of genoscapes are enhancing our understanding of migration by providing information on migrant non-breeding ranges, and on the connectivity of breeding and non-breeding sites (Finch et al., 2015;Renfrew et al., 2013;Robinson et al., 2010;Ruegg et al., 2014;Trierweiler et al., 2014), enabling migration routes to be linked to adult survival and population trends (Hewson et al. 2016). Currently, however, such data are available for only relatively small numbers of individuals, and for few species and populations. Expansion of this monitoring will enhance our understanding of specific habitat use on the non-breeding grounds and of linkages between migratory end-points .
Our pan-European analysis provides valuable insight into the largescale drivers of the population trends of migratory breeding birds.
Future linkage of population-specific breeding and wintering areas, migration routes and stopover locations, as well as improved data on trends in habitat conditions, will enable a better understanding of the mechanisms driving the declines of long-distance migrants at a finer scale (Ockendon et al., 2012).
We have demonstrated that, at a continental scale, the population trends of European breeding migratory birds are more closely associated with long-term climate change than land cover changes on their breeding grounds, but the converse is true on their non-breeding grounds. Importantly, we have shown that, in recent decades, long-distance migrants have benefited less than short-distance migrants from any improvements in climate suitability across their breeding grounds. Nonetheless, our understanding of the drivers of population trends of long-distance migrants remains incomplete because we lack crucial information on migration stopover locations (and associated threats) and on the linkages between breeding and non-breeding localities.
Population-level studies into the use of non-breeding areas by migrant species are urgently required if we are to understand the drivers of migrant population dynamics and to prevent, or reverse, further declines.

This work was supported by the National Environment Research
Council grant NE/T001038/1 and, in part, by a previous NERC CASE studentship NERCNE/J500215/1 to CH (with the British Trust for Ornithology as CASE partner). CH was supervised by SGW, PAS and JPH. We would like to thank the EBCC and the many thousands of amateur ornithologists who contributed to the data collection. The Pan-European Common Bird Monitoring Scheme is part-funded by the Royal Society for the Protection of Birds and the European Commission. We thank Prof. Will Cresswell for helpful insight and comments on an earlier draft of the manuscript.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13144.